Watch on Youtube, Apple Podcasts, or Spotify.

Introduction

After having written long-form essays over a weirdly diverse number of areas of the life-sciences, I am increasingly confident in my status as someone who knows a little about a lot of things. But every now and then, you meet someone who casually reveals to you an entire subfield who, up until your conversation with them, you’d never even thought of before. This happened to me when I met Sterling a few months back. We met in the elevator as we were both leaving an event, and by the time we’d reached the bottom floor, the conversation had become so interesting that we stood in the lobby for an hour as I pestered him with more and more questions.

Sterling runs a company called Iku Bio. Iku ostensibly does something quite simple: it helps biologics manufacturers figure out what to feed their cells. This is called media optimization, and it is done in an astonishingly old-fashioned way. An engineer runs a handful of experiments in a benchtop bioreactor the size of a Fiji water bottle, waits days for analytical results, and repeats, maybe three or four times before timelines force them to stop searching.

Sterling’s solution was to use printed circuit boards (PCBs)—the same green wafers inside your phone and your microwave—as the substrate for microfluidic bioreactors. Because PCBs are made via lithography, you get complexity for free. Because they’re already mass-manufactured at planetary scale, you inherit sixty years of cost optimization. And because they’re literally designed to carry electrical signals, you can embed sensors directly into the thing rather than cramming them in after the fact.

The result is a device that costs $8 per experimental lane versus $20,000 for the nearest comparable microfluidic system. And there are many, many ways for to improve from here on out.

This conversation covers the full stack: what cell culture media actually is and why it’s so much more than sugar water, why biologics manufacturing has more in common with semiconductor fabs than chemistry labs, how Sterling arrived at PCBs, and at the end of the talk, why he thinks a fair bit of lab automation is “philosophically a crime.”

Timestamps

[00:00:48] Introduction

[00:01:26] What is Iku Bio?

[00:05:00] Media optimization as the biggest lever

[00:06:23] What actually is media?

[00:13:07] Fetal bovine serum and the move to synthetic media

[00:15:10] Walk me through a media optimization workflow

[00:18:49] Why biologics manufacturing is closer to semiconductors than chemistry

[00:21:50] Matching the phase three batch and generics

[00:24:12] The 200-dimensional search space

[00:37:02] Printed circuit boards as a medium for microfluidics, and the utility of lithography

[00:40:48] Anatomy of the Iku device

[00:57:09] What sensors are on the device today?

[01:01:36] How do you use the Iku device to perform media optimization?

[01:14:44] Does media optimization survive scale-up?

[01:24:32] $8/lane vs. $20,000/lane: the economic utility of Iku’s device

[01:32:05] Why PCB microfluidics didn’t exist 10 years ago

[01:39:24] Who is the customer?

[01:43:14] What is the ultimate goal of Iku?

[01:49:07] What does the validation evidence need to look like?

[01:52:14] What would you do with $100M equity-free?

[01:57:31] Lab automation is in a strange place right now

Transcript

[00:00:48] Introduction

Abhi: Today my guest is Sterling Hooten. Sterling is the founder of Iku Bio, where he is building a microfluidic bioreactor built on a printed circuit board that cultures, senses, and streams biological data in real time, claiming 10,000x higher experimental throughput at a 100x lower cost. It is one of the most niche areas of wet lab automation that I think I’ve ever discussed on this podcast, and I don’t think I would’ve ever learned about it had I not stumbled across Sterling at an event a few months back where we had a conversation that was so fascinating that I immediately wished we had filmed it. Sterling, welcome to the podcast.

Sterling: Thank you for having me. Very big fan. Really enjoy your articles.

[00:01:26] What is Iku Bio?

Abhi: Thank you. So I’ve given a brief introduction of what you’re working on at Iku, but I’m sure I oversimplified some things. I’d like to hear your own pitch for what you’re doing there and why is it so valuable.

Sterling: So the largest problems of the 21st century — things in medicine, for climate, for material optimization — all of these are predicated on our ability to manipulate and control living matter. So advancing our understanding of biology is just so fundamental to these problems in the future, and yet the tools that we use right now to interact with biology are primitive. They’re primitive in an absolute sense, and they’re primitive in a relative sense to what we could be doing. At its core, biology is time varying, it’s parallel, and it’s sensitive. And yet the tools that we use right now — that interface destroys at least one of those properties. And in principle, advances in AI also would be an excellent connection with biology. But that interface is fundamentally broken. So lab automation right now is stuck at the Petri dish and the microtiter plate level. It’s equivalent to handwriting manuscripts in the 15th century, sometimes. And so what we’re building is a printing press for biological data. And the way that we’re doing that is we’re rethinking that interface between compute and biology, and we’re replacing traditional microfluidics with a printed circuit board that allows you to embed the fluidics — cells can live inside of it. And that allows you to communicate and control cells in a way that has not been possible before at high throughput. And the largest application that we see for that is in biologics manufacturing. Right now, biologics — it’s a half a trillion dollar industry and it’s supply limited. So every year, Samsung Biologics has to build a new $400 million facility. The reason they’re doing that is because you can only get so much out of a traditional fab plant. They’re closer to silicon fabs actually. And the largest lever that they have is in yield — so how much can you get out of these things, are they producing, and also what are the costs. The core of that comes down to literally how many of these dynamic cell culture experiments can you run. And that’s a process called media optimization. And it ends up that that one problem ends up being connected to this half a trillion dollar industry.

[00:05:00] Media optimization as the biggest lever

Abhi: So to paraphrase, if I wanted to increase biologics manufacturing by an order of magnitude — at least my capacity to produce like antibodies and the like — the lever that is most easily pushed on and most likely to give you the most bang for your buck is media optimization.

Sterling: It is the most bang for your buck. You are unlikely to get 10x on that. What you’re looking at is how much can I produce per unit time, and then how consistent is that. And if you can produce more per unit time, you get higher throughput for the entire facility. And then if you have more stability in the product — for biologics and for things that go in our bodies — that’s a desirable outcome.

Abhi: And so my conception of these bioreactors that are producing antibodies is you have a bunch of CHO cells maybe sitting in a very large tank. They’re sitting in a fluid of media and they’re constantly just excreting out these antibodies that are later purified. Iku comes in at the step of deciding what media to actually put into this tank. Is that fair to say?

Sterling: Correct. Yeah.

Abhi: What is — well, like I’ve never worked in a wet lab before.

[00:06:23] What actually is media?

Abhi: My conception of media is that it is sugar water that cells are generally fine with drinking up. I’ve learned that this is incorrect and I’d like to hear your take for what actually is media.

Sterling: I would say that that is a very limited view of what media is — not incorrect in that, if we were talking about media for growing yeast, sugar in water is pretty close to sufficient. But the more powerful way of thinking about media is that it is a very high dimensional control surface for what you can get cells to do, right? Cellular communication comes through things in the media, right? The media actually is the communication channel in a sense between cells. It’s also what carries nutrients into the cells. In mammalian cell culture, it’s closer to serum in blood. So it has either many different types of proteins in it. It’ll have different metabolites. It’ll have salts. In defined media it’ll have buffers to keep the pH. It basically has a lot of components — and there are hundreds of them really, down to things like magnesium. And each of these are really communicating and interacting with the cells. And they also work across different time periods. So you’ll have growth media, which is when you’re building up the cells, and then there’s media when you really just want them producing these particular things. And right now, if you buy or produce media internally, it tends to be connected to a particular clone or particular cell line. And so you will optimize the media for that particular cell line, or you’ll optimize media for — if you’re growing neurons. And so every — it’s complicated enough and important enough to the results that you get that exploring it is very valuable.

Abhi: Like I know that there are a few companies that have popped up claiming to technically redesign cell lines to make them better at biologics manufacturing. Does that also demand a change in media?

Sterling: It can demand — the key thing is that the biologics that we are producing now are becoming more complicated, and that is making media optimization more difficult. So you do tend to pair the cell line with a media line, both for repeatability and ease of use, also just for commercial reasons — that’s a better business. But you can — what really happens is you tend to take a standard growth media or something off the shelf, and then you will customize it for this particular thing that you’re trying to make. Because ultimately, productivity is really the interaction of these three or four things: it’s the cell line, it’s the media, it’s the process conditions or the tank that you put it in, and then the actual compound of interest and things that you’re trying to do.

Abhi: You mentioned earlier about like media is both a way — like nutrients for the cell — but is also the substrate upon which they actually communicate with each other. That second part was surprising to me. I did not naturally conceptualize cells in a tank actually talking to each other while they’re churning out antibodies. What are they communicating exactly? Does that question make sense?

Sterling: I think it’s maybe easier to think about it in the sense of our bodies, right? Cells will send out or communicate through different hormones, right? Those will get released. There are small signaling molecules that get broadcast — those are carried through the media. Well, in the body we call it blood serum, right? But in the sense, it’s media.

Abhi: You mentioned also that you have different stages of media that you want to introduce to the cells depending on the cell’s actual life cycle. Is that also true for serum in the human body? Does the body constantly adjust its own serum to whatever the cells need?

Sterling: Yeah. I mean, that is the way that cells differentiate, in a way. You’ve got some gradient that will happen, and then that gradient — that’s basically saying you’ve got different media, and that gradient can tell cells how to orient or can tell cells how to develop. And from stem cells, triggering when — what they’re going to end up being — that’s also basically — it becomes media as you add things into the cell environment there.

Abhi: So why — what’s stopping me from just replicating human serum for mammalian cells? Is that not the best substrate to use?

Sterling: Well, the first question is, where are you gonna get it?

Abhi: Well — I guess this is a more basic question. Do we understand human serum well enough to perfectly replicate it?

Sterling: Replicate it? I don’t know. What I will say — and that gets closer to what you were talking about originally — is that’s what we’ve been doing historically. But instead of using humans, which — not that — very limited supply, or limited willing supply —

[00:13:07] Fetal bovine serum and the move to synthetic media

Sterling: we’ve been using fetal bovine serum, so from calves. There are problems with that. It is highly variable. And for all of biologics manufacturing, the goal is reduce variability. And if one of your largest inputs is variable, that’s a problem. It’s also a challenge because things like — you can’t sterilize it in the traditional way. You can filter it, but you can’t heat it up without destroying — and things like prions, which could be quite bad, you would need to prevent those coming in. So the industry has really moved much towards formulated medias. So you’re building it up from the constituent parts, and that also allows you to — it reduces variation and gives you a lot more control over how you are particularly tuning that media.

Abhi: When you say like at some point fetal bovine serum was being used —

Sterling: Still. It is still in use. It’s mainly in use in research. I think — I’m — maybe there are some biologics manufacturers who are using fetal bovine serum. I don’t know. But I think the industry has pretty much moved to —

Abhi: At this point, would you consider that the synthetic serums that are attempting to recapitulate the biochemical properties of fetal bovine serum — the synthetic stuff is better? Or is it just like it’s easier to get, so you’re okay with not perfectly recapturing fetal bovine serum?

Sterling: I think it’s better.

Abhi: Okay.

Sterling: I think it’s better, and I think it’s better in that you again get to tune it.

Abhi: And so attempting to be more concrete about —

[00:15:10] Walk me through a media optimization workflow

Abhi: what is a media optimization engineer exactly doing? Let’s say I have a plate of CHO cells. I want to produce Keytruda, so pembro. I have a bunch of cells. I have all of them willing to produce the drug. They’ve been genetically edited to do that. What’s the next step?

Sterling: So the process in general is guess and check. So you will take a cell line that you’ve edited or produced for this. Most of the time it’s just — and then you’ll take it out from the freezer. You’re gonna grow it up a little bit. And then you will probably take four or five of those because you don’t kind of know yet, right — which particular strain will do best.

Abhi: So you’re trying with multiple strains.

Sterling: You’re gonna try with multiple strains. And then you will run experiments that allow you to — first you’re gonna run in microtiter plates normally, right. And you’re going to just see where are we, which of these cell lines seems like it fits best with these. After you’ve narrowed it down, you’re going to move to something that has more control. And the reason that you’re gonna move to something that has more control is that what happens in a microtiter plate is extremely disconnected from what happens in any kind of production environment. And the core reason for that has to do with flow. So in a microtiter plate, you get a lot of capillary issues, right? It changes the — you’ve got the surface tension kind of comes up, that changes the gas exchange rates. You get evaporation. And you don’t get any of the different gradients or different little bits of shear forces — all these things that actually affect how cells grow in large reactors. So what you do is you put it into what’s called a benchtop bioreactor. And so this is a little bit bigger than a Fiji bottle in terms of what it’ll contain, and it’s got an impeller in there and it’ll spin it around. So now you’re going to grow those cells in that media for 10 days or something, right? And during that time, you’re going to also change or control the pH level that’s in there. You’re going to control the temperature. You’ll set different impeller rates, seeing what’s optimal. And you’re going to run that for — one person can maybe run 12 of those experiments, 15 of those experiments. It’s pretty laborious right now to actually set those up. It’s gonna run, and during that time, you’re gonna pull off some samples. You’ll take those to the analytics section, depending on how booked up that is — that could be three days to a week sometimes to get all of your answers there. And then you’ll do that.

[00:18:49] Why biologics manufacturing is closer to semiconductors than chemistry

Abhi: I’m sorry, what questions are you asking at that point? What are the samples meant to answer?

Sterling: So ultimately, your sample is meant to answer how much total biologic did we produce in here, at what quality, right? And then the other question there is how overall — how consistent is it? Will it be — that’s actually a large sort of hidden cost, as I said. The best way to think about biologics manufacturing is to think about it as high precision manufacturing, closer to semiconductor manufacturing. That’s really the reason why Samsung Biologics is in the position that they are — because they took what they learned in terms of process control and brought that over. The reason that Fujifilm is a large manufacturer is because they took chemical process engineering and brought it over. Now, these were not biological companies, right? They are industrial manufacturing companies. And when you think about reducing process variability, one way of looking at that is how precise is the part that comes out. But then what makes up that, right, is like how much variation can we absorb without it affecting the end product? And so if you can come up with media and process conditions that are more forgiving, you’re relaxing it a bit, right? You can still end up with something that’s very precise at the end, but oh, we didn’t actually need as much — we were more forgiving over here. And that can be important because if you lose a batch of biologics, it’s very expensive. And that can happen. And it does happen. And so the way to reduce that is through media optimization. And so to finish on this — you’ve run that set of experiments, you’ve got your readout there. And those readouts, although those are the most important, you’re also going to characterize kind of everything in there that you can, because you want to see how those are affecting that actual result. Then you will repeat this. And depending on how much time you have, maybe you will get three or four runs at that, and then that’s it. And that comes down for biologics manufacturing to the regulatory reasons.

[00:21:50] Matching the phase three batch and generics

Abhi: So how much of — would you say the optimal cell lines and the optimal media — it’s like there is a threshold of quality you want to meet and after that you’re done, versus you are trying to make this as perfect as possible? Is it kind of dependent on what drug you’re trying to produce?

Sterling: I think the goal is match what was in the phase three trials. So in the process of taking a drug to market, during your phase three trials, the batch that you produced there — that is what all of the FDA’s evaluation was based on. So they want to keep that the same. So anything that deviates from that is undesirable.

Abhi: Is this true even when the drug goes off patent and the generics manufacturers — are they trying to make it even — they’re trying to improve the process even more, or even for them, they’re trying to replicate exactly what went on with the original company?

Sterling: That is a great question. I should look into that because — no, truly, because they do have to go through — so they have a couple options. The first thing is that they will basically just license the cell line and the media from the existing pharma company, right? Pay them for that. And then that way the pharma company can still get some revenue from that. The alternative is they need to come up with their own cell line and — I think the regulations are such that there’s a way of — I think it’s like if you can prove that it’s similar enough, then it just counts as a process change.

[00:24:12] The 200-dimensional search space

Abhi: And getting back to the question of actual media optimization — the media optimization person goes to the analytical chemist. The chemist tells you all you need to know about the samples that you’ve been given. You repeat this five to six times. What are the levers of change that you have over the media?

Sterling: So media is best thought of as this control surface for affecting what the cells are doing. What are the levers in there? You can change the components, and then you can change the concentration of those components, and then you can change timing of those things. And if you start with 200 or more — let’s start with 200 components that you could put in there, and then the different concentrations that they come in, and then the timing — that already is quite a large space to explore. Then you have that interacting with the cell and the different cell lines — larger space. And then with that fixed compound that you’re looking for. So the standard things that people are going to change or tune, right, is when is a carbon source coming in, and when — as you start producing different proteins, the needs of the cell change. So if you shift into a different mode for the cell — you can signal it to shift into a different mode, starts producing these other — all of a sudden its needs change.

Abhi: Mm-hmm.

Sterling: And being able to anticipate, buffer, and meet those needs — that then has a lot to do with the output.

Abhi: How much of the optimization — like even the direction or specifics of the optimization — can be theoretically known and applied versus just always empirically determined? I guess the more specific question I’m asking is, does a media optimization engineer — are they coming to every new problem almost like tabula rasa? Whatever experience they had in the past does not apply to this new cell line with this new drug.

Sterling: So the question of how tractable is this of a problem and what’s the current state of the art — the current state of the art is that best practices live in the mind of the practitioners. And a lot of that comes down to familiarity with that cell line, familiarity with the media they already have. And most manufacturers are working in a particular kind of domain or specialty, right? And so as you’re constraining that search space, it does make it easier to operate in there. However, it is not the case that you will one-shot it coming through. And then the second thing is, it’s actually reasonably easy to get caught in a local maxima. And if the cost of running those experiments or experiments themselves are sort of precious, you’re really not going to push very far out. The lever they currently use is mainly in strain engineering. And so they’ll try to select strains that’ll have the highest performance. But once those cells that you’re using are set, it does all come down to the media for optimization. In a model sense, it does seem that it’s tractable. It does seem like there’s transfer learning. How broad that really comes down to what experiments have we been able to feed into these models so far? And the answer is not very many. The largest facility that I know of for running sort of like dynamic cell culture experiments — they can run like 300.

Abhi: In parallel at any given time?

Sterling: Yeah. 300. And that’s like, the entire company is just doing that. So that’s the state of the art. And a lot of that comes back to the fact that it’s so manual.

Abhi: So the one last question I have before we move on to how Iku is fixing this — I can understand being able to easily modify concentration of the media. I understand being able to modify the timing of when you’re giving which media to the cell line. The components, the constituent components, feels a lot more complicated. Because that’s like 200 components. How much of that is like — in practice there’s 10 of them you modify at any given time, and the other 190 are pretty standard and all cell lines will need this.

Sterling: Yeah. So how much is like — what’s the core? Is there some —

Abhi: Dimensionality reduction?

Sterling: Yeah, like is there an 80/20 thing going on? Oh yeah, absolutely. Absolutely. Which, as I said, the glucose — your sugar source or carbon source, energy, the pH that you’re running at — those are, yeah, there probably are 10 that are dominating. But that’s why it’s actually so challenging — because there are 10 that are dominating, but because the system that we’re controlling is quite non-linear, it can amplify what are sometimes in certain conditions some small change. And my favorite example of this is that — this was in industrial manufacturing — but changing the amount, just changing the amount of magnesium at a particular point doubled the output. And it didn’t necessarily need — there was no a priori way of knowing that it would’ve been magnesium that went in there. And you can say, oh, okay, sure, that’s a lever and we should do that on each of these. But the problem is that potential exists for all of those other 190 things, right? So it’s like, sure, there are these core things that tend to dominate —

Abhi: But those 10 things could vary based on what the problem actually is.

Sterling: Yeah. Well, those core things of like — you do need to, the salts that are in there, right, and when energy comes into the system — those are definitely floor level. You have to figure those out. But then — and if you get those wrong, basically those are controlling the — where the floor is. So if you get those wrong, it kind of doesn’t matter what you do in these other areas. You’re not going to have high performance. But just because you get those right doesn’t mean that you have high performance at all. They’re just table stakes. You need to get those done.

Abhi: That makes sense. And so we mentioned this engineer who’s trying to produce Keytruda.

Sterling: Sure.

Abhi: They’re evidently building, at the very beginning, in a Fiji-shaped bioreactor.

Sterling: Yep.

Abhi: Doing these rounds of iteration, trying to get to something good. What is Iku’s proposal for a better way to do it?

Sterling: Our proposal is to rethink what it is that you’re trying to do when you run that experiment. So that Fiji bottle device gets used for two purposes, one of which is you want to grow cells and you want to grow them to feed a seed train. So you’re growing them, or you need that quantity of those cells. That’s one. And the second is that you need information and you need to be able to control the environment that the cells are in over time in order to get it. And so for this first set of things where you’re trying to grow a lot of cells or grow them up — great, perfect use for it. If you’re trying to extract the most amount of information and trying to control the cells, it’s a very limited way of doing it. Before starting on any of this, I’d actually seen some of these benchtop reactors and I asked them — if the thesis is that it gets better when you go smaller, why did you stop at the Fiji bottle? And the answer was, well, if we go any smaller, our sensors won’t fit. And that’s because they’re using off-the-shelf sensors. And if you ever see a photo of these things, it’s a hodgepodge of different things that have been kind of crammed in there. And that literally is — doing sensor design is its own field. And you need to design not just one type of sensor. You need to design many different types of sensors. And there’s also not that much of a benefit going from a Fiji bottle to half a Fiji bottle in size because of the manual labor and all these things. So our solution is to think about what’s actually the best platform for building sensors, and then can you put cells inside of it? And my last company was a robotics company. Any of the humanoids now that you see going on — I’m highly skeptical of the economics on these things — but any of the humanoids that you see, the core technology that enables them to move and interact with the environment — that was what we built. And that is a sensor problem. And it’s a sensor in a high-noise environment. And that is abstractly quite close to what we’re doing in biology, right? So the idea is, if you have a good place for building and placing sensors of different types around, now you’ve reduced the problem. And so, easy place to build sensors — now you just have to figure out how to grow cells inside of it and keep them alive. And if you pick a mass-manufacturable technique for doing that, it also solves some of the scaling problems. Because the challenge with controllable systems right now is that they still literally require somebody to come over, unhook everything, set it up. You can use disposables to take that down a bit. But it also takes — when you go larger, it takes more media. It’s more expensive to run it. It’s less repeatable. None of it makes sense except that it’s a difficult engineering problem.

Abhi: In a practical sense — I can buy that this form factor was chosen purely because our sensors aren’t small enough to fit in something smaller. What is the form factor that you guys have?

[00:37:02] Printed circuit boards as a medium for microfluidics, and the utility of lithography

Sterling: So the core differentiator is that we are reusing printed circuit boards, which are ubiquitous. They are in your phone, in your microwave. And we put microfluidic channels inside of them. And by doing that, it allows you to then have cells live inside. They can pass through, they can live inside there. And it turns out that making microfluidics previously that integrate those types of sensors is extremely awkward. And so you either don’t do it, or if you do do it, it’s still hand-finished. And so the big differentiator is everything comes straight from the fabricator ready to go. And this is a theme that has happened before. So in silicon photonics, which is where you take existing silicon fabs and you say, hey, can we use this in a new way? And not just to do integrated circuits, but can we now do things with light in it? Or in your iPhone, it has a light detector. That was a new way of using that. And the core there is that the process that’s used is called lithography, which is where you’ll take a mask, kind of like a snowflake, you project light down through that or something, and that causes certain things to react and certain things not. And lithography is a really powerful manufacturing technique because you get complexity for free. What that means is, normally if you’re doing traditional subtractive manufacturing, as your part gets more complex — you’ve got more nooks and crannies in here — it takes more time to make it, or you’ve got more tool changes, all these things. But with lithography, you pay that cost once. You pay that cost when you make your snowflake. But it actually doesn’t matter how complicated you make the snowflake for what’s down here. And so it pushes you to say, what’s the most complicated thing we can make here that has the most value? Because it literally costs the same. It doesn’t matter if it’s one line through here or some complicated maze. So that’s what semiconductors are doing. Then they apply that to photonics, right? LIDAR — printed circuit boards are made the same way. It’s lithography. And if you can leverage that in more complicated ways, you start both enabling capabilities that weren’t possible before, and also are riding a cost curve that’s really beneficial. So the idea is, every time that we have found as a society a new use for lithography, large industries get built off of that.

Abhi: And sorry, so where’s the lithography component coming in when you’re talking about building a new bioreactor?

Sterling: So the way that we make our chips — which you have, right?

Abhi: Yeah. Let’s — do we? Oh man. Here it comes out pretty small.

Sterling: Yeah.

[00:40:48] Anatomy of the Iku device

Abhi: I am seeing that there’s a bunch of circuits coming on from here. Walk me through the anatomy of this device.

Sterling: Sure. So the first thing is that it looks kind of cohesive, but it’s actually six layers. And each layer either is carrying electrical signals or fluidics, or routing fluids in there. And so for this particular chip, it has a channel that’s a millimeter wide and about a hundred — about the size of a human hair — tall. And that’s actually a great size for cells. And you can flow media and cells into it. And then it has all of the components that a benchtop bioreactor or a more controllable system would have. And the way that you make these is through lithography. So these lines and all of the features that are on here — there’s a snowflake kind of pattern that’s made for that. And then they will put what’s called a resist and an etch on. And so it will keep those lines where you want them and etch away everything else. And then you make the next layer, and then you make the next layer, and then you compress all of those together. And so the way to think about it is, it’s like a 2.5D space. So you’ve got a two-dimensional plane, but you’ve got multiple two-dimensional planes. And so topologically that’s going to allow you to do things like take a spiral and get to the middle, and you need to get out of it. So you can come up and out in a way and around. And it also allows you to put electrodes or different sensors in relation to the fluid, in relation to the cells in different places. And that’s kind of abstract, but let me give you a very concrete example, which would be — if you want to have a readout of electrical signals of heart cells, cardiomyocytes, you want to read across those cells. Well, you need to be able to put electrodes above and below them normally, right? Or you can put them side to side, right? If you’re trying to do these things, that’s like a primitive — that is really, it sounds very simple. And yet I will tell you, that is, with other techniques, a difficult thing to do. And so by switching to this new substrate, a whole class of problems that are traditionally quite difficult become substantially easier.

Abhi: And sorry, I don’t have a great conception of where do the cells — on this green thing, are those holes where you put the cells?

Sterling: It is, it is. And I actually have a drawing I should send to you. You can put up a drawing on this screen.

Abhi: Yeah.

Sterling: Because that is also part of the problem — from the outside it literally looks the same as any printed circuit board. Second thing is, in biotech, a printed circuit board looks like alien technology. But yeah, it has actually small holes. There are ways of getting fluids into the actual device. And then you can run them past sensors, or you can — it’s often easier to run the fluid past the cells. And then you’re kind of reading things out on the fluid.

Abhi: And so there’s not a specific chamber here where the cells sit. They’re literally in a line formation as you run fluids through them.

Sterling: In this particular chip — this particular chip is like a year old. In newer designs, you have more like a chamber. And you’re seeding that chamber and then your cells are growing over it. But the powerful thing about using this technique for making microfluidics is that you can make a large number of variations, and it’s a difficult problem in traditional microfluidics because you would need to make new molds. And a new mold is $25,000, $40,000 — you need to get a mold maker to come in and machine it. Your economics on that mean that you need to make a lot of them. With printed circuit boards, it’s easier to make variations to them and just do it. So we have a core catalog that we’re building — these are the designs for particular applications. But every new printing, it’s relatively easy to change it to whatever the condition is.

Abhi: Sorry, is it fair to say that typically microfluidics are not built using lithography, but you are building them with lithography?

Sterling: Microfluidics historically started with lithography. They were built using similar techniques used for semiconductors. And in most research labs, when people build microfluidics, that’s still the way it’s done.

Okay.

What you’ll do is you will make a silicon mold and then you cast a polymer over it. This polymer is called PDMS. And the desirable properties of it is that it’s optically — not transparent, but you can at least see into it, and it’s gas permeable. And so that allows you to have exchange of gases without — you can put it in an incubator and you can use it there. Downside of that is you can also get evaporation. The problems with that is you end up with a fragile output, and it’s also fairly labor intensive to do that. But people like it because you can do it in your own lab. The difference comes down to the use of lithography for the sensors and fluidic channels together in this thing. And critically, for silicon fabs, you need to be really careful about contaminants. So if you need, for example, a gold-plated electrode, you cannot do that in a silicon fab because you will contaminate — it’s not allowed at all. Very bad. So with the printed circuit board as a medium, basically you can integrate many more different types of sensor modalities than are possible with silicon. And then the second thing is just — the reason to use silicon is because you want extremely fine features and detail. Once you need something on the nanometer scale, it’s kind of the only option. But our thesis is that cells themselves are more on the five-micron scale, which is a few orders of magnitude difference.

Abhi: Yeah.

Sterling: And that’s actually the domain where printed circuit boards are a better place.

Abhi: Is there — so if historically people do use lithography for microfluidics, but they only use it for the channels and not the actual electronics — what innovation allowed you to actually include electronics in the design of the microfluidic?

Sterling: Yeah, so let me state that. Microfluidics is a really broad term. For example, DNA sequencing — Illumina, right? That’s using silicon for a microfluidic system. And doing the sensors. It’s a really useful place for doing that. But it has limitations in terms of where in space you can place things. The example I gave earlier about trying to read across these cardiomyocytes — you can’t do that with silicon. There’s no way to build a channel that size that you need for the cells to go through it, but it’s buried and you have electrodes above — it just — you just can’t make it that way. So the core innovation is, first of all, just conceptually thinking about printed circuit boards as a medium for making microfluidics. I’d been working with circuit boards for 10 years or something. Never occurred to me to put fluidics into them. Been talking to people about this for three years. Never met anybody who was like, oh yeah, I’ve seen that before.

Abhi: So as of today, there’s no one combining circuit boards with microfluidics?

Sterling: Not for — there is for diagnostics.

Abhi: Oh, okay.

Sterling: Yeah. So Professor Moschou at the University of Bath — she’s really the pioneer of putting fluids into the circuit board from the fabricator. And the reason that’s so important — that I keep coming back to it — is you can do a lot of things and, academics are prone to this, you can do a lot of things by hand that does not scale if you need to make hundreds of thousands or a million of things, right? If you’re doing that, you need to pick something that is mass-manufacturable. So in terms of cost and complexity, the cheapest thing to mass-manufacture for microfluidics — it’s either paper or molded things when you build a lot of it. But if you try to make microfluidics in a PCB in a lab, you can do all kinds of weird things. Getting it so that it’s compatible with the standard fabrication process — that’s a different ask, both because they’re not terribly keen on changing their processes for the most part. But then the second thing is that when you do it by hand, you’re introducing variability from the beginning. When you have it done in a fabricator, you’re inheriting the hundreds of billions of dollars that have been spent cumulatively on printed circuit board development. It’s been around for 60 years. Entire industries are built upon it being already very good. So let’s just reuse that thing that’s already quite good and low variability.

Abhi: Could you give me some intuition for how the device is actually put together? So my mental conception of lithography is you’re able to create these very fine channels in the silicon via shining light through a mask. What’s the next step after that? Maybe you do this on multiple layers to have this multi-layered system of channeled —

Sterling: Yeah. So for traditional silicon fabrication, it really is a mask and then you etch and then it’s a mask and you etch and mask and you etch. With printed circuit boards, it’s more like each layer can be made out of different materials. So this is where there’s an enormous amount of flexibility in terms of — it’s a much richer palette to start building out of. So the foundation is what’s called FR-4, which is a fiberglass structure. That’s why they’re normally green. It’s a fiberglass structure. And on top of it, it’ll come coated in a layer of copper, layer of copper on the bottom. And that is the simplest circuit board that you will buy. The cheapest one is just that, and it’s just been etched. And then they will put down what’s called basically a protective layer on it, so that you don’t just scratch off the copper. And then you’ll silkscreen it, which is if you want to put labeling and all these things. But at its core, that’s what the process is. When you add in microfluidics, there are techniques for being able to make the fluidic channels on one layer. And then as you need, you can just stack on another layer, and then that layer has fluidics, or in between them now you can route your heaters, right? You need to put your heaters there. Or if you want to put the electrodes or whatever your end sensor is, you’ll pattern that on that layer and then you sort of build it up and then you stack them together. You close it and then —

Abhi: So in V2 of this device, you have this chamber where the cells live. You have microfluidics connecting this internal chamber — maybe it’s external — to a bunch of pipes that feed in some particular axis of variation that you want to control during the media optimization process. And you also have embedded or maybe external sensors that are connected to the circuit board to have some sort of readout of what’s going on in this chamber where the cells live as the media is being applied. And what’s the output? What do you actually — what is the output of the system? I imagine one is maybe temperature, maybe another is internal humidity. What other axes are there that you can actually get straight off the sensor and straight off the device?

[00:57:09] What sensors are on the device today?

Sterling: So the way to think about it is that if you’re going to do any kind of cell culture, there are a set of table stakes that you need to be able to do in there. And those are temperature, pH, dissolved oxygen — we’re flowing things through, so you need to be able to measure flow rate. And those together — that’s the core set of things that our system is currently reading from. The next layer are the electrochemical sensors. So being able to read impedance is actually very useful. If you can read impedance for the media itself, you can detect some changes in how the media is adapting. And if you place them in relation to the cells, you can also correlate cell growth with impedance, which is based on how these charges sort of end up hitting against cell walls at different frequencies. So that’s a core thing there. You can do conductivity through it, which is partially used for offsetting where the impedance reading is coming from, because it can get interfered with in a lot of ways. And so you sort of need a reference point in order to do that. And then you can do other electrochemical techniques, like cyclic voltammetry. But the readouts right now are the impedance, flow, dissolved oxygen, pH, and temperature.

Abhi: Theoretically, I imagine all of these sensors already had miniaturized versions of them available. Is that true? Not true?

Sterling: Not the case. Not the case. Nothing that our system can do at the moment is anything that you couldn’t have done by hand or with a very custom setup. The challenge is, how do you do more than two of those, three of those, at a time? How do you build them economically? For example, the chip that I showed you, in any kind of reasonable quantities, it’s like $4 or something, $3. And that’s actually still even — you can get it down to less than a dollar on that. So if you’re buying sensors off the shelf, the economics are going to start killing you very quickly. And then the second thing is, it’s a challenge to integrate those things. So a big idea in robotics or engineering — any kind of real system — is that interfaces and connectors are what will kill you. They’re very common points of failure. So the best solution is no connectors. When you build sensors all in the same platform, you essentially get to do it with no connectors. So that’s the trade-off — harder, more difficult engineering from the outset, but lower variability and better economics at the outset.

Abhi: I imagine you get dissolved oxygen, pH, and a few of these other parameters. I imagine there’s still some you’re missing in the sense of — is the protein that I’m expecting to produce actually being produced?

Sterling: Yeah.

[01:01:36] How do you use the Iku device to perform media optimization?

Abhi: So it sounds like you’re allowed to optimize to a threshold and then after that you need the analytical chemist to come back in and do their thing.

Sterling: So our goal is to make the analytical chemist kind of a confirmation rather than be limited by it. And the reason comes down to lessons from control theory. So the first is that any system that you’re trying to control — in this case, cells — if they move at a certain rate or certain speed, and you want to be able to dampen that or amplify it, right? You need to be able to read it fast enough that you can come in and make an intervention. Anytime that you take a sample of something and do an offline measurement, that loop is normally too long, right? Sometimes that loop is five minutes or two minutes — okay, maybe you can work with that. If you need to take something to your analytical chemist, it’s probably hours or days. That information is not useful to you in the actual control of the culture, right? So what you want are real-time sensors. You want sensors that are truly integrated into the thing. For the sensors that we’re using now — that really is just the table stakes to enable us to start building in these other sensors. If you don’t have those core sensors, you can’t even keep the cells alive. There’s just no point. But being able to have live readouts of monoclonal antibodies — that is what we’re building towards in the device. It’s being able to have the optical sensors built in. It’s being able to leverage the biological techniques or chemical biology techniques that we have right now for getting signals out of cells. All of those are compatible with our system. And that’s where I think the real value starts becoming unlocked, because there’s a large difference, sort of philosophically, between just reducing the cost of something versus what questions become askable now. And the questions that become askable and the experiments that you could run — that’s what I think is so powerful about using this substrate as a technique. You make this core thing — can you grow cells in high throughput in this dynamic way? Okay. Once you have that, every new sensor system you put in gives you more lenses into it. And this comes back to why lithography is so powerful — normally you have to make a trade-off, right? Every sensor I put in, it costs me money. And so I’m only going to put in the sensors that I need here. But if it doesn’t cost us anymore, or if it’s basically trivial, then the idea is actually let’s just instrument it. Let’s just keep instrumenting it. And classically you would say, well, I don’t really care about those features and those things. Those things don’t matter. But what we’re moving towards is more of having fewer priors and having less human interpretation on the streams of data that are coming in. And so for example, the impedance sensing does not give you a simple number that comes out. It’s a complex number that comes out. Okay, whatever. You could still deal with that, but there’s a complex number across hundreds of frequencies. So you’re getting back this large readout. And then it’s changing over time. So if you and I try to decode that, it can be difficult, right? And we can argue about this, but machine learning is getting pretty good — arguably quite good at handling those types of things. And so the way that I separate these two — they’re what are called narrow-band sensors, and then they’re broadband sensors. So a narrow-band sensor is, for example, readout on temperature. You’re gonna resolve that to some either resistor variable or some Celsius basis, and you want that to pretty much just respond to temperature, not respond to anything else, right? Very easy thing to interpret. Same way with your lactate — you want something that only responds to the lactate in the media, nothing else coming out. These are narrow-band sensors. They’re meant to reject everything else. And then there are what I’m gonna call these wider-band sensors, which is — if you take a microscope and put it on something, that’s a fairly wideband, right? There’s a lot of stuff going on in there. There’s not just one answer about what’s going on. And you can sort of select — I think these things are more relevant to the questions I’m asking, or not. And things like optical, the impedance, some of these other electrochemical techniques, the magnetic fields that are in there — when you have machine learning on the other end to interpret that, it would be surprising to me that that’s not useful.

Abhi: This is maybe a naive question, but at the end of the day, all the signal you’re able to extract from this device is gonna be some electrical property of the tiny little bioreactor you have in there. Is that correct?

Sterling: No, the big picture is that we’re integrating all of these different modalities. So we are integrating the optical modality. My dream here is to get Raman sensing into — multiplexing Raman sensing across this, right? Having that method of looking at it. It’s having those with the lactate and the glucose and the monoclonal antibody readout, right? Or whatever those domains are — in an instrument sense, that’s extremely powerful. So that’s the goal.

Abhi: Okay. Interesting. I imagine some of these variables — you mentioned — are immediately interpretable. There’s a good value you should be reaching. I imagine dissolved oxygen is one of those. For the more complicated ones where you don’t know whether this is a good value or a bad value — like glucose or some other mineral — where does the ground truth come in? Is that where the analytical chemist comes in and they give one singular data point, like what’s good? And then the purpose of the system is to correlate everything that you put into the system and all these output variables you got out to that ground truth? Or something else?

Sterling: So I think a useful lens for this is from a book called How to Measure Anything. Highly recommend. This book changed my life. And the idea is the expected value of perfect information — that any reduction in uncertainty has some cost to it. So when we’re taking a measurement, there’s an economic aspect to that and therefore a trade-off. So knowing the temperature of this room — there’s not much value to us, right? Doesn’t matter whether we’re off five degrees or 0.1 degrees. For semiconductor manufacturing, matters quite a lot, right? You need really, really tight value there. So if you take that lens and you say — certainly overall, there’s a need to have precision on the readouts of how much antibody do we get out of this, and the quality of that, right? But earlier parts of the process — do you need that level of precision?

Abhi: Well, I guess at the end of the day, I imagine the whole purpose of the process is to get to antibody production. But I guess, is part of what you’re saying that there are earlier intermediate benchmarks you want to hit before you get to the antibody?

Sterling: What I’m saying is that your ultimate readout, right, is yield, titer, quality, and stability over these things. Those are the things you care about. And pretty much in that order. Even on the yield though, you’re still going to get — there’s still variation inherent in cells, right? Every batch you run, even though they’re trying to reduce variability, you’re still going to get some variation in there. So if you take a sample and you learn to two decimal points the titer that came out of that, the yield that came out of that — okay, great. But your process variability is 1% anyway, or something, 2% anyway. So knowing it to three decimal places doesn’t really help you. And then the second part of it is — if every measurement has a cost in some sense, can you change your measurement system such that you get the information that you need in a more economical way? And part of the way of doing that is by loosening constraints when possible. So ultimately, certainly you still need — you’re still gonna run it on your benchtop and your pilot things, and you are going to characterize it there, right? Because you do need ground truth from those things. But in terms of which is the right media or conditions to get to — okay, do you need two decimal points of accuracy on that? Do you need all of those readouts to do it? No.

Abhi: Is a good way of thinking about this — you start with the Iku device at the very beginning, and then once you’re happy with what you see, then you move on to the benchtop device? Allowing you to narrow your search space down to a very small number of parameters.

Sterling: Right. It would basically be like — you’re still going to end up — the process looks pretty much the same. The difference is what is the quality and speed that you came to that answer. What’s the quality of the answer you came to? What’s the speed that you came to it? And then the second part is, how many of those benchtop experiments did you need to run? Because there’s a difference between running them in an exploratory sense versus running them in a validation sense. In a validation sense, you’re just trying to make sure that things are repeatable. So you need to run, let’s say, three to five copies of it or something. But if you’re already quite confident that you’re at the optimal point, it doesn’t make sense to do the exploratory experimentation there anymore.

[01:14:44] Does media optimization survive scale-up?

Abhi: Moving on to — okay, you’ve done the Iku optimization and now it’s time to move on to the bigger things. How worried are you that moving the cells to a physically larger space forces the media optimization to move into a completely different direction?

Sterling: It’s definitely possible, and every time that you change physical shape and geometry, you do get some variation there. The confidence comes from understanding that — first of all, empirically, every microfluidic system that has flow integrated into it ends up correlating quite well with the larger system. The reason that people have hesitation about it is because they think about microfluidics that doesn’t have flow, and the recirculation effects. And that’s actually the key thing, right? It’s a question of, do you have flow in this thing or not? And how does that flow and those shear forces and the oxygen transfer rates and the gradients that you create — how are those representative of what’s going on here? So that’s one part of it. But let’s say you don’t buy any of that. The easier way is that it actually decomposes into two broad parts. There are parameters that change with scale. So these are things like your hydrostatic pressure — definitely changes with scale, right? You’re not getting away from that. Certain mixing times — these change. You can get pockets in very large reactors, right? These change. But then there are a set of parameters that empirically don’t seem to be scale-variant. And for the most part, media optimization seems to be scale-invariant.

Abhi: Do you imagine in the ideal setting that this is a closed-loop system that just continuously tries different media optimization parameters, feeds it all into a model, it plans the next round of media optimization, and that just goes in a loop?

Sterling: Yeah. So how does the — aside from running the experiments, how do you actually interpret and decide with it? So clearly the entire zeitgeist right now is about replacing the control layer with AI and models. And whether you can do that on experimental design from reading a bunch of papers and then this is the thing I’m going to build — I’m less convinced that that’s necessarily the best way. But for these types of experiments, certainly seems the way. It’s actually key for making the whole product, because otherwise you’re handed so much information back that the problem then shifts to processing it. So one of the lessons that I’ve taken from talking to people who have tried things in media optimization, tried doing cloud labs or doing these things — there’s a lot of hesitation around sharing cell lines. Understandable. And it also comes down to information about what the result of those cell lines are. So for example, a company that was running experiments externally was not allowed to look at the results of some of these analyses. It was in their contract that they’re not allowed to actually look at the results. So it’s really hard to improve or build your own model if you cannot look at the results. What we’re building is a federated model that allows the customers on-site to run the device. They can pull the model, get a new experiment design, that runs in there, and then the model weights are updated, right? This is the same way that the Tesla self-driving was trained, right? Federated learning resolves that IP-sharing complaint or constraint. And the reason that’s so powerful is that now you have a model that is learning from diverse experiments across different cell lines, at different places, but still on the same hardware. That’s really key, because otherwise there’s too much experimental variability in the data you’re getting back. And so you’re not gonna generalize well on that. And the sort of hedged bet here is that if it’s not tractable through machine learning and models, we are still building the highest throughput, most economic, and fastest way to get to that answer through still running experiments. And if it is tractable, we’re going to have the best model for running those experiments. And I think the answer is actually going to be a blend of both. I do not believe that experimentation is going away. But I do think that we will be able to get to much better answers much faster, because that’s really the ideal, right? The ideal is, once you have that model, now you can feed it in even earlier in the process, right? When you’re doing your strain engineering. So coupling those together becomes possible once you have a model.

Abhi: What parameters does the model actually intake? I imagine it takes all the inputs you’ve given into the system, all the outputs you get out of the system, and maybe what the system is actually meant to produce, and the strain itself. Is that everything or are there others?

Sterling: That’s — I think that’s a complete view.

Abhi: Okay. If the belief is that you’ll probably still need human experimentation to help the system along, and maybe the ML won’t fix everything zero-shot — can I conceptualize this as like there are 10x media optimization engineers, and they’ll be able to iterate much faster on this model system as a result of that? Or do you imagine media — bioprocess engineering is a pretty standardized field where these are the first 10,000 things you try, and maybe in the old world you get to try like 5% of that, and in the new world you try those 10,000 things? But ultimately it’s the same set of parameters that the media optimization engineer is tuning.

Sterling: So are we tuning a different, a larger set of things rather than just the engineer?

Abhi: Yeah. Like, all the knobs that the engineer usually gets to tune — do they also get to tune in the system? Or is it a subset, or maybe even larger?

Sterling: It’s a superset.

Abhi: Superset. Okay.

Sterling: You’re getting to tune far more. And it’s a superset in a few different senses. The first is that just bringing the economics down, making it automatic, allows you to — even if you had the capability previously to change a variable, you wouldn’t have essentially the budget or the time budget or the capital budget to actually exploit it. That’s one sense. The second is that it allows you to make finer interventions, with more feedback built in. So the reason for having the real-time sensors, why that’s important — what you actually want to do is be able to anticipate what the cell wants before it needs it. Because there’s always a delay between when something gets introduced into the environment to when it gets uptaken by the cell, right? So ideally I actually want to see those signals happening before the cell needs it. Now, in order to do that, you need real-time sensors that are picking up on that and starting to match that. So that’s a domain that’s just not possible —

— in other systems.

[01:24:32] $8/lane vs. $20,000/lane: the economic utility of Iku’s device

Abhi: I’m curious about — I assume there are microfluidic bioreactor systems that at least exist in the literature. How much improvement do people generally see by going to these systems versus the Fiji-shaped benchtop?

Sterling: Right now? I would say close to zero. And the reason is economic. So the one metric or lens for looking at it is just what is the all-in cost to getting that dynamic cell culture data — that one experiment, that data. And there’s two components to that. The first is, what’s your CapEx, right? How much did it cost to actually get this device in here and use this thing? And it’s really this CapEx per experimental lane. And then the second is, what is the OpEx on that? Every time that we run the experiment, how much does that cost? And so to give an example — the benchtop reactors, depending on whether you’re going with the gold standard or some of the derivative ones now, let’s say the CapEx is between $5,000 to $15,000, $20,000 for each experimental lane. And then your OpEx is — you’ve got not just the media, you need to also take the time to grow the cells up to be able to seed it. You’ve got the human coming in and running it, and then you’ve got the actual disposable, or you’ve got cleaning the thing and sterilizing it. So it ends up being around $1,500, $2,000 every experiment that you run. The closest microfluidic system in capability — it’s only four lanes and it’s $80,000. And so that gives you a per-lane cost of still $20,000. And then the disposable costs are I think still around $500, $700 for each thing. So there’s no — there’s not much economic reason to it. The reason that that product is on the market is because it cuts down on media utilization. But that’s why I think that’s not a very successful product. What we’re building is — in philosophy, there’s a difference between changes in degree and changes in kind, right? So it’s like, okay, you take a little step, you take a little step, and it’s just, okay, it’s different, but it’s not qualitatively that different. And then when you 10x or you 100x something, right — all of a sudden new things get unlocked. And so we’re looking at a CapEx of $8 a lane, and we’re looking at an OpEx per experiment of like $20 or less, right? And so those two things together really transform what’s — and then if, as I said, you start integrating more sensor systems into it, those two parts are kind of fixed, right? The CapEx and mostly OpEx on that. But the amount of data and the amount of value that you can get out of it — that’s where I think there’s much higher place to go.

Abhi: Instinctively — if I understand correctly, both existing microfluidic systems and your system have lithography as the underlying manufacturing component. And yours has circuits integrated, so you can get these sensors. But if the underlying creation process is the same, why are microfluidics so much more expensive than your device?

Sterling: So that device I was just referencing is not made with lithography. It’s a molded device. But the key thing actually is that they don’t have active — there’s a big divide in microfluidics between passive and active microfluidics. So passive is like paper microfluidics or something, right? Your pregnancy test — that’s paper microfluidics. It just does one thing, doesn’t have feedback in it, doesn’t really have control and regulation. And then really separate is, can you come in here, can you sense things and change things as they’re going on? And most of the systems right now do not multiplex the control aspect across a large number of things, and the sensing part of it, and some of the actuation part of it. If you have to use molded plastic, there’s kind of no way to integrate sensors easily from molded plastic. It doesn’t come out of the factory with all these things into it. You still have to go and add all these things together, so then you’re adding in labor costs there, right? And all that. So even if some of the end result is, in certain capabilities, similar, the upstream manufacturing of it — because you can’t integrate everything together — really constrains your economics on it.

Abhi: And so even if the lithography-produced microfluidics device that’s potentially on the market — that alone may cost something similar to the Iku device. But all the sensors that are added on increase the cost.

Sterling: Right. Let me back up here and say that lithography as a technique does have this property where the cost doesn’t scale with how complicated you make it. The big difference is, in silicon, the base cost for making it is substantially higher than the base cost for making things in printed circuit boards. So in general — this is true of almost all forms of manufacturing, to my knowledge — as you increase precision requirements, you increase cost. And it tends to scale logarithmically, right? So if you — there are two ways that you’re using silicon and lithography, which is either you will make it as a mold — so you’re really just using the lithography as a mold, and then you’ll peel this casted thing off of it. Or people will actually use the silicon and make the channels in there. But the problem with silicon is it’s really expensive. In general, we do not make disposables out of things that are made in silicon lithography. Because to make something this size — probably $400 or something.

[01:32:05] Why PCB microfluidics didn’t exist 10 years ago

Abhi: Why — if it seems like the big innovation here is combining lithography — or doing lithography on the circuit board as opposed to doing it either in silicon or via a mold — both of which seem more expensive than the printed circuit board — was it simply a matter of realizing that you could do this on circuit boards and dramatically reduce your costs? What — why did this not exist 10 years ago?

Sterling: Right. So I think the first is that different worlds don’t talk very much, and in this case, the tool-builder world and the tool-user world are very distinct. And the second is that — to answer the question of how did I come to it — I was in my apartment in São Paulo, and I’d been really digging into biofilms. I was like, okay, so much of this is about the concentration of these things, and they’re creating these little microenvironments and all of this. And then I was really — at the time there was this concern about, are we going to have enough bioproduction capacity? And what I’d seen work before is in traditional chemical synthesis — they switched to continuous flow microreactors. So Corning Glass, that makes the glass in your iPhone, they also make chemical reactors. And the benefit of this is that you can flow things together. They react quite quickly. You can pull the heat off and things, and it’s really consistent. The reason you can’t use that in biology at the moment is because, in order to — traditional chemical synthesis, you really are pretty much just controlling flow rate. And the reactions happen really fast normally, right? You just mix them together and it’s done. But in biology, right, you need sensors in order to see what’s going on. The environment is much more tightly controlled, right? There’s more aspects to it. And cells themselves are again perturbing the environment around them. So that was the lens I was looking at — how do you bring this thing that clearly worked in chemical engineering to biology? And also thinking about these biofilms. And so I studied mathematics. I literally wrote this down as a set of axioms. I was like, what do you need? You need to be able to hold fluids apart. You need to be able to combine them together, right? You need to integrate sensors of different modalities so that you can adapt it. It needs to be small, both for mass transfer reasons — because as you get smaller, there’s more surface area around. And the limitation from any reactions is literally just how fast can you get things from the gas phase into the liquid phase. And that’s purely a function of surface area. Even in large reactors when they’re using bubbles, the bubbles are just creating surface area. And it’s about diffusion across that. So if you go small, you get that. You go small, you also get laminar flow, which is really, really nice because it takes problems that are normally chaotic and it linearizes them. So there’s a great experiment everybody should watch on YouTube of — you put a couple drops of dye into this gel, and the gel has a really high viscosity, and then they stir it up this way, right?

Abhi: And they go backwards.

Sterling: Yeah. And they go backwards, right? And that idea — well, why can you do that? You can do that because in a sense it’s linear, right? Whereas in a chaotic system, you’ll get to some point and now you can’t tell which path you were at before, right? So these are things. And then you need to be able to run a lot of them, both for — originally it was for throughput, but that throughput idea also translates to data parallelization. And then if you need a lot of them, you also need it to be manufacturable, right? Mass-manufacturable and needs to come down. Okay, those are the axioms. I was like, these are the things I need. And then I literally went through every manufacturing technique that I could find. I mean, truly everything, down to like, what are they doing with 3D-printed glass at the moment. And you can just knock these out for a variety of reasons. The molded polymers don’t work because you can’t integrate the sensors in them quickly. 3D printing doesn’t work at all — it doesn’t matter what the modality is, because the infrastructure isn’t already there, right? So if you need to make a bunch of disposables — which, great business, always make disposables — if you need to make a bunch of disposables, then you should pick something that you don’t need to have a lot of capital in order to scale, right? So you need an existing manufacturing industry for it. And all these came back, and then ultimately I was like, let me just reframe it. I was like, let’s just pick one of these and optimize for that. What’s the best way to build sensors? I was like, well, printed circuit boards are really good. And I was like, okay, can I then build the rest of this in here? Let me just take a common technique — can I just select some subset of this problem, optimize for that, and then force the other ones to fit into it? And I was like, yeah, okay. Sensors are good there. It’s good on manufacturing. Okay. And then after that, went to the literature. It was like, okay, here’s the one person who’s actually done this. Go fly to England, go work with her, and then —

Abhi: The University of Bath person.

Sterling: Yeah.

Abhi: Okay. Interesting. One person in the world has stumbled across this idea. Well, I guess if every technique seems to have its mild drawbacks and there wasn’t a single optimal one that you stumbled across, what is the drawback of going for printed circuit board?

Sterling: Okay, well, I will tell you — from a — there’s the problem that you might think, and then there’s the problem you’ll discover. The problem you would think is that it’s a kind of weird thing. You have to get people to adapt to it, or — also, you do have to design each of those sensor domains. Just because you pick a good palette to work with, you still have to do a bunch of work. You don’t — all this — those all end up actually being not that big of a deal. The harder problem is this, which is that nobody understands it.

Abhi: That’s true.

Sterling: Truly, nobody understands it.

[01:39:24] Who is the customer?

Abhi: I guess, who are you selling these to? I can imagine one customer is academic labs. Maybe — and I imagine the much bigger customer are people either preparing drugs for clinical trials or generics manufacturers. How — one, how willing are they to buy this stuff? And two, is there a customer base I’m missing?

Sterling: Yeah, so I’d say our first customer is actually the US Army.

Abhi: Oh.

Sterling: And that’s for doing something quite different from media optimization, but still within the realm of — you need to explore a larger space and current ways of doing that are insufficient. The broader answer here of who’s the customer — the customer who feels the most pain for this are the large CDMOs. I’ve spoken to people who have worked for those places. What is the thing that they talk about every year? It’s yield. That’s it. They actually don’t have — if we’re talking about degrees of freedom for them as a company, they don’t have that many, right? They don’t come up with their own products. They aren’t allowed to innovate on it once the process is set. They have extraordinary downside risk if they make a mistake. And they are in a competitive marketplace with — the pharma companies are taking the bulk of the — the pharma companies are getting the value capture, right? They ultimately own distribution. And so those features make them very desirable buyers for it. But media optimization — if you — it both happens within pharma companies for their — sometimes pharma companies manufacture their own things — but also the process of running dynamic cell experiments, that dynamic cell culture, that is pervasive. That’s where I think the largest opportunity really is — all of these problems in biology, many of them ultimately just reduce to, how many dynamic cell culture experiments can you run? And so this is true for new antibiotics discovery. It’s true for doing things in organ-on-a-chip. It’s true in cancer research. If you actually just take the lens of, what are people trying to get out of this experiment? Well, they need to be able to come in, they need to be able to perturb things over time in this, and they need to be able to read out during it. Maybe that’s too big of a lens, right? Maybe there are particular areas where our system is not going to be compatible. But there’s enough of a core there. And the justification for this empirically is that you already see it — every time that bioreactors have gotten smaller and more automated, they diffuse more into the ecosystem. It gets adopted more and people continue to want more automation, more experiments, and cheaper on it.

[01:43:14] What is the ultimate goal of Iku?

Abhi: Do you view Iku as not just a media optimization company? The hope is that whatever the final device ends up looking like, it’s useful for almost anything that’s an in vitro system where you’re trying to screen many things across it.

Sterling: Yeah. Our goal is to produce 99% of the world’s dynamic biological data. And the reason that that’s achievable is because we do not produce that much right now. And by increasing the throughput, by increasing the relevant modalities that we’re putting in and those conditions, I think that is a totally achievable thing. That’s where I started in the beginning talking about this interface between computation and biology and there being that mismatch. That interface, that layer — that’s what we want to build and that’s what we want to own.

Abhi: I’m curious — among the customers right now — maybe the military project is its own direction — for selling this to either CDMOs, pharmas, generics manufacturers — my impression is that all of these groups, like you said, don’t like variability and so they’re very hesitant to buy new technology that promises the sky and the moon. What’s the hardest part about selling to these people and how do you reassure them that things are gonna be fine?

Sterling: I certainly underestimated the importance of that aspect here. I’ve sold a lot of things in my life so far in very different domains, and I will say that not only in biotech, not only in pharma, but for biopharma manufacturing, the level of conservatism and scrutiny is extraordinarily high. So the wedge or way of getting into that distribution — there are a few examples. The first one is the kind of traditional way, which would be, who do the CDMOs look towards? The CDMOs are not going to adopt it until they’ve seen a pharma company use it. Pharma companies are not going to talk to you until you have a paper published from probably a premier lab of some sort, right? The premier lab is not going to touch anything until at least you have a white paper and some connection. In order to do that, you need to build the device. So how do you resolve this problem of getting to that end customer? The first is that there are ways of augmenting existing instruments. So the advantage of it being a standalone sensing system is that you can come in as just an add-on to something — you’re still gonna have the same economics, but now we can offer you some more data out of that same thing. And you can — that’s a lower threshold for them and it’s not involved in the actual — they can just throw that part of the data away if they don’t like it, right? If it’s not useful. So that lowers some of it.

Abhi: I guess it’s cheap enough such that it’s not a major investment to try.

Sterling: Right, right. The second is — and a big — I was just rereading Geoffrey Moore’s Crossing the Chasm, which — have you read this?

Abhi: I have not.

Sterling: Okay. I highly recommend it. It’s been on my bookshelf for eight years, 10 years. The other day I was just like, I should reread this, and — my God. And the big idea is that what counts as a market — what counts as a market is not only that people are buying something repeatedly, but critically that it’s a group of people who talk to each other and look at each other, right? And so you’ve got the academic labs who look at each other and talk to each other. And then the pharma companies look at each other and talk to each other. And then the CDMOs look at each other. But the key thing is actually the big CDMOs — they don’t talk that much. They don’t associate that much with the little CDMOs. But those are ones actually that we can sell to and get some evidence coming in there. So there’s building ancillary systems that can tack on to existing things for getting in there. And then there’s the other way, which is just — be so good they can’t ignore you, in a sense.

Abhi: I was gonna ask — I imagine the gold standard here is you show one of these CDMOs, here’s the cost and titer of expert-produced media versus the cost and titer of expert plus Iku media.

Sterling: Right. Well, actually the gold standard is not that we say it — the gold standard is that Eli Lilly says it.

Abhi: Sure. Yeah.

Sterling: Right. Because that’s their customer. And those pharma companies own those CDMOs.

[01:49:07] What does the validation evidence need to look like?

Abhi: Yeah. I guess, has any pharma done this and produced — or even you internally have done this side-by-side comparison and you have this very clean result to share to them? Or is it more like you’re still in the phase of seeing the magnitude of improvement the system gives?

Sterling: It’s more like — it will be extremely surprising if you do not get the — first of all, if you don’t get the economics. And then also, all evidence points to being able to run more and different experiments gets you to a better answer. So you can kind of work back from that.

Abhi: If you follow the trend lines, it almost necessarily has to be the case that this is better than what’s currently being used.

Sterling: Yes.

Abhi: Okay. Yeah. Has there been — in the early initial deployments of this — and like, will there be a white paper coming out in the next year of, here’s what we found by using the Iku system?

Sterling: Sure. I would say it will necessarily be more dull than that. I would separate it between two — there’s the hype marketing stuff to do, and then there’s what a CSO actually looks at, right? And from my interaction with scientists as a breed — first of all, they are a breed, and secondly, they are allergic to any hype and any kind of promotional stuff here. So what they want to see — and I don’t need to be creative here — what they need to see is your experiments running your device with some readout, and then you take the gold standard and you replicate that, and it needs to be at the same facility, right? It needs to be that. You need to show those two. And basically the graph needs to be obvious enough that it’s like, okay, I can see how these correlate and they scale. They don’t actually need to be perfect. None of them are perfect. This is true for doing scale-up from the benchtop to the pilot and all these things, right? It’s really just a series of graphs. Like, okay, this thing maps onto this, maps onto this thing. And then the next step is, actually you need that replicated at another facility. So for pharma to adopt something, it’s not even just that — you need one lab, you need it to be out of three different labs who all get — because ultimately their thing is about repeatability.

Abhi: Reducing variability. Okay. Yeah.

Sterling: Reducing variability, but then also repeatability. Yeah.

[01:52:14] What would you do with $100M equity-free?

Abhi: If someone were to hand you a hundred million dollars equity-free to push forward the mission of Iku as much as possible — one, I would be curious where you would spend the money, and two, what are the axes of improvement that still lie ahead for the future of the device?

Sterling: Yeah. I think the first thing we would do is really build this high-throughput perfusion system. I would integrate Raman sensing, and I think that’s the — I think that’s a killer app. I think if you do that, it unlocks so much. But also, if you go back through the literature, people have been talking about the value and use of having a high-throughput perfusion device for a quarter century, and that was before we had the machine learning or AI to also interpret that data. That was before the problems that we’re encountering are also getting harder to manage. So I think that’s very clearly there. Along the way there, you build organ-on-a-chip, high throughput. That’s also a constraint right now. One of the larger manufacturers — they’re moving towards it a little bit, but they still have some trade-offs as they try to move to it. Where I think actually really interesting, and I hadn’t gone down until recently, is in droplet microfluidics. So the idea of — in some sense, what we’re doing with perfusion is, okay, let’s take a benchtop bioreactor and all that control, and let’s shrink that down. The droplet microfluidics is more like, let’s just take a test tube and shrink it really small, right? And if you shrink — that’s where 10x Genomics — that’s a form of droplet microfluidics. It tends to be more of an integration of microfluidics with some chemistry, some chemical technique to help with signaling or help with the formation of particular types of droplets that allow memories that you can diffuse through and things. But I think what’s really underexplored are two things. From the customer side or from the data side, it’s higher resolution, more temporal datasets from these, right? Getting back to this idea that cells are time-varying and sensitive and highly parallel in a bunch of different ways. The ability to shrink that experimental system down that much, explore the space, but not lose the temporal element the way that it is right now for the most part — I think that’s really, really powerful. And there are a couple of techniques people have been trying to get down to it. There’s a technique for getting it down to like seven minutes now. But it’s still — there’s still trade-offs. When I look at it, I’m like, oh, they still haven’t resolved these trade-offs. So that’s one aspect I think could be enormously valuable. And then the second thing is, the droplet microfluidics right now — they’re really focused on the formation of the droplets and these things coming through. They are not really chaining things together. And in the literature there are all of these almost like transistor parts, right? Little parts that people have built. And you can see there’s this dream of building truly lab-on-a-chip, right? And the problem is that right now, as you try to build a lab on a chip, you try to do these things — there just aren’t enough of the subsystems or steps that you can link together on there. So it’s like, you do a set of these and it’s, okay, we gotta come out of the chip, right? And then you kind of lose all of it. And so I think it’s really only in the past five years, and then with our technology for being able to actively manipulate things in there and do the feedback — I think rather than conceiving of lab automation as automating manual tasks, which has a hard upper bound on how much efficiency and capability that you will get out of it — let’s just start doing what we did in other industries, which would be, no, no, no. Okay, we have to start over and we have to build some of these things in here, but we’ve already built a lot of them. Now why don’t we actually start building that lab-on-a-chip?

[01:57:31] Lab automation is in a strange place right now

Abhi: I remember, for my lab automation article, one person remarked to me that it’s a shame that liquid handlers have become so popular, because biology happens at much smaller scales than that. So you’re making a system very large when it doesn’t need to be that large.

Sterling: It’s — okay. I don’t know if you’ve ever seen these robot arms that get a cup of coffee and then — they’ve got them in the San Francisco airport. Terrible idea. And it’s like, okay, you go over, and the machine picks up the cup and then puts it over here and does the grinder and brings it to you. What that’s doing is automating a manual task. It’s taking the way that humans have just done something and then been like, I’m just going to throw an arm or an anthropomorphic thing on top of it and then duplicate it. And the result of that is honestly not that great, right? There’s a reason that those things will continue to not take off in any sense other than novelty. And compare that to your Nespresso, which still has an interface — you still need to get your cup — but far better, right? The Nespresso, they’re like, oh, actually, let’s integrate the actual keeper and the automatic dispenser and all of these things, right? And they made it much more compact. Or your coffee vending machines — also works for this, right? Neither one of those are trying to just take the human steps and then be like, literally wherever the human is, we’ll just put this thing in here. And that’s what I’m seeing happening right now in lab automation in general. And I don’t just think it’s lazy. I do think it’s lazy. I don’t just think it’s lazy. I think it’s also close to philosophically a crime. I think it’s a crime —

Abhi: Because you think for automation to truly be useful, there needs to be a new way of interacting with the underlying systems.

Sterling: Yeah. It’s like, they’re just not really thinking through the problem.

Abhi: Well, I guess one argument is that it’s easier for these things to get adoption if you are allowing them to work in the exact same environments that humans are able to work in.

Sterling: Yeah. And I think that makes sense for things like machine tending for 3D printers or for CNC machines, right? But what’s the difference? Well, the CNC part is a hundred millimeters, right? So it necessarily has to be closer to human scale. But look at what’s happened in industrial space — the most useful places for robotics — and a heuristic you can use is, if it says “robotics” in it, it’s not really that useful. Whereas if it says what it just does, then it’s successful. So a dishwasher is a very useful robot, right? Self-driving car — very useful robot. And in the industrial space, it’s mainly around logistics and moving things, right? So the really successful ways of actually leveraging automation — first, they respect the real goal, and they respect the limits of the thing you’re trying to manipulate. So if the things you’re trying to manipulate are grocery things, one way could be — let’s take a humanoid and it goes to the grocery store and picks up things off the shelf. That’s what people do, and that’s what these humanoid companies want to do. And the alternative would be — actually, if you look at the logistics companies that do the best of it, it looks nothing like that at all, right? It’s some huge grid. It has these things running around like crazy, and all they’re doing is picking up these things and setting them down. And there is no way that a humanoid system can compete with that, right? There’s no way. And you can let the economics decide that over time. I just — this idea that it’s actually pushing the lab forward — I don’t really buy. I also do not see Eli Lilly or Johnson & Johnson putting a robotic arm near their lab bench. I think what Kao’s doing with their lab automation system — those carts — right. I think that’s at least sort of a reasonable compromise in a sense. We don’t need to go and re-engineer each of these things that already exist. If we can literally just make the interface easier. But then I think the real goal should be, as much as possible, if the economics fit, just think through the problem correctly. Just put it on chip as much as possible.

Abhi: That makes sense. I think those were the last questions I had. Thank you so much for coming on.

Sterling: Thank you for having me. Yeah.

Spotify: https://open.spotify.com/episode/6BLZph2uGGUVphbNQ8NGPd?si=SVBSKJM8RdO4AhYzDa-ZfQ
Apple Podcast: https://apple.co/3OU5Zse
Transcript: https://www.owlposting.com/i/189602943/transcript

Introduction

This is an episode with Ben Woodington and Elise Jenkins, who are the cofounders of Coherence Neuro. The pitch for Coherence is as follows: a brain implant that treats cancer with electricity. When I first learned of the company in mid-2025, it was such an alien thesis that I instinctively wrote it off entirely. This surely isn’t clinically plausible at all, maybe it will be one day, but certainly not today.

Then, while I was in San Francisco, I met up with Nicole, Coherence’s chief of staff. After that, I was far more convinced that there was something real here, especially after she told me that the electricity ←→ cancer thesis already has some merit: Optune, an FDA-approved medical device developed by Novocure. This has been on the market for over a decade, and uses externally delivered alternating electric fields to treat glioblastoma. And it works! If Optune is consistently used, glioblastoma patients can live up to twice as long compared to chemotherapy alone. How does it work? Simple: the alternating electrical fields prevent fast-dividing cells from replicating by interfering with the physical process of cell division (specifically, mitotic spindle formation).

After this, Nicole connected me with Ben and Elise, the cofounders of the company. It was an incredible conversation. During it, I was informed that cancer cells behave eerily similar to neurons: hijacking neural pathways, attracting nerves into their microenvironment, and forming synaptic connections with surrounding tissue. Given this set of evidence, none of which felt particularly controversial, an easy logical leap is to ask the question: why can’t you throw neuromodulation at the tumor? Maybe not even just for treatment, but monitoring as well? Optune was a step in the right direction, yes, but surely it can be pushed even further.

So Coherence was born, the only (neurotechnology x oncology) company in existence. Ben and Elise met during their PhD’s at Cambridge, spinning up the startup with the belief that a modality long assumed to be exclusively for neurological conditions like Parkinson’s, epilepsy, and chronic pain, may have a profound role to play in cancer. And perhaps even conditions outside of it.

And during my last trip to San Francisco for JPM 2026, I had the honor to sit down with Ben and Elise to talk about it all.

This conversation covers how Coherence’s first neurotech device (SOMA) works, the molecular reasons behind why neuromodulation affects cancer at all, what the biomarker readouts look like, the obvious Michael Levin comparison, and a lot more. Coincidentally, Ben helped me out a fair bit for my neurotechnology piece awhile back, and that article may be helpful reading material for this episode.

Enjoy!

Timestamps

(00:00:00) Introduction
(00:01:42) How is SOMA different from Novocure’s Optune?
(00:08:57) Why does neuromodulation affect cancer at all?
(00:13:28) How was cancer-nervous system crosstalk first discovered?
(00:15:42) Anti-epileptics and beta blockers as accidental cancer drugs
(00:17:38) What is molecularly happening when you block cancer-neuron crosstalk?
(00:19:50) What is SOMA actually reading out as a biomarker?
(00:20:44) What does it mean that cancer is “very electric”?
(00:22:02) Can you derive universal biomarkers across patients?
(00:23:09) How is the device placed?
(00:24:45) How does the blocking stimulation regime work?
(00:26:43) Is it fair to say this is closed loop?
(00:29:05) Why not just spam the tumor with constant stimulation?
(00:32:31) Why MRI safety is non-negotiable for oncology devices
(00:33:35) Walk us through the patient journey from diagnosis to implantation
(00:36:13) The Michael Levin question: can you reprogram cancer back to normal?
(00:42:29) Efficacy, hospice settings, and the utility of the neuromodulation literature
(00:45:52) Why start with glioblastoma instead of an easier cancer?
(00:48:57) Regulatory strategy and the reimbursement threat
(00:55:37) How well does mouse-to-human translation work for neuromodulation?
(00:58:09) Why didn’t this exist 10 years ago?
(01:01:48) The founding story
(01:06:38) Why build your own device instead of using off-the-shelf arrays?
(01:08:35) Speaking with glioblastoma patients
(01:12:04) What was it like to raise money for this?
(01:13:56) Beyond cancer: TBI, lung disease, and the pan-disease argument
(01:17:40) Hiring at Coherence + what is the hardest type of talent to find
(01:23:17) What would you do with $100M equity-free?
(01:27:15) Are you a neurotech company or a cancer company?

Transcript

[00:00:00] Introduction

Abhishaike Mahajan: Today I’m going to be talking to Ben Woodington and Elise Jenkins, who are the co-founders of Coherence Neuro, a startup that is building therapeutic neurotechnology that manages cancer from inside the body. I first want to talk about what specific device they are building, because I think it really sets the stage for how interesting the Coherence pitch is.

Ben and Elise, welcome to the podcast. Your first device is called SOMA. What exactly does it do?

Ben Woodington: That’s a brilliant opening question. Thank you so much for having us here. To rewind slightly, you’re right, we build technologies that interface with cancer, surrounding biology of cancer using electrical stimulation and recording. We’re really leveraging the intrinsic electrical properties of cancer, and also the way that they intersect and interact with our nervous system. We have a lot of programs ongoing, and I’m sure we’ll talk about some of them today. But as you correctly identified, our first product and program is SOMA, which we’re using in brain cancers. This is a tiny device, a BCI-like device that sits in the skull, and it can deliver an electrical stimulus to a tumor in the brain. We can also record the electrical activity from the tumor and around the tumor for readouts. And we’re working very hard to investigate what those electrical readouts can mean for diagnosis of the patient, for the prognostication of the patient. And of course, the therapeutic potential of that device as well.

[00:01:42] How is SOMA different from Novocure’s Optune?

Abhishaike Mahajan: The interesting thing when I was first researching Coherence is that this is not the first device that uses some notion of electrical fields to interact with cancer. There’s another one called Optune developed by a company called Novocure. Is SOMA fundamentally different from the technology employed there?

Ben Woodington: Fundamentally different, yes. I think the natural development of — we have often said that Novocure is a 25-year-old technology. Optune is a 25-year-old technology. And if Optune is a Walkman, we’re going to be the iPod. It’s a lot more technically dense. We’re doing readout capabilities, recording capabilities. The thing is much, much smaller. We get much closer and in contact with the tumor and the body itself.

Abhishaike Mahajan: Maybe just to give some context as to what the Novocure device actually is, my understanding is that it is a non-invasive device — you stick it to the side of your skull, meant for glioblastoma patients — that emits a low frequency electrical field, preventing fast dividing cells from dividing. What is the evolution of SOMA? What about SOMA is an evolution from that?

Ben Woodington: I’ll let Elise clarify the Novocure point because that was the background of her PhD.

Elise Jenkins: Yeah. So Novocure, as you said, is a wearable device. It’s actually four patches or arrays of electrodes that are positioned on the scalp, on a shaved head. And it delivers alternating current electric fields. They’re actually more intermediate frequency electric fields. And they are proposed to interfere with mitotic spindle formation, the way that cells divide. That’s what Novocure’s technology does.

Abhishaike Mahajan: And what about SOMA ...what is it an improvement on?

Ben Woodington: Though the Novocure device has powerful overall survival statistics, there’s a very steep usage effect curve. When patients use the device around the median time, which is around 18 hours, the overall survival in those glioblastoma patients is about four months. But when they are super users of that device — when they use it 22 hours, 23 hours, or even more — that overall survival goes through the roof and those patients are getting maybe nine months, maybe more, in median overall survival. Which is almost doubling their life expectancy, which is huge. That’s probably the most impactful thing in glioblastoma in the last 20 or 30 years. Now it is a very natural progression to say, okay, why aren’t patients using that device 22 to 23 hours a day? It’s a large compliance issue.

Go inside and you’re in charge of how much stimulus the patient is getting and for how many hours of the day. You offset a lot of those systemic effects — the skin irritation, just the fact that they have to wear something on the head all the time. Patients tend to not like wearing large things on their head. That’s why there’s been so many EEG device failures. Just being able to justify going inside and treating those patients 24 hours a day is a huge benefit already, before you even start talking about the data elements and recording elements that you can introduce, or the novel stimulation regimes that you can start using once you’re inside.

Abhishaike Mahajan: What other electrical things can you take advantage of when you’re actually physically inside the body?

Elise Jenkins: I think there’s a few things. If you think about the way that an electric field is delivered through Novocure’s platform, they require a very large voltage to overcome the skull. There’s a big loss component there. And because of that, they have to carry this very large backpack with a big battery pack, like a car battery, in order to deliver the electric field threshold that they need to enable that interaction with the cell, whether that be mitotic spindle formation or one of the other mechanisms, which is around the process called dielectrophoresis that happens in the cell during metaphase. There are two interactions that happen when you expose that kind of electric fields to cancer cells. And they do that externally using this field. The advantage of going into the brain is that you no longer have that barrier anymore. You don’t need these extremely large electric fields. You don’t need these car batteries. You can have a very small wearable. The interface is non-obtrusive for patients. There’s a big argument around non-invasive versus non-obtrusive. And that’s one of the natural progressions in terms of using this type of technology continuously, which is shown to have the biggest benefit in patients. They’re going in already, they’re doing surgery already on these patients. Let’s put a device in that can deliver that kind of field or other types of electrical stimulation. Let’s do it locally. Let’s also record what’s happening because we’re right there. We’re interfacing with those cells and tissues. And we can do it without being obtrusive to patients’ lives.

Abhishaike Mahajan: And the actual device itself is emitting the exact same type of field that the Novocure device is emitting, or something else?

Elise Jenkins: It can do either. We’ve been looking at — a lot of my PhD work was looking at Novocure’s types of stimulation, tumor treating fields. But the advantage of being closer is that you can start to look at different types of stimulation. Neuromodulation is something that we are particularly interested in. We’ve been exploring different ways of optimizing electrical stimulation in these types of cancers. Neuromodulation is a really interesting one. When we started development of the SOMA platform, we were really interested in the data that you could actually record. That was — we knew that there was a therapeutic intervention that you could use. Novocure had already shown that clinically. We were really interested in the data element. When we started looking at what happens longitudinally when you record electrical activity from these tumors, based on a lot of history of neural interactions that happen with these cancer cells, we started to discover that there were these very interesting biomarkers that are really relevant to what people target with neuromodulation. And that was what drove us to consider beyond just what Novocure is doing with electrical stimulation — field-based mitotic spindle, cancer cell focused — to what’s happening in the rest of the environment and how can we actually target or look at the rest of the environment, modulate that behavior, and how would that affect cancer cells. That’s what we’ve been looking at, optimized strategies for stimulation.

Abhishaike Mahajan: Just to have a good mental model for what the SOMA device actually is and where it is placed in relation to the cancer itself — should I imagine it as like you have a little head right here and a bunch of spikes coming out of it that poke directly into the cancer?

Ben Woodington: We wouldn’t use the word spikes. We would use leads or threads. But yes, the brains of the device is that part that you see that’s anchored in the skull. And as Elise said, that’s delivered at the time of surgery through a very small perforation in the skull. Then the front end of the device is modular. We have leads and threads that come off the front of that device and we can position them in and around the resection cavity after a tumor has been removed, or into a tumor maybe without resection. Then in the rest of the body, we can look at targeting specific nerves or tumors there as well. And as Elise alluded to, we’ve seen some pretty promising results looking at neuromodulation — slightly lower frequency regimes rather than super high frequency regimes — in those peripheral cancer indications as well.

[00:08:57] Why does neuromodulation affect cancer at all?

Abhishaike Mahajan: Whenever I’ve brought up Coherence to other people and mentioned neuromodulation in combination with cancer, there’s always this surprise that neuromodulation does anything to cancer. What is the intuition for why you would expect neuromodulation to do anything to cancer? I can buy that monitoring the nervous system nearby the cancer helps you have some notion of biomarker, but why does performing neuromodulation at all affect it?

Elise Jenkins: I think maybe the same surprise that others might have when they first hear this is the same surprise that people who discovered these interactions had. Cancer cells behave and act a lot like neurons. And I think that was a big surprise for the entire field that started to make these discoveries. A lot of cancer cells — and not just in the brain, this happens in other organs as well — mimic a lot of the behavior that neurons have. They hijack neural pathways. They have an ability to attract neurons into their environment. They also have an ability to attract nerves into their environment. If you’re outside of the brain, all of these properties make a really nice opportunity for you to then consider neuromodulation or other targeted strategies that are not just looking at the cancer cell itself. You could consider a similar analogy with the immune system. When people are looking at immunotherapy, they’re not targeting the cancer cell. They’re targeting a completely different subsystem in biology that they can leverage and tune in a way to target cancer cells. And when people have started to unpack this opportunity — that cancer cells are behaving so similarly to neurons — it massively opens up the therapeutic opportunities that you can exploit, things we have used for decades in other indications. I think that similar surprise was also a surprise to the people who discovered it.

Ben Woodington: And that’s recent. These are recent discoveries over the last five to ten years. Not mechanisms that have been uncovered and explained for 50 years, which is the exciting thing.

Abhishaike Mahajan: Is it fair to say that glioblastomas have the most nervous system interaction and maybe prostate cancer has the least, or is it not that clean?

Ben Woodington: I can say one thing and then maybe Elise can say the other. It’s difficult to draw a side-by-side comparison between glioblastoma brain cancers and peripheral cancers in the body simply because of the nature of those tumors. The tumors in the brain are in a sea of neurons. It’s a volume of conductive tissue, neural tissue, whereas tumors in the rest of the body are heavily innervated with nerves but are not existing within this sea of neurons. So it is tricky to draw exactly a side-by-side comparison. It does change how we design the devices and how we introduce them to the body. But yes, they do still have these neural features.

Elise Jenkins: I think if you’re studying gliomas or tumors of the brain, it’s a natural curiosity to imagine that cancer cells might have some type of similar features because they’re an extension of oligodendrocytes or astrocytes or whatever other brain-type glial cell. It’s not so unbelievable to think that cancer cells might behave similar to the environment they’re in in the brain. And I think that’s also why a lot of the research has been more well-established in the brain — in gliomas and diffuse intrinsic pontine glioma [DIPG], which is a pediatric glioma. A lot of the research is very well-established in those regions. But I also think that’s just a consequence of proximity — how close they are to that environment. When you start to look at these interactions happening in other organs, I don’t think it’s a matter of proximity. I think it’s just a consequence of the fact that people have done a lot of research already in glioma. And now that people are starting to observe these interactions happening in other organs, it’s exploding. If you were to do a PubMed search on cancer neuroscience and look at the trajectory of publications coming out in this space, it’s exponential. I think we’re only at the very start of that now. I think we will start to see that these interactions are possibly happening in every cancer in the body, not just ones in the brain.

[00:13:28] How was cancer-nervous system crosstalk first discovered?

Abhishaike Mahajan: You mentioned that a lot of this research is relatively new, past five to ten years. How was it first established that there is any crosstalk between cancer and the nervous system?

Ben Woodington: By one of our idols and collaborators.

Elise Jenkins: It’s possible that this initial discovery really is a byproduct of over 50 or a hundred years of bioelectricity research. If you look at some of the really early research where people are talking about voltage-gated channels and interactions, the idea that cancer is very electric — that’s not a new phenomenon. That’s been something that people have been talking about and studying for a very long time. I think the discovery that Michelle Monje’s group at Stanford made, where they were able to show that there were basically similar behaviors in cancer cells that were very similar to neurons — and then they actually started looking at, well, if you patch neurons and you start to depolarize these neurons, what happens to the cancer cell? They started to see that not only are they mediated by neural interactions explicitly through synaptic interactions, but they’re also mediated by paracrine signaling. When neurons release specific factors into the environment, or even cancer cells releasing those types of factors, there is this network effect. And that network effect is actually really bad . There’s this reciprocal engagement that they discovered, which has now caused a bunch of researchers to say, there’s so much going on here. What are those individual mechanisms? Which ones are happening through neurotransmitters? How many of them are actually synaptically integrating? It creates this opportunity for a whole new host of targets, whether that be new drugs to discover or new ways to therapeutically intervene, but also looking at repurposing. People have been looking at repurposing epilepsy drugs as neural inhibitors. People have been looking at retrospective studies in gliomas — what happens if you were on an anti-epileptic and you also had glioma, and how is their survival different? You do see these types of differences retrospectively. So now people are starting to do those kinds of studies forwards, which is awesome.

[00:15:42] Anti-epileptics and beta blockers as accidental cancer drugs

Abhishaike Mahajan: Does that mean you can imagine at some point it will become standard of care for all people who have glioblastomas to be on an anti-epileptic, or it’s not that open and shut?

Elise Jenkins: They’re running these trials at the moment. I think if there is a significant benefit, I can’t see why they wouldn’t add that as a standard of care. We’re talking about patients who have such poor prognosis, such poor survival. The standard of care has not changed for 25 years. If there is anything that’s beating or contributing to current standard of care, I can’t see why — if the side effects are not completely debilitating, the quality of life is still very important — but if they can do that, then I don’t see why that wouldn’t be a natural progression as well.

Abhishaike Mahajan: That’s cool. I usually don’t hear about free lunches like that. When I think of anti-epileptics, they aren’t super side effect heavy.

Ben Woodington: Another famous example from another one of our collaborators, Erica Sloan at Monash, where they’re looking at beta blockers — another relatively innocuous drug, generic, widely available. Again, looking at retrospective studies on the outcomes of patients that happened to be on beta blockers and showed pretty profound impacts, reducing the chance of metastasis from breast cancers, I believe, in a way that you wouldn’t necessarily expect from something innocuous and that has not typically been used in cancer therapy. It’s exciting. It’s this uncovering of biology that we haven’t thought about before. There’s drug repurposing opportunities, but also forward looking — what does that mean for our understanding of biology and cancer and the system itself and how we can design new therapies as well.

[00:17:38] What is molecularly happening when you block cancer-neuron crosstalk?

Abhishaike Mahajan: Before we move on to using these signals as a biomarker of cancer itself — let’s say SOMA works, you’re able to prevent the cancer from interacting with the rest of the nervous system, and somehow that improves prognosis, which given the evidence you’ve given so far feels like not too large of a logical leap. What do you suspect is molecularly going on that is actually helping the patient? What about breaking the crosstalk is actually benefiting?

Elise Jenkins: I think there’s a whole host of things going on. There’s not a single molecular interaction that happens between neurons and cancer cells. There are interactions that happen with what they call pacemaker cells, which are specific cancer cells in the network that have synaptic integration with neurons. And then they have this gap junction-mediated network that happens between cancer cells. When we’re looking at neuromodulation as a regime for electrical stimulation, we look at something called a blocking regime, which essentially is a depolarization block. You either try to stop the neuron from being able to propagate action potentials, which in turn could mean things like blocking neurotransmitter release. Glutamate is a primary example — a primary excitatory neurotransmitter that has a really profound impact on cancer cells. There’s a lot of glutamate in the brain. Being able to reduce some of that — when people have used pharmacological blocking of AMPA receptors, they see around a 50% reduction in tumor volume in mice. So we imagine that there is a whole host of things that you’re interacting with when you block that.

Abhishaike Mahajan: Maybe a dumb question, but why does reducing the levels of the neurotransmitters lead to a reduction in tumor volume?

Elise Jenkins: They’re growth factors. Glutamate and neuroligin-3 is another one — another paracrine signal that is picked up by gliomas — they are growth factors.

So you’re blocking growth factors.

[00:19:50] What is SOMA actually reading out as a biomarker?

Abhishaike Mahajan: And with regards to the actual biomarker aspect of SOMA — you implant the device, maybe it’s post-resection or before a resection — what is the actual readout that you’re getting?

Ben Woodington: Fundamentally we’re recording electrophysiological signals as well. For all intents and purposes, you can think of this like a brain-computer interface. We’re recording electrically, we’re stimulating electrically. We’re reading out the same endogenous electrical activity of the brain as a device that’s trying to read out motor intent, for example. We’re looking at spike rates across the brain and more broadly, local field potentials — frequency shifts in the brain. Cancer cells do have this intrinsic electrical property as well. And you can measure that and we have measured that.

[00:20:44] What does it mean that cancer is “very electric”?

Abhishaike Mahajan: Before you move on — you mentioned that cancer is very electric. What does it mean that something is very electric? Why is cancer very electric?

Elise Jenkins: They have a very high expression — or they retain a very high expression — of voltage-gated ion channels. When they say they’re very electric, they hold a particular membrane potential. They depolarize under certain events. Those depolarizing or hyperpolarizing events usually occur during the stage of the cell cycle. That’s what we mean by that — they’re very electric.

Abhishaike Mahajan: Sorry, go ahead.

Ben Woodington: It’s important to clarify. We are looking at this electrical activity of the brain. We use that as a proxy for what is going on in the tumor or near the tumor. Because of these interactions between the tumor and its microenvironment, we can use that electrical proxy to inform ourselves of what’s going on in the tumor. We’re not measuring specific biomarkers or specific proteins in that microenvironment. We’re measuring the electrical activity of the microenvironment and what that means. Trying to correlate that to tumor volume, or drug responsivity, or seizure activity, or maybe the aggressiveness and growth rate of the tumor — trying to do all of that from electrical proxies.

[00:22:02] Can you derive universal biomarkers across patients?

Abhishaike Mahajan: That sounds very custom from patient to patient and tumor to tumor. Are there actually some universal properties that you can derive from the EEG-esque readouts or spiking that you’re getting?

Ben Woodington: We’re going to find out in humans. Yes, in rodents. But of course rodents are very homogeneous between mouse and mouse. We’re going to find out in humans. I will say, using electrical biomarkers as a proxy for something else is not new. People are doing this in other areas of medicine. People are now looking at this in Parkinson’s, in pain, in depression. And in those spaces, there’s still a lot of variability between patients. We expect to be able to demonstrate the same thing in humans between patients by training one model and applying it to multiple patients. That’s the end goal. Of course, proof will be when we run our long-term human studies, because these have never been done before. It’s one of the most exciting things — people have done this intraoperatively. You can take a recording from a patient, from a tumor while they’re under in surgery.

[00:23:09] How is the device placed?

Abhishaike Mahajan: We’ve been talking about when you actually implant the device — it could be pre-removal or post-removal. The pre-removal setting makes some sense to me. You attach the leads into or around the tumor itself. In the scenario where the tumor has been removed, are you just placing it into the cavity where the tumor was found, trying to see if there are any tumor cells left over? What’s the use case there?

Ben Woodington: In the margin. In the brain cancer case, most recurrences — glioblastoma, to paint a picture, is a horrendous disease. Always terminal, unfortunately, and very poor mortality rates. Those patients, even after a resection — somewhere between 70 to 80% of patients are having a resection, are fit enough to have a resection — all of them will get a recurrence again. And of those patients that get a recurrence, most of them, I think 90%, are in the margin of where the resection was made. That’s where your residual cells are. That’s where your tumors start again. That’s where we’re targeting first. Now that’s not to say we can’t go more broad area, and we have internal programs where we’re looking at broad area coverage, maybe from the surface of the brain, maybe deeper in the brain. The goal is to get as much coverage of the brain as possible so you can control distantly as well.

[00:24:45] How does the blocking stimulation regime work?

Abhishaike Mahajan: This is somewhat clear in the case where you’re purely recording and you get this spike readout. For the case of intervention, what does the perturbation you’re applying actually look like? You mentioned something about blocking. Could you talk a little bit more about that?

Elise Jenkins: There is a multitude of stimulation regimes, but one of the ones that we’re really particularly interested is this blocking regime. When we started looking at biomarkers — when we started recording longitudinally — what are the electrical biomarkers that we’re seeing change in the brain? We saw a really interesting peak in hyper-excitability. If we look in the high gamma range, you see this week-to-week increase in high gamma activity in tumor-bearing mice. Frequencies above 70 Hertz. That’s where we tend to see very distinct changes in brain activity during progression of disease.

What led us to thinking about blocking was, well, if we know the general spike rate of these kinds of neurons in a particular region of the brain — it varies throughout — computationally, we started deriving some neuromodulation parameters that looked at modeling the neuron and essentially looked at, can we block, can we stop this activity from happening? If you generate an action potential, how do you stop it from generating again? And what consequence might that have on cancer growth? That’s essentially the types of perturbations that we’re looking at. If you’re not familiar with blocking, you can think about it similarly to how you would evoke a neuron. If you stimulate a neuron, you evoke an action potential. When you block, you essentially stimulate at a faster rate such that when it tries to regenerate or repolarize, it can’t. You leave it in this constantly depolarized state. It can’t reach a threshold. It can’t activate. That’s the kind of perturbations that we’re interested in, in the brain and outside of the brain.

[00:26:43] Is it fair to say this is closed loop?

Abhishaike Mahajan: Is it fair to say that this is closed loop and that it is not actively learning about the electrical state of the cancer?

Ben Woodington: Currently, no. This is an open loop system. Rather than thinking of this as a closed loop therapy, we prefer to think of this as a theranostic platform. We have some electrical stimulation where we’re treating this disease and we have a diagnostic function where we’re reading out and we present the clinician and the patient with what is going on. You can imagine that would be overlaid on an MRI to say what’s going on with the tumor in real time, very fast. That’s great — it’s a new diagnostic weapon. Eventually, of course, you train these models and these systems to be better than clinicians so that you can make these systems closed loop. That’s the holy grail across many aspects of neuromedicine, where you no longer have to have a second diagnostic readout and say, how do we tune the stimulation, where do we position the next device, the next electrode. Instead, the system’s doing that for you, optimizing exactly where it’s treating, where it’s stimulating.

Elise Jenkins: It might be helpful to understand that when we started trialing these types of blocking stimuli in the brain — high frequency blocking is not new, people have been doing this for a while throughout the body — there is a way that you can understand whether or not you’re having an effect in the brain. What we typically do is record a segment of electrical activity, process that data, look at the frequency band. We apply stimulus, we then record again, and we see that there is a transient response when you deliver this type of neuromodulation, where you can see a decrease in the high gamma range that we’re interested in. And that’s also how we threshold. That’s also how we would work out — do we need to increase the stimulus, do we need to change the stimulus, do we need to change which pairs of electrodes we might be using to achieve a certain area of activation or blocked regions of the brain. We can use those kinds of methods of pre and post recordings to tell us whether or not we’re having the effect that we desire.

[00:29:05] Why not just spam the tumor with constant stimulation?

Abhishaike Mahajan: Incredibly naive question on my end, but it sounds like you just want to be constantly stimulating all points of the tumor as much as possible to prevent it from ever being able to fire off an action potential.

Elise Jenkins: Pretty much.

Ben Woodington: In some scenarios, yes. But we are also running studies where we’re looking at dosing, because there are other mechanisms at play as well. Some of those mechanisms — you don’t need to be stimulating and spamming them constantly. You can perhaps dose once a day and elicit some local biological response as well.

Abhishaike Mahajan: I’m curious, if you’re actively going to be working with a clinician to tune the actual inner workings of the SOMA, what are the knobs of control that the clinician is actually allowed to tune? If it seems like just overloading the tumor with stimulus is what you’re doing in practice.

Ben Woodington: I would draw a parallel between radiotherapy. Whole brain radiotherapy is really fucking grim. If you’ve ever seen a patient go through whole brain radiotherapy, it is gnarly. It is awful. It affects their whole brain, as the name suggests. And those patients are never quite the same afterwards. Clinicians don’t want to do that. They do it as a last resort and they try to use very focused technologies where they can hit the tumor very hard and spare the rest of the brain. That would be the approach that we would take as well. And it’s the approach that Optune takes — they focus that field, focus the stimulation to as much of a concentrated point as they can, so they can hit that area as hard as possible. That’s what we would want to do as well, rather than targeting the whole brain.

Abhishaike Mahajan: Instinctively, why is there any off-target effects if all of the threads are around the tumor?

Ben Woodington: Because current spreads.

Elise Jenkins: The network in the brain is crazy. You have long-range projection neurons. You might be affecting the body of a neuron in one area that has a projection going very far away. There is definitely a network effect. One of the things that we’ve been looking at is how you can computationally model what the affected area of tissue might be — affected meaning the area that would be blocked. We’ve integrated neuron models into our computational models that essentially tell us, this is the threshold that we need to hit in order to block a neuron X distance away. You build these really nice balloon-type shapes around the electrodes that tell you how far you’re actually going to reach if you want to block neurons. And then of course there are probably some neurons, as an extension, that are connected somewhere else in the brain’s network. At the same time, you’re also imagining that in glioblastoma, when these patients are having resections, they’re having big chunks of tissue just taken out. Trying to preserve function and being aware of which areas of eloquent cortex you want to try to avoid so that you’re not inhibiting movement or inhibiting speech or inhibiting critical functions — trying to design the stimulation parameters or the way that you might activate those electrodes, where should they be in order to avoid those spots. You can do that by taking the MRI into account as well.

[00:32:31] Why MRI safety is non-negotiable for oncology devices

Abhishaike Mahajan: Actually, I’m curious — you do mention that SOMA is MRI-safe. Why is that important or particularly useful?

Ben Woodington: It is absolutely critical for oncology. MRI is not going anywhere. It’s the gold standard imaging technique. It’s used in the brain more than anywhere else. In glioma cases, they ideally want the patient to be having an MRI every three months. If you introduce a device to the body that is non-compliant with MRI, that’s a massive problem. That’s one thing. The second step is not just inducing compliance into the device, but making sure that device doesn’t cast any artifacts. MRI relies on magnetic fields, and if you have a magnet in your device or large chunks of metal, that will affect the MRI image. You start casting shadows, and your clinician’s not going to like that. We’ve spent a lot of time engineering this device and using technologies in this device to overcome this issue.

[00:33:35] Walk us through the patient journey from diagnosis to implantation

Abhishaike Mahajan: External from the actual inner workings of the device with regard to therapeutic interventions and the biomarker readouts, I am curious about the practical use of this device in a clinician’s workflow. What will it look like? You’re diagnosed with glioblastoma — you immediately have this put in, or what?

Ben Woodington: We’re going to work our way up through patients. Our first patients will be the most sick, recurrent patients who are probably coming in for their second surgery at this point, and the device will be left behind. Then we’d be moving towards newly diagnosed patients. Let me walk through what that generally looks like for a patient. A patient would usually present with perhaps a seizure — fit, healthy, 42-year-old man or woman, has a seizure. They go to A&E, the doctor will say, I think you should have an MRI. You have an MRI. They spot the tumor. Pretty quickly you’re brought into surgery — within a few weeks, ideally — for a resection. Our best clinical access point would be right there. Leave that device behind at that first surgery. The patient will then have radiotherapy and chemotherapy, temozolomide usually. Eventually we want to work our way up and be at the top of that pile.

Abhishaike Mahajan: Let’s say you run the clinical trial with SOMA. The thing that I would be instinctively curious about is that the patients who are most willing to have this device put in are also the sickest, and potentially the device might not be able to do anything at all. Is that at all a concern?

Ben Woodington: No. We will have done a lot of work preclinically to validate this technology. We’re adopting some of the lessons from Novocure as well, which is now very clinically validated — it’s been on the market for 20 years. Of course, early feasibility patients are always signing up for a clinical trial. We won’t be signing those patients up saying we guarantee this is going to end your disease and you’re going to live forever. That’s not something we can do, just like any other drug or device trial. We’re going to be working within the bounds of ethics of clinical trials as well. But there’s a hell of a lot of work that leads up to that so that we’re confident we’re going to have a clinical and therapeutic effect.

[00:36:13] The Michael Levin question: can you reprogram cancer back to normal?

Abhishaike Mahajan: One thing I did want to ask — whenever someone outside of the bio field hears about bio, their first thought is Michael Levin. If I put my Michael Levin hat on and look at Coherence Neuro, my thought is: well, if you put SOMA into a glioblastoma, why can’t you just reprogram it back into a normal neuron, because all cancer is membrane depolarization gone awry? That probably isn’t true, but what is your view of the Levin-esque understanding of bioelectricity?

Ben Woodington: I’m going to hold back for a second.

Elise Jenkins: The way I interpret Levin’s interpretation of what’s going on in cancer is essentially that cancer is maybe mostly influenced by external cues, less so by genetic abnormality. I think that aligns a lot with the way that we’ve been building this technology and how we would use it — we’re trying to influence the environment, given that that’s a dominant factor for how these diseases are able to thrive. If you can believe that cancer cells are able to modify themselves to thrive in a particular environment, it shouldn’t be so far-fetched or impossible to believe that the same can be said in reverse. I think the challenge against some of this thinking is that cancers are normally diagnosed at a really late stage. At that point you have a significant number of driver mutations that have happened. Saying that you can essentially nudge these cells back into a healthy state is a bit of an oversimplification. However, I do think that by using something like a bidirectional interface — where we’re no longer just relying on what we see in a cell at a specific point in time, a snapshot at day one or day five, and we miss everything that happens in between — I think there is a lot of information that we can get out of longitudinal data. What happens to that cancer cell or that environment over the course of its evolution? We don’t have that yet, and that’s exactly what we’re trying to build. And I think it’s not infeasible to think that these types of devices will be able to have single-cell resolution at some point. So while I don’t fully buy just yet that you can just nudge these cells back into a healthy state, I don’t think it’s so far-fetched. If we understood what was happening that makes them change through time — which we still don’t know — perhaps if we listen to them and read from them and learn what’s happening, I don’t think it’s that crazy that we can start thinking about what kind of nudges we need to make to put them back into a healthy state. This is not super crazy.

Abhishaike Mahajan: So the argument is that at the very start there is genuinely membrane potential gone wrong, and then driver mutations are acquired, and then it’s irreversible.

Elise Jenkins: I think that in a simplistic view, you could say that, but there is so much going on in a cell. It’s an oversimplification to say that one single voltage channel is driving this entire process. I think there are multiple things going on — multiple different membrane potential-mediated interactions that are happening that drive the change in DNA or whatever else it might be that says, now change this expression, express more of this protein, so that you can leverage the environment. As they’re growing, that growth happens exponentially. You’re having so many more of these mutations happening. And as the environment changes, they change again. They’re really clever at figuring this out. I think that because we catch it after so many things have happened, it’s very hard to work out how you’re going to go back and change 15 or 16 different steps with one single application.

Ben Woodington: In a highly heterogeneous tumor environment that now has 40 different cell types or however many.

Elise Jenkins: And I don’t think you can make the claim that if you stimulate that environment — let’s say you try to target just the cell — that there’s no consequence to the neighboring non-cancerous, healthy participating cells. You’re going to modify them too. How do you design a protocol or a system that essentially only targets those specific channels, for example? If you look back at around 2010, there was a really incredible review article that looked at what happens in cancer cells — what happens to the membrane potential, what happens when they depolarize just before they enter a certain stage of the cell cycle. It’s called cell cycle-mediated membrane fluctuations. There’s so many things involved in that process. I don’t know how you could really specifically target one specific version of the ion channel that can mediate that change. It’s complex.

Ben Woodington: I’m going to go one level higher. Ion channels are important. Membrane potential is important. And our best chance to mess with it and target it is by using electrical biological interfaces like high-density BCIs. That’s exciting. I think there’s a lot of potential there. I don’t think we know yet what the downstream effects can be, because a lot of this has been done in a dish. A lot of this has been done maybe in simplistic rodent models and not a lot of this has been done at the network level and single-cell level in a human brain. So I think it’s exciting, there’s a lot of potential. Jury’s out on whether you can fully reverse cancer back to a healthy state.

[00:42:29] Efficacy, hospice settings, and the utility of the neuromodulation literature

Abhishaike Mahajan: At least for SOMA-like devices that have been tried in mouse models — has there been anything more complicated than mice, or is it just mice?

Ben Woodington: Mice for cancer. All of our safety work is done in larger mammals. The cancer models in larger mammals are less useful, let’s say. Spontaneous models in some companion animals can also be used.

Abhishaike Mahajan: How well does this work in mice? Like neuromodulation for cancer.

Elise Jenkins: In pharmacological settings, what people have shown is around 50% reduction in DIPG, so pediatric glioma. That’s essentially our target. We’ve been looking at how different types of stimulation parameters work in glioblastoma versions of those models. We’re still working on that right now.

Abhishaike Mahajan: Is that for monotherapy or is that combined with something else?

Elise Jenkins: We do both. We’re looking at combination treatment with the standard of care, which is temozolomide, and we also do standalone treatment with a host of different neuromodulation parameters.

Abhishaike Mahajan: Actually, this is something we completely did not discuss. Is SOMA useful even if a patient is going to live only two more months — just for reducing symptoms? Is there a good argument that could be made there, or is it iffier?

Ben Woodington: There has been work that has shown that neuromodulation can be used to reduce seizure activity. And electrical stimulation can be used to reduce seizure activity. There’s a hell of a lot of work that’s been shown that you can reduce pain. You’re talking to some of the same nerve bundles that the sensory neurons are traveling down. I think it’s very likely that we will end up reducing seizure burden, pain burden, et cetera. But again, we can’t speak to our rodents. We’ll have to find out when we do our intraoperative and safety work in humans.

Abhishaike Mahajan: My impression is — are you able to just automatically take advantage of all the neuromodulation literature that’s out there when using SOMA, or do you need to build up your own corpus of knowledge because it’s a brand new device being used for cancer at the site of where cancer just was?

Ben Woodington: Both, right? We massively leverage elements of Optune and Novocure and elements of the neuromodulation world. This isn’t coming out of thin air. There is a body of work that’s been around in the neuromodulation world for 60 or 70 years. We leverage a lot of that, looking at how electrical stimulation routines affect biology and neural firing. We are adopting some of that and applying it to new diseases.

Elise Jenkins: The only thing I would add is that especially on the device product development for the human device for SOMA, there are so many neuromodulation and electrical neuromodulation devices that exist that we can absolutely learn from — to the extent of how do you characterize the electrodes. This is really important to make sure that they’re safe. All of that literature is directly relevant and useful, and we use it all the time.

[00:45:52] Why start with glioblastoma instead of an easier cancer?

Abhishaike Mahajan: This is more of a broader question, but if it turns out that cancer broadly interacts with the nervous system, why go after glioblastoma specifically? Pan cancer would be too ambitious of a goal to start with, but alternatively, why not go after a cancer that’s perhaps less fatal and a bit easier to work with?

Ben Woodington: I think pan cancer is the right amount of ambition. We do want to go pan cancer with this. I think what we’ve seen historically in cancer treatments — the ones that really move the needle are the pan cancer approaches that tackle some fundamental mechanism. Cut the thing out — one of the most effective approaches for cancer treatment. Burn it with radiation — one of the most effective treatments. Chemo. Immunotherapies. Things that affect lots of tumors. We’re going for the same thing. We want to build devices — one for the brain, one for the torso — and we want to go after as many solid tumors as we possibly can. Any of them that we see an effect in, we’ll be pushing forward. Why start with the brain then is the next question. Because it’s hard. There are a number of reasons for this — economic, technical, and cultural. Number one, Novocure has set the stage for the use of electrical devices in brain cancer. The FDA regulators and payers are comfortable with the use of electrical stimulation devices now in glioblastoma. That’s a big cultural moment for these kinds of devices. Clinicians are comfortable with the use of these devices and with physical modalities of treating the disease as well.

Elise Jenkins: 70 to 80% of patients are having surgery.

Ben Woodington: Number two is the surgical elements. For our first devices — and this may not be the case forever — we’re implanting, we’re going inside. So we want to be looking at diseases where surgical intervention is not uncommon. Bring the barrier right down. The risk floor is established. Leaving something behind is marginal risk there, versus trying to justify with many of these neurotechnology companies trying to justify new surgery for a patient — the risk-reward starts to get complicated. For us, we don’t have that issue. And then the final large reason — there are many, but the final large one — is how many therapy options are out there, and what are the macros looking like for new interventions coming to these diseases. Glioblastoma, unfortunately, is not a pretty picture when it comes to what’s on the horizon. The standard of care has not changed for 25 years. Optune is probably the most transformational thing that’s happened in those 25 years. Outside of that, there isn’t a lot of hope. There aren’t many clinicians singing the praises of other technologies coming online over the next 10 years. We want to be at the top of that pile. We want it to be resection, radiation, radiotherapy, chemo, and us. That’s not as easy of an equation for, say, breast cancer.

[00:48:57] Regulatory strategy and the reimbursement threat

Abhishaike Mahajan: You mentioned that Optune has paved the way for medical devices to be used. I’m assuming that by virtue of this being invasive, there will be some new territory that you have to navigate. What is that new territory? What are the logistical and regulatory challenges ahead?

Ben Woodington: Regulatory — we’re not scared of regulatory. We will get this device approved. It’s very likely that we’ll get breakthrough designation for this device. I’m not concerned about that. Reimbursement is your biggest threat and challenge in devices, always. We need to be designing trials, designing the device, designing how the patient interfaces with these devices, and how that also makes payers happy. Novocure has laid some of that groundwork, but there will be different costs. There will be different costs involved in the surgery with the patient, how the surgeons are interacting with the device, how the external components are being supplied to the patient. We need to design trials and a go-to-market strategy that lends itself to that.

Abhishaike Mahajan: This is maybe related to the actual implantation process itself, but do you need a Coherence employee alongside the surgeon, helping guide how exactly the device is put in?

Ben Woodington: It’s a good question, but no. We’ve been designing these technologies alongside clinicians, neurosurgeons, and neurologists to make sure that we are compatible with what they already do. How they plan surgeries, how they implant devices — so that we’re not having to build a hundred-million-dollar robot.

Abhishaike Mahajan: The Neuralink way.

Ben Woodington: There’s obviously some incredible engineering that’s gone into that, but right now we want to get into patients as quickly as possible. They don’t have much time and we want to get there fast. The best way to do that is by giving a device to a clinician that requires minimal surgical training to start

implanting in patients.

Abhishaike Mahajan: That makes sense. Well into a question I’ve had for 30 minutes now. What are the axes of improvement that are on the table for SOMA to be improved upon?

Ben Woodington: Size is critical. The smaller you can go, the wider your patient population and the safer these technologies are. Everyone’s trying to move to more and more minimally invasive approaches where eventually, as Elise said, you’re going through some very small, single-digit millimeter access point into the body.

Abhishaike Mahajan: If it’s at a certain size, are you limited to the most severe, largest tumors?

Ben Woodington: We’re already very small. Our device is about half the width of a Neuralink device, which is compatible with standard burr perforations into the skull — the kind they’ll do for a biopsy, for example.

Abhishaike Mahajan: Is there a good visual indication?

Ben Woodington: A thumbnail. About the size of a thumbnail.

Yeah. Now there are other improvements, of course — power efficiency, electric coverage, all these kinds of elements. Eventually maybe you want high-density electrical coverage to get more and more precision in your stimulus and recording. There’s of course always improvements to be made.

Elise Jenkins: Big one for me is access. Right now, one of the limitations that you might see across a lot of neurotech platforms going out today is how much access of the brain can they get. Neuralink has a really high-density multi-thread device, but they’re all going into a specific region of the brain. One thing that we’ve been really focused on is how do you get multiple access points? How do you create a device or a platform that can access the front of the brain, versus the side of the head, versus somewhere at the back of the head — so you can access multiple areas, but your surgery is still very minimally invasive. By shrinking everything down really small, you can imagine not just having one of them — maybe you can have multiple of them. Now you’re not only accessing this specific region of the brain, but also this region and this region, or across hemispheres, to see if it’s migrating across. That’s something that is pretty hard but quite interesting on our end.

Abhishaike Mahajan: Actually, if a tumor is on one side, why would you care about what’s going on on the other side? The tumor is at the occipital cortex — why would you care about what’s going on in the frontal cortex?

Elise Jenkins: Because of these network effects in the brain. You have a crossover point in the corpus callosum where you have neurons — motor activity that might be happening on one side is actually projecting over.

Abhishaike Mahajan: Okay, they’re projecting on over.

Elise Jenkins: So you can imagine that if you have a very diffuse tumor that is making its way across the brain and actually going to project into the other hemisphere — if, long down the line, you had a device on the primary side of the resection with some electrodes or probes in that region, but you know that they are going to at some point migrate across, you can also put an electrode there and pick up the signals before they start moving across, and maybe start stimulating earlier on that side of the brain.

Abhishaike Mahajan: Wait, what do you mean by migrate across?

Elise Jenkins: It’s not uncommon for very diffuse tumors to move from one hemisphere across to the other hemisphere.

Abhishaike Mahajan: I did not know that. It’s terrifying.

Elise Jenkins: It’s really terrifying. And you can’t see this on MRI for diffuse tumors because they don’t pick up the contrast. You can’t see them, which is a problem. But you can record them. And we know that we can record them. So if you were able to implant in multiple regions of the brain across hemispheres, you can start to actually record when that is happening. You can pick them up from long-range projection neurons as well. We could start recording that information and also start intervening at a much earlier time point.

Abhishaike Mahajan: Is it obvious how many SOMA-like devices you would want in a glioblastoma patient’s brain? Is there a max — like seven of them is enough to cover all the important spots?

Elise Jenkins: It would be entirely based on their MRI. In the pre-operative setting, you would take their MRI. The surgeon will know the extent of resection that they will likely be able to perform, and you’d pre-plan with software that essentially tells you: position the electrodes in this position to get this coverage. That would be how we would do that.

[00:55:37] How well does mouse-to-human translation work for neuromodulation?

Abhishaike Mahajan: Returning back to an earlier thread about all the mouse discussions we’ve been having — how big of a concern is translatability from a mouse platform to pig, to human? Is membrane depolarization a pretty well-conserved phenomenon across all life, or is it case by case?

Elise Jenkins: Particularly in neurons, it’s very well conserved from mice to pigs to humans. We started almost all the way in computation — in silico — then we went into in vivo models. In vivo models for cancer are mouse models.

Abhishaike Mahajan: What do in silico models look like for neuromodulation?

Elise Jenkins: You model the neuron using a Hodgkin-Huxley model. You can computationally, mathematically build that model. You can get that model to generate a specific spike rate. Those are quite well characterized depending on the region of the brain you’re in. Then you can start applying stimulus in silico — computationally — that helps you with selection of what kind of stimulation parameters you think might work best for the region of the brain that you’re in. Then you go into mouse models. These are the most relevant models we can use for oncology. We use orthotopic models, xenografted models. We take human cells, put them in the brain. It’s quite a hard model to do. Then we take a device that is already very difficult to make small for humans and we make it 10 times smaller and put it in a mouse brain. We do a number of tests over short durations to work out the optimal stimulation parameters and the effect of those. Then we go to large animals. We’ve done large animal studies. We’ve been able to show the same suppression effect in large animals in healthy brain. And the natural progression from that is to go into humans. We just got approval to do our first-in-human study to try the stimulation parameters in humans. So far the trajectory seems as good as we can possibly expect.

Abhishaike Mahajan: That’s exciting. Does that translate to a Phase 1 trial?

Ben Woodington: It’s our first-in-human safety work. It’s not a Phase 1, but we’ll be doing recording, mapping, and stimulation safety across the brains of patients.

Abhishaike Mahajan: And this is for glioblastoma patients?

Ben Woodington: It’s for glioblastoma. And that will lead into our next phase.

[00:58:09] Why didn’t this exist 10 years ago?

Abhishaike Mahajan: Exciting. Why does this not exist today? Why doesn’t every glioblastoma patient have this?

Ben Woodington: There has been a lot of innovation in neural implants over the last 20 to 25 years. Miniaturization of electronics, better powering methods, new electrode materials and lead materials. There has been a hell of a lot of innovation — things that didn’t exist in the early 2000s, frankly. On top of that, with Neuralink coming to the table in BCI, there’s obviously been a lot more focus on the use of these kinds of technologies across diseases. Many diseases. And a lot more cultural acceptance from clinical centers to adopt them. I think that’s one of the reasons.

Elise Jenkins: I also think that for us, the scientific underpinnings of these interactions are still very new. The discovery of bioelectricity is not new, as we said before, but the neural interactions that are happening and observed in cancers are really new. I think that in combination with the ability to miniaturize technology and get it implanted chronically and record and stimulate these environments for patients who have literally nothing else — those have been the limitations before now.

Ben Woodington: And to underline how new that is — we sometimes present to academic cancer groups or cancer neuroscience groups. We show our mouse setups that we’ve developed. And it’s a bit mind-blowing for them that you can do high-density neural recordings across the brain of a mouse over months — four, six months. These are technologies that haven’t quite existed in that way. They didn’t exist in that way 30 years ago. We are really at the early stages of that.

Abhishaike Mahajan: When you go to something like the AACR and present your results, it seems like such an interdisciplinary field. There probably can’t exist that many people in the world who really understand the intersection of cancer and neuromodulation, and whatever other fields you’re intersecting with. Do most people seem convinced today that there is something here, or are there still skeptics?

Ben Woodington: I think if there are not skeptics, you’re not working on the bleeding edge. You want people to not agree with everything you’re saying. Our interactions with clinicians and cancer biologists, I would say, usually go like this: “I’m not sure.” And then we show them data. We talk through it. We show them devices, we show them work. And then there’s a big buy-in — people are very excited. I think they see the same things that we see. By the way, I had the same interaction. I come from a neurotech background, a neuroengineering background. I was working in spinal cord and spinal cord injury for years. I had the same response when Elise showed me this about four years ago. It took me a while to digest the papers and read the research and then go, “Oh man, why is no one looking at this? There’s so much opportunity here. I would do this device and this device and this device.” And now we are having the same effect with clinicians who start saying, “Well, hang on — this is how I would design the device to do this.” There suddenly becomes quite a lot of buy-in. I think we’re just at that takeoff point right now. I think we’re going to see a lot more attention clinically, and probably some companies as well, take off. And we’re excited about that.

[01:01:48] The founding story

Abhishaike Mahajan: Similarly, when Elise showed you this — these results from three years ago?

Ben Woodington: Four years ago.

Abhishaike Mahajan: Do you think that was the only moment a company like Coherence could have been founded, or was it just right place, right time to discover this information and put all the pieces together and think there is an unmet need here that’s filled very cleanly by this device?

Elise Jenkins: I felt like I was very lucky because I was really interested in — actually, I was bought into the PhD to look at a drug delivery implant. I knew nothing about glioblastoma. I knew nothing about cancer in general. I’m an electrical engineer. I really wanted to understand the problem. When I started looking at this problem, it’s horrific. I started looking at the potential of a drug delivery platform — an implantable drug delivery device — what is it going to offer here? These cancers don’t respond to these drugs. Maybe you can repurpose drugs that can’t cross the blood-brain barrier — that’s one advantage — but they still just manage to evade these drugs and kill patients. Then I heard about Novocure’s work. I was like, absolute bullshit. No way this works. This doesn’t make sense. I built a platform to try and replicate the work. There’s a whole history to that. I was like, I’m going to figure out what’s going on here. I’m very curious. And I could not disprove it. I tried, and I kept seeing what they were showing in their data. These cells would halt when you would deliver this type of electrical stimulus. My PI, George Malliaras, at the time — we were talking about, well, what happens? There’s an infinite parameter sweep that you can do here that looks at uncovering how cancer cells behave when you put them under certain electrical stimulation parameters. And at the time was when the work from Michelle Monje’s lab came out. I think it was 2019. One of my other advisors had pointed me to this work. I was quite into the membrane potential. I was like, maybe that’s what tumor treating fields are doing — they’re modulating calcium ions or calcium modulators in the cell.

Ben Woodington: Levin is right. Novocure just don’t know it.

Elise Jenkins: I was convinced that that was what was going on with Novocure. At that time —

Abhishaike Mahajan: The mitotic spindle theory was not proven out?

Elise Jenkins: It’s definitely been hypothesized.

Abhishaike Mahajan: Even today, it’s not known for sure?

Elise Jenkins: There’s been a lot of evidence that suggests that’s what’s going on. Yes. But from an engineering perspective, it was not making a lot of sense to me. You’ve got a very weak force acting on a very strong force happening inside this protected barrier in a cell. I was struggling to fully comprehend it, but it worked. And under certain directionality — actually, it was a piece of work that we worked on together — it does work. If you can control the direction of an electric field, you have a really profound effect on tumor treating fields. That was what we found. But Michelle’s work came out and that was my holy shit moment. I was working in a lab full of amazing people doing neurotechnology — making wearables, making implants, making spinal cord stimulators, everything you can think of that interacts with the body. Our lab was building it. And I was this weird person doing cancer in the group. This paper came out from Michelle’s group. I invited her to give a talk to our group — totally fangirling. I love her work. I was like, there is such an opportunity here. Initially it was actually more on the recording side. I was like, these neurons are interacting with these cells. You can read this. We’ve always struggled to get single-cell resolution, to reconstruct that, because it’s very noisy in the brain. You’re getting all of the neurons telling us most of the information that’s going on. The cancer cell signals maybe are a lot weaker or at a much lower frequency. If we can listen to the neurons, that’s amazing. That was when I went to Ben and said, let’s use your device, let’s put it in here, let’s listen to what’s going on. And Novocure works, so we’ll stimulate using that. And now it’s like, well, there’s way more opportunities that we can do now, because look at all of these interactions that are happening, and all of them are a function of neuromodulation or something that we can modulate with neuromodulation.

Abhishaike Mahajan: At the time, you were working on novel devices for measuring and stimulating?

Ben Woodington: For spinal cord injury — brain interfaces and spinal cord interfaces.

[01:06:38] Why build your own device instead of using off-the-shelf arrays?

Abhishaike Mahajan: Super cool story. This leads well into a question I had that we chatted about previously — why build your own device for this? Why isn’t there some standard like Utah arrays that you can hijack and use? You don’t have to build your own thing. It doesn’t seem like anyone does that — everyone hand-rolls their own thing for their own purposes. Why is that a practice in this field?

Ben Woodington: It’s a good question. There are white-label device manufacturers where you can take a device and stick it into a neuromodulation indication, stimulate some nerves, and do your thing. But there are indications where it really does make sense to create your own device. You need a certain density of electrodes. You need to be compatible with the clinical workflow — for our case, MRI. We can’t just take any of those off-the-shelf devices. You can’t just stick a Neuralink device in a cancer patient because they’re going to have to have an MRI, and the magnet in that device is going to affect the MRI.

Abhishaike Mahajan: And there’s no off-the-shelf device that’s also MRI transparent?

Elise Jenkins: Definitely not transparent.

It’s really hard to build.

Ben Woodington: So it makes sense for us to design purpose-built devices for the treatment of these diseases rather than taking off-the-shelf devices. That’s not an easy lift. It takes a lot of engineering effort. And we have a very excited but exhausted engineering team who are doing this. But it’s necessary for us.

Abhishaike Mahajan: Returning back to the original story of Coherence — you showed Ben your work, you decided to form Coherence four years ago. What were the initial set of milestones you had set up to prove whether this is a real thing that could be scaled up into a company?

[01:08:35] Speaking with glioblastoma patients

Ben Woodington: For us it’s — do clinicians and patients want this? It’s very easy for engineers and scientists to start creating things that they like, that are passion projects, without actually speaking to the end users. We see this all the time. We went straight out and started speaking to clinicians, and the pull is huge. I don’t think we’ve spoken to a single clinician that has said they wouldn’t use that. Every single one is like, tell me when I can run a trial with this. I want to run a trial with this sort of technology. Then we went out and started speaking to patients. I’ve become friends with a number of glioblastoma patients. I just hang out with them and drink coffee with them and watch them interact with the technologies that they’re using. And I would say it’s pretty universal — this disease sucks and my options suck. I’m using this piece of technology and it’s horrible and I don’t like it. And if you tell me right now that there’s something better, I will go back and get another surgery. That’s a big barrier. People do not like going in for surgery. So getting that clinical and patient pull was huge. Now how do we transform that into tangible milestones? We build the technology that they need. We’ve been doing that now for almost three years, running safety animal studies. As Elise mentioned, we’re then doing a first-in-human safety study. The next piece is — what does our early feasibility look like? Get it in patients. First 10, then a hundred, then maybe 500. And show that there is a meaningful clinical, therapeutic, and diagnostic benefit to these patients.

Abhishaike Mahajan: I’m curious — these glioblastoma patients that you’re friends with — one device they interact with is probably Optune, and I’ve heard it kind of sucks because you have this constantly heated device near your head 24/7, above 18 hours a day. What other technology do they have that they potentially use to help their disease?

Ben Woodington: Not a lot. There are some things on the horizon that people have started experimenting with, that people have been on in trials — looking at ultrasound-type devices, blood-brain barrier disruption-type devices, convection-enhanced delivery devices. They’re not great.

Abhishaike Mahajan: No silver bullet.

Ben Woodington: It’s also just the quality of life and the patient impact. I don’t want to sit here and talk negatively about Novocure. I’m really happy that company exists. I’m happy that technology was created for those patients who are in desperate, dire need. The engineers, the scientists, the people that run Novocure — kudos to them for bringing a novel technology to those patients who desperately need it. And of course those patients are using it because it is extending their lives. But there’s so much more you can do to enhance the quality of life for those patients who don’t want to spend the rest of their lives traveling with companions to align stickers on their head and being affected by skin rashes and the pain associated with all of that. There is much more we can do for those patients.

[01:12:04] What was it like to raise money for this?

Abhishaike Mahajan: Back to the creation of Coherence — you talk to the providers, you see there’s demand. You talk to the patients, you see there’s demand. Now it’s time to raise money. Coherence feels like it’s in this weird place where the thesis is so strange that there’s not really many investors I can imagine off the top of my head who instinctively... did their PhD in this area and understand what you’re talking about. How difficult was it to raise money for a thesis like this?

Ben Woodington: They’ll come around. They’ll see what we all see and they’ll realize how large the pan cancer opportunity is. How hard was it? Both hard and easy.

Abhishaike Mahajan: What was your seed?

Ben Woodington: We did a pre-seed in the end of 2022, early 2023, which was about $2.5 million. And then we’ve just very recently closed our seed round, which was another $10 million. The investors that we’ve brought in follow the same trend that scientists, patients, and doctors all have with us. They’re cautiously skeptical at the beginning — hang on a second, does this work? — and then go in a very big way, get very interested, obsessed, both on the therapeutic opportunities but also on these data creation opportunities. We’re living in a world now where there are a lot of AI bio companies out there, and they desperately need data — novel datasets that are showing progression of disease and novel insights from human biology. That’s exciting to a lot of our investors as well. A lot of people are excited by BCI, but they’re all looking for what’s going to be the killer application. When people get that impression of us, they go all in.

[01:13:56] Beyond cancer: TBI, lung disease, and the pan-disease argument

Abhishaike Mahajan: Speaking of TAM expansion, one market is pan cancer. But I imagine there is a very reasonable logical leap you can make that membrane potential is probably important for a lot of diseases. Is that true? Could you make a reasonable argument that you don’t really need to do these five-year-long Alzheimer’s disease progression readouts — you can get a decent proxy from electrical readouts? Is that at all an argument people are trying to make?

Ben Woodington: It’s an argument that people are trying to make. It’s not something that we’ve done inside the company. But people are exploring electrical and other physical stimulation modalities in Alzheimer’s. People are looking at recording readouts for Alzheimer’s, Parkinson’s, other neurodegenerative diseases. And then of course there’s a whole host of neurological disorders that people are looking at — both electrical readouts and electrical stimulation — and other systemic diseases like diseases of the immune system and other things as well.

Elise Jenkins: The nervous system is involved in everything. I feel like it would not be a surprise to me that these types of interactions — you can pick up a whole host of things going wrong just by looking at nerves or neurons.

Abhishaike Mahajan: Is there any convincing evidence that, outside of cancer, if someone gave you a few million dollars to throw at another indication on top of what you guys are already doing, what would be the next thing?

Ben Woodington: There’s a company that just launched thats looking at targeting the nervous system for treatment of asthma and COPD. Before I did my PhD, I was actually working in lung diseases, drug delivery for lung diseases. I actually think that’s a pretty big opportunity. Chronic diseases — many patients are not managed particularly well. A lot of hospitalizations. There is evidence that you can stimulate certain nerves to relax the lungs, to bronchodilate. And closed-loop opportunities as well, predicting when someone is about to exacerbate. I think there’s a big opportunity there.

Elise Jenkins: Chronic stress or TBI. TBI is interesting.

Abhishaike Mahajan: Why TBI? Actually... I could fabricate an intuition for myself. I’d prefer you guys give one to me.

Elise Jenkins: I think traumatic brain injury is really interesting and has similar attributes to what you can leverage from glioblastoma. A lot of the time when someone has a traumatic brain injury, they’re already going in to put something into the brain — usually a shunt or something. There’s an obvious access point, which I always think is the biggest barrier to entry right now. Until this becomes more mainstream, that’s the biggest barrier. So that’s an obvious one — they’re going in and doing something already. And biomarkers.

And you can do similar types of strategies to suppress activity there. I’m also really interested in the data side of all of this — how diseases, degeneration, whatever else evolves. I think what would be really interesting with TBI is looking at how you can watch a brain go back to its normal, healthy state — what type of biomarkers give us that indication that something is going right, and how do you steer that. That’s really interesting in TBI. Chronic stress is because it’s regulated by adrenergic signaling. You can just target the vagus nerve or something else. I think that’d be quite cool.

[01:17:40] Hiring at Coherence + what is the hardest type of talent to find

Abhishaike Mahajan: That makes sense. One thing I’ve been curious about — I think I interviewed Hunter Davis a few months ago, the Until Labs cryopreservation guy. His company shares some similarity with yours in the sense of being wildly interdisciplinary in a way that very few other companies in the world are. He had some interesting thoughts about how hiring works in companies like that. I’d like to get both of your philosophies on what makes for people you want to join Coherence.

Elise Jenkins: I think probably curiosity and taking lessons from other industries. Some of our engineers come from the robotics industry. Some come from the med device industry. Some are scientists from completely outside of cancer neuroscience. And then we also have cancer biologists who really know cancer, but also know immunology and also know neuroscience. We look for people who are experts in their domain but also have demonstrated interdisciplinary overlap with multiple things. Robotics is a really nice example of that — you have mechanics, electronics, spatial interactions, and those types of things you have to consider in your design. The scientists are some of the most fun to find, because a lot of them are coming from the neuro background — that’s the kind of talent we seem to attract. But when you introduce them to this concept of these interactions that happen in cancer, people’s minds massively expand. Watching that process — when you start going through that in the hiring, or when you bring them on board, and how quickly they go from never hearing about it ever before to being so bought in, building and designing these crazy experiments to try and uncover some new neural biomarker — that’s been really cool to watch. Especially when you have this crazy idea many years ago that no one’s ever heard of, and you’ve got all these people that are super pumped about that discovery and want to build something that interfaces with that discovery. That’s been really cool. Mostly I’m looking for interdisciplinary. Yes.

Ben Woodington: Code-switching across disciplines is super important. It’s the same as a lot of deeply technical companies — it’s about your ramp of being able to learn. How steep is that? Because we have electrical engineers that need to come in and learn biology really fast. We have computational neuroscientists that come in and need to learn how to run what would be adjacent to clinical studies really fast. BCI and neurotech is a field that covers so many touch points — electrical engineering, neurobiology, to the sort of stuff that you do. It’s hard to find people that are willing to spread themselves across that many fields.

Abhishaike Mahajan: What do you think is the rarest skillset to find and/or to teach?

Ben Woodington: We know this because it’s the person we’re always trying to hire. Very good electrical engineers and embedded systems engineers. They’re hard to find.

Abhishaike Mahajan: Is it that there aren’t many hardware people?

Ben Woodington: I think a lot of the electrical engineers that have come out of Stanford or wherever get attracted by the tech industry. They’re often good programmers. So they go to Google or Meta or wherever. We need them when they’re at least a few years into their career, with a few projects behind them, a few product cycles, if we’re lucky. And most of them have gone into tech. Bringing them back into hardware is tricky. I think we’re seeing a shift now. Hardware is kind of hot again. Maybe in a year or two, there’ll be a bit of lag and then we’ll see more hardware people that we can bring into the fold. But it’s always the positions that we’re fighting most for.

Abhishaike Mahajan: I think some of the most talented people who have joined the companies I’ve been a part of have been ex-engineers at places like Cerebras, Uber, or the big SaaS companies. What is the big company in your field that you wish you could just pull all the engineers from to come work for you? Is there one, like Neuralink?

Elise Jenkins: I think Neuralink could be a good one, given that they’ve just taken strides in being the first ones to take both a high-density BCI and a robot into trial in a really short period of time. There’s a lot of things that those people would have learned along the way that could definitely be leveraged at a company like ours. I think there’s a challenge when you’re trying to do something really new, but it’s also a regulated technology. There’s this balance of being able to bring in people who really know how to build medical devices that are not scared of things that are new. That’s a really hard balance to find. There’s no company, maybe apart from Neuralink where that exists.

[01:23:17] What would you do with $100M equity-free?

Abhishaike Mahajan: The last question I have — if you were given a hundred million dollars, equity-free, to push this work forward as fast as possible, but you had to spend it within the next year, what would you spend it on?

Ben Woodington: Can I give one and a half answers?

Abhishaike Mahajan: You can have as many answers as you want.

Ben Woodington: This technology exists. The technology that we’re building fundamentally — there are no more science challenges. This is an engineering optimization piece now. Being able to get those technologies into as many human beings, as many cancers as possible — we could build such insane datasets. We could build such incredible real-time, real-world datasets that would blow a lot of people’s minds for what you can access from that data. You just can’t run that many trials all at once if you don’t have a hundred million equity-free cash. If you’re offering, I will take it. The other super exciting thing would be — fab floor, engineering integration floor, clinical scientists, clinic — all in one building. Everything in house.

Abhishaike Mahajan: Including a clinic?

Ben Woodington: A neuro-oncology clinic. That would be insane. I think you could do that for just about a hundred million if you did it maybe not in America. That would be incredible. Being able to highly iterate — build devices, build them in your own clean room, validate them, get them in patients really fast and start running studies, collecting data, and becoming that hub of those studies.

Abhishaike Mahajan: Why does it matter? Why do you care about having a clinical oncology suite inside the building?

Ben Woodington: So you have some control over the functions, the implants of the device, the same surgeons, quick readouts connected to your teams. When our preclinical and engineering teams are working in unison, it’s humming. You’re getting data out that the engineers and the computational scientists are analyzing overnight, feeding back into the next day’s experiments. That’s not really possible in clinical studies. There’s this barrier between you and the hospital, where you’re waiting for data, then you have to wait, then you have to submit new ethics to run a new study. Being able to turn that wheel super fast would be pretty exciting.

Abhishaike Mahajan: This is leading into a lot more questions, but I am just now realizing I never actually asked — is there an experimental loop that goes on? In rodents, at Coherence — where you design one version of the device, implant it, see how well it works?

Ben Woodington: Constant iteration. Both on our preclinical devices — where we’re recording data from these animals, running new stimulation regimes — and on the primary product development pathway as well. Both of those have tight iterative loops.

Abhishaike Mahajan: You exist amongst many other neurotech companies, and you’re probably the most alien amongst them. Do you pay attention to most of the neurotech research that’s going on outside of your immediate field, or is it not super applicable to what you are doing?

Ben Woodington: Firstly, I take it as a great compliment to be called the most alien neurotechnology company. That’s good. Secondly, both of us are having conversations almost every day about what’s going on in the field. It’s entirely relevant, both from a technology landscaping exercise and from a cultural landscaping exercise — which indications are getting more heat in the use of neurotechnology, where are people most excited, what are the innovations convincing more clinicians and patients to adopt these technologies. We need to be abreast of all of this, because there are some similarities with the technology stack and how it’s introduced to the patient as well.

[01:27:15] Are you a neurotech company or a cancer company?

Abhishaike Mahajan: Do you think you’re a neurotech company with ambitions to attack cancer, or a cancer company with ambitions to use neurotech?

Ben Woodington: I personally am a neurotechnologist that wants to develop technologies that can help a lot of people. And oncology seemed like the fastest and highest-impact route to get there. If I can speak on behalf of Elise — and maybe she’ll say I’m wrong — I think Elise comes more from an “oncology matters, and I’m going to use whatever tool I can to help these patients, and this makes sense” perspective. Is that an accurate read?

Elise Jenkins: I think so. I don’t know why the two have to be separate. There are so many debilitating conditions and diseases that need attention. There are two major diseases in the world that are causing death or suffering for a lot of people — cardiovascular disease and cancer. I feel like we can have a really big impact here by leveraging technology that is well-established in other indications that could have huge potential in cancer. I fit in either camp. I want to develop technology that will benefit people.

Ben Woodington: That’s fair. I’m more neurotechnology-pilled. Neurotechnology is crazy. Why are we not using it in all these indications? It’s amazing. It’s going to change everything — from the extreme cases that some of the neurotechnology and BCI companies are making to just day-to-day medicine. I just think that cancer is an extremely promising way to get there and to scale these technologies into a lot of people.

Abhishaike Mahajan: Do you suspect that the full landscape of possible perturbations is pretty limited and you’ve discovered most of them, or you may actually expand that over time?

Elise Jenkins: Initially it’s looking at well-established regimes. If you were to take what Setpoint or Galvani were doing in vagus nerve stimulation for rheumatoid arthritis — they’re targeting immune response there, with well-established parameter sets that are published in literature and have been done in humans. Those are the safer bets that you’d want to try in a novel indication. We are starting with those types of things, with some variations that depend on the nerve that you’re targeting, whether you want to increase immune function or immune activity or decrease stress. They’re very different types of stimuli that you’d apply, but they are well-established.

Ben Woodington: It’s actually a real problem in clinical programming generally — not in our field, but in other fields, for example, pain. The clinical programming profession hasn’t caught up with the engineers. You’ve got more and more complex devices. You’ve now got hundreds, in some cases, of electrodes with thousands of different potential waveform characteristics that you could apply to each electrode. Which gives you this multi-billion parameter operational space. And then you’ve got a clinical programming nurse sitting there saying, where do I even start on this? There’s this massive space now that I have to operate in to try and treat the pain of this person. It’s a job that probably will end up being done by some AI model down the road, using some sort of Bayesian optimization — not a nurse going, “does it feel better or worse by doing this?”

Abhishaike Mahajan: Is that currently how it’s done?

Ben Woodington: It’s currently how it’s done. You would be quite surprised how much human-in-the-loop there is in electrophysiological medicine, where you’ve got people watching a screen saying, “I think they’re going to have a seizure soon.” And someone else going, “well, better stimulate their brain to stop that happening.” And there’s no model, no computer really in the loop giving early indication.

Abhishaike Mahajan: You mentioned a bit about what gives you anxiety, Ben. I’m curious what gives you anxiety, Elise, or if you’d like to add to your answer.

Elise Jenkins: I think for me it’s maybe a combination of anxiety and frustration. You want to move as quickly as humanly possible. The impact that we need has to happen in humans. You need to get to humans as quickly as possible, but you don’t have all the answers in that design process. That is frustrating and can be anxiety-inducing. You’re having to make some assumptions about what might happen in certain scenarios, or how to design this implant, and it has to be safe, of course. That iteration — you just want to get to humans as quickly as possible, but you have all of these things that you need to consider. That’s frustrating, drives me a little insane.

Ben Woodington: I totally agree with that. You work in the cancer field yourself, correct? We’re not in the ads business. We’re not interested in pumping out a few extra targeted ads to people. We’re in the game of actual human beings who are dying quickly. And we’re trying to get technologies that can help those patients as quick as possible. That is frustrating. That is anxiety-inducing, especially when you maintain a close connection with those patients and you see those patients dying. And then you’re screaming at people in the office to move quicker because you’re very connected to that.

Abhishaike Mahajan: I don’t think I have any other questions. This has been an amazing conversation. Thank you so much, Elise and Ben, for coming on.

Ben Woodington: Thank you so much. It’s been a pleasure. Had fun.

LatchBio is building agentic scientific tooling that can analyze a wide range of scientific data, with an early focus on spatial biology. Check out their agent at agent.bio! Clip on them in the episode.

If you’re at all interested in sponsoring future episodes, reach out!

1. Introduction

2. Links

3. Timestamps

4. Transcript

Introduction

This is an interview with Matthew Osman and Fabio Boniolo, the co-founders of Polyphron.

The thesis behind Polyphron is equal parts nauseating and exciting in how ambitious it is: growing ex-vivo tissue to use in organ repair.

And, truthfully, it felt so ambitious as to not be possible at all. When I had my first (of several) pre-podcast chats with Matt and Fabio to understand what they were doing, I expressed every ounce of skepticism I had about how this couldn’t possibly be viable. Everybody knows that complex tissue engineering is something akin to how fusion is viewed in physics; theoretically possible, but practically intractable in the near-term. What we can reliably grow outside of a human body are simple structures—bones, skin, cartilage—but anything beyond that is surely decades away.

But after the hours of conversation I’ve had with the team, I’ve began to rethink my position. As Eryney Marrogi lines out in his Core Memory article over Polyphron, there is an engineering system that has reliably produced viable human tissue for eons: embryogenesis.

What if you could recapitulate this process? What if you could naturally get cells to arrange themselves into higher-order structures, by following the exact chemical guidelines that are laid out during embryo development? And, most excitedly, what if you didn’t need to understand any of these overwhelmingly complex development rules, but could outsource it all to a machine-learning system that understood what set of chemical perturbations are necessary at which timepoints?

This does not exist today, but Polyphron has given early proof points that is possible. In their most recent finding, which we talk about on the podcast, their models have discovered a distinct set of chemical perturbations that force developing neurons to arrange themselves with a specific polarity: just shy of 90°, arranged like columns. This is obviously still a simple structure—still a difficult one to create, given that even an expert could not arrive to that level of polarity—but it represents proof that you can use computational methods to discover the chemical instructions that guide tissue self-assembly.

We discuss this recent polarity result, what the machine-learning problems at Polyphron looks like, and the genuinely insane economics of the whole endeavour. The last of which is especially exciting; it is rare you hear biotech founders talk about ‘expanding the Total Addressable Market’, and actually believe them. But here, it is a genuine possibility if the Polyphron approach ends up working.

Enjoy!

Links

Spotify

Apple Podcasts

Substack/Transcript

Timestamps

(00:02:16) Introduction
(00:02:37) Why replace tissue rather than the whole organ?
(00:10:34) Why not do simple stem/progenitor cell injections?
(00:13:51) Can organs repair themselves naturally?
(00:18:21) What does “structure” actually mean in tissue engineering?
(00:21:04) Why are skin and bone the only FDA-approved tissues today?
(00:23:45) What exactly are tissue scaffolds?
(00:27:52) Why are organoids a “dead end” for this field?
(00:35:08) The argument for recapitulating developmental biology
(00:40:28) Walk us through the Polyphron experimental loop
(00:47:56) Can you simulate morphogenesis with only small molecules?
(00:49:49) How large is the set of possible tissue scaffolds?
(00:52:32) How reliable are developmental atlases?
(00:56:45) What is the machine learning model actually optimizing for?
(01:04:04) Polyphron’s first big tissue engineering result: polarity
(01:15:33) What comes after polarity?
(01:17:09) Why is vascularization the hardest problem of tissue engineering?
(01:20:33) Why can’t you just wash angiogenesis factors over the tissue?
(01:22:25) How does the graft integrate with the host’s blood supply?
(01:25:45) How do you validate tissue function before implantation?
(01:29:01) How do you design a clinical trial for a biological pacemaker?
(01:37:01) The argument for being a pan-tissue company
(01:41:57) What are the biggest scientific and economic risks?
(01:45:23) Who are Polyphron’s competitors?
(01:47:07) Expanding the TAM beyond transplant lists
(01:52:28) Autologous vs. Allogeneic approaches
(01:55:07) Is a 3-year timeline to the clinic realistic?
(01:56:28) Cross-species translation
(01:58:05) What would you do with $100M equity free?

Transcript

[00:02:16] Introduction

Abhi: Today I’ll be talking to Matthew Osmon and Fabio Boniolo, who are co-founders of Polyphron, a startup working to grow ex vivo tissue to use in organ repair. We’ll be talking about the history of the tissue engineering field, the science of existing approaches, and their argument for why computation is necessary for the field to move forwards. Thank you for coming onto the show.

Matthew: Thank you for having us. Happy to be here.

Fabio: Happy to be here.

[00:02:37] Why replace tissue rather than the whole organ?

Abhi: The first question I have is somewhat of the obvious one: why do tissue replacement as opposed to either cell replacement or full-out organ replacement?

Matthew: Right. So, first, let’s give an overview of the kinds of conditions and diseases where tissue replacement is potentially warranted as a strategy. There are a lot of them. So it’s any indication where there has been tissue loss, 3D architecture loss, and where function is downstream of that 3D architecture. You’re looking at indications where really it’s not plausible that you could drug your way out of the fibrotic tissue. So you just can’t plausibly drug scar tissue in the heart back into being a beating heart.

Now on the other end, you have kind of whole organ transplantation, which is the existence proof that tissue replacement at all—writ large—works as a strategy. There are obviously some serious limitations there, particularly around supply, eligibility, of course, the need for lifelong immune suppression, and ICU time. Organ transplantation is enormously invasive; thoracic or abdominal surgery.

And so our thesis is that actually for certain indications—a lot of the chronic age-related diseases—what you should really try and do is build functional units of tissue and intervene a lot earlier. So instead of waiting for total organ collapse, which is basically the organ transplantation model, what you do is you identify early focal forms of damage, which are highly predictive of eventual failure, and intervene at that level. With modular tissue replacement, that can essentially prevent the collapse of the organ occurring in the first place.

So places where this makes sense: heart—so scarring after a myocardial infarction, left ventricle tissue there. Advanced CKD, liver, fibrosis in the lung, COPD, pancreatitis, a bunch of CNS disorders around trauma, and in the retina. Am I missing any? There are like a hundred million people in the US with these chronic diseases that involve organ damage. And so drugs, we don’t think can modulate; devices, we think can route around the problem, but never restore full function.

Abhi: It makes instinctive sense why you would want to replace individual aspects of an organ. If most of it works fine, there’s just a tiny few bits of it that don’t... The reason I thought people usually opt for full organ replacement is more that they don’t understand what bits are damaged and what bits are not damaged, and so they just want to replace everything outright. But you told me that that is in fact actually not true. There are usually pretty clear markers of damage that you could just excise and replace, and I’d love to just hear you repeat that.

Matthew: Yeah. So, that’s absolutely the case. So you are really looking for places where there is focal damage before it gets fully diffused. Now, obviously when an organ is close to full collapse, damage is probably so widespread that it would be difficult to select individual portions to resect and replace with replacement tissue. But there are many, many steps before that.

So you will see clearly in imaging, damage to heart tissue in ischemic cardiomyopathy, and also all the other indications that I mentioned, with a couple of exceptions. So you should be able to locate focal lesions, focal damage that you could potentially treat with this replacement strategy—maybe across more than one site, right? We’re not saying it’s necessarily one site per organ, although within the heart I think it probably would be one site, but in the kidney it’s probably multiple sites. But you would intervene early enough that it becomes a tractable problem.

Abhi: And for the example of left ventricular damage... those are pacemaker cells, correct?

Matthew: Uh, so left ventricular tissue is actually not pacemaker cells. Left ventricular tissue is the tissue that is very often damaged in a myocardial infarction, in a heart attack. So what you get is reduced ejection fraction, so your heart stops pumping enough blood out of the ventricle. And that causes obviously a huge amount of health problems down the line, leading eventually to death.

Abhi: And the damage you would see is so visually apparent that a doctor could just see like black speckles or something and just extract that out?

Matthew: It could be visually apparent to a surgeon. It’s certainly apparent to all of the diagnostic tools currently used in modern surgery. One of the things that we wanted to do when trying to build a platform—because we are trying to pioneer tissue blocks, functional replacement tissue as a new modality—is require as few people to change what they do as possible. And so we’re always looking for ways that we can piggyback off existing reimbursement pathways, surgical workflows, et cetera.

So yeah, all the indications that we’re initially looking at have whole organ transplantation as a covered treatment. Because what you really want to do is intervene kind of much, much earlier. Just to make it super concrete: there were 48,000 organ transplants last year, give or take a few hundred. Of those, maybe 10,000 might have been heart transplants. You have 6.7 million people in the US with heart failure. So that’s anything from New York Heart Association category one all the way up to category four.

What we are proposing initially doing just in the heart case—and we have multiple tissue types we’re going after—is to intervene just before they would need a heart transplant. So it would be a deferral strategy for that specific product. We have other heart products we’re working on, including a pacemaker actually, which we can chat about. But for that product, what we’re trying to do is act as a bridge to prevent them from needing a heart transplant. So it’s either a deferral strategy for someone who needs more time, or it is an alternative strategy for someone who is not able to get a heart transplant—so either for comorbidities, for age, for adherence, or psychosocial reasons, which is one of the eligibility criteria that would prevent you from getting a heart transplant.

And then what you would do is you would kind of move earlier and earlier in the progression of the disease and much more minimally invasively. So all of these tissue products are delivered through surgical workflows that are considered to be minimally invasive. So in our case, it’s not a full thoracic—you’re not cutting open the sternum. So it’s much, much easier to slot into that existing surgical workflow. And it’s already reimbursed and there’s an anchor price. So you know how much a heart transplant is going to cost, which is 1.6 million in the US plus everything else. So you have, for the insurer, a really strong argument about “this is how much you should pay to defer that happening.” And it means that you get a product in the sort of low to mid six figures, which is important because you’re having to figure out how to manufacture this stuff at scale.

[00:10:34] Why not do simple stem/progenitor cell injections?

Abhi: I think the economics here are really crazy in the sense that you get access to this entire patient population that currently no one is really able to touch. I think that is something I will want to talk about a little bit later.

I think the craziest part about Polyphron is this extreme importance of tissue structure. Which I did not naively appreciate before talking to you. Before speaking to especially Fabio, I kind of had the belief of, well, why can’t we just squeeze in some progenitor or stem cells into the site of the tissue damage? Like excise the tissue, squeeze in the stem cells. The body will figure out how to work with it. Why doesn’t that work? Obviously it doesn’t seem to work, but *why* doesn’t it work?

Fabio: No, absolutely. It would be great if it worked. Unfortunately, it does not. And it has been tried, especially in the heart. So people have tried to inject cardiomyocytes into the infarcted area to see whether there was any regain on function. And this did not happen. And it really goes back to the fact that in nature, in vivo, cells really exist in a specific environment within which they can perform whatever function they’re supposed to perform—meaning they can proliferate, they can commit or differentiate into a specific lineage, they can grow, they can assemble, so on and so forth.

And this relation between the microenvironment and the architecture within which these cells grow, and the cells themselves, is something that is established throughout development. And of course, this is something that is lacking at the injury site after any type of traumatic event. Therefore, cells that are just injected there are not able to assemble properly, are not able to signal the microenvironment their presence properly, and they’re not able to learn what the microenvironment is telling them properly. And therefore they’re just basically unable to understand what to do. They’ll just be either washed away or they will start moving around and then they either die or they are killed by the organ.

And this is quite important because, again, one of the things that we really have at the core of the company is this idea that structure is the fundamental underpinning for any tissue engineering approach that has the potential to cure people.

Matthew: I would just kind of piggyback off that—Fabio mentioned the work that was done in cardiomyocytes. So structure is important for function; it is also incredibly important for safety. So particularly where you derive the function of the tissue from its structure—so anything that has signaling or conductivity—if you don’t have structured tissue, what you get is incredibly aberrant effects that are really, really damaging. So in the iPSC cardiomyocyte work that was done, you get arrhythmias because heart tissue is part of an electrical system. And likewise, if you try and inject excitatory neurons that aren’t in a proper structure, you get epilepsies. So it’s very, very important from a safety profile to have as close as possible to native in vivo morphology.

[00:13:51] Can organs repair themselves naturally?

Abhi: Is there any organ that, if damaged, will be able to repair itself to some reasonable degree? Like most of organ development happens while you’re an embryo... is there any ability for repair to happen after you’re born?

Matthew: I mean, the liver is probably the case of fairly persistent regenerative capacity. It is an extreme outlier. You just do not see that in the heart or the lung, for example, to nearly the same extent. So the liver is a very specific case. For what it’s worth, the liver is a really interesting example of an existence proof of not needing to replace the entire organ in order to get clinical benefit, because you have had successful liver segment transplants for a while. So that actually gave us sort of comfort that there’s not something that we would be missing by not having to fully recapitulate the entire organ before doing the transplant. There are also examples in the intestine, I think, and a couple other organ systems as well where segmental transplantation gives you a huge clinical uplift.

Fabio: So we... I think we will be talking about quite a few different axes of complexity today, but I think one of the most interesting ones is the regenerative potential of these different tissues. You can put them on a continuum going from, as what Matt was saying, from the liver to the heart. Now the interesting thing is that people noticed quite early that a few organs could actually be intervened on by having some type of structured framework scaffold with isolated cells coupled in. So much so that the first real applications of tissue engineering, which were probably in the 1980s, were all about getting specific plastics or biomaterials, filling them with cells and putting them in animals to see whether there was any regenerative potential.

And you know, these applications were called *Chimeric Morphogenesis*, just to remember this idea of trying to recapitulate what happens in development that leads to tissue formation. So all of these kind of small elements point to the fact that we can indeed regenerate organs. And I think the exciting thing and cool thing about tissue engineering is that we are not necessarily asking the organs themselves to regrow tissue, but we’re rather using engineered solutions to support regain of functionality.

Abhi: The primary thing I was trying to question was: on one hand, the heart is not able to repair an aorta by itself, so you need to go in and replace it. On the other hand, given the fact that sometimes you do see this—like the liver is able to partially repair itself—is there any pathway to being able to convince an organ, primarily through genetic or chemical means, to repair itself? Or is that just out of the picture? Is there no developmental pipeline that does that, or is it kind of unknown?

Matthew: I mean, it hasn’t worked. I can tell you that. I think that if it were to work, it is most likely to work in the liver, but unlikely to work in other organ systems where the niche of the damage is so fundamentally changed from the developmental program that it’s hard to know how you would kind of act on it in the way that you’ve just described reliably. So it’s somewhat of an unknown right now, but I can tell you it hasn’t worked.

Fabio: I would speculate there is some threshold below which the regenerative potential is not enough to actually bring back functionality. In many of the chronic inflammation indications that Matt was mentioning, and especially in their acute phases, there is necrosis happening. So the actual focal location in the organ dies, it stops working. Therefore, there is not much the body can do.

[00:18:21] What does “structure” actually mean in tissue engineering?

Abhi: That makes sense. And when we like vaguely gesture to “structure” and “tissue”, what does that actually mean? One axis is clearly that there exist multiple cells and multiple cell types in this environment. What other types of structure exist?

Fabio: This is quite important and actually it is so important that at Polyphron what we are doing is trying to establish metrics that can tell us how close our grafts are to in vivo structure. And the way we look at structure is really at the three-dimensional architecture and composition of the tissue. Where by this we mean we look at how different cell types are patterned within the tissue, how they locate themselves with respect to each other, and whether they have specific orientations or polarities—so where there are specific distributions of proteins that tell us what is up and what is down. And we can see that these elements—so polarity, multicellularity and cell composition, and of course also the layering and the geometry of the tissues—are something that happens across different organs and different tissues. Of course with different features, but macroscopically, we can identify these. And one of our hypotheses is that really we should try to recapitulate all of these single steps towards our goal of recapitulating tissue structure.

Abhi: Do you think there are dimensions of structure that are not well either currently unknown today, or not legible entirely and you need a model to encompass it all for you?

Fabio: As many things in biology, it really depends at the resolution level at which you look at tissue. We have had decent ways to measure tissue morphology or tissue structure for quite some time, either imaging-based or fluorescence-based. And what we’re learning with more and more powerful technologies such as electron microscopy, for example, is that we can really go down and look at the nanoscale organization of these tissues. Now one question is how relevant it is to understand all of these different scales to actually be able to recapitulate structure in the lab or manufacture it. But it is for sure true that we can see this continuum of complexity scales. And as for many complex systems, the macro features and behaviors we see really arise from this continuous scale of complexity.

[00:21:04] Why are skin and bone the only FDA-approved tissues today?

Abhi: There is an existing proof point today—beyond just an organ’s ability to regenerate or organ transplantation—that you can do this sort of fractional replacement, and it has only popped up in, as far as I can tell, three areas: skin, cartilage, and bone grafts. These all exist. There seem to be FDA-approved products in the market that do this. Why hasn’t the tissue engineering field moved beyond these three?

Matthew: So, I can give the very naive response, which is that those are easier to engineer for a couple of important reasons. One, skin is very thin, which means you don’t have to solve the vascularization problem of perfusing vasculature that you have to solve for thick tissue. I mean, something like cartilage as well often doesn’t have blood vessels, so there you aren’t having to solve the vascularization problem either. They’re metabolically less demanding tissues to produce as well. And you can do a lot of this work in thin 2D sheets, which is what some of the original skin work was done with.

Fabio: Following up on what Matt is saying, he has identified two more axes of complexity, which are metabolic demand—we don’t need vascularization, which is a bottleneck for any tissue engineering approach, and I’m sure we’ll discuss that. So we don’t need vascularization as much for these products. And also they have a relatively simple structure again. So there’s a relatively small number of cell types. They’re organized in a very, in a relatively simple configuration—so there are ways... it’s simple layers.

And for example, for bones... I actually had my first experience in tissue engineering in a bone graft production company. And the incredible thing about bones is that there is this mineral inorganic component that we can find elsewhere in nature. Bovine bone is exactly the same basically as human bone in terms of mechanical features. Corals can be used as bone graft, and they’re just the perfect scaffold to put in the body. There are maxillofacial applications, there are spinal cord applications for the bone component of course. And it is quite simple to insert them and have the body repopulate them and basically make them their own. So you know, in a way they were the low hanging fruits of tissue engineering. And now the challenge is kind of on us to go to the more complex structures.

[00:23:45] What exactly are tissue scaffolds?

Abhi: We’ve mentioned scaffolds a few times in this conversation. I probably should have asked this question earlier. What *is* a scaffold in the world of tissue engineering?

Fabio: So the scaffold is basically one of the key elements of tissue engineering. If we look at the field in general, you basically need three elements. And then every single approach combines these in different ways. These elements are: some type of isolated cells [whether iPSC derived or primary cells] , you need some type of bioactive factors that help make whatever you want to make, and then you need this kind of scaffold, which is basically the framework or the structure that cells need to develop, mature, and make the tissue you want.

Now, in tissue engineering, this scaffold is looked at as a structure that can be either a plastic or a biopolymer that can be transplanted, or it is a nature-derived material such as collagens. And you can imagine them as sponges or basically 3D porous matrices that you can use to seed cells in. You can use them to create gradients. And you can also use them to tune the mechanical and physical properties of the sponges so that cells receive very well-defined stimuli. And the hypothesis in the tissue engineering field has been: if we give the initial structure to the cells and then we let them do their thing, basically, they will remodel the scaffolds on their own. And then what we will get out at the end is the desired graft that might be then transplanted. And what people quickly realize is that unfortunately, this is too much of an artificial setup for the cells. So they will not be able to actually go and remodel and restructure these scaffolds. They will kind of go midway, and the product will not be as effective as a real graft might be.

Abhi: What do you mean when you say the scaffold is “too artificial”? Like what does artificial concretely mean?

Fabio: So in this case, what I mean is that what happens naturally in vivo—once again looking at development—is that the scaffold [which in this case for natural tissues means the extracellular matrix and the environment within which cells grow and proliferate] evolves and changes together with the cells that are developing and committing to specific cell lineages.

Abhi: So it’s not just secreted during morphogenesis and then populated by the cells and it stays static?

Fabio: No, it actually changes throughout development. So the physical properties, the stiffness of the scaffold changes because it has to support different cell types and different functions. And this change will be impacting developing cells, but will also be impacted by developing cells. And what we are seeing and what we hypothesize is that this complexity is really a process that basically reaches equilibrium through these different steps in vivo. While what I was describing earlier of this artificial setup where we give the scaffold from the outside and hope the cells will grow inside, is a very kind of non-natural setup where we’re trying to define complexity from the top down and not having it grow and stabilize on its own.

And one of the things we’re trying to do at Polyphron, or that we’re actually building with our technology, is really a way to allow cells to create their own three dimensional microenvironment, their own scaffold, and therefore also the structures that are relevant for function.

[00:27:52] Why are organoids a “dead end” for this field?

Abhi: And I think gesturing back to the current FDA approved products, most of the way that those worked is like bioprinting—layering on one layer of cells at a time works well for those particular cell types. Obviously it doesn’t scale to more complicated ones. One of the other approaches people seem to be working on is organoids, and Matthew has not positive opinions about organoids with tissue engineering, and I’d love to get your take on them.

Matthew: Uh, so I think that organoids are a dead end for therapeutics. I think that as a strategy, it is just not gonna lead to meaningful therapies that could bring the need for organ transplantation to an end. I think that there are definitely some useful drug screening use cases with organoids, but for a bunch of reasons they lack the complexity and in vivo structure that you would need to get any of these functional restoration effects that we think that we need.

Abhi: But you do see the organoids are willing to like mangle themselves into some sort of structure. And so you do get something that’s clearly better than single cell replacement. Why is that not enough? Like where does that start falling apart?

Fabio: Yeah, so you are right in saying that what we’re seeing with organoids is some type of self-assembling behavior. Um, and this is mainly dictated by trying to intervene on typically pluripotent stem cells in a way that simulates how different cell lineages come to be during development. The, there are multiple issues though, with this approach. Um, the first one is that, once again, development really is successful because cells develop not in a vacuum, but in a very specific microenvironment. And this is not recapitulated in typical organoid cultures. Cells will grow in collagen or in some type of extracellular matrix, but they will not be receiving the chemical and especially the mechanical stimuli that cells need to create structure in vivo.

Secondly, what you typically see in organoids is that due to this self-assembling behavior, you will see sporadic structures arising. And here with structure, I mean micro features that resemble natural structure. What we’re missing though is the micro features that kind of put all of these smaller components together. So I can give an example. If you take for example, kidney organoids, you will see within an organoid some cells that make glomeruli-like structures, some other features that will make renal tubule-like structures. But we will be missing the union between structure one and structure two. And this of course, is quite important for what we want to put in in vivo because we need that graft to be able to accomplish its function.

And this is not always possible, actually, it is not possible with organoids. Um, and one more complexity is that organoid cultures are still an intrinsically stochastic process. And of course this puts quite some limitation in terms of approval and in terms of manufacturing. So one of the things that we’re thinking quite often at Polyphron is how to make the whole process of production of replacement tissues as robust as possible so that it can be, you know, a proper technology that can scale.

Matthew: Sorry... I was just going to—maybe we’ll come to some of the commercial manufacturing challenges later, but one of the things I also wanted to point out that I think is unusual about the way that we are designing the various loops that will allow us to build functional tissue units is that we are taking cost into account as part of the cost function of the overall loop. So we’ll talk about like how we try and recapitulate morphogenesis, I’m sure. But one of the things we’re really trying to do as well is to select the cheapest way down the mountain. So we take into account the price of reagents, we’re optimizing the pathways to recapitulate morphogenesis because the last thing we want is to pull off this kind of technical miracle and have a commercially non-viable product at the end of it. So we’re like trying to build in manufacturing COGS viability even in our initial ML approach.

Abhi: Have potentially organoids... like do they not work at all? Has it ever been successfully—or like ever a transplantation has ever happened and it just didn’t take? Like the native functionality was not restored, or has it still never been tried?

Fabio: There are examples in rats, specifically in the intestine where there seems to be integration. Restoration of function though is, has not been proven yet. Okay. So these, these organoids are recognized as self, they’re integrated, but you know, there is no real restoration of the functions that they’re supposed to carry.

Abhi: Yeah. So it sounds like there’s multiple hard problems. You first need to grow the tissue in the first place. And then the second hard problem is you need the body to be willing to accept that tissue and to integrate into the rest of the body. Is that connected to the structure problem or is that an independent thing?

Fabio: No, it is fundamentally connected. I would actually say it is... you know, structure is the underpinning element to being recognized and integrated properly. And this is because if you think of organs, they work as this extremely well integrated kind of setup. And the question is what is the smaller functional unit we can use that can be recognized, connect and basically restore function. And we believe that, you know, in order for this to happen, the graft should be as similar as possible to what was lost. In terms of structure, in terms of cellular composition. And this should ease the way that the body integrates and makes the graft its own.

[00:35:08] The argument for recapitulating developmental biology

Abhi: Okay. So it seems like bioprinting is too simple for more complicated things. Decellularized scaffolds are also potentially too simple to do the most complicated things. Like scaling up organ transplantation via xenotransplantation is both super technically risky and also... you don’t wanna do organ transplants for everyone. What is the way out of this conundrum that you’ve set up where every approach is either too simplistic or too complicated?

Matthew: Our approach...

Abhi: What is it?!

Matthew: It’s to make much smaller functional units of tissue that are recognized as self and integrate and restore function. So to give you a rough sense of the order of magnitude, these are tissue chunks in the sort of centimeter-cubed volume or less, right? We’re not trying to build entire hearts in terms of biomass. But yeah, so that’s the way out of the conundrum. You have to be able to do it in a repeatable way with exceptionally low variability. You need to be able to control COGS. And ideally you should try and get this as much clinical effect as you possibly can with the smallest unit of tissue, because the less biomass you have to produce, the cheaper it’s going to be. Whereas we think the pricing will probably stay where it is, because there are these anchor prices for transplantation, for assisted devices.

Abhi: But it seems like the core tenant is almost like: if everything else is too simple to recapitulate developmental biology, and now your answer is basically “let’s just recapitulate, let’s just do developmental biology straight up.”

Matthew: So if I had to kind of sum up like one of the precepts of the company, it’s that there is an engineering system that has already produced functional human tissue—and that’s human development. And why don’t we try and recapitulate that as much as possible? And that over the past few years, the data sets that allow us to have at least a fuzzy starting prior of what development does have come online and become available. So these are developmental atlases that are often multi-omics based or including increasingly spatial transcriptomics. And so the prior kind of set of tissue engineering approaches involve this highly mechanistic understanding where you’re trying to smooth out the complexity of what’s happening in development to fit it in the brains of the scientists that you have working on the problem and to kind of fit it experimentally.

Um, our view was that now the data sets are rich enough and wide enough that you can start throwing them at some of the exciting new architectures that we are seeing and have models learn latent rules of morphogenesis in a way that doesn’t need to be legible to a human. This obviously needs to be very, very tightly paired with wet lab validation, which is something that we are super explicit about. We have a kind of a closed loop between the developmental references—which are kind of our fuzzy priors—and what’s being tried and validated in the wet lab in this loop.

But our view is that now there’s a plausible path to us being able to start from what happens in development, potentially find alternative pathways to achieve the same goal—which is like super, super exciting—and eventually end up with a functional unit which is similar to what nature produces. And one advantage, and I’m sure we’ll talk about kind of like model architecture and some choices that we’ve made there, is that we have a strong hypothesis that there will be like an ultimate latent logic of development that models will learn when they see more and more tissues. And there are kind of already hints of that in some of our experimental data, that you might be able to kind of transfer things across tissues. We don’t yet know, I should be clear, whether, you know, you only need to do three tissues and then you can do everything. Or whether it’s you can do 10 and therefore you can do 20. But we do fully expect there to be transfer learning across tissues as we go.

Abhi: I think the pan-tissue aspect of Polyphron is something I really wanna talk about because I think it’s one of those crazier...

Matthew: It’s kind of insane. But actually I think it makes sense because the human body doesn’t produce a liver separate from a kidney, right? The human body is an engineering system which is holistic and comprehensive. Um, and therefore you would expect there to be redundancies and kind of the same techniques across different tissue types. And it’s very, very important for us that the space of interventions that you can use to manipulate morphogenesis is bounded. It is finite. And in natural development it is by definition.

[00:40:28] Walk us through the Polyphron experimental loop

Abhi: To look at Polyphron’s experimental loop with more of a concrete lens... Like what do you start... like you have a box of like collagen or something, you start as the existing scaffold. You seed that with induced pluripotent stem cells, iPSCs. What’s the next step? Like, let’s say you’re trying to produce some functional heart tissue. What would you do next?

Fabio: Yeah, so the first step really is to look at developmental atlases, where we are looking at single cell atlases, spatial and transcriptomic atlases of the developing human heart. And this allows us to first understand which developmental time points have been sampled. And, you know, this then dictates what type of dynamics and what type of lineages we can try to recapitulate in vitro. We then use different types of computational approaches to mine these high dimensional data sets and extract temporal trajectories and dynamics that we care about—these being specific lineages, when they arise and when they commit, or specific microenvironmental interventions or perturbations.

And then we move to our in vitro setup where we have, as you’re saying, these kind of three dimensional boxes within which we can use different types of extracellular matrices depending on what microenvironment—or what developmental microenvironment rather—we might want to try to simulate. And we then seed these scaffolds with different types of either progenitors, pluripotent, or committed cells, depending again on which type of cell type we want to recapitulate, which type of structure we might want to achieve. And what happens then once we have our cells in this kind of 3D box, is that we can start perturbing them using the same kind of perturbations that we have learned are effective during development based on our atlases.

Abhi: But the developmental atlases are, as far as I know, telling you like the ligands that it sees on like day 15 of heart development. How do you relate that back to like a causal relationship that like these ligands caused—like was essential to day 15 of heart development?

Fabio: So, couple of things. What basically state-of-the-art computational biology approaches allow you to do right now is to go from a discreet sampling of a developmental trajectory to a continuous trajectory. So you can really start to see kind of continuous dynamics, whether there are peaks, valleys, whether there is a steady state at one point in development. And this then tells us basically which molecules to apply when. But one important thing to your point is that we really do not care about understanding the relationship. Because what we want to do is to just define the broadest set of interventions that might matter for the structure we want to recapitulate. See how those perform on our cultures and then optimize based on that. So that is what Matt was saying: our atlases just become our basically first and initial prior and then we quickly move into the lab, we start generating relevant data, and then we optimize on this data only so that we can see how different interventions when combined in a certain way, give us certain structures that we can then optimize on, select for, et cetera.

Abhi: So you have like a box of scaffold with cells on top of it. You have this developmental atlas that tells you like at each of these time points what ligands was noticed in that developmental environment. And you sample those and apply them to your Polyphron sample and just see how well it recapitulates like the native tissue.

Fabio: That is correct.

Like, what you can do potentially is actually do this at a whole transcriptome scale. You don’t necessarily need to focus on ligands. We focus on ligands because again, we have this quite strong hypothesis that the microenvironment is what matters. Not only cell intrinsic transcription factor related dynamics. But yeah, the other advantage really with ligands is that you can get small molecules for them to simulate their activity oftentimes. So it’s okay to actually go out buy them and then apply them to our experimental setup.

Abhi: My impression is that there are just like thousands upon thousands of small molecules going on inside an embryo while it’s developing. Do you guys have the ability to also put in thousands upon thousands of small molecules in your tissue sample? Or is it like you picked up like a dozen?

Fabio: Yeah. Okay. So, let me first say what happens in vivo and then there’s kind of a jump to what we do in the lab. But spoiler, the jump is mainly due to our current constraints in terms of teams and instrumentation. Um, but what happens in development is that there is actually—and also Matt was referring to this—there is quite a finite space of molecules and pathways that is activated at specific time points to get the, basically to get cells through morphogenesis. And the other interesting thing is that there is a lot of redundancy. So there are parallel pathways where one might be active, the other one might not be active. And all of this basically becomes a quite constrained space to start from.

What we then do is to pick from this list of molecules and proteins that are activated, the ones that we can easily source. The ones that we can cheaply source. And the ones that actually have quite a clear MOA [mechanism of action] , so that we can at least predict what we are really perturbing in vitro. Let us say that this brings us to a hundred molecules, we can create specific sets of combinations from these hundred molecules and then use them to perturb our cultures. Um, what we want to do moving forward is to scale up in terms of automation. And, you know, the more we have robots that allows us to make more and more complex interventions, the more we can explore different regions of this space. For now, we’re limited in that as we are doing this kind of semi manually. But the idea is to potentially browse the space using automation and robots.

[00:47:56] Can you simulate morphogenesis with *only* small molecules?

Abhi: My impression is that while morphogenesis is going on, there’s a lot more going on than just small molecules alone. There’s electrical fields, there’s mechanical forces. Like right now, are you just thinking “well, small molecules get us 80% of the way there, we’ll deal with the other 20% later”? Or what are your thoughts on the subject?

Fabio: Yeah, so it is exactly as you’re saying. Um, and we were discussing this earlier also when we were talking about organoids... what really pushes tissues across the line in terms of functionality and maturity is something that goes beyond chemical perturbations—with this being either mechanical stimuli or electrical stimuli. And this actually has been proven where, for example, to get mature cardiomyocyte fibers, you need basically periodic electrical or mechanical stimuli that can basically bring your tissues to function.

We are fully aware of that, and this is something we are taking into account for our next generation of experimental setup where we will try to integrate chemical perturbations and mechanical and electrical ones. What we can do for now is to push as much as we can with small molecules and also be smart about the way we design our extracellular matrix. So one of the cool things with these kind of gels or plastics is that you can play with their chemical structures or you can kind of embed them with different molecules so that their chemical or physical features change. And by putting together different types of molecules with different type of ECMs, we’re actually able to find proxies for most of the knobs that one might want to tune during... while growing a tissue graft.

[00:49:49] How large is the set of possible tissue scaffolds?

Abhi: I can vaguely understand like, oh, there are this universe of small molecules that happen in developmental biology. Let’s just recreate those for ours. For ECMs, how much do you need to stick to the world of natural things versus explore that into novel chemical territory?

Fabio: Yeah. There’s a trade off and a fine line to thread there in the sense that natural ECMs are better used to culture cells in, so it is easier to culture and grow cells in these natural ECMs. They recognize, you know, again, the familiar kind of microenvironment and they will be happy basically, and grow. The kind of other side of the metal is that you even do not have the ability to really fine tune the features you might want to get as you might have for example, with biopolymers or any type of, again, plastic that can be used. Therefore, for us it has been easier for now to use natural derived extracellular matrices. It might be that at one point in the future we start playing with biomaterials, which I find as a very, very interesting venue or direction of research and application. But for us, natural extracellular matrices have worked quite well. We can now control and engineer them quite well.

Abhi: How large is like the universe of natural ECMs? Are there like a flat dozen or like... hundreds?

Fabio: Yeah, so there are kind of macro categories or buckets, but then there is a plethora of modifications you can apply. So you can take different types of collagens, you can take different types of any type of extracellular matrix that you know is present in other tissues. And then you can start tuning it in terms of how cross-linked it is. And this will determine the physical qualities. You can start again, embedding different types of molecules. You can have different types of porosity of the matrices. And all of this moves on different continuums, right? So you can potentially tune it for as long as you want, and to the resolution you want.

[00:52:32] How reliable are developmental atlases?

Abhi: With regards to like... you’re treating the developmental cell atlases as almost like a ground truth of the real system. I’m curious as to how trustable are those? How heterogeneous are they amongst different embryos? Like do you have one golden standard data set that you can derive everything from, or do you need to average across like hundreds of these?

Fabio: Yeah. So, just a few words on how these data sets come to be. So for the most studied tissues, such as the brain or the cortex, really, there is quite a lot of public available data out there in terms of single cell developmental data sets. So what is possible to do is to compile all of these data sets together and create what is called a gigantic or very large atlas where different samples come from different studies, and of course you will have multiple individuals from every study, right? So you have a multi-donor, multi study starting atlas that gives you some confidence that you can actually—that you’re actually capturing enough heterogeneity and cell types for whatever your goal is. And of course this is also function of the actual number of cells in the dataset, due to single cell technology themselves. But also, you know, due to basically random sampling, you will have an overrepresentation of some cell types versus others. So the hope is that by accumulating again, different data sets and different donors, you’ll be able to have a good representation of the cell population of the tissue of interest.

Unfortunately, this is not the case for every tissue. And incredibly, there are some tissues that are quite important in terms of human disease, such as the kidney, for which there is not much data out there. Other tissues such as the heart... we’re somewhere in between the cortex and the kidney in terms of representation. So a lot right now comes in identifying which data sets we can use and how to integrate them properly in our kind of technology.

Matthew: I’ll just add that as a company, we’ve signed a partnership agreement with, I think it’s one of only two places that you can get developmental tissue as a commercial entity. It’s extremely hard to get it. And we essentially have tissue that arrived maybe a week ago. It was an extraordinary freight process. Um, so there’s the possibility that for some of these tissue types we’d want to tackle like the kidney—because CKD by itself would be a mega blockbuster product as a tissue construct—we may want to create our own analysis.

Abhi: Is the usual process for creating these developmental atlases... like you get an embryo at like day 30 of development and then you sequence stuff from it? Or do you sequence continuously while the embryo is developing?

Fabio: No, so these assays are destructive. You get one sample per developing tissue. I mean, you can get couple of samples if the tissue is big enough, but once it is sequenced, it is over. Um, so when you are really looking at different time points, every time point is typically actually from a different donor. Also because, you know, of course developmental tissues come under very, very tight regulation and control. So it’s not very easy to source them. And then there are also like quite brutal tissues to handle. So there are specific sequencing protocols one needs to apply. And there are consortia out there that have optimized the whole process. Uh, and yeah, the idea is you get them, you isolate the cells, you in some way, you sequence them, and then you have these kind of gigantic tables that you have somehow to make sense out of.

[00:56:45] What is the machine learning model actually optimizing for?

Abhi: And so like I... we haven’t actually... like we’ve just discussed a lot on the data collection problem here, where you have like... you pick from many ligands through the days of trying to recapitulate the developmental process. Is the machine learning task you’re trying to solve selecting like the minimum number of ligands you need to reconstruct the native tissue? Is that largely the primary problem?

Fabio: Actually it is quite the opposite of the limit in the sense that what we might want to do if we had infinite manpower and infinite automation would be to try every possible combination of ligands. What we must use the developmental reference for is to go from all the potential ligands in the human genome to the set of ligands that matters for that cell type in that tissue. Right? So there’s already like a funnel there happening. And then from there we don’t really care of understanding what each ligand does, but it is rather: can we try them in our cultures and see how they perform? And this, in my opinion, is quite important because what we want to do is to extend our protocols to different cell lines, for example, right? And different iPSCs. And that has been proven in different settings—we respond differently to the same ligands. So the idea is: can you build enough redundancy in your set of interventions to take into account, for example, donor variability or interpersonal differences?

Abhi: So is it fair... actually, if a model is trying to help you decide which ligands I should be... what ligands should I introduce at this time point for this specific tissue to lead to this final outcome... What is this model actually trained on in terms of the labels? Like what is the final readout of this whole tissue creation process?

Fabio: Yeah. So there are two things that have to be clear here. On the one end we have the reference developmental atlas that is only used to create this list of perturbations and maybe the order. And after that we move it to the lab, right? We move to generating our own data from these 3D structures and boxes that we were discussing earlier. And this is where the actual optimization and model training happens. So what we do is we have ways to non-destructively monitor what happens in our cultures. Right now we are looking at microscopy mainly, but the idea is to extend this to other modalities to get more and more complex readouts.

We semi-continuously basically take pictures and videos of our cultures, and then we use these images and this data to build basically a digital twin of our experiments. We look at embeddings we see how different cultures and cells are growing. We see how these differences are linked to different ligands. And then we define one direction that matters for us. So one basically cost function that we want to minimize to get to the tissue of interest.

How do we do that? I think this is one of the most innovative approaches in our platform. In order to identify what we want to maximize or minimize, we have to first identify what feature of natural tissue we want to recapitulate. Once we do that, let us take for example, the heart. As we have been discussing the heart for quite a few times today. We take the heart and we want to try to recapitulate fiber orientation and alignment. We quantify what these measures look like in a mature, healthy human heart. That becomes our quantity we want to get as close as possible to. And then we optimize with an active learning setup what happens in our culture so that whatever chip or whatever combination of interventions brings us closer and closer to this quantity of interest.

Abhi: Is the cost function or the loss or whatever the model is optimizing for... is it usually right now, as of today, like a single metric of interest? And in the future you’ll extend to multiple things, but for now it’s just a single metric?

Fabio: Uh, so it is a single... yeah. A single structural feature.

Abhi: Kind of relatedly, we... this is something I guess like the conversation didn’t naturally lead to, but I think it was fascinating enough that I wanna divert back around to it: in the limit case, you can imagine that whatever Polyphron comes up with will create the natural end result that developmental biology does, but go about it in a way that’s potentially more compressed and cheaper than it is in the real world...

Matthew: We hope.

Abhi: Are there existing proof points that this is indeed possible? Like you can have a model that instead of these tens of thousands of ligands, it compresses down to like 50 that do most of the work?

Matthew: I mean, iPSC differentiation protocols, this probably is like an existence proof. Transdifferentiation... these are like non-developmental pathways that get you to a cardiomyocyte that on qPCR and sequencing looks like a cardiomyocyte. So that’s a pretty strong existence proof in our view.

Abhi: And with regards to the experimental loop, how long does each experimental loop take? And is it the sort of thing where you need to like... like each one costs a million dollars, you need to really think about it each time before you go in? Or you can kind of just throw it and see what comes out?

Fabio: Yeah, so, um, what we’ve been trying to do so far—I think I should preface this—is to again, recapitulate one specific feature of structure, which is polarity. Okay. So what we set out to do to de-risk our platform was to say, okay, can we control polarity across different cell types? Once we have identified polarity, we can then say... we can first decide which tissues to try. And then we can basically define the set of ligands and the set of differentiation protocols that allows us to get to this, to basically to try to control this cell type for the feature of interest.

This is where the time comes in. Depending on which tissue we’re looking at and what polarity looks like for that tissue, the time for one experimental loop, one experimental round might change. Um, what we’re actually seeing is that it is pretty fast. Okay. So, like we are running two programs right now. Uh, we’ll be publishing about them, but one is in the cortex, one is in the Heart with cardiomyocytes. And what we’re seeing is that for both of them, we can run hundreds of experiments really in the span of one week. And within this week we will basically try the first pass of our developmental inspired interventions.

[01:04:04] Polyphron’s first big tissue engineering result: polarity

Abhi: Is that... I think this actually lends naturally well to the next question of what is the first interesting slash promising result that Polyphron is willing to share? And it sounds like it is along the lines of this polarity thing.

Matthew: Yeah. So I mean, to put some more numbers on it. So we started with the cortex as our first program. We have a cortical program, a cardiac program, and then potentially have other couple programs, which we’re not kind of being public with right now. But we started with the cortex as the sort of proof of concept in part ‘cause of data availability. It is the most mapped tissue type, like really, really beautiful deep developmental atlases. So if you wanted to kind of prove that that could be your sole prior, it’s a good place to start.

And what we did was we identified a key feature of native morphology in the cortex, both developing an adult, which is neurite orientation. So there are these things called neurites, exciting neurons, that in the cortex have a polarity, like an angle relative to the apical surface of the developing cortex or the adult cortex, which is about 90 degrees. So it looks like it’s kind of beautiful row of neurons. Now it’s a specific subtype of neuron. It’s not all neurons in that kind of section of tissue. So if you were ever going to try and recapitulate cortical tissue, for example—which we actually don’t think is a good initial therapeutic, which we can discuss later, but it’s a very, very good proof of concept for the platform—if you ever wanted to produce cortical tissue, you need to be able to have those neurites have be 90 degrees to the—give or take five degrees—to the apical surface. And just those neuronal subtypes.

And so what we did is we basically created a developmental atlas out of all the available public data. And it was a 10 week period. The total experiment took 10 weeks. It was three loops. We went from a starting orientation, which is measured in an angle. So the, you know, in vivo is 90 degrees. An organoid—going back to our favorite approach—gets you about 45 degrees on average. So the neurites are random, but it kind of averages out to 45 degrees. And we took it from 45 degrees to 82.2 degrees, which is damn near close to in vivo morphology in a three iteration loop that took 10 weeks. Um, we’re extending that right now to the heart. That experiment is ongoing. It looks like we will have sped up experimentally, which is good. Like one of the things that we’re trying to do here is to make it easier and cheaper to onboard each incremental tissue.

Abhi: Is polarity a pretty important phenomenon in a lot of tissues beyond like... it seems like the brain and the heart? Is it important in a lot of other tissues besides that?

Matthew: I mean, it’s important in all tissues, but it’s definitely like the polarity of specific cell types is super important to any tissue that has like an outside and an inside, for example, any tissue that is conducting a signal, be it electrical or otherwise. It’s kind of most of them. And it’s also one of the first macro features of tissue-ness that emerges during a kind of a classic developmental pathway. It’s like one of the first things that’s laid down in development is figuring out... like, development, you’re just this one long tube and you need to figure out which way is up and which way is down. Like, that’s one of the first things that is done. So it made sense to start there for a bunch of reasons.

Abhi: And how long did it take? So like you had scaffold, you seeded with these neuronal subtypes. Eventually polarity emerged after experimental iteration. What was that experimental iteration process like in terms of time? In terms of I guess cycles? Or is there like a single foundation model at Polyphron that decides all ligands, or is it all like you have a new model for each new experimental loop?

Fabio: Yeah. So... so. Keep in mind this was our first program, so there was no model before this one. But the whole idea is, we train our first—let’s call it V zero model—after the end of the first iteration of the neuronal program. And this was a model based on imaging data that was supposed in a self supervised way to learn features across all of our experiments. We then use this model to basically dictate what the next set of interventions might be to optimize for polarity. This then led us to round number two. And then the same happened between round two and round number three, which was our final round. And that’s where we got to. So we started from 45 median round one to 82 max in iteration three. This all happened in 10 weeks. Of course, we are basically still working a relatively low data regime, so we’re not using most cutting edge type of architectures approaches. But one of the very cool things about the way the active learning field is moving is that these are relatively non data hungry approaches. So they’re really effective even if they do not see very, very vast amount of data.

Matthew: I just wanna add something about why starting with polarity is both... so I think we’ve covered why it’s kind of useful from a technical perspective. I think we always have our eye on the clinic. And so something else that we considered as well is: are there tissue types and cell types within that tissue where solving polarity gives you a huge clinical unlock that was otherwise not available? And so that’s why we have a cardiac program. Because in myocardium, a couple things are interesting about myocardium. One is that you have to have like the right alignment. And if you have incorrect alignment, you have arrhythmias. Also the contractile tissue has this kind of helical arrangement, which is kinda interesting. So most of the approaches in cell replacement for heart failure and also some of the engineered muscle patches have not successfully solved this alignment problem. And we believe that... we actually will have a number of advantages relative to those approaches beyond solving alignment. But we believe that if we can solve the alignment problem we’ll have a much, much better safety profile. So even though it’s like the first element of tissueness, just solving that gets you something that is potentially clinically transformative.

Abhi: Sorry, I may be bit confused. Alignment is equivalent to polarity here?

Matthew: Uh, in this case, alignment is a sub feature of polarity. Polarity is like a broad category of directionality.

Abhi: Is polarity the sort of thing where getting like near native polarity after three experimental cycles just feels like crazy given how large the initial search base is? Is it that you suspect polarity is like a pretty low dimensional thing? Because the way that I’m imagining is like first experimental loop, the model gets like... okay, it seems like the model’s getting three data points in total. That’s a lot of extrapolation.

Matthew: Well, I should point out that this is being done in a high throughput chip.

Abhi: Oh, so this is not like you apply a bunch of perturbations, you get a single readout at the end?

Matthew: Sorry. No, no, no. This is like a relatively high throughput. Right now it’s a microfluidic system. We’re gonna build our own slightly more than microfluidic, like meso-fluidic system. So the model at each round is seeing like per plate, there are what? 40. And we do it in duplicate or triplicate. So you’re not just seeing one per round. It is a significant compression. I think we worked out the total combination set of all the ligands and conditions, et cetera, like cell density, ECM, was like 1.7 million or something. And in total, the model probably saw like 90 different experimental conditions. Less than 99%.

Abhi: So when you refer to three experimental loops, what does the three refer to?

Fabio: Yeah, absolutely. So imagine that taking the mental framework of what we are doing. We start from a developmental atlas. We have this list of molecules that we might care about. And then we define random sets of these interventions at the very beginning. So we start from a hundred different molecules. We want to perturb ourselves with three molecules at a time. And then we define all the potential triplets from this list of a hundred ligands, right? We have our microfluidic high throughput set up that allows to try as many of these triplets in parallel as possible. Taking however many plates it takes. And this gives us many, many data points from which we can learn which interventions are more efficacious and which interventions are less or even deadly for the cells. And that’s what we then use to optimize. And that’s why the active learning setup is very useful, because not only it will tell us which triplets that it has seen are interesting, but also it will predict which unseen triplets might be very cool to try. There of course is this play between exploitation and exploration. But you know, all in all, what we see is that we can get enough triplets to go to round two. We already see an improvement in round two, and then we can have further improvement going to the one additional round, which is round three.

Abhi: Or for the case of polarity, is that like a single step perturbation in that like you’re not doing like one set of perturbations and then tomorrow you’re doing another set of perturbations?

Fabio: So right now it is, we’re looking at one time point. And then this time point, it’s one set of perturbations that lasts a few days. In the future, you know, the more complex the structure we will have to recapitulate is, the more complex this kind of protocol will be. So we’ll have different interventions at different time points. And we are already playing with this a little, but the idea is for now was, okay, let us see if the active learning setup and the closed loop system can actually work.

[01:15:33] What comes after polarity?

Abhi: And so you mentioned that one of the reasons you opted for polarity is that like alone polarity is like cool. It’s almost like sufficiently MVP to some capacity. What is the second lowest hanging fruit that you would wanna optimize for after polarity?

Fabio: So there are three things that we mentioned that I discussed at the very beginning that we believe are structure. One is polarity, two is multicellularity, and three is basically reaching the size and the shape you want to achieve for the graft to be clinically meaningful. And that involves vascularization. But our next step will for sure be multicellularity. Again, these things will not happen sequentially, right? We will optimize polarity first and multicellularity... they will rather happen altogether, so they will be optimized as one system, but I think that conceptually it makes sense to think at them as three kind of things we need to care about. So yeah, we’ll basically start adding multiple cell types and seeing: can we preserve the polarity structure we defined in our first programs while having multiple cell types that interact with each other?

Abhi: Do you imagine like... at what point will you need to move into the realm of like multiple time points? Is it kind of like unclear?

Fabio: Right now it’s happening. So as soon as you go above polarity and really... for some tissues, we’re already past that. Just for polarity, you need to have multiple, multiple time points.

Matthew: Lots of robot arms.

[01:17:09] Why is vascularization the hardest problem of tissue engineering?

Abhi: No, automation seems like pretty essential for this. It’s a very interesting direction and cool initial result. And you mentioned vascularization as the final thing that needs to be done for any tissue engineering company to eventually take off. And I think while I was researching for this, like, it just seems like everyone is like talking about like vascularization is like a fundamentally unsolved problem in the tissue engineering field. Why is it so hard?

Fabio: Yeah. So let me go once again back to our reference, which is development, human development. Vascularization is an incredible system and the way it arises throughout development is incredible because as you can imagine, every developing organ needs nutrients. So as soon as the first stage of development is over when basically diffusion is enough, organs and tissues need vascularization. So it is incredibly complex. It has very specific phases that it uses to first define kind of the general vascular framework of the body, and then to generate all the capillaries and this incredible dense network. And once again, we have not been able so far to recapitulate this complexity or to trace this complexity properly in the lab. So all the approaches that have been tried so far have been quite simplistic and reductionist. So we were not able to really achieve the vascular complexity needed to feed growing tissue grafts and make them, you know, and bring them to the necessary size and shape.

Interesting... what has been happening in the space is that people have started understanding how to use different approaches to have vessels grow and how to engineer them. Where the key insight has been: we have to have a clear starting point and we have to have a final point towards which the vessels can grow, right? We have basically have some attractor that the cells can use to point at. Our insight is, once again, you cannot use or recreate this complexity artificially or top down. You have to grow it. So our approach is as we are growing different tissue structures by tracing development similarly, we are also trying to grow vessels in these three dimensional boxes full of extracellular matrix of some type. And again, the way this will happen is that we will have computational models that try to drive vascularization and optimize for different features. And this is important because one of the very cool things about vascularization is that it varies across different organs. So the brain with a blood brain barrier needs a very specific type of vascular network. The heart, a different one. Kidney, pancreas, liver, different ones again. So there is a lot of optimization to be done there as well.

Matthew: And we have a vascularization program underway. Like, we know it’s a showstopper. We’re working on it. It’s not an afterthought.

[01:20:33] Why can’t you just wash angiogenesis factors over the tissue?

Abhi: I spiritually get why you guys want to just like mirror developmental biology. ‘cause that’s kinda like the thesis of Polyphron as a company. Why, why, like, naively, why can’t you just like seed the whole thing with endothelial cells at the very beginning and wash over angiogenesis factors to create the vessels? Like why doesn’t the naive solution work?

Fabio: Yeah. So the problem is that cells need to interact in a very specific way in order to create the structure first and then to gain functionality. If the first steps are skipped or not properly recapitulated, you will not be able to obtain a functional network by the end. That’s why the majority of vascularization approaches first implies some type of kind of blob of endothelial cells that need to exist. And then from this blob, you know, the vessels will start to arise and kind of diffuse throughout the growing graft.

Abhi: Is there like any world in which you grow the vessel separately and then you can join them back in?

Fabio: It is extremely difficult. Imagine that for many tissues there is basically one capillary per cell.

Abhi: I was not aware of the complexity. It’s not like flooding a rough neighborhood of cells...

Fabio: It’ll change from tissue to tissue, but the density you need to achieve is astounding. So the chances of being able to do this ex-graft and then plug this in, in a way I think are relatively low. Um, and again, it’s really difficult to reach equilibrium of a complex system by combining things. It’s just easier to have the features arise together with complexity.

[01:22:25] How does the graft integrate with the host’s blood supply?

Abhi: Makes sense. Um, and let’s say like you do solve this like grand challenge of the field. Um, you’re able to get vascularization working. You give it to a surgeon, they’re about to implant it into a patient. Once it’s in the patient, it’s not like integrated with the rest of the vascular system of that patient. How does that integration occur?

Fabio: Yeah, so there is a process that needs to happen and this regularly happens in the surgical rooms during organ transplant or any type of surgery, which is anastomosis—where the vessels in the entering body have to be connected to the existing vessels. One way is to do so is surgically. The other way is to try to exploit the natural way the body reacts to foreign bodies, which is basically by perfusing them with blood or with fluids having their own immune cells and cells kind of try to colonize the graft and try to find anchor points that can be used to make basically, um, integrate this, this addition.

Abhi: By anchor points... is that like a physical vein that they’ll attach?

Fabio: So it’ll be first cells and then it’ll be, you know, it’ll be either like some type of fibrotic tissue. I should preface, I’m kind of speculating here. It hasn’t been, I think, fully proven across grafts. But what I can tell you, for example, in bone replacement, this is exactly what happens. So the bone replacements are designed to have basically to be exposing anchor points that the body sees, recognizes. It latches onto, and this basically helps the body integrate the new graft. Right? So we can imagine something similar happening with our grafts, whether we have to insert these kind of anchor points—this might be like peptides, so nothing too problematic for the body—or we should find ways to stimulate angiogenesis, which of course is kind of tricky due to oncogenic potential. But you know, one way to approach the problem will be how can we exploit the way the body reacts and basically use it as a way in.

Abhi: I didn’t know about that bone thing. That’s interesting. Like, I know like skin doesn’t really need vascularization all that much. ‘cause it’s thin enough that like, diffusion just works fine. I didn’t know bone grafts even had like needed blood flow into it.

Fabio: They do. So imagine that a bone graft is basically again, this rigid sponge. And it’s immediately perfused. And then you have these tiny pores that kind of are exposed and cells will just pass by, latch onto them. There’s a small period of inflammation. And this then will have other host bone cells that colonize the graft and then start basically remodeling it and making it new bone from the host.

Abhi: This is just my own curiosity, but like, in bone grafts, do they also replace the stem cells within it, or is that ignored?

Fabio: No, no. You can just put the graft. This mineral kind of thing.

[01:25:45] How do you validate tissue function before implantation?

Abhi: Interesting. Um, okay. I wanna zoom out a bit, like more the future. And so like right now you don’t necessarily have functional tissue in like an absolute sense, but someday you will. When you get to that point, are there functional assays for this sort of thing that you can use to prove out that like, oh, this chunk of cardiovascular tissue is actually gonna be useful? Once I implant into someone, like, how do you prove that out in advance? Beyond like measuring polarity, the proteins are there, is there anything else?

Matthew: Yeah, I mean, so in vitro you have like cell identity sequencing, qPCR, all of that stuff. But you also have electrophysiology. You have functional readouts where you can measure the function of the tissue. So like contractility in heart tissue, pacing if it’s pacemaker type tissue. So that’s all the stuff that you have in vitro. And then obviously you have animal models. So in the heart you can’t really do functional readouts in rodents because the physiology is so different. Like the heartbeats are like an order of magnitude apart in terms of beats per minute. So it’s a very poor translational model for cardiac interventions, for most cardiac interventions for that reason. So you would probably do it in a pig. So there is a panel of fairly robust functional assays. You can do the in vitro level and then you would do a large animal study in a pig first before you put it in a human.

Abhi: I remember when I interviewed Hunter, the Until Labs guy, my last episode... he said that, “oh, well we’re good with functional assays because the organ transplant field has already like, figured out most of them.” Is that also true for the tissue side where like... I know that there are discrete tissues that are transplanted from like one person to another person. Are those metrics like pretty well flushed out?

Matthew: So, I mean, my understanding is that—and I’m not a complete expert in organ transplantation to the degree that Hunter must be by now—but my understanding is that there are pretty minimal assays that are done on those organs before they’re transplanted. So you’ll probably have like an ischemia time window check. You probably have like a biopsy, maybe you have some imaging. But these are basically being done in a helicopter, so your QC/QA is: “it’s human tissue and human tissue is good, so let’s put it in and this person’s gonna die otherwise.” So that obviously changes your risk profile.

Abhi: He did mention like the big advantage of cryopreservation was like, you get to do more testing right now.

Matthew: And actually it’s an equivalent advantage of being able to do these ex vivo grafts, which is that you get to do QC/QA like really, really deeply.

Abhi: Like as much as you want.

Matthew: And I think that’s gonna be a core advantage going forward.

[01:29:01] How do you design a clinical trial for a biological pacemaker?

Abhi: Makes sense. The other big question I have is like, this is a brand new therapeutic modality, in basically every fashion. And so how do clinical trials work for this sort of thing where you were going up to a patient who—like for the pacemaker case—like they already have a pacemaker? How do you convince them? Like, “oh, can we put this engineer tissue into you to see if you like, can go without the pacemaker?” How do you recruit these patients in the first place?

Matthew: Yeah. So... so we have—before I get to that step, I just wanna kind of touch on how we think about indication selection. ‘Cause I think it’s quite important. So because we are you correctly pointed out, trying to pioneer a completely new therapeutic modality, we have a barbell strategy for every new tissue we approach, which is that we are looking for two potential products. One we call a Regulatory Pathfinder. One is like a Commercial Workhorse.

So in the cardiac case, our regulatory pathfinder is a biological pacemaker. And what you’re looking for for a regulatory pathfinder product is the ability to have a unbelievably unambiguous clinical readout, incredibly fast, with the circumstances that would allow you to have a very, very small clinical trial. And then the commercial engine, which in our case is a left ventricular muscle patch for heart failure with reduced ejection fraction in stage three and stage four... that’s basically 95% of the revenue, but the readouts take a lot longer because the clinical readout for the reimburser for your commercial engine is reduced hospitalizations, which you just have to measure over the course of a year or two years even. We need to find proof of efficacy as quickly as possible in the most unambiguous way we possibly can to give a kind of a halo effect to both products in the barbell strategy.

So in the biological pacemaker, what you are looking for—and back to your kind of the initial challenge that you gave—is you’re actually not looking for patients where they have something that’s working perfectly well. You are deliberately trying to find cases where this can be a salvage therapy. And because you have the commercial engine, you don’t really mind about the size of the patient population that you’re gonna go after there. Because these things work together in tandem. So you are really trying to optimize for: can I have a first, like a phase one trial that is like almost N of 1, potentially even compassionate use so that you can get as much regulatory speed up as possible and have an unbelievably clear binary signal.

So for the biological pacemaker, what we’re looking for is “hardware exhausted” patients. Uh, they’re often pediatric or neonatal, sometimes they’re adult. So these are patients for whom they have a device that is going to fail. It’s either because of infection or its device rejection in some form. They’ve had multiple surgeries. There’s nowhere to put a lead left, like occluded veins. There are a bunch of reasons that this could happen, but basically your quality of life is gonna be anything from very, very bad to palliative care. And you need to find one of those people.

From a patient recruitment perspective, a lot of these cases are clustered in the same three or four surgical centers in the US. So you need to basically build inroads with the cardiologists who are doing these lead extraction, lead placement operations. But it’s essentially the same surgical workflow that they’re already doing. So again, no one needs to be like retrained on anything. Um, but the way that the trial would be designed would be as like a weaning trial. So you would have someone who is on device who has a pacemaker right now, but they are hardware exhausted—that device is gonna fail.

Abhi: And that’s usually predictable.

Matthew: It’s incredibly predictable, right? They will be classified as a hardware exhausted patient. That is known. There’s a code, like you can find them. And you essentially run a parallel trial where you have the engraftment of our biological pacemaker and they have the device still there as a safety backup, which again makes it should make it a lot easier to get buy-in from all of the patient groups and FDA and cardiologists because you have a kind of a baked in safety valve, which is that you already have this device and what you do is you toggle the device on and off and you see how much time can you have off device. And that readout you should find out whether you’ve built a functional tissue that does the thing that we said it was gonna do within days, hours, certainly, certainly weeks. And so it is the fastest possible way that you can get a human clinical readout. And so patient recruitment, you are really looking for like one, maybe two people for that trial.

Abhi: Why are these types of patients congregated at three medical centers? Is just like the rarity of this condition?

Matthew: Yes, because so much of it is pediatric. So it tends to be clustered in like where there are pediatric cardiologists, which is obviously a specialism within a specialism.

Abhi: This is maybe... these are difficult questions to answer because it’s something that’s gonna take a while for you guys to get to, but finding these sort of patients, or like convincing clinicians that they should be using this seems challenging. How much do you convince the cardiologist themselves that this is a worthwhile approach versus like, you just file a clinical trial and if like the parents of the child are interested, then they’ll sign up for it?

Matthew: So, I mean we think there’s a pretty good shot that... obviously we wanna engage early and often with the cardiology community, and we want ‘em to be very, very supportive of this as I think they should be. When it comes to the first use, it could be done through a compassionate use pathway, which has a much, much lower regulatory bar. There are large number of functional assays that we would do before even attempting to put this in a human, including successful large animal trials, obviously. And then the good thing about the way that the pacemaker trial would likely be designed is that there are no other options for that patient. Um, and there’s already a device in place that you could switch on to take over from the graft if something were to go wrong. So you are looking in these regulatory pathfinder indications—which we have for every tissue—we are looking for something with these features where you can try and have safety built in almost by definition. So looking to take someone off a device is a good way for looking at indications.

Abhi: Is there some analog to that in anywhere else? Like “you are on a device right now, we’re gonna introduce an intervention to try and get you off the device”?

Matthew: Is there an analog? I mean, I would guess that in kidney there must be... you know, to get you off dialysis. Time off dialysis is an endpoint. And that would probably be our endpoint for what it’s worth for nephron units and stuff.

[01:37:01] The argument for being a pan-tissue company

Abhi: One of the other... this is something we touched on earlier, but right now you have cortical programs going on right now. You have heart programs going on right now. And like the hope for Polyphron is that you become this pan-tissue company that you are involved in every single organ, every single system in the body at once. And when I first talked to you guys, it seemed like clearly insane. After you explained to me the rationale, it makes a bit more sense. I’d love for you to just like, repeat that basically.

Matthew: I mean there’s a... there’s I suppose a technical dimension to this and there’s like a commercial dimension as well. Um, so from a technical perspective—and Fabio can chime in if I absolutely mangle this—but as we said before, we have a relatively strong hypothesis that there are gonna be some latent rules of morphogenesis and development that extend across tissue. And we think that there’s a critical mass of tissue systems and tissue products that we can onboard, after which we could probably do most other products. So we actually think that there’s a race to reach that critical mass and that the company that does it first likely has a very strong competitive position relative to the market. Helped by the fact that most of the data is actually being driven by the lab. So after you have this initial bootstrap from reference data, it’s really, really hard to come in as a fast follower and attempt to try and produce this stuff. You would have to build the loop and run through the loop the same way that we’ve done. And we think that that will take a lot of guts and capital and maybe wouldn’t even work if someone were to try it afterwards. So the race is on in our view to try and get to that critical mass of tissues.

Not only that, but from like an organizational perspective, you need to master the manufacture, production, deployment of new tissue products at commercially reasonable cost of goods, which is something we’ve built into the loop from day one. And we’ll get better and better at that over time. And then we’ve also built these kind of nested flywheels where we should not only get better as a company and then at the company level, at the model level, each new tissue type we onboard, but each graft, each tissue product that is sent out into the world. So we’ll be very, very careful about collecting super deep telemetry data about what these grafts are doing, engraftment rates, clinical readouts, et cetera, all of which will be incorporated back into our loop.

And then the last thing I’ll say is that from like a company compounding perspective, we believe that these tissues will not suffer the patent cliff problem. So the value is not a composition of matter, which you would normally protect with a patent—which is the kind of the grand bargain of pharma is that you get 10 to 12 years of like monopolistic profits, but you have to disclose exactly how you’re doing it. Process power beats patents basically every time, and ours is a very, very process power driven company. So we anticipate for each product to have quasi-perpetual revenue streams at the limit. Like a bunch of stuff can happen, people can produce better products than us, but I don’t think we will have this kind of drop off. And so we should get quicker and quicker and faster and faster with every new tissue type and every new tissue product we produce. And the company should get more and more defensible as we go.

Abhi: Is that last point along the lines of like, oh, the defensibility of what we produce at Polyphron will be a combination of one, it’s just very hard to produce, and two, how we made it is a trade secret?

Matthew: Basically. So it’s model weights and process. It’ll... it’s closer to a semiconductor fab, honestly.

Abhi: That’s interesting. Um, I do have a friend who actually considers that like this will happen to the biology field in general because... it’s just like, like if China can just like pick up the molecule, then like why would you ever give away the... yeah.

Matthew: I would anticipate actually across the field, like classic pharma and biotech approaches working on regular modulations will start to move more and more into trade secret. Particularly with all of the target crowding that we’re seeing.

[01:41:57] What are the biggest scientific and economic risks?

Abhi: What is the most risky thing that could possibly happen at Polyphron that is either scientific or economic or both?

Fabio: Scientific... I think we made this explicit, but we have quite a few technical problems to solve. Adding complexity in terms of structure. We will have to see how well our platform gets to that level, or which knobs we will have to tune. Of course vascularization, we have this very strong hypothesis, but I can’t lie, it is something that we’ll have to face and that we are already starting to think about. The more on theoretical side, I find this idea of learning these latent rules of development as quite an interesting challenge. It’s not really a roadblock, but I think it’s something very fascinating that we will try to prove.

Matthew: Economic... Um, so I think I’m gonna kind of give like a non-answer. Here’s my non-answer. So when investors invest in biotech or tech bio companies, I think there is a delusion... which I don’t know who it serves... but there is a belief that you are only taking technical risk and you’re not taking market risk. And that’s what kind of deep tech investment is. That’s what biotech investment is. That’s absolutely not true. You’re taking both. And if you are not appreciating the commercial risk you’re taking, then... well, it’s not ideal.

So the worst position to be in as a company that has pulled off technical miracles—and we’ve pulled off maybe like one or two of the technical miracles we need to, but there are plenty more we need to pull off to make this successful—the worst thing that could happen is that we end up in a position where we don’t know how to manufacture this profitably. We don’t know how you would sequence the regulatory strategy, the market rollout strategy. Like Bluebird Bio is the now, I think, canonical example of what happens where you can do unbelievably good science and build like an unbelievably impactful product that can meaningfully benefit lots and lots and lots of patients, and you get sold in a fire sale for like 5% of the money you raised. So as much as possible, not just in how we build the organization, how we think about things strategically, but actually in what we are trying to optimize in the lab, we are baking in, as I said before, like COGS, vendor redundancies, like where we are sourcing this stuff from, to ensure as much as possible that we’re building a technically viable and commercially viable organization at the same time. So I suppose the risk is I’m wrong about any of that. But we’re trying as much as possible to kind of think around the corner even well in advance of when it would actually be going up for reimbursement by a payer.

[01:45:23] Who are Polyphron’s competitors?

Abhi: Relatedly, who do you need to worry about in terms of competitors in the tissue engineering space? I can’t really think of anyone. You the only...

Matthew: So... I mean, there are people working on tissue. I think we are the only ones with this pan-tissue approach. I think it’s because it probably does seem insane to people initially and you have to think about it a little bit to get more comfortable with it. But there are tissue specific companies that are working on things. It’s difficult to find out exactly what their approach is from the outside sometimes. But Bob Langer has a company called Satellite Bio that, as far as I know is doing liver. Aspect Biosystems, I think is taking a bioprinting approach. They’re doing some endocrine stuff, liver as well. And then they did a Novo Nordisk deal, which I think is obviously gonna be obesity related. And then you have, I suppose at the other end you’ve got the Xeno companies as well. Again, like I don’t believe that if you could produce human tissue or you could have pig tissue, you’d pick human tissue every time. And I actually believe that the pace of the field in sort of human tissue engineering is gonna exceed xenotransplantation pretty quickly in terms of technical maturity. And so I think that you’ll always want human tissue. If you had like a head to head, you’re always gonna want human tissue for a bunch of reasons. Not least of which is the genetic engineering, which adds some safety concerns and may not even be transferrable patient to patient.

[01:47:07] Expanding the TAM beyond transplant lists

Abhi: You mentioned earlier on about how by going after this non-super advanced stage of patients where they need organ transplants, you get access to this patient population that there is not really any therapeutic intervention available to them other than perhaps like medication. I’d love to get your take on that because I think it does dramatically change how I think about the economics of Polyphron.

Matthew: Yeah. So I think that’s something that you probably should believe in order to be bullish on like the extremely successful case of Polyphron. So let’s take an example. Let’s take the heart, which I know we’ve been talking about a lot, but we know a fair amount about it. So there are 6.7 million Americans with heart failure. It is the second biggest cause of death in the US I think. And of those 6.7, maybe 10,000 roughly will get a heart transplant. There are a large number of people who just cannot get the heart transplant even if they kind of were on the list. So either for frailty, age... like you’re just not gonna give it to someone who’s over a certain age. Comorbidities, lifestyle. You know, there are potentially other kind of exclusionary factors that would prevent you—like say you can’t take immunosuppression, for example. You can’t deal with being in the ICU because the heart transplant involves this like full sternum being cut open.

So of those 6.7 million, only 10,000 will make it to a heart transplant. But you have 670,000 roughly—so about 10% of those—are in stage three or stage four of the categorization that is used to categorize heart failure. And of those 670,000, we think that automatically with our first product, you can start to kind of address most of that population. So these are people who would not be able to get a heart transplant, who would be kind of allowing to defer that potentially forever. So you can meaningfully expand the population by, I mean, 10 to 20x roughly.

And it remains to be seen how much you can push it, but to kind of give you a sense of like TAM expansion: Okay, so let’s imagine that you start with a patient population who have heart failure with reduced ejection fraction below 35%. So there will be like a kind of a cutoff limit. And these are people in stage three or stage four. Your initial patient population is gonna be about 40,000 people. And the pricing for that product, you probably would anchor against a left ventricular assisted device. So call it 150 to 200,000 dollars, let’s say 200,000 dollars. Again, you’re probably anchoring against deferring a heart transplant and not having a left ventricular assisted device. So you have... let’s take the midpoint between 20,000, 40,000, which is number of these patients. You have a TAM just there of like 6 billion dollars in the US alone. That’s probably 10 globally. But you can meaningfully expand that just by going up. You can potentially even get to 50 plus billion dollars of US TAM by being able to address like a fraction of the people who have like progressive chronic heart failure. So right there, just in like expanding the patient population window slightly to say, you know, a hundred, two hundred thousand people, you begin to be in a revenue range of like Eli Lilly. And we believe that for most of these tissues, right of the list that I gave of like 10 different tissues we could go after, we believe that there are these mega blockbuster products waiting if you can just slightly—not even fully—just slightly increase the patient population window. And at the limit we think that we can increase it significantly.

Abhi: Is this particularly true in heart or do you imagine like similar dynamics would occur in almost every organ? Maybe except the brain?

Matthew: Almost every organ except the brain. Maybe eventually the brain, but that requires a bunch of technical work. There’s no brain transplant, right? So you take a lot of reimbursement risk and then there’s a huge amount of translational risk that you probably don’t wanna layer on top of the other risk that you’re taking here. But yeah, conceivably for pretty much every other organ that we currently have transplantation for.

[01:52:28] Autologous vs. Allogeneic approaches

Abhi: In the limit case of like Polyphron is now outwardly open to patients in need, would you go autologous or allogeneic for whatever your tissue construct is and why?

Matthew: So I think it depends. I think it depends on the indication. I think that there are manufacturing considerations obviously around autologous. So the optimal COGS scenario is probably immune-cloaked allogeneic, so that you can amortize the cost across multiple units. Because our system is designed from the get go to be resistant to starting donor line heterogeneity, you could actually just as easily imagine doing HLA matched allogeneic. So there are 40,000 roughly different subtypes, but that’s not normally distributed. There’s a kind of a fat head of subtypes that people belong to. Um, you could have a kind of an inventory of like a hundred different cardiac products and cover 70, 80% of the US population. It’s lower in more homogenous, genetically homogenous countries. So like Japan, I think you can do like 20 and you can cover 80% of the population. So I think that really, if you’re gonna be able to produce this at scale, you are better off going with a kind of immune matched or immune cloaked allogeneic product.

Our system is designed to be able to produce both hypothetically. Now this is like a bit more of like a sci-fi case, but if Hunter is successful, you could imagine biobanking frozen tissue well ahead of time, right? We know that there’s cluster dysfunction. That’s how age-related chronic diseases work. We don’t necessarily know the order, but we know that your kidneys are gonna fail, your lungs gonna fail, your heart’s gonna fail. And you could imagine a future where you would create a bank of constructs that you think you might need, and those are kept on ice for you. But I think for near term commercial solutions, you would want to do immune compatible, allogeneic.

[01:55:07] Is a 3-year timeline to the clinic realistic?

Abhi: The future of Polyphron in terms of like... when will the first clinical trial start? To me, when I first talked to you, I thought like, okay, it’s gonna like six to 10 years away. You had a rather aggressive timeline: three years. What’s the rationale on that?

Matthew: So now just to be clear about what the claim is. I am not claiming that within three years we could produce every single tissue. I’m also not claiming that any tissue that any of all the possible tissues could get into to a human within three years. I am claiming that if you are very specific about indication and tissue product selection and you parallel track some stuff—you are properly financed to allow you to parallel track some stuff—then yes, you could put the Regulatory Pathfinder product in a human within three years. So I think you would do, you know, large animal takes you some amount of that, and then you would jump straight to a large animal, and then you would go into human from there. And I think you could do it in less than three years if you had a compassionate use.

[01:56:28] Cross-species translation

Abhi: This isn’t something we actually talked about, like how how big of the translational risk is there that you might optimize something really good for large animal, but might not be very good for human?

Fabio: So our whole computational setup is designed to optimize for human structure. We are using human cell lines. So there’s always some type of translational gap. There is the limits of predictive validity of animal models. But you know, in terms of what we’re trying to recapitulate, we’re looking at like human tissue.

Abhi: But like in that case, that means it’s very difficult to do this like large animal study because you don’t know how to create tissue to their species. Is that not fair?

Matthew: I mean, you’ve got a decent similarity. Enough of a similarity between like in the cardiac case between pigs and humans. So the heart is similar enough in terms of size. Alignment is not perfect for what it’s worth, which is another issue with like xenotransplantation is that there is like a slightly non-natural morphology that you’re having to deal with. But that I believe would be considered kind of the best in class model. Porcine... maybe, but probably pig.

[01:58:05] What would you do with $100M equity free?

Abhi: The last question I have is: if someone gave you a hundred million dollars equity free tomorrow, where would that be best allocated to make the company move faster?

Matthew: Doing that thing I said in three years? No, in all seriousness, I think that a hundred million dollars would be best spent on... because we believe that this company is truly a platform and that there are some very particular moats that will emerge at a certain scale... sizing up the automation fairly aggressively early on. Uh, it can be sequenced. It doesn’t all need to be done at once, but a lot of the investment would go there. A lot would go on compute, honestly. And then the remainder would be the steps that you cannot skip and you should not skip, around like CMC and QC/QA and like all of the regulatory stuff to be able to put these products in humans. But I mean, with a hundred million dollars equity free, we would imagine that we could try and have two products in human, one within IND underway. And of those two products, they’d be in two different tissue systems. So we wanna prove that this... it’s not something weird about the heart that allows this to be the case. That this transfers across germ layers. Because it transfers across tissue systems, it transfers across germ layers. So that would be my use of funds.

Abhi: I’m a little bit curious... I felt for like your take of like on one side, I can imagine this a hundred million dollars is like really well spent in gathering like a lot more embryo data. But is that actually like not super useful? Like you have plenty like of set line data that you need?

Fabio: Yeah. So. We do not have it right now for sure. So there is gaps both in terms of stages for one specific tissue as much as we do not have pretty much anything for other tissues. But I do agree. My sense would be that after a certain size of the data set, the amount of information you can extract kind of plateaus. Just you need the feedback loop instead. So what we really have to do is to scale up as much as possible the scale of our experiments. So having automation that allows us to create very, very complex intervention kind of groups. And secondly, it is how many readouts we can get out of our experiments. What we’re doing right now, it’s mainly imaging based and it has of course its limitations. There are solutions out there that allow you to do spectroscopy or collect ‘ome or any other type of sensor based data. And you know, by building this kind of 360 view of our experiments, we can really start exploiting our computational approach and then really try to optimize for the structures we want to get.

Abhi: Okay. Cool. I think those are all the questions I had. Thank you so much for coming on, Matthew and Fabio.

Matthew: Thanks for having us.

Fabio: It was a pleasure. Very fun. Thank you.

Rush is augmenting drug discovery for all scientists with machine-driven superintelligence.

LatchBio is building agentic scientific tooling that can analyze a wide range of scientific data, with an early focus on spatial biology. Clip on them in the episode.

If you’re at all interested in sponsoring future episodes, reach out!

1. Introduction

2. Timestamps

3. Transcript

Introduction

This is an interview with Yunha Hwang, an assistant professor at MIT (and co-founder of the non-profit Tatta Bio). She is working on building and applying genomic language models to help annotate the function of the (mostly unknown) universe of microbial genomes.

There are two reasons you should watch this episode.

One, Yunha is working on an absurdly difficult and interesting problem: microbial genome function annotation. Even for E. coli, one of the most studied organisms on Earth, we don’t know what half to two-thirds of its genes actually do. For a random microbe from soil, that number jumps to 80-90%. Her lab is one of the leading groups working to apply deep learning to solving the problem, and last year, released a paper that increasingly feels foundational within it (with prior Owl Posting podcast guest Sergey Ovchinnikov an author on it!). We talk about that paper, its implications, and where the future of machine learning in metagenomics may go.

And two, I was especially excited to film this so I could help bring some light to a platform that she and her team at Tatta Bio has developed: SeqHub. There’s been a lot of discussion online about AI co-scientists in the biology space, but I have increasingly felt a vague suspicion that people are trying to be too broad with them. It feels like the value of these tools are not with general scientific reasoning, but rather from deep integration with how a specific domain of research engages with their open problems. SeqHub feels like one of the few systems that mirrors this viewpoint, and while it isn’t something I can personally use—since its use-case is primarily in annotating and sharing microbial genomes, neither of which I work on!—I would still love for it to succeed. If you’re in the metagenomics space, you should try it out at seqhub.org!

Youtube:

Spotify:

Apple Podcast:

Transcript: https://www.owlposting.com/p/we-dont-know-what-most-microbial

Timestamps

00:02:07 – Introduction

00:02:23 – Why do microbial genomes matter

00:04:07 – Deep learning acceptance in metagenomics

00:05:25 – The case for genomic “context” over sequence matching

00:06:43 – OMG: the only ML-ready metagenomic dataset

00:09:27 – gLM2: A multimodal genomic language model

00:11:06 – What do you do with the output of genomic language models?

00:17:41 – How will OMG evolve?

00:20:26 – Why train on only microbial genomes, as opposed to all genomes?

00:22:58 – Do we need more sequences or more annotations?

00:23:54 – Is there a conserved microbial genome ‘language’?

00:28:11 – What non-obvious things can this genomic language model tell you?

00:33:08 – Semantic deduplication and evaluation

00:37:33 – How does benchmarking work for these types of models?

00:41:31 – Gaia: A genomic search engine

00:44:18 – Even ‘well-studied’ genomes are mostly unannotated

00:50:51 – Using agents on Gaia

00:54:53 – Will genomic language models reshape the tree of life?

00:59:18 – Current limitations of genomic language models

01:08:54 – Directed evolution as training data

01:12:35 – What is Tatta Bio?

01:19:02 – Building Google for genomic sequences (SeqHub)

01:25:46 – How to create communities around scientific OSS

01:29:06 – What’s the purpose in the centralization of the software?

01:35:37 – How will the way science is done change in 10 years?

Transcript

[00:02:07] Introduction

Abhi: Today I’m gonna be talking to Yunha Hwang, an assistant professor at MIT, applying machine learning to microbial genomes. She’s also the co-founder and chief scientist at Tatta Bio, a scientific nonprofit dedicated to building tools for genomic AI. Welcome to the show, Yunha.

Yunha: Thank you. Thank you for having me here.

[00:02:23] Why do microbial genomes matter

Abhi: First question, what makes microbial genomes so interesting to you?

Yunha: So yeah, I get this question a lot. If we think about the history of life, microbes have dominated that history of life, which means it’s the most diverse, it’s the most flexible, and in terms of the chemistry that it can do, it’s the most divergent you can possibly imagine. When we think about diversity of sequences, that’s where you’re gonna find most of the diversity of sequences, in microbial genomes. Yeah.

Abhi: And so it feels like a natural place to take AI and ML tools to just throw at it.

Yunha: Yeah. That’s one way to look at it. I think when we think about using biology to do cool things, I think about doing cool chemistry. So there’s like a utility aspect there as well.

Abhi: Were you focused on this topic since your undergrad days, or was it something you switched to during your PhD?

Yunha: Yeah, so I was a computer science student in undergrad, and I learned about the human genome. So I was interested in biology, but I got really hooked when I learned about this field of environmental microbiology, which sounds really niche, but it’s essentially... you’re looking at life in very extreme environments or places that you wouldn’t really typically look for life, such as the deep sea or deserts and so on. And then you’re finding new types of life, all through sequencing, through different kinds of methods. And that’s when I really got hooked in terms of scientific interest.

[00:04:07] Deep learning acceptance in metagenomics

Abhi: I think an interesting trend in a lot of people applying AI to at least somewhat niche fields in biology is that they are usually amongst one of the first people to stand up and say, “Hey, deep learning could be really useful here.” And the culture around that field is usually not pretty accepting of deep learning. How much did you find that when you were applying AI to metagenomics?

Yunha: Yeah, that’s a good question. I think.. . I think at the beginning, people were a little skeptical. But I think people were also quite open to it because when you’re studying metagenomics, you basically scoop up dirt and then you sequence everything out of it and you use computation to piece them together. So essentially you’re looking at billions of base pairs, and there’s no way a human can do it. There are people who are really good at it and who can piece together entire genomes using manual curation who are just pattern recognition geniuses. But for the most part, we’ve been using computation to study these billions of base pairs of divergent data. So in that sense, people are not so opposed to the idea that, “Wow, maybe we’re not very good at doing this. Maybe we do need machines. We do need some extra layer of understanding in order to understand this massive amount of divergent data.”

[00:05:25] The case for genomic “context” over sequence matching

Abhi: Traditionally—I’m not super familiar with the field—but my interpretation is that the traditional bioinformatic tools for studying metagenomics are like... you’re literally matching nucleotides between sequences that you found in one pile of dirt to another pile of dirt. What is your pitch for a better way to do it?

Yunha: Yeah, that’s a great question. So sequence matching is definitely part of the workflow. I think what’s really interesting is when you can look at a sequence and also consider the context it’s found in, and then understand that sequence within that context. And then also do basically comparative work between that sequence found in different contexts, and how the differences in the sequence can be made sense of using that information. So if you just take out the sequences, then these are just two sequences that are a few mutations apart. But then if you consider the full context of either the sample or the genomic context or the taxonomic context, then you’re actually answering a much more biologically relevant question.

Abhi: So you’re adding multiple layers of information on top of the raw sequences alone? And seeing what else you can pattern match from that?

Yunha: Exactly. Yeah.

[00:06:43] OMG: the only ML-ready metagenomic dataset

Abhi: And I think that leads well to perhaps probably your first big paper in the space. Maybe there’s others. But I think the first one that I was made aware of is a paper that introduces two things. One is a really large metagenomic data set called OMG. The second, included in the same paper, is gLM2, a genomic language model. I’ll separate my questions for both of those. The first one is OMG. Why did you release another metagenomic data set? Because from my outside view, right, there’s already a few out there. Why was there a need for another one?

Yunha: Yeah, that’s a great question. I would argue there were none out there—none that was useful for machine learning. Yeah, so there are public data sets. That doesn’t mean they’re useful. That they can be used immediately for machine learning purposes, for language modeling, for instance. An example is, metagenomic sequences can be very poor in quality, so you do need to do a lot of quality filtering. Also, there’s a distribution effect where you have a lot of really short sequences. ‘Cause as I said, you’re doing shotgun sequencing and piecing them together. So the curve, if you look at the distribution, it’s just like this. So you get a lot of really short sequences that don’t even contain a single gene, so you have to throw them out. If you modeled using that, then you’ll be basically modeling nothing. So there is some sort of filtering that you need to do with quality control.

There’s also two major big public databases. One is JGI’s IMG database. And the other is EMBL’s [European Molecular Biology Laboratory] MGnify database. And there is overlap between the two, and also a lot of biases. So for instance, people like to... it’s much easier to sample human feces, compared to deep sea ocean, even though that has a lot more diversity. So you get hundreds of samples of the same sort of human gut sample, but then very few of the very diverse deep sea sample. So by putting them together and then doing dereplication and semantic deduplication and various sort of methods in order to de-bias the data set, we’re making it actually a resource that’s useful for machine learning as opposed to its raw state, which was not really useful.

Abhi: So OMG was for the most part a combination of the existing data sets with a huge amount of pre-processing on top.

Yunha: Yeah.

[00:09:27] gLM2: A multimodal genomic language model

Abhi: I think that dovetails well, and you mentioned semantic deduplication. I’ll have questions about that later. But first, maybe we can start with... you created this data set, you built a model on top of this data set called gLM2. What is gLM2?

Yunha: gLM2 is a genomic language model, but it’s actually not a DNA language model. So, it’s trained on metagenomic data. It’s a multimodal model in that all the DNA sequences or all the intergenic regions are encoded in DNA nucleotides, and the coding sequences are encoded in amino acids. There was a reason why we did that. We actually wanted to make sure that we can model amino acid interactions across protein sequence boundaries. So if it’s a protein language model, it is not gonna learn protein-protein interactions, because you’re not seeing multiple proteins in the same context. Whereas a genomic language model that contains multi-protein context, you’re actually able to model multi-protein interactions or intergenic region-to-multi-protein interactions, which I think was what we wanted to do. And that was like what we wanted to do from the beginning. That’s why we modeled it that way.

Abhi: What is the actual task for this language model?

Yunha: It’s a masked language model.

Abhi: Like given this protein sequence, inter-genomic sequence, protein sequence and so on... you mask out like 15% of that. The job is reconstructing?

Yunha: Yeah, exactly.

[00:11:06] What do you do with the output of genomic language models?

Abhi: At inference time, what do you do with the output of the model?

Yunha: Yeah, so we were mostly interested in representation learning as opposed to generation, for instance. Because our goal was... there were two main tasks. One was we wanted to see if it learns inter-element interaction. So that’s one thing we wanted to learn.

Abhi: By inter-element, does that mean  inter-protein...

Yunha: Inter-protein-protein is definitely one. So multi-protein. So protein interactions, but also we wanted to see, can we actually detect RNA-protein interactions? That’d be pretty cool because then you can find new types of RNA-guided systems, or can we just find like promoters for sequences, which we should be able to do, but we still don’t know how to do for a lot of divergent sequences. So that was what we wanted to do as like our primary task.

The secondary task was, we wanted to improve sequence representation such that we can propagate annotations better. So by that... so basically we have this problem where we have a lot of proteins and sequences, but we know less than 1% of what they do. Because we laboratory validated less than 1% of these proteins. So the problem is, there’s no way we’re gonna be able to laboratory validate all of these functions when we don’t even know what the assay is. So then the problem is we... the thing that you have to do is you need to propagate that information as much as possible, and then help that information guide the next set of experiments. And that’s the only thing we know how to do. And the only method that we’ve been doing it with was sequence similarity-based propagation. So if things are decently similar, we just call it the same thing, which is true for... to a certain degree. And then now you can do it with structure with FoldSeek and so on. If things are similar in structure, we just call it the same thing, which is also not always true, but it’s the best attempt at doing what we have to do.

So you can think of that as we’re basically compressing information across these different axes of information, which is sequence, and the other one is structure. The question is, can we do that across context? And that was a sort of motivating factor for genomic language modeling. Can we infuse like contextual information such that things that are similar in context would be pushed together in representation space, such that we can actually propagate information from one protein to another protein because they share the identical semantics in terms of context. So that was the sort of main motivator for why we wanted to do representation learning.

Abhi: And instinctively what’s the intuition for why just because two proteins are near each other, it means anything?

Yunha: Yeah. That’s a great question. So this is actually going back to why microbial genomes are cool. Unlike mammalian genomes or like anything that’s eukaryotic, microbial genomes can exchange DNA almost stochastically. That’s just part of its evolution. So things that are really far apart can exchange genomic information, which is not something that humans can do. We cannot exchange DNA with plants, right? So what that means is because there’s all these stochastic processes that’s happening in orders that we can’t even think about because there’s just so many microbes with really short, much shorter lifespan compared to our lifespan. These processes that are happening have been happening for the past billions of years.

So there’s selection pressure that’s keeping these sequences together in a certain order. And this is probably... some of these things are the things that we can rationally understand, as in these three proteins must be kept together because they literally form a complex that if one fell apart by chance, that organism just would not live and therefore would not propagate that particular arrangement of the genome. So certain ways in which genomes are arranged—gene content and genomic organization—all of these things have some sort of meaning. Some of them we can’t understand. Some of them we might be able to understand. So it is just... there’s patterns there. So how do we extract that pattern? And that is all selected upon. Some of them are random, so what we’re assuming is that the language model, by finding these patterns that are really salient, those salient patterns are probably not gonna be random. So then how do you extract noise from signal using language models?

Abhi: Yeah, it makes sense. Yeah. Like, the explanation of why protein-coding genes exist near each other means like some functional... has some functional meaning. Alternatively, I could imagine one explanation being that, oh, the microbial genome is just gonna be filled with a bunch of nonsense stuff. Like there’s one explanation of, yeah, nearness of protein-coding genes mean something because they need to travel together. Alternatively, it could be that even if one of them traveled to another bacterial genome, it’s just not used and it just sticks around there, like taking up space. Is that ever a concern?

Yunha: I think it’s less of a concern. So we talk about this junk DNA; we don’t really know what they do in like human genomes. I think for microbial genomes... there is... so no one really knows what junk DNA does, so that’s a separate conversation. For microbial genomes, if you have a gene that is not being used, there is a cost. So in order to be able to carry this forward, there is energy that’s required. There’s information burden, there’s just mutational burden. It’s just better to get rid of it.

Abhi: Yeah. That does make sense.

Yunha: Yeah. So I think it’s really difficult to conceptualize this because we’re thinking of it as, oh, like there’s gonna be so many random things that happen. But if you look at it from across samples, across history, the patterns that get picked up... there is a reason for that pattern.

[00:17:41] How will OMG evolve?

Abhi: Going back to the OMG data set, ‘cause I realized I have more questions about it. I imagine OMG is not gonna be like the final iteration, like the final metagenomic database. What do you wanna improve about the next version?

Yunha: Yeah, that’s a great question. So metagenomic databases are exponentially growing, so there is the sort of the size consideration. So I think since... I forget when exactly OMG came out. I think it was like a year ago. It basically grew almost like twice. So you can imagine like that being a big piece of what OMG-2 might be.

I think there is also sort of new types of data that’s being generated. So when it comes to things like epigenetics, so like methylation signal... it’s not as prevalently available as the raw sequence data or like the assembled genomic data. But I think that subset of data that has methylation calling done by the sequencing technology itself, I think that’s a really interesting data set to include or to subset. So I think ideally, OMG extends beyond genomic data into transcriptomic data and other types of omics data. So that’s the vision that we have down the line. But that does require many more iterations.

Abhi: Are you not a “DNA is everything you need” maximalist?

Yunha: No.

Abhi: I guess has anyone trained a DNA-plus-epigenetic or some other type of modality model and seeing that there are vast improvements in being able to represent something? I guess like you did that with genome and proteomic stuff. But has anyone else extended beyond that?

Yunha: Yeah, I think there was a new paper that came out recently. For human and mouse genomes where they included a bunch of like functional genomic data. I think it came outta Genentech actually. That was an interesting paper. I think it’s exciting, because you are basically adding genomic data with a bunch of other tracks of information. I think the sort of limitation there is you can’t do that for a vast majority of life branches. So you can’t call it like a foundation model for biology because we simply would not have that data for most branches of life. Like basically everything except like human and mouse and maybe a few things that we can culture.

[00:20:26] Why train on only microbial genomes, as opposed to all genomes?

Abhi: Why—this is almost like a cultural question—why is there the separation of like metagenomics and human genetics? Why isn’t there, like, why isn’t gLM2 trained on all genomes?

Yunha: Yeah. That’s a good question. So, all genomes as in mammalian and... yeah. Okay. So I think there are some practical reasons why we didn’t extend our model to eukaryotic genomes. One reason is like for plants, you can’t even call genes for a vast majority of their genes. So calling genes is not a trivial task for even some microbes actually.

Assuming that a sequence that you currently have in front of you is a protein sequence or protein coding sequence, that is not an assumption we can always make for a lot of genomes. Given our sort of data structure, we couldn’t make that assumption for plant genomes, fungal genomes, or mammalian genomes. There’s that consideration. Also, there is... microbial genomes are really tightly packed. So there’s very few intergenic, or very small intergenic regions that you have to consider. Whereas for eukaryotic genomes, there’s really long intergenic regions. So in order to be able to model multiple proteins at the same time, your context length has to increase significantly. And that was just not a very... it was not a practical thing to do for our model.

And I think in terms of... if you’re thinking about like data, like bang for the buck kind of situation, you’re getting so much more from microbial data, not just because it’s things are more packed, but it’s just way more diverse. So if you were... if you had a pool of data that was organized in terms of diversity and you were picking things out, like vast majority would just be microbial genomes and microbial genes. So why inject human bias and then add a human genome when it’s not really for understanding human genomes in its innate purpose? So that was the reason why we didn’t include mammalian genomes.

[00:22:58] Do we need more sequences or more annotations?

Abhi: Is it fair to say at this point, the thing you need to turn up is quality of the existing data rather than quantity of like more sequences? Or is it like non-obvious?

Yunha: I think it’s very obvious we need more labeled data and I think everyone probably agrees there. The question whether we need more metagenomic data or more unlabeled data... I’m probably... it’s probably nice to have. It can’t hurt. But it’s just a matter of... you have a lot of metagenomic data and then you find patterns that are becoming more and more salient because you have data that’s less sparse and therefore you are recognizing cooler patterns. But there’s no way of understanding what those patterns are if you can’t match it to any labels. So that labeled data is a lot more valuable, in my opinion.

[00:23:54] Is there a conserved microbial genome ‘language’?

Abhi: What do you think is... I guess this is a good question the protein people have also, but I imagine like proteins are a little bit more conserved. There’s 20 possible amino acids. Maybe not. Maybe that’s also a contentious point, but... At like gLM2, how close to like full universal microbial... or how close is it to like fully understanding the universe of microbial genomes? Like if we take gLM2 and we apply it to say the genome of like a hydrothermal vent bacteria... how good is it at representing that particular genome?

Yunha: Yeah. So if it’s in the training dataset...

Abhi: Sure.

Yunha: It will be better at it than when it’s not in the training dataset, as with any language model including protein language models. If you throw in a sequence that is very different from seen sequences, then ESMFold will fail. AlphaFold will fail. Same with representations for gLM2 and so on. So yeah, I think there is value in training this sort of like base layer, going from sequence to some sort of representation or some sort of understanding. Because yeah, if you have a really divergent sequence that’s out of the training set, then it will not generalize to that particular sequence.

Abhi: Yeah, I guess like the dream for the non-MSA protein language models is that it has this like universal understanding of proteins, regardless of like how many MSAs actually exist for the protein. Like, for Alphafold, as the MSA depth goes down, performance gets worse. Do you see something like that also for GLM where like as a sequence gets further and further away from the training data set, it also goes down?

Yunha: Yeah.

Abhi: Do you think you’ll ever escape that? That you’ll ever discover some like universal grammar for microbial genomes? Or it’s just so diverse, it’s like unlikely.

Yunha: I think it’s probably the latter, but maybe there are cool new advances that prove me otherwise.

Abhi: Moving away from the actual dataset and like more closer to the model... what was the context size for gLM2 and why did you pick it?

Yunha: Yeah, I forget the exact context size, but the benchmark was we wanted to include about 10 genes. And the reasoning there was, we’re looking at sort of an average length of operons, or average gene number for operons, and then we wanted to have multiple operons. And so that came down to about nine or 10 genes.

Abhi: Do you see, or do you intuitively expect as a context window expands you see better and better representation performance? Or does it probably max out?

Yunha: Yeah, that’s a good question. We experimented a little bit with varying the context length for the tasks that we benchmarked against. We did not see a significant improvement as we increased the context length. But that’s the benchmarks that we used, which is limited because what we know is limited. So yeah, it’s all against what you’re measuring. So if you’re measuring against something that’s super obvious, then the model is gonna learn something without needing a lot of context. But if you’re measuring for things that require multi-protein context across multiple proteins, across interactions that are really far apart, then maybe it actually benefits from including that context. I think the things that we’re measuring are too shallow. And too... we are trying to understand biology and we’re chipping it away at emergent properties that come from biology and these are really obvious patterns that we’ve observed. So no wonder these obvious patterns are the first ones to be picked up, without necessarily requiring like large context.

[00:28:11] What non-obvious things can this genomic language model tell you?

Abhi: When you say obvious patterns, I’m curious, like what did gLM2 tell you about microbial genomes that was like interesting? Like you mentioned like it was able to pick up inter-genomic elements and like what each one of those inter-genomic elements potentially mean. What could it do besides that?

Yunha: Yeah. So one thing that we were able to showcase was protein interaction. So it’s not just about “oh, these genes co-occur.” But these genes actually have co-evolving residues that goes across multiple proteins that actually maps to the protein interfaces that are known. So if you apply that to things that we don’t know much about, then we can actually resolve new types of PPI interfaces.

Abhi: Can you walk me through like how you extract PPI information from a model like gLM2?

Yunha: Yeah. This actually was in collaboration with Sergey’s Lab. Sergey’s Lab showed that you can use this method called Categorical Jacobian, where you are getting out co-evolving residues within a protein. And you can basically use that in order to identify residues that are close together and therefore co-evolving. And then you’re basically like turning the 3D structure into a 2D space. B

ut you could technically do the same thing for protein-protein interaction. It’s just two things folding together. But in order to do that, you need to have an understanding of which of these protein variants co-occur in the genome, right? So that if you just have protein A and 50 variants of protein A and protein B and 50 variants of protein B, but you don’t know how these two things are connected, that kind of signal goes away. But then if you know that A-dash and B-dash go together, A-double-dash and B-double-dash go together, then you are able to actually resolve that statistic where things are co-evolving across protein A and protein B.

Abhi: Yeah. When you say like they co-evolve... Is that translate to like they’re close in the embedding space?

Yunha: No. So co-evolve literally meaning if one residue changes in A then another residue that it’s in contact with changes too because of the biophysical sort of...

Abhi: Oh, so this is like relying a little bit on gLM2’s ability to generate genomic sequences or...

Yunha: So it doesn’t generate. So yeah. So basically if PLMs, or ESM, essentially learns the compressed MSA [Multiple Sequence Alignment], right? gLM2 also learns compressed MSA, but it’s paired. Which means... if you just had A and B together and then you concatenated them and then ran MSA, you would actually get similar signals. So you’re basically finding that kind of signal because you’re incorporating genomic context into modeling.

Abhi: Sorry, I’m just like mentally trying to walk through, because I think in... I may be incorrect, but like Sergey, the Categorical Jacobian paper was like mutating residues and seeing like what does the model think. But here it seems like it’s something different. Oh, it is the same thing.

Yunha: It is the same thing. Yeah. It’s just... think of it as like a sort of interpretability method.

Abhi: Okay. Okay. That makes sense. Yeah. I would define the Categorical Jacobian thing as like a mech interp, outside of that... was there anything else interesting you could pop out? You also mentioned you were able to derive like RNA-protein interactions. Is it using the exact same method?

Yunha: So we have seen some evidence of RNA... yes, using the same method... RNA-protein interactions. We haven’t been able to validate them. So I can’t speak for it.

Abhi: I was gonna ask like, how are you with identifying genes that’s perhaps a little bit easy... How do you identify purely RNA coding regions?

Yunha: Yeah. So small RNAs and tRNAs have such conserved structure. It’s actually very easy to spot them using GLM’s way of looking at the data. So if we ran like Categorical Jacobian on a stretch of DNA that contains RNA coding region, I guess RNA sequence, then it lights up immediately because of the hairpin structures. That’s really salient in RNA.

[00:33:08] Semantic deduplication and evaluation

Abhi: And one thing we have been continuously talking about is like models like these are potentially really useful for annotation of existing metagenomic sequences. And I think there was this really interesting thing you did with the OMG dataset that actually relied on the gLM2 model that it was trained upon called semantic deduplication. Would you be able to like just walk through what you did there?

Yunha: Yeah. So this was to tackle the exact problem where we have arbitrary chunks of DNA. And because of the way... so the classical way of deduplicating would be sequence alignment. So that’s what protein language models do. So you cluster and then you cluster using basically sequence similarity, and then you pick from the cluster. And that’s one way to make sure that you’re not over-representing your training data with one cluster. So you can’t really do that with arbitrary chunks of DNA because... assume that you have a chunk of here and then you have a chunk of that’s like this. It will align here, but it won’t align there. And also, because it’s so long, you can’t align... you can’t cluster DNA so quickly because alignment gets really expensive as you increase the length of the DNA.

So that was like a problem that we needed to solve in order to de-duplicate or de-bias the data set as much as possible. So we were actually looking at computer vision literature, and they have the same sort of problem where there’s just a lot of images and how do you make sure that you don’t have a model that’s only trained on like cats and dog images, because that’s what people like to take photos of? Then you need to like either classify them or... but then if you just classify everything as dogs, then maybe you wanna keep some of the diversity in dogs. So there is... how do we de-bias the data set with as little bias as possible, as little human bias as possible?

So I think one method that people have used in DNA language models is, okay, let’s just like use taxonomy as label, and then we’re just gonna sample one from this genus, one from this genus, which I think is a fair thing to do. But the problem with metagenomics sequences is that you don’t always have taxonomic labels. You’re literally getting sequence from like a pile of dirt.

Abhi: Which may include like brand new genus, is that right?

Yunha: Exactly. And you don’t wanna bias against those either. So we wanted to basically... we trained a small model that was essentially the same thing as gLM2. And then we embedded all of these contexts and then we sampled from those contexts in order to be able to de-bias the model as much as possible.

Abhi: How do you judge whether this like works?

Yunha: Yeah, so we had a benchmark. So we basically designed a benchmark that was actually quite a lot of work because if you just rely on existing benchmarks, then the model seems to be doing worse once you make the data set more diverse. And the reason is... the data itself is over-represented with e.g. E. coli because that’s what’s most studied, but also what the benchmarks are based on is also E. coli. So then might as well just train an E. coli model. Why do you even go about training a metagenomic model? So what we did was we actually, before we even trained the model, we actually worked on getting together a really diverse set of embedding benchmarks. And this is really like going... we are, when we are sampling sequences, we’re sampling across the tree of life, not just from E. coli. And that was like a very deliberate thing that we did before we even started training gLM2.

[00:37:33] How does benchmarking work for these types of models?

Abhi: What does benchmarking even look like for a microbial language model? Are you purely measuring yourself by your ability to reconstruct the genome or is there something else?

Yunha: No. So we don’t even actually consider perplexity as like a good metric. So what we did was... we looked at how good the representations were, or as in the embeddings were, for various tasks. So one is a classic task of: does it actually capture phylogenetic relationships between sequences. So there are statistical models that you can use in order to resolve the phylogenetic distances between sequences. And I guess the important thing to do there is to make sure that these sequences are sampled across the tree of life. So we did that and then we basically compare the embedding distances to phylogenetic distances between the sequences. That’s one benchmark. Another benchmark is: can this embedding represent... can this representation space actually compress information such that sequences that are far away in sequence space or structured space, but actually do the same thing in function, bring them closer together? So that’s... you’re using like metric that is like “nearest thing in space” in order to retrieve. So it’s a retrieval-based benchmark in order to be able to find things that are similar in function that we’ve hand curated across the tree of life, to see if you can do that using embeddings only. So we are benchmarking against ESM and other types of embedding to see if it performs as a retrieval task.

Abhi: The thing I would be like very... I think like protein, like, RMSD benchmarks are oh, like fairly trustworthy. ‘Cause you can trust that the x-ray crystallography was like correct. With function annotation, how much can you trust that like these papers that you’re pulling the functional annotations from actually did their job correctly?

Yunha: Yeah. That’s a good question. So we... I mean it’s like really hand curated. We do look at the papers. We make sure that the function that we are looking at is correct. So for instance, like enzyme functions. So people... so I think that was actually one of the benchmarks. So given the sequence, you’re trying to predict the EC number, which is an Enzyme Commission number, which represents what kind of reaction it can catalyze. But the problem is there is positive data... but one enzyme can actually do multiple enzyme reactions depending on the context. So just because it wasn’t documented doesn’t mean it’s not possible, right? So it’s actually very common for a single sequence to be able to confer multiple enzymes in certain hierarchy, that are in the same class, but different substrates. So it’s... so there are cases where our model actually predicted, “Oh, this sequence is likely to conduct both of these reactions with equal probability or similar weight to each of these reactions.” And there’s only data for one, but not the other. So we cannot really say for sure that this is wrong. There are definitely gaps in the data that we need to be aware of, even when you’re really carefully curating this data set. But that’s also an interesting case to look into because it’s... yeah, it’s spotting things that we didn’t spot it before.

[00:41:31] Gaia: A genomic search engine

Abhi: That makes sense. Yeah. And actually OMG and gLM2 are actually some of your earlier work. I think your latest paper is about another genomic language model called Gaia. Could you walk me through what Gaia exactly is?

Yunha: Yeah. So Gaia is actually not a genomic... actually it’s a... I would call it more of a system that’s built on top of gLM2. Gaia is essentially a search engine. So what we wanted to do was demonstrate that gLM2 embeddings can be used to find sequences that are similar in function. And the way we did that was: okay, it needs to definitely find sequences that are similar in sequence, because otherwise... that’s like the least you can do. And then you should find sequences that are also similar in structure. But also you should find sequences that are similar in context. So that’s what we wanted to do. And gLM2 representations were suited for that because it has all that information as part of the training. So Gaia stands for Genomic AI Annotator. And the first thing it does is it retrieves sequences that are similar in gLM2 embedding space. And then the next thing it does is it actually maps that embedding to text descriptor so that we can annotate more rapidly.

Abhi: So you input in a genomic sequence, you find all the nearest proteins via gLM2 embeddings. And then how do you convert that to text? You just like pick the closest protein?

Yunha: Yeah, so we use Swiss-Prot as our sort of golden dataset. That’s probably the best curated data set that we currently have. And so that is pairs of protein sequences to a text descriptor, right? So we train a CLIP model on top of that.

Abhi: Yeah. Okay. And so like... so you’re relying on the full universe of proteins in Swiss-Prot to represent also the full universe of possible functions.

Yunha: Yes.

Abhi: While that very well may be valid... do you suspect that there are possibly like microbial proteins or inter-genomic elements that are not cataloged within Swiss-Prot?

Yunha: Yes, certainly. Vast majority. That’s the whole point. Yeah. And we also choose not to... there is a threshold where we say “no function” or “no known function.”

[00:44:18] Even ‘well-studied’ genomes are mostly unannotated

Abhi: I am not aware of this literature at all. How often is it that people find some weird microbe able to do something that no other microbe can do?

Yunha: Very often. So if you look at a microbial genome, and even for really well-studied microbes such as E. coli and Mycobacterium tuberculosis, you’re finding half to two-thirds of their genes being unannotated we just don’t know what they do. And that’s not even including things that are just like “this is a membrane protein,” which still doesn’t tell us anything about the function. So there’s that problem. But if you look at a random microbe from soil, 80%, 90%, if not 95% of their genes will have no annotated function using basic like sequence-based methods.

Abhi: ...when a microbe can do something that’s never been observed before.

Yunha: Yeah. So that happens. I would say that’s why environmental microbiology was so interesting. There were literally microbes that were being discovered like left and right that can do crazy chemistry. Like literally live off of... it breathes rock as opposed to oxygen. Or it converts disproportionate sulfur, like elemental sulfur, into like sulfite and sulfate... that kind of reaction. We just don’t even know how to do What else? Just things that are like living off of uranium and using that energy, or harnessing that energy to live. Microbes that just live for a million years and we don’t know why and how.

Abhi: There seems to be like two elements here. Like one is trying to annotate functional genomic elements that we reasonably understand... like, “what’s this?” like “this exists somewhere else in the microbial kingdom. Maybe this does it in a different way, but like the function is conserved across other domains of life.” And on the other side, which feels like the far more interesting bit, is that there are microbial functionality that exists uniquely within the species and exists nowhere else. How common is that latter bucket? Like you mentioned like uranium eating bacteria, like rock eating bacteria. Is it usually there are very specific species that do this exact thing and nothing else does it?

Yunha: So interestingly, there’s more and more cases of convergent evolution happening where there’s multiple ways of doing the same thing, which is not that surprising if you think about it. So that’s why I think this idea of compression is actually an interesting idea. If there is like more like a sort of layer to biology that we didn’t fully understand... so we know how to look at sequence pretty well now. But if there are patterns underlying those, and then if we can use those patterns to actually match functions, so that we can actually discover new functions that have conversely evolved to do the same thing. That would be really cool.

Abhi: Going back to Gaia now... I imagine you have this setup for turning the pre-trained GLM embedding into like functional annotations for this like dark universe of microbial genomes. Have you done that? Have you gone through every single un-annotated genome, applied Gaia to it, and is that all just stored somewhere?

Yunha: Yeah. So we did that experiment with Mycobacterium tuberculosis, where two-thirds of the genes we don’t know what they do. And we actually developed... so it was like hard to do this manually, because it’s still... you’re still looking at 2000, 3000 genes, and then you’re trying to figure out what it is using Gaia annotation. So we actually built like “Gaia agent,” which would then try to validate what the Gaia annotations are, given the context. So we basically ran the whole pipeline in order to discover new sequence functions in this really well-studied microbe that thousands of labs have studied for tens to hundreds of years. And yeah, we were able to find four proteins that we could actually validate in silico. And I’m like, “Why didn’t we know this before?”

Like one example is... it’s two proteins that each were annotated as uncharacterized protein in literally every single database that we looked at. And then when you search it individually, you don’t get any matches. But then if you fold them together and search, you actually get a match to an Archaea, which is an entirely different domain of life that have diverged billions of years ago. And you get very little sequence similarity to the extent that you won’t be able to find it using typical tools. But if you look at the structure, it’s actually almost identical. And that’s like a membrane transport protein complex. And then another one was... that one was really interesting because it was like a very small ORF that was never annotated in Mycobacterium tuberculosis because it was really small, but then it also had two other proteins that transforms that tiny little protein peptide into something that’s antimicrobial. So that’s something that’s three systems that we weren’t able to identify previously because we are only looking at each one separately instead of looking at the full picture.

[00:50:51] Using agents on Gaia

Abhi: Could you walk me through Gaia as a platform... makes a lot of sense to me. What does Gaia agent exactly do?

Yunha: Yeah. So Gaia agent, what it does is what a really good microbiologist would do in silico, but just automates the whole thing. So Gaia agent looks at the full context, which is what microbiologists would do. So you see a protein and you look at its annotation. You look at all the motifs that this protein has alongside all the motifs for other proteins, and all the DNA sequence motifs. And then you’re like looking for patterns across the tree of life. Oh, these two things co-occur together, or there’s a co-orientation and very small spaces between the genes, which likely means they actually travel together. And then you’re doing reasoning across the functions of... “this reaction happens and this reaction happens. Most likely this gene is probably doing the reaction that goes from this product to this substrate.” So if you have a reaction chain, for instance, then you can actually figure out... So you have product A and then substrate A going all the way to product D, and then there’s steps B and C. And we have reaction enzymes for reactions in the first part and then the last part. But we don’t know what’s doing the middle part. You can make a reasonable guess that the protein that’s found somewhere near those two proteins might be doing that particular reaction. And you can actually use that kind of reasoning to be able to essentially fill the gaps and de-orphan this particular enzyme reaction.

Abhi: So does the reasoning... so like Gaia agent treats like gLM2 as a tool alongside like the rest of the literature?

Yunha: Yeah. And also other tools such as like FoldSeek. So we give it FoldSeek and you give it other types of bioinformatic tools that you know you can access in silico. Ideally you also have access to like automation labs. We’re not quite there yet.

Abhi: Why is like... is it just like too computationally expensive to just let this rip over the entirety of all un-annotated microbial genomes?

Yunha: Yes. It’s not cheap to run this. And we’re looking at a lot of genes. So one thing we’re actually looking into doing right now is we are gonna look at a few hundreds to like few thousands of genomes that are like on the wishlist of all of these biologists. So we’re just gonna run it and then see, and then also share that result so that people can use it.

Abhi: I’m curious... I’m completely unfamiliar with what the typical metagenomic workflow of a biologist looks like. What’s the fundamental difference between just like providing a gene sequence into gLM2, seeing what proteins are nearby in Swiss-Prot and like nearby in the embedding space... picking up the nearest Swiss-Prot protein as “okay, this is what this protein does”... versus using Gaia agent? Why do you need reasoning on top of that?

Yunha: Yeah. So if it’s a sequence that has a good match to a Swiss-Prot sequence, then you know...

Abhi: You go home after that.

Yunha: Yeah, you don’t need to even run Gaia. You can just do this with BLAST. I think the problem is for a vast majority of genes, you don’t even have that match. That’s why when you run a typical genome into like genome annotation tool that relies on BLAST, you will get 80 to 90% of the genes as unannotated or something that’s meaningless. So how do we make that 50% or 40%? And that’s done by compressing that space so that we can make more associations faster.

[00:54:53] Will genomic language models reshape the tree of life?

Abhi: You had this offhand comment about like how you discovered an Archaea-esque protein within this very well-studied protein that is distinctly not Archaea. And you’ve also mentioned in the past that like how potentially models like these can dramatically change our understanding of what the Tree of Life or phylogeny in general looks like. I’d love to get just like your take on that subject.

Yunha: Yeah, so I guess on the sort of Tree of Life side... So I don’t think the language models will replace phylogenetic trees. Phylogenetic trees are a lot more complex... I mean this is a whole discipline that’s built on top of like how things mutate, what are sort of models of mutation that we should be using...

Abhi: But still all sequence based, right?

Yunha: It’s all sequence based. Yeah. But there’s just a lot of modeling that’s there. And, yeah, I think you should almost see the phylogenetic trees as almost like ground truth to how things evolve. Just also because these things also take a long time to compute as well. So I think there is a future where we can get like cheap and easy phylogenetic trees using language models and embedding spaces, and that would be like an easy way to get a quick look at how things are related. But in the end, phylogenetic analysis have its own space in science literature and science analysis.

I think what’s changing though is as new sequences come about, and as we sample more, the tree is shifting. Because you are only constructing trees based off of what we can sample right now, right? But if you add new branches, the branch structure changes. So for instance, like an example is... we don’t know if eukaryotes... the traditional way of thinking about the Tree of Life is that there’s bacteria, there’s Archaea, and then there’s like a special branch of eukaryotes. What we were actually realizing is that actually the Eukarya are just like a single branch from Archaea. And that has like fundamental change in how we think about the Tree of Life. And that only happened because we actually sampled this hydrothermal vent that contained this Archaea that was closer to eukaryotes, but also still part of the Archaeal tree. So now humans and eukaryotes, the entire branch of eukaryotes, belong to Archaea technically.

Abhi: That sounds like a dramatic reshaping of how we think about... so in that sense, why don’t you think the same thing will happen if you bring in genomic language models? Like why won’t it dramatically change that tree of life in a similar way to that Archaea discovery?

Yunha: Yeah. So because I think that discovery, the amount of information that both models, whether it’s a language model or a phylogenetic model has access to, is the same.

Abhi: So sequence alone gets you like 80% of the way there and like whatever genomic language models bring to the table... it’s probably not like a massive amount...

Yunha: Yeah. I don’t think it’s gonna shift the shape of how things evolved. And we also don’t have a way to validate any of that.

Abhi: Interesting. Do you think you’ll ever want to do phylogenetic research?

Yunha: So I did some of that when I was more in the environmental microbiology research. I think it’s really fascinating, the kind of work that you can do in retracing what happened across the tree of life and the history of Earth. I think that’s really cool. I do also find it a little bit frustrating that you can’t be entirely sure, because you can’t go back in time. But it’s... I think there’s really cool science that comes out of doing phylogenetics.

[00:59:18] Current limitations of genomic language models

Abhi: It’s interesting ‘cause I think also like Sergey [Ovchinnikov] has an evolutionary biology background. It’s interesting how these paths are converging a little bit. One thing I did wanna ask is we’ve talked a lot about the extreme promise that all of these models have. One thing I’m wondering about is where do they currently fall apart? What particular like species genomes problems do these models not currently work well today in?

Yunha: Generally they don’t do well when the training... when it’s on a problem where, or on a genome where it’s not well sampled in the training set. So that’s... I think everyone knows that now. There’s no surprise there.

I think in genomic language modeling, DNA language modeling, what we wanna do with these genomic language models are not still clear. And I think that’s largely because we don’t have a lot of paired data. So when we think about protein language models, it’s pretty clear how you can assess the quality of the protein language models because you’re trying to go... there’s a pair data of structure, right? So you have a lot of protein sequences and there is really good set of structure from very different systems and so on. So you can actually benchmark against structure. But for genomic language models, I would argue we don’t have that data to benchmark against. And I think everyone likes to talk about function, but I think that data set is still very much limited and extremely biased. And it doesn’t really... it doesn’t like do the justice of showcasing that GLMs are learning functional information. It’s just impossible to utilize this model because there’s nothing to pair it to. So like for protein language models, you can use it to design a new structure or new sequence. But for genomic language models, because we don’t have this other modality to condition it on, we don’t know how to use it yet.

Abhi: Do you think we’ll ever get to the world of like single “model to rule them all” ? Like maybe gLM2 also spits out protein structure and like maybe that’s an area you can like check. Does that make sense? Like you have these auxiliary outputs that help you ground... help you understand what is the model able to understand versus where it’s like a little bit up to vibes and like you’re unsure as whether it’s understanding it.

Yunha: Yeah. I think that’s how we’ve been benchmarking a lot of these models, right? Like Evo and gLM2... we can make gLM2 generative as well, and then we basically generate a protein and see how good the protein is. And then we benchmark against the protein language models. We can do all of these things, but what’s the point? Like you can just have also a protein language model. So I think we’re still figuring out like... what is the problem that we’re trying to solve with genomic language models? For us, we’ve been focused on like annotation. How do we make annotations better? How do we make representations better? But one thing that we’ve realized is, yes, we can make representations really good, but we still need better golden data set in order to make a bigger dent in how we are understanding genomes. So it’s like a... you need to attack it from both angles, like more labeled data, better models and keep going in both directions. So that’s one sort of area that people can work on. I think there’s also like genome design, is another. I think the same problem comes into play. Like what is a “better” genome? For proteins, I think you can... there’s an axis that you can optimize on. I don’t know, like binding affinity or something. Thermostability. Like things like that. For genome, I think that’s a lot more... I think there are ways to fine-tune it to do one thing. But there’s no general sort of axes that you can like optimize generations for.

Abhi: I know that this is something you’ve mentioned in the past about how like microbes are often capable of chemistry that is either almost impossible for us to do, or straight up just impossible for us to do. Is it not a clear benchmark, just being able to generate a microbial genome, which like innately allows you to sustainably produce something that we otherwise cannot do outside of that microbe? Do you think like we are close to that at all? Like for gLM2, how good is it at generating microbial genomes outright?

Yunha: So in order to do what you said just right before—which is, wouldn’t it be the benchmark to be able to show like, “oh, this generation can do something that nature cannot do, or something that we wanted it to do, that doesn’t already exist”—then you need to be able to condition.

Abhi: It needs to be in your train set.

Yunha: Yeah, but what I’m trying to say is that conditioning signal or conditioning dataset doesn’t quite exist at its full scale to be able to do that.

Abhi: Let’s say that you just wanna replicate something. Like there is like this one microbe that like feeds off of uranium. You wanna be able to create a microbe that is very much like it, but perhaps is as easy to grow as E. coli or something. How well can you do that today?

Yunha: Yeah. That’s a great question. I think that still comes back to the annotation problem. Where given an your like microbe that can feed off of uranium, we don’t know which parts are important. Which parts are not important.

Abhi: Yeah. I guess this is why you potentially would want to max out the context length of a model like this. So you can just feed in... either you can get the model to spit out an entire genome and then you don’t need to know what is important, what isn’t important. Is that a fair way to think about that?

Yunha: Yeah. So then... what would the training objective look like? You will have genomes that can do a like chemistry X. And then you need to generate a sequence given this like chemistry X and then you need to make it also like E. coli.

Abhi: Yeah. I think that second part’s a bit difficult.

Yunha: Because otherwise if you just say, okay, like we already know this genome Y can do chemistry X. And if you tell the model to build a genome that does chemistry X and it will just output something that’s similar to genome Y, and you could say, “Oh, that works.” Like maybe you get really lucky and it’s a few mutations, synonymous mutations away, such that it doesn’t actually change the biology at all. But all you’ve done is just like maybe I don’t know, learn synonymous mutations.

Abhi: One thing I was surprised by by the Evo-2 paper and perhaps all genomic language models is that it is difficult... there’s no way currently to condition it on anything other than sequence. Why hasn’t someone built a model that could be conditioned on function?

Yunha: Yeah. Because there is no good pair data sets.

Abhi: But there’s some. You’re just saying like there’s not enough?

Yunha: Yeah. There’s not enough. And also I think paired dataset exist for proteins. Not really for genomes or segments of genomes, right?

Abhi: Especially for segments of genomes. But if you have a model ingest the entire genome, maybe the functional annotation could just be like: “Eats this, grows this amount.”

Yunha: Yeah, I think that... so if somebody curated that data set and did it, and it’s accurate, which I think is a big if, then I think it’s possible. You can basically build a database of natural language description of a genome. But that also relies on us understanding the genome, right? So okay, so you have a genome and you’re like, okay, there’s a cellulose degradation pathway. There is like a carbon fixation pathway. So you already know okay, this organism is gonna grow like this. So then in order to condition a generation on that function, then the only vocabulary that you can use is the vocabulary that you’ve used to annotate that genome. So you’re completely limited by the capacity to be able to annotate that genome, which comes back to the annotation problem.

[01:08:54] Directed evolution as training data

Abhi: Have you heard of like Pioneer Labs? This like forcing microbes to evolve down a certain path. And then evolving... observing like what the genome looks like after that. Do you think that’s a particularly interesting way to gather data and it’s maybe like what more people should be doing?

Yunha: Yeah, I think... so like more on the directed evolution side?

Abhi: Maybe I’ll give like a quick description of what Pioneer Labs is. It’s a company that basically wants to create microbes that are able to survive... in I think Mars-like environments, which is just basically just extreme environments in general.

Yunha: Yeah. I think it’s really interesting because it gives another sort of dimension to the data that we didn’t have readily available. So it’s the same thing as if you’re learning how to drive a car, it’s much better to see how the car drives than see the final state of where the car is. Like I think you could potentially learn how the car drives by seeing a lot of photos of cars in different contexts.

So that’s what we’re doing. But then if you had more trajectories and you learned more from trajectories, I think there is a path forward in learning something that’s more meaningful. And that can be modeled better. So I think that... I think there’s a lot of potential there. I think one caveat there is you can’t do this kind of directed evolution for all types of functions, nor all types of organisms. So you’re... but I think that’s fine. It depends on the question. If your application is in an organism that can be cultivated and for a function that can be optimized for, then it’s the right approach to do it. You just can’t apply that for Archaea where it doesn’t grow.

Abhi: Makes sense. How much of your research... I think you’ve focused on the kind of two different axes of this like genomic language modeling problem. Like one, like the data’s not fantastic, we need to get better data. Actually maybe three. The second is like maybe the modality, like we need more modalities of microbial genomic data. And the third is the models which, Gaia agent is maybe like an improvement over just like gLM2 alone. Which of these three are you most interested in personally pushing forwards?

Yunha: Sorry. The three were... one, what was... yeah, sorry.

Abhi: The one is like the total quantity of like labeled genomic data.

Yunha: Oh, quantity of labeled genomic data. Yeah.

Abhi: Or potentially unlabeled as well.

Yunha: Oh, yeah.

Abhi: The second one is like modalities beyond genomics. Third is like the model itself and pushing on that direction.

Yunha: I think they’re all tied. Because the label data is like... you’re labeling and therefore you’re adding another modality to your dataset.

Abhi: That’s fair. Yeah.

So yeah. One and two are the same.

Yunha: Yeah. Yeah. So I think for me, I guess adding new sort of data modalities to genomic data, I think is the most exciting path forward because then you can start actually conditioning things on function, like you can actually imagine being able to do things that we can’t do with the toolkits that we currently have and the knowledge that we currently have. I think that’s just the most exciting path forward.

[01:12:35] What is Tatta Bio?

Abhi: Yeah. That makes sense. And so yeah, we’ve talked about OMG, gLM2, Gaia and also Gaia agent. Many of these things were spawned from Tatta Bio, which you’re one of the co-founders of. It’s a scientific nonprofit dedicated to developing like tools for genomic intelligence. Why is it a nonprofit?

Yunha: Yeah. Tatta Bio is a nonprofit because we’re trying to tackle a problem that maybe too big to tackle for an academic lab in an academic setting. And also very interdisciplinary in terms of... it does require a lot of software talent and machine learning talent, which there are plenty in academia, but it’s difficult to just organize that team in an academic setting. But also there’s no immediate incentive for the market forces to solve this problem. So, say for instance, like the annotation problem... It’s clearly a really important problem because it limits what we can study and what we can understand, and it obviously is gonna underpin new research directions that have unknowable like value. But neither the market nor academia are tackling this in the sort of the scale that we wanted to tackle it at. So that’s the reason why we are a nonprofit.

Abhi: And what is the actual... like I mentioned like Tatta Bio is developing “genomic intelligence.” I think that’s straight up like on the website. What is the... what do you consider the purpose of Tatta Bio to be in terms of what is it delivering to people?

Yunha: So what it’s delivering to people right now is helping people to better understand their genomic sequences. I think it’s clear that genomic sequences cannot be understood by humans. So human-machine sort of collaboration has always been the case for understanding genomic sequences. And how do we make that better? How do we augment that? So that’s the big mission that we have. So that’s how we... what we mean by genomic intelligence. Being able to truly, truly understand genomes, but not necessarily in the sort of like the rational sense that we have. It’s like “this part does this and this is evolved because of that.” It’s really being able to harness the genomic information that’s currently available and engineer it and modify it in the way that makes sense for applications. So yeah, so that’s what we are currently doing. I think within that there’s like the tool building, there’s infrastructure building, there’s community orientation. Like all of those things are sort of part of our mission.

Abhi: Actually one question I wanted to ask for a while, why is it called Tatta Bio? Because actually when I’ve brought up the company to other people, they thought “oh, is it tied to that one like Indian consultancy company?” [Tatta Group]. Why that name?

Yunha: Yeah. It’s... I guess it’s like reference to “TATA box”. And TATA box is like a literally a sequence motif in DNA that’s rich in TA or T-A-T-T-A in this case, that signals the start of a gene or like a reading frame.

Abhi: Yeah, that was a good name.

Yunha: I don’t think everyone got that memo.

Abhi: What would you... what would make you think that like we’ve succeeded at Tatta?

Yunha: Yeah. For us, if we could... I say for instance, if we could double the number of sequences that can be annotated. I think that is a success.

Abhi: To some degree it feels like with Gaia agent, you can do that today; you’re almost like just like compute limited. Is that fair to say? What else needs to be really be done?

Yunha: Yeah. I think there are just real dark patches of the sequence space that we haven’t fully explored. And I think... so if you can imagine like if it was literally just a map and there are complete dark map patches, and if we can figure out a way to generate hypothesis for any one of those sequences, that’s gonna make a big impact because now we’ve already built a very good way to compress that information so that we can propagate that information really quickly. So then... yeah, so then there are definitely like areas that we should really be studying because it’s gonna make a big impact in how we understand sequences. So that is how I see it as a sort of next step. How do we identify those areas that are really poorly characterized, but has high impact potential, and go about experimentally validating some of these sequences and functions.

Abhi: So is it like... I guess I keep returning to this question. The reason you don’t wanna let Gaia agent just run over the entirety of un-annotated sequences is that you’re unsure about the validity of any one of those given predictions, and there’s like more work to do as to figure out like where is Gaia agent reliable and where is it not reliable? Or is there something else?

Yunha: So well, I guess like you can always generate hypotheses. But the question is how many of these can we actually validate? And how many of these is it worth validating given the sort of resource limitations that we currently have?

Abhi: Like I imagine one thing you could do is like let it run across all microbial genomes and then just give that information to the community. And see what they’re able to come up with.

Yunha: Yeah. Yeah. So we are basically trying to do that. But we can’t do it across the entire trillions of genes. So we’re making... we’re trying to make a good selection of either genomes or genes that are like on the wishlist of people and scientists.

[01:19:02] Building Google for genomic sequences (SeqHub)

Abhi: Do you imagine like... FROs [Focused Research Organizations] have a specific like specific like length of time they exist before which they become for-profit? Or they just die entirely? Because they fulfilled their mission. What do you think the future of Tatta is? Yeah, eventually there’s a for-profit or at the end of it, it just like winds down because you’ve annotated the sequences. You’re done.

Yunha: If we could figure out a way that we annotate every single sequence, which I think is very ambitious and probably not possible in the next X years, then that should be our goal. We take a stance that this is going to be an evolving like database of sequences to function and how do we best optimize this database so that things don’t get lost and things are optimally propagated across scientific literature and across scientific discourse.

One of the sort of like latest projects that we’ve been working on is called SeqHub. It’s literally like GitHub for sequences or Google for sequences. So in an ideal world, you can type in the sequence and you get all information, not just the annotation, but what papers refer to it, who are the best people to ask about it, what kind of discussions have been had about this particular sequence and what obviously what other sequences exist that are in that provenance and what kind of genomic context is found in. So we with Gaia, we tackle the genomic context problem. With SeqHub, we’re basically tackling other types of sort of infrastructure problem, because way too often people make discoveries all the time, but it’s not... that information cannot be propagated like readily, because it doesn’t fit into certain database that people have built like 10 years ago. And it just doesn’t fit. And that database doesn’t get propagated to what people use all the time.

So how do we build this more real time understanding of sequences? So that’s a big part of our mission. How do we build a better software infrastructure for sequence understanding and data sharing? And so as part of that mission, we can’t... if we wanted to fully fulfill this mission, and we have the assumption that this is gonna take a long time, we actually want to maintain this infrastructure for as long as we can fulfill this particular mission. Which... so as part of that, I think what we still need to figure out is how do we build sustainability into our operation and business model. And our goal is to remain fully non-profit, and still build in ways to generate enough revenue so that we can maintain this scientific software and infrastructure, which by the way, has been very difficult to maintain in this current funding environment. Traditionally I think it was funded by the government. But that also means certain types of innovation is difficult to switch. You can’t build a fast-paced team in a lab that is either getting funding that is not enough to do this kind of work. So we are also like thinking really creatively about how do we maintain scientific infrastructure and software infrastructure because so often good softwares get made, but are not maintained. Or good ideas transpire... like okay softwares, but doesn’t get scaled up and deployed into production level software. So this is another sort of aspect of work that we’re currently doing.

Abhi: I’m not sure if you’re like able to talk about this, but... PyMol was a really great piece of software, Schrödinger just acquired it... they have a private version that you have to pay Schrödinger to use, but they also have this very nice open source version [PyMOL] . Do you think you could imagine Tatta Bio going down that route where they’re acquired by some existing like Basecamp or someone who really cares about the information that Tatta is gathering and they allow this shaved off like open source version?

Yunha: Yeah. I don’t know. Yeah, we haven’t fully thought about that. I think what right now we’re more focused on is how do we become entrenched in this like scientific ecosystem. And I think a key sort of difference here is it’s not just a software. If it’s software, then you can just copy it and then you can improve it, and then you can share it. But if it’s an infrastructure that needs the community to deposit data, share data, then as soon as you close source any part of it, then the value of that particular infrastructure goes away. I think the only sort of big... the only sort of parallel that I can think of is like PDB. Or you could argue the same thing about Google. If you didn’t have Google that was free... to just deposit in the internet was free. But then you can’t build LLMs if you didn’t have that, the internet. Same thing with like AlphaFold and PDB. So yeah,

Abhi: Like all of it needs to be open sourced for like the network effects to actually start thinking...

Yunha: Yeah.

That’s how I think about it. That’s why I think it’s really important for us to stay open and stay like free for the vast majority of the functionality.

Abhi: Have you seen the XKCD comic? That’s like, you identify some universal problem everyone has and that says “I’m gonna build a solution to it”... and now you’ve just added another universal standard to the 13 others that existed prior. Like what other quote-unquote universal standards are there besides SeqHub and like where do you think they fall short?

Yunha: So in the space of like sequences, I think UniProt is a great example. It’s what people go to when you have a protein sequence.

Abhi: Sorry, specifically for genomes.

Yunha: Oh, genomes. Oh, like specifically like a... Oh, I see.

Abhi: What almost like network territory is SeqHub encroaching on? Are there any... or is like SeqHub unique and there is no other... there’s no other platform for something like this?

Yunha: The only other genome centric like existing platform that’s widely used is NCBI.

Abhi: And that’s not... there’s not really network effects there.

Yunha: No. Yeah.

[01:25:46] How to create communities around scientific OSS

Abhi: Okay. That makes sense. Okay then yeah, it seems like ripe territory to capitalize on. How do you... how have you typically found the process of gathering a community around a brand new piece of open source software? I imagine it’s like a relatively new experience for you.

Yunha: Yes. Yes. Certainly.

Abhi: How has that been?

Yunha: Oh, very interesting. A lot of learning on our side. It’s... yeah, it’s different in that so it is a self-serve software. And it is also B2C in some ways.

But it’s a very small community of people. We’re not tackling the general public here. We’re also currently really focused on microbiology community. And hopefully we can expand out to other communities like in plants and fungi and so on. So that’s our sort of roadmap.

Yeah, but it’s... we need to get in the head of scientists and think about what... why do we do what we do and why do we want to contribute? And how do we contribute and where do I spend most of my time? And what are the most biggest pain points that we have? So all of these things that we need to think about when we design the software and the platform. And building good software is one thing, but building a community is just an entirely new thing that we’re literally just figuring out as we speak.

Abhi: Especially if it’s yeah, like you mentioned, the community is so small. Like I can’t imagine the people who like would actively be power users of the software number more than a few thousand people worldwide. How do you like... how do you get in touch with all of those people and tell them like, “oh, you should be using this thing that we built.” Like how do you convince them that this is worth their time?

Yunha: Yeah. For us, it’s truly... so I think there’s been a lot of attempts at encouraging people to deposit data better, add more data, metadata, blah blah blah. I think one thing... we need to make it really easy. So it should be depositing data should be super easy.

And we shouldn’t require them to do a bunch of things, so that’s just a basic thing that we can build in. Another is we need to give them what they really want the most, and for us it’s better annotations. When I was a student, it’s like the most frustrating thing when you have sequences that you’ve waited so long to get into your hands and you look at it and so much of it is just hypothetical and you’re like just banging your head against the wall to understand what these sequences do. And that is the biggest motivator. If we can give them better annotations, if we can give them more insight into what they’re looking at, that’s what’s gonna bring them here. And those are gonna be the people who are gonna be the most incentivized to contribute because it will come back to benefit them and the community. So that’s our hypothesis. We’ll see how that goes.

[01:29:06] What’s the purpose in the centralization of the software?

Abhi: That’s fun. Like you have this platform which is really hard to populate to start off with, but the draw... like the reason you’d want to interact with that at all is because you get access to Gaia, basically. As like a way to help you interpret what’s going on.

Why... what’s... this is maybe something I should have asked before. Why even care about having something like SeqHub? Is it like... yeah, like maybe you want more people to use Gaia, but like alternatively Gaia could just be like a standalone GitHub thing? Like why do you want a central place to deposit sequences?

Yunha: Yeah. Yeah. That’s a great question because we’re trying to expand this labeled data set. This gold standard data set that we have, which is currently Swiss-Prot... we think there is actually quite a lot of information that’s outside of Swiss-Prot. Swiss-Prot is human curated by the way, which is incredible. There are curators whose full-time job is to look at papers and validate, “oh, this is like a new sequence. We should add this to Swiss-Prot.” I think there’s just a lot of knowledge that’s hiding in labs and hiding in people’s brains and hiding in papers and supplements that can be organized a lot better so that we can actually improve sequence annotation without even having to do any experiments. And I think that is like... if we organize ourselves properly, with infrastructure that is up to date and with correct incentivization schema, then I think we can... we might be able to like double the number of sequences that we can annotate without having even having to do any experimental workflows. And I think that is like what we’re trying to build right now.

Abhi: What’s the... yeah, you said Swiss-Prot is human annotated, which makes sense why it’s so low throughput. I’m curious like how much realistically... how much knowledge is like hiding in the heads of people at these microbial genomic labs who simply like don’t have the results necessary to write a paper about it and get it like deposited somewhere? So like how strong... what... when you talk to these people, is it usually that they have like tons of things in their head that like they’ve been thinking about it for decades, but like they just don’t care enough to write a paper about it?

Yunha: Yeah, I think that definitely exists. And I think this is also byproduct of the publication system. As in, if it’s not a big story, then where do you share this information? And when it’s not gonna be really cited, and when things are not gonna be discoverable... so there’s no incentive to write a single paper just to say “this is something.” You might be able to say, “oh, like we have experimental results.” But it’s just not gonna be a very highly cited paper. So what happens typically is either it’s like a tiny little section in a large like paper. So you write a whole paper and then there’s like a tiny little thing. It’s “oh, we think this is this, or we have like high confidence this is this, based on this tiny little supplemental figure that no one looks at.” And that never gets propagated to central database.

Abhi: Is it like the Swiss-Prot annotators just have so many other things they want?

Yunha: Yeah. So there’s that. And then there’s just internal knowledge. Like people do experiments all the time. Like we do a lot more experiments than what gets published in the paper. So I think there’s both of those sort of like at play, in terms of what is a publishable unit, how can we make knowledge transfer be more efficient across people. So imagine if you had to write a publication for every single bug fix in software. That just doesn’t make sense.

Abhi: And so like SeqHub, I think you guys officially released a month ago. Am I correct? And so a month has passed. What’s next on the roadmap? What do you... what have you seen the use cases are so far?

Yunha: Yeah. So what we... so we launched SeqHub about a month ago. And a key sort of difference between SeqHub and Gaia is that SeqHub can do like whole genome annotation. And it’s also a place where you can deposit data.

Abhi: Sorry, how does it do whole genome annotation? Just split it up into...

Yunha: Yeah. So basically, you can pull a... so Gaia is a sequence, like protein search. But we’ve extended it across like the full genome. So if you put multiple sequences, which is a genome, then it does automated annotation.

Abhi: Gotcha. Okay.

Yunha: So then now you can automatically create collections or data sets, right? So you have a data set for each genome, and then now we’ve integrated Gaia agent into SeqHub agent, that can do multi-gene reasoning in a genome that is native to your particular data. So, given a genome that I’ve sequenced from soil. I have high conviction that this soil... this genome can produce a molecule or degrade a molecule. I can ask SeqHub agent, “go through 5,000 genes that I’ve sequenced here, in this particular order that is found in... use all the tools that you have and find me the set of genes that’s gonna be involved in degradation of this particular compound, or synthesis of this particular compound.”

Or “this thing is found in this kind of environment.” So you basically can do reasoning that’s a lot more complex than “what does this protein do?” So that’s something that we’ve implemented for SeqHub. Essentially all of that is just... it’s aligned with our mission and that we wanna help people understand their sequences better, but it’s also to make sure that we can bring in this community of people who really care about their sequences and want to share their knowledge. So the next step for us is to build this community of scientists who will generate this paired information with sequences to either human understanding or experimental data or sample data. We’re just trying to get as much information as possible publicly for sequence to a label that matters in science.

[01:35:37] How will the way science is done change in 10 years?

Abhi: When we last spoke, you mentioned that you think the way that science gets done will look very different in 10 years. What do you think changes?

Yunha: So one idea that I have... I don’t know, like this is changing all the time... but I think there’s been a lot of focus on scientific narrative. So, how you tell the scientific story is really important in science, in the scientific enterprise. So even when it’s like a small finding, you write a whole like narrative...

Abhi: Amplify it.

Yunha: I think... you contextualize it so that it’s impactful and that’s really important. You might find like “this protein does something” and alone that’s just “okay, sure.” But if “this thing does something, then this means that this can do something else and then that means we can use it to do fix this particular problem.” So that’s contextualization of scientific discovery. And that narrative has been really important. And I think almost overemphasized. And I think that’s also... I think that’s not a... maybe in to the extent that I think it’s overdone.

And I think in the future as machines are more involved in scientific discovery, perhaps data is gonna be a lot more important. And how we... I think currently the narrative is more important than the data. Data is just like a zip file, and then people read the narrative and AI agents read the narrative, right? So that is... that’s become really important part of science. But I think as we do more science with the data itself, not with the narrative linking, I think the data sets are gonna be a lot more important. And maybe in the future we’re just gonna be like depositing data and calling that a scientific product, which is not something that’s being done today. And the sort of innovation is in how you generated that data, how meticulous you are, how innovative you are. I don’t think like the human role is gone, but it’s just the data generation is done in a way that’s so sophisticated that it has a big impact on the conclusions that we can draw from that particular data. That is like scientifically salient.

Abhi: Do you think we’re like currently poking at that with the release of Future House’s Kosmos? Like the existing like AI co-scientist stuff... and were you gonna just plug in your data? How much do you... have you used those? How much do you trust them today?

Yunha: Yeah. So I think it goes back to the same question of like human language and narrative, and how much emphasis we wanna put there. I agree that language is a like a very important medium in which we understand things and then link concepts. But overemphasis on narrative and using only agents to like natural language agents... I’m not saying the current agents are like this... the worst case scenario is the AI agents only read and it doesn’t do any data analysis. I’m sure it’s still gonna find something new, right? It just read a lot of papers and then you chat with it and you’re like, “oh, like what does this protein do?” It probably doesn’t... it probably does this.

I think in an ideal world, there’s more emphasis on the data part and the understanding of the data without the sort of biases of language. Whereas the language is how it communicates with humans. So I think we’re not quite there yet in terms of how do we build like scientific systems.

I’m not even gonna call them agents because I think that places too much emphasis on the narrative. But how do we build systems that can conduct science and scientific inquiry that can go beyond like human narrative and human understanding. So that’s... yeah, I don’t know. I still think about it a lot.

Abhi: In some sense, like I almost imagine the natural language agents are like—also like Gaia perhaps, or Gaia agent perhaps—are like somewhat poisoned by the fact that they have read narratives and have like hyper-focused on certain things that perhaps not actually that useful or interesting. When you look at Gaia agent’s reasoning traces, how much do you see this, that it’s like focusing on what you personally would not have focused on?

Yunha: I see. Okay. Yeah. And sometimes that’s a good thing. Sometimes it’s not a good thing. I think, yeah, so I’ve seen cases where Gaia agent just doesn’t focus on what it’s supposed to focus on. And there’s no reason for it, like it’s just doing what it wants to do. And I can’t really... I don’t know if this is something that can be solved with like better prompt engineering, giving it more tools, and how to rescue it going down a path that is just too obvious or too... yeah, like how do you make it more like rebellious against the existing knowledge? I don’t know, because it’s so reliant on what it knows. So I think I’m sure there are like a lot of like agent-based research for how to make agents more, yeah, more creative I guess. So I think there’s like definitely work that can be done.

Abhi: Have you seen that one like Andrej Karpathy tweet about him really desiring some LLM that knows nothing about the world, but is like maximally intelligent and is able to go out and gather information as it needs?

Yunha: Yeah.

Abhi: And I heard that like GPT-OSS was actually like this, it had incredibly low benchmarks on like general world knowledge. But it was really good at math. And it was really good at just like the CodeBench or the software engineering stuff. I’m curious, have you tried GPT-OSS in Gaia agents?

Yunha: Okay. I have not.

That would be pretty interesting.

Yeah.

Abhi: Cool. I think that’s all the questions I have. Thank you so much for coming on.

Yunha: Cool. Thank you.

If you’re at all interested in working together for future episodes, reach out!

1. Introduction

2. Timestamps

3. Transcript

Listen on Spotify/Apple Podcasts/Youtube:

Introduction

This is an interview with Hunter Davis, the CSO and co-founder (alongside Laura Deming) of Until Labs, which you may also know by its prior name, Cradle. They are a biotech startup devoted to organ-scale cryopreservation. They raised a $58M Series A back in September 2025, and are backed by Founders Fund (especially interesting!), Lux Ventures, and others.

In this interview, we mainly talk about the engineering and scientific difficulties in the cryopreservation field, including some background details on their September 2024 progress report on neural slice rewarming, how they characterize tissue damage in their attempts to do kidney cryopreservation, the potential economics of future cryopreservation protocols, and lots more.

One of the most interesting conversations I’ve had in a long time. If any of this work seems interesting, Until Labs is actively and aggressively hiring!

Enjoy!

Timestamps

[00:01:50] Introduction

[00:05:00] Why don’t we have reversible cryopreservation today?

[00:07:05] Why is freezing necessary at all for preservation?

[00:08:23] Let’s discuss cryoprotectant agents

[00:14:09] Until Lab’s 2024 progress report on neural tissue cryopreservation

[00:20:28] How do you measure cryopreserved tissue damage?

[00:22:34] Translation across species

[00:26:04] Why was the cryopreservation storage time so short in the progress report?

[00:30:47] Nuances of loading cryoprotectants into tissue

[00:37:03] Let’s discuss rewarming

[00:43:02] What scientific problems amongst vitrification and rewarming keep you up at night?

[00:45:58] Why are there so few cryoprotectants?

[00:48:11] How can you improve rewarming capabilities?

[00:53:03] What are the experimental costs of running cryopreservation studies?

[00:57:49] What happens to the cryoprotectants and iron oxide nanoparticles after the organ has been thawed?

[01:01:34] Cryopreservation and immune response

[01:03:25] How do you filter through the cryopreservation literature

[01:05:54] How much is molecular simulation used at Until Labs?

[01:10:04] What are the (expected) economics of Until Labs?

[01:14:49] How much does cryopreservation practically solve the organ shortage problem?

[01:17:04] Synergy between xenotransplantation and cryopreservation

[01:21:12] How much will the final cryopreservation protocol likely cost?

[01:21:58] Who ends up paying for this?

[01:23:28] What was it like to raise a Series A on such an unorthodox thesis?

[01:27:49] What are common misconceptions people have about cryopreservation?

[01:29:58] The beginnings of Until Labs

[01:34:07] What expertise is hardest to recruit for?

[01:39:27] What personality type do you most value when hiring?

[01:44:17] Why work in cryopreservation as opposed to anything else?

[01:46:26] Until Lab’s competitors

[01:49:30] What would an alternative universe version of Hunter worked on?

[01:51:33] What would you do with $100M?

Transcript

[00:01:50] Introduction

Abhi: Today I’ll be talking to Hunter Davis, the Chief Scientific Officer of Until Labs, previously known as Cradle, which is a startup working to build reversible cryopreservation for use in organ transport, and eventually medical hibernation. Prior to starting a company, Hunter received a PhD in physical chemistry from Caltech and got a postdoc in neuroscience at Harvard. Today we’ll be talking about the recent progress report that Until Labs put out, what cryopreservation problems keep Hunter up at night, and a lot more. Thank you for coming onto the podcast, Hunter.

Hunter: Thanks for having me.

Abhi: So just to set the stage, I’d love if you could give me a very general overview of what cryopreservation exactly is and the primary roadblocks to outright solving the problem.

Hunter: Sure. If we look back to the origins of using temperature as a knob, you can go all the way back to 1776. At this point it was observed that sperm under a microscope could be cooled down to hypothermic temperatures and then motility would decrease. And then when it was rewarmed, it would come back. And this led to some hypothesizing about what might be causing this. Fast forward to around 1950, and people discovered that not only could you cool things to hypothermic temperatures, you could add cryoprotectants into these molecules and you could cool them entirely. Go down to the point that you completely arrested molecular motion. Fast forward again to more contemporary processes. And what you’re trying to do is slow down molecular motion inside of cells in a variety of different contexts. So one that you might envision is for a hypothermic use case. You have a patient who needs a surgery. In many of these cases, the ischemic time is far too short to complete the surgery at 37 degrees.

One example of this is aortic arch surgery. This is a heart surgery that can’t be bypassed and by default what would happen is the brain would go through ischemia because you can’t actively perfuse oxygenated solution into the brain during the surgery. So what the surgeon will do is they’ll take the patient and they’ll cool them down to around 15 degrees Celsius, and then they can complete the heart surgery and then rewarm the patient. And it shows no long-term neurological side effects. Then you can look at something like an organ. Here we might want to slow down the metabolism of an organ during transport to be able to increase its viability window. Here we can take the organ and literally just put it on ice. And what this does is it just decreases the rate of all these metabolic processes that are happening inside of the organ or in the case of the patient, inside of the body. Just reduces all of those by reducing the temperature.

. You can think of it as between zero and minus 130 is a danger zone where ice can form; below minus 130 is safe perpetually.

So what does the process look like in practice? You could take an organ, you load it up with some cryoprotectant molecules that reduce the rate of ice formation during this danger zone. Then you rapidly cool your sample from zero to below minus 130, store it there, rewarm, and then unload the cryoprotectant from the organ or tissue. And then you can bring it back up to 37 degrees and biological function will resume.

[00:05:00] Why don’t we have reversible cryopreservation today?

Abhi: What, why don’t we have this today?

Hunter: Yeah. There’s a lot of challenges with scaling this up. We do have it today in really simple systems. One example of this would be cryopreservation of embryos for in vitro fertilization. Here you have something that’s very small. Normally they’re operating on embryos at this stage where it’s four to six cells. And here you can take the embryo, dip it into DMSO, dimethyl sulfoxide, and then quickly dip it into liquid nitrogen. This rapidly cools it through this danger zone and embryos can be stored for decades in this state and then rewarmed and implanted and using the contemporary methods, the viability of embryos that are going through this process is a little over 95%. So we have seen that the live birth rate that occurs after vitrified embryos has started to exceed those of fresh implantation. And the reason for this is the allowance for additional genetic testing while the embryo is cryopreserved.

So this does exist for really simple systems. There’s also been some proofs of concept of bringing it up to more complex systems, like a rat kidney. The John Bischof lab at the University of Minnesota has shown that you can take a full rodent kidney, vitrify it, bring it down below this minus 130 degree state, bring it up, reimplant it into the rodent, and then it’ll support life for this rat. The challenge is that as you scale up, as you try to go up to something that’s the size of a human kidney, all the thermal transport becomes much more complicated. If you try to imagine cooling something that’s the size of an embryo, I can go really fast. If I do something that’s the size of a human kidney, which is around 150 grams, that’s going to be much slower. So then in this competition between you getting down to the safe zone, minus 130, and the rate of continuous ice formation in that zero to minus 130 range, you just start to lose out because you can’t cool fast enough. And similarly, rewarming becomes a challenge.

[00:07:05] Why is freezing necessary at all for preservation?

Abhi: When you mentioned earlier about how you can’t just perfuse a transplant organ with oxygenated blood to prevent ischemia, intuitively, why is that the case? Why do you need to freeze to preserve the cellular state of the organ and why can’t you just continuously pump oxygenated blood through?

Hunter: Yes, that’s a great question. There are technologies that exist on the market right now that will perfuse hyper-oxygenated fluid. And they do this in either a cold perfusion context traditionally. There’s a bunch of these products that exist on the market that can increase the viability of the organs that will then be donated. But all of them still have time... organs still time out.

At some point.

Abhi: I guess what are the specific reactions?

Hunter: What are the mechanisms that can account for here?

So one thing is that now the organ is in isolation. So let’s think of a kidney in isolation. Now I don’t have a liver. So any toxins that need to be cleared by the liver, the circuit isn’t going to be able to take care of this. We don’t have dialysis for a liver, for example. So I think that part of this is that you have taken the organ out of the context of being in a multi-organ system that is responsible for clearing a bunch of the toxins that are generated from the metabolism of any one organ outside of the concert of the other.

[00:08:23] Let’s discuss cryoprotectant agents

Abhi: That makes sense. And moving a little bit onto the cryopreservation step. First you have this vitrification step, second you have the rewarming step. Focusing on the vitrification part. You mentioned that you physically stop the water molecules from interlocking with one another and one way you can do this is by freezing really fast, such that they don’t have time to lock into one another. Another way is adding in cryoprotectant agents, which prevent the water molecules from doing that. What are cryoprotectants doing beyond that very simple mental model?

Hunter: Yeah. Yeah. That’s great. I think there’s a couple things to keep in mind here. Most cryoprotectants that you’re going to add are not going to completely prevent the formation of ice in perpetuity. You can think of it as reducing a reaction rate. If you think of the liquid to solid water as a first order rate equation. And this is the there’s this thing called classical nucleation theory. It predicts the rate of conversion from liquid water to solid ice.

Cryoprotectants increase the activation energy of that liquid to solid transition. There’s still a very molecular question of how do they do that?

There’s a couple of different mechanisms of action here. So one that you can look at is direct hydrogen bonding of water molecules. You can imagine the rotational tumbling that’s required for water molecules to align into hexagonal ice. If I can just slow that down by hydrogen bonding to the water, then I’ll reduce the rate of that alignment and buy myself more time to be able to cool down below minus 130. Other things is just generally increasing the viscosity of the solution. If you look at the Stokes-Einstein model for diffusion in a fluid, it’s obviously inversely proportional to the viscosity of the fluid. So these cryoprotectant molecules tend to directly hydrogen bond water and be viscous.

There’s also some interesting alternative mechanisms of action that you can explore. Looking to nature, for example, there are these things called antifreeze proteins. So instead of directly hydrogen bonding to liquid water, what they do is they preferentially bind to solid-phase water. So they allow some ice nuclei to form, and then by cooperatively hydrogen bonding, these macromolecules will stick to the ice surface and prevent it from extending. So there’s a few different mechanisms of action, and you could think to exploit all of them together in a cocktail.

Abhi: And I imagine the type of cryoprotectant agent that Until is most concerned with is the former category of directly binding to hydrogen and you don’t care too much about the antifreeze proteins.

Hunter: Yeah, I think that we’re open to exploring all of these mechanisms of action. I think in the end, when you want to do vitrification, the thing that’s going to have to do the dominant work is going to be these colligative agents. These are the small molecules that are actually directly interacting with liquid water. If you think about the antifreeze proteins that nature uses, most of the time, always, these organisms don’t care about surviving down to minus 130. What they care about is surviving in equilibrium at minus five degrees. These are very different processes. To try to prevent the extension of ice in a supercooled state, just below zero, is just a fundamentally different problem than trying to survive all the way through down to minus 130 and back. So yeah, the primary focus is on these colligative agents.

Abhi: When you look at the cryopreservation literature, has there been any evidence to suggest that using antifreeze proteins and being okay with just mildly arrested, or mild cryopreservation, is good enough in some cases or there hasn’t been too much work in that direction?

Hunter: Yeah, so not specific to antifreeze proteins though. I think that there are some applications here where people have been trying to use either things inspired by antifreeze proteins is mostly it. But there is a whole stack of products that will hit the market that are these supercooled organ solutions. I think of this as an extension of hypothermia where you can go instead of to four, now you can go to minus four. And there’s a few different tricks that people have used here. All of them have their trade-offs. But I would imagine that similar to how Until will be bringing a vitrification product to market, I would imagine there will be some near-subzero storage products as well that will have their own trade-offs. And in the end, I think the thing that’s going to matter for what wins out and becomes the dominant way of transporting organs is going to be the thing that gives the patients best outcomes.

Abhi: Yeah, that makes sense. Amongst the cryoprotectant agents that you guys are actively developing, how much work goes into improving those agents versus improving the next step, which is rewarming?

Hunter: Yeah. So we view this as a very multifaceted problem, and I think that one of the things that originally fascinated me about it is that it’s one of the few scientific problems that I’ve seen before that brings together things from applied physics, chemistry, transplant, biology, all the way to hardcore electrical engineering and power electronics. And we work on all of these things simultaneously. I view them as relaxing each other’s constraints. If you have devices that are good at rewarming quickly, then my molecular agent doesn’t need to be quite as performative and vice versa, right? If I have the killer molecular agent, then maybe I need a very facile, easy device. In the end I think these constraints are pretty hardened, so it’s going to be some combination of these solutions in the middle. But yeah, we work pretty intensively on both improving the cryoprotectant agents and on improving the devices that do the cooling and rewarming.

[00:14:09] Until Lab’s 2024 progress report on neural tissue cryopreservation

Abhi: Okay. I’m going to wrap back to these questions about cryoprotectants and rewarming in a bit. But what I really wanted to talk about, what actually started this conversation to begin with, was you guys released a progress report in September 2024, alongside your series A announcement, that described how you recovered electrical activity from a slice of mice cerebellum, I think, that were frozen, rewarmed, and you observed some level of electrical activity in the neurons there. This is obviously incredibly impressive work. But you did mention in the paper, in the progress report, about how there is more to neural functionality than simply recovering neural activity. What else is there?

Hunter: Yeah, so I think that, first of all, I think it’s maybe worthwhile to go into why we did that as our first POC. When Lauren and I met and we were talking about the idea of reversing, reversibly pausing biological function eventually for a whole organism, we had the initial conversation of what are the falsifiable hypotheses here? What are the experiments that we could do that would prove that this will not work? I think one of the first ones that we wanted to do was look at these very delicate pieces of tissue, being neural tissue. And so the easiest thing that we could come up with was doing these cerebellar slices. It’d be very easy to load these diffusely. We didn’t need to do any perfusion. You could cool and rewarm them very quickly. And the cerebellum is known for having very periodic firing. So you would be able to know if you got, at least on the single cell level, you would be able to tell if you were getting any action potentials that made sense. And so we performed that experiment and like you were saying, we saw the recovery of some electrical activity of action potentials in the slice.

But yeah, there’s the question of what comes next? Actually, a piece of very impressive work that developed along this axis came out of Alex German’s lab in Germany this year. And he was able to show using a very similar protocol that you can recover long-term potentiation from these acute slices. So this is... you take one canonical example would be like take the Schaffer collateral in the hippocampus. Here there’s a bundle of axons that are all synapsed in a very similar location in the hippocampus, and you look for potentiation of those synapses. This is like a bit flip for memory. And what he was able to show was that you can recover LTP in these slices. So this is all really interesting and I think it’s useful as a micro-circuit level problem, but if you were to talk about what does it take to preserve full neural function? This is a much more complex question than the micro-circuit inside of a slice. And eventually you have to go to the brain as a whole and to the organism as a whole. But this is a very deep challenge that I think will take quite a bit of time to get to the point that you can make traction against it, is my presumption.

Abhi: Yeah. One pretty shocking thing I found after we had our initial conversation a while back, is that the ultimate goal, or at least the short-term goal, of Until Labs right now is not whole brain preservation. Right now it is, I think you said kidney preservation. Why switch to kidney? When I read the report, I thought, okay, this is one step to whole brain cryopreservation.

Hunter: Yep.

Abhi: Why move to a different organ?

Hunter: Yeah. So I think the other thing that we have been interested in as the moonshot from the beginning is reversible cryopreservation of an entire patient. And this would involve allowing someone to pause their biological time as a whole. There are some really useful things about going through the process of doing this on an organ-by-organ basis. First one, this allows us to get therapies to patients in need in a very immediate way. You can deliver care to people who need it and use cryopreservation to do it. There’s also a natural scientific roadmap that’s built out of this, where you can start to learn technologies on isolated test beds where you can learn about how to preserve kidney, heart, lung, these kinds of things. And we view this as a natural foundational platform on which we can build towards our eventual goal of being able to do hibernation of an entire person.

Abhi: Why choose kidney over anything else?

Hunter: Yeah. So we’re actually pretty organ-agnostic.

Abhi: Okay.

Hunter: I think that kidney is the one that’s commonly talked about and I’m happy to discuss nephrology as a particular application. But I think one thing that’s nice about vitrification is because it leverages physics instead of going after very specific biomolecular pathways, it’s somewhat organ-agnostic. Okay. Which is, I think, exceptional compared to some other chemical strategies for preservation.

Abhi: Yeah. Returning back to the progress report, just because I imagine a lot of the questions, at least external people may have about cryopreservation, are probably answered or at least raised by the report. In the report, you did cryopreservation across four mouse samples and found, I think the primary results were over one mouse, but I think there were discussions over the other three as well. But I did want to ask how much heterogeneity was there in your success in being able to recover electrical activity from all four rats?

Hunter: Yeah. We were able to see electrical activity recovered from all four rats. There was a large degree of heterogeneity in the amount of electrical activity that we recovered. The traces that were placed into the report, I think are representative of the group. But particularly in that iteration of the device, I think we were very early and I definitely think that our QC was not as dialed. So there was quite a bit of heterogeneity in things like the cooling rates and the rewarming rates that were coming out of this little cartridge that we had.

Abhi: Why? Would you imagine the primary axis of variability... how much of it is just an experimental batch effect versus some rats are perhaps better at being cooled than others?

Hunter: I wish that I could tell you that it was down to the rats. I’m pretty sure at this point, at least particularly at the point that we were filing the report, I would chalk it up to experimental variability.

[00:20:28] How do you measure cryopreserved tissue damage?

Abhi: Okay. That makes sense. Returning back to the kidney cooling and rewarming moonshot, what is your metric for... you freeze a whole kidney, you rewarm it back up. How do you tell whether this kidney is good?

Hunter: That’s a great question and this is something that actually I didn’t appreciate until we started working on it, which is that there’s this whole literature and an entire field of organ evaluation in isolation, outside of a body. These are normothermic machine perfusion. So this is for the electronics people. This is like a test bench for your organ. You can hook up fluidic circuits into, if you imagine a kidney has one inlet and two outlets.

It’s got its arterial inlet and it’s got a ureter and a venous outlet. I can press certain fluids into the arterial inlet and then measure the fluids that are coming out of both the ureter and of the venous effluent side.

This allows you to measure things like what is the uptake of glucose in the organ. You can measure things like what is the lactate concentration in the stuff that is coming out of the venous side. So there’s a whole host of these biomarkers that have been established by the transplant community and I think that one marker of good cryopreservation work, as you guys are looking through the literature, is how well do they reproduce the metrics that have been established and known to work as predictors of transplant outcomes, which there’s a whole literature on this.

Abhi: How... are these metrics for kidney damages established by the transplant community? Are they fully dialed in? This is as good as it gets. Or there’s still room to improve there.

Hunter: I would argue that there’s still room to improve. It’s still an active area of research, improving the correlation of these NMP assays to transplant outcomes. I think it’s established that it is possible to correlate these things. I think that no one would claim that we are done

Abhi: Okay.

Hunter: With how to draw the best conclusions.

[00:22:34] Translation across species

Abhi: And right now are you still doing mice kidneys, or are you moved on to other species?

Hunter: I think you need to move up to pig as the standard preclinical model, and it’s for the reason of the size. So many of these constraints have to do with matching the size of the organs that you’re interested in.

Abhi: At what point... maybe this is an unanswered question in the transplant field, but how much do what you learn from pig kidneys transfer to human kidneys?

Hunter: Yeah, so this is a classic question of the translatability of any of these assays. And I think that particularly for something like vitrification, we’re going to have to see and we’re going to have to be intelligent with our trial design. I think that there are ways to access things like human organs that would otherwise be discarded to check for some of these questions. But this is a standard translatability question and particularly because we’re going to be pressing vitrification through for the first time, I think that we’re going to get to learn the answer to that question.

Abhi: At what point in time... are you still right now operating on... I’ll just ask the question. Slices of kidneys or are you doing the full kidney at any given run?

Hunter: I think you need to be doing the full kidney.

Abhi: Gotcha.

Hunter: And the reason for this... so there’s ways to chop this up and do these things on tissues, but in the end, vasculature is a really critical part of this process. I think this is probably something that I should have gone through in more detail as I was describing the original protocol, but the only reason this is possible is because of vasculature.

You need to be able to do mass transport of this cryoprotectant deep into tissue. The heat diffusion that I was complaining about previously because this kidney is 150 grams, the same logic would apply for loading the cryoprotectant if you weren’t to use vasculature. If I were to try to load it from the outside, it would take forever. But we are perfusing through the vasculature. So I think a lot of the development that needs doing is on how do you efficiently transport cryoprotectant agents into a kidney or another organ.

Abhi: How much of the... where do most of the ideas in the room usually come from? Are they from the nephrologists? Are they from the electrical engineers? Are they from physics people? What plays into pushing forward the kidney preservation goal?

Hunter: Yeah. The only reason that we’re able to make any progress is that the answer is all of them. Okay. I could give you... there are examples of times where I think really great ideas have come from unexpected places. So for example, Andrew Ted, who’s our head of applied physics was previously running battery materials research at Tesla.

Abhi: Interesting.

Hunter: And has a very interesting perspective on material discovery, which is critical to what we’re working on. Yeah. And then there’s stuff that comes from Gerald Brander who just joined as our Chief Medical Officer who has decades of experience in transplant. And he has obviously his own lens on the way that these things need doing inside of organs. So I think that the magic that I have seen always happens when these people get in a room and they have conversations and they can relax each other’s constraints. You have this specter of another field where it’s a black box and it’s oh, this seems quite challenging or hard because you have some perhaps overestimation of the level of constraint that collaborator has.

We just get these people in a room and they have a conversation. I think that’s where you can unlock a lot of upgrades to protocol.

[00:26:04] Why was the cryopreservation storage time so short in the progress report?

Abhi: How much... returning back to the report? You said you kept the neural tissue at negative 196 degrees centigrade for about a minute. I imagine you’re going to do the exact same for kidneys as well. Why not extend that out to a day, a month, a year? Why choose a minute? Was it just you don’t expect anything to change after that point or...

Hunter: Yeah, so that’s a great question. For the tissue slices, it was a very simple thing of just, this is the way that we built the device. I literally cannot express to you how quickly this device was thrown together to be able to do this report. And it took a lot of optimization on the protocol side. But literally we had this idea for this device and then threw it together mostly with thinking towards screening, not towards, Hey, we need to go get this milestone. And be able to report it out. Mostly to try to screen cryoprotectants. But for the kidney, I imagine that we’ll store it for longer. There’s a natural question that I think is at the core of what you’re getting at though, which is why does it matter or not matter?

Abhi: Yeah.

Hunter: And I would contend that it actually doesn’t matter.

Abhi: Okay.

Hunter: You can store it down there to be able to prove out the theory. But the reality is, because water is 13 log units more viscous down in this vitrified state than it is at room temperature. There’s a massive time dilation. Think seconds at room temp to millions of years in vitrified state for equivalent diffusion distances. So for any reasonable timetable, think days, months, years, doesn’t really matter because nothing’s going to have moved in this glass state.

Abhi: Yeah. One instinctive thought I would have is the reason you may care about a day or a month is... when you’re freezing something, you remove all of the blood from it and replace it with cryoprotectant. At least that’s my mental model of it. The reason you may care about testing it out for a day is what if you left a little bit of blood in there? And that blood will continue to gather, accrue damage over time. Is that at all...

Hunter: I don’t think so. So actually if you got... so let’s say I left some random red blood cells inside of the vasculature. It’s what’s going to happen? first of all, the water inside of those red blood cells is still going to exchange. Sure. So they would still equilibrate with all the cryoprotectant that you’re perfusing in there. And also, let’s say that maybe there’s a water pocket inside of this. if I got any ice on the way down and let’s say I induced damage as a result of that ...that damage is now there.

Abhi: Okay.

Hunter: it’s... there’s no period of time. Because even if I have ice that is nucleated it’s still not extending below minus 130. It really is the case that all the damage is getting done in this minus 130 to zero range.

That definitely doesn’t mean that you’re okay. if you have nucleated ice on the way down, there’s actually an annoying/interesting asymmetry in the physics here where you tend to form more nuclei as you’re cooling and then you tend to extend them while you’re warming. And that extension is really what kills the tissue because that’s where you start to tear through quite a bit of cells. yeah. And tear things up.

Abhi: One question I initially had when we first had this conversation was, let’s say centuries from now, we’re all cryo-frozen on a ship going through the stars. How much do we need to worry about background radiation of the universe still affecting our genomes?

Hunter: Yeah. I can’t prevent the DNA from getting nicked from radiation as a result of it being vitrified. But I think that at the time that we have all of the technology available for vitrification and deep space exploration, I would sincerely hope that we could figure out how to line the ship with lead or something.

Abhi: Yeah. Yeah. Yeah. That makes sense. Returning back to the question of metrics of damage you’re looking for when you are cryopreserving an organ. For the brain, it’s neural activity, it’s LTP. For kidneys, it’s this platform you just talked about. Do you think you’ll have to custom make this for every single organ or pretty much every organ in the human body has some sort of well-established protocol as to how you assess damage?

Hunter: I think that at least the donor organs, Okay. which are the first ones that we’ll be interested in, Yeah, I think there’s pretty canonical ways of evaluating their function. I think the things that we have to build out bespoke are things that are not so much for proving out a milestone of, is this thing healthy or not? There we can always reference back to the transplant community.

Abhi: Yeah.

[00:30:47] Nuances of loading cryoprotectants into tissue

Hunter: I think the thing where we’re both working on our own and also looking to the cryobiology community as well, is how do you assess things that are very specific to the cryo process? There are damage mechanisms that you want to be looking for and optimizing against that are specific to vitrification and cryoprotectant loading.

Abhi: What’s an example?

Hunter: So an example here might be figuring out the time constant that you can load the cryoprotectant in. So obviously you would like to get the cryoprotectant in there very quickly.

You don’t want to put it in too quickly.

Abhi: Why not too quickly?

Hunter: Yeah. That’s a great question. So I want you to, let’s do a liberally simple thought experiment first. Okay. So I have a cell in a test tube, by itself, just a cell. And to start it is in a solution that is isotonic with the interior of the cell. So water is exchanging into the cell just as quickly as it’s exchanging out of the cell. A key part about cell membranes is they’re really good at exchanging water. So there’s a protein that is in the membrane called aquaporin. And its entire job is being just a water-specific transporter. Just water freely diffuses through it in and out. So now what happens is, if I take, let’s say half the water out of the extracellular space, and I replace it with cryoprotectants.

Even very permeable cryoprotectants are still going to be, 100x slower at diffusing across the cell membrane compared to water. So what invariably happens is initially you have an influx of water... that water runs out of the cell. And then cryoprotectant can slowly get into the cell.

So the cell immediately shrinks initially and then starts to re-expand. if I shrink too much, then I die of osmotic shock. And similarly on unloading, and unloading actually is an even stiffer constraint. If it’s loaded with cryoprotectant and then I add a bunch of water on the extracellular space, then the cell inflates like a balloon and then pops. So there’s that constraint on not loading or unloading too quickly. You can imagine extending this very simple thought experiment of a single cell up to that first layer of the vasculature, that first endothelial layer of the vasculature. It’s seeing that maximum osmotic shock as you’re trying to increase the cryoprotectant. So when you’re loading this, you don’t just go from zero, isotonic solution, to full-blown cryoprotection. The protocol looks like a ramp. Gotcha. Where you’re linearly increasing the cryoprotectant as a function of time.

Abhi: Is there established theory as to what the slope of this line should look like or is it empirically determined?

Hunter: Yeah, so you can calculate what it should be. The thing that is challenging is that you don’t necessarily know all the transport metrics that you need to be able to establish. But I think it’s still useful to go through the first principles of how you would think about this. And it’s a balancing between the relative permeability of the cryoprotectant compared to the water. And then you also need some transport model of how is this cryoprotectant getting through the vasculature and perfusing along the flow path.

If you have these together, then you can establish a rough heuristic of what it should look like. Things that make this really challenging are that, yeah, not having the permeability coefficients for all the cryoprotectants in formulation for the cell types that it’s going to be seeing. That really ends up mattering. The other thing is that for some of these cryoprotectants, the permeability is actually a function of the cryoprotectant concentration, and then it’s just incredibly hard to model. So I think transport... one of the interesting biophysics questions that’s left in cryopreservation is good transport models for the cryoprotectant into the tissue.

Abhi: How in practice, how much do you use models versus just empirically determine that?

Hunter: Yeah, so I think that you want to use both, definitely don’t want to just write it down on pen and paper. Definitely the biologists in the company would prefer that I used a whiteboard marker less than I do. But I think that oftentimes you can make these things sing together. So for example, we really prioritize making simple assays, not just for how well are we doing, but also to try to establish these metrics, try to build models because you can’t proceed with, for example, how do I know what slope to load the organ? We’re pretty committed to having a first principles, let’s say hypothesis that’s built off of some simple cellular model. Some simple reduced experiment that can be done at high throughput. And then we make a best guess and adjust from there. But it always requires adjustment in the context of an organ because it’s so complex.

Abhi: How much do different cell types matter? Are all human cells, let’s say, all happy with the exact same slope or are some cell types especially sensitive to osmotic pressure and they will almost always pop if you...

Hunter: Yeah. Yeah. I think so there’s... talking like a physicist, I would say, to first order, you can consider them as all the same. And then there’s corrections. And one way of thinking about this is it is the case that aquaporin is the dominant way that water gets into and out of all cells.

So that story that I told you, that’s true across all the cells.

Abhi: Yeah.

Hunter: But the specific permeability of a given cryoprotectant is going to be different for different cell types because, for example, they use things like... they tend to hijack things like urea transporters to get into the cell. Obviously the density of urea transporters is going to be different in different cell types. So the relative permeability of these cryoprotectants in different cell types is going to be different.

Abhi: Yeah. How much do... if you move beyond organs and you start to consider, I’m trying to think of something that isn’t an organ... like blood vessels. For example. There probably are blood vessel transplants.

Hunter: Yeah.

Abhi: Transplantation. But it doesn’t sound super trivial to just perfuse blood through and see what pops out the other side. It’s more of a morphology question. Are there good metrics there? And if not, do you have to generally create your own?

Hunter: Yeah. So fortunately, again, if you want to do... so vessel cryopreservation is its own field.

Abhi: Oh, okay.

Hunter: It exists. Yeah. And there are things like H&E staining is really useful. This is also a place where you can lean on the medical field and look at established protocols for histology. And look to what is the inter-lumen supposed to look like? There’s a bunch of these structural assays that you have to lean on. Because you’re right. You can’t do the same thing of asking the question of, is the organ functional? Yeah. ‘Cause it’s just a pipe, so it’s not going to give you any information.

[00:37:03] Let’s discuss rewarming

Abhi: Yeah. We’ve spent a long time talking about the freezing process itself. Moving on to the rewarming side. What does that actually look like in practice?

Hunter: Yeah. So I think that, first let’s start with why do you need to rewarm quickly? We already talked about this, but we want to outcompete the formation of ice on the way up. And this is particularly challenging because ice nuclei that have been formed on the way down but are very small, are going to extend really quickly during the rewarming. So this placed a huge pressure on the field to create methods that would homogeneously and rapidly rewarm the tissue. So you can envision a few different ways that you could try to rewarm something that’s the size of a kidney. Maybe the most naive one is I just take it and I’ll just put it in warm solution. Here again, we have the problem of a heat diffusion equation. Everything tries to diffuse in from the outside. It’s going to be way too slow. Maybe you’d be able to warm the surface really quickly, but the core is going to stay cold. Similarly, maybe you could think of, okay, throw it in a microwave. But there again, you get cold and hot spots. If you’ve ever tried to rewarm something that’s very extended in your microwave, the surface gets warm, but the core does not.

So one of the innovations that came out of the Bischof lab at the University of Minnesota was they figured out that you can use biocompatible iron oxide nanoparticles that are perfused into the vasculature of the kidney. Fill it with metal and then put the entire kidney into an alternating magnetic field. And what this does is it rewarms the kidney, somewhat homogeneously, much like an induction heater in your kitchen. You’re flipping the magnetic dipoles back and forth. That generates some heat that tends to heat the organ relatively homogeneously. As it would happen, this was actually the topic of the second half of my doctorate. Not for the applications of organ rewarming, specifically. I was studying it in the context of cancer hyperthermia. It was the application I was looking at. But it’s a full circle moment for me that we’ve circled back to this. So this is one thing you could think to use is alternating magnetic field heating. This is one thing that we’re studying at Until. We’re also studying some other methods for volumetric rewarming. But this is the canonical one that the field is settling in on, is this idea of using alternating magnetic fields in combination with magnetic nanoparticles.

Abhi: Just to set the order of operations. The cryoprotectant agent and iron oxide nanoparticles are both given at the same time.

Hunter: So yes, you think of it as if we have this ramp of cryoprotectant that has to go into the organ, imagine the last part of that ramp, you also dope in the iron oxide nanoparticles into your solution. So it’s a colloidal suspension of iron oxide nanoparticles inside of cryoprotectant.

Abhi: Is there no interaction between the iron oxide and how the cryoprotectant actually works?

Hunter: That’s a great question. Iron oxide nanoparticles can be ice nucleating agents if they’re not coated properly. So the surface chemistry here becomes really interesting ‘cause you need things that are colloidally stable in cryoprotectants, which tend to be very high salt and kind of messy. And then you... I should say, high concentration of non-water substances. And you also need something that’s not going to nucleate ice. You don’t want to create surfaces that tend to allow for this nucleation process to occur.

Abhi: That makes sense.

You mentioned earlier, and this is something I hadn’t naively thought of, was as you freeze, I can vaguely understand water molecules interlocking with one another to create these ice crystals. How come during rewarming, there’s also a chance for nucleation to happen? It just feels...

Hunter: Yeah, it’s deeply counterintuitive, right? Yeah. But you get more ice that forms on rewarming. We can maybe walk through why.

Abhi: Sure.

Hunter: Basically between zero and minus 130, the chemical potential energy of ice, solid water, is lower than liquid water. This is true in that temperature range, whether you’re warming through that range or cooling through it. But there’s something that’s this asymmetry that we’ve been getting at. I guess we can go into a little more detail. What ends up happening is as you’re cooling, the temperature at which you get maximum nucleation, that’s the formation of these tiny nuclei, is actually colder than the maximum rate of extension of nuclei.

So it ends up happening is on the way down, you go through the temperature zone of maximum extension, then you get to the temperature of maximum nucleation.

So you go through extension, but there’s nothing to extend. And then you nucleate. On the way up, I nucleate a little bit more, and then I extend everything that I’ve nucleated on the way down and on the way up. So there’s this constraint where because you go through on the way up the nucleation and then this maximum extension phase, which is relatively warmer, you produce the majority of the volume of ice on rewarming, assuming you do symmetric cooling and warming rates.

Abhi: Okay. So if you have no ice crystals in your solution and you rewarm, there’s no chance for further crystals to form.

Hunter: Oh no, you can. Sorry. Just to clarify, the, if you think of it as the rate that you’re warming or cooling is separate from the question of just is the equilibrium of water, ice, or liquid at a given temperature, at a given absolute temperature. Not at a dT/dt, not at a change. Just think about in terms of the physics here, think of it as if there’s a water molecule that is at minus 80 degrees, it doesn’t know or care if you are currently cooling it or warming it.

So it is equally likely to do nucleation or do extension during the cooling or rewarming phase. You can think of just running a path integral over this entire curve to get the amount of ice that is formed.

[00:43:01] What scientific problems amongst vitrification and rewarming keep you up at night?

Abhi: Combining these two areas of... it seems two very difficult problems to solve. Building very good cryoprotectant agents and also building very good ways to rewarm a frozen tissue or organ. Which of those two axes do you consider hardest, easiest, or what problems in both keep you up at night?

Hunter: Yeah, both keep me up at night. I think that one thing that is nice about the rewarming system is we can bring to bear some pretty strong power electronics expertise that we have brought into the company to be able to build out a few different technologies there. And I think that we’re pretty committed to building out a wide technology platform. So we look at a few different mechanisms of action here. On the molecular side, I think there’s some give and take here because the reality is that we’ve been using... we, the field, have been using the same cryoprotectants... same few dozen cryoprotectants have been used since the 1950s. Like glycerol and then DMSO were quickly discovered. A lot of people still use glycerol and DMSO.

Abhi: Yeah.

Hunter: There’s obviously a very large chemical space that you could search for cryoprotectants.

And I think that you can find things that work better, but the question is how much better and how much less toxic can you get? And so because these things relax each other’s constraints, I don’t think of one as necessarily being harder than the other because if I were able to get traction on it and make progress on it, then I would just demand more of it.

Abhi: Sure.

Hunter: Yeah. And that actually builds a really cool team environment. Practically it builds a very cool team environment because there is no... for our applied physicists who are thinking about these molecular interactions, there is no “we’re done.” It just... you just keep chasing

Abhi: Yeah.

Hunter: continuously better versions of this. For the people who work on our cooling systems, there is no “we’re done” with improving the homogeneity and rate of cooling. I think this is true of the field writ large. This is why I think this is a relatively... yeah. This is why this is a relatively interdisciplinary, multidisciplinary field as a whole, is that everybody can find their angle through which their expertise can help with cryobiology.

Abhi: Yeah. With the development of better cryoprotectants. I think the one that you used in the September 2024 report was a well-established one with one particular ingredient removed. Yep. I’m curious about the rationale. It was just... you just wanted to get the report out of the way and there wasn’t that much work on fiddling around with the cryoprotectants. Yep. yeah.

Hunter: Yeah. So the report actually came out before we had a molecular development team.

Abhi: Okay. Okay.

Hunter: So we now have a molecular development team that looks at novel cryoprotectants. So we were working off of a sheet of stuff that had been previously used.

[00:45:58] Why are there so few cryoprotectants?

Abhi: When you say that only a handful of cryoprotectant agents have been developed since the 1950s, is that because finding better ones has been too difficult or more because it’s been pretty high-hanging fruit that people haven’t really been attracted to solving?

Hunter: You know, it is is a little bit complicated for me to understand the relative social milieu that has driven this, because I think that there actually has been quite good work on screening new cryoprotectant agents. I’ve been thinking about people like the Toner lab, the Higgins lab. There are labs that have been really focused on how to screen better cryoprotectant agents.

Abhi: Yeah.

Hunter: And yet everybody’s still using ethylene glycol, propylene glycol, formamide, DMSO. These are... there are various combinations of these. And so you’ll see like VS3, VMP, all these things thrown around in the literature. But fundamentally we’re just mixing around a few different agents and occasionally there’ll be a new one that will come in or a new additive that will come in. Some of them are quite toxic. But yeah, it is surprising to me because there has been really quality academic work on screening these things.

And yet people are still using the same stuff. So it’s an interesting contradiction.

Abhi: Is it potentially because there’s no... there’s not that many groups focused on actively translating this to a full actual organ?

Hunter: I think it’s a big lift to do this in whole organs. There are examples of people doing the full stack. So there’s great work coming out of a few different labs to try to scale this up. I think that John Bischof’s lab at UMN would be a good example of this. The Toner and Tessier labs at Harvard would be another good example.

Abhi: On the flip side of how do you improve the rewarming process? Is iron oxide like... will we ever get better than that?

Hunter: Yeah. I mean there are examples of... or is iron oxide it, meaning is magnetic warming it or...

[00:48:11] How can you improve rewarming capabilities?

Abhi: yeah. Is magnetic warming it, even where... even the agent as to how you do the magnetic warming, is that also as optimized as it could be?

Hunter: No, I think that you can continue to make better versions of this. And the... John Bischof’s group has continued to do work on improving the core compositions. We have done our own work on improving core composition materials. There’s a whole field obviously that pre-exists for how to do really good coupling of magnetic fields with nanoparticles. This is not specific to the problem of cryobiology. I think honestly, as with many questions inside of cryobiology, one need only look outside of the field, and this is probably true of any discipline in science, look outside of the narrow aperture and you will find someone else who has solved a similar problem. This is a good example of that.

Abhi: I guess even outside of magnetic induction, is there any other upcoming or perhaps already well-established approach that you view as particularly promising?

Hunter: Something that I was interested in that came out recently was the use of electric fields and microwave... actually very tens of megahertz frequency electric fields for rewarming a rabbit kidney. This was Greg Fahy was the person who led this study, and I just don’t see how this scales up.

Abhi: Okay.

Hunter: I think that maybe one day it will be doable, but it is quite challenging to think of how do you scale this from something that’s rabbit-sized up to something the scale of a human organ?

Abhi: Is it just because you mentioned earlier about how microwaving has the chance to introduce such large thermal gradients or something else?

Hunter: It... there’s there’s a bunch of... and some of the physics here is a little bit subtle, but yeah, there’s a tendency for hotspots to form inside of the organ when you’re using electric fields. Which couple... The way of thinking about this very intuitively is in one case I’m trying to directly couple to water, the thing that is everywhere inside of the organ. And so by definition, my energy is going to get attenuated just a little bit as I’m trying to go through. Whereas for the iron oxide, this is an exogenously introduced material that has a very high coupling coefficient that is very localized to the particle itself. So that... there’s some arguments for increasing the penetration depth using that.

Abhi: For the magnetic stuff, what are the waves actually interacting with? If you replace all the water...

Hunter: What are the waves interacting with? With the magnetic fields? Yeah. Yeah. So they’re interacting with the nanoparticles.

Abhi: Oh, okay.

Okay. They’re still... they’re still using the iron oxide or something...

Hunter: In the... sorry. In the case of any magnetic field stimulation, you’re going to have to use iron oxide nanoparticles.

Abhi: Okay. Okay. Yeah.

Hunter: For electric field stimulation, you can directly excite water molecules themselves. Gotcha.

Abhi: Okay. One thing that has been you’ve hinted at this, or perhaps explicitly mentioned at one point. All this inherently depends on the vascular system. Distributing everything nicely and equally. How does this work for tissues that don’t perhaps, or are not perhaps as heavily vascularized? I think even within an organ, I’m sure not every single little bit of it is... there’s potentially fascia that does not have that much vasculature attached to it. How much work goes into thinking about that? Is that kind of a long-term thing? Like we don’t need to worry about that right now? Or that’s on your mind?

Hunter: Yeah, I mean it is on the mind insofar as we are continuously thinking about the entire scientific program and not just what is right in front of us, but it is not right in front of us. In fact, yeah. There are areas of the kidney that are less vascularized than others, but all of them are sufficiently vascularized that you can, with some intelligent protocol development, get the cryoprotectant where it needs to go. And get the heating power where it needs to go. Yeah. But yeah, you’re right that the vascular density even inside of a given organ is not homogeneous.

Abhi: How are... you’ve mentioned a few metrics you use thus far for measuring function of the organ post-warming. Are there metrics to establish how well the perfusion process works in the first place?

Hunter: Yeah, so there’s a few different ways of doing this. So you can do... there’s a variety. One is I can just look at when I cool the entire thing down and vitrify it. This is a midpoint assay. You can actually use micro-CT to figure out, are there ice nuclei somewhere inside of this organ.

Abhi: That’s interesting.

Hunter: Because you can look for the crystals themselves. There’s also some crazy applications of MRI where you can look for chemical exchange with your cryoprotectant agent inside of the MRI. And that can give you an idea of local concentration of a given molecule. Actually, Alex German, the guy that I referenced previously, was a radiologist first before becoming a cryobiologist. He published a really cool paper on using something called CEST imaging to look for the concentration of given cryoprotectant species deep in extended tissue.

[00:53:03] What are the experimental costs of running cryopreservation studies?

Abhi: Okay. And one question I had while reading the September 2024 report was this... this experiment was done on four mice, which is a relatively low N. Yeah. And that made me think, how expensive was this whole process? How many shots on goal do you have when you’re running Until Labs? Do you do... you guys all sit in the room, you think very hard and then you run a million-dollar experiment? Or is it more there’s a bunch of intermediate assays you can do to get sanity checks?

Hunter: Yeah. Yeah. This is great. So definitely the mice are less expensive than the larger preclinical models for sure. But the way I like to think about this is as a funnel. And at the top of the funnel you have in silico models. You want to do as well as you can with in silico models because they are very cheap to run.

Abhi: Sure.

Hunter: Compared to everything that’s going to come downstream in this pipeline. Then below the in silico models, you have things that you can do in a test tube or things that don’t require any interaction with biology. A canonical example of this would be things like differential scanning calorimetry. So here you can take a cryoprotectant that is in water, you can put it in a tiny little sample and cool it down to minus 130 degrees. And you can look for an exothermic peak. This is heat flow out of your sample during the cooling or during the rewarming process. If you see this exothermic peak, that’s indicative of ice formation. The liquid-to-solid transition in that temperature range is exothermic. So you’ll be able to tell, oh, I got ice in this little solution, and the solution still hasn’t seen cells yet.

You can then imagine screening many of these such compounds in cultured cells.

Abhi: Yeah.

Hunter: And that gives you some idea of toxicity. And then from there, I think that this is where a lot of our work is done, it’s what is the gap? How do you bridge between doing cells and going all the way to an organ? Because setting up organ experiments, like you were saying, is both expensive in terms of capital, but mostly it’s that they’re very time-expensive. These are very long protocols. And I’m incredibly proud of our team for their ability to work through some of these arduous protocols, but you don’t want to waste that.

Abhi: Yeah.

Hunter: So I think that there’s a lot of focus on how can we move stuff up that funnel to more translational assays.

Abhi: To give some sense of how expensive this protocol is from a time perspective. What did it take to get these rat cerebellums frozen and rewarmed?

Hunter: Yeah. So the rat stuff, the rat acute slices, that’s not so bad.

Abhi: Okay. Okay.

Hunter: That’s not so bad. The challenge is when you want to do a preclinical model of an organ. And that obviously requires actually taking the organ out of the animal or getting it. And then you’re going to have to do things like flush the blood out of the organ. You need to load with cryoprotectant, you need to do the vitrification process, do the rewarming process, unload the cryoprotectant, and then do the assay.

And the assays themselves take several hours to be evaluating the function of the organ. So it’s just, it’s a very long process if you want to get down to an entire organ as an evaluation.

Abhi: That makes sense. One question that immediately popped up in my head was, I imagine you want to rewarm uniformly to prevent these thermal gradients and prevent mechanical stress. Is there any world in which you have a molecule that you can inject alongside the cryoprotectants to help with that? Are there such things as cryoprotectants that help prevent mechanical stress?

Hunter: Yeah, so I think mechanical stress is one that would be interesting to look into. I haven’t looked specifically at that, but what we have looked at... there’s a set of molecules that would help with not necessarily blocking ice, but help alleviate any damage that would be induced.

Abhi: Yeah.

Hunter: And we’re definitely not the only ones who are looking at this. I think the field as a whole is looking for these additive molecules that can help with things like the shock that is induced by going down and coming back up again. There are also some cryoprotectant agents whose mechanism of action is strictly what I would call biomechanical. That instead of blocking ice, what they do is they actually interpolate into the biological membrane and strengthen it. So you can actually have these, they’re polymer molecules that Medevelop developed that will actually sit in the membrane and will strengthen the biological membranes. This obviously works much better in small cells or embryo context. Pretty hard to get that to work at the scale of an organ, but it is an active area of research.

Abhi: Why is it difficult to get to work at the scale of an organ? Just...

Hunter: Loading things that are that large homogeneously outside of the vasculature becomes challenging. The transport here is one of the key limiting factors for any of these approaches.

[00:57:49] What happens to the cryoprotectants and iron oxide nanoparticles after the organ has been thawed?

Abhi: I think when we first talked, one of the most interesting questions I had and most interesting answers you gave was, you fill up this organ with cryoprotectant and iron oxide nanoparticles and you rewarm it back up. What... how does the body deal with all this stuff that’s left over?

Hunter: Yeah. Yeah. So you really don’t want to leave anything behind at 37.

Abhi: Yeah, for sure.

Hunter: And so I think a lot of the work that we do is making sure that you’re able to effectively clear all these cryoprotectants. I think that there is sometimes a misunderstanding of a beautiful symmetry here. So when you go down to around four degrees, which is where we load the cryoprotectant into the organ, obviously the metabolism is much slower.

Abhi: Yeah.

Hunter: And as a result, the toxicity of any of these molecules is substantially reduced. As long as you’re loading down at four degrees. And we load between four and 10. So you hold there, you load your cryoprotectant up. Similarly, when you want to go back and you want to unload, I think that oftentimes there’s this misconception. It’s oh, I need to add a chelator or something to take the cryoprotectant out. But you can unload the cryoprotectant the exact same way that you loaded the cryoprotectant,

Abhi: Just flush...

Hunter: Just slowly reducing the concentration of the cryoprotectant in what’s being flowed through the vasculature. And then you get passive diffusion that starts to take it out.

But yeah, critically we don’t want to leave cryoprotectant behind and then warm the organ back up to 37 degrees Celsius. Because then we eat all of this toxicity that’s going to be a result of being warm and loaded with cryoprotectant.

Abhi: Yeah.

Hunter: Similarly on the nanoparticle side, I think a lot of work in the field has been focused on passivating the surface of the nanoparticles to make sure you’re not leaving iron oxide behind in the kidney.

Abhi: You’ve mentioned a few times in the past about, you don’t want your cryoprotectant agent to be too toxic, but why does toxicity actually matter if it’s so cold that no reactions are actually occurring?

Hunter: Oh, interesting. Yeah. So the thing is that some reactions are occurring at four.

Abhi: Okay.

Hunter: That’s the trick. Gotcha. Yeah. It doesn’t matter if it’s toxic once you’re down at minus 130.

Abhi: Okay.

Hunter: For sure. And there have been some strategies that have been brought up previously where you can actually have very toxic cryoprotectant agents that can be loaded even colder. So one example of this would be M22, which is a formulation that must be loaded at minus 22 degrees. That’s where it got its name. Very effective cryoprotectant, also very toxic.

Abhi: What does toxicity usually... yeah. What does toxicity usually manifest as?

Hunter: Yeah, there’s a variety of different... when you say manifest as, it’s like in cultured cell context or...

Abhi: In full organ context, or I guess has there ever been a particularly toxic cryoprotectant used in organs?

Hunter: I think one tends to not use a really toxic cryoprotectant in whole organs. Obviously, the outcomes here are going to be different based on different organs.

Abhi: Yeah. That makes sense.

Moving... perhaps relatedly, there was this kidney... mouse kidney transplant thing that you discussed earlier. How did that go? When they... everything was an autologous thing. Where they took the kidney out, froze it, rewarmed it, put it back in. Was the mouse just perfectly fine or... I think, was it a rabbit or mouse?

Hunter: It was a rat.

Abhi: Rat. Okay. A rat. Okay. Did the rat live?

Hunter: Yes. The rat in fact lived. Yeah. It was a cool study. They showed that you can actually regain normal function in the kidney. I think that if you look at the recovery time, it took some time for the organ to recover, a matter of weeks for the organ to fully recover normal function. But they were able to get a viable kidney through the cryopreservation and rewarming process.

[01:01:33] Cryopreservation and immune response

Abhi: Gotcha.

And I remember you mentioned at some point that there was this concern of an immune epitope exposure as a result of the rewarming and the freezing and rewarming process. Would you be able to walk through that again? Because I thought that was a very interesting anecdote.

Hunter: Yeah. I think that anytime that you’re going to induce cell lysis in tissue, you can invoke an immune response. And so there’s actually been some examples of this in human pediatric cases. Where you can take some heart tissue out of a person, it can go back into the person afterwards. So obviously there’s no self versus other response. But the strategy that you use for cryopreservation affects the outcomes of these very tiny little tissue extracts and then reimplantation. And one hypothesis for the mechanism of action here is that because you get some cellular lysis during the cryopreservation process, you expose a bunch of the intracellular molecules that are then going to induce some immune interaction. And can cause rejection of the graft, even though it’s an autologous transplant.

Abhi: I imagine there haven’t been so many cryopreservation transplant studies done such that it’s pretty clear when this will happen and when it won’t happen.

Hunter: Yeah. I think that this is such a nascent field that I wouldn’t say... The thing that is not nascent is embryo cryopreservation.

Abhi: Okay.

Hunter: There, we really, I think, have many N on that process. Right now, I think this year there will be something like 150,000 embryos... 150,000 live births this year in the US alone that were previously vitrified as embryos, which is a wild number to me. And you can store for... they can get stored for 30 years, which is crazy to me.

[01:03:24] How do you filter through the cryopreservation literature

Abhi: Cryopreservation is a nascent field. It does seem like it’s ramping up in recent years. And now that there’s this probably deluge of cryopreservation papers consistently popping up all the time. I’m curious, what is your own filter when you’re reading through these papers? What marks a paper as particularly high quality and you pay attention to this versus something that you can probably just skim and not super pay close attention to?

Hunter: Yeah. I think that there are fortunately a whole host of groups that are doing things that would meet this bar. But I think that the bar that I’m particularly interested in for when you’re doing things that are going to translate into an organ is to ask the question, are the assays that are done to demonstrate the viability of this organ in concert with what the field that they’re trying to interact with has established? So for example, if they’re doing kidneys, are they doing standard assays that would be accepted in nephrology to evaluate the viability of this organ? And I think that if you go all the way down to the molecular study, I think the molecular studies that I’m particularly interested in are ones that don’t do random scattershot screening, but instead really get into what’s the mechanism of action of the interaction of this cryoprotectant with the water molecule. Water is an under-hyped molecule. It’s actually incredibly subtle and complicated to interact with it. I’m biased, but it is currently becoming my favorite molecule. And I think that you can write really deep interrogations of the interactions of these cryoprotectant agents with the water molecules. I think that those can be deeply informative for follow-on study.

Abhi: And sorry, these are papers specifically trying to create new cryoprotectants or even in the context of no novel cryoprotectant work, you still want to know what the interaction of the cryoprotectant was?

Hunter: I think I learned stuff even when people just look at DMSO interacting with water. And people have done great things of studying the Raman spectroscopy of water and DMSO interacting to try to figure out the actual coordination, like how is this hydrogen bond actually coordinating? I think that sometimes these things are academic and can’t be translated into better engineering, but I think oftentimes we can lay better foundations for our mental models of coming up with better cryoprotectants and definitely we can come up with more interesting metrics we can be calculating in silico. I think that’s where a lot of these physical chemistry studies become really helpful is it’s oh, how can I think about simulating that? How can I think about driving this onto a chip?

[01:05:54] How much is molecular simulation used at Until Labs?

Abhi: On the topic of in silico simulation, how much molecular dynamics, molecular simulation goes on internally at Until?

Hunter: Yeah. I think we think a lot about how to do in silico screening well, and I think that there’s a variety of different ways that you can do this. One that is quite common is you can think about molecular dynamics simulations where you’re trying to look at what is the extension rate of an ice nucleus that is placed inside of some solutions. These are coexistence simulations. The way that you can think about this is just set up a box. This box has hexagonal ice in one half of it, and the other half it has some mixture of cryoprotectant and water. And you can look for, does the ice tend to extend or does it tend to melt?

And this is a good way of measuring the equilibrium between the liquid and solid state, given some cryoprotectant mixture. And we found that this is somewhat interesting to look at for in terms of ice formation.

Abhi: I’m obviously not at all a molecular simulation expert. My interpretation was that phase changes, especially of water in molecular simulation, is really gnarly and not very well modeled. Does it... do you see that? And even if you do see that, you still think these simulations are pretty predictive of what happens in real life.

Hunter: Yeah. I think one nice thing about trying to be an engineer here and not trying to be a scientist is that you just want to find things that correlate.

Abhi: Okay.

Hunter: And we have the thing in the lab. It is true, the thing that you’re getting at, which is a real problem for the field as a whole, is it’s quite hard to simulate actual ice nucleation. So this is a very... the thing I just pitched you is super contrived, right? I literally just set up a boundary of ice and I’m trying to look at the thing that is interacting with it. This obviously has some deep limitations on its translation to the actual test tube that I’m doing the experiment in, and it’s because I haven’t allowed for nucleation. Nucleation has been taken completely out of the equation.

And that’s the whole thing that you’re trying to get at, is how to suppress the nucleation. The reason nucleation is hard to simulate is that it’s actually... it’s too fast for us, but it’s too slow for molecular dynamics. Because if your time steps are one femtosecond and you need to simulate for a nanosecond or something to be able to see nucleation, for every cryoprotectant that you want to interrogate, this is not tractable. So I think that it’s a challenging thing to get around is what is the right thing to simulate? And I think the things that our applied physics group spends a lot of time thinking about is what are the right things to simulate to be able to have some predictive power around what is the efficacy of this cryoprotectant going to be?

Abhi: How much do you personally or perhaps anyone in the company pay attention to the neural network potential research that’s coming out?

Hunter: Yeah, I think that this is actually really helpful. There’s some versions of this that are more useful than others, as it would always be the case. Some of the neural network potentials are still very costly to simulate compared to classical force fields. And so I think you want to pick and choose which ones you use and in what context you use them. It’s I think there’s no, at least in our hands, there is no skeleton key for, oh, you do this exact simulation and it works. I think it’s more of an intelligent ensemble of simulations to try to get some interpretation. But yeah, I think the work that’s been coming out of ML-related tools for better simulating interactions is going to affect cryopreservation, just like it affects drug discovery. This is fundamentally improving our ability to model the physical world in silico. And I think that is great, and I would encourage those academics to continue to push hard.

Abhi: Maybe a naive question, how much does QM matter for these sorts of simulations? Or is molecular mechanics fine.

Hunter: Yeah. I think that there are some things that you really care about quantum mechanics for, and this is where the neural network-based potentials can be helpful is the actual hydrogen bonding of the stuff to water obviously involves things that are not well simulated just with a standard Lennard-Jones potential. Sure. So there, I think that it is useful to think through some more complex interactions than just a simple Lennard-Jones system.

[01:10:04] What are the (expected) economics of Until Labs?

Abhi: Okay. Yeah.

That makes sense. Moving on to almost non-scientific questions about Until. One immediate question I had upon learning about you guys when you were named Cradle was, what are the economics of this setup? My mental model of organ transplantation is that there is this organ waiting list. When your name gets called, you go and get your organ. There’s no direct-to-consumer setup. Who are buying these organs?

Hunter: Yeah. So the transplant pipeline is very complex as I have been learning. And the thing is there is no direct-to-consumer. I can’t call up and say, “Hey, I need a kidney.” That’s not how this works. As you were mentioning, there is a transplant list. There are the organ procurement organizations who are responsible for facilitating the transaction of organ comes out of this donor, goes into this recipient. And there are a host of companies that specifically handle the logistics of transporting the organ from one place to another viably. A really big player in this space would be TransMedics, a publicly traded company, who literally... they have a suite of private jets that will fly around and pick up the organ from the donor and bring it back to the recipient. This is all very heavily coordinated and the logistics are certainly not trivial.

Abhi: We were just talking about this before we started this conversation. I had this question about, right now putting an organ on ice is incredibly cheap. Perfusing with oxygen is incredibly cheap. What’s the value proposition to go for something like Until Labs? A potentially much more expensive protocol. And you mentioned that you’ll need to rely on this incredibly expensive transport chain less heavily.

Hunter: Yeah. I think that one of the nice things about doing something like organ vitrification is that because it takes urgency out of the process, it just relaxes all of these logistical constraints. So for example, I don’t need to have a private jet to go get the organ anymore. I also don’t need to wake up a transplant surgeon at 2:00 AM because that’s when the organ became available. Yeah. We now have in our vision a process that can be much more, let’s say, disciplined about bringing the organ from the donor to the recipient. And this has a bunch of knock-on effects.

So one for example is that it could increase testing. If you look at the outcomes for living versus dead donors for kidneys, if you look at 10-year graft survival rates, the 10-year graft survival rate in the US for a kidney recipient from a dead donor is around 50%. From a living donor, so this would be you get it from your brother or something. Sure. It’s about 60%. So literally just the increase in immune matching of getting it from a living donor... and it may also have some other logistical constraints there of you can literally do it on the table that’s next to the person. But there’s that much benefit to get just from improved matching in the biological sense.

If you look at the reference that I made previously about the fact that in vitro fertilization of embryos... you now have a higher chance of getting a live birth from a cryopreserved embryo than from a freshly implanted embryo. And the reason for this is, again, increased ability for testing. So we think one, the cryopreservation process can lead to better outcomes for patients because we have this time that we have bought to be able to improve matching. We think that it can improve the equality of organ allocation by allowing us to respect the transplant list more, and have fewer open calls where the organ just needs to go to someone because we don’t want it to get wasted. And then, yeah, no private jets required because we can get them there. No surgeons woken up in the middle of the night.

Abhi: When it comes to what the supply chain would look like if Until Labs ruled the entire system. How careful do you need to be with a vitrified organ? Can I put it in a truck and just have the truck go? Or does it need to be in a very specific, very special container?

Hunter: So I think we would manufacture a container that was sufficient such that it could go on a truck. I think the things that are significant is you don’t want thermal gradients to be able to come in. You don’t want to thermally cycle the organ ever up to above minus 130 degrees. There’s some stuff around tight temperature control.

But I think that these are all highly solvable problems. And I think the marginal cost of doing this on a program basis is pretty trivial.

[01:14:49] How much does cryopreservation practically solve the organ shortage problem?

Abhi: Yeah.

And again, in this hypothetical of Until Labs is everywhere. How much... is the organ shortage problem solved overnight?

Hunter: No. Okay. Unfortunately not. I wish that were the case. But in the end, this is still a supply-limited market.

Abhi: What gets better? Let’s say 10% of people who need... 10% of names are crossed off every year from the organ transplant list and everyone else dies. What does that number rise up to, at least for kidneys?

Hunter: Yeah. So I think that what’s going to end up happening is you’re going to have initially, let’s say, a few thousand organs, which would be the ones that should be going into patients, would be viable, but get lost due to logistics. It’s like the organ is on the plane, needs to get de-iced, the organ expires while it’s on the plane.

Abhi: Does that happen?

Hunter: Yeah.

Actually, Laura, my co-founder, was literally talking to a transplant surgeon the other day that was recounting this exact story. I’m not making this up. This is an actual thing.

Abhi: And this is not a particularly rare incident.

Hunter: I think that this particular annoyance of plane de-icing maybe is rare. But I think that the idea... if you can get a few thousand additional organs, if you reduce logistical constraints. And these are publicly available figures.

Abhi: Yeah.

Hunter: The thing that is an interesting unlock in the long term is if you can start to relax the supply constraint.

Things like being able to get organs that are from... there’s a concept of DCD versus DBD donors. A DCD donor is death by a cardiac event.

These are more, let’s say challenging logistically to get out. The time constraints are tighter.

Abhi: And just because ischemia happens immediate...

Hunter: Okay. Exactly. It’s an ischemia issue. The clock starts earlier. And so as a result, that is a nascent field where people are trying to push into these DCD organs to try to increase organ availability. The other in the long term that I’m particularly excited about is xenotransplantation.

[01:17:04] Synergy between xenotransplantation and cryopreservation

Abhi: I was going to ask about that.

If we had sufficiently good xenotransplantation, do we need cryopreservation?

Hunter: I think sufficiently good xenotransplantation mandates cryopreservation, but that’s my perspective.

Abhi: What’s the rationale there?

Hunter: I want to envision a future for xeno that is maybe more aggressive than has previously been talked about. I want to envision a world where someone who has a heart attack and maybe would not even have time to get a heart transplant, can now get a heart transplant. Where you go into an ER and there is a liquid nitrogen dewar sitting there that is filled with hearts that are ready to go, they can be transplanted in. Now all of a sudden, organ donation is no longer something for chronic conditions. It is now something that is also for acute conditions. And this is the thing that will dramatically increase the availability of organs is if xenotransplantation could get solved. I think that there is some reasonable, let’s say, skepticism about the timetable on which xenotransplantation will come online. But yeah, I am, let’s just say I’m cautiously optimistic that those guys will make progress. And I think that cryopreservation would be a natural fit for their logistics supply chain.

Abhi: I think bringing acute conditions onto the table is really fascinating. I had never really thought about it that way. What do you think of the xeno... I’m not super familiar with it. I just know there was that pig heart that was CRISPRed to be a little bit more humanized. It was implanted into a human patient. The patient ended up dying, I think, but potentially for reasons unrelated to the heart.

Hunter: Yep.

Abhi: Do you think that field is going to rapidly mature over the next five years or there’s some big insurmountable problems there?

Hunter: Yeah. I should clarify. I’m certainly not an expert on the process or the progress of that field. And I think that if you ask five people, you might get six opinions on the future of xeno. I’ve heard everything from, “Oh yeah, it’s right around the corner.” To one transplant surgeon told me, “Xeno is the future of transplantation and it always will be.”

Abhi: It’s perpetually five years away.

Hunter: Exactly. It’s perpetually... Yeah. I think that was his perspective. Okay. So I think that there’s a variety of perspectives on that one.

Abhi: One thing I was really curious about, this X number of people die per year because they are unable to receive that organ. How has X changed if Until Labs really succeeds? I’m not sure if... the answer you gave was, oh, several thousand because we’re still figuring things out. I’m not sure if you had an exact... there’s been almost models drawn up as to how much can we put a dent in the organ shortage crisis if this really takes off.

Hunter: Yeah. This is a really, unfortunately, it’s a super complicated question to answer. And really a complicated, even more complicated question to answer well. And the reason for this is that the data here is just, it’s very challenging. Yeah. I imagine I should tease out what is the counterfactual of the organ going into a person or not going into the person. So we have some statistics on the expiry of organs during transit, which is where I got the few thousand organs metric. You can actually look at a pie slice of what’s the outcome of various organs. And the ones that are basically, were viable but expired in transit... if you add that pie slice with some other ones that are clearly logistically related, that’s where you arrive at a few thousand organs a year.

Abhi: Gotcha.

And so it’s not necessarily the case that, let’s say Until succeeds, 10 years goes by, we’ll have a surplus of organs, like every organ that’s currently in transport right now will be given to someone.

Hunter: Yeah. I think, that for the time being, this is still going to be a supply-limited problem.

Abhi: Gotcha. Okay.

Hunter: So it will still be the case that there will be a waiting list. Unfortunately, it’ll still be the case that there will be people on dialysis in this country. And I think that it will have to be cryo plus some other technology that will need to come online for that to not be the case anymore.

[01:21:12] How much will the final cryopreservation protocol likely cost?

Abhi: Yeah, that makes sense. How much do you envision the Until Labs protocol costing? Is it an undecided figure or...

Hunter: It’s an undecided figure. Partially because we don’t know what the Until Labs protocol will be.

Abhi: That’s fair.

Hunter: I think it’s... we have such a large suite of potential technologies that are brought to bear, but what I can say is that I don’t see any obvious place where this is going to be a gene therapy that costs a million dollars a patient. That’s not what we’re talking about here. And I think that the parts that are expensive are primarily the devices.

Which will be amortized over many, organs. These perfusion devices that we’re talking about, you’ll have a disposable component to it for sure. But that’s not the expensive... the expensive part is the part that’s reusable.

[01:21:58] Who ends up paying for this?

Abhi: When it comes to companies like TransMedics, who’s paying? Are they primarily contracting with insurance companies and you are also planning on contracting with insurance companies? Yeah. Who... yeah. Whose job is it to ensure almost there’s alignment between what the patient wants and the money that you expect to get from this.

Hunter: Yeah. So there are Medicaid reimbursement rates that are set up for organs, basically. And so there’s a public payer. Obviously that’s setting some market baseline.

And then yeah, there’s insurance companies that you have to be able to figure out what they’re willing to compensate the transplant centers for. The customers that you’re actually working with though, these are transplant directors. And you’re going to be working with some OPOs. That’s the people who you’d be directly interacting with.

Abhi: And what’s the approval process for this? Because it’s not a drug, it’s almost a procedure more than anything else.

Hunter: Yeah.

Abhi: Is it a... what is it classified as exactly?

Hunter: Yeah. So we don’t know yet. But I think that medical device is probably how it’ll be classified. Yeah.

Abhi: How... and I imagine there’s not that much precedent for something like cryopreservation or would... was the ice stuff also... that had to go through its own approval process?

Hunter: Yeah, I think there’s a different... there’s going to be different approval processes for each of these. I think that there is some precedent if you look at things like hypothermic machine perfusion. I think that could reasonably set some precedent for the vitrification process. But again, we’ll have to leave that to the FDA to decide.

[01:23:28] What was it like to raise a Series A on such an unorthodox thesis?

Abhi: Yeah, that makes sense. Moving on to the actual raising journey for Until Labs. What was the series A again?

Hunter: We did 58 million.

Abhi: Okay. Okay. Until Labs is a little bit of a strange thesis for a company. It’s a biotech, but it’s not a therapeutics company. It’s serving transplantation, which I had not conceptualized as, oh, there could be a for-profit company really playing and doing innovative biomedical research here. And as far as I can tell, you guys are one of the very few people playing in that area. What was the fundraising journey like?

Hunter: Yeah, so I think that, first of all, I would say that while it may appear that we are one of the few people working in the area, my guess is that is not for long.

Abhi: Okay.

Hunter: And in terms of the fundraising journey, I think it started about a little over a year ago, not where we were actively looking to raise. We had just completed our seed and the neural slice paper came out right after we announced our seed.

Abhi: So what neural... the German lab?

Hunter: The one... no, the one you were referencing.

Abhi: Gotcha.

Hunter: The one where we did the rat cerebellum. We used that and announced the seed at the same time. And that was 15 million for the seed. So, yeah, about a year, a little over a year ago, we had this idea that maybe we wanted to start playing around with this donor organ problem.

And as I said, there had already been some of this work from the University of Minnesota, which we had been looking at. And I think it just became clear that this was going to be on the roadmap anyways, this is going to be on the scientific trajectory that we would want to be on and we get to start helping patients. So Isla, who’s currently our director of preclinical research, came to me and pitched this whole cloth, was saying, look, we have this long-term roadmap, there’s this obvious use case in getting this into patients, we should go after this seriously, and we should start scaling right now to go into preclinical models. And so that kicked off this journey where we started to initially just kick the tires on, could we take our protocols and move it over? Could we take our engineering team, task them off to this? And now this is the dominant focus of the company is how to get this translated.

One thing that we wanted to be able to do there is raise to be able to accelerate that process of getting this done and towards the point of doing a first-in-human trial. And so I think that the primary purpose of the raise is to bring on additional capital to be able to parallelize a lot of these processes and get things to market quicker.

Abhi: Was... what was raising like? How many questions did you get over the economic thesis versus the scientific thesis versus some other thesis?

Hunter: Yeah. I think we got questions along both axes. Okay. And I think that we are very fortunate to have excellent partners. Excellent capital partners who I think understand the thesis really deeply. I think actually in one of the pre-conversations you had asked me, what advice would you give to people who are in similar positions? And I think I didn’t give you a particularly satisfying answer because I don’t think that I have one. And the reason for that is I think that we were really fortunate in that the people who... so the raise was led by Founders Fund with Field Ventures and Lux joining.

And all three of these groups are able to do exceptionally detailed diligence on their own. Having technical conversations with their technical team was like you and I sitting here talking about science, plus on the level of depth that was required. But it wasn’t some story that needed to be packaged. For them, it was just an actual conversation about the technical risk.

Abhi: So it wasn’t necessarily that, oh, we want to do brain preservation and we’re moving over to kidney preservation because we’re bringing in outside investors...

Hunter: Yeah. No, that was... it was very much the reverse process there. It’s like we went to go get more capital to accelerate being able to get the donor stuff to market.

[01:27:49] What are common misconceptions people have about cryopreservation?

Abhi: What are common misconceptions that people have about the cryo field, especially people who are external to the field, and perhaps not even laymen, but people who you consider smart and what their misconceptions?

Hunter: Yeah, I think that I had a misconception when I originally jumped into this, which was that it was not doable because of ice. That you would always get ice that would form and irreversibly damage tissue. And in reality, you have this minimum temperature for ice formation at minus 130 degrees Celsius where water turns to glass, not ice. So if you can traverse this danger zone between zero to minus 130, then the tissue will be safe and you can rewarm it without damage.

Abhi: But you did mention about how these cryoprotectants have existed since the 1950s.

Hunter: Yeah.

Abhi: And so there, I imagine there should have been time for the rest of the scientific field to be aware that, oh, ice nucleation is a solvable problem. Why do you think there still is this misconception that it’s fundamentally unsolvable?

Hunter: I think that part of it probably has to do with the fact that when we actually do cryopreservation of things like cultured cells in the lab every day, we use a completely different process that does allow ice to form, that would not be compatible with tissue. So you imagine, take some cultured cells and you want to store those in vapor phase storage.

Abhi: Yeah.

Hunter: There you use a process called slow freezing. You’re allowing ice to form in the extracellular space and it slowly expands and hyper-concentrates the cryoprotectant elsewhere and then forms the glass. I think intuitively people understand this should not work for tissue. Because you’ll tear up the extracellular matrix. Yeah. So I think that probably has facilitated a misunderstanding of fundamentally how the cryopreservation process works when you want to go to these more tissue-specific or organ-specific processes.

Abhi: Does... it sounds like people don’t actively think of the vascular system as a very good transportation medium. Is that fair to say?

Hunter: So I think that’s also a common misconception. It’s oh, if I need to diffusely load this object, isn’t it going to take forever? Because they think about the equivalent of, okay, I have in vitro fertilization, I have a tiny little embryo that’s six cells that needs to load cryoprotectant. People can envision in their head, okay, I load the cryoprotectant, I cool really fast. That makes sense. And I think that a key unlock is hijacking the vascular system for mass transport.

[01:29:58] The beginnings of Until Labs

Abhi: Why... at the very early beginning when you and Laura first founded this company, was the plan, oh, we’re just going to push forward on this brain thing until it’s done?

Hunter: Yeah. So when Lauren and I met... that’s a funny story. So Lauren and I met via her cold emailing me while I was a postdoc in Adam Cohen’s lab at Harvard.

And she hit me with a very open-ended question, which was if I thought it was possible to reversibly pause biological function. And my initial response to her was the same response that any reasonable person has, which is basically, are you kidding me? Of course not.

Abhi: Yeah.

Hunter: And there was a couple of days after there where something was eating at the back of me, which is... I was a physicist by origin. And it’s okay, if you have this kind of response to someone who has this history, right? Laura had a very established history at that point of making strong scientific bets.

Abhi: Yeah.

Hunter: It’s like, okay. I should be questioning my assumptions here. And so I went back and looked at it and was like, oh, this problem is fascinating. And I think that there’s real traction that’s been made recently and I have some ideas about how we might be able to continue to do that.

And so initially it was very much this curiosity that drove me into wanting to join up with Laura and get this going. And that curiosity manifested in a slowly expanding way where I realized the places that this could help. As a bit of personal context, in 2016, my father-in-law, Mark, was diagnosed with a terminal case of cancer and was given a six-month prognosis. And he lived almost exactly six months. And near the end of his life, a clinical trial for Keytruda came online for his disease. But he was too sick to qualify for it. And I think in these conversations with Laura, something that hit me was this is a technology that could help people like Mark.

People like Mark and their families who they don’t need some hundred-year jump into the future. They need six months.

They need a year. And this continued conversation with Laura led me to understand, oh, Mark’s case is not isolated. This is not some exceptional thing. It is the case that oncology drugs are forever getting better and survival rates are increasing. Things like pandemics, think about AIDS. Okay. In the 10 years between the onset of the AIDS pandemic and the creation of combination antiretroviral therapies, 9.7 million people died. 10 years, 9.7 million people. And I think there’s just this overwhelming sense of... there were no other things that I had touched before scientifically that had this kind of a lever arm on them to be able to affect this many people just by time. It was a very skeleton key solution to healthcare. So yeah, I flew out to meet with Laura and basically, flew back, quit my job, wrapped things up and moved out. And I think initially we really were committed, and still remain very committed to this day, to this hibernation as the long-term goal.

And this is the idea of taking an entire person, taking an entire patient who would be terminal with some disease and giving them access to cures that are right around the corner.

We started with that in mind and that was why the neural stuff, that was the first thing we wanted to do because it’s like the hypothesis is, oh, that’s going to be the hardest thing. Yeah. That’s going to be the thing that for sure breaks.

These are insanely delicate tissues. In the lab we would have challenges just keeping them viable on their own, much less cryopreserved. So yeah. That’s what we came out to originally do.

Abhi: I think prior to even learning about Cradle existing, whenever I thought of cryopreservation, I just assumed, oh, that’s thermodynamically impossible to do. It’s not... no one’s ever going to crack the problem. Everyone working in it are grifters.

Hunter: Yeah.

[01:34:07] What expertise is hardest to recruit for?

Abhi: Yeah, it... I think when I read that progress report in preparation for this interview, I was like, oh my God. That’s obviously small scale, how do you scale it up, but still a crazy achievement. One question I had was, you’ve mentioned over and over again about how multidisciplinary cryopreservation is as a field. What expertise is hardest to recruit for?

Hunter: Yeah, I think that again, you want top talent across everything. And so getting top talent in any field takes a disciplined approach. And I think that you want to show people that this is for real and show them how they can contribute. I think the thing that practically is very hard to onboard is medical talent and expertise.

And I think that this is primarily driven by the labor and economic incentives of being a doctor in the US. No one wants to stop doing their clinical practice, which means that it is really challenging to be able to get good medical talent to come and work on it. We’ve been very fortunate to recently have some awesome talent on, not necessarily even from MDs, but from PhDs who’ve been working at Hopkins for example. So Dr. Amanda Lofton, who has recently joined us, she’s a PhD at Hopkins who studied things like the perfusion of these donor organs at hypothermic temperatures. So there are... you can find it. It takes a disciplined approach. And you’re definitely looking for personalities that have a high degree of, let’s say, openness to change.

Abhi: How important is it that the people that you recruit for these positions have actual... have spent several years in an organ transplant clinic versus they have done an MD and know the biology?

Hunter: Yeah. so I think that we tend to across the board recruit for people who understand the platform, understand the biology as a platform and not very specific... So I think hiring too specifically in a company like this is a deep mistake.

Abhi: Really? Why? Why is that?

Hunter: And so I think, for example, you could go and say, oh, we’re just going to hire... let’s, if we shifted over to the molecular development side. Sure. You’d say, oh, we’re exclusively going to hire people who have developed cryoprotectants before. Yeah. I think that it’s reasonable to hire people who have developed cryoprotectants before. But I think that those are certainly not the only people who are capable of contributing well to this problem.

And so I think our approach on recruiting has been to go find not people who specifically have thought about this problem, but have thought about the facet of the problem that we need, maybe from a different lens. So Andrew maybe is a great example of this, very talented material physicist, both previously thinking about x-ray scattering and...

Abhi: This is the Tesla guy?

Hunter: This is... Yeah, exactly. He was doing battery research previously. He’s... you wouldn’t naturally think thinking about lithium ions is very similar to cryopreservation, but he’s been a very effective leader for that group. And I think that is just one of many examples of highly agentic people coming into our group and being able to make real progress against this problem.

Abhi: I imagine the mental leap to finding Andrew out in the wild working at Tesla and thinking, oh, maybe he’d be really useful for a cryoprotectant group is a big jump. Did he reach out to you first? Did you reach out to him first?

Hunter: Andrew and I go way back, so that was helpful. We used to race bicycles together actually when we were at Chicago. So that was helpful. And honestly, that was one of the benefits of having come up through a scientific background that put me in academia for a long time. Yeah. Is that I have the benefit of knowing who is very good from having seen them work before. And Andrew always struck me as someone who was both exceptional at physics, but also was deeply operationally minded. Very no-BS problem solver. And he was a natural early hire for the company.

Abhi: How much of your recruiting process is reaching out to people who seem hyper-talented in this orthogonal discipline to cryopreservation and saying, Hey, have you ever thought about working at Until versus them reaching out to you and saying, I want to turn my talents into this field I know nothing about.

Hunter: Yeah. I think that in the beginning it was almost all the former. Because we’re in stealth mode. And no one knows who we are. I think that we are slowly seeing it shift to be the latter where... Laura and I are getting cold emails from some thermodynamics expert who’s “Hey, I saw this.” Yeah. “And this is crazy. Can you tell me more about the glass transition in water?”

And I think that’s part of why I want to have conversations like this. This is part of why I think it’s my job to go around and talk about what we’re working on, is that I do think that there are likely very many intelligent physicists, biologists, chemists who would think this problem is interesting, but maybe don’t see themselves as being able to contribute to it meaningfully and I think I would want to make the vociferous pitch that is not the case.

Abhi: Okay. That’s an open call to everyone watching this to apply to Until.

Hunter: Indeed.

[01:39:27] What personality type do you most value when hiring?

Abhi: What, amongst all the talented people who end up in the recruitment pipeline for Until, is there a particular personality archetype that you think is most useful to have around?

Hunter: Yeah, I think that the first thing that I’m looking for is a highly agentic personality. And what I mean by this is a willingness to take on a high degree of responsibility for solving a really hard problem. I think that’s one. The other that I always look for is people who understand failure deeply. Because I think the one thing we pride ourselves on at Until is failing really fast. One of the things about building out a really wide tech platform, the way to do that effectively is if there’s an idea that seems good, design the experiment that proves that it’s bad, and do that at a very high rate. And I’ve found that people who have simultaneously a high degree of agency and a low enough ego to be able to design something to prove themselves wrong, those are the people who end up actually being able to, I think, move really quickly and contribute really meaningfully to the problem.

Abhi: You had, I’m not sure if I’m hallucinating this, but I think you mentioned your postdoc advisor or your PhD advisor had this one quote about, you should strive to treat everything as the same field or something along those...

Hunter: Oh, okay. Adam. Yeah. Okay. So I, this is... I have deep respect for the people that I’ve gotten to work for. I actually started... my first research job was with an ultrafast optics person, a professor at the University of Chicago, Greg Engel. Greg was great and gave me a job when I knew less than nothing. You’re an undergrad and you think you know something, so you’re actually less useful than a complete naive human being. And then in grad school, I got to work for Mikhail Shapiro at Caltech, exceptional scientist. And was the one that converted me from physics over to biology. And then my postdoc, I actually got my dream postdoc, which was to go work for this guy named Adam Cohen. And Adam is both an exceptional physicist and a great biologist and neuroscientist. And there’s this interdisciplinary way that his mind worked that just fascinated me. I had been following his publication history from even before he was a professor. And I’m looking at this guy as someone who had been able to traverse these discipline boundaries.

And so when I showed up, I asked him, “How do you do this? How is it possible that you seem to be pressing things all the way from single-molecule biophysics all the way up to solving neural circuits to building crazy microscopes?” I was an optics... I thought of myself as an optics expert and I’m looking at the microscopes that are in this guy’s lab and it’s completely blowing me away. The answer that he gave has stuck with me so deeply, which is that nature does not care. Nature doesn’t respect these boundaries. There is no such thing as biology to nature, or chemistry or physics. This is all just fundamentally the same at its core. And I think he and I share a physics background, so I think at his core it’s just physics at the core. With more scaffolding built up around it. And I think oftentimes when physicists say that, they’re trying to give themselves priority or something, but that’s very much not how he meant it. What he meant was these discipline boundaries are just purely human-imposed. Yeah. Yeah. So yeah, that has totally stuck with me. And Adam’s mentorship in particular was incredibly helpful in setting up for what I’m doing now.

Abhi: Do you think the only... the primary way one can imbibe that mindset is just to simply learn about many different fields? Or... I’m curious, if you were mentoring someone how do you encourage them to think about the scientific process?

Hunter: Yeah. I think that in terms of learning many different fields, first of all, just get comfortable with being an idiot about different disciplines, particularly early on because you have to be deeply, comfortable with the idea that if you’re going to move through a discipline with which you are unfamiliar, you will be a fish out of water in that space for a while. But I think that what... and this is one of the things that I particularly love about physics, is that it sets this beautiful foundation through which you can attempt to apply analogy to a whole host of different disciplines. And this isn’t always functional. This obviously breaks down in some very complex systems. But it gives you a language for thinking about the physical world. But I’ve also seen people who are very good biologists use their own version of analogy to try to understand nature. And have similar, conversations. Mandy, this person who I was talking about, who does a bunch of these organ perfusions for us.

She and I can have a conversation. And it’s interesting because we will find a way to find common language to describe something from completely different backgrounds. Yeah. So that’s how I think about it.

[01:44:17] Why work in cryopreservation as opposed to anything else?

Abhi: And you’ve mentioned you’ve worked in a lot of different fields. I went through your Google Scholar and I saw papers on biophysics, neuroengineering, bioimaging, molecular engineering, and nanotechnology. Why... you did mention this thing about your father and the field that Laura pitched to you as especially interesting, but I’m curious why start a cryonics company as opposed to any other company?

Hunter: Yeah, critically the stuff that we do, we consider slightly different than cryonics, which is a different process. But why start a cryopreservation company as opposed to any other company? I think a few things as we’ve gone over previously, I think one of them comes down to impact. I had actually not even thought about starting a company. I thought I was going to be an academic when Laura and I met. I was looking forward to starting my own lab, becoming a professor. I think that it was really a pull and not a push. I enjoyed doing academic research. But there was the sense of, oh, the scale at which we can have impact with this problem, the level of neglect that it has had in terms of being able to amass the amount of capital required to really go after it hard. These things were really contributing.

And then there’s this... and again, I learned this next part from Adam. Adam always applies a three-part test to figure out if he should work on something. So the first one is, do I find it interesting? Do I think that it’s something that I want to work on? Do I have the skills that are required to execute it well? And then the final one was the removal test. If you deleted me from the Earth, is someone else likely to do it? And I applied the same criteria here in my own way. And I think for me it passed those... that three-part test of should I pivot into this? I don’t think that I could have appreciated what it would look like when I did. I really was casting myself into an unknown. But yeah, that was fine.

[01:46:26] Until Lab’s competitors

Abhi: One thing I perhaps should have asked earlier is, I kind of view Until Labs as existing as perhaps one of the very few cryopreservation companies. What does the landscape of competitors look like? Do you view TransMedics, the current dominant organ supplier company, as competition? Or do you view other cryopreservation companies as competition?

Hunter: That’s a great question. I think that in the end, there’s going to be a marketplace for all of these technologies to coexist and there will be a question of which ones are used in what context. That’s all going to come down to what are the relative viabilities of all of these strategies. But yeah, there’s a whole host of companies that are coming up for doing near subzero storage. These store at minus five. There’s already established... there’s a lot of already established devices that do zero degrees up, like four degrees, hyper-oxygenated machine perfusion at cold temperatures. It’s very well established all the way up to normothermic machine perfusion. So this is at 37 degrees actively perfusing the organ during transplant. These all have different trade-offs. Some of them are organ-specific trade-offs. So yeah, it’s hard to say yet who our competition is because it’s hard to say yet who is going to be able to press those viability metrics as high as possible.

Abhi: If you view what Until is working on as, you can store an organ infinitely You can leave it for centuries and it’ll come back just as fine. Yeah. And the very top is, the organ needs to immediately go into a patient, otherwise it will go ischemic and just die. Yes. What does the time look like for each? For putting an organ on just pure ice. How long does that organ last versus if you go a little bit up or down the hierarchy, how long does it last?

Hunter: Yeah, so it’s obviously very organ-dependent. Okay. And so for something like a kidney, kidneys are incredibly robust actually to ischemia. So you can keep a kidney on ice for 72 hours and still it will think back up and recover. If you go for tissues like the lungs, that is not the case. You have, a few hours to be able to get it transplanted because of really short ischemic windows. So it’s going to be a very organ-specific question.

Abhi: And even if you have this hypothermic blood diffusion, it’s still max 72 hours.

Hunter: Yeah. I’m not actually sure if there’s an established machine for hypothermic machine perfusion of lungs. If there is, I’m not familiar with it.

Abhi: Oh, even for kidney?

Hunter: Oh, for kidney?

Abhi: Yeah. You’re saying it’s extending beyond 72?

Hunter: So I think that there’s no... yeah, I think there’s not a ton of research on pressing kidneys out beyond the 72 hours. There is, but in terms of established places where people are really looking hard at perfusion, it’s mostly actually liver is something that’s had quite a lot of perfusion work done for it. But all these things, you’re trying to press out those time windows to make it progressively longer. Yeah. But in the end, if you get to vitrification, then that just... you take time completely out of the equation. Yes. Yeah. Yeah.

[01:49:30] What would an alternative universe version of Hunter worked on?

Abhi: That makes sense. I’m curious what... yeah, you’ve worked in a lot of different fields. What would an alternative Hunter have done, if not for cryopreservation?

Hunter: Yeah. It’s a challenging question because it was such a sliding doors moment for my life that sometimes it’s hard to look back and think about it. But I guess in a complimentary world, things that I was interested in at the time that I was working with Adam. A lot were mostly revolved around two questions. One was advanced sensing and the other was neural computation. I think I was drifting very much towards advanced concepts of sensing. So something I did in my PhD was building out small-scale devices to look at really tiny magnetic fields inside of cells.

Which was mostly an academic study at the time that I did it. But I think that there are some interesting downstream applications for being able to do something like MRI on a single cell. To try to look at diffusive transport of tiny little molecules label-free inside of a cell. I think that this is one of those things where it was beautiful and academic and I think that the relative impact was not quite as high as what I could be doing outside. Yeah. And I think it’s a really natural fit and I feel really blessed to have had it walk into my life.

Abhi: Do you think if this Until Labs opportunity did not pop up, you would’ve felt pretty content not doing a ton of translation work throughout the rest of your career, or you think there was always something in the back of your head thinking, I should turn this into something that’ll reach a patient?

Hunter: Yeah. I think that was one of the things that was always challenging for me in academia actually, was that there was always this biting thing in me that was a need for a large impact. I think all of us want this, right? Sure. All of us want to make a larger impact. And I’m not alone in being an academic that’s I wish that I could find some way of translating this work. I also think a lot of academics do a really good job of eventually finding that in their career.

[01:51:33] What would you do with $100M?

Abhi: But yeah, that opportunity was just presented itself to me directly.

Yeah. That makes sense. And I think that perhaps the last question I have is if someone handed to you a hundred million dollars equity-free. You can spend it on Until or you can spend on some other scientific field that you’re very interested in. Yeah. Where would you allocate the money?

Hunter: Right now? Right now I’d allocate it directly into Until. Okay. And I’m not saying that as a cop-out. I actually think that having seen the interior of the organization, it’s like... for me, this is the obvious place where we can make rapid advancements for humanity.

And yeah, that’s where I would want to spend it. I think that in terms of where we would allocate it inside of Until, I think parallelizing more to be able to go after more kinds of organs and get these things to clinic faster. Because I think there is some real urgency to try to... there are a hundred thousand people right now in this country waiting on organs on a waiting list. And I think that I have... my personality is very, let’s say I have a bias to urgency.

Abhi: When it comes to actually using the hundred million dollars, what is the primary bottleneck that the money will be used to help solve?

Hunter: Yeah. So I think that the primary bottlenecks are a couple fold. One is in vivo experiments are hard. It’s not even in vivo experiments, organ experiments are complex, require lots of people and are capital intensive. And to just... in the end, some of these things don’t translate unless you scale up to there. So it’s a deep focus on that translational research. I also think that it allows us just to explore more of these fundamental questions. Work with academics to explore more of these fundamental questions, and build out the foundation that actually predates Until substantially around some of these more fundamental scientific questions, to try to figure out what is the best avenue for us to build off of as we go forward. Because I think that even the solution of doing this on isolated organs does not solve the long-term problem. I think that the capital injection is also quite helpful for helping us lay a foundation to be able to eventually do this on an entire organism.

Abhi: Sure.

Is the... does it make sense to say that... the way that you’ve explained this sounds like both sides of... you need people to figure out the theory for everything and then you also need a small army of RAs to actually do the experiments.

Hunter: Yep.

Abhi: Would you allocate the money equally or you think just the empirical animal studies are infinitely more valuable than people theorizing?

Hunter: I think that it’s a weird cost equation because it’s I wouldn’t say that one is more valuable than the other, but empirically one is much more expensive than the other.

Abhi: Yeah.

Hunter: That’s... and it is simply the case that doing preclinical model experiments is much more expensive.

Abhi: Okay. Cool. Hopefully someone gives you a hundred million dollars equity...

Hunter: I think we’re pr... I think we’re pretty good on capital for a while and thanks. If you’re showing up with equity-free checks, I’ll take them.

Abhi: Okay. Thank you so much for coming onto the podcast, Hunter.

Hunter: Thank you.

Abhi: Okay. Think we’re good.

1. Introduction

2. Timestamps

3. Transcript

Youtube:

Spotify:

Apple Podcasts:

Introduction

Ellen Zhong is perhaps one of the only people in the ML x bio field to have created an entirely new subfield of research during her PhD: the application of deep-learning to cryo-EM particle images.

If you aren’t familiar with that field, I luckily have a 8,000~ word article covering it, which walks through a lot of Ellen’s papers. If you don’t have time to read something that grossly large, the general breakdown of the problem is as follows: cryo-EM can give you thousands of 2D views of a 3D protein from many different angles, from that data, can you discover what that 3D structure is?

Ellen, who is now a computer science professor at Princeton University, has spent her academic career investigating that question, and now has an entire lab at Princeton (E.Z. Lab) focused on that and related ones. Including, as the title mentions, the possibility of performing cryo-EM structure determination at ultra-high scales.

In this podcast, we talk about her research, what she did during her recent sabbatical at Generate:Biomedicines, her recent interest in areas beyond cryo-EM (cryo-ET and NMR specifically), and more!

Timestamps

[00:00:00] Introduction
[00:02:43]  What does it mean to apply ML to cryo-EM?
[00:04:28] Ab initio reconstruction and conformational heterogeneity
[00:15:41] Can we do multiplex cryo-EM structure determination?
[00:22:19] Datasets in cryo-EM
[00:26:25] Why isn’t there a foundation model for cryo-EM particle analysis?
[00:33:07]  How much practical usage is there of these cryo-EM models amongst wet-lab cryo-EM researchers?
[00:40:34] Where can things still improve?
[00:46:57]  Has deep learning done something in cryo-EM that was previously impossible?
[00:48:22] Ellen’s experience in the cryo-EM field
[00:53:40] Deep learning in cryo-EM outside of structure determination
[00:57:32] 3D volume reconstruction versus residue assignment in cryo-EM
[01:00:26] What did Ellen do during her sabbatical at Generate Biomedicines?
[01:07:07] Ellen’s research in cryo-ET
[01:13:54] Ellen’s research in NMR
[01:21:05] How did Ellen get into the cryo-EM field?
[01:26:57] Why did Ellen go back to graduate school?
[01:32:17]  What makes Ellen more confident about trusting an external cryo-EM paper?

Transcript

[00:00:00] Introduction

Abhi: Today I’ll be talking to Dr. Ellen Zhong, a computer science professor at Princeton University. Ellen’s research focuses on applying machine learning to protein structures derived using cryogenic electron microscopy, also known as Cryo-EM. In fact, she was one of the first people to ever apply deep learning here, and her lab at Princeton remains at the forefront of this field.

Today we’ll be talking about her research, what the future of protein structure determination looks like, and her recent sabbatical at Generate Biomedicines. Thank you for coming onto the podcast, Ellen.

Ellen: Yeah, thanks for having me.

Abhi: So my first question is, I gave a very brief overview of what your own personal research interests are, but I’d love for you to give your own take as to what you’re most interested in these days.

Ellen: Sure, yeah. So, right now I’m a professor of computer science at Princeton University and our group works on various molecular machine learning problems. So really interesting problems at the intersection of AI and biology, and specifically we focus on structure determination. So can we actually analyze actively collected experimental data, in particular cryo-electron microscopy or Cryo-EM, to solve the actual 3D locations of the atoms of these macromolecular machines and be able to understand how they work and discover new kinds of proteins and their dynamics.

Abhi: When you first got involved in this field, like back in your PhD days... it seemed like you were the first person to ever take Cryo-EM particle images and apply deep learning there. When you were doing that back then, did it feel like you were taking a really big risk or was it... you were kind of sure of yourself, like this was the right direction?

Ellen: I guess at the beginning of my PhD I worked on a couple of different things and then I learned about the Cryo-EM problem, and I thought this 3D reconstruction problem was really interesting. And in particular, this problem of reconstructing continuous motions was unsolved at the time.

And actually I was sent this paper by Roy Lederman and Amit Singer, who are mathematicians at Princeton University, about this mathematical introduction to this problem. And after reading that paper, I was like, “Whoa, this is a really cool problem.” And I guess I didn’t know at the time that the rest is history, but it was very much like, “This feels like a problem, a very juicy problem that I can work on for the next five years or something like that.”

And I didn’t really think about anything else. But when I, at the time, was learning more about modern machine learning techniques, it seemed very natural to apply these deep learning models, which can learn super complex distributions from large-scale datasets, to this Cryo-EM problem. But yeah, it was just super fun.

[00:02:43]  What does it mean to apply ML to cryo-EM?

Abhi: When it comes to actually applying machine learning to Cryo-EM particle images, I have a vague sense of what that means. Would you be able to give a brief overview of what it actually means to take the Cryo-EM particle images you get out of an electron microscope and throw ML at it?

Ellen: Yeah. So the specific problem that was the focus of both my PhD work and is very much a central research question in our group right now is this 3D reconstruction problem. So the electron microscope, you take a bunch of 2D projection images of your vitrified or flash-frozen biological sample.

So you have a bunch of 2D pictures of your protein. And then there’s the algorithmic challenge of reconstructing the 3D structure from all these 2D particle images. And so it doesn’t have to be approached with deep learning approaches. There’s a lot of classical algorithms.

There’s state-of-the-art tools that are not using neural networks to do this 2D to 3D reconstruction problem. And the central computational challenge is inferring both the 3D structure, but also the unknown camera poses and orientations of each of the molecules. And the specific new challenge we’re interested in is what happens when the different proteins are in different conformations.

So not only do we want to reconstruct the 3D structure, like a static picture of the protein, but we want to reconstruct a movie of the dynamics of these macromolecules. Which, you know, to me that seemed like the coolest thing, right? Structural biology was... it’s already so interesting to be able to see these 3D structures, but can we actually see how they work and how they function from experimental data?

[00:04:28] Ab initio reconstruction and conformational heterogeneity

Abhi: I’ve written about Cryo-EM before, and it seems like there were three distinct phases to the field, at least the part of the field that you work in, where at the very beginning you’re focused on ab initio reconstruction, where you have no prior knowledge as of what the actual protein structure looks like.

Then conformational heterogeneity, where there are multiple conformations in the particle images. How do you tease out all the possible conformations? Before I get to the third one... for ab initio reconstruction, is that a solved problem? Like you no longer need to have any prior knowledge?

Ellen: For well-behaved samples. So, because this is always collecting active new experimental data, you need to solve this 3D reconstruction problem for every single protein. And so for these new complexes that people had no idea what they looked like before, there’s so much upstream work both to make sure the sample is expressing well and is well-behaved, make sure the ice is very thin and you can have high-quality images.

And then there are existing methods for ab initio reconstruction. And so assuming that your input data is reasonable, then you can do this pose estimation in 3D reconstruction to get the structure. But that’s very much still an open problem.

Abhi: Is it an open problem in the sense of like sample prep can mess it up? Or is it an open problem in that, even if the sample prep is perfect, there are some proteins which, like CryoDRGN, which is the method you created to solve ab initio reconstruction... it still doesn’t work.

Ellen: The... I would say both. So like, definitely if the sample’s not good, the algorithms are not going to solve that for you. But the part where it becomes interesting whether you can jointly design the experiment and the algorithms are when you have lots of dynamics in your protein or maybe a complex mixture of multiple things. If you just have a very purified protein that’s very much like a rock or something like that, or a ribosome, which is one of the canonical examples in Cryo-EM, then it is shockingly easy and absolutely fantastical how we can solve the structures of these molecules in like... in like a day.

Abhi: Before I move on to conformational heterogeneity, one thing I wanted to ask was, I saw that you posted on Twitter recently about CryoBoltz, which is trying to use existing protein structure predictions... it’s almost as a starting point for the Cryo-EM modeling. If you have that, do you even need ab initio reconstruction, or are the prediction models are actually pretty good and...

Ellen: Yeah, that’s like another super interesting question for the field right now. So I feel really lucky that currently we have AlphaFold 2, AlphaFold 3, all these really powerful protein foundation models that can just predict structures from sequence. So like, why do we even need to do experiments?

And the answer is, okay, these prediction models are just predictions. There’s a lot of proteins where you can’t actually... a lot of specifically large complexes, which are the functional machines in vivo that do these essential biological processes. So for those you can’t predict the entire structure. If we want to understand function, if we want to actually be able to see what’s going on and the mechanism of these essential molecules, then we still need experimental data. And especially for training the next generation of these models, we’re going to need to understand what are the interesting actual ensembles or dynamics and larger complexes to be able to have these structure prediction models reach the next level.

Abhi: Okay. I’m going to have more questions about that later, but... But moving on to conformational heterogeneity, that was the subject of CryoDRGN 2... which released in 2021, is that correct?

Ellen: So the... yeah, I should give a whole talk or something like that. But the Cryo-EM reconstruction... so CryoDRGN is our model for reconstructing... it’s a VAE-based neural model for reconstructing these distributions of structures. And at this point we have so many extensions of CryoDRGN that tackle more challenging settings of the problem or in situ data with Cryo-ET, but in our group, we’re just all calling it CryoDRGN. Okay. To have it in one software package that people know where to access the model and where to use the method.

The original CryoDRGN paper... the original machine learning paper was very much trying to tackle the whole enchilada of ab initio reconstruction of complex mixtures or of distributions of structures. And then when we were creating the software tool for the actual structural biology community, then I realized, “Wait, wait, wait. That’s actually too hard.” And already if we just tackle this conformational heterogeneity problem where you assume you already have the camera poses and you just want to infer residual heterogeneity, we can already discover interesting things like missed structures, like new conformations, continuous motions, which was the original motivation for the method. That was CryoDRGN, I guess V1. In CryoDRGN 2, we revisited this ab initio reconstruction problem...

Abhi: Oh. So conformational heterogeneity came first and then ab initio came second?

Ellen: Yeah. Okay. So the original... well, the original ICLR paper in 2019-2020 was trying to do everything. Okay. So it was like... I guess at the time I didn’t know what the scope was for a machine learning conference paper. I just want to solve everything. Yeah. So that was doing ab initio heterogeneous reconstruction. And then later when we developed the tool, we just focused on the conformational inference. And then I guess in my research group, in the last couple of years, we’ve released an extension, CryoDRGN-ET, to Cryo-ET data and CryoDRGN-AI, which is ab initio... not artificial intelligence. And those two methods are designed for modern Cryo-EM datasets to actually tackle ab initio reconstruction of super large, complex Cryo-EM data.

Abhi: So, yeah. Okay. I guess I mixed the two up, in terms of timeline for the conformational heterogeneity bit, is that entirely solved? Like given a protein that is very flexible and displays many conformations, if a Cryo-EM person took images of that protein, applied the CryoDRGN software package to it, could they reliably grab out all the conformations or are there still ongoing issues?

Ellen: I would say, yeah. None of these problems, I would say, are solved. Okay. Right. These are all very much open challenges that any new dataset could present completely new challenges. There are now... I mean, it’s an exciting time because there’s a lot of new methods, both neural network-based and also classical signal processing-based algorithms or computational numerical linear algebra methods to tackle this heterogeneous reconstruction problem.

But yeah, I would say one of the main challenges is that this heterogeneity problem in Cryo-EM is not super well-defined, right? It really depends on what the practitioner, what the structural biologist wants at the end of the day. Like, do they want an ensemble of structures? Do they want an ensemble of atomic models, which is what CryoBoltz is trying to do? Is this atomic modeling part of...

Abhi: Sorry, what? Like in my head, the output of Cryo-EM... the output of this is one singular ground truth. What is the full ensemble of truths out there?

Ellen: So I break down this Cryo-EM problem into... there’s a couple of stages of image processing. There’s the pre-processing of the raw micrographs, and you know, there’s a lot of steps to that. And then there’s this 2D to 3D reconstruction problem. So that’s already with segmented single particle images in 2D to resolve the 3D structure or an ensemble of structures. Those structures... so “structure” is an overloaded word... in the Cryo-EM setting, that means a density volume.

Abhi: Okay. Right.

Ellen: So it’s just the electron scattering potential of the molecule as felt by the electron beam in the cryo-electron microscope.

And then assuming you have a sufficiently high-resolution volume that’s not filled with artifacts or too low resolution, then typically what’s done is you manually build in the atomic model. As of recently, now there’s some tools based on these modern deep learning and protein-based models that can do atomic modeling, but it’s still very much a manual art to actually placing the atom locations to then deposit into the PDB.

Abhi: That was surprising. I kind of always thought... I guess it makes sense that the output of Cryo-EM is an electron density field, because what else could you possibly gather from shooting electrons at something and seeing what comes out on the other side on the electron detector. But I also assumed that assigning atoms is a trivial process. Why is it... intuitively, why is it hard?

Ellen: So it’s very trivial if the volume is high resolution. Yeah. So if you have an atomic resolution volume, then you can just see where the nuclei are and place the atoms. But that’s extremely hard. And for most, a high-quality Cryo-EM structure will be maybe three angstrom resolution, which is enough to resolve maybe secondary structure features and some side chains. And then based on our existing prior knowledge on the geometry of protein structures or nucleic acid structure and composition, then you can mostly unambiguously place the atomic models.

But also, this is just a messy, hard scientific problem. The structures are not necessarily uniform in resolution. The parts that are moving will be lower resolution because we’ve blurred it together in the standard reconstruction problem.

Abhi: So like if there’s a bit floppy over here, it’ll just smear the electron density. Yeah. Okay.

Ellen: Yeah. You’ll smear it together, assuming you’re doing a homogeneous 3D reconstruction. So the whole promise of CryoDRGN and these heterogeneous methods is that we can actually model the ensemble and then we’re not averaging together all these different conformational states.

[00:15:41] Can we do multiplex cryo-EM structure determination?

Abhi: Yeah. So, we’ve discussed ab initio reconstruction and conformational heterogeneity.

Ellen: There’s a lot of jargon.

Abhi: One of the... I think one of the craziest things you’ve written about was a paper at the end of NeurIPS 2024 called Hydra.

Ellen: Yeah. Multi-headed dragon.

Abhi: Yeah. Which is a method that... I think the phrase used in the paper is “compositional heterogeneity,” where you have multiple proteins on the same micrograph, and you are trying to multiplex the structure determination of all of them at once.

Ellen: Yeah.

Abhi: This was back in December 2024. To me as an outsider, that just feels clearly extraordinary. You’re able to 2x to 3x how many proteins you can shove through this incredibly expensive and manual process.

Ellen: Yeah.

Abhi: Is it as revolutionary as it seems on face value? Like how much work is there?

Ellen: Yeah, so it’s a very hard problem and that’s very much still what we’re working towards. Like, can we do high-throughput structure determination, not only of ensembles and these dynamic atomic models, but just shove multiple proteins into the sample, into the electron microscope, and then simultaneously solve their structures?

So that’s very much something that we’re still working towards. Hydra, we demonstrated maybe three or four different structures on real data at the same time. And it just becomes a much more challenging optimization problem with more and more proteins if we’re using classical-based approaches where we’re trying to infer both now the identity of the protein and the pose or the 3D orientation relative to the camera pose... the orientation of the particle within the ice and also the conformational state.

So I think there’s probably some inherent limit to these classical-based approaches, and so very much as a moonshot in the group right now, we have a bunch of new projects that are like, “What happens if we want to solve a thousand structures at once? What happens if we want to analyze a cellular lysate or a Cryo-ET sample, which is like a slice out of a cell, and actually solve all of the...

Abhi: Dozens of biomolecules in there, maybe hundreds?

Ellen: And that’s... I would say, it’s both a good thing and a bad thing, but it’s very much still a moonshot, right? And it’s very exciting times.

Abhi: In the paper that was at NeurIPS, there were three proteins determined. Is three as of then kind of the upper limit, or is it more like we haven’t tested up to five or 10 yet and we just chose three?

Ellen: I would say it very much depends on the experimental sample. So definitely people are doing structure determination from native extracts and things like that... in actual structural biology papers, and using classical-based reconstruction methods you can simultaneously solve a handful of different structures.

The unfortunate... the status quo though is that it’s very manually driven and it’s very... an expert structural biologist or cryo-electron microscopist is going in and arbitrarily subsetting your dataset to find reasonably pure classes where it’s mostly just one complex or not. And it becomes this very user-driven process.

And so even in Cryo-EM papers where you’re focusing just on a single structure, if you look in the supplement and you look for one of the data processing figures...

Abhi: Sorry, what’s SI?

Oh, supplementary.

Ellen: Yeah. If you look in the supplement, there’s always going to be one of these figures that show the image processing pipeline, and it’s just this crazy flowchart of all the different steps and all the different user-chosen subsets that are taken to get to the final particle stack.

And so, very much, one of the interesting computational challenges is, can we automate that process? Can we make it either one-shot from an algorithm that is trying to do all of the processing or the optimization in one go, or, these days, can we have an autonomous or an agentic approach to tackling it? So there’s so many interesting directions right now.

Abhi: What is currently the... if I just come at it from an incredibly naive perspective, if the problem is, “Oh, there are so many... let’s say you’re trying to do a hundred proteins in one shot and there are so many different conformations each of these unique proteins have, how do you subset them into classes?” My instinct is just, “Oh, expand the micrograph and have more shots...”

Ellen: Just collect more data.

Abhi: Yeah, just collect more. Is that... is it just... yeah, what’s the primary bottleneck in getting to a hundred-shot Cryo-EM?

Ellen: Yeah. I think the number of images that we’re taking right now is... people can collect a lot of data. The microscopes are becoming very automated. So I would say data collection is not necessarily a problem. I think people are not actually actively working on it from the experimental side, which is interesting.

And there are also experimental challenges too. There’s computational challenges. You really need to do both. I’m excited to do that. And actually along these lines, my group and the Flatiron Institute and CCP-EM, which is this structural biology consortium in the UK, are putting together this challenge, we’re calling it CAHRA: Community-wide Assessment of Heterogeneous Reconstruction Algorithms, where we’re collecting datasets of complex mixtures and just seeing how people do with existing methods or workflows.

Abhi: Up until Hydra came out, or maybe methods like it, did datasets of heterogeneous mixtures not exist?

Ellen: No, I do think that’s one of the challenges for the field is people are just working on their bespoke protein complexes and then they deposit the data, but it’s not targeted towards algorithms development or pushing the capabilities of maybe extreme compositional heterogeneity. That’s something that we want to work towards.

[00:22:19] Datasets in cryo-EM

Abhi: When people deposit their data in the Cryo-EM field, do they deposit... obviously they put the electron density structure on the PDB... or is there some practice of also giving the ice crystal images?

Ellen: Yeah, so we deposit... people deposit both the atomic models and the density volumes. And something that is common but not ubiquitous is depositing the raw electron microscope data to EMPIAR, which is another publicly accessible database. And that’s where things get a lot messier. ‘Cause the raw data is usually much, much larger, like terabytes of raw imaging data. And the processing is challenging. And so... I think that’s a major challenge.

Abhi: So even if we lived in a world where everyone deposited their raw data onto these platforms, even then it would be such a headache to deal with the data that it’s not necessarily clear that it would translate to all that much utility.

Ellen: Yeah. And also, yes, both that, and there’s no standard metadata and things like that. So there’s maybe just some system-wide logistical challenges and the fact that the data quality is very heterogeneous. So sometimes you have really high-quality micrographs and other times it’s not that great.

Abhi: How much is the generation of the micrograph almost a skill issue on the person who actually generated the data versus... there are just some proteins that are very difficult to characterize well, and the data will always be low quality.

Ellen: I would say both. Okay. Right. Yeah. This is why experimental wet lab biology is just wild and really challenging. It’s definitely both. The hands matter, right, in terms of making the sample. And then it’s totally the case that some protein complexes are just really sticky. They adhere to the air-water interface. The images just look weird and, you know, there’s a lot of sample preparation challenges that are very system-specific. So in that way, it’s similar to X-ray crystallography where you just need to do brute force guess-and-check to get your protein to crystallize.

Abhi: Do you think we’ll ever live in the universe where Cryo-EM and maybe NMR will be the only determination methods that anyone does? Or will there always be a place for X-ray?

Ellen: Oh, I think there’s always going to be a place for X-ray.

Abhi: Okay.

Ellen: Yeah. I think Cryo-EM is very good at large complexes and the sample prep is easier ‘cause you don’t need to crystallize. But if you want to do rational-based, med-chem-driven, structure-based drug discovery, you probably just want to crystallize your protein target with your small molecule.

Abhi: Oh, there’s no... even if we had the greatest whatever version of Cryo-EM in the world, it probably would not be able to resolve a small molecule.

Ellen: It’s not the best for small things because the smaller the target, the less signal you’re going to get... because the less electron scattering you’ll get from the microscope.

Abhi: Has there ever been a paper that has tried to actually detect whether you can find a small molecule from a Cryo-EM image?

Ellen: Yeah. So you can definitely solve small molecule structures. It just becomes literally exponentially harder. Okay. Because you need more and more data to get higher and higher resolution... because you’re just fighting against the noise and there’s an exponential decay in the signal as a function of resolution or frequency. So you just need more and more data to solve higher and higher resolution structures. So if you really want to see atomic-level detail of small molecules, you just need a ton of data.

[00:26:25] Why isn’t there a foundation model for cryo-EM particle analysis?

Abhi: Okay. So we went through ab initio reconstruction, conformational heterogeneity, compositional heterogeneity. I think an interesting quirk of perhaps all of the models that you created is that there is not a single universal pre-trained model for any of this. They’re rebuilt for every new protein that comes in. Why? This just feels alien compared to most machine learning problems. Why was it initially formulated that way when you first came up with the idea and why has it stayed like that?

Ellen: Yeah, I think that’s a really interesting question. And it is very different from the rest of bio-ML, right? Where you have a large, maybe pre-trained model that you can now... like AlphaFold, you train on the PDB, and now you can use it to predict new structures. And in Cryo-EM, and these CryoDRGN-based models, we’re always training a new model from scratch. So when people are analyzing data, it’s some structural biologist who’s training this deep learning model, which is super cool, to analyze their data.

And there’s two things. One is... we’re always... the problem is that we want to infer the structure. We want to infer... we want to solve this inverse problem of what is the actual structure from the experimental data. So it’s very much this active problem of inferring the signal from the experimental data and...

Abhi: I guess instinctively, surely there is translatable information from one protein structure determination problem to another. Has anyone tried to build a pre-trained foundation model for all of this? Or maybe they have and it just doesn’t work that well in practice?

Ellen: Yeah, I think there’s definitely aspects of the problem that make it really hard to train a general model. The first is that the images are extremely noisy. And so we’re already trying to do this very hard inverse problem of what is the 3D structure given these noisy 2D images. And traditionally, the field has been very averse to bringing in prior knowledge because then you can really easily trick yourself.

Abhi: Oh yeah. Okay.

Ellen: And the thing we’re interested in is solving the *new* structure. Right. We’re not interested in pattern matching existing structures. And so, because of the high amount of noise, you can really easily just overfit to noise, align all the noise, and get whatever you want from the data.

Abhi: But I would imagine... yeah, like pre-training people would’ve said that also pre-AlphaFold. Is it... how real is the paranoia that you may potentially hallucinate something as part of the electron density that doesn’t actually exist?

Ellen: So I think it comes back to what is exactly that you want to get out of this experiment. Like, what is it that you want to get out of this multimillion-dollar microscope. And there’s definitely... I do think something our group is working on are these generalizable models that can actually learn across datasets. It’s a very hard problem. It’s unlike any other domain of machine learning. So I think it’s interesting to work on.

It has all these caveats and disclaimers of, “Okay, if you do this, the prediction from maybe this AlphaFold-informed reconstruction model... is it actually the structure from the data? Or is it a structure from this extremely knowledgeable prior that knows everything about proteins that currently exists in humanity?” And I think it’s more interesting to be sure that this is what’s in your data.

Abhi: That’s fair. This is so expensive, so time-consuming that whatever the output of it has to be correct.

Ellen: Right. And like, why else are we spending millions of dollars on these microscopes? Like just tell me where the atoms are. Don’t tell me what AlphaFold already knows. And I think from the discovery, from the actual scientific perspective, that’s where we can discover new things. Like new things that we don’t know currently about structural biology or protein structure, or maybe antibody CDR loops... things that AlphaFold can’t predict or these structure prediction models can’t model that well.

So the whole promise of structure determination is to be able to determine these new structures. And so the generalizable foundation models for Cryo-EM, I think, would be very useful for speeding up the process and would be very useful as... I think it’s an interesting research direction, but I’m very much of the mind that we want to let the data speak... and to let the data show experimentally what’s going on with these proteins.

Abhi: I’m not sure how much you’d be able to talk about what you think the next generation of methods would look like. But if you have any insight into where things will go beyond Hydra and CryoDRGN, I would love to hear your thoughts.

Ellen: Yeah. So I do think it’s interesting right now that there’s a lot more methods for heterogeneity and there’s a lot more deep learning approaches that have different kinds of architectures and different kinds of inductive biases for the type of heterogeneity or things like that. I do think at this moment, this heterogeneity problem is very heterogeneous. And so it’s kind of dependent on what people want to get out of the experiment.

Abhi: Sorry, what do you mean by that?

Ellen: Like, do you want... right now there’s all these different... now we have this potpourri of different methods. You can get atomic... like maybe dynamic atomic models. You can get simpler, linear subspace methods that will give you maybe larger-scale motions. So I guess at this point, what I tell the Cryo-EM community is that now people have a lot of choices, right, in terms of what reconstruction methods they want to use. And so, it’s on them to understand what the priors are in these different methods. Like, is this a method that only models conformational heterogeneity? So if your data has any compositional heterogeneity... so these are mouthfuls... then you’re going to...

Abhi: it’s easier to shoot yourself in the foot.

Ellen: Yeah, so maybe we’re in this risky time. I don’t know.

[00:33:07]  How much practical usage is there of these cryo-EM models amongst wet-lab cryo-EM researchers?

Abhi: Maybe that lends itself well to the question I had, especially while I was writing a Cryo-EM article. How much use do these methods have today? Amongst non-machine learning people who... they don’t really believe in machine learning at all. Their job is pure structure determination. How much do they use stuff like CryoDRGN?

Ellen: Yeah, that’s a great question. So I’m always delighted when people are actually using these methods. And so there’s the Cryo-EM methods community and machine learning community which are actively working on new methods, and that’s maybe more cutting edge. And then there’s this much, much larger community of actual practicing structural biologists. And I would say there’s definitely overlap in the methods, but it takes longer time to adopt.

And something that I am very satisfied with during the whole CryoDRGN development was actually the educational part of... when we’re training these deep learning models and learning this low-dimensional latent space describing different conformations, like what does that even mean? Right? So there was a period of just going on a speaking tour and talking to a bunch of research institutes and research groups to tell them just how these methods work and that they’re not magic, they’re not hallucinating things. There’s art to training them for sure, but how do you even interpret the outputs?

So, I would say they are used, but there’s for sure a challenge in terms of understanding the outputs of these models and really taking advantage of them to extract all you can from your data.

Abhi: I don’t know too much about the culture amongst Cryo-EM people, but I know a little bit about the culture amongst medicinal chemists and it seems like molecular simulation is really useful in that area. But if you are doing molecular simulation in medicinal chemistry, you’re usually a pure computational chemist who thinks and breathes computational chemistry. And if you’re very much into doing things in the real world, you don’t know how to do molecular simulation at all. So there’s one group at the company devoted to just pure molecular simulation, another group devoted to giving results to the molecular simulation people. I guess this is all to set up the question of: amongst the people who are actively users of these models, of these techniques, how often is it that they need to deeply understand machine learning and that they are kind of brought up in the culture of computation?

Ellen: Yeah. I think there’s two ways. I think this bifurcates. If the methods get good enough, if the models get good enough, then it becomes truly democratized, right? And then you don’t need to understand how they work. And you can just black-box it, use the methods, use AlphaFold, look at the outputs, and as long as you know enough about how much to trust it, then it’s great.

And I think that’s for sure a direction that’s worth going towards: having robust enough models with reasonable defaults that people can just run and not need to bother themselves with all the details of how it works. On the other hand, if you understand how these models work, you can get so much more out of it, right? You can debug, you can really get a lot more out of your data.

And so I do think right now... I hope that people are in this place, and maybe these days with AI-based tools it becomes a lot easier to learn about all these different areas that one needs to be adept in to use these models. But I think Cryo-EM structure determination is still very much an area, it’s a niche where you need to deeply understand so many different aspects of this pipeline vertically, and if you do that, you have this superpower. And I think that’s still very much where we are. And I think maybe that limits the accessibility of the method, but it’s where you can actually make magic happen.

Abhi: Yeah, What do you view as the biggest barrier to adoption? Is it entirely just education, the models getting better, or some secret third thing?

Ellen: I think it’s definitely... usability is huge and also interpretability of the output.

Abhi: Aren’t the outputs immediately interpretable because it’s speaking the exact same language as the practitioner? What’s the interpretability issue?

Ellen: I think the... I mean... yeah, this is... so now that we have... now that we live in this world where you can reconstruct these molecular movies... the challenge is now what? Like, what do you do with the molecular movies? How do you actually then analyze the ensemble? And depending on the question that the structural biologist is asking, then maybe they just need the two end states or something. Or maybe they actually want the movie. But I think this is actually where it becomes perhaps overlapping with the MD community just in terms of the way they analyze the ensembles. But how do you extract the insights from the distribution of structures?

Abhi: Yeah. I mean, I guess you alluded to this earlier about how there’s currently a ton of methods each with their own biases and failure modes. Do you think we’re slowly tending towards an AlphaFold-esque or AlphaFold 3-esque era where there’s a single model that does it all and it’s push-button?

Ellen: Maybe the field is moving a little bit towards that because of the software packages that exist. So I guess the other thing about whether people use the models or not, and I think this is probably true for computational biology in general, is if you want people to use your methods and tools, they have to be easy to use.

Abhi: Sure.

Ellen: Just from an actual UI perspective and software design perspective. And so right now we’re trending towards these universal tools based on the ones that are actually easy to use... for people who don’t know how the command line works and things like that.

Abhi: Do you think where the CryoDRGN software package is today is... people who have never used GitHub before can just use it just fine?

Ellen: Hopefully. I think that was definitely something that I cared a lot about is the usability of the tool. And I have gotten good feedback on it, but, you know, it’s still a command-line-based tool. It’s still training a deep learning model. And I guess on my end, I had to learn so much Cryo-EM to get into Cryo-EM. There’s so much jargon in this field. There’s so much to know. And so... yeah, I think it’s a good life skill to force these structural biologists to learn command-line-based tools... and to learn maybe Jupyter Notebooks or Colab-based things. And so I think this is a win-win.

[00:40:34] Where can things still improve?

Abhi: Are there some niches even within Cryo-EM that there are no useful deep learning tools available for today? And it’s very much like the art that... whatever they’ve been doing for the last 30 years, it’s probably the best thing.

Ellen: That’s an interesting question. You know, neural network-based models have definitely taken over a lot of parts... have eaten a lot of parts of that computational pipeline, but definitely not all. I think the image processing pipeline, there’s a lot of different stages and a lot of them are just simple function-fitting classical tools. And there are parts where neural-based methods have... particle picking is one where it’s still very much a hard problem to identify the protein within the larger micrograph, the larger field of view.

Abhi: Like, what part of this ice image actually contains the protein?

Ellen: Okay. Yeah. So when you put your sample in your microscope, you take a picture. And so the picture is called the micrograph and there’s, depending on the concentration of your sample, all the individual particle images of your protein floating around in the ice. And then you have particle picking algorithms, basically segmentation algorithms, that identify... that detect the locations of the particles and segment them out for this 2D to 3D reconstruction problem.

And that’s still a very hard problem. You know, you have... very immediately there were these computer vision-based, CNN-based tools to particle pick. There’s still classical, just cross-correlation template matching tools that are used. And, you know, depending on that particle picking algorithm, do you see the rare views, do you see the rare conformations? And that’s still a hard problem. And so people are using all these different tools, but definitely just going back to the data a lot to reprocess and...

Abhi: Well, I guess the way that you’re phrasing it right now makes it sound like the simpler methods work fine. Is that the takeaway? That deep learning or the inclusion of deep learning here has not yet demonstrated extraordinary results beyond the usual ones? Is that fair to say?

Ellen: Yeah, I think... I mean, I think there’s... it depends on... there’s pros and cons. So the deep learning-based tools work, but they require a little bit of fine-tuning actually.

Abhi: Okay.

Ellen: Or you need to retrain a new model every time. And actually, these days, the microscopes are super automated, are really fast, and the image processing is now a huge bottleneck in terms of solving structures. And so if you want speed, you just use maybe the simpler-based tools.

Abhi: Okay. Oh, sorry, go ahead.

Ellen: Oh yeah, just the simpler tools.

Abhi: The one other question I wanted to ask was, beyond machine learning, how much innovation do you expect to happen in the physical processes of collecting the data?

Ellen: So that’s also very much still actively being developed in the field. So like there’s people at Berkeley working on these crazy laser phase plates to just fundamentally increase the amount of signal in the collected images, and that will just solve all the downstream problems. Like if we could magically get less noise or be able to irradiate our sample with more electrons before blowing it up, then that solves all of our downstream problems. So that would be great. So people are very much still working on the hardware, which is why this field is so rich in terms of the different areas that you need to understand.

Abhi: I was going to... if someone is at OpenAI and working on pre-training, maybe three years ago, all they needed to care about was, “Oh, how do I distribute my data across multiple GPUs better?” But now maybe things are getting so close to the edge of what’s possible that they now need to start deeply understanding kernels. They need to start paying attention to every software update that Nvidia turns out. How much do you personally stay on top of the hardware innovations that’s going on?

Ellen: So I try. I think one of the privileges of my job right now is that I can do the research, but I can also... I also, you know, both get access and can talk to all these people in all these other areas. And I think it is really important to stay abreast of all these other directions to understand which ones will completely change our field.

Abhi: When I look at the field of at least protein structure prediction, it feels as if to me there is not much room left for machine learning advances. And most of the innovation will just come from being able to collect the data better and faster.

Ellen: That’s a hot take.

Abhi: Yeah. Potentially it’s wrong, but I’m curious on your side... how much alpha do you think there is on improving... churning the ML crank faster and faster versus just trusting the people at Berkeley will solve the problem by better data collection? Like how much room is left to go with stuff like CryoDRGN?

Ellen: I see. I think there’s still lots of interesting moonshot problems on the ML side. And one of the main directions that we’re interested in is can we innovate on the algorithms enough to be able to change how data is collected? Right. So one major trend of the field is moving towards less pure samples. So before it was completely pure. And now with Hydra and these kind of extreme compositional heterogeneity, we are moving towards dirtier lysates or just cellular fractions or the in situ slices. So that’s definitely a trend. And that’s only going to be possible if the algorithms can either keep up or can the algorithms themselves motivate new ways of collecting data.

[00:46:57]  Has deep learning done something in cryo-EM that was previously impossible?

Abhi: I think... all my examples are coming from the protein structure prediction field. But... I think it took a few years post-RFdiffusion being released, I actually saw for the first time that, “Oh, these models have given us research papers... like clinical-stage research papers... that otherwise could not have existed if this model did not exist.” Are there some Cryo-EM-determined structures out there that exist and are deposited that you suspect would not have come about had stuff like CryoDRGN not existed?

Ellen: Hmm. Oh, I don’t know. I feel like it’s too...

Abhi: I guess this is a strong claim to make.

Ellen: Yeah. It’s too strong of a claim to make. I mean, the hope of these tools is yes. Right? Like, and in the original publication, the satisfying part was re-analyzing previously published datasets and finding structures that were missed.

Abhi: Okay. That’s something.

Ellen: So that was awesome. Yeah. That was super cool, especially because at the time I didn’t even have the taste to understand what I was doing in terms of reanalyzing the data. And I was like, “Oh, this is weird. I guess it was not there.” And then it was like, “Wait, this was not there before. That’s cool.” Yeah. So I think that... yeah, that was, I guess, super gratifying.

[00:48:22] Ellen’s experience in the cryo-EM field

Abhi: Going back to when you began in this field entirely and you need to come up to speed on Cryo-EM and all the jargon that that entailed... what was the biggest, almost like culture shock that you had coming into the Cryo-EM field versus... I’m assuming you were primarily like a physics ML person prior to this.

Ellen: Well, I was actually an MD person, or molecular dynamics person prior, so yeah, very new to Cryo-EM. And I talk about... so there’s a bunch of mathematicians working on the Cryo-EM problem and those papers are a lot easier to read because there’s less jargon. So that was maybe my initial foray, but then, you know, just reading all the papers in the field and just trying to make sense of the jargon was the biggest shock originally, just as a graduate student trying to figure things out.

My first time actually talking to... meeting people in this field... was also just interesting in terms of... I guess the jargon was the main thing. Of just like, “Okay, what is... what is pseudosymmetry?” You know, “Is it symmetric or is it not symmetric? Like what the heck is pseudosymmetry?”

And so... and I remember one of the first times I gave a talk on CryoDRGN to a primarily structural biology community, they were really kind of suspicious of deep learning. Yeah. So I think there was a major sea change in the structural biology community because of AlphaFold.

Abhi: Okay.

Ellen: So that was a pretty major shift... a vibe shift... of whether people trust deep learning-based approaches or not. And so definitely before that, there was a lot of skepticism, but there was a lot of excitement too, because heterogeneity was the next frontier of the technology and “how do we model continuous motions?” And so I think, yeah, in the beginning people were extremely excited and I was trying to not overhype it, you know, of just like, “Don’t get too excited.” This is not going to solve all your problems. Like, if it’s garbage in, if your sample is really, really challenging, it’s not going to necessarily solve all the problems.

Abhi: Do you often find that, at least in... actually I’m not going to make any broad statements about proteins again. At least within your group and perhaps other Cryo-EM ML research groups, how much of the innovation comes from talking to Cryo-EM people and listening to their ideas and their research ideas? Versus looking at the ML literature. Like, I’d see... I think to take a concrete example, there feels like a pretty strong connection between your work and the NeRF work... Neural Radiance Fields. How much of it is... is that where you’re taking inspiration from the ML literature versus taking inspiration from traditional Cryo-EM?

Ellen: Yeah. So I think now... I think my group has people who are very much computer vision, machine learning people, computer scientists, people who are more mathy. We have a biochemist, a biophysicist. So I really love how interdisciplinary or multidisciplinary the group is because you need to draw from all these different kinds of areas.

Originally, CryoDRGN was... I had no idea about NeRF. And so at some point... CryoDRGN preceded NeRF. And then when I saw... at some point somebody was like, “Oh, you should check out this Neural Radiance Field.” I was like, “Whoa, that’s super cool.” And I had no knowledge of computer vision. And I remember my PhD advisor at the time being like, “Oh, you should think about whether this type of architecture can be useful for other domains.” I was like, “Nah.”

Abhi: You could have been a graphics researcher.

Ellen: I could have done graphics. Yeah. So I guess... yeah, definitely it’s good to stay abreast of what’s going on in machine learning. What are the trends, what are the latest and greatest architectures and training paradigms and activation functions and things like that. But I think it’s extremely important to stay focused on the problem at hand and focused on what are the challenges associated with a specific problem... and whether the techniques can be cargo-culted or not.

Abhi: Gotcha.

Ellen: And most times they can’t, but... related things.

Abhi: So I’m assuming you aren’t going to SIGGRAPH every year to see what inverse problems people in the graphics field are working on?

Ellen: No, I’ve never been to SIGGRAPH, but I sometimes go to the computer vision conferences. And I think it’s nice to establish crosstalk between all these different communities, because the problems are so related. But yeah, the thing to fight against is just chasing the trends in these other areas and applying them to Cryo-EM. I don’t think that’s the way to go about it.

[00:53:40] Deep learning in cryo-EM outside of structure determination

Abhi: Outside of structure determination...

outside of applying machine learning to particle analysis, you mentioned that there is machine learning being applied to actual particle picking itself. What do you view as the most promising direction outside of particle analysis? Is there anything that you view as particularly interesting outside of what the Ellen Zhong lab is working on?

Ellen: Oh, well I was going to interpret your question as outside of the 3D reconstruction problem.

Abhi: Sure. Yeah.

Ellen: Which is also something that our group is working on right now. But outside of... so really we’ve been focusing mostly on this 3D reconstruction, so just 2D to 3D image processing. The direction that we’re getting into now is this CryoBoltz paper. Yeah, so the atomic modeling. And I think that’s something that’s really interesting. It’s an unsolved problem.

It’s largely done manually still. It’s unclear how to do it for low-resolution maps, but how do we actually build the atomic models into the Cryo-EM density volumes? Especially when we have an ensemble or distribution of structures from CryoDRGN. That’s currently unsolved. How much should we rely on priors from structure prediction models? Building in the atomic models...

Abhi: How much do you personally trust them?

Ellen: I think we should use them. We should definitely use them for everything they got. I think now the challenge is validation. Like once it spits out a structure, how do we actually validate at scale? If we have a thousand Cryo-EM volumes that we’re building atomic models in, how much information do we share across those thousand structures? And now how do we automate the validation procedure? Because any practicing structural biologist, when they deposit an atomic model, they’re going through every single residue or every single atom, residue by residue. And... so should we still be doing that? That’s an open question.

Abhi: Earlier when I asked about why pre-trained foundation models aren’t used in Cryo-EM, you said you want to actually recapitulate the data as you measured it and not potentially hallucinate something. But it does sound like if you’re going in the CryoBoltz direction, it is on the table to hallucinate something.

Ellen: Yeah.

Abhi: How much have you seen that happen in practice?

Ellen: That’s a great question. I think one... I guess the way that I’m approaching it right now is separating the reconstruction problem from the atomic modeling problem.

Abhi: Okay.

Ellen: People are definitely trying to bring them together and having reconstruction models that are deformations of an atomic model. And that’s where I think you can get hallucinations. And I have seen... there was this really interesting case study that one of our collaborators at Princeton showed us, where if you have a homogeneous dataset... so synthetic dataset, you can create a synthetic dataset of just a static structure... and it’s extremely noisy. And if you fit one of these heterogeneous reconstruction models, you get... you hallucinate conformations. Especially if it’s one of these models that only models conformational heterogeneity. You just get these flexing motions that are definitely not in the data. And so yeah, that happens. Right. And that’s something that is a worry.

[00:57:32] 3D volume reconstruction versus residue assignment in cryo-EM

Abhi: I actually hadn’t mentally separated this 3D volume reconstruction problem as something you do first and then you assign residues. I assumed... I guess now looking back at the actual papers, it does seem like they are two sequential steps. Is it obvious... going back to proteins, there’s a similar question of, “Should you design the structure first and then the sequence, or maybe you do both at the same time?” And the canonical thought has always been, “Oh, you want to design them at the same time because they rely on each other.” Is there any similar undercurrent of thinking within the Cryo-EM field where you want to be able to do both at the same time because it just helps you succeed at each of them individually?

Ellen: Yeah. So I think by separating them out, you are separating the 3D structure determination from the sequencing of the actual protein. And obviously, maybe you needed to sequence something to make the sample in the beginning... but not necessarily, especially if you’re just extracting from the wild.

Abhi: Yeah.

Ellen: And so it’s been used... Cryo-EM has been used in this bottom-up fashion just to discover new proteins, to discover new complexes, new interactions. And then it becomes a hard problem of how do you sequence the thing that you’ve solved the structure for. And I think now in this post-AlphaFold era where we have structural hypotheses, it becomes less of a completely crazy problem. But I think that is one of the exciting things. And then in an actual design context, you could imagine high-throughput structure determination of an ensemble of different design sequences and then seeing afterwards which ones bind or something like that.

Abhi: Yeah. How much... actually, this is a question I probably should have asked a long time ago... but when you’re actually assigning atoms to this three-dimensional model, how often is it the case that you know what residues are on the table to start off with versus you don’t know anything about the sequence?

Ellen: I think most of the time you’re trying to solve the structure of something whose sequence you know.

Abhi: Okay.

Ellen: Yeah. You want to determine this new protein structure, this is the sequence that you’ve purified out. But especially if it’s from an endogenous source, you never know what else is going to be there.

Abhi: I imagine like lysates...

Ellen: Yeah. If it’s from a lysate, like whether there are ligands bound, whether there are other complexes that are co-purified. And so that’s the thing that from the machine learning standpoint is an interesting axis: is how much do you rely on priors and how much do you allow for discovery?

[01:00:26] What did Ellen do during her sabbatical at Generate Biomedicines?

Abhi: Okay. That makes sense. And I know that you posted on Twitter a while back that you recently took a sabbatical from Princeton to go to Generate Biomedicines for a bit.

Ellen: Yeah.

Abhi: And I learned this morning that that sabbatical has now ended.

Ellen: Yeah.

Abhi: What were you working on there? And I guess... so for some context, Generate Biomedicines bought this 70,000 square foot facility, I think in 2023, and just filled it with Cryo-EM machines. So clearly there’s an overlap as to why they actually need your exact skillset.

Ellen: Yeah.

Abhi: What were you working on there?

Ellen: Yeah. So at Generate... so that was an amazing sabbatical. Generate has four cryo-electron microscopes, which is double the number of Princeton University. And all these challenges that we talked about in the past of, “Okay, the datasets come from all these different research groups around the world.” I mean, they’re aggregated in these databases, but the quality is kind of different. And Generate, you know, that’s a single organization where they can actually standardize everything and pipeline everything for high-throughput structure determination.

At Generate, we were working on structure determination and solving the ensembles of antibodies... the conformational ensembles of antibodies, and trying to ask interesting questions about the role of antibody structural dynamics in binding, in affinity maturation, and things like that. So that was super cool, but very much from... “Can we just actually observe?” Right. So... I learned a lot about antibodies. It was amazing. But yeah. So... is there a logic or grammar to antibody CDR loops?

Abhi: Why... what’s the utility of structure there?

Ellen: Hmm. Yeah. Great question.

Abhi: ‘Cause in my head, CDRs are floppy. Is that just completely incorrect?

Ellen: In my head, they are too, but no one knows. Right? Okay. Because no one’s observed the floppiness because we haven’t had tools to be able to observe conformational dynamics or conformational ensembles until, I guess, Cryo-EM. And it’s still a very challenging sample prep problem, especially if you’re trying to scale it up. But now can we actually just observe from experimental data what the actual floppiness of the CDRs are? Is that actually the case? Or are some more rigid than others? What’s the sequence determinants of that? And obviously, there’s lots of therapeutic relevance if you understand... if you can now rationalize the structures of the loops.

Abhi: How much are you able to share about how floppy they actually are?

Ellen: Oh, well, very much still a work in progress. But it is publication-oriented.

Abhi: Okay. Gotcha.

Ellen: Which is why I’m really excited. And definitely why I was interested in doing the sabbatical there.

Abhi: What is the clinical relevance of it? Like, let’s say you are able to determine that the CDR loop is floppy, it binds in this specific way. What do you tell a therapeutics team as the result of this process?

Ellen: Yeah. I feel like the therapeutic angle is above my pay grade or outside of my actual area of expertise. But I would imagine that the floppiness has some functional relevance. And if you can better understand, characterize, or design the function of the binder, maybe you would get fewer off-target effects or you would just actually have more useful functional data about your antibody or about your therapeutic.

Abhi: Did the experience make you desire building therapeutics or did you go the other direction, thinking “this kind of sucks”?

Ellen: Yeah. No, no, no. It was actually extremely inspiring. I would say my prior has always been more on the academic side of, “Oh, basic science. There’s so much we don’t understand about biology. I think working more upstream on the fundamental things is more... I don’t know, just interesting.” And then actually being there and realizing that these are real drugs that are actually helping people was very inspiring. I was like, “Oh, yeah.” So that was cool. That was super cool.

Abhi: The... when I saw that they acquired all of these Cryo-EM machines... I did not know it was only four. In my head, it was dozens.

Ellen: Four is a lot!

Abhi: Yeah, four. Four seems to be a lot. My initial thought was like, “Oh, they want to... they’re the ones who made that protein structure model Chroma. Maybe they want to scale up protein structure determination using these.” Was that also on the roadmap or is that...

Ellen: Yeah. I guess... so I’ve known some of the people at Generate for a long time. I think Gevorg Grigoryan, the CTO, John Ingraham, the head of ML there... their papers I read when I started as a PhD student. I was like, “Whoa, these are super cool protein ML papers.” And I think their hypothesis... okay, who knows? I don’t know. But I think definitely structure prediction is a huge capability, or generative models over protein structure. And then can we use those models for design? But obviously... I think very much this area, you don’t want to stay purely in silico, because if you actually want to make real therapeutics, you want experimental data, ideally in high throughput. And so I think probably Cryo-EM was a major part of that and that investment was... ahead of its time, I would say.

Abhi: Yeah. There’s one other therapeutics company that has made a really big bet on Cryo-EM. I think it’s Gandeeva Therapeutics. Do you happen to know what their particular bet is?

Ellen: I think there’s a couple. There’s definitely others that I’m aware of that are using Cryo-EM. And I think the idea there is very much to use Cryo-EM to understand maybe cryptic pockets, or... I don’t actually know about Gandeeva in particular, but just my understanding is with Cryo-EM you can actually get the ensembles, and with the ensembles you can maybe find new binding sites or better understand allosteric mechanisms and things like that.

[01:07:07] Ellen’s research in cryo-ET

Abhi: You’ve done a few things besides Cryo-EM as well. I’m going to talk about one thing that your future work is going to be focused on, and one thing that your past work is focused on. The future thing seems to be cryogenic electron tomography.

Ellen: Yeah.

Abhi: The basics of the technique, I think, is just exactly the same as Cryo-EM but you’re tilting the sample so you get... actually, maybe it would be best if you explain it.

Ellen: Yeah, yeah, yeah. So cryo-electron tomography or Cryo-ET is using the same microscope as in single particle Cryo-EM. But there’s two main differences. One is the tomography part is now, instead of just taking a single projection image through your ice, you’re now tilting the imaging stage, like plus or minus 60 degrees is practical, and taking a series of images at different angles. And so then you get a 3D tomogram. And so you can see things in 3D. So that’s one main change is the tilting.

The other change is that it’s usually used to image in situ samples, so like thin cellular sections. So instead of a purified solution, you’re looking directly in situ.

Abhi: Sorry, that I guess maybe answers my follow-up question of, “Can you not approximate this three-dimensional view from Cryo-EM?” But I guess the in situ part demands that each molecule is unique and you’re not going to see it repeated across the entire data.

Ellen: Yeah. So the spatial scale is different. So now you’re seeing subcellular architectures and membrane morphologies or organelles and things like that. You can still identify individual particles... and then pick all those particles. And the individual particles are super noisy and super low resolution... but you can do the same trick as in single particle Cryo-EM, combine them all together and computationally amplify the signal to get near-atomic resolution structures in situ. And so the extremely exciting part there is that now we’re determining the structures of these protein complexes instead of in a crystal, which is from X-ray crystallography, or instead of in a purified solution. Now we’re actually looking at the functional structures in the native environment.

Abhi: One instinctive question I have is...

Ellen: And you also get the spatial organization too, which is super cool.

Abhi: How large are these samples? Like, are you able to see them with your naked eye?

Ellen: No. So I think the field of view is still pretty small. So there’s many orders of magnitude between angstrom-scale things that we resolve to maybe nanometer, hundreds-of-nanometer scale sections, to whole cells. And definitely, whole-cell visual proteomics is a huge research direction for the field. Can we have those beautiful David Goodsell images of these cellular landscapes, but from experimental data?

Abhi: I was going to ask, how far are we away from being able to image an entire cell at once? Has that already happened?

Ellen: So that has already happened. It depends on how big the cell is. Okay. Yeah. And so people have done this with bacteria, which are much smaller, but for larger, more complex eukaryotic systems, it’ll require a lot of data. And then it’s an interesting question of whether you want to be hypothesis-free about it and just see all the things, or whether you want to test a specific hypothesis and have “with a drug” / “without a drug.” And I think that’s a challenge or maybe a research question for the field.

Abhi: What’s an example of a hypothesis-driven question you’d have about a Cryo-ET result? Is it like, “Does this particular protein exist in this native environment?” or is it something coarser?

Ellen: I think people are interested in personalized medicines. Right? Okay. So like, does the morphology of some organelle change? And so it is tackling this length scale that’s unique, right? Light microscopy can only get you so far, can only get you the wavelength of light, which is thousands of times larger than the wavelength of an electron.

Abhi: Yeah.

Ellen: And so with Cryo-ET, you can actually see maybe subcellular features and architectures. And then for certain complexes that are abundant, you can get molecular details and atomic details. And so really the dream would be to be able to span all these different length scales from atoms to cells.

Abhi: The idea of a Cryo-ET result being clinically relevant is interesting. I don’t think I’ve ever heard a claim like that before. Is there some particular disease you’re aware of where if we could exactly measure the physical structure of the mitochondria or something, that would help a lot? Or is it kind of a basic research thing where we’re not sure how this will help?

Ellen: Yeah, I am not sure. I feel like there’s probably scope, but it’s way above my... or I don’t know enough about the clinical side of things. It is an expensive experiment. So I think that’s one of the main challenges is the facilities... it’s expensive, it’s hard to do, it’s immature still in terms of... you know, you need a lot of... the sample preparation... everything is very much still... people are working out how to do it the most effectively and efficiently.

Abhi: So it’s like, it’s a newer method than Cryo-EM.

Ellen: Yeah. Okay.

Abhi: Is it... do you imagine... the road to getting more traction is there? It’s already shown early promising results. Maybe like Cryo-EM won the 2017 Chemistry Nobel Prize. Is this on track to win the 2035 Nobel Prize?

Ellen: I have no idea. I have no idea except that I do think because it has this unique vantage point in this length scale of cellular and molecular biology... it really spans molecular and cellular biology... it can be used to find new things and make really important discoveries for sure.

[01:13:54] Ellen’s research in NMR

Abhi: Yeah. That makes sense. Another thing you mentioned to me just when we were talking this morning is that you were also working on NMR.

Ellen: Yes.

Abhi: Which is new to me. Yeah. What are you doing there?

Ellen: Yeah, so that’s been super exciting just from a personal level, because I’ve been working on Cryo-EM for a long time and there’s so much more to do there. There’s a lot of really interesting open research questions. But we do have a new direction in the group that is small molecule structure determination.

Abhi: Mm-hmm.

Ellen: So I guess that’s the commonality is: can we analyze... can we just do useful things and help chemists? Now we’re talking to chemists. Can we help chemists elucidate the structures of novel natural products and novel molecules? That is still very much a bespoke method... done manually, requires expertise to just stare at these NMR spectra and figure out what the graph or the structure of the molecule is. So yeah, I think that’s super interesting. When I first learned about that problem, I was like, “Ooh,” I had the same feeling as the Cryo-EM problem. Like, “This is a really cool problem and I’m not aware of anybody else working on it right now. And this seems like a really useful problem to solve.”

Abhi: So the mental framework... the mental framework necessary to understand Cryo-EM was a little bit alien to me when I started reading about it. You have a bunch of these 2D images from an electron beam passing through a protein. You have an electron detector that records a signal and your job is to reconstruct all these 2D images into a 3D structure. What is the mental framework for applying ML to NMR?

Ellen: Yeah. So it’s another... I guess... it’s another inverse problem where we have imperfect experimental measurements. The spectra itself is... the NMR or the shift of the resonance of these different nuclei, either protons or C13 or carbons... isotopic carbons. And depending on the shift, that tells you about the local chemical group and the local chemical environment.

And then the puzzle that usually a human, an expert, will do when they’re looking at the spectra is like, “Oh, there’s a peak around here. So I know that this is this type of hydrogen in a benzene ring or this type of hydrogen. And there’s a peak around here and there’s some decoupling, so it must be near some other type of chemical group.” And maybe you know the chemical formula, like you’ve done mass spec or something. So you know the total composition of the molecule, but you don’t know at all how it’s connected together.

And then from the spectra you just stare at it and from all of your training, you know that these peaks are artifacts, these peaks should be there but it’s missing because this always happens. And so when I learned about this problem, I was like, “Oh, this seems perfect for machine learning.”

And I guess what was most inspiring to me in the beginning about this area of natural products is: if you look at the structures... well one, they’re oftentimes the most bioactive and therapeutically relevant molecules. They’re isolated from natural sources instead of made from rational drug discovery or these drug-like molecules. And they look crazy.

Abhi: Yeah, I’ve heard this from natural products.

Ellen: Yeah, the... if you actually read the papers and look at the structures, I’m just like, “Oh my God, these molecules are gorgeous and also totally gnarly.” They look really exotic. And I’m just like, “Who... this is what the bacteria are making when they’re conducting chemical warfare against each other or whatever they’re doing.”

And then you read the machine learning papers in chemistry and it’s a really interesting modeling problem, but most of the datasets are super small molecules that have fewer than maybe 20 heavy atoms or something like that. And I’m like, “There’s a huge gap here in terms of what we’re actually modeling from a machine learning standpoint and what are the super-cool, active molecules that we’re just extracting from nature.” So that was the main motivation.

Abhi: I imagine... whenever I see a chemical structure of a natural product... yeah, one, it does look often pretty crazy, but two, how does anyone come up with those structures today? Like, in Cryo-EM, when you arrived to the field, there were a bunch of traditional methods being used. Are there a set of traditional methods also used to determine the structure of these really crazy molecules? Or is it like they need to learn how to synthesize it and then from that they naturally get the structure?

Ellen: No, you very much don’t know how to synthesize it. So the synthesis problem is usually downstream of discovering and characterizing the molecule. Because oftentimes, all these really expensive cancer drugs, they’re still extracted naturally because we can’t figure out from a chemistry perspective how to do the total synthesis. And... yeah, I’m not a chemist, I’m just a computer scientist. But I think natural products is interesting because we just mine nature and maybe figure out how to perturb these bacteria to produce all these crazy chemicals. And then we just extract them from natural sources and then do all the biochemistry and physical chemistry and analytical chemistry experiments to figure out what it is... and figure out how it works, figure out how the enzymes that make it work... and it’s just a whole super interesting field that has been... yeah... great to get a sneak peek to in the last year or two.

Abhi: I know very little about this field, but I do remember reading these Mass Spec Foundation Model papers from Enveda Biosciences, if you know them. And they claim to be able to turn the mass spec peaks into the actual structure of the molecule. How trustable are those methods?

Ellen: I think the mass spec stuff, I don’t actually know that much about. I do think metabolomics... there’s a lot... that’s an interesting problem, especially when you have complex samples and things like that. But from our chemistry collaborators, I think when you’re trying to solve the structure of a new natural product, you typically use NMR.

Abhi: Okay.

Ellen: Which gives you more information on the actual connectivity of the atoms.

Abhi: And so the mass spec information is underdetermined for the structure problem?

Ellen: Yes. There’s just not enough.

Abhi: Okay.

Ellen: I think mass spec is usually used to identify composition or just identity of, “Is the molecule there or not?” if you already know the molecule, you know its spectral signature.

[01:21:05] How did Ellen get into the cryo-EM field?

Abhi: Right now I’ve talked about cryo-ET and NMR, both of which you’re working on right now. But for a long time... three years you were at D. E. Shaw Research before going to graduate school. And I have two several questions on that front. The first of which is, why start with molecular simulation? Why was that the thing you originally were working on when you were at University of Virginia?

Ellen: Yeah. I would say luck. You know, originally, how did I get into research and how did I get into the area that I was in, is just luck. And I’m extremely fortunate to have met the undergrad research mentor I had at UVA, and from him, that’s where I learned about protein folding and molecular dynamics. And we wrote this super cool... or I thought it was super cool... model to simulate protein folding. And that’s how I learned about Anton and D. E. Shaw Research. And then I was very lucky to work there for an internship and then full-time for three years. And so, getting into that area, I think, was just luck in the beginning. But I’m really glad I did for sure.

Abhi: Do you ever... I think there’s a common pattern of a very smart person who touches molecular simulation for a few years and then gets out of it because they think “this is not going anywhere.” Did you have a similar reaction or was it more like, “Oh, I’m curious about other things besides molecular simulation.”

Ellen: I think a little bit of both. I think working at D. E. Shaw Research, I was working on free-energy calculations and binding affinities. And then through that was working on originally the sampling problems there, which is really interesting technically. And I learned so much. And then after characterizing it on more systems was like, “There’s a force field problem here.” And then worked on some quantum stuff to better estimate the force fields. I was like, “This is a really hard problem.”

And then was like, “Okay, at the end of the day, all these predictions need to be validated experimentally.” And then I had no idea anything about biology at that time or experimental biology. And so that was just this huge area that I was like, “Huh, that’s the... you know, that’s the answer key.” And it’s a huge mystery. And so I think after a couple years, it was definitely following what’s next and wanting to understand the experimental side.

Abhi: Yeah. And were you like a pure physics person by undergraduate training?

Ellen: In undergrad I was doing chemical engineering. And so that’s how... and that’s where I was exposed to stat mech and this protein folding area. And yeah, so it’s interesting. Now I’m in a computer science department. I took the CS classes, but wasn’t that interested in some areas of computer science. And it was too hard to get into the classes ‘cause it was really oversubscribed. And I don’t know, I liked chemistry for sure.

Abhi: Yeah. Do you ever imagine returning back to molecular simulation in any capacity? Or are you kind of set on this Cryo-ET, NMR track? And maybe even those will involve some molecular simulation.

Ellen: Yeah, definitely. I am not against it at all. I think, especially now, there’s so much interest, right, in molecular simulations with neural-based approaches eating various areas of MD. And again, it’s all about what are the questions that this technique can answer. Like Cryo-ET gets this unique length scale. MD also, I feel like, accesses a unique length scale that you can’t get with experimental data. Right?

Abhi: I think, yeah, I saw this really interesting graphic that says the timescales of NMR are incredibly coarse to what you can achieve with sufficiently coupled cluster....

Ellen: Yeah. Right, right, right. And the dynamics... I mean, what originally drew me to the Cryo-EM problem was that, “Oh, we can actually get the dynamics of these proteins, but from experimental data instead of from simulating these very, very simple physics-based models.” And so that was super cool. It’s like, “Oh, this is directly from the data.”

However, in Cryo-EM, the ensembles are still just larger-scale, slow-timescale conformational changes. And so I think you’re never going to get super-fast kinetics from a Cryo-EM dataset. And so there’s still... there’s interesting areas of overlap or complementarity between MD and Cryo-EM.

Abhi: Have you ever felt... does anyone in the Ellen Zhong Lab focus on applying molecular simulation to Cryo-EM? Or it’s somewhat of a nascent field where you’re not sure where it could actually be applied?

Ellen: Some of our collaborators are definitely applying MD to Cryo-EM right now. Oh yeah. And my lab, I call EZ Lab.

Abhi: Oh, okay.

Ellen: For, yeah. Which is a fun acronym. So MD, I feel like, is also... it’s even more so this bespoke method. So... Cryo-EM is bespoke for each dataset, but MD is even more so because there’s this upfront cost of setting up a simulation and all this stuff to validate the force field terms for a particular system. So right now we’re not directly integrating MD-based modeling and Cryo-EM, but definitely using it as an interesting testbed for conformational ensembles from Cryo-EM.

[01:26:57] Why did Ellen go back to graduate school?

Abhi: During your time at D. E. Shaw Research, and the subsequent decision to go to MIT for a PhD... what was the primary motivating factor? I do know there is this track of scientific associate at D. E. Shaw Research to PhD track. Was there anything else that was on your mind of what you could have done besides?

Ellen: Oh. Yeah. I guess at that time... I wasn’t sure. Right. I wasn’t really sure exactly what I wanted to do with my career. But I thought learning more things would be great. And I think concretely, my GRE scores were going to expire, something like that. So I was like, “I might as well apply.”

And I remember going through the process and it was really useful to think about what exactly I want to... what would I want to study for a PhD? And what kinds of things am I interested in? And then after going through that process, it was pretty clear. And after starting my PhD it was so clear. I was like, “This is really cool.”

Abhi: Okay. Have you ever gone back to look at your personal statement, your statement of purpose, and seen how much you’ve deviated since then? Or was it pretty spot on, like, “Oh, I care a lot about validation, I care a lot about validation today as well.”

Ellen: I definitely remember going back to my statement of purpose to look at which groups I was interested in, and it was totally different.

Abhi: Okay. You were not even... were you aware of Cryo-EM at the time?

Ellen: No, not at all.

Abhi: Okay.

Ellen: So that was only after a year... my first year of my PhD, after doing rotations in all these different areas. And I tried experimental stuff, which was cool. Learned how to pipette, which was cool. Did some neuroscience, which was crazy. And I was very open to many different things. And then when I learned about Cryo-EM, I was like “okay”.

Abhi: The last questions I had was, if you had a hundred million dollars to spend on anything, someone gave you that money, no strings attached, what would you spend it on?

Ellen: I think I would be doing exactly what I’m doing right now.

Abhi: Well, sure, sure. Not in terms of, “Oh, I’m going to retire,” but...

Ellen: Oh, like I need to spend it or something.

Abhi: Like, would you buy a hundred Cryo-EM machines? Is each one a million dollars a piece? I’m not technically sure. But what would you be spending that on in terms of accelerating your own lab?

Ellen: Hmm. Yeah, that’s a really interesting question. I definitely wouldn’t change anything about what I’m doing right now. I think the problems are interesting. The people, my group is amazing. And the research... I don’t feel particularly limited. Okay. Maybe if I had to spend a hundred million dollars, I would branch out into more of the experimental side and have more data-generating capabilities to either test hypotheses or to design the experiments. But yeah, I feel extremely lucky right now that we get to just work on the cool problems that are hopefully useful.

Abhi: I feel like that’s a rare answer. Usually when given a hundred million dollars on the table, people are like, “Yeah, I’d buy so-and-so machine, I’d start so-and-so research group.” And especially Cryo-EM feels like the sort of thing that is so bound by the world of atoms. But yeah, I guess in practice, not.

Ellen: Yeah, I think... concretely, right? Like, then I would need to know how to do Cryo-EM experimentally or do NMR experimentally or something like that. And so that would be... that’s a whole other thing that...

Abhi: Well actually, I think that’s an interesting question. Would you trust yourself to do NMR or Cryo-EM independently today?

Ellen: Me personally, no.

Abhi: Really?

Ellen: Like I have no idea how to do that.

Abhi: Really?

Ellen: Like there are people in my group who do collect data. And that’s great, but I’ve never collected data.

Abhi: Wow. Is that the sort of thing where it’s like you genuinely need months upon months, maybe even years of training to be able to do it well?

Ellen: Yeah. I think, like I was talking about earlier, the field is so... it’s so deep. You do need a lot of expertise. And I think there was an opportunity at some point during my PhD to actually think about collecting data or something like that. But it was just never... I don’t know, I liked coding too much.

Abhi: Yeah, that makes sense.

Ellen: And I think... it’s interesting, money is not really the bottleneck. It’s time and bandwidth. And what was so amazing about the sabbatical is like, “Oh, I don’t have to teach now. Okay. And now I can code myself again and run experiments... or computational experiments.”

[01:32:17]  What makes Ellen more confident about trusting an external cryo-EM paper?

Abhi: Perhaps a question that I should have asked earlier is: you were probably the person who made the particle analysis Cryo-EM machine learning field big in the first place, and now there’s entire workshops devoted to the research field. With the rise of many more talented people entering the area, I imagine there’s also some deluge of more noisy, lower-quality papers. Is there some easy way for you to tell whether to take a Cryo-EM paper seriously or not?

Ellen: The... I don’t like looking at the field with that lens. I’m like... I think everybody is doing interesting stuff as long as people have scientific integrity. And I think people are interested in different aspects of the problem, and I like that we have that flexibility, right? Of like, “Oh, they’re more interested in proof of concept.” Sure. And are writing conference papers and can get these concepts out there more quickly. And then there’s papers that are more targeted towards users. Right. And actual tools that people will use. And I think as long as there’s a healthy ecosystem where people have the flexibility to work on the style of research and, we’re making forward progress as a field... then, yeah, then I think that’s great that there exist these different types of research directions.

Abhi: I guess concretely... I know, especially in both the field that I work in right now and the field I worked in prior, there are known datasets that are... like people claim very strong things in either toy datasets or known datasets that have really big fundamental problems. I think maybe most overlapping with our work is... I think PoseBusters is known to be kind of a bizarre dataset.

Ellen: Oh yeah. I’ve heard this. Yes. I don’t know why, but I’ve heard this.

Abhi: I don’t know why either. I just... yeah... I’ve had colleagues who say this. I guess the Cryo-EM dataset world is a bit smaller than the protein world and the small molecule world. But are there any issues like that? Actually instinctively, what I thought your answer was going to be was, “Synthetic data is really hard to take seriously.” But maybe that’s not the case.

Ellen: Yeah, I think the proof is in real data and seeing whether a method works on real data. And there’s some established benchmark datasets that people are usually analyzing. So that’s definitely one of the markers of, “Does this... how well does this method work?”

And I mean, one of the challenges for especially new people coming to the field is that you need a lot of expertise... domain expertise... in interpreting the results and interpreting the structures. One of the things that one of my group members did last year, it was at the NeurIPS benchmarks track, was CryoBench, or our benchmarking effort, which was hopefully designed to both have simple diagnostic datasets that people can use for methods development that can tell them whether their method is working or not, from both qualitative and quantitative metrics... and challenging datasets that will motivate new methods development. And so we were very deliberate with the design of these datasets.

And so that’s been super cool to see where people are actually using these datasets and computing the metrics. Metrics is a huge problem for the field because... yeah, I think that continues to be an open problem for the field. And so, yeah, I think maybe the lack of metrics helps guard against blindly following a number. But yeah, I think it makes it harder for new people, like maybe who are more expert in other areas like graphics or vision, to get into the field. And so with all these things, I think there’s just nuance and it’s important to keep that in mind.

Abhi: Metrics... like what’s the story on Cryo-EM metrics? Is it just RMSD? Is there something else beyond that?

Ellen: The main metric is resolution.

Abhi: Okay.

Ellen: For a given volume, but that’s also estimated a bit heuristically. And a notion of global resolution for the structure doesn’t actually hold. And the actual computation of resolution can be hacked a bit. So that’s for sure a challenge in the field. You know, there’s enough people who have the expertise to judge the quality of a structure that you’ll get called out for it. Right. So that’s one of the main metrics is resolution, and computed via something called FSC, or Fourier Shell Correlation. For heterogeneity, for ensembles, it doesn’t exist yet.

Abhi: Okay.

Ellen: And we have some stand-ins that are FSC-based. Some people in our group are taking some of the shape-based metrics from computer vision, like Chamfer distance, just volumetric IOU, and things like that. But there’s a lot of just caveats and challenges. And so I think metrics for conformational states and measuring what is it that you care about is not solved.

Abhi: Naively, I think, Why isn’t the conformation problem just like, ‘Oh, there are these five conformations in the dataset. If your model doesn’t give me back all five or it gives me four out of five... why don’t you just report that accuracy?’ Like, what’s the nuance there?

Ellen: Yeah. I think even RMSD, right... so that doesn’t capture the specific conformational state. And I think, you know, people are using maybe local-type metrics or maybe that would be a direction to move forward. But that is very system-dependent. And so I think a general metric for conformational state just doesn’t exist right now.

Abhi: Yeah.

Ellen: Perhaps against a target dataset, you could define something like that.

Abhi: Was that done for CryoBench?

Ellen: We didn’t have dataset-specific metrics. That’s definitely something that I want to move towards. Like, especially one of the datasets is based on an MD simulation, so there’s 46,000 distinct structures in the dataset. And you’re not actually going to be able to recover 46,000 distinct structures. So how do you actually characterize the distribution? And you have distributional metrics, right, that is established. But, you know... what is the actual... yeah... what is the representation? Like there’s all these actual details in terms of the metrics and computing distances that can make a difference.

Abhi: I think that was the other last question I had.

Ellen: Yeah. Thanks for the conversation.

Abhi: Yeah, absolutely. I am struggling to think of anything else. Oh, one hour 47 minutes. Okay. That’s not bad.

Ellen: That’s pretty good.

Abhi: Yeah.

Ellen: Cool. Awesome. Okay, ship it.

Here it is on Youtube:

And Apple Podcasts, and Spotify.

Introduction

The current in-vogue thing to do for most longevity companies is to go for partial cellular reprogramming. As in, fill a cell with the necessary transcription factors for long enough to reduce epigenetic noise, restore mitochondrial dysfunction, and so on, but not long enough to completely change the cells identity. I’ve written about the promise there before, it’s definitely an exciting field.

So, when I first met Alan— who told me that he was a longevity researcher — last October, I naively assumed he was also on the reprogramming train. But he told me that he was investigating something a bit different. His pitch was that, instead of reprogramming the cell to fix age-related damage, what if you just protected it from (genetic) insult first? It’s an obvious idea, but one that I’d never really deeply considered. He sold me on the concept, and I was very curious to hear what he’d do next to push it forwards.

A few months after our chat, he spun up a company to pursue this line of thinking: Permanence Bio, which develops molecules that stabilize/protect the genome. They are just about eight months old, but there are already some exciting results coming out. I’m a sucker for people doing ‘contrarian research in consensus fields’, and I immediately knew I wanted to have Alan on the podcast. He graciously agreed and, during my trip to SF last month, we sat down and talked for a few hours.

In this episode, we talk about why DNA protection is so important, what indications is it useful for, how to mentally conceptualize the idea of a molecule ‘stabilizing’ a genome, what it was like to raise money for a company pursuing such an out-of-distribution thesis, and lots more.

Finally, Alan has a really great blog (something I mention in the video), and I wanted to attach a much longer article he’s written about the topic here.

Timestamps

[00:00:00] Teaser clip

[00:01:39] Introduction

[00:07:32] What is Permanence working on?

[00:11:48] What does DNA protection actually look like?

[00:27:12] Why is DNA protection not focused on as much?

[00:41:03] The utility of epigenetic clocks

[00:46:47] Do you need multimechanism approaches for longevity?

[00:51:58] Longevity outside of DNA protection

[00:55:57] What's going on inside of Permanence?

[01:05:54] How could Permanence fail?

[01:09:03] How do you stay optimistic?

[01:10:26] Why work on aging?

[01:15:26] What are you bearish on?

[01:19:12] Weirder types of aging beyond 110

[01:21:37] How did you decide on DNA protection and what else would you have done?

[01:25:27] What was it like raising money?

[01:31:48] What do you think of past cancer prevention trials?

[01:34:12] What does good wet-lab talent look like?

[01:37:02] What does your information diet look like?

[01:40:06] What's it like going from research to being a CEO?

[01:42:20] What happens after cancer prevention for Permanence?

Transcript

[00:00:00] Teaser clip

Alan: The fascinating thing is that a lot of organisms have mechanisms for preventing DNA damage. If you look at epigenetic clocks and how they measure aging, you can create an epigenetic clock that measures aging consistently across tissues and cell types. That was Steve Horvath's big finding in 2013. But if you take his clock and you apply it to sperm cells, it's always off. The sperm cells either don't seem to be aging or are aging 70% slower than every other cell type.

Abhi: This is something common with germline cells, right?

Alan: Yeah, exactly. Oocytes have similar things, although oocytes sort of cheat a little bit. I think there's a really cool paper that showed they just turn off part of the electron transport chain and so they kind of get into this hibernation stage. But sperm don't get to do that. They have to keep dividing, they have to stay alive, they have to stay metabolically active, and yet their genomes don't seem to age nearly as much as other cell types and tissues.

Escaping humans a little bit, there are these deep-sea bacteria that live in temperatures that are crazy high and very damaging to DNA. They create small molecules that protect the genome. Of course, there's the naked mole-rat story too, where they have these long lipids that protect DNA from damage. So it's not like an unthinkable idea. We're not building something that has never existed before. I think we are just taking the best parts of everything and seeing what we can translate into humans.

[00:01:39] Introduction

Abhi: Today, I'll be talking to Alan Tomusiak. He has a PhD in the Biology of Aging from the USC Buck Institute and is currently founder and CEO of Permanence Bio, a longevity startup developing genome-stabilizing therapeutics. Today, we'll be talking about DNA protection mechanisms, why preventing somatic mutations will be extremely important for longevity treatments, and much more. Welcome to the podcast, Alan.

Alan: Thank you. It's amazing to be here.

Abhi: So I think my first question is to set the broad picture of what we'll be talking about today. One of the theses for the field of longevity at large that you seem to be really interested in is that protecting the genome is of far higher importance than almost any intervention we can do. You have a really great blog post that walks through all the reasons you believe this, which I'll attach to the description of this video, but I'd love for you to just give that as a recapitulation.

Alan: Yeah. On a high level, there's no such thing as a more important or less important part of aging. I think almost any biologist who's working in the field of aging will tell you that all of the hallmarks of aging exist. I think the one dimension where genome stability has a certain degree of primacy isn't the question of what is the order of events. So you have different hallmarks of aging. You have extracellular matrix breakdown, mitochondrial breakdown, genome stability, and epigenetic alterations. But the question is more, it's not so much which of these is most important to aging, but which happens first. When you're 20 and 30, what are the first things that go wrong that lead to the other things? What is the sequential order here? In that case, it seems to me, from everything that I've seen, that genome stability stability seems to be a first thing that breaks down.

There are a few different areas from which you can look at this. For example, one Twitter thread that I want to write at some point is looking at every single hallmark of aging and how a genetic defect in each hallmark of aging looks like aging. For example, there are people out there who have a mutation in elastin. You can ask, if you have a mutation in elastin, how much does that look like aging? Of all the progeria, the accelerated aging disorders, every single one of them has a mutation in some DNA repair or some genomic maintenance gene. So I think that's how I think about it. I think about it less in terms of what is the most important, and more in terms of what is the causal relationship and what is the most causal upstream factor for aging.

Abhi: When I say important, I do mean that there's this base layer upon which all the other hallmarks of aging seem to derive some sort of causal link. If you have some somatic mutation, it causes this hallmark of aging. Is it correct to think of it that way, where somatic mutations genuinely are the cause for the six other hallmarks?

Alan: I'm very careful about saying genome instability. Somatic mutations on their own are a very controversial field. To what extent somatic mutations, specifically a change in the DNA code versus damage happening to DNA and then all the downstream effects of that. As an example, the somatic mutation field used to be one of the most dominant theories of aging. More recently, there's been a lot of work showing that there are individuals out there with certain mutations in a polymerase enzyme that leads to far higher rates of somatic mutations. They do not look like they have accelerated aging. Specifically, I think it's polymerase epsilon, and they have a 15 times higher basal somatic mutation rate in the colon. Sure enough, they get a lot of cancer, but they don't age faster.

There are still ways that you can think around this. I wouldn't say the somatic mutation theory of aging is dead necessarily. There are people who, you know, there are certain papers that are really compelling, showing there are selective mechanisms. In clonal hematopoiesis, you have that advantage of certain somatic mutations that then leads to aging-like phenotypes. There was recently a very fascinating paper that came out looking at chronic kidney disease, and in mice, they show that somatic mutations lead to defects in, I want to say the laminin gene, that then leads to dysfunctional senescent cells that look a lot like aging cells. But I think overall, the state of the evidence right now doesn't quite support the fact that somatic mutations are the cause of aging. I would say it's one factor upstream. The way that I think about it is that you have DNA damage that then leads to somatic mutations, which leads to cancer, but you have DNA damage that leads to other things, including epigenetic loss of information, and that is more concretely an aging phenotype.

Abhi: One potential mistake I've been making is I'm equating genome instability with somatic mutations. Is that incorrect? Is there actually some nuance there?

Alan: There is a lot of nuance. I'm thinking, is there a way to rephrase genome instability? Because a lot of people will say, "Oh, genome instability, mutations," but genome instability means a lot of things. Even on a meta-level, thinking about the 3D architecture of the DNA, that changes so much. Senescent cells will just throw out DNA from the nucleus into the cytosol. That is definitely genetic instability; that's not a mutation. It's not sure whether it's a controlled process or not, but telomere attrition is a form of genome instability in its own way. You can think about abasic sites, crosslinks, double-strand breaks. All of these things are derived from genome instability, but they're not mutations. So I would take it one step higher level than looking at mutations specifically.

I do care a lot about mutations. I think mutations are incredibly important from the point of view of preventing cancer. For what I'm working on right now, to some extent, mutations are a good enough proxy for genome instability because at the end of the day, genome instability and mutations both cause cancer very directly. Therefore, if you prevent one, you're fine. But I think if we're thinking on a bigger level in terms of aging, it's not quite somatic mutations. I think it's one step higher.

[00:07:32] What is Permanence working on?

Abhi: So the primary lens by which Permanence is working is trying to prevent somatic mutations from occurring in the first place?

Alan: I would say so. Of course, if you can prevent other types of DNA damage... as an example, one thing that comes from genome instability is inflammation. A cell picks up the fact that the genome is damaged and then it releases all sorts of different factors that lead to aging, but also cancer creation. I would say mutations are a very, very useful readout for us, specifically because it's also relatively easy to measure, and it's a binary thing. You have a mutation or you don't have a mutation. How many mutations do you have? It's relatively easy to interpret. If you protect the genome, then you shouldn't be getting as many mutations. But it's one of many things that we care about.

Abhi: For Permanence, you guys are building chemical treatments to prevent the genome from undergoing mutation. Could you walk me through the general lines of medicinal chemistry reasoning when developing these sorts of things? What are the primary directions you've historically seen in the field and where do you think the field is going?

Alan: That's a very good question. Where the field is going or has been going is finding ways to boost DNA repair. I think that's by far the most common thing. What people have looked at is they found all sorts of different natural compounds that improve DNA repair, and certain drugs that seem to increase the fidelity of DNA repair. I think that's been by far the most common way in which people have gone at the problem. We're doing things a little bit differently in that we're trying to actually prevent the damage. Instead of trying to improve repair, we're seeing if we can just stop the DNA from being damaged originally. In my mind, this is going down the key theme of how upstream can you go. The most upstream way of preventing genome instability is to prevent the damage in the first place.

Abhi: It feels like at least some of the sirtuin stuff was concerned with trying to boost natural DNA repair mechanisms. It doesn't seem like it actually worked out particularly well. One, is that correct? And two, why do you think it is the case that boosting existing DNA repair mechanisms hasn't empirically worked all that well, or have they actually worked well?

Alan: That's a very good question. There's a lot of interesting work in this area. The sirtuin story is definitely a compelling one. Most recently, the DREAM complex story out of Bjorn Schumacher's lab is very, very exciting. I think the biggest challenge for everyone building in that field is that DNA repair is so intrinsically tied to cell cycle state or cell cycle arrest. If a cell knows that it is about to divide a lot, then there are certain mechanisms that it turns on in terms of DNA repair and certain ways that it turns off. If you're boosting DNA repair a lot, that is a pretty clear signal to a cell that it should probably not be dividing because clearly there's damage happening.

The most direct example outside of the therapeutics realm is looking at p53 mice, where if you turn up p53 in mice, the idea was, "Oh, you prevent cancer because elephants have like seven copies or whatever." The result is that these mice are actually less healthy. They do get less cancer, but to my understanding, they have hematological deficiencies because the cells that should be dividing a lot are not dividing because they're recognizing, "Oh crap, there's a danger signal here." That's just across the board, no matter what DNA repair complex you look at, there is definitely some tie to cell cycle state and stress state. I think that's been the main reason why it's been tricky. I don't think that it's impossible. For example, a number of companies now are thinking, "Well, if you look at cells that are already out of the cell cycle, what if you specifically boost DNA repair in neurons? Would that work?" I think that's pretty interesting. We haven't seen a big success yet, but we very well could.

Abhi: You are going the opposite route, well not the opposite route, but the route of trying to prevent the mutations from happening in the first place rather than fix mutations that have already occurred.

Alan: Precisely.

[00:11:48] What does DNA protection actually look like?

Abhi: What's your mental model for thinking about what protection actually looks like?

Alan: The fascinating thing is that a lot of organisms have mechanisms for preventing DNA damage. If you look at epigenetic clocks and how they measure aging, you can create an epigenetic clock that measures aging consistently across tissues and cell types. That was Steve Horvath's big finding in 2013. But if you take his clock and you apply it to sperm cells, it's always off. The sperm cells either don't seem to be aging or are aging 70% slower than every other cell type.

Abhi: This is something common with germline cells, right?

Alan: Yeah, exactly. Oocytes have similar things, although oocytes sort of cheat a little bit. I think there's a really cool paper that showed they just turn off part of the electron transport chain and so they kind of get into this hibernation stage. But sperm don't get to do that. They have to keep dividing, they have to stay alive, they have to stay metabolically active, and yet their genomes don't seem to age nearly as much as other cell types and tissues.

Escaping humans a little bit, there are these deep-sea bacteria that live in temperatures that are crazy high and very damaging to DNA. They create small molecules that protect the genome. Of course, there's the naked mole-rat story too, where they have these long lipids that protect DNA from damage. So it's not like an unthinkable idea. We're not building something that has never existed before. I think we are just taking the best parts of everything and seeing what we can translate into humans.

Abhi: To create a mental model of the system, you have this nucleotide and phosphate backbone, and you have potentially reactive oxygen species kicking nucleotides out. You have the background radiation of the universe damaging the nucleotides. What are these phospholipids in naked mole-rats or any of these other DNA protection mechanisms physically doing to the DNA to prevent the damage from occurring in the first place?

Alan: It depends. There's a whole spectrum. Some of them don't interact with DNA at all. A lot of these compounds are just basically very powerful antioxidants that exist inside of a cell. The downside is that you need a very high concentration of them to do anything. You can imagine a cell is very big, and the amount of distance that a reactive oxygen species takes to actually hit the DNA is very, very tiny. So you have to be either very lucky or you have to have a very high concentration of antioxidant to see an effect. Or you have to localize it to the right place. Some compounds are very directly DNA binding, so they will change the 3D structure of the DNA in order to make it more resilient against damage. Or they'll compact it in some ways. Again, you kind of need high concentrations of the compound for that. But for a lot of these, especially bacteria, that's the interesting mechanism.

Abhi: Potentially a dumb question. If DNA is compacted into chromatin, is it more resistant to damage or less resistant to damage?

Alan: It's more resistant.

Abhi: Okay. I guess that makes some intuitive sense. On one hand, you could perhaps build better antioxidants. On the other hand, potentially you find some ways to stabilize the phosphate backbone, or interact with the chromatin in some way. What is the primary axis of variation and what is something you think is the most important thing to poke at?

Alan: This is a very complicated answer. I'm looking at it from both a top-down and a bottom-up angle. The top-down angle is not even thinking scientifically, but thinking about indications. For example, if you have a human disease, which human disease can you treat such that a genome-stabilizing compound will be effective? That's what landed me on cancer prevention. Cancer is a very debated topic to what extent genome stability is relevant to aging, of course. I'm on the side that it's very, very important. But for cancer, it is not disputed. Cancer seems to be largely driven by mutations and inflammation. If you could prevent mutations, then there's a lot of cancers that you can stop almost entirely. That became a bit of a no-brainer. Plus, the sheer impact that you could have on people by modulating and improving DNA repair for cancer prevention is huge. We're kind of driven by that as well.

Abhi: When you were developing DNA protection molecules, how are you actually testing that whatever protection you're conferring to the molecule actually works? Is it just deep sequencing of single cells or is there some other proxy measurement you use?

Alan: Everything. This is where Permanence Bio really stands out and what we're doing is very special. The biggest challenge in the genome stability and DNA repair space has been figuring out how you screen and test things. If you just blast cells with an enormous amount of radiation and see which ones survive, the hits that you get are not DNA repair-related. They're hormesis proteins that shut down the cell and protect it from doing anything. They completely shut down the cell cycle, they completely shut down everything. So how do you tune an assay to tell you the genome protection mechanism, but not anything else?

People look at gamma-H2AX a lot. For listeners who don't know, when there's a double-strand break, there's a little marker on a histone nearby where the cell tags it. That tag is a sign that there's damage happening here, and that recruits a ton of DNA repair machinery to go to the area. We screened a whole bunch of compounds in the beginning, basically trying to compare the things that we're developing versus the best in class that other researchers have found. We found that some of the best DNA repair molecules are not only not hits, but they actually increase the gamma-H2AX signal. For example, nicotinamide, which is very well known to be a very good DNA repair booster, I think canonically accepted as such, and we see a 20% increase in gamma-H2AX signaling after we induce damage in a cell. You'd immediately think, "Oh, that's terrible. That means that you're actually increasing DNA damage," but you're not. You're actually improving signaling. That's also a good thing.

The long way around to getting to your question is that we're doing a lot of things. Gamma-H2AX is one of them because if we can actually reduce the damage, then we reduce the damage signaling. Therefore, that's good. We're also looking at broader things. We're experimenting with hematopoietic stem cells. They're very genomically unstable. In the bone marrow, there's a very small percentage of oxygen, so it's very easy for them to exist there. As soon as you take them out of the bone marrow, suddenly they're exposed to a ton of oxygen, so their genome instability jumps up. If you have a compound that can, for example, improve genome instability in hematopoietic stem cells, they should be able to grow better. You should be able to see a whole bunch of different flow markers that indicate the cell is in a healthier state, and they should remain high even as the cells are dividing in these unnaturally high oxygen conditions.

There's also a diversity of insults that you can do. People like radiation a lot because that's fairly direct, but of course, the type of damage that you're seeing with radiation is not the type of damage that you see with formaldehyde, which is not the type of damage that you see with H2O2, which is not ENU. So what we're trying to do is a whole battery of things at the same time and then see if we can get hits in everything.

Abhi: Is there a way to discretize different types of DNA damage, or is it kind of all on a spectrum and it's hard to categorize what is what?

Alan: This is also a very good question. I think there are two ways to do it. One is you have insults that you know should be causing mostly one type of damage. There are very good papers that have mapped out every single type of DNA damaging drug and roughly what type of DNA repair mechanism it's impacting. So you can just screen across, let's say, six or seven insults and say, "Hey, which one of these seven is our compound most effective against?" The only downside to that approach is that a lot of these will not be very specific. Some will damage one but also another mechanism, but it's very efficient. The other way you can go about it is you have some cells and then you knock out their ability to repair one type of damage. Then you cause general damage and then you see to what extent your compounds can rescue those particular cell lines. I think both approaches are good. We're leaning more on the first.

Abhi: Is there a particular cell type that you're most interested in working on, or is Permanence Bio really spread across many different cell types?

Alan: It has to be many different cell types, especially for cancer prevention. What I've noticed is that there's a big distinction between how slow-dividing cells, non-dividing cells, and fast-dividing cells deal with DNA damage. As an example, I mentioned earlier that there are people out there with a disease where they have a mutation in a polymerase that then leads to a 15 times higher mutational burden. You mostly see that mutational burden in the colon and very fast-dividing cells. They primarily have very high rates of colorectal cancer. In contrast, if you mess up another genetic pathway that leads to DNA repair in humans that leads to Cockayne syndrome, they don't get colorectal cancer at all. Instead, it's a major neurological disorder. These are mostly impacting non-dividing cells, so it's mostly autoimmune and neurological. Werner syndrome impacts a third batch of DNA repair, and there you see slow-dividing cells get messed up. These patients have something that looks a lot like diabetes. Their skin starts to look wrinkled, and it's the most physiologically accelerated aging-looking model that I've ever seen. So we need to be effective against many different cell types and many different types of damage.

Abhi: How should I think about DNA damage not in the context of environmental insults, but in inherent genetic defects in an individual? The example that I'm thinking of is someone who has a BRCA mutation and is unable to repair double-strand DNA breaks. Do you imagine something like what Permanence Bio is doing is amenable to fixing that problem, or if there is a genetic defect that is causing DNA damage, is that a trickier problem to fix?

Alan: No, we think about that a lot, mostly because when we think about clinical strategy. If we want to go big picture for cancer prevention, to run a cancer prevention trial in the general human population without stratifying populations at all, that's somewhere north of a hundred million dollars and five years and tens of thousands of patients. Looking at individuals who have a high risk of cancer is absolutely something that we're deeply interested in. There are sort of gradations here. You mentioned BRCA carriers have a much higher risk of breast cancer. We also have thought a lot about Lynch syndrome, where these patients have a 70% lifetime risk of colorectal cancer. So that's an even higher risk of cancer. Then there are more rare or ultra-rare diseases like Fanconi anemia, where they're guaranteed to have cancer. As we're thinking about our path to clinic, we'll very likely want to start at these higher-risk patients and then gradually move on to lower-risk patients to prevent cancer in them as well.

Abhi: The larger question is, are there some forms of DNA damage that aren't amenable to these protection mechanisms, or are all forms of DNA damage amenable to DNA protection?

Alan: In theory, you can only get DNA damage by virtue of having some insult that causes DNA damage. Some are trickier than others. The main one that we think a lot about from the point of view of this question is replication stress. Is replication stress inevitable? Or is it purely dependent on polymerase fidelity? If it's only polymerase that can cause or contribute to replication-mediated DNA damage, then DNA damage protection mechanisms are not going to do a whole lot. It's going to have to be, at that point, some sort of gene therapy or some sort of interesting drug that boosts DNA polymerase fidelity. But aside from that, no, I think DNA damage is pretty central and will impact everything.

Abhi: There are some scenarios in which learning some aspect about a certain drug modality is fine to do in an in vitro setting because it transfers pretty well to in vivo settings. I think one common example people use is that you don't typically need to worry too much about antibody toxicity, so it's fine to test in vitro, and it'll probably translate in vivo. With DNA repair mechanisms, what is the situation like? Is in vitro usually perfectly fine to test everything in, and you expect that it'll translate pretty well, or are you primarily testing in animals and even then there are worries that it might not work across species?

Alan: That's a very good question. I think you learn a lot from cells. For example, especially from the point of view of things that could go wrong. Things that could go wrong with DNA repair, a lot of them have to do with cell cycle arrest, as I mentioned. There's this deep, entrenched talk between not replicating and doing DNA repair. You can do experiments in rapidly dividing cells, activated T cells, hematopoietic stem cells, you name it, in a cell culture model. That will be really, really informative as to whether or not you'll see that kind of problem in the mouse. So there are a lot of things that are going to be helpful.

Others, less so. My big questions are kind of like the ones I asked before. Are there certain types of cancer that are going to be more amenable to treatment than others? That's hard to tell unless you test against every single type of cell in an in vitro model.

Abhi: I've always been a little bit curious, and again, potentially a very, very dumb question, how well does in vitro cancer mirror real-life in vivo cancer? What are the big distinctions there? What are you not accounting for, and what are you accounting for?

Alan: The immune system is the big thing that you miss. That's the toughest one. I kind of mentioned before, cancer is largely described by two things: inflammation and mutations. By doing an in vitro model, unless you have a very sophisticated organoid model, you're missing the inflammation part.

Abhi: But does the immune system still matter for the problem of DNA damage protection?

Alan: Oh, absolutely. There are even cases where, let's say that you don't prevent mutations at all, but you modulate the immune system's response, and you have a huge effect. One thing that people, even I, have thought a lot about is autoimmune disorders. A lot of those are driven by genome instability, but they're not driven by genome instability because the cells themselves are dysfunctional. They're theorized to come from the cell sending out dangerous signals to the immune system, which then comes in and floods and causes chronic inflammation. So the actual DNA damage and mutations barely matter. It's just this overactive immune system cycling that is causing the issue there.

[00:27:12] Why is DNA protection not focused on as much?

Abhi: Speaking a little bit big picture, you make a pretty strong case for why preventing DNA damage is really important. But it does feel like historically the longevity field at large has not focused on this particular subfield and has focused a lot more energy on, I think, largely epigenetic reprogramming and also lifestyle changes, like eating less and exercising. Why do you think epigenetic reprogramming has been given the spotlight, and why has this particular subfield of yours been kind of left to the wayside?

Alan: Epigenetic reprogramming is really, really exciting because it promises reversing aging. If you're already pushing your later years and you want to live, especially if you're motivated by living forever, epigenetic reprogramming is one of your very few choices. The field of replacement is coming out that's becoming more and more exciting, where the idea is you just replace body parts. But reprogramming, I think, is the most mature field that has that idea of, "Okay, maybe you can do something right now and then live forever." There's this whole concept of longevity escape velocity. If you're going to hit that, you probably need something that will reverse aging, especially for the people who are already older. That's the biggest reason why reprogramming is exciting.

In addition to a few mouse experiments that I think were really, really compelling. The biggest reason why genome instability is toughest to reach, and we think about this a lot in terms of clinical strategies, is that prevention in general is hard. Prevention is expensive. Prevention takes a long time. By definition, it is very hard to do. Most genome instability diseases are going to be inherently prevention-oriented because in most cases, once the genome instability has already occurred, either through mutations or inflammation, it is not something that you can reverse. Maybe you'll slow down the progression at best, but you're not going to reverse it by preventing more damage. The damage is already there.

You can think about this in the simplest form, like Huntington's. If you already have the crazy high somatic expansion, if you already have a lot of copies of HTT—Huntington's is caused by an expansion of a region in this gene called HTT—then preventing more expansion isn't doing anything for you because you already have all that. Most of the thinking around how you treat genome stability relates to prevention, and prevention is just so much tougher to do.

Abhi: If I think about what the major prevention drugs are, I'm thinking of anti-hypertensives and statins. But even then, those aren't a close analog to this because they have pretty easy biomarkers to tell quickly how well this is working or not. Is there any drug class in history you can point to as what you think the parallel journey for Permanence is going to look like?

Alan: Statins are actually a great one. The most direct are vaccines. That's the easiest, right? There are two drugs approved for cancer prevention. One of them is Gardasil, which is the HPV vaccine. There's one other cancer prevention drug that exists to my knowledge, and that's Tamoxifen. They showed that Tamoxifen is effective as a breast cancer treatment, but then there was a huge study that showed it can also prevent breast cancer. The clinical path has already been set. Statins, GLP-1s, these are actually very similar to what we're trying to do in that they were originally approved for very narrow indications, for people who are at very high risk of a disease. Then gradually, as people saw that they were also safe and effective across a whole broad range of different indications, they gradually label expanded. It's the same thing that we see here in cancer prevention where we'll probably start out with people who are at very high risk of disease and try and prevent them from getting cancer. But then slowly, as these drugs are shown to be safe and working in more and more people, we just expand.

Abhi: I'm curious, for Tamoxifen, I was not aware of that. Do we know what's going on there? Does it have a DNA protection mechanism, or is it doing something else entirely?

Alan: I can speculate. What's known is that Tamoxifen modulates the estrogen receptor signaling. That's also why it's a drug that prevents breast cancer. The conventional wisdom for breast cancer prevention is that in women, you have this idea of a pre-cancer, where it's not quite cancer yet, but estrogen signaling can help push it over the edge. If you block that signaling, then you prevent it from going over the edge into being cancer. That's why it's preventing cancer.

With that being said, there's a lot of fascinating overlap between estrogen receptor signaling and DNA repair. There are only hints of it out there. As an example, if you look at GWAS hits for the age at which a woman undergoes menopause, almost every single hit is a DNA repair gene. It's strong; these are massive effect sizes. There's one very common variant in the GWAS hit where if you have it, you undergo menopause a year and a half earlier or a year and a half later, regardless. There's a very strong correlation, a link between estrogen signaling and DNA repair.

Abhi: Does this imply that if you're a woman and you're above 40 or 50, you should be on Tamoxifen?

Alan: Probably not. To be honest, I'm not knowledgeable enough in that field. Menopause itself may or may not be estrogen receptor signalling mediated.

Abhi: I'm curious, and related to this whole cancer tumor thing, do you imagine that DNA protection chemicals would also help in preventing cancer from adapting to whatever therapy you throw at it? Or is there something about cancer that's a little bit weird, and these methods may not necessarily work for the cancer itself?

Alan: Yeah, I've thought about this a lot: can you prevent cancer progression? In fact, there's a very interesting study done with nicotinamide, which I mentioned earlier as a pretty good DNA repair drug, in terms of preventing, I want to say, secondary malignancies from a melanoma treatment. It was effective. It worked. The patient population was 400 patients in each arm. The trial was a year and a half, and they showed that sure enough, if you give patients who had melanoma once, you have a 30% decrease in the risk of a second malignancy two years later. I think it'll depend a little bit on the cancer. Some cancers are known for getting a lot of mutations throughout the course; some cancers don't. For a lot of cancers, progression is more defined by immune cell exhaustion than it is by adaptation of the cancer itself. So I think it is going to depend a little bit on the specific cancer type.

Abhi: Do you think the first indication for Permanence is going to be for stopping the progression of existing advanced cancer or for preventing cancer from occurring in the first place amongst some high-risk population?

Alan: Probably the latter.

Abhi: Why not the former?

Alan: The former is exciting. When I think about indication selection, I think about reasons you could fail. If there are patients out there who are going to get cancer and we have a drug that should prevent mutations, and they're getting cancer from mutations, then it should work. If we fail, then that means that our drug just didn't work. That is a very clear indication. If the indication is prevention of progression, progression can happen for a lot of different reasons. I would hate for our drug to work really nicely and prevent mutations and then that not impact the endpoint.

Abhi: There are a lot of reasons a patient could die for reasons unrelated to mutations. What were the other things on the table for indications other than cancer?

Alan: We've thought a lot. If you look at which drugs right now are genome-stabilizing or antioxidants, for example, things that prevent damage, to my knowledge, there are two indications right now that people are pursuing. Epsilon Edavarone is a drug for ALS, and neurological diseases are very well known to have DNA damage as a driving mechanism. ALS was interesting to us, especially because there's already been a clinical success for an antioxidant there. Dry AMD, currently the main treatment, or not even treatment, but the main way that you prevent that is through this cocktail of antioxidants. When we think about what is something that we can move into where we see there's already success, there's already a path, those two stand out because that's been done before.

From there, there are other indications where there's a very clear causal link to genome stability. I cited Huntington's before. Alzheimer's is looking increasingly interesting, and kidney disease, where there isn't currently a beaten path, but genome stability is clearly related to progression or initiation of the disease. Those were interesting. I think the main thing for us was mission alignment. Our vision is to do cancer prevention and longevity. What a lot of longevity companies, I think, run into is that the way that you treat a disease ends up looking a lot different than the way that you address longevity. What I mean by that is that, let's say that you have a drug that you want to put in as a cancer therapeutic, and you think that it also has a pro-longevity mechanism. The downside to that is that the way that you think about a cancer therapeutic is very different from a safety point of view. For example, you can dose extremely highly if you have a cancer therapeutic because otherwise the patient will presumably die. That is not something you can do for longevity. You want to have the lowest possible dose where there are no side effects and the patient's just healthier.

When I think about overlap with the mission of doing something that's pro-longevity, cancer prevention is really well aligned. Similar to longevity, in cancer prevention, what you're trying to do is prevent a new disease from occurring. You want something extremely safe. You want something that you can take orally. You want something you can take for a very long time. You want something that has minimal drug interactions with every other drug. In terms of just thinking about the problems that we're solving on one path versus the other, they're basically the same problem. We like that as a case, but of course, we could go into multiple programs. I don't think we're married to only doing cancer prevention.

Abhi: On the topic of actually taking the drug itself and it systemically going throughout your body, if you do go with the cancer route, do you imagine your drug will be injected at the site of the cancer? In the case of taking the drug to prevent cancer in the first place, you have this systemic application of the therapeutic. Where do you want the therapeutic to get into most? Do you want it to get into hematopoietic stem cells? Do you want it to get into muscle cells? Where is the primary place that you think is important either for longevity or for cancer?

Alan: It really depends on the indication. For example, if you look at xeroderma pigmentosum patients, they have an extremely high risk of cancer. It's a really rare disease that's mostly caused by sun damage. You really, really, really need the drug to get into the skin. For Fanconi anemia, these are patients that are mostly getting hematological disorders and squamous cell tumors. In that case, you want it to get to the bone marrow, but you also kind of want it to get everywhere. If you're thinking about more common diseases, Lynch syndrome, that's the colorectal cancer I mentioned earlier. You need it to get to the colon. So it really depends on the indication in question.

Abhi: To some degree, when you say the indication for Permanence Bio is cancer, it's probably going to be more specific than just cancer. It's going to be like a very specific organ, for example.

Alan: In the beginning, I imagine it's going to be a lot more specific.

Abhi: Do you think we'll ever get to the point of having this systemic negative 10% reduction in cancer across all types of cancer because of genome-stabilizing drugs? Some context for this question: I know that for at least epigenetic reprogramming, different cells reprogram at different rates. There's this weird problem where if you want to reprogram a neuron, you need to almost go slower in the reprogramming process versus a muscle cell. Is there something akin to that for genome stabilizers where different cell types need a different level of dosage?

Alan: That's a good question. On a very, very high level, yes. The cells that are going to be accumulating a lot of mutations tend to be the ones that are dividing a lot. You want to get into your bone marrow, you want to get into the colon. From that point of view, yes.

The other thing is, and this is both a blessing and a curse for us, when I mentioned that we have this cancer prevention/longevity alignment, one of the biggest things for us is safety. We need a drug that is fine in everything. If it gets into any cell type, no matter where it works, it should still work. That's very, very important. And that's why we're testing so many different cell types, because if we find out that we're working in four cell types but we're toxic in the fifth one, we just cannot pursue it.

[00:41:03] The utility of epigenetic clocks

Abhi: You did a lot of research on epigenetic clocks during your PhD, but from just scrolling through your Twitter feed, you're pretty pessimistic on the overall utility that they can provide. I'd love to hear your take on the current state of things in the field and perhaps on biological clocks more generally.

Alan: I've been thinking about this a lot lately. I think clocks are very promising, perhaps not right now, but probably very soon, from a clinical biomarker point of view. So looking at, "Hey, we have this intervention, we give it to patients, what do the clocks say after two years?" I think there's a primary challenge there in that a lot of these clocks are difficult to interpret, but I think slowly the field is solving that problem and their power, the clocks are getting strong enough that they have real meaningful predictive power for functional endpoints that would just take too long to get to using a conventional trial format. I'm very bullish on that front and excited to see where that goes.

That's especially helpful for prevention. What would be nice is if we had a compound that is not only cancer-preventing but also longevity-extending, and you can tell that with a biomarker off the blood that you trust, so you don't have to run the full, long cancer prevention trial. From a prevention point of view, I think biomarkers and clocks are a godsend. As basic research tools, they're also very helpful. Most of my PhD was trying to use epigenetic clocks to understand to what extent aging is a mechanism of cells getting old and not working versus a mechanism of the system collapsing.

The place where I think I'm a little bit more pessimistic is in direct-to-consumer, like an individual person takes a clock. I'm not sure to what extent that is very actionable and it doesn't seem very accurate. A lot of the technical noise there is higher than anything that you could possibly impact. Unless you have the money to buy a hundred of these things, there's not a whole lot of value that you get out of buying a genetic clock off the internet. I've never done my own tests. That's where I'm pessimistic. But in terms of where the field is going with the biomarkers of aging consortium and looking at clinical trials, I'm very, very excited.

Abhi: How much does the biological clock you use matter based on which therapeutic you're throwing at the problem? I imagine for the Retro and Altos of the world, who are all epigenetic people, epigenetic clocks are the most useful for them. Are they also useful to you, or are there some genome-stabilizing clocks that you're more interested in?

Alan: That's a cool question. I think we're getting to a point where you could actually use mutational burden as a clock. We are getting there. We're not quite there yet, but I think within the next two or three years, we'll start seeing that. I think what's most exciting are the systems clocks.

Abhi: What are those?

Alan: Basically, there are I think five papers out now, but there's a flurry of them. So far, everyone's been mostly looking at epigenetic clocks in terms of blood. You take someone's blood sample and then that tells you their epigenetic age. But the question is, you're not—there's not just one dimension to aging. There are 60-year-olds out there who have perfectly healthy lungs, but they have a failing heart, or vice versa. The question was, can we use epigenetic clocks or proteomics clocks to understand to what extent someone's at high risk of a certain type of disease for a specific tissue or organ? Those are exciting. When you're thinking about diseases, those are very, very good because let's say that you have a drug that's supposed to improve sarcopenia, well, maybe the muscle clock is going to be the most effective at predicting how well the trial will go a year before you hit the endpoint.

Abhi: Let's say you meet a hundred people, all of whom are 110 years old. Do you suspect there is a primary protective mechanism that all of them have, or is it likely very heterogeneous among them?

Alan: That is a good question. There's this saying in the longevity field that I like a lot, which is, "If you want to live to be 90, what do you do?" And the answer is, "Oh, you should diet and eat well and exercise and sleep and do all these healthy things." And if you want to live to 110, what do you do? It's, "Have parents that live to be 110." If you look at 110-year-olds specifically, I think it is a genetic improvement in genome stability. I think that the aspect of aging that fixes your lifespan is likely mostly genome-stabilizing. That would be my prediction. There's some evidence to support that. People have found there is a variant in, I believe, (SIRT6), which is a known genome-stabilizing protein, and that is enriched in people who live a very long time. But I think there aren't even enough individuals to really have a strong enough study to predict that.

Abhi: I suppose an easy way to study this is to look at the average mutational burden of some sample among these hundred people who live to be 110. Is there some reason that is not a trustworthy number, or is it just that we don't have enough samples?

Alan: We do not have enough samples.

[00:46:47] Do you need multimechanism approaches for longevity?

Abhi: Somewhat related to this, it feels like longevity startups in general have pursued singular mechanistic hypotheses at a time. Some groups are pursuing only replacements, some groups are pursuing only epigenetic reprogramming, and you're pursuing DNA stabilizing stuff. It has long felt to me that aging almost wants to occur, and so if you plug one of the seven holes that cause aging, the pressure on the remaining six holes will simply intensify. Is this an accurate mental model of aging as a concept? And two, do you imagine you'll be forced to move into other therapeutic areas for longevity, or is genome stabilizing a really good base to start off on?

Alan: My own mental model is almost the opposite. People have shown that a lot of these hallmarks of aging, these aging dysfunctions, feed off each other. If you remove one, you'll probably actually release the tension off a lot of the other ones. As an example, inflammation leads to dysregulated nutrient sensing because suddenly your immune cells are eating all of the different nutrients that other cells should be getting, which then causes mitochondrial stress. Mitochondrial stress induces genome instability, which leads to telomere attrition. If you just didn't have the inflammation in the first place, probably the other ones would be a lot better off too.

I definitely share the core sentiment that there are a lot of longevity companies that focus on one thing. Even when we think about genome instability, I think by the time that someone's quite older, let's say 85-plus, doing genome instability alone isn't going to be enough. In fact, I'm not sure to what extent you'll see a big effect. That's kind of why we're focusing on this prevention angle, because the hope is that early on, the primary thing that's affecting you as you age is genome instability. So the earlier you treat that—for example, inflammation is not that much of a thing when you're 25 or 35 or 45 or 55. It only becomes a real hallmark of aging past the age of 65. It probably doesn't make sense to prioritize genome-stabilizing drugs for someone who's 85. At that point, I would probably hit the other hallmarks. But for someone who's 30 or 40 or 50, absolutely, that would make a huge difference.

Abhi: A very, very dumb question. When we talk about a cell's ability to repair its own genome, can that happen hours after the insult has taken place? Or is it kind of like the repair mechanism is only available a few seconds after the damage has been done, but after that the information is lost?

Alan: That's a very interesting question. It depends. The concrete answer is all of it. What I mean by that is there are stem cells out there, quiescent cells that get DNA damage that just sit there with the DNA damage for months. Because if they're not doing anything, that doesn't really matter. But as soon as they emerge from quiescence, suddenly they turn on their DNA repair machinery. They open up their DNA, all the DNA repair proteins come in, fix all of it, and so on and so forth. Versus there are certain types of damage, like a double-strand break, you better fix immediately or the cell is going to immediately become senescent and then die.

Abhi: The more specific question I'm asking is, you say if you're 80, we probably don't have anything to offer you. But can you not imagine that if you upregulate their DNA repair mechanisms and allow them to go through this rough period of cells not dividing because they permanently think they're in a permanent repair state, and then you give them the genome-stabilizing drugs, do you think there's anything there, or is it just like their cells have been through so much information loss that isn't recoverable?

Alan: The way I think about it is, my mental model of aging is like a boulder rolling down a hill. In the beginning, there's not a whole lot going on. The boulder's not moving very fast, and at the bottom, it's going at full speed. Trying to reverse entropy is tough. Inherently, that's just a fact of nature. Biology does have mechanisms for doing it, and that's kind of what the reprogramming field is pursuing, this idea of returning back to a natural state. It's just very, very hard.

You could definitely imagine some synergy, like a reprogramming therapy plus a genome-stabilizing therapy to kind of push someone back and then keep them there. I definitely can't say that we will have things to offer for people who are very advanced in age. Plus, if you stabilize a genome, there are just certain diseases that elderly people get that could be very helpful, like autoimmune disease, as I mentioned earlier, or Alzheimer's or something. It's just much harder at that stage.

[00:51:58] Longevity outside of DNA protection

Abhi: Outside of genome-stabilizing drugs, what is the second most promising area of longevity therapeutics that you're most bullish on?

Alan: I like reprogramming, but everyone likes reprogramming. I think just in general, hematopoietic stem cell treatments excite me.

Abhi: What is that exactly? What are you actually doing to those cells?

Alan: I mentioned earlier the replacement field is very big, and so I'm thinking about hematopoietic stem cells. If you were to take a human body and you were to replace something in them, and you were to weigh the cost trade-off of how hard is it to replace, how difficult would the surgery be, versus the maximum possible impact that could have in terms of rejuvenating that organ or that system, that's where I'm really excited about bone marrow and hematopoietic stem cells. They're responsible for your entire immune system. They're responsible for your entire blood system, and they're in a very relatively small niche in a relatively small part of your body. So if you could actually replace all of them, which is feasible, or at the very least, theoretically very feasible, then the benefit that you would see would be huge because you just rejuvenate your entire systemic circulation.

Abhi: Going back to this group of 110-year-olds, do you imagine that the vast majority of them have extremely detrimental somatic mutations in their hematopoietic stem cells, or is it kind of like if you're prone to having mutations in that area, you're probably not going to live to be 110?

Alan: The second one. There have been studies looking at this. There have been studies done on people who are centenarians and looking at their immune systems, and their immune systems are weird. As an example, there's this relatively rare subpopulation of cells in younger people or even elderly people, just everyone in general that isn't one of these crazy centenarians, called a cytotoxic CD4 T cell. CD4 T cells are usually helper cells, so in some ways, it's an oxymoron; they have cytotoxic helper cells. They have cells that kill and also help. This is a very rare cell type. It's weird, it's grossly understudied. People don't know what it does. It probably helps against cancer is one of the theories. It's a very understudied cell type. Turns out that there are people who are 105 years old where something like 30% of their T-cells are this bizarre, weird cell type that is maybe 0.01% in a healthy individual. Those people tend to live longer. They're enriched in these supercentenarians.

That could mean a lot of things. It could mean that they have a mutation or some genetic variation that leads, that makes them more predisposed towards having this cell type. It could mean that every other cell is just so vulnerable that they die off and this one weird cell type survives, so there could be a selection bias. It could be that as they age, they got a mutation in their hematopoietic stem cell system that led to them producing more of this very beneficial cell type. None of it's known, but the hematopoietic system does weird stuff in very old people.

Abhi: Is there anywhere else in the field of longevity where among the people who live a really long time, there's something unique going on about them biologically and it isn't just that they're better at doing the normal things that everyone can do?

Alan: Nothing immediately comes to mind.

[00:55:57] What's going on inside of Permanence?

Abhi: What is currently going on inside of Permanence? Obviously, we have discussed your thesis at length, but I'd love to hear about any details that you're able to share about what's going on inside.

Alan: A lot of it's what I've shared already. A lot of testing, a lot of screens, a lot of assays, just looking at the battery, ordering every single best genome-stabilizing thing that we could find and then putting them head-to-head and then doing some iteration on those and trying to improve them. Different cell types, different tissues, different types of damage, just doing the whole battery of assays. I think our first thing is just to look wide, what general patterns do we see? What are the most effective things in what disease models do certain types of compounds perform the best? As we kind of build up a database, as we build up a knowledge base internally, then that is going to be putting us towards specific compounds. We've gotten some incredibly exciting results lately where we have something proprietary that seems to be outperforming the best of the best that we've tested. So we're starting mouse studies, and that is a very, very exciting thing, especially a cancer prevention study because doing a cancer prevention mouse study is a lift. It takes a long time. So what we are doing a lot is asking, "What is the best model? How do we screen this? What do we look at? What do we model? Where do we look at mutations?" The details of how do we turn something that looks really compelling in cell assays into something three steps closer to being a drug that we can give to patients.

Abhi: When it comes to designing these DNA-stabilizing molecules, I tend to think of molecules and proteins as attaching to some cellular receptor. One way you can conceptualize how well this is going to work as they transfer from a mouse to a monkey is how structurally conserved this receptor is. Is there an analog to that in the DNA protection world?

Alan: Not really. It depends. Broadly speaking, yes. I kind of mentioned there were a few other genome stability companies that are looking at hitting things like DREAM or sirtuins or whatever. From that point of view, yes, you're still going to be titrating binding, you're still going to be finding, hopefully, the best possible binder that is going to be attached to this particular thing. If you have something that's mopping up damage, not really. You still have to do your dose curves to figure out what is the minimum dose that leads to the best improvement, but your readouts tend to be less biophysical, like "enzyme plus compound, how do they bind?" and they tend to be a lot more functional, of "you have a cell, how is it reacting molecularly to damage to this compound? Is it staying alive? What does it look like phenotypically?" and so on.

Abhi: When you're quantifying how well a particular molecule is working, is it kind of a binary yes or no on these functional assays, or is it a very wide spectrum?

Alan: It's a wide spectrum. The biggest thing that we're thinking about right now in terms of assay design is how do we get the cleanest separation between the things that we're testing? Some types of damage are just going to kill everything, and some types of damage are going to do nothing. You can't see a difference between your compounds. But how do you hit that sweet spot where you can see that some things are working but others aren't?

Abhi: The way that I've been conceptualizing DNA protection mechanisms is that they affect, for a given cell, the entire genome all at once. Is that a correct way of viewing it, or is it often the case that you want to protect specific segments of the genome at a certain point?

Alan: It depends on the mechanism, it depends on the drug. Some really do not do anything next to histones and they only have an effect on open areas of DNA; other ones, vice versa. If you're improving DNA repair, then certain types of DNA repair mechanisms are more involved in certain parts of the genome than others. So I think it's going to be very molecule-dependent.

Abhi: I imagine, just given what you've said about the DNA protection field, it feels like it's almost certainly relatively immature compared to most other longevity fields. As such, I imagine Permanence as a company is in an area where there's relatively little established ground for what empirically works, what theoretically works, etc. Have you learned anything that you can share about preventing DNA damage that perhaps runs a little counter to how biologists might naively think about the problem?

Alan: Yes. I will say that we're very lucky. In some ways the genome stability field is nascent, in some ways, it's very old. The idea that DNA damage is tied to aging is something that a lot of thinkers have thought on before. There's just an incredible number of people who have developed antioxidants and mouse models. I mentioned that we're screening the drugs that we're developing against the best in class. Every single one of those things that we're testing against was developed by teams of people trying to find the best possible protecting agent against something. There is a lot that we've learned from everything and everyone that has come before us in the cancer prevention space, in the genome stability space, and in the aging space.

In terms of what is something that's counterintuitive and different that we're doing, I think it is really the recognition that different types of damage lead to different effects in different cell types. I think there's been a very common trend of, "We're going to take some fibroblasts, we're going to look at gamma-H2AX, and just hope, cross our fingers and know that's going to translate to everything." I don't think that's how it works. I think what we've learned, where our intuition was and that has absolutely panned out in our assays, is that you need to look at a lot of different things to get a very big picture view of what DNA damage is doing and how you can prevent it.

Abhi: And there's no functional readout that you can trust above everything else.

Alan: Right.

Abhi: What is the most exciting internal result that you're currently seeing and why do you consider it exciting?

Alan: I mentioned this earlier. It's just the fact that we're seeing some proprietary compounds just work across everything. What I was really scared of, and this was absolutely possible, is that we would have four types of DNA damage that we would be screening to protect against, and then you would have like three compounds that would work for one, three that would work for two, three that would work for three, and three that would work for four, and then none of them would overlap. The fascinating thing is that we do actually see that with some of the control compounds. We're screening nicotinamide, we're screening astaxanthin, which is this incredibly powerful antioxidant that algae that salmon eat produce. To my knowledge, I think it's one of the most powerful antioxidants that has ever been known. We were screening a whole lot of different genome-stabilizing compounds, and we do see that in some of them where they're really effective in like two out of four or three out of four, but they're really harmful in the fourth one or something. I'm really excited to see that we have potential compounds that we want to move into mice that are helpful in everything. They're not necessarily number one in everything, but they're high up there in everything, and that's good.

Abhi: When you're designing these molecules at Permanence, how do you select the molecule in the first place? If there's no cellular receptor that you think, "I have this ligand that's binding to this one epitope," and you're inside this world of, "I want to bind to this backbone," or "I want to bind to the chromatin," what's your mental process for selecting these molecules in the first place?

Alan: A lot of things. We have a hypothesis for what makes a really good genome-stabilizing drug. We are hoping that that hypothesis is correct and is backed by literature. It is a little bit contrarian. We're testing out whether or not that hypothesis pans out in our cell screens and in other screens that we're doing. So far, it seems to. Given that it works, then the question is, can we improve upon that mechanism? We're doing a lot of things for that. For example, we're doing some first-principles AI prediction algorithms to see, given that this is our hypothesis, can we create new compositional matter that does that particular hypothesis better?

Then there's a second layer of AI that we have where the idea is every single assay that we have that is finalized and every single result that we get for every single compound, we feed that all into a massive database. The plan is in the next three months to go, "Okay, now that we have this huge amount of data and structural data on all these families of compounds, can we actually generate something new that a really powerful medicinal chemist could not have thought of?" We'll get there as well.

Abhi: When it comes to Permanence's usage of AI, the way that you phrased it implies that it's primarily LLM-based and not molecular models. Is that correct?

Alan: We do both.

Abhi: When you're using these molecular models, are you taking the DNA and trying to find something that binds to the DNA, or am I conceptualizing it incorrectly?

Alan: I'll not tell you. But it's very cool. I wish I could. I genuinely wish I could.

[01:05:54] How could Permanence fail?

Abhi: I think it's helpful to think about ambitious companies in terms of the bets they're taking because I think it's difficult to do ambitious things without taking a bet on where the future is going. Clearly, it is known that DNA damage prevention is important and that there are mechanisms that help prevent or fix it. What do you think is the bet that Permanence is making, or phrased differently, if Permanence completely fails, what do you suspect the reason is?

Alan: The clinical trials will be tough. The clinical trials for cancer prevention have been done, but they're all difficult and they all are expensive and they take a long time. The upside is massive, both in terms of good for humanity but also in terms of market. There are little weird things that can go wrong. For example, one of our advisors is Professor Dr. Sir John Burn. He's an absolutely incredible man. He has run the world's longest cancer prevention trial ever, clocking in at 20 to 25 years. He ran the CAPP series of trials, which were basically testing aspirin for preventing colorectal cancer in patients who have Lynch syndrome. Five years into the trial, they wanted to look at actually preventing colorectal cancer, but they had some secondary endpoints, and one of those was whether it could prevent polyps because polyps are on the path to progression of becoming colorectal cancer. Five years in, they saw no decrease in polyps whatsoever. They were really hoping that this polyp result would work because you don't want to wait until patients get cancer, because that's going to take another five years and be a mess. They almost killed the trial at that mark because they didn't see any decrease in polyp count. They let the trial run another 15 years, and the result was a 40% decrease in colorectal cancer in the patients who were taking aspirin versus the patients that weren't.

The bet that we're taking is that there's not going to be something ridiculous that falls out like that. A lot of clinical trials, especially ones that go long, tend to fall apart for reasons that have very much nothing to do with the mechanism or being wrong on the core biology, but just something else like that. That's where we're the most worried, I would say.

Abhi: Would you imagine the length of whatever clinical trial you run would be five years, 10 years, even longer?

Alan: Definitely not 10. It's got to be less than five. It's a little bit of a math problem. The more patients you have, inherently the less long it takes, but the more expensive it is. The more highly cancer-predisposed patients you have, also the shorter it takes and the fewer patients you need. We probably don't want to run a cancer prevention trial that lasts longer than three years. There are some cancer prevention progression trials that we can do in as little as one or two. But it's going to be a long clinical trial for sure.

[01:09:03] How do you stay optimistic?

Abhi: Working in the longevity field seems like an excellent way to grow pessimistic as to anything working out at all. You've posted also in the past on Twitter about how it is a shame that the ambition of the field has dramatically lessened than what it was 10, 20 years ago. How do you personally stay motivated in such an extraordinarily difficult field to work in?

Alan: It is just so worth doing. I got into the longevity field because I used to work in the late-stage cancer field, and I was just made aware of the idea that the way that we set up our healthcare system, the way that we treat patients, is that we wait until they get a very serious disease, and then we hope to bring them back from it once they're at the brink. Most of these drugs are operating on very specific mechanisms that don't help anything else. The promise of the longevity field is that you can actually modulate something that's way upstream of everything. It could prevent not only one disease but nearly every disease. What else is worth doing? As far as time immemorial, people have been having this idea of how to extend lifespan. It's the most effective revolution of healthcare that we can imagine. I think it's worth fighting for.

[01:10:26] Why work on aging?

Abhi: I mentioned in your introduction that you did a PhD at the USC Buck Institute of Aging, which is obviously an aging institute. Walk me through the journey leading up to getting the PhD. What made you realize you wanted to spend a significant fraction of your life working on this?

Alan: I used to work at Arsenal Bio. I think that's where I got my intellectual start. Arsenal Bio is a CAR-T and cell therapy company that has incredibly advanced technology in that particular space. My job was improving screening. The question that I would come into work every day thinking is, "Okay, here is a way that we can test 10 ways of improving a CAR-T to kill a cancer cell better. How do I make that a thousand ways in two months?" That was a very fun problem and it taught me a lot about the importance of the details in screening. One of the biggest values that Permanence has is that we think about screening and testing things for genome stability in ways that are, in many ways, better than other people. That came from my time at Arsenal Bio, where I was really enmeshed in that screening mentality.

The thing that led me to aging was, I looked at all of that and I saw they're doing incredible work and are incredible people, but could we have prevented that cancer in the first place? There's a personal overlap there in that the very same time that I was at Arsenal Bio, my father was experiencing cancer, and he passed away from cancer right before I started my PhD. Strangely enough, he had the signs of melanoma for a decade prior to it actually becoming malignant and then eventually killing him. I thought to myself, "Why do we wait so long? Why are we developing therapies that you can only introduce into patients once they're already six months away from dying? Why can't we go more upstream, and what is more upstream than aging?" I was like, "I've got to do aging."

When I thought about aging, I was like, "Okay, where do I go? Do I start a company immediately? Do I go into VC?" I realized that the aging space was very confusing and it was very hard for me to see what was the signal, what was the noise, what was real, and what wasn't. It felt worth it for me to spend years of my life on a PhD to figure that out. So I went to the Buck and I loved my time there.

Abhi: I haven't looked too much into the Buck. Do they have a pretty dogmatic thesis on what the causes of aging are and how to fix it, or is it pretty freeform and they don't take any particular stance?

Alan: No, not at all. That's the beautiful thing about the Buck, and that's the beautiful thing about the PhD program. Almost every single day, we would have conversations with my colleagues in the cafeteria or somewhere, just chatting about what aging is, what is the best way that you can interface with it. We had this program called "Think and Drink," where we literally just got wine, oftentimes with a relatively well-known person in the aging space, and just debated and lobbed questions at them, thinking about what aging is, how we can intervene here, and what's the best thing that we can do. We hosted the Buck Student Aging Symposium. We had our own in-house conference where we had speakers and presenters. So there is no particular thesis I think the Buck has. They're a little bit more healthspan-focused than lifespan-focused. They lean more on how can we live better, still longer, but on the living better aspect as opposed to the mortality crowd. But in terms of scientifically and intellectually, that was a place where all opinions were accepted.

Abhi: Do you feel that DNA protection mechanisms were covered well during your time at the Buck, or was that an area that they could have focused on more? How you've described this so far does feel like a relatively, even though the field has existed for a long time, it's taken a bit of a backseat to other fields. Does Buck give a very holistic overview of all the available approaches, or does it really focus on the things that feel the most promising in the last five years?

Alan: It's definitely much more all-encompassing. I'll say that there's no concrete genome stability lab at the Buck, for example, or a genome stability center. But genome stability pops up all the time. As an example, the research the Buck is really well-known for is this work in the senescence field. Of course, how does a cell become senescent? A whole variety of things, but genome stability is one of the main drivers of senescence. So you had a lot of people—postdocs, students, PIs—working on understanding genome stability in the context of senescence. Ditto, I was working with Dr. Eric Verdin, mostly focusing on immunology, but what drives a lot of immune dysfunction? What drives clonal hematopoiesis? That's mutations. So, wherever you look, whether you're primarily focused on genome stability or not, it always comes up.

[01:15:26] What are you bearish on?

Abhi: What is a subfield of longevity research or longevity therapeutic research that you don't personally think there's that much promise in, or perhaps is the furthest away from being in the clinic?

Alan: Those are two very different questions. The furthest away, on the most radical side of aging, there's this idea of doing whole-body replacement or even BCI, but from an extreme angle of you just chop off the body and have mind upload. That's super far away. I think a lot of people are excited about that field because they come into it from an engineering mindset. I think there's a very high representation of engineers who are really excited about replacement because they see it as an engineering problem. The reason why I think it's a long ways away is it's not an engineering problem, it's actually a biology problem that is masquerading as an engineering problem. If you talk to a doctor about how you would replace a spine, that is just a whole other beast. That is not an engineering challenge yet; it is very difficult medically and scientifically. Not to say that we won't get there and not to say that maybe that's the most promising path towards immortality for the individuals who are mortality-oriented, but it is a very long way away.

Abhi: What about the former question of something you view as not very promising?

Alan: I've been bearish on—there is very little evidence that thymic involution drives aging.

Abhi: I have no idea what that is. That's a brand new word to me.

Alan: Thymic involution. The first organ to age—this is also a debatable field; people will argue about what the first organ to age is—I will say the first organ to age is the thymus. You start to see the thymus falling apart even as early as the age of five. The thymus is really, really important. The thymus is where all of your naive T cells spawn from. The thinking is, if you've run out of naive T cells and you don't have the organ that creates naive T cells, then it's clearly going to negatively impact immune function. The downside is that there have been studies that have been done on patients who have had a thymus removed early, even as early as five years old, and it doesn't seem to really impact aging a whole lot. I think the reason why is that you make a lot of T cells and humans are pretty big, so as a result, you have enough T cells to last you past where you would naturally die.

I once chatted with someone, and he said, "Thymic involution would be a serious problem if we live to 150." There are a number of problems like that in aging where if we solved aging up to 150, then you suddenly get a whole new suite of problems. It seems like thymic degeneration is one of them. I'm less excited about ways to bring the thymus back.

Abhi: Purely because it's a problem—it's like an overpopulation on Mars problem. It's not something you expect to be an issue for a very long time.

Alan: That's what the literature says to me. Of course, the contrarian argument is that having more naive T cells is always better. So if you had more new naive T cells as opposed to old naive T cells, then that would be beneficial. But naive T cells also don't age a whole lot. So I've had a hard time with that.

[01:19:12] Weirder types of aging beyond 110

Abhi: This is interesting because I've long wondered, let's say we solve all six to seven hallmarks of aging, and we get to the 110 mark. Thymic degradation is the first thing I've heard of as where it really starts to be a problem. Is there something else that is like an end-game boss with regards to aging that only rears its head long past the natural lifespan?

Alan: In the immune space, that's the only one I can think of. I'm trying to see if there's anything else that also comes up, and nothing immediately comes to mind. I think the more time goes on, the more the extracellular matrix plays a role. You just have a lot of extracellular matrix proteins that have a certain half-life, and that half-life is measured in many decades.

Abhi: Do those turn over?

Alan: That's the thing. It's complicated. A lot of them do, but when they get turned over and get replaced, they get replaced poorly. Over time, you see problems with elastin, but the problem is that it's not even an end-game boss because you start seeing these problems even when you get to be 80, 90, 70, 60. We wrinkle, and a lot of that is ECM problems. I wouldn't say that's an end-game boss, but it becomes increasingly hard to deal with, I think, the older we get.

Abhi: Do you imagine that stuff like epigenetic reprogramming and DNA protection can only go so far and the final longevity treatment will be replacement? Is that how people in the longevity field generally think about it, or are there people who are full on "epigenetic reprogramming is all you need," "genome stabilizing is all you need," and if you have this in sufficient quantities, you will live to be 200, 300?

Alan: I think almost everyone thinks that it's going to be a combination of things. It also depends on the age. Again, if you're super old, if you're pushing your nineties, then probably replacement/reprogramming is the way to go. If you're in your twenties, then you're better off preventing aging and doing genome stability work. So there are a lot of ifs involved in there.

[01:21:37] How did you decide on DNA protection and what else would you have done?

Abhi: You've mentioned in the past that you decided on DNA damage prevention relatively early in your career. I think it was something you happened upon while you were at the Buck Institute. Was there some less-discussed approach to longevity that you were really wondering about and thinking that you should start a company in, or was DNA damage prevention the main thing you were curious about?

Alan: The way in my mind that I broke this problem down was as a healthspan or a lifespan problem. If I was trying to optimize for how do we make people who live better, longer, do you emphasize "live better" or "live longer"? Genome instability is harder, but I think it'll do both and it also will extend maximum lifespan. If we're thinking purely from the healthspan angle, I'm just very excited about targeting inflammation. BioAge has a LRP3 drug that I'm very, very excited by. I think that one of the primary reasons why people feel bad as they age is because of chronic inflammation. If there's one lever that I would pull to address that, it would be that inflammation side, especially innate inflammation.

Abhi: What would an alternative universe Alan have worked on that is perhaps not even in the longevity space at all? Is there such a thing?

Alan: There is. Conceptually, I love oncolytic therapies. This is coming back from my CAR-T world background. I love CAR-Ts. Again, Arsenal Bio was a great time, but you always have this antigen escape problem. If you train a T cell to recognize a single thing on a cancer cell and then kill it, then the cancer can just not make that thing, and then the CAR-T is not particularly effective. But there's the concept of oncolytic viruses. For everyone who's not familiar, you just take a virus or a bacteria and you put it inside of a tumor, and that leads to a number of things happening. One is that you kill the tumor itself using your virus, but the other thing is that your immune system sees the threat and it wakes up. It turns a cold tumor into a hot tumor because suddenly it sees not only is there a bacteria or a virus that's attacking the tumor cell, but suddenly there's also a tumor cell, and it only starts to recognize that once there are other immune signaling threats happening there. As a more generalized cancer therapy, I'm incredibly excited about that field, even though clinically it's been a little bit of a mess.

Abhi: I have heard a lot about this particular side of oncolytic therapies. Why do you think it has been such a mess?

Alan: It's just so hard to—you're toggling three things at the same time. If you have a virus that is too good, that is incredibly effective at killing the cancer cells but then doesn't get recognized by the immune system, then the tumor can kind of adapt to it and then the immune system never gets turned on and the virus slowly dies. You deal with your primary malignancy, but you don't do anything with the secondary malignancies because the immune system doesn't recognize the threat. If you have a virus that isn't strong enough and is too safe, then you don't have an effective therapy. If the virus is very effective and not safe, then you start seeing actual side effects in patients where you not only have cancer but you now have a very serious infection, which is also not good because cancer patients have notoriously not good immune systems. Dialing all of those to be the right level of safe, effective in terms of turning on the immune system, but also effective enough in terms of actually killing the tumor cells has proven to be very difficult. I think a lot of companies have worked on this and it's still a very promising field, but it's going to take some iterating, I think.

[01:25:27] What was it like raising money?

Abhi: I'd like to talk to you about your process of raising money for Permanence. Permanence sits in this place of not having a clear parallel to any existing therapeutic that's out there. At the same time, I also see a lot of VCs saying that they want new targets, but the people who are developing new targets often struggle with raising money. I'd like to hear about your fundraising journey and whether that was an issue for you.

Alan: I think there's a little bit of a difference between the tech world and the biotech world in this regard. In the biotech world, people have tweeted about this all the time and posted on LinkedIn of, "Look how many PD-1 drugs there are now, look how many TIGITs and CTLA-4s." Because biotech VCs, especially on the East Coast, really care about validation. They want to have something that they trust will work simply because every stage of the biotech/biopharma process is so expensive. You want to de-risk as much as you can, especially from the beginning. It's difficult to go to a biotech VC and fundraise on something very fundamentally new because it's just technical risk on top of personal risk.

For tech VCs, the way that tech has been so successful, you hear about the power law a lot. The idea is 95% of things will fail, 5% of things will be incredibly successful and become unicorns. You really only need to bet on—if you have a VC fund and you bet on 99 things that go to zero, but you bet on one thing that is a 1,000 to 10,000x, then you've returned your fund 10 times over, so you're very happy. The appetite towards risk is extremely different in the tech world. Biotech VCs have been incredibly helpful and were very useful in terms of, "Hey, look, here are the things that we would want to see when we invest eventually." But early on, it was the tech world that was the most helpful to me because what they saw was, "Hey, there's a guy who's trying to prevent cancer. There's a 95% chance it's not going to work because it's a very difficult journey, and that's entirely valid, but there's a 5% chance that it succeeds and you prevent all of cancer." That is going to be massive. It's worth investing in. So I think leaning into that was very helpful for me and was how I managed to fundraise.

Abhi: The way that I've heard one tech VC actually describe the biotech VC world is that the managing partner at these firms looks at what the top 20 pharma chiefs of science are interested in and just does aggressive M&A to try and find the people to invest in who will appeal to one of those 20. Do you imagine that whatever Permanence produces will eventually be acquired by a big pharma, or will you take this drug to the finish line?

Alan: That's a good question. We'll see. The way that I've always presented it, especially to investors, is that there are different paths that Permanence Bio could take. Our dream is cancer prevention. Our dream is longevity too, to prevent disease as early as possible. At this stage, that is a hard sell and I am contrarian enough to try my best to do it. It seems like we have some early success and hopefully, that success continues. So long as we can see success moving into cancer prevention, we want to be the number one cancer prevention company.

There is a world out there in which genome-stabilizing drugs turn out to be very effective for a whole variety of different diseases. I mentioned dry AMD and maybe ALS. But the cancer prevention thing is just too much of a lift. In that case, we'll have a very effective drug against, insert acute indication here. Maybe we'll get acquired, and that can be more of the strategy. But I think so long as cancer prevention is on the table, we want to do it.

Abhi: When you were raising for Permanence, how difficult was it to tell the story at first?

Alan: Very. There was a huge shift between how the narrative played out. The challenging aspect was that I came from a PhD for three years and then did VC for a very brief period of time and then dove right into building a company. That PhD mentality, even the kind of conversation we're having now, it's hard to unwind. You have to unwind it when you're trying to raise for a company, especially among tech investors. I remember my first pitches, the first narratives I had were like, "Oh, you know, aging and genome instability are really important in aging, and if you can prevent genome instability, then you can prevent other forms of hallmarks of aging that then leads to other diseases." Boy, everyone was bored immediately.

So what I pivoted into instead was focusing on the cancer prevention aspect because people really thought that was interesting and different and unique, and it had a huge potential upside. The narrative became much more into, "Cancer is a function of mutations and inflammation. No one's targeting that mutations aspect. We can. We know how." So let's just do that. Let's build drugs that prevent mutations and then we prevent all of cancer. That was a really compelling story.

Abhi: Was it a matter of removing the science and not explaining the exact specifics of it? How important was that? How much did tech investors practically care that you're doing one modality versus another modality?

Alan: It's both. I think it's both making the vision grander. It seems weird, right? Because cancer prevention is theoretically a less grand vision than longevity, but longevity means a lot of things to a lot of people.

Abhi: It's kind of a squishy definition. It's like you can put whatever you want inside. Cancer is a little bit more definitive.

Alan: Cancer prevention specifically is a much more defined big goal. Then, taking the science down a notch from genome instability—I mean, even the first 10 minutes of our conversation today was like, "What do you mean by genome instability?" It's this whole thing. But instead, just "mutations." Focusing on mutations because people know what that is and everyone can agree on a definition of what a mutation is. I think that grounded it for a lot of folks.

[01:31:48] What do you think of past cancer prevention trials?

Abhi: Why do you think the existing cancer prevention trials, or the past ones, have failed?

Alan: It really depends on which ones. There's a very long cancer prevention history, and to be honest, I'm only familiar with some of it. For example, people very early on realized that aspirin is very effective at being a chemopreventive agent. Big pharma also realized that aspirin seems to prevent cancer by pretty significant margins, specifically colorectal cancer. Aspirin has a few less-than-ideal side effects, especially at high doses. What a lot of big pharma companies did was they were like, "Okay, let's take aspirin and then improve its targeting, I want to say its COX-1 or COX-2 targeting, but one of the two specifically, and then improve that angle to make it a safe thing and then really target that one mechanism, and then that will be the best ever cancer prevention drug that will be like aspirin." The sad thing is, and this is still, I think, very controversial, there isn't a clear answer, but from reading the literature deeply, it seems like aspirin prevents cancer through a weird off-target mechanism. So pharma, having spent hundreds of millions, possibly even billions of dollars into cancer prevention, improving these drugs for a mechanism that was incorrect, the appetite just died after that because there was a lot of investment and a lot of hype and a bunch of failure, but they were just going after the wrong thing. So I think that's part of the reason.

Abhi: When was that? Early 2000s? Do you think we're coming back into the cycle of people being really interested in cancer prevention, or is it still a little bit outside of the Overton window?

Alan: It's coming back. The first-ever, I want to say, the first-ever cancer prevention conference was hosted a year ago. The second one is actually happening, I think, possibly right now in the UK. More and more people are thinking about it. The big boon here is that we're getting the technologies to measure genome stability coming online. Of course, there's Alex Cagan's big paper looking at aging and somatic mutations with aging and showing there's a crazy strong correlation between how quickly a species gets mutations and its lifespan. That technology just didn't exist even five or 10 years ago. The fact that the tools are coming online to develop these drugs, to measure these things, the biomarkers in the blood exist now, I think it's bringing a lot of fresh air into cancer prevention.

[01:34:12] What does good wet-lab talent look like?

Abhi: Permanence Bio, I think, just hired its first person, at least according to the LinkedIn post. You posted a job posting, and last I checked, it seemed to be offline. I assumed you hired someone. If so, congratulations. I'm curious, what were you looking for in a founding scientist, the first few employees?

Alan: Details. The main thing that is challenging about cancer prevention and genome stability, I would say, is how do you set up these assays? How do you measure mutations? That is, if you've looked at a whole-genome sequencing pipeline specifically looking at DNA mutations, it is complicated. Every single little enzyme that you use, the steps that you use, the read depth you analyze it at, all of that matters so, so much. We brought on our first hire, Carlos, who I knew from my PhD. He's the smartest guy that I knew in the PhD program, and I felt very lucky that he offered to join. He's got such an eye for these particular details. He knows exactly, "This experiment failed. I'm going to test these four things and that's going to make it work." Detail orientation, I think, helped a lot.

Abhi: What do you think leads to that? Is it just extremely high conscientiousness, a lot of past experience, very high fluid intelligence, or something else entirely?

Alan: Obsession.

Abhi: Obsession with getting things correct, obsession with the field as a whole, or what is it exactly?

Alan: It's more of that engineering obsession of how do you solve a puzzle, how do you troubleshoot this thing and make it work.

Abhi: Do you think there is such a thing as 10x wet lab biologists?

Alan: Yes. Easily. Yes. Absolutely.

Abhi: And you think it comes down to this detail-oriented approach?

Alan: I think it comes down to probably two things. One is the detail orientation. I think that helps a lot. There are a lot of scientists who can spend a lot of time and effort on something that, because of the details, fails. But it's also about asking the right questions. I think probably asking the right questions is the more important part of that. It's so easy to spend, especially because biology just takes so long, months, if not years, beating your head against the wall for something that, in the end game, doesn't really matter. The people who I think have been incredible scientists are the ones who see a gap that they can solve and they figure out how to solve it.

[01:37:02] What does your information diet look like?

Abhi: What does your information diet look like? If I scroll through your Twitter page, you often post links to papers that are completely unrelated to the field that you're working in, but you find some way to connect it and you say, "Hey, this is super relevant for DNA protection mechanisms." Do you spend relatively little time reading papers in your own immediate field and instead search for "alpha" in other fields, or is it a mix?

Alan: It's a mix. I think this is also more of a deeply personal thing and not my philosophy towards life in general, which is to surround myself with things that I like. To a large extent, we are a product of our environment. To what extent can you surround what you see and what you experience with the kind of person that you want to be? My Twitter is actually a really good example of that. I follow a relatively small number of people, and for everyone that I follow, all of them talk science. That is their main thing. There are a number of people that I would feel awkward not following because they're like a close friend or something, and they've posted other stuff. In which case, every single person that I have there that doesn't post on science, I mute. So my Twitter feed, and I never go into the "For You" tab, I only go into the "Following" tab. When I scroll Twitter, I only see scientific papers, scientific discussions, maybe some founders, and now kind of more oriented in that direction, biopharma discussions as well. Just surrounding myself with that knowledge.

With regards to papers, I have Google Scholar alerts for cancer prevention, lifespan, DNA repair, kind of everything that I'm interested in there. I also try to surround myself with people who think about these things a lot. I'm part of a genome stability journal club group chat on WhatsApp where it's just like 30 people who are all obsessed with genome instability. We meet once a week to discuss some paper. We literally had that today. I spent an hour debating whether or not mitochondrial DNA mutations are a causal driver of aging and what the evidence for that is right now. A lot of my other communities are also oriented in a similar direction. So it was just a non-stop input stream of relevant information.

Abhi: I know at least one person in the pure machine learning world who says that a lot of the best papers of today are worth reading, but the best papers of 30 to 40 years ago are also worth reading, and no one is reading those. Is there something akin to that also in your space where you read papers from the eighties and nineties, or are they not super relevant?

Alan: They're very relevant, but I think much more from a—I've found a lot of joy in reading old theory of biology or theories of aging papers. I've been trying to dive into those less now that I'm very much heads-down operational, but still, thinking about how people saw things big picture is pretty important, especially as we go into, I start to sound very old now, but as we go into a world where everything is sped up, going back to a time when people thought very big picture, I think, is helpful.

[01:40:06] What's it like going from research to being a CEO?

Abhi: You went from three years of doing research in your PhD and I imagine also your time at Arsenal, and I imagine you did some research during your undergrad. I imagine your day-to-day is not too much bench work research and is more being a CEO. What's that transition been like?

Alan: It's both. It's the best of both worlds. My day-to-day is I come into work at 9:00 AM and then I take my meetings from 9:00 AM until 1:00 PM, and then I put on a lab coat from 1:30 until 10:00 PM and then I do science. So it's both, actually. It's a beautiful mix of both. I think it's both things that I feel very comfortable in. I love operating and I love science, and at least for this period in Permanence Bio's history, which I think will probably not last a whole lot longer, the best use of my time, the most alpha I can bring to the organization, is in my scientific knowledge. Right now, I'm still harnessing that. I imagine two years from now that's going to be much less the case and it's going to be mostly CEO hat.

Abhi: You guys are roughly eight months old, I think. Is that correct? Incorporated in October, so eight months. At some point, I imagine you're going to have to make a decision of whether you want to be the full-time operator versus the full-time scientist or the full-time CSO of the company. Do you know if you're leaning one way or another?

Alan: I think much more the CEO hat than a CSO hat. I think I'm very good at vision, like seeing and connecting the dots between, "We have a vial of a hundred micromolar of some compound that we want to test in some assay," and seeing how the dots connect towards how can we prevent cancer in the general human population. I think that is kind of, exec coaches like to say, the "zone of genius." The zone of genius that I think I have is I can connect those dots, and I think that's very hard for a lot of other folks. I'm a pretty good scientist, but I wouldn't even—I think there are a lot of very, very, very good scientists out there who are much better than I am that I would love to hire and bring on. I don't think I'm as needed there.

[01:42:20] What happens after cancer prevention for Permanence?

Abhi: Big picture view, and I think this is the last question that I'm going to ask. Over the next 10 years, let's say whatever Permanence is working on works really well for a specific type of cancer, do you imagine you'll keep poking away at improving, like moving whatever therapies you have into other subtypes of cancer, or would you switch to a different indication entirely?

Alan: We'll follow the data. If our assays show that it's really powerful in some acute disease model, then we've got to do that. But if it doesn't, then yeah, just keep moving into cancer.

Abhi: Cool. Thank you for coming onto the show, Alan.

Alan: Thank you so much. This was a lot of fun. Hours of yapping on my favorite topic, so thank you so much.

1. Introduction

2. Timestamps

3. Transcript

Watch on Youtube, Apple Podcasts, or Spotify!

Introduction

To those in the protein design space, Dr. Sergey Ovchinnikov is a very, very well-recognized name.

A recent MIT professor (circa early 2024), he has played a part in a staggering number of recent innovations in the field: ColabFold, RFDiffusion, Bindcraft, automated design of soluble proxies of membrane proteins, elucidating what protein language models are learning, conformational sampling via Alphafold2, and many more. And even beyond the research that have come from his lab in the last few years, the co-evolution work he did during his PhD/fellowship also laid some of the groundwork for the original Alphafold paper, being cited twice in it.

As a result, Sergey’s work has gained a reputation for being something that is worth reading. But nobody has ever interviewed him before! Which was shocking for someone who was so pivotally important for the field.

So, obviously, I wanted to be the first one to do it. After an initial call, I took a train down to Boston, booked a studio, and chatted with him for a few hours, asking every question I could think of. We talk about his own journey into biology research, some issues he has with Alphafold3, what Alphafold4-and-beyond models may look like, what research he’d want to spend a hundred million dollars on, and lots more. Take a look at the timestamps to get an overview!

Final note: I’m extremely grateful to Asimov Press for helping fund the travel + studio time required for this episode! They are a non-profit publisher dedicated to thoughtful writing on biology and metascience, such as articles over synthetic blood and interviews with plant geneticists. I myself have published within them twice! I highly recommend checking out their essays at press.asimov, or reaching out to editors@asimov.com if you’re interested in contributing.

Timestamps

[00:00:00] Highlight clips

[00:01:10] Introduction + Sergey's background and how he got into the field

[00:18:14] Is conservation all you need?

[00:23:26] Ambiguous vs non-ambiguous regions in proteins

[00:24:59] What will AlphaFold 4/5/6 look like?

[00:36:19] Diffusion vs. inversion for protein design

[00:44:52] A problem with Alphafold3

[00:53:41] MSA vs. single sequence models

[01:06:52] How Sergey picks research problems

[01:21:06] What are DNA models like Evo learning?

[01:29:11] The problem with train/test splits in biology

[01:49:07] What Sergey would do with $100 million

Transcript

[00:00:00] Highlight clips

My big goal in life has always been to come up with a unified model of protein evolution that accounts for all these different effects. And so what may appear to be creativity is just trying to tackle every part of the problem…

But I think one thing that maybe computer scientists don't quite realize yet is that all of biology is related. Every biological data point, there's no IID, every sample is related to another sample out there. And so if you do like a random train-test split, you might actually have overlaps…

And in some ways that you can think of that, that's what essentially Alphafold is doing. Like Alphafold will say, I'm going to make a guess, that's like zero recycles, and then you iterate and you sort of move around. But maybe if you do many, many independent seeds. And I think that's actually what some of these models like o1 and o3 are doing, like they have many, many independent starting points and they explore. And so I think in some ways, I guess we could say we've been already doing that for a while in the protein world. And they're kind of catching up…

I've never sort of sat down and said, okay, this direction is probably the most meaningful thing to do. It's more just like, okay, this is like a puzzle and there's no solution here. I'm just trying to figure out what's going on here.

[00:01:10] Introduction + Sergey's background and how he got into the field

Abhi: Today I'm going to be talking to Dr. Sergey Ovchinnikov, a recent biology professor at MIT. Sergey is an easily recognizable name to those in my own field, as he is one of the undisputed greats in the world of machine learning assisted protein engineering. His prior research includes ways to make protein folding more accessible, models that can generate de novo protein binders with massive success rates, and methods to help scientists learn what protein language models are actually learning about protein folding.

Today we'll be talking about issues with existing protein models, what future protein models may look like, Sergey's own journey in this field, and lots more. Thank you for coming to the show, Sergey.

Sergey: Yeah, thank you so much for inviting me to be here. Excited to talk to you about those topics.

Abhi: And so just to start this off briefly, I'd love for you to give us an overview of what your historical research focus has been and what types of problems you're most curious about.

Sergey: Well, I would say my journey started in trying to understand the relationship between species, how different organisms are related to each other, but more specifically trying to compare their DNA sequences, their protein sequences to each other. One of the things that we encountered early on is that sometimes you might think two things are highly related to each other based on the similarity of their protein sequences, but in fact, it's because of convergent evolution. There might be similar selection pressures in two different organisms that make their protein sequence look very similar, when in fact it's essentially just convergent; there's similar selection going on. To be able to separate that, you really need to understand the underlying protein structure, and also understand the protein function for these protein sequences.

My journey started off actually in phylogenetics, and then I transitioned during my PhD to try to say, okay, I need to learn everything about protein structure in order to sort of go back and maybe correct the signal so we could do more proper phylogenetic analysis. In some ways, I'm still trying to do that. That's one of the areas we're actually pursuing. We're thinking about how do we do better phylogenetics? But along the way, we've done a few side projects, you could say, like try to, hey, maybe we get into protein design, but that's all sort of related to building models and evaluating these models for this general goal of understanding protein evolution and evolution in general.

Abhi: It's interesting that your original background concerned phylogenetics, and you're still really curious about phylogenetics, even though you're probably most known as being the protein design guy. Do you plan to do that many pure phylogenetics projects in the future? Or is that kind of on the back burner for now?

Sergey: We do have a few projects actually going on in the group where we're trying to actually do better phylogenetic reconstruction. But even for the reverse problem... so one thing I've always thought of it as sort of: structures getting in the way of phylogeny. But now we're also beginning to believe, for example, closely related species could actually influence extracting coherent coevolution signals from multiple sequence alignments. And so it's also becoming an issue where maybe for bacteria, it turns out it maybe is not a big issue because most bacteria are highly diverged and you can almost think of it as a star tree; they're all equally separate from each other. But when you start to deal with eukaryotic organisms, like even fungi, you may have random mutations that are propagated and might mislead some of the calculations in terms of structure. So in some ways, not just purely trying to understand phylogeny, but it turns out phylogeny might actually become an important thing to think about when doing even studies more on the structural side or even protein design side.

Abhi: And for people who aren't super aware of how Sergey's work translates to the current state of the art in the field, there's a pretty clear direct line between the co-evolutionary work you did in your PhD and a relationship between that and how a model like Alphafold2 actually works. In many ways, you were dramatically ahead of the curve. I'm curious, during the mid-2010s when you wrote those papers, was it clear to even you that this work would be particularly useful for the problem of protein folding, or did it feel very much like a pure phylogenetics problem that had no relationship to any translatable research?

Sergey: Well, I think in the early 2010s or so, other folks besides myself had been thinking about: can we somehow extract covariance signal and could that be used to predict structure? So I think that's always been on the radar of people: this covariance signal will be useful for structure prediction.

At the time, I was mostly thinking from the perspective of using that signal to subtract it out and do better phylogeny. That was my initial goal. But one thing we found out is that actually the signal is useful, especially if you start to look at metagenomic sequences. What I mean by that, one of the interesting things people have found was that when you start to ask the question, okay, where is coevolution useful in terms of being able to predict contacts, to be able to predict structure? Often those structures were already solved; somebody's already determined that structure because if there's a lot of proteins for that protein family, very likely that somebody's already predicted the structure of one of those sequences or actually determined the structure experimentally. The only things that were left unsolved were membrane proteins, just because those were really, really hard to crystallize. But with metagenomics, what happened was suddenly protein families that had only few sequences suddenly had huge amounts of sequences. And so now these coevolution methods suddenly became more and more relevant for those kinds of protein families.

Abhi: I remember when we first talked a few weeks ago, you said coevolution for the problem of protein structure prediction was one of a few other parallel directions that were going on. I think you named two others. I'd love to hear you recapitulate that, because I think that was a really interesting story.

Sergey: I think maybe I brought up the fact that... well, maybe I talked about different groups working on this. Is that what you're referring to? Maybe that's what I was... okay. Yeah. So essentially, the idea of extracting coevolution from multiple sequence alignments has been around for a while. Actually turns out this continued to be pursued in three different fields, essentially: people in the physics world work with things like Potts models and Ising models, how we transfer this to this problem. People in the computer science world were also thinking like Markov random fields, Boltzmann machines, how do we use that? And then people in the, I guess you could say more computational biology field, were also thinking along those lines, like mutual information, and applying those kind of approaches.

I think the part that I found remarkable or interesting looking back is that often these people didn't cite each other because they didn't know about each other. It turns out, if you actually look at the math, the math is almost identical, but they just never sort of talked to each other because they used completely different terminology for the same concepts.

Abhi: But the underlying data that you used was all identical? The data was used...

Sergey: Even the algorithms were the same. Okay. It's just that I think in computer science, people call it Markov random fields. And then in physics, people call it Potts models. And then other places were just calling like coevolution models. But turned out a lot of the math was actually identical. They just used different symbols. Like one field will use W to represent coevolution, another would use J. And it's like, okay. But if you look at it, it's the same equation. They're just using different symbols, different words for the same thing. But I think on my side, I worked more I guess with the folks coming from the computer science side. So I worked with Ti-ti from CMU in Pittsburgh. And so I came from it from the Markov random field perspective. And I think in our papers we always call it Markov random field, but then I realized sometimes people who do Potts models, they get a little confused because they're like, Hey, is this something different? But it's like, no, it's the same thing. So now I have to say both terms when I refer to it.

Abhi: Do you think the field has generally consolidated into the pure computer science direction or are there still computational biologists and physics people who are pursuing their own parallel paths?

Sergey: I think they've all sort of... I think past 2011 or so, when things suddenly started to work, I think all these groups sort of became aware of each other. Some folks actually started collaborating together. And so there's been... so now I think they're all kind of aware of each other, but I'd say pre-2011, it was kind of multiple parallel efforts of maybe people not recognizing that they're working on the same thing.

Abhi: And speaking on your own computer science background, there's relatively little publication history on your own backstory and the lore of Sergey. I'd love to hear about why you decided to study biology, what made you focus on phylogeny, and what led to the eventual pivot closer to computer science applied to the whole subject.

Sergey: Let's see. Where, how far should I go back here? We could maybe even start in college?

Abhi: Maybe even high school? If you think that's a good place to start.

Sergey: So I started off back actually when I was still in high school. I was on robotics teams. So I used to program robots. There's this thing called US FIRST. And it's essentially a team of, I think it's international now, but at the time they were doing... high school students would build robots and they would compete with each other. And I was the one usually involved in programming these robots. But the reason why this sort of led to biology is to me, when I started to think about biology from the perspective of code, it kind of all clicked for me. And so what I mean by that is: you could think of all organisms, they all have some kind of code, or you could think of all organisms maybe as robots to some extent. There's some code that essentially codes for how they act, how they're developed and so on. But to me it kind of felt like we didn't really have a good understanding of the compiler or the compiled code. We have all this code, we don't know the syntax. We knew roughly where proteins start and end, but no idea at the time. And so to me, really, that's why I got really excited about biology because I was like, Hey, this is like unknown code. We have no idea what it's doing. What if we start to compare these codes to each other?

And so like right now when we think about GitHub, we get to probably look at some project and ask everybody who's cloned this project and who's modified it. You could check to see which parts were modified and which parts were not modified. And that would quickly tell you, okay, these parts that were probably not modified are important for this overall project. And parts that were heavily modified, those parts are probably not as important. Or maybe they were modified and they've gained new function, I guess you could say. And so in some ways, when you're comparing different source codes to each other, to me it felt like, oh, you could compare a bunch of genomes to each other and figure out what...

Abhi: And this naturally leads to phylogenetics.

Sergey: Exactly, exactly. So you start to compare all these genomes to each other, I guess you could say, reconstruct the GitHub history of all genomes. But then that lets you start to understand the syntax and so on. And so that's sort of what got me initially. So then when I went to college, I was like, you know what, I'm going to learn about biology. At one point I was... I think I moved a little bit into history because I was like, maybe I'll do history of science because I wasn't sure if I was good enough to do science yet at that time. But then eventually I transferred and started doing more biology. But over time, I joined a couple of labs, like one lab that worked on milkweeds, another lab that worked on various arachnids. And I was actually participating in extracting the DNA, sequencing those guys, and then getting all the sequences. But once we started getting sequences back, I think my advisor at the time quickly realized that I had some computational skills. So like, okay, maybe you could help assemble some of these sequences. So I worked on genome assembly and building some algorithms to be able to do that at the time. And that sort of, I guess you say, brought me back into computation because initially I was like, I'm going to do biology, and then I'm now back to using these algorithms that I worked on in the past.

Abhi: And from there, you also did your PhD work in phylogenetics.

Sergey: Oh, yeah. Maybe I should clarify. So for my undergrad education, I joined a couple of different labs that were working on phylogenetics. But then for my PhD, that's where, during my undergrad, I started to realize we can't really do correct phylogeny without understanding structure, without understanding covariance patterns. Because one of the things that's interesting in phylogenetics is when you build, when you compare a bunch of sequences to each other, there are certain sites that have high entropy and mislead phylogeny. And what I mean by high entropy, there are certain positions that just change rapidly. There are certain organisms that also evolve rapidly. And those organisms that evolve rapidly, they would appear just by chance to be highly related to each other. There's this process called long branch attraction in phylogeny where highly evolving species suddenly start to get grouped together.

And so what to get around this problem, what folks sometimes do is they always say, okay, let's remove positions that are not consistent with each other. Because if there are multiple positions that are kind of consistent, then you would say, okay, this is probably due to phylogeny. Positions that are inconsistent, those are probably just random, and so we should just remove those. But turns out these metrics of looking for these self-consistencies between positions is actually very, very similar to coevolution. These are sites that are covarying with each other, but it's not clear is this covariance due to phylogeny or is this covariance due to coevolution? And the only way to tell that is to say, can we look at the protein structure? Like if those two positions are consistent and they're close together on a structure, that's a strong indicator that maybe there's some coevolution going on and that could mislead your phylogeny signal. Yeah. And so it sort of turns out these signals are completely entangled. And so that's when it was like, okay, I'm going to get my PhD, I'm going to learn everything about structure so I could disentangle this effect. And that was sort of, I guess you say, my journey into that space.

Abhi: And then you graduate from your PhD and during your postdoc you continue this exact same line of work.

Sergey: Yep. Yep. I guess now that I sort of, I felt like as an undergrad, I understood phylogeny. Then as a grad student, I understood structure and I was like, okay, now I'm going to combine the two and finally try to resolve this problem. And so then I went on to become a fellow at Harvard University. And there I actually started to say, okay, how do I actually combine these things? How do I build a unified model that understands conservation, coevolution, phylogeny? And that was my work during, I guess you could say during my fellowship. But during that time, large models started coming out. Like, for example, folks started training giant, giant language models for proteins. Things like initial versions of Alphafold started coming out. And there I was thinking, okay, is it possible these models have already learned to do that? Like, did we kind of get scooped without realizing? Like, are these models learning phylogeny? Are these models learning coevolution? And so in some ways, my work kind of partially pivoted towards: let's fully understand what these models are actually learning. We really need to dig into them because that would tell us, one, did we get scooped? And two, do we still need to work on this problem?

Abhi: I'm curious. I think you're often... whenever people think of Sergey Ovchinnikov, they often think of deeply creative papers. I think papers that you wouldn't really expect to come from any other people. Do you think there's an aspect to... do you think a lot of your quote unquote, alpha as a researcher comes from your background in Phylogenetics and that people who work at Isomorphic and EvoScale could stand to learn a little bit more about phylogeny?

Sergey: I don't know if that's where it's coming from, but I guess for me, I guess my big goal in life has always been to come up with a unified model of protein evolution that accounts for all these different effects. And so what may appear to be creativity is just trying to tackle every part of the problem. Like for example, we're trying to extract evolution signal, but then we also need to think about alignments of sequences, right? So for example, maybe we're extracting the wrong coevolution signal because sequences are misaligned. And so we venture into the alignment problem. But then once you start thinking about alignments, then you're like, well, how do you know you got the right alignment? Yeah. Well there, that's where it's like, well, maybe a structure prediction model could tell you that the alignment's correct or not. Right. And so... I guess what you could say, what may look like creativity, it's all just trying to solve this unified model problem, I guess. That would be one way to put it.

Abhi: How often do you return back to your background of phylogeny when you're looking at these problems? Are you often thinking from a phylogenetic lens or you often thinking from a pure machine learning researcher lens?

Sergey: I guess I'm always thinking from the, I guess you could say phylogeny or protein evolution lens would be one way to put it. It's... so I'm always coming back to this thinking like, how do we solve problems in this bit, like constructing this unified model that I keep talking about?

[00:18:14] Is conservation all you need?

Abhi: And the, like one of the axioms you have in your head when you're working on these problems is that you can learn almost everything you need from conservation. Is that a fair way to put it? I do notice you have this side interest in molecular dynamics. I sometimes see you post papers in that realm on your Twitter. But you don't actually ever seem to publish in that area. Do you think molecular dynamics will actually become really important in the future when Alphafold 4 or 5, 6 comes out? Or do you think for the moment conservation is the most important thing?

Sergey: Let's see. I guess when I say conservation, what I mean is like there's certain, I guess you could say there are certain very important positions that are highly conserved for purposes of function. So maybe to step back a little bit, we have, I guess you say there's lots of sequences out there and every sequence has some amount of selection going on. And some sequences are maybe in one particular organism, and that organism needs to do some function. And maybe a group of organisms, they all have the same function and so they have a highly conserved position. But it's not because there's coevolution there, it's just that that position is super, super important for that group of organisms.

But it could get misleading from the perspective of phylogeny, meaning like maybe you have some random mutation that happened early on in the tree of life and now because of this doubling effect, like you have one speciation event and now you have another speciation event, and every single time, I guess you say half the organisms now have that random mutation. And that could be a little bit misleading from the perspective of saying, is this really conserved? Or is this just a signal that gets propagated and propagated? And so sort of decoupling those effects.

But maybe coming back to your question of molecular dynamics, it's one of those things where we do... when we're interpreting models, we're thinking about how are these models working? Like, for example, language models. If we go step back to language models for a second, are they learning conservation, like learning each group of sequences have different levels of conservation just because they belong in a certain space of sequences? Are they working because we have coevolution going on and different groups of sequences have different coevolution? Or are they somehow internally in the model like solving the protein folding problem, or maybe doing molecular dynamics? Like is each layer sort of learning at different steps of folding or physics?

And when these first models started coming out, it was a little unclear which of those would be true. Like, is it, is each layer sort of folding up the protein, or is it sort of picking up on all these statistics of evolution, I guess you could say? And so that's one of the problems we're trying to separate. But that being said, of course, if we want the model to learn physics, then maybe we need to get into molecular dynamics in terms of trying to get these models to reason over molecules, reason over interactions of atoms. Yeah. I'm not sure if that's what you're asking or getting to, but that's...

Abhi: I guess the broader strokes of the question is do you think we'll ever escape the well of having... because you published this paper last year about how this pattern of a model always learning co-evolutionary statistics even continues if you don't have multiple sequence alignment. Even if you have a pure language model that has only seen sequence and no conservation, it still is learning folding via evolutionary statistics. And we obviously care more about the scope of all possible proteins rather than just the proteins that are near evolutionary proteins. Is this a problem in your mind that we need to find some way to move beyond the well of evolution? Or do you think for now it's actually completely fine?

Sergey: I think it depends on what you're trying to do and what your claims are. And what I mean by that is, I mean, I guess from the pure engineering perspective, there's probably nothing wrong with saying, Hey, let me grab a piece of evolution here, piece of evolution here, and sort of stick it together and make some kind of Frankenstein protein. And I think protein language models actually seem to be really good at that. They've learned different parts of proteins, which motifs tend to be the same across many different proteins. And you can imagine using such a model to sort of stitch different things together.

But then of course for somebody who's maybe coming more from a first principles point of view, they're like, Hey, this is kind of cheating. We want to be able to understand why these sequences code for these things and why are we able to stitch these things in a certain way. And maybe that would allow us to move into space that nature has never explored.

But then again, there are people that argue also that maybe any protein seems to be a combination of fragments. And so maybe nature's already explored all possible fragment space, in which case, having a model learn fragments is not a terrible thing. You just need to be able to sample these recombinations of things.

Abhi: Which do you think is true?

[00:23:26] Ambiguous vs non-ambiguous regions in proteins

Sergey: I think there's definitely space still left to explore. The way I like to think about it is that any given protein sequence is sort of composed of ambiguous and non-ambiguous regions. And so what I mean by that is: there are certain regions where there's a sequence that always codes for a helix. There's a sequence that always codes for a particular turn. And there are parts of the sequence that are these ambiguous motifs that you have actually no idea what they're coding for. It could be a helix, it could be a beta strand, it could be a loop, it could be a break there or so on.

And the only way you'd know is in the context of the full protein. The tertiary structure folds up and you're like, okay. It's almost like this region is some kind of chameleon sequence that could sort of adapt to different things depending on its context. And I think any natural protein is sort of a combination of these two things. There are certain parts that you say, I want this part to be rigid, and so you probably want to have this non-ambiguous sequence. And then there are some regions that are, you say, you know what, maybe this part needs to be flexible. And so maybe I'll put in some ambiguity here that maybe could respond upon a ligand coming close.

And what that means is that now of course if we try to predict these proteins, things that are made up of purely non-ambiguous sequences, Alphafold, ESMFold, like all these fold methods can predict them really well because I think they've already learned all these sort of non-ambiguous sequences, motifs that all these proteins share. But then for regions that are more ambiguous, that maybe have to do with function, these models, unless there's evolutionary information associated with them, are unable to predict it.

[00:24:59] What will AlphaFold 4/5/6 look like?

Abhi: Do you think the future looks something like you have one model to predict the non-ambiguous parts and then like a more physics-based model to predict the ambiguous parts? Or yeah, like, I'd love to get... I guess the broader question here is: what do you think Alphafold 4 or 5, 6 looks like? Do you think it goes something in that direction or somewhere else?

Sergey: Probably. I would say that if you construct a protein that's purely from these, I would argue non-ambiguous sequences, and I would say all of de novo design people who do de novo design often completely just exploit these non-ambiguous sequences. Like every single stretch, you can actually even do this experiment: if you take any little cut and try throwing it into Alphafold, it'll predict that without the context of the rest of the protein. And I would say it's a very simple search problem. You're essentially just stitching fragments together, and that's why it's really, really easy to predict these proteins.

But of course now once you start moving into this more ambiguous space, it becomes a really large search problem. And there you sort of do need to maybe start to say, okay, maybe we need to put physics in here. Maybe one way to put it would be like, there's this global search problem where you do huge fold exploration, and then there's more like a local search. Like once you roughly know what the fold is, you sort of fix all the little details. And I think Alphafold has learned to take a few steps along some kind of energy function that it learned. But it has a really hard time doing a kind of global search.

Abhi: Like I remember you said there was this figure in one of your papers where the Alphafold uses the MSA to find itself roughly where it is on the energy landscape. Yep. And then does some local energy minimization from there on out.

Sergey: Exactly. So I guess our current hypothesis is that multiple sequence alignments sort of give you the global... or give you sort of, I guess you could say you can skip the global search and you're just focusing on the local search. And sometimes you can mess with the multiple sequence alignment. Like people have found, if you subsample the MSA, sometimes you turn on dropouts, you enable sort of random masking, you could sometimes explore other parts of this starting space. And in some cases, like what we try to do is say, well, what if we just give it like a template structure? So this was work with James Rooney, and we're like, oh, we could give it some starting point and try to almost move around this space and see how well it is actually able to know which part's correct or which part's incorrect.

But coming back to your question of like what's the next version of Alphafold, I do think it's probably going to be some version where there is some global search exploration that gets baked into the model. And so what I mean by that is: you can imagine in every instance of Alphafold is sort of a few moves of a game, but it's not the full game that we're playing. And so, but there are probably many, many starting points. And so having some way to sort of guide or move into this global space is going to become quite important as we move into space where we say, how about proteins where we don't have evolutionary information and so on. Or if you want to design these proteins that look completely different.

Abhi: By many different steps, could I mentally equate that to inference time compute? Or are you referring to something else?

Sergey: I guess two things here. Because, so I guess want to step back a little bit. There have been researchers that have shown that you can run Alphafold with thousands and thousands of seeds and you could sort of think of each seed as sort of being seeding some independent MCMC trajectory and some search. And so you have a bunch of random starting points. And for example, even Alphafold 3 for antibody-antigen where there's no evolutionary information, they found they have to go up to like a thousand, and even after a thousand you can continue adding more and more seeds. And so those are kind of things where those are kind of random starting points. And so if you're lucky, one of those things will get to the right answer.

But one could imagine doing something, say, well, what if we have a smarter way of seeding, like some smart seeds? Is there some way to bypass? So I could almost imagine some kind of a model that sort of says, okay, here's where you need to explore or here's some hypothesis where to explore. And these are some of the directions my group is currently exploring, saying like, could we somehow seed or do some smarter seeding of the space?

Abhi: Going back to what your lab is working on, a theme I've seen in a lot of your lab's work and people who have collaborated with you is a belief that the existing base models are actually really powerful. And if you use them in interesting ways, you get a lot of value out of it. Do you think that there is that much value in pre-training from scratch, or do you think a lot more work could be done with the existing models and there's a lot of really interesting work that could be done there? Like you have Bindcraft, which for people who haven't even heard of Bindcraft, it basically allows for you to create de novo protein binders at ridiculous success rates, far higher than anything else that had come out before.

Sergey: Well, I guess coming back to a little bit, the earlier point that I was trying to make is: we know that Alphafold is highly limited to a certain space. Like there's, I guess you could say this idealized, non-ambiguous space. And so the question is, well, one question is: do we even need to move into this ambiguous space? Or alternatively say, well actually, you know, if we're happy with this space and there's a lot of things to accomplish here, why not just limit ourselves there? Why even explore to other things? And if we take that philosophy, then I would say, well actually current tools are fine. We could just invert Alphafold. That's what we do in Bindcraft. You can essentially use ESMFold to sort of score these sequences. We think all these models have learned this sort of non-ambiguous, idealized, I guess you say low contact order space. And let's just fully exploit that and just design within that space.

And there's nothing wrong with that. It's just a little bit less satisfying from the perspective of saying, Hey, what if we want to move into something more complicated? And I guess some people could argue that, and we argue this, is that, well, if you start to move into more complicated function... so what I mean is: if you're binding to something, we're not too much worried about, Hey, does the binder change upon binding? Is there some flexibility to the binder? The fact that it's like a rock or doesn't unfold is fine. So binding seems to be a relatively good application of these highly idealized structures.

But then if we start to move into, say, now if we want to design enzymes, now maybe we have to start to understand this sort of ambiguity. But maybe there's some sort of middle ground where you say, I'm going to try to make it as rigid as possible, but maybe destabilize a few little things. And so then becomes like a local search problem. Maybe that's one thing Alphafold can do. And so there's been sort of these combinations of methods that have been coming out, like from David Baker's group, it's like, well, we could restrict some spots based on what we believe nature has optimized for. We're going to keep that fixed and then we redesign everything else. And maybe we don't fully understand the part that we kept fixed, but we also just say that we know it moves somehow, we just have to keep it all in that same spot. But everything else we can redesign and make it rigid. And then that creates like new enzymes.

Abhi: If it turns out at least for the rigid binder world, if it turns out non-ambiguous sequences are just kind of intractable for humanity as a whole to deal with, do you think that's a big loss for anyone? Is it a big deal or is it kind of fine?

Sergey: Well, I guess it depends why you're getting into this problem. I think there are some people who go into protein design, they're coming more from like a, I guess you could say, more of a biophysics background. They're like, Hey, I want to understand this problem. And often people like to quote Feynman saying, Hey, what I cannot create, I don't understand. And so the idea there is: you want to be able to, if you could say, Hey, my model can make something and it works in lab, then I fully understand this problem. And so there you kind of want to have the model work for the reason that you understand, like the mechanism.

I mean, I guess now we're in this weird space where now we could actually make things and we still don't understand it. So it's like, what's going on here? So it's like, how do we check understanding anymore? And I think for those folks that want to understand things on more, I guess you could say on the biophysical side, they do want to sort of make it work for the right reason, I guess you could say.

Abhi: But for the people who are most interested in, I want to create a protein for a specific functionality. I don't care if I understand how it works or not. Do you think the non-ambiguous space is perfectly fine?

Sergey: I think for the most part it is. Okay. Depending on the application. So I would say like for binding, I think we're definitely... I think don't necessarily need to worry about maybe this ambiguous space.

Abhi: for enzymes it's a little bit more complicated.

Sergey: For enzymes, that's... I think when, like for example, I think some of the dreams people have is to design molecular motors, for example. And then there it's like, well, this is where maybe we can't get away with this hack. Or folks say, well, maybe upon binding you have a conformational change.

I mean, there has been work where folks have done these locker kind of proteins where they... it's like you have one helix flipping out and the ambiguity is mostly in that loop. So you essentially have one helix and then you have a loop. And that loop could potentially have a lot of ambiguity in there and it could get displaced by another helix coming in. And so I guess there are these hybrid things where you sort of combine things that are ambiguous and non-ambiguous. And so maybe there's still a lot to push in these hybrid sequences.

Abhi: If you look at the problem of at least rigid binder design, do you mentally consider that problem solved or are there still edge cases where you think there's a lot of work to be pushed on?

Sergey: I would say that it doesn't work for everything. And part of it has to do with the fact that... so one thing we find is like if you take a target that maybe has a little bit like a hydrophobic patch somewhere, interestingly, if you just run Alphafold with any random sequence, Alphafold always puts any random sequence near that patch. And then, of course, once it goes there, it's not really confident. And as you optimize, that sequence becomes more and more confident. But then there are some targets, like you give Alphafold random sequences and every iteration it just places in a different location. Alphafold doesn't see a clear signal anywhere where something can go. And I think those are regions that are more hydrophilic. So for example, if the surface is completely hydrophilic, essentially there, unless you already have like a perfect sequence on the other side... by perfect sequence, what I mean is like you have maybe the correct hydrogen bond patterns that can maybe detect one part on the surface... it's really, really hard to optimize. It's like you almost need the right answer before you even try to design or optimize.

There, this is where we think maybe methods like diffusion or flow could be useful because you sort of target, say, Hey, let's just explore here. While with hallucination, you're constantly... it's almost like you're hoping that it already knows where to go before you even start. And that becomes a bit of a limitation because you maybe want to restrict yourself to a certain spot.

[00:36:19] Diffusion vs. inversion for protein design

Abhi: That makes sense. And actually, kind of on that point, I think the last time we talked, you had a lot of really interesting things to say about the value of diffusion versus hallucination or masking. I'd love to get your pitch as to why diffusion is the way forward.

Sergey: Well, I guess I wouldn't say diffusion is necessarily the way forward, but it is definitely a step in the direction. So maybe just to step back a little bit, ultimately the protein design problem is find a sequence that folds into one structure and no other structure. But also not just folds into that one structure and no other structure, but also where the conformational landscape, or I guess you could say the folding landscape has a sort of, I guess you could say, a smooth landscape where you can actually get there. Because you could imagine you can have a sequence that has very low free energy, but there's a huge barrier you have to go over to actually be able to fold into that structure.

And the reason why we like things like Alphafold or inverting Alphafold is because we say, well, actually during design we're at every single step of design testing for that condition: does it fold into that structure and no other structure?

The problem with the method like diffusion is that you're sort of coming at the structure, the sequence is not yet implemented. But even if you do joint sequence and structure optimization, you still, at the end of the day, you're only evaluating that sequence for that one structure that you're diffusing. And you're never asking the question, does that sequence fold into something else?

I mean, of course you can run it thousands and thousands of times and then after the fact, run Alphafold on all those sequences and check for that condition. But if the goal is to, if at the end of the day you're going to be using Alphafold, why not just use it as the oracle itself?

So, I guess to describe to people who are not familiar: you can either generate a bunch of sequences and then check for this sort of, I guess you could say inverse folding property of making sure the sequence only folds into one structure and no other, or you could say, well, let me just evaluate that condition at every step of design.

And in the past, people sort of abandoned this idea of inverting structure prediction models, just because it was really, really hard to optimize. Like if you try to backprop through Alphafold, it gets very, very unstable once you start to only work with a single sequence. And it takes too many steps. And so I guess you could say diffusion was sort of invented to try to, or at least not necessarily invented, but was introduced into this field in order to get around that problem of sort of instability of back propagating through structure prediction models. But I think now we're sort of getting to a point where I would say, actually, you know what, with some of these tricks like relaxing the sequence representation, maybe removing some of the recycles from the model, we could actually do... like it's not as expensive as it used to be. And now we could actually maybe move back into this kind of approach.

Abhi: Like the inversion approach?

Sergey: Yep. Yep, yep.

Abhi: Okay. So actually you... I may have misinterpreted your original viewpoint. You don't think diffusion is actually particularly... I guess like do you have a strong stance one way or the other as to whether inversion versus diffusion will ultimately win as the primary protein design method?

Sergey: Well, I guess what I was trying to get at is that, I think ultimately it's testing for the same... even when you run diffusion at the end of the day, you still use Alphafold as your oracle to check. That's true. And so one could argue like, well if that's your end goal is to pass the Alphafold test, why not just use Alphafold throughout the whole process? And in some cases... and so usually people didn't want to do that just because Alphafold was really, really difficult to work with, in terms of inverting it and propagating through the signal. But I think the reason why hallucination has been coming back recently is because people have finally figured out, hey, actually there is a way to use it as the oracle and use it during optimization.

But that being said, there are still some problems, like what I described a little earlier, where let's say you want to design a binder to a particular location, and you give Alphafold some starting sequence, and that sequence just doesn't go to that spot. And you could add some hotspots. You can say, Hey, I really want to go there, but unless just by chance Alphafold predicts that protein to be close in that region, there are no gradients to push that sequence in that spot. Yeah. And so there's just no way to tell Alphafold, here's where I want to explore. You're almost hoping just by chance it appears there and then you optimize there. In the case with diffusion, since you're working explicitly in structure space, you could say, Hey, I have my structure. I'm going to initialize my noise here. And now you're kind of forced to explore there.

Abhi: You can guide the diffusion process like the tiniest bit.

Sergey: Yep. Yep. So you, I guess you could directly steer the structural components. And that I think is, could be powerful in that context.

Abhi: I saw that you had this, like, this Boltz Inverse Design thing that came out recently. Do you think there's going to be that big of a step up from Bindcraft? As in, some sense, Bindcraft is using Alphafold2 as an oracle. This new approach is like using Boltz as an oracle. Do you think there's that big of a step up in actual improvement in terms of generating binders? Because I remember I watched a lecture by Martin a few months ago and he said initial tests with Boltz actually showed that it wasn't that much more successful than Alphafold2. Does that also continue to be empirically the case?

Sergey: That's a great question. We haven't really done any kind of benchmarking in terms of protein-protein interactions. And was it... does it able to predict things better? The reason why we... so Yilun, in my group has been working on this... is we're thinking like, how about other problems where we don't actually have proteins that we're binding to? So let's say the protein might be, instead of protein, you say might have DNA or RNA or a small molecule, or maybe a protein with a modification. And so I guess I was thinking the... what we call Boltz Design was more, not necessarily because we think it could do better than Alphafold2, but more because it could now address other classes of problems that Alphafold2 cannot do. So anything to do with small molecules or anything to do with non-protein.

But just coming back to the protein-protein interaction problem, I guess my immediate guess would be that it wouldn't do much better. But I don't know. I guess that's something worth exploring.

Abhi: Yeah, it was surprising to me. Because instinctively I would expect, oh, Alphafold 3 has seen more molecular interactions. Thus, it probably learns to do protein-protein interactions better, and it seems like that hypothesis hasn't really proven itself out. Is there good intuition as to why?

Sergey: Well, I mean, at the end of the day, like if you sort of look at Alphafold2 versus Alphafold 3, and then all the different implementations of Alphafold 3, the core parts of the model are the same. Like you still have the Evoformer, and then you still have the Pairformer, which removes parts of the MSA. And the only thing that really got replaced was the structure module got replaced with the diffusion module. But to me it seems like both of those things are kind of doing the same thing. Okay. Like it's just one is a bit more deterministic than the other. Like one, you start with the noise and then you try to satisfy some constraints coming from the pair features. One, you start all 0, 0, 0 and try to match those constraints.

And I guess if I was to guess, I would say maybe the DeepMind team probably thought that by doing the first part and the more expensive part, we're going to kind of like separate it as a separate component and then do a lot of diffusion afterwards, like many, many iterations, that you would get many different solutions. But in reality what happens often is the first part just gives you back the same constraints. And so then you always sample the same structure. And so I think there were great plans there, ideas there that might have improved things. But ultimately I think the structure prediction part maybe didn't get helped too much by these updates. But that being said, like all small molecule, non-protein things definitely improved.

I mean I might take a little part of that back because I think in the Alphafold 3 paper they did show for like, for example, antibodies, you can do far more recycles... I mean not, sorry, far more different random seeds and you could find the right solution. So maybe there are some improvements there.

[00:44:52] A problem with Alphafold3

Abhi: On the topic of Alphafold 3, I remember you talking one point, and I might completely misremember this, that you had some quibbles about how the recycling process is being done in Alphafold 3. I'd love to hear you just expand on that.

Sergey: Alright. So yeah, so I guess for those of you who are not familiar: Alphafold2, you make a prediction and then you take that structure and you feed it back into the model and you say, make another prediction. You sort of keep cycling this over and over and over. And each step, the structure kind of improves and improves.

And one thing we've noticed with Alphafold2 is if you try to do like homooligomer predictions, if you look at recycle zero, all the copies are literally overlapping on each other completely, overlaid on each other. And then each recycle kind of pulls the copies apart so they're no longer clashing with each other. So what we think is happening is the Pairformer, or I guess you say the Evoformer in the context of Alphafold2, sort of learned some constraints, but these constraints are physically not possible. Like you try to actually realize that structure and you realize that you create a bunch of clashes. And I think one of the cool things about Alphafold2 is like you see that there are these errors in your structure and you feed those errors back into the model and you refine those constraints and you keep refining it.

One thing people have noticed with Alphafold 3 is that you start to see these clashing problems. Like if you try to predict like a homooligomer structure, these copies are all overlaid on each other. So there's no feedback in Alphafold 3 back from making an attempt to predict the structure and then refining the constraints.

Abhi: I thought Alphafold 3 does have a recycling component.

Sergey: So it has recycling, but it's only within the MSA and Pairformer.

Abhi: Gotcha.

Sergey: Like there's no... the structure that gets predicted doesn't get fed back into the model.

Abhi: So there's almost like an original sin problem going on a little bit where it has no chance to try and modify what it predicts.

Sergey: I mean yes, exactly. Exactly. So it's like you try to realize those constraints, but there's never feedback that you failed to realize these constraints. And I think the reason they did this was well-meaning because I think they were hoping, Hey, let's do this expensive... this whole cycling or this recycling one time and then run diffusion thousands and thousands of times. Yeah.

Abhi: The hope is that diffusion does what you're asking for, right? Like it realizes the constraints early on and then refines it. Why do you think... because in my head, I've always mentally thought of recycling as like a pseudo-diffusion process. How come the actual diffusion process doesn't seem to do what recycling is doing?

Sergey: I suspect the reason is because the diffusion part doesn't have a chance to sort of remodify how it's interpreting the coevolution signal. And so what I mean by that is we think like the MSA module and then the Evoformer or the Pairformer is taking the evolutionary data and trying to clean it up or trying to disentangle some of maybe ambiguity in the constraints. And but once you've disambiguated all those constraints, you're kind of now stuck to just try to satisfy those final disambiguated constraints. But what if the ambiguity didn't work out correctly? Like you didn't correctly disambiguate these things. And the only way you would know is to try to embed those constraints in a 3D structure. And I guess if one was to go back to the diffusion part and let it reprocess the MSA at each iteration, then maybe that would work. But then you're back to what you call the pseudo-diffusion part of Alphafold2. Yeah. Because that's essentially what Alphafold2 is doing. It's like you're doing this iterating process and you're fixing the constraints, refining, I guess you see the potentials, and then rediffusing the structure and you're iterating on that over and over and over.

So yeah, but I would say like one simple fix to all this is say, Hey, let's just feed the structure back in. My guess is the reason why they didn't do this was because of that thing that I talked earlier about these hopes. But that would be one, I think an obvious thing to sort of say, Hey, let's bring back this iterative refinement step. And that would be one way to maybe fix that problem.

Abhi: Has any of the Alphafold... I haven't looked too closely at any of the Alphafold 3 replications. Have any of them tried to poke at this problem or they haven't really touched it?

Sergey: They, from what I've seen, they haven't, no. Okay. And I think part of the reason I think everybody right now is just focusing on saying, let's reproduce the baseline. I'm sure that some of these groups are already thinking about how to improve on top of this, but I think right now the goal is to reproduce the baseline.

Abhi: That makes sense. Um. And kind of like on this note of replications, we seem to be kind of entering... the field at large seems to be entering something similar to the GPT-2 moment in biology, or phrased more differently, a period where the best models are being built by private companies and are for the most part, never open sourced. I'm curious to hear your own thoughts about this as someone who has built probably the most popular open source implementation or open source plugin to Alphafold2. Where do you think the trend of open source protein models and biology models is going?

Sergey: Yeah, I think it's definitely sad to see companies now not wanting to release their code. Because I always felt there was a synergy between industry who have a lot of resources like GPUs and building models, but often having no idea what the models are doing or how to interpret them. And then academic groups taking these models and trying to interpret them. And so there's this synergy: somebody builds a model, a bunch of people are trying to interpret it, figure out what's wrong with the model, what is it learning, what is it not learning, and then feeding back into the industry and the industry could try to fix those models.

I'm guessing maybe industry has reached a point where they're like, Hey, we've now... it works too good. We can make billions of dollars on this. And so they have less incentive to release these models for academic research. And yeah, so it's definitely... I mean, I'm personally not too happy that they're not releasing these. Yeah.

Abhi: Do you... like in the NLP world, people thought things would go the exact same way, but you had entities like Meta release the Llama models. Do you think there'll be some entity like that in the biology world where some either very strong academic group or some very caring for-profit institution releases good open source models? Or do you think that's also unlikely?

Sergey: Well, I mean, like if you look at what happened with Alphafold 3, there have been indeed groups like, I think like ByteDance for example, released Protenix, and I think they've now also made it for both commercial and non-commercial use. Chai released their model. So, and like HelixFold, I think there have been a few groups, I guess similar to the whole DeepSeek kind of thing, where they're releasing everything for people to be able to use these kinds of things. And so, yeah, I think there are always going to be some groups out there that are just releasing.

I'm not sure how much it'll continue. I think sometimes what people do is they... industry, I'm guessing, is they would release some early versions of things just because it's really good for PR and so on. But then once they get things working much better, they might not want to release it anymore. But then maybe some other group catches up and releases everything. So it's kind of a... so I think there are going to be cycles where people try to... they think they have something cool and they're trying to hide it, and then somebody else will catch up and eventually get there.

Abhi: Kind of on this note, at least amongst NLP frontier AI labs, there's some level of homogeneity amongst all the approaches they're going. For the most part, everyone's doing autoregressive models, but there's also like one group doing diffusion language models because they think that's how you get better language generation. It seems like for the most part, almost everyone in the frontier biology labs are pursuing the exact same approach, from my perspective. Do you see strong levels of heterogeneity in the underlying details, or would you agree that they all seem to have a very similar thesis?

Sergey: I think sometimes it looks like everybody's doing the same thing because everybody's just trying to reproduce the baseline. But once the baseline is produced, then people start to explore more. So, I mean, it's true that once something comes out, like somebody says, Hey, we have a working solution, everybody's trying to catch up to it. But once people have caught up to it, then people say, Hey, you know, how can we improve it? How can we change it? But every year you always have somebody else say, Hey, there's another way of doing it. And so there are always these shifts. And I guess in our world we do see, like people are doing some kind of MCMC hallucination and then people switch to diffusion, and now people are moving into flow and now hallucination's coming back. And I'm guessing there's going to be a sort of... but there's always like some group that shows something and then everybody tries to catch up to it. Yeah. So I don't think there's that issue per se.

[00:53:41] MSA vs. single sequence models

Abhi: Yeah. It does feel like EvoScale is the most unique amongst everyone by eschewing MSA. Do you think getting rid of MSA is going to be a good long-term solution to not having to deal... because the MSA track is annoying to deal with in practice. But do you think it's so essential that it's very hard to escape from needing it?

Sergey: I guess our argument is that even the models that do work on single sequences, they are learning something akin to an MSA.

Abhi: Mm-hmm.

Sergey: Not necessarily memorizing MSA per se, but the statistics that you would have for a given protein family, it's learning these things. So. And I guess it's sort of a debate between... I like to think of it as sort of a debate of storage versus lookup. Like you could have a model that just memorized all the statistics, or you have a model that's much smaller and can retrieve some information. And then using that information, recompute some of the statistics.

And in some ways you could sort of see parallels to that in the OpenAI field, where they're often thinking like, Hey, instead of making these models bigger, let's just give it access to Mathematica or give it access to, I think in their case it would be Bing, instead of Google. But it's like giving it access to search. And MSA is actually sort of, I guess you could say essentially a version of search. Like you search for a few things and based on that information you compute the statistics.

And so I guess the question is: what makes more sense? Do you make models larger and larger, or do you make smaller and smaller models, but give it access to information? And I personally prefer the latter saying, well, could we just make smaller models that have access to information?

But that being said, I think folks that do evolutionary scaling kind of stuff, I think Alex Rives and so on, they do believe that as you scale eventually, maybe at some point it has some kind of emergent property. Like right now, to me, it seems like it's mostly learning conservation, coevolution, maybe a bit better coevolution in terms of maybe being able to share statistics between protein families that we previously were not sharing. But I guess the hope is at some point it snaps out of that behavior and says, Hey, now to make it even better, you need to learn physics. Yeah. And so that, I think that's... I think it's worth exploring. I mean, it's going to be really expensive. But if we don't believe in emergent behavior, then it's the question: why bother scaling? Why not just give it access to a bunch of sequences and then let it search for some of those, that information itself.

Abhi: And like one alternative thought. Maybe instead of providing MSA during train time, you provide it during inference time. Like, I've seen a few papers try to do like a RAG approach when it comes to actually providing MSAs. Do you think that approach has much steam in it or there's like some reason it doesn't really work that well in practice?

Sergey: Could you explain the process again?

Abhi: Yeah. Like the idea is like you provide the sequence to the model and the model can refer to like a retrieval augmented generation database which contains embeddings of homologous or aligned sequences. And it can just feed that information in and that's how it computes things.

Sergey: Yeah. Yeah. I mean that seems like one way to do it. I mean, I guess you could always... I mean, to step back a little bit, one of the issues with MSAs is that one is often you don't actually know what sequences to include. Like you get a bunch of sequences, but some of those sequences might not belong. Like maybe they have different conformations or maybe they're completely from a different protein family. In some cases it's just really hard to find these sequences because maybe they're less than 20 or 10% identical to the query sequence. And so some of the typical methods people used might not be able to find these sequences. And then finally there are problems of just aligning the sequences.

And so I guess the retrieval methods could be useful because they're able to potentially say, well, instead of doing sequence alignments to find those sequences, maybe there's some barcode I can use to find those guys. Yeah.

Abhi: But you're probably trading off on something and you're losing... You're using a heuristic, really. You're unable to do the actual alignment and you're trading off on some level of accuracy.

Sergey: Yeah, yeah, yeah. I mean, I think if you try to actually take every sequence and compare it to every sequence, it's just going to take forever. It's going to... like even making one multiple sequence alignment on hundreds of CPUs would take probably multiple days to do a search against all the sequences out there. And there are all these heuristics like, Hey, let's try find... like for BLAST the way it works is sort of, you chop up things in small words and you say, do any of the words match? And MMseqs2 does this to a similar extent. And then ungapped alignments, then gapped alignments and so on. But you're probably losing a lot of things. And so you could always find more sequences by turning off all these pre-filtering. And so in some ways, one could also argue that, well, maybe there are just issues in making multiple sequence alignments. And so having a model that just stores evolutionary statistics is better... Yeah. Or alternatively, maybe there are other ways to retrieve sequences that don't require explicitly aligning things or looking for matching words.

Abhi: It feels like, I imagine what some people at EvoScale think about is if you try to do the alignment yourself, you can't include every possible alignment. You have to make some sort of cutoff as to how many you include. And like, how exactly do you do the alignment? And it feels better to have the model decide what is important versus what isn't important. I guess this is a question more for your phylogenetics background. Do you think the way people handle alignments with models like Alphafold2 are actually pretty sophisticated and there's not that much further optimization that a zero priors ESM approach could really improve upon? Or you imagine there's still a ways to go to improve things?

Sergey: I think there are still ways to improve things. The reason I'm saying that is actually in the most recent CASP, people have found the best working methods is to essentially try to create multiple sequence alignments in thousands and thousands of different ways. Like people try lots of different methods, different E-value cutoffs, different coverage cutoffs. And you could always make better predictions by sort of sampling lots of different kinds of MSA generation methods. Or also what you include and don't include.

And in some ways we've also found this before. One of the projects we worked on before, we said, well, what if you make MSA generation differentiable? Like let's say you have a collection of sequences but you don't know how they align to each other. Could we pass them through some kind of differentiable Smith-Waterman module and then try to figure out the best way to align them to maximize the pLDDT in Alphafold? And we did actually find that for proteins that have very few sequences, sometimes just realigning one sequence made a huge difference. And so there's still room to improve there.

In some cases it worked for the wrong reason. And what I mean by it worked for the wrong reason is it turns out... this is work from Sam Petti, who's at Tufts now, but she looked at some of these alignments that we generated, like these multiple sequence alignments, like, Hey, we got a better multiple sequence alignment. Was it actually better? And it turns out in some cases, the sequence got misaligned and that improved Alphafold because right now what happens, this currently aligned sequence was influencing the covariance statistics. And so if you misalign it, it sort of gets averaged out. And so in some cases, some sequences just do not belong. And by realigning them, we actually removed it. And so it's like in some cases... so I guess you could say the MSA problem is still unsolved.

Abhi: Yeah. Like, I think one question that I have tried to figure out myself, and I've kind of come up empty-handed is: is there a principled way to improve the MSA? And it seems like the answer is no, really? Like the CASP result AFsample where you're just rerunning things like a billion times with different MSA subsets. It's kind of under-discussed in the paper as to what makes one MSA better than another MSA. Is there actually a principled approach to decide which one is better other than just looking at the confidence metrics? Or is there some often like some phylogenetic theory that you could rely upon to prove that it's clear why this one MSA is better than another?

Sergey: Yeah. Yeah. Well. Yeah, I guess it's an unanswered question, like why does it get better? I guess one hypothesis is like, well maybe each MSA is essentially initializing this global search. Coming back to the earlier topic, like each MSA is sort of perturbing it enough to start somewhere else, and now you found the solution. And it's not really the fact that the MSA is any better. You just added a little bit of noise and now you're starting somewhere else. Yeah. And so maybe if you just use the same MSA with lots of random seeds, maybe we would have gotten the same answer.

That being said, other researchers like Hannah Wayment-Steele found that actually if you cluster the multiple sequence alignments, you can look at different clusters and different clusters seem to have different maybe evolutionary signal, and that could push Alphafold to do one thing versus another thing. And I guess originally I would've thought that wouldn't be the case because with Alphafold, you have attention between the sequences and so you're like, well, Alphafold should be able to figure out which sequence to include and not to include. But in practice it looks like you can get a better signal by essentially subsampling the MSA closer and closer to the query.

That being said, I think other researchers have shown that actually sometimes you can just make random MSAs and you can still explore that. It doesn't necessarily mean that one invalidates the other. It just turns out there are lots of ways to get Alphafold to predict alternative structures. Yeah. Like you can maybe get it to predict the alternative structure because you're just initializing in a different search space, or because there's some bias in one place or another. And so there are a lot of ways to sort of push Alphafold to do different things. And whether or not they're doing it for the right reason or wrong reason is sort of still a debated question, I guess you could say.

Abhi: I think kind of related to this point, I have been... a lot of your papers kind of poke at the problem of mechanistic interpretability when it comes to these base models. But as far as I can tell, you haven't clearly tried to train an SAE on ESM or anything like that. Do you think there's a lot of steam in these approaches being applied to protein models or there's some gap that you're unsure about and that's why you haven't tried exploring it yet?

Sergey: I mean, we have looked more specifically at different attention layers and seen different attention layers maybe learn different things. I guess the reason we never really... I mean there have been recent papers, I think a couple papers where they look at these sparse autoencoders applied to the... not our papers, but other folks have done this now.

Abhi: Yeah. There are people from Reticular, I think. I think also from MIT, they started a company around figuring out which layers of ESM correspond to the specific secondary structure.

Sergey: Yeah. I guess, I mean, it's definitely an interesting sort of thing you could look at to see like, is there parts... I mean, the reason why we personally haven't done that was because we're like, well, the attention layer already tells us there are different layers learning different things. And you sort of see secondary structures popping up and so on. So that, I guess for us, the attention layers seemed to be already interpretable from that perspective. But that being said, maybe there are certain features that are not necessarily in the attention themselves. That may also be... I guess it would be interesting to compare these, I guess you could say sparse autoencoder features to the corresponding, I guess you could say attention heads. And see are they picking up something different or not.

I mean, I guess to step back a little bit, folks have in the past, like Martin Weigt for example, have looked at sort of taking these POTS models and decomposing them and finding that different sort of, I guess you could say components of the POTS model have sort of connection to different features of the protein. So I guess one thing that might be fun to explore might be to say, well, what if we just take the categorical Jacobian from a POTS model, I mean from a language model, and decompose that and see do these correspond to similar kind of features that SAE models are picking up on?

And I think, I think there's still room to explore. That being said, I think that one exciting thing about these sparse autoencoders is that perhaps if you could figure out exactly in the model what is being activated to make a specific prediction, then that could be some kind of steering approach. Like you could say, well, now I know if this is activated, it means that this is going to be an enzyme or something. Yeah. Or this is going to be a secondary structure. And so then maybe during hallucination, if you want to use that approach, or maybe you could somehow enable those things. And so if you could figure out those connections. And so, so I guess what I'm saying is that there might be actually some benefits to it that are beyond what we're working on that we're not considering at the moment.

[01:06:52] How Sergey picks research problems

Abhi: And kind of on that note, this is leaving the realm of science a little bit, not entirely, but if you were a PhD student today, what do you think you would be working on? It can be either in this field or some other field.

Sergey: Let's see. It's a great question. I haven't really thought about it too much. I think for me, I've always... yeah, I don't really have a good answer for that for you. I'll have to think about it some more.

Abhi: Well, you can come back to that question. If I guess to give some hypotheses, it actually feels like when you were describing your original personality when you were back in high school, it actually feels like the type of mind that would've gone to physics much more than biology. Do you think... and like it turns out you studied phylogenetics, which is kind of like a field which lacks a lot of translation to the real world. Do you think it's good for someone to focus on something that has very theory heavy and relatively little immediate application in the hope is that you can eventually convert that into actually useful research during grad school or during further research? Or it's good to immediately start with the immediate applicable stuff straight up?

Sergey: It's a good question. So I think for me, I've always done things that I felt were more fun or more curiosity-driven research, I guess. I've never sort of sat down and said, okay, this direction is probably the most meaningful thing to do. It's more just like, okay, this is like a puzzle and there's no solution here. I'm just trying to figure out what's going on here. And I think when you go down the path of saying, okay, what's going to be the most... I don't know, going to make me the most money or something... I mean, I guess people are driven different ways. For me that's not really what I think about when I think about thinking of a new topic to work on. For me is like, is there some unsolved problem here? And like, why does this work the way it works? And sort of... I think I just like to solve puzzles is maybe where...

Abhi: So it's not even necessarily like, I want to have this grand scientific impact. It's just like I get nerd sniped by something and I spend years working on that.

Sergey: I think so. Yeah. Yeah, yeah. Exactly. Exactly. I mean, sometimes you're lucky and you happen to get curious about a topic that actually turns out to be really popular and takes over. I mean, when I was moving into this whole coevolution field, I didn't think that this would be some hot topic. So maybe I got pretty lucky in picking an area that happened to have exploded recently. I mean there's definitely like a chance or like a lottery thing here. Yeah, like, for example, when I was moving into coevolution, a lot of people actually told me like, don't do this. This didn't work because people have been exploring this since the nineties and seventies and so on. It never quite worked. And so they said this is the wrong area to be in. But I still believed I could figure it out. And so... but it's also those areas, people didn't think it was going to work at the time. And so... I mean like right now, if I was to give somebody advice, I would say, well, if it's an area everybody thinks is going to work, there's probably going to be a bunch of people doing it because everybody else thinks probably the same way this is going to work. And so it's almost more fun to say, okay, here's something that no one thinks is going to work.

Abhi: Okay. That's good advice. So, not field specific, just pick something that doesn't seem to have billions of dollars worth of attention being focused on it.

Sergey: Yeah. Well I'm not sure if it's always good advice. Like, maybe ultimately you want to pick something that's not completely crazy, but... yeah. But I would say whoever's listening to this, don't listen to me because I'm probably giving really wrong advice here. You probably want to pick something that will secure your future somehow. But at least for me, I've never... every time I pick a problem, I always think like, it's more like, let's explore it. Find something cool and see what we find. And sometimes you get some cool things that come out of these things.

Abhi: Do you think... I know you mentioned you had this passing interest in pursuing history when you were in college. What would an alternative Sergey look like? If you had not touched biology at all, what field do you think you would have focused on?

Sergey: It's a good question. I mean, I did find that I enjoyed history, in terms of understanding what happened in the past. Maybe this is also why I was interested in phylogeny because that's also history. I'm just sort of making this connection now, potentially. But one thing I found is that when I went to college and I started taking some of these history courses, I quickly realized I didn't have the right sort of, uh, I guess you could say background, because at least my colleagues were all like, they all learned Latin at some point. I was like, and I don't know that. So I quickly realized this is not something I was pre-trained to be able to handle. And so, but I really enjoy just learning about the history of science because like what people have done before, how people decide to work on DNA and so on and so on. That's kind of... and so, but I think it's kind of fun to see it from that perspective. I don't think I would be able to get into it now because I probably have to go back and learn Latin again, but...

Abhi: Well, at least for history of science, you probably don't need to learn Latin. Do you think you're an active consumer of, like the Genentech book that came out that explained the history of recombinant insulin? Do books like that particularly appeal to you and you enjoy reading about how scientists in the eighties and nineties really worked on things?

Sergey: Yeah. Yeah. I definitely find it really exciting just thinking about... because I think one thing that I found quite fascinating as you sort of look at all these different stories of how people made breakthroughs, often it wasn't because somebody actually had a very specific thing they wanted to test. Often it's like somebody does something completely random and they see some interesting unexplained signal and they start to pursue it, and they find and they get a really cool discovery. And so I think a lot of the most important science have been sort of made that way. Like people sort of see a random signal and they pursue it. And that of course sort of brings up the question of like, how do you actually get this research funded? Right? It's like, if you don't have a direct question and you just want to explore in the cloud and find something... but that being said, I mean there is research for... there are questions in terms of basic... basic research, I mean, there's funding for basic research. What I'm trying to say. And like for us, often what I do is we try to propose a general area to explore. And every time we start projects, we start exploring areas, but then if we see something interesting along the way, sometimes we change direction to that area.

Abhi: When I think of your lab, I consider it much more from the outside an applied biology lab. In your head, do you still consider yourself a basic research scientist?

Sergey: Yeah, I would say we're more... I like to think we're more basic. In terms of thinking about more theoretical problems, big picture problems. We do collaborate with a lot of people who do have specific applications. So for example, some people want to do protein design and we say, Hey, maybe some of these tools or some of these recent hacks we found could be useful for protein design. And then we collaborate with people like that. But yeah, I like to think that we're more on the basic sciences. Yeah.

Abhi: Do you have a... when it comes back to these... you pursue problems that you're often just feel innately interested in, and it just makes you want to pursue it for years on end. Do you think you have a good sense of taste for what makes for a good problem and what doesn't make for a good problem? I remember reading this article a while back about types of problems you should focus on during your PhD and what makes for something good to focus on for the next five to six years. What's your own internal sense for what those types of problems are?

Sergey: That's a good question. So I'm, disclaimer, I'm a new PI, so I haven't seen... I only have a sample size of one being myself. So I don't know if these are good answers or not. It's... I mean, I guess one could say that... I mean, I think it's good to start with some problem, regardless if it's that interesting or not. Because I think you sometimes just need to get yourself familiar with the field. Right? So like, you start with something, say, Hey, I don't know, protein docking. Right? And so regardless if you're interested in that problem or not, but as you start to explore the different tools, you start to realize different limitations of these tools. And then something you might come across, something interesting that you pursue further. And so, at least the way I've currently been doing this with my current students, we'll see, we'll see five years from now, if it's actually a good idea or not, is to say, Hey, let's just start working on something. Like, here are some interesting areas that I find interesting, but with the intent that at some point we'll see something else. And sort of thinking back now, when I look at all the things that we have published, I don't think there's been a single thing where this is what we thought we'll be publishing five years from now. It's always like we start with something and we're like, oh, we found something really cool along the way, and then we realized, hey, this could be applied to this. And so on. And so I think it's just like getting started, starting to explore. And I think curiosity is probably a really important component. Like you see something that doesn't make sense and you sort of keep pursuing it.

Abhi: So you typically don't have like a two to three year plan for what paper do I expect to publish in three years?

Sergey: Well, we do start with that. Like the idea is we start off with a plan. It's like, okay, here's a problem. But I guess what I mean is when you start that problem, you might end up deviating from that problem. And that's totally okay. At least the way I view it. It's like you start... you have a... I mean, I guess in a maybe more ideal world, one would say, okay, let's start with maybe it's good to have two problems, like one that's a little more safe, you know? Okay. If you do these steps, it'll work. And then you say, and then you sort of have like, maybe, I guess maybe similar to optimization, you have like a, I guess you could say, I was thinking more like simulated annealing. It's like you have different chains. Like you have sort of a fast moving chain, a slow chain, and then you could potentially do some kind of recombination, but Okay. But this is kind of getting a little... but I would say it's good to have a plan, just in case. Yeah. But I would say don't deviate, but don't be afraid to deviate from the plan. Because sometimes you might come across something cool or interesting or unknown and it's okay to shift over and explore that unknown. And that's usually where the most interesting impact will come from.

Abhi: Do you think it's generally when you're setting up this initial plan to start off with, do you think it's good to focus on incredibly ambitious, largely intractable problems, or it's good to start with like a pretty close-ended thing that you know if you apply enough engineering effort, you will have a paper at the end of it?

Sergey: Yeah, that's a good question. I don't think I have a good answer for you because... I think right now I've, at least anytime I've done any research myself personally, I've never said, okay, let's, if we do this, we'll get a paper. Okay. I was like, Hey, let's explore this area. I know eventually we'll get a paper about something, but whether or not we'll get the paper for that specific thing we started with... And at least that's the way I've been doing science whether or not it's a good idea, if that's a good idea or not, I'll find out later. But I feel these are kind of questions that might be more relevant to ask somebody who's been a PI for many, many decades, they could probably tell you, Hey, I've seen all my students and these are the things I experimented with, these are the things that worked and didn't work. But just coming from myself personally, I've always felt like it's good to have a goal in mind. But deviating from that goal is going to be an important component to actually find something really cool.

Abhi: I imagine being a PI, you're a relatively new PI, you've been around, I think at MIT since January 2024. And I imagine as PIs must do, you have to specialize a little bit? You can't kind of be all over the place, especially early on in your career. Do you think there are areas outside of protein structure determination and protein binders that you wish you could have an extra 12 hours of the day to focus on?

Sergey: Well, I would say if me personally, I mean we are somewhat protein-centric. But we are trying to expand a little outside of that in terms of thinking about genome scale related things. Because ultimately proteins don't exist in isolation. But even for the same protein, the RNA itself might actually have other components there that determine the structure. So what I mean by that is, for example, like different codon usage could potentially influence how a protein folds, just because they might stall the ribosome in different parts. There might be some things upstream or downstream of the sequence that maybe changes the influences the expression levels, or maybe even changes where things change. So I guess in some ways we are also exploring that other side.

Abhi: Do you think like your lab will start poking at RNA and DNA models in the near future?

Sergey: Yeah. Yeah. So, so we, for example, have been exploring models like Evo coming out from the Arc Institute trying to understand what are they learning, are they learning something different? And yeah, this is part of the, I guess you could say... I think we're kind of moving more and more and thinking about proteins in the context of a genome as opposed to proteins in isolation.

[01:21:06] What are DNA models like Evo learning?

Abhi: I'm curious, what's your take on Evo1, Evo2? What are your general thoughts about it?

Sergey: Yeah, it's quite interesting. So we, I mean, we've tried applying some of these techniques that we developed, like the categorical Jacobian to Evo. And we are currently not seeing that it's learning contacts in terms of protein interactions. We're still internally debating, is this because the approach that we develop, this categorical Jacobian, just doesn't work for these kind of models being autoregressive and being predicting single nucleotides? Or is it because the model itself is learning some other signal that's dominating the protein signal?

Abhi: Sorry. When you say that, are you saying it's unlikely Evo2 is actively relying on co-evolutionary signal?

Sergey: So what I mean is: to step back a little bit, when you do have a model that takes a sequence and returns the sequence, you can ask the question as you perturb the inputs, how do the outputs change? And so if you think two positions interact or are dependent on each other, when you perturb something, let's say in position 10, the logits or the outputs should change differently in position 1 versus position 10. And when you do this kind of perturbation experiment in protein language models, we clearly are able to recover contacts like saying these are interactions.

But when you do the same thing with models like Evo and Evo2, you see it's able to recover RNA interactions like RNA stems and so on. But we don't really see a strong signal for protein-protein residue-residue interactions. And so one hypothesis is like, well, maybe there's another dominating signal that's completely sort of washing out. Like it's possible it's still learning contacts, but maybe other forces at play, for example, encoding start and stop codons, encoding the starts and ends of genes, maybe that signal's much more powerful. And so it's kind of washing away the contacts. Gotcha. So maybe we just have to disentangle the signal. Or maybe it's not learning any contacts at all. And so this is sort of still an area that we're investigating.

I guess another hypothesis might be that maybe the... if you sort of look at any given codon, and you sort of... if you remove, for example, the third codon position, you could still recover what the amino acid is. So like, for example, masking the third nucleotide of a codon is not really informative in terms of learning coevolution because it recovers it. Same thing, even if the first position, like if you mess with the first position of the codon, you often stay within either hydrophobic or hydrophilic. You're actually not changing the property of the amino acid.

Abhi: So there's something off about the masking strategy that Evo2 is using that leads to these strange results?

Sergey: Well, that's one hypothesis. Like maybe the reason why for RNA works perfectly, because with RNA, you're actually every nucleotide, or I guess ribonucleotide is interacting with something else. Yeah. And so if you mask it, then of course you need to change the other guy to compensate it. Or if you, not mask, but if you mutate it, I guess. But with protein sequences on the DNA level, it's possible that maybe you could just recover a lot of the statistics of things that you've masked just based on local properties. Like if you're in this codon, you know you're going to be an alanine, so the third one is not going to tell you much difference. Or if you mask or if you mutate the first one, you could look at the other two to figure out what the first one should be. Even the middle position, like for example, transition versus transversion keep you in the same amino acid property. And so maybe it's just a little too easy to reconstruct the masked tokens without having to actually fully understand the protein interactions. But that's something we're still investigating, so we'll have to see.

Abhi: And kind of on this note of like potentially off masking strategy for Evo2, clearly modern day biology models have a lot of pathologies and edge cases where they don't quite work as well. Do you think the fault of that can be largely placed on the internals of the model itself? Like the engineers and scientists need to fix it? Or do you think it's often much more like a data problem? Because I think for a long time, the data problem was clearly focused on by a lot of people. And it seems like you can actually get pretty big gains by just messing around with the model, which is what your lab does a fair bit.

Sergey: I mean it could be that... so I guess another hypothesis of why we don't see contacts is that maybe the model just needs to get much larger. Yeah. So what I mean is like if we do believe that all language models are doing is just storing information or storing evolutionary statistics, there's probably a lot of evolutionary statistics that you need to store about the genome. And suddenly there's just too much things to store. And so you're probably going to store some of the, I guess you could say the low rank signals like conservation and so on. And then maybe as the models keep scaling, you might start to learn more of these sparse signals, which are contacts.

And so it kind of makes sense as the first signal you learn is conservation. And then as you scale... and that's actually what we see, for example, with some of the earlier protein language models. Like TAPE, a lot of times just learned conservation. But then as we moved on to ESM2 and you start to scale up, you start to see like, hey, now if you start comparing the different models that were trained, some of them don't learn any contacts and some of them start to learn contacts as you make them larger and larger. So it's also possible we're just like the very early stages, like Evo2 is just the very first early model. And that maybe it needs to be like 10 times bigger or more to finally start to learn those details.

Abhi: Do you align with the idea of DNA is all you need? Or do you think at some point you need to bring in structural information?

Sergey: It's... I mean, I guess one doesn't necessarily need to bring it explicitly because I guess the idea is like, if structure's important, then the model, the attention would learn those structural constraints implicitly. Like if that... like for example, if structure's important for the reconstruction task, it should be able to pick up on that information. And that's what we see with ESM2 and so on.

That being said, there's possibility that you could unify some distance space by introducing structure. And so what I mean by that is, like, for example, people have trained SaProt from a group in Westlake. And then ProST T5, for example, they started introducing structural information into language models. And the SaProt team, for example, found that sometimes you could do much better at predicting effects of mutations by introducing structure. But the structure resolution is so low. Like they're using 3Di tokens from Martin Steinegger's group. And so it's kind of like, if the resolution is so low, how is it able to use that information?

And so one of the current hypotheses is like, well, maybe what's happening there is that you currently, like maybe let's say in the language model space, you learn that this is protein A, this is protein B, and they're far away from each other. But in reality, they might have really similar structures. And by maybe introducing even some really low resolution structural information, you sort of start to align those spaces together. And so maybe the model is sort of learning different categories of proteins, and those spaces are not well aligned to each other, and you're sort of bringing them close together. And so now you can borrow statistics from neighboring families that you didn't borrow from before. And so I think structure could help maybe sort of learn a maybe a better latent space of protein sequences.

Abhi: So it feels unlikely that a single modality will dominate, like there'll be a single multibillion parameter model using only one modality of data that'll dominate every single benchmark.

Sergey: I mean, theoretically it should be possible. Like it should... technically, DNA should be all you need. But in practice, I think if you don't have a large enough model, or maybe if you don't have enough... something wrong with the... like maybe you get stuck in a local minimum. I think sometimes structure could help steer the training in the right direction. But theoretically it shouldn't need, like, you shouldn't need it, I think.

[01:29:11] The problem with train/test splits in biology

Abhi: Yeah. Makes sense. And one question one of your old students actually told me to ask you is, first context for the question: Making test-train splits for molecular and protein engineering is really hard. Since there are so many possible ways to leak data. You need to simultaneously think about homologous sequences, structure similarities, and functional similarities between proteins in your test set and your train set. What do you think most papers get wrong about splits in this field and how could it be improved?

Sergey: Well, I think one of the issues is that... we have, it's not an issue. I think it's good. We have lots of people coming from the computer science world into biological problems. And so I think it's great that a lot of people are coming and helping us solve all these problems. But you do have to sort of consider the relationship of things. But even when you do consider the relationship and things like where do you set the cutoff? And so this is where the debate in the field is like, well, you can use sequence identity of 30%. And this is usually people say, but where that number 30 comes from is usually people say, if it's 30 or higher, it's probably doing the same thing. But you can still have things less than 30 and do the same thing. It's more like 30 is more of a cutoff of something being positive, but it's not a cutoff of something being negative. And so I think people often saw the number 30 and they're like, okay, we're going to split at 30% identity. But in fact, you could have proteins that do the exact same thing, have the exact same structure, and with sequence identities as low as 10%. And so that's where then you guys say, Hey, I split by 30. But in reality, you still have an overlap in the train test set.

And I guess one could say, well, let's use structure, let's try to use structure here. But the problem with structure is that you also have a similar problem where maybe in one organism, there are domains that are in different orders. So for example, you might have the exact same essentially, I guess you could say protein, but this protein has a different arrangement of the domains or has a couple more extra domains, or has some disordered loops or something. And then you compare those two things and say, Hey, they look different. But in reality, they have some distantly related relationships because any given protein is actually made of a bunch of domains. And these domains could be in a completely different order between different organisms. And so if you do a structural superposition, you might also be misled to think that they're more different than they actually are. Yeah. So it's almost like you first have to split by domain and then do domains. But then the problem is that within domains you could also have rearrangement of secondary structures. Like, you'd have the same exact protein, but maybe the order of the secondary structures are actually in different orders. And so then... so that creates a bit of a problem, like how do you... but then at some point if you start to, like, you say, well, let's just split by amino acid. And then it's like, well, everything's related. So that, and so that's kind of I think one of the tricky things is you have to be a little careful in that space.

It depends what your claim is though. I think for example, to step back a little bit, I mean, before deep learning, there's always been a field of remote homology search. Like say, Hey, if I could find a similarity between one thing and something super distantly related, like less than 10% identical, and I could say that they're the same, it's actually still a very, very important problem to work on. Because you could say, well, if I can find remote similarities, maybe I can make hypotheses about those things to say, if this looks like this, maybe this also does the same thing. And so there is benefit to remote homology search. And so if you say, Hey, I have a model that does better remote homology search, that's actually okay. But if you say, Hey, I have a model that works because it learned physics, then suddenly now you have to be more careful how you're going to split your data. You have to say, I have to completely make a huge giant hole in the sequence space and say, does it still generalize there or not?

Abhi: Do you think there's... I haven't actually seen that many people do this, but I've thought in the past like, why can't you just use the embeddings of a language model as a way to stratify things? Is there like some pathology there as to why that actually... like the issue you pointed out with domains, is there some fundamental issue with relying on embeddings?

Sergey: Exactly. So it turns out it's because of domains. Okay. So if you take the embedding and you average it, you've essentially already lost the information about what the order of the domains are. Okay. Yeah. In some ways you could say that's a good thing. Like maybe you don't want the orders, like if you average you got rid of the order. But the thing is, sometimes they would have a scenario where in one organism the domains might actually be in a different protein. Like there could be three proteins coming together with different arrangements of domains. In another organism it could be one protein with all those domains stitched together. Like all the domains are fused in one protein. And so if you average, all those things will look different from each other. It's almost like you want to first take the original embedding and chop it down and then compare all the different chopped segments to each other. But as soon as you start to do that, then it becomes... then you essentially are back to aligning sequences. And if you start aligning, then if they're in different orders, you can't really align them anymore because you can't use dynamic programming. And so it creates a bit of a challenge there.

Abhi: Yeah. Like the way you have set up that dilemma feels unsolvable. Like what do you think researchers should be doing?

Sergey: Well, like in our example, for example, we had a recent paper with Nick Polizzi, where we had this paper AF2Bind. And so there, what we did was say, you know what? Let's make sure there's no overlap on the Pfam level, like the protein family level based on sequence HMM comparisons. Let's make sure there's no structural overlap, but also let's cut out the actual binding site and make sure there's no overlap in the binding site. And so we went through levels and levels and levels of trying to sort of say, make sure, make sure there's absolutely no similarity between these two things. And so then we were more safe to say, okay, these are probably not... but once you start doing that, like for example for protein-small molecule interactions, I think we found total, there's like only 500 proteins at the end of the day.

Abhi: Mm-hmm.

Sergey: Like independent samples that don't look like each other anymore.

Abhi: In terms of the binding site?

Sergey: Yeah. Yeah. Well, so it's like if you start to go through, say let's cluster based on sequence, structure, binding site, turns out there's not that many examples anymore.

Abhi: Gotcha.

Sergey: But of course you don't have to remove the data. Like you could still keep the members of that cluster within your training, but you quickly realize there's not as much data as you used to think there is.

Abhi: Do you think this stratification problem also pops up with DNA language models? I didn't actually look too closely as to how the Evo team stratified things and it's kind of unclear how DNA language models work in this capacity.

Sergey: Yeah. I think for language models, people are not as concerned in terms of train-test split. But what I mean by that is like, there are concerns. Like you definitely don't want just to memorize all the sequences. So you want to definitely remove some sequences to make sure you didn't memorize. But in... I guess if we come back and think about language models just storing protein families and their evolutionary statistics, then in some ways you don't actually want to remove sequences from an entire protein family. You want to remove just enough to confirm that you're not just memorizing sequences, but you don't want to remove enough that you sort of obliterate an entire protein family from the training.

Abhi: Well, the hope is that there's like transfer to like an unknown protein family, I guess this is the whole ambiguous versus non-ambiguous thing.

Sergey: Yes, exactly. So like, I think you're right. So if you want to make the argument that the protein language model has learned sort of a new space, then yes there you want to actually make sure you remove anything that's remotely similar to it. But if you say, Hey, I just want a language model that sort of stored all the information and statistics of all the protein families, then in some ways you don't actually want to remove too much information. Like you kind of want it to see sequences from every single protein family because you want to learn an embedding of all these protein families. So I guess it depends on like what you're using the model, what claim you want to make for those models.

Abhi: And I imagine if I'm to infer your viewpoint correctly, is it that... do you think trying to train a model such that it'll be able to divine like a new protein space entirely, like a new protein family entirely isn't actually all that fundamentally useful because most of the known protein families are in the space of things we actually care about?

Sergey: I mean, I would say it would be very useful. Okay. I mean, ultimately the goal is to be able to have a model that can generalize to all proteins. And so I think one example of plots that we like to make and other folks too is like the number of sequences versus the performance of the model, by number of sequences meaning how many sequences are there in the protein family? And currently, like most protein language models, I think all of them actually have like a curve this way. Right? So it's like if that family doesn't have that many sequences, the model has very poor performance. I mean, eventually we'd want the model just to be... this line to be flat, it shouldn't matter how many sequences there are in that family.

And one could imagine keeping like, I guess you could say, hiding some sequences or one could just alternatively just keep evaluating if that curve is shifting. Yeah. Like if that curve starts becomes something flat, then you're like, okay, it's learned something interesting, fundamental.

But this also sort of brings up the question sort of a fundamental difference in how sometimes people in the computer science world think about problems versus biologists. I think sometimes when I talk to my computer science colleagues and I tell them, Hey, we want to predict things out of distribution. And I think they get a little bit like, what? It's like that sort of goes against the fundamentals of machine learning. You want to learn things within distribution and you want to be able to sample things within distribution. And the thing is, in biology, often we want the out-of-distribution things. And so like how do you get a model to go out of distribution? And I think the things that I find recently more exciting is like, well, things where you could iteratively reason over things might be a way to sort of move into that space. But...

Abhi: Going back to this idea of many different seeds and like seeing where on the fitness landscape the model places you?

Sergey: Yep, yep, yep. Exactly. And I think there's been, in other fields, for example, people are now introducing like chain of thought and so on, where maybe now if you have a model that says, Hey, return something in one go, or one shot or zero shot, it sort of goes through and makes the best guess it could. But if you sort of iterate on that, maybe you have it be able to explore outside of its training. And in some ways that you can think of that, that's what essentially Alphafold is doing. Like Alphafold will say, I'm going to make a guess, that's like zero recycles, and then you iterate and you sort of move around. But maybe if you do many, many independent seeds. And I think that's actually what some of these models like o1 and o3 are doing, like they have many, many independent starting points and they explore. And so I think in some ways, I guess we could say we've been already doing that for a while in the protein world. And they're kind of catching up. But I guess one could also say, well, maybe some of the techniques learned in that space could be applied now in the protein world as well.

Abhi: Do you think there's that much utility in integrating... like instead of having... actually more straightforwardly, do you think there's useful information in integrating wet lab assay data when... I guess like one example of this is introducing binding affinity into Alphafold 3 instead of just having like two protein structures together, you also say like, oh, this is the KD that came alongside it. Do you think there's that much value in doing something like that, or empirically it doesn't turn out to actually be all that important?

Sergey: I think in practice it could be useful because right now the model, like for example, methods like Alphafold, assume anything you give it is a good thing. Like it's a real thing, it's a real protein. But turns out in some organisms, maybe you want it to have higher affinity. In some organism you want less affinity. Like if you're in a hotter environment or colder environment, maybe you want the protein to be less or more stable. And so right now the model just assumes everything you give it is a good thing. And so it sort of learns some interesting average of all that stuff, but being able to sort of tune that label during prediction or something or during design, could be actually useful.

Abhi: Do you think there's anyone trying to integrate almost like human labels? Like because alongside a lot of these protein structures that are deposited in the PDB, there are some semantic labels that go alongside it of like, this is a particularly well-resolved structure or like this residue was hard to resolve. Do you think integrating that sort of information, like bringing in semantic level information is actually all that useful? Or is it just kinda like so far beyond what anybody else in the field is focusing right now that it's not worth poking at too much?

Sergey: I mean, I think it's definitely an exciting direction to look into. I don't know if anyone's tried that yet. I mean, people have tried filtering on that labels. Like, for example, I guess the most recent example I can think of that was really cool is the Soluble-MPNN paper, where folks say, Hey, let's retrain ProteinMPNN, but only on soluble proteins. I guess a newer version of that is say, well, what if you just provide a label? Say, is this protein soluble or non-soluble? Yeah. Like right now, the model sort of implicitly... like the default model, ProteinMPNN, even Alphafold by default sort of sees that sequence and maybe tries to infer if it's soluble, non-soluble. But now if you retrain on that kind of label, or introduce this label, you could essentially now tune and say, Hey, this is a protein in the membrane, and maybe you should be folding it differently or filtering the constraints differently. So I think there's already been demonstration of this in the context of Soluble-MPNN, this is Justas and Bruno Correia and so... Yeah. Yeah.

Abhi: On the subject of Soluble-MPNN. One question I realized I should have asked much earlier is what are your thoughts about sequence-structure co-design versus like one step structure, second step sequence?

Sergey: We recently... so Yilun in my group, we recently put up a preprint where we sort of explore this a little bit, where we co-design both at the same time. And our current hypothesis is that there is actually a benefit to co-design from the perspective that, like for example, let's say if you start with a structure and you come up with a sequence for it, maybe there's a better structure that can encode even a better sequence. And so having ability to move the structure closer to the sequence space, and then finding a better sequence, there's some benefit to that. Because maybe there's... like it's possible for this fold that you could find even a better free energy sequence. But unless you see that structure, you won't be able to come up with a sequence. So I think there's benefit in that space.

But it doesn't necessarily solve the problem of does the sequence fold into that structure and no other structure. So you still need to have the, I guess you could say the structural prediction module in place.

So I guess there's been some work by folks like Jason Yim at MIT where he looked at say, Hey, could we co-design things? But I feel like when you want to co-design things, what you really want to do is you want to co-design sequence, structure, and folding. Yeah. Because the way I like to think about it is when you're designing a protein, you're not designing a sequence or structure. You're designing essentially a folding landscape. Yeah. And one of the problems... so it's not about co-design of sequence and structure, it's co-design, or I guess you maybe tri-design, I don't know if you want to say it that way, of sequence, structure, and folding. And it's kind of hard to think about how models like diffusion could incorporate folding because they're considering that one structure, but they're not considering the ways that protein would fold into that structure, or how a protein sequence would fold into that structure during the process. And it's not super clear to me how you would actually incorporate that at the moment.

Abhi: Like, as opposed to pure structure models?

Sergey: Well, I guess there have been approaches where people say, well, let's diffuse the sequence and structure at the same time. And then the question is like, are you both satisfying the folding and inverse folding? And so there I would argue like, no, you're not satisfying inverse folding. Because inverse folding requires that you come up with a sequence that only folds into one structure and no other structure. And so, unless in your diffusion, you're somehow accounting for all the other ways the sequence can fold, you're not actually optimizing for the inverse folding problem.

Abhi: I guess ProteinMPNN is trying to satisfy that.

Sergey: Well, this is what I would argue. It's not really yet. Okay. Uh, well, so I think there's been... unfortunately in the field, people have been using the word inverse folding incorrectly because formally inverse folding, previously, and I think folks like Kendall have defined it as finding a sequence that folds into one structure and no other structure, but also at the same time making sure it's actually accessible.

Abhi: Like, thermodynamically accessible?

Sergey: Or I guess I was thinking more on the kinetic side, but what I guess I was... like during folding, you want to make sure it folds. Like, for example, there's not a huge barrier. Yeah. Unless you have a chaperone. Maybe that's one way to lower the barrier. And so that's what formally inverse folding was defined as. I think unfortunately folks in the computer science world when they're like, Hey, let's develop a method that takes a structure, returns a sequence. We're just going to call it inverse folding. And I mean, it is a nice term, but unfortunately it's not actually the real inverse folding from what the term was originally defined. And so I think in the ProteinMPNN paper, they were pretty careful not to call it inverse folding, but I think other folks who've developed similar methods have called it inverse folding. And so I think on, at least on Twitter or X, I always, every time somebody says inverse folding, I try to correct them and say, Hey, this is not really inverse folding. You're not actually satisfying the inverse folding question, I guess you could say.

Abhi: I guess though, but if you use that definition, no structure-to-sequence or sequence-structure co-design method is satisfying that metric.

Sergey: Well, this is what I would argue: inverting a structure prediction model is sort of maybe implicitly trying to satisfy that condition. So like you say, I have a method that takes a sequence and predicts a structure, and if I invert that model, then as you're changing the sequence, you're essentially every step along the way implicitly checking, are you satisfying that condition?

Abhi: Gotcha.

Sergey: And so I'd say that is closer to inverse folding than a method that takes a structure and tries to predict a sequence from that.

Abhi: Gotcha. Um. And so one thing I am still a little bit unclear about is why do you think that structure-sequence co-design models are not able to do this?

Sergey: Well, because they're not at any point evaluating along the way if that sequence won't fold into something else.

Abhi: Gotcha. Whereas the sequence-to-structure model is during a training process doing that. And then if you invert it, you hope the one-to-one relationship stays the same kind of?

Sergey: Exactly. Exactly. Exactly. Gotcha. Okay. Now, I mean, people who actually study protein folding might tell you again, no, that's not true because the model doesn't explicitly fold proteins. Yeah. And so maybe you do actually need a model that does protein folding first to be able to do that. But we think it is sort of implicitly maybe pushing in that direction.

[01:49:07] What Sergey would do with $100 million

Abhi: One of the last questions I had is, you've discussed a lot on Twitter about recent funding cuts to academia in general and specifically concerns about your own lab. One, I kind of want to talk about the best case scenario of like, let's say you did have a few hundred million dollars you could spend on any type of basic or applied research you wanted. What would you want to work on?

Sergey: That's a good question. I think for me, if I had lots, lots of money, I always try to think if there are some kind of experiments we could do, like wet lab experiments to collect information about protein folding. I think that's one of the things, like right now we have absolutely no experiment... I mean, there have been people like... I'm trying to remember his name. It's a guy in Canada... it's not Lewis Kay, it's like K. Lewis or something. Okay. Maybe I'll start with that question again. So, so I think... let me think of a good way to explain this.

Abhi: When you say measure protein folding, is that like NMR level, like the actual process of the protein folding? Yeah. Okay. And you would want to just collect thousands of these results?

Sergey: Yes, yes. Okay. So, I think one of the problems right now is: let's say we... okay, so let's just say we step back and say we want to solve the actual protein folding problem. And by protein folding problem, what I mean is starting from an extended chain, find the right conformation of the structure, step by step. And actually thinking about all the steps required to get there. One of the issues is that we don't actually have any truth. Like we can't really train such a model.

I mean, there were experiments where people look at transition states. That's the researcher, I was trying to remember the name of, Lewis Kay, I think his name is. But that just gives you, for just like one or two proteins, I think they did some NMR transitions that they've measured there that are sort of what they believe to be intermediate folding states. And so I think if there was lots of money, I am wondering if there are ways that we could develop experiments that could actually measure large scale, about how different proteins fold.

And I've been sort of thinking about how we would go about doing this and it's a little tricky because it's like you have to kind of measure it on an individual molecule level to some extent. Yeah. Because it turns out every protein will fold differently. So I don't know if there's a one particular way. Maybe if there are some ways to synchronize it. Like, for example, people have done these molecular tweezer experiments. And so in some ways you can unfold and refold the protein by pulling on the N and C terminus. If there's some way we could image those things.

Abhi: Like just a line of amino acids to the final state.

Sergey: Yep. Yep. Yeah. I mean, the other problem is also pretty exciting. Like, hey, how do you extract all the different dynamics a protein could have, or maybe all the different conformations it could fold into. But I guess what I'm saying is like, it would be great if we could somehow actually get step-by-step snapshots of the structure folding. And so if I had money, I think I'd throw it all on that.

Abhi: In some sense, is that not dynamics?

Sergey: It is, it is. It is dynamics. Yeah. Yeah. Yeah. But usually it just happens once and it stays there. Yeah. And so then with NMR, you're kind of stuck watching that one structure vibrating and moving around. But the question is like, could we sort of take snapshots of it going from an extended down to the final structure?

Abhi: Do you think the... like let's say we did have a perfect NNP that was capable of taking this string of amino acids and seeing femtosecond to femtosecond how does it fold into a final structure? Yeah. Would you still want this NMR dataset? Is there something useful about measuring it in the real physical world that you cannot get from a simulation?

Sergey: I mean, if we do have a good simulator that could actually do the whole process. I mean, I guess I was thinking more like could we generate data to be able to train a better simulator to be able to do this? Gotcha. Okay. Because I guess I'm under the impression that the current, even if you had infinite compute time and MD trajectory might never have found the solution just because there are sort of inaccuracies in the energy function and so on. But if we assume that everything's correct in MD and then you're right. I mean, if we have infinite compute, we just do that.

Abhi: Yeah. But like, yeah, I think empirically you are correct in that they are nowhere near that place of being able to go from...

Sergey: I mean there, I think D. E. Shaw showed they could do like really, really small domains and they could get them to fold up, but that's... we're quite far from that. I mean, I think we could approximate a lot of this with neural networks, but we sort of need some of these intermediate steps that we could train on. And so I think that's... but the question is like maybe why would you even need this? Like if we already have methods that can predict a lot of protein structures' end states? Yeah. Like the end state. Right, right. But I guess I'm under the impression in order to be able to predict some of these really, really complicated protein structures, it would be good to be able to see some of those intermediates because maybe you actually want to push the model to explore those step by step.

Abhi: Returning back to the NMR research approach, are you currently pursuing any attempts to really scale up wet lab efforts in your own lab? Or are you largely right now entirely focused on computational work?

Sergey: I guess just to clarify one point, I wasn't saying NMR was necessarily the solution to the problem. I was just saying this is one area that maybe be able to get there. But besides that, so we're not trying to do NMR right now. Hannah Wayment-Steele is actually starting a lab where they're going all in on the NMR stuff. So she'd be a cool person to talk to at some point. But at least in my group, we've always thought of it more... I've always thought of it as more of a sort of theoretical computational group. But that being said, we have started some efforts in the space of sort of building up maybe some robots that can do experiments for us. Like maybe some automated lab kind of things. But this is just more the exploratory phase at the moment. So, but primarily we're a computational group and we collaborate with people. Yeah.

Abhi: Do you think there's that much value... you're really interested in the pathway from start to end protein structure. Do you think there's that much further value in mass collecting end structure data? Like further... let's say we double the PDB overnight. Do you think it would actually make these models all that much better?

Sergey: I think they could. So, for example, we have, I mean we have a lot of structures for prokaryotic organisms, and we can now predict a lot of prokaryotic organism proteins. And because we have lots and lots of sequences, we can get coevolution from them. But I think one of the challenges has been for eukaryotic organisms where you have multiple domains and we still don't quite know how those domains come together. And often when you make a multiple sequence alignment, they would cover one domain but not the other domain or separately cover each domain. And so then when you give these to Alphafold, it sort of arbitrarily places these domains in 3D space. And so these sort of large multi-domain proteins is areas where we actually have no information how they come together.

Abhi: Gotcha.

Sergey: And so I think if we... but besides that, there are also a lot of proteins that are not multi-domain, but only have a few sequences. And there's sort of debate like, are these mostly disordered and that's why we don't have that many sequences? Or maybe they actually are structured or maybe they're only structured upon binding another protein? So that, I think, I think there's still a lot of proteins that we, it would be useful to have, especially for this problem of trying to go from single sequence to structure where we have no evolutionary information. And so the more we can collect of those, the more we can sort of create data sets to be able to maybe start venturing into this problem.

Abhi: So if you have the choice between just randomly doubling the PDB overnight versus doubling it in the areas where proteins are either... protein structures are either very large or they're suspected to be disordered. You'd prefer it in that latter category?

Sergey: Uh, I mean, I guess I said both, but I think for me personally, I would say, I mean, for me, the more exciting part is regions that currently have really low pLDDT in Alphafold that might actually be ordered. Yeah. Like we think they're disordered, maybe because there's a lot of correlation between pLDDT and disorder, but I think a lot of that low pLDDT is not because of disorder, but because just this lack of MSA information. And so I think what would be... if someone had infinite money to give us, it'd be kind of great to say let's go ahead and try to solve all those structures that we think are actually ordered, but have really low pLDDT.

Abhi: Gotcha.

And I mean, in some ways people used to do that back in the day with like structural consortiums or say, Hey, if there are no homologs, but I think now we can use Alphafold to quickly tell us, okay, what problems are worth maybe putting efforts into solving.

Abhi: Triage and decide these are the structures I actually want, and everything else is information that the model already kind of implicitly knows.

Sergey: Yeah. I guess in some ways it's almost like active learning, I guess you'd say. Yeah. Predict everything in UniProt, look for everything that Alphafold failed on. Those are the proteins we should be trying to predict the structures for. Yeah. But that being said, I think we just have to be careful because there are cases where the pLDDT might appear to be high. But if you start to look at the PAE matrices, the predicted alignment errors, there might be no information how the domains come together. And so even cases where we think we know all the domains, it'll be great to know how they come together. And I think those are also useful problems.

Abhi: This is one thing I've been... again, a little bit of a deviation from what we've been talking about. One thing I've long been confused about is there's this implicit trust in what Alphafold2 is confident about and what it's not confident about, or what it claims to be. And it feels like people don't actually seem to run into adversarial optimization all that often. Why is that? It feels like in almost every other field, if you try to really trust the model in one specific metric and just keep doing that over and over again, you eventually run into adversarial edge cases, but that doesn't seem to happen with Alphafold2. Is there a strong reason you suspect why?

Sergey: Um, I think it depends on the problem. So, for example, there have been people who have been showing that, for example, if you try to predict protein-peptide interactions, you get a lot of false signal in terms of you can have really, really high PAE at the interface, but the prediction is actually completely wrong. Like you have really high... and but I think it depends how you got there. And what I mean by how you got there is in some cases, like for... one thing we saw recently in some of our work is if you try to score things with Alphafold, and you use a bunch of recycles, it sometimes gets confident about things that are completely wrong. But if you dial down the recycle, like go to recycle zero, you actually see that it's actually a bit more correlated with known measurements because there's the question is: did Alphafold think they interact because there's some evolutionary information or maybe some motif matching or just it sort of reinforced itself and now it's really, really confident? Yeah. And so... I mean, it's true that a lot of times the confidences are pretty good. But I think the reason why people do trust it is because often they're not subjecting Alphafold to sort of unknown problems. So what I mean by that is often when people say Alphafold predicts something, they already know these two things interact or they know this protein folds into a structure and they want to know what does it actually look like? How do they interact? Where it fails is when you don't know if they interact and you're asking the question like, Hey, do these two things interact? And sometimes it will give you the false impression that they do interact.

Abhi: And...

Sergey: And I think part of the reason why this happens is because Alphafold was only trained on positive data.

Abhi: I was about to mention this. Yeah, what does negative data look like in this case?

Sergey: Well, so negative data would be like, let's say if you purposely take, hey, two proteins that should not interact, can you fine tune methods like Alphafold or RosettaFold? There's a researcher, Qian Cong, at UT Southwestern. Um, that has been actually exploring this a bit. I think they had a preprint recently where they say, Hey, let's purposely mispair things. And then tell Alphafold that these should not interact. Yeah. And she saw some... her group saw some success in that area, where they could potentially try to now do a better job at picking up what does and doesn't interact.

And I think something like that could also work in the design space where right now, any sequence you give Alphafold, the assumption is that this thing will fold into a protein. Yeah. And so even if the sequence is very suboptimal, it will find a way to sort of internally fix itself, which is great for structure prediction. Like you want a model to sort of maybe only see a few key residues and make the right prediction. But for a design problem, you actually want it to be sensitive to point mutations. Like you want to say, Hey, you put a, I don't know, hydrophilic in the core, it should not be predicted well. But it kind of assumes everything you give it is actually a valid answer.

Abhi: Yeah. That's a cool idea. Has anyone shown that... like you said this one UT Southwestern professor... it feels like it hasn't really percolated to the field at large. Do you think long term there will be like positive sets and negative sets for structured data or it's unlikely that that will be the case?

Sergey: Um, I mean, I would assume that people are going to be doing this more and more. I mean, I think for structure prediction it doesn't seem necessary because everything is positive. That's right. But I think for example, for people that do care about PPI prediction, like what's interacting with what... I mean this is why, part of the reason Qian Cong was doing this because she actually wanted to know, hey, which protein interacts with which protein? In which case she saw a lot of false positives from just looking at the confidence metrics. And so she tried to fix that problem there.

But I would imagine in other cases where maybe you want to say which ordered or disordered region actually folds into a protein. There might... like, for example, people do see, for example, with Alphafold 3, you give it regions that are completely disordered and it will still predict a bunch of helices. And there that's where some negative training could be useful. And to some extent they did that. Like the DeepMind team did do that. They actually took Alphafold2 predictions and used that for training to say, repeat the same extended sheet, extended loops and so on. But it turns out there's some kind of mode thing where if you make the protein too large, it sort of snaps back into making helices everywhere. And so that didn't completely work out the way they were hoping.

Abhi: The... I forget... there was a really good reason why they needed to use Alphafold2 predictions to bootstrap Alphafold 3.

Sergey: Well, the issue was that... so the problem with diffusion is that, or at least training with diffusion is that it sort of learns some distribution about how things tend to interact with each other. And it's only trained on ordered things. And so its prior is to make everything ordered. And so one of the things was like, if you start to use...

Abhi: It is unable to be not confident about anything?

Sergey: And it's still not confident. Like if you look at the confidence metric, it's still not confident. But I think the problem was that a lot of researchers sometimes don't even look at the confidence. They look at the structure and say, Hey, if it's a giant long disordered loop, then it's probably not confident. But the problem with Alphafold 3 was, at least before they did this sort of trying to correct it with Alphafold2 predictions, every single prediction always had helices. Yeah. Like anything you give like completely random sequence, it'll always put secondary structure. And so then they're like, okay, how do we fix that? Well, maybe we could try to get it to reconstruct how Alphafold2 used to do, like if it was not confident, it sort of extended out into a loop. Yeah. And so they started training on Alphafold2 predictions to try to get around that problem. Um, it seemed a little hacky, but it seemed like...

Abhi: It feels like a bizarre design. Surely there has to be a better way to do it than just bootstrapping from...

Sergey: I'm not sure how you would do it. So I think the problem is diffusion. The problem is you need it to give you the wrong answer to tell it it's doing something wrong. Yeah. But the problem is it's always... the last step is always the right answer. And so I think with like structure module, you can have it have, I don't know, the loop get clashing with everything and you add like a clash loss.

Abhi: Yeah.

Sergey: But like how do you add a clash loss in diffusion?

Abhi: That's a fair point. Yeah. Yeah. I guess having this physics-based oracle would help things a lot.

Sergey: Yeah, yeah, yeah. Yeah. Well, it's like, it's almost like you want it to move into... so I think this is why they had to put in the Alphafold2 structures in, because there, it's like you train it, say, Hey, if there's no information, extend it. And it worked to some extent, but it turns out if you make the protein too large, it just starts making helices again everywhere. Yeah. So it's sort of learned to do that for small proteins, but not for very, very large proteins.

Abhi: That makes sense. Yeah. And I think I am largely out of questions. Was there anything else you wanted to talk about?

Sergey: Uh, I think we're probably good.

Abhi: Okay, cool. Thank you for coming onto the show, Sergey.

Sergey: Of course. No problem. Glad I could chat.

Note: if you’re a stickler for video quality, watch this on Youtube instead! The max video size for Substack is 20 gigabytes, and this video, sadly, came out to 20.5G. I had to mess with alternative video codecs (H.264 → H.265) to get it smaller. This shouldn’t have strongly noticeable differences, but maybe you have a keener eye than I do. The original quality video is on Youtube!

Youtube

Apple Podcasts

Spotify

Twitter thread

Introduction

There's a lot of discussion these days on how China's biotech market is on track to bypass the US's. I wondered: shouldn't we have observed the exact same phenomenon with India? It has seemingly all the same ingredients: low cost of labor, smart people, and a massive internal market.

Yet, the Indian biotech research scene is nearly nonexistent. Why is that?

To figure it out, I had a two-hour discussion with Soham Sankaran, the founder and CEO of PopVax, an mRNA vaccine development startup. Amongst those in the know, Soham is well understood as one of the most talented biotech founders in India, and his company has had a genuinely incredible underdog success story. This story is still being written, but there's good reason to be bullish, given that PopVax has an (in mouse) influenza vaccine that is 250x better than its competitors, multiple large research collaborations, and their first upcoming US based phase 1 clinical trial being fully sponsored and conducted by the NIH.

We discuss so many things. Including policy prescriptions for Indian R&D, why PopVax's vaccines are so good, how machine-learning is changing vaccine development, and much more. Transcript below, and links in thread (including a jargon explanation).

Timestamps and transcripts are below. Just as in my last episode, I’ve included a ‘jargon explanation’ as a quick primer for some of the subjects discussed in the episode.

Some final bits: the studio rental costs were kindly covered by Dylan Reid at Zetta Partners! Huge shout-out to him for making this episode possible. Also shout-out to Samarth Jajoo, Reha Mathur, and David Yang for some very helpful discussion about the Indian biotech scene. And, if you think PopVax is interesting, here is their Substack which has some articles on their results, their job section (they are actively hiring), and can be reached at contact@popvax.com.

Jargon explanation

Antigen versus immunogen: This is explained again in the podcast, but, concretely: the antigen target you want antibodies to recognize, like a viral protein or bacterial toxin. An immunogen is what you actually put in the vaccine to generate those antibodies. Just because something can be bound by antibodies (antigen) doesn't mean it will generate good antibodies when used in a vaccine (immunogen). This can be for a lot of reasons, which is discussed in the episode.

Epitope: The specific part of an antigen (or immunogen) that an antibody recognizes and binds to.

Antibody elicitation is the process of getting your immune system to produce specific antibodies. When you give someone a vaccine, you want it to "elicit" (cause the production of) antibodies that can recognize and fight off the target pathogen. The challenging part, as Soham explains in the interview, is that just because something can bind to an antibody doesn't mean it will cause your body to make that antibody when used as a vaccine.

This is why vaccine design is so tricky. You need to carefully design your vaccine to elicit the right kinds of antibodies. Remember, your adaptive immune system will undergo affinity maturation to adapt to whatever immunogen it sees, it won’t simply produce something that binds well! Your immunogen is simply a guiding force in this stochastic, antibody evolution process, and it must lead that process down the right pathway for good antibodies to be bad.

Effector function refers to how antibodies actually fight pathogens beyond just binding to them. In antibody structure lingo, it specifically refers to the the ‘Fc’ part of this diagram. It is, for the most part, constant and doesn’t interact with the pathogen itself, but rather the rest of the immune system.

How does this effect immune design? You need the antibody elicitation caused by your chosen immunogen to not only bind well to the immunogen, but also, one, be a type of antibody that has a ‘desirable’ Fc region (e.g. IgG1, IgG2, IgG3, IgG4 have different Fc regions), and two, have the antibodies cluster around the immunogen densely enough to activate the immune system.

A good case of not taking effector function into account is mentioned by Soham: HSV-2, where vaccines focused only on neutralization failed because effector functions turned out to be crucial for protection.

Timestamps

01:31 Introduction

02:38 Why is there such little biotech research in India?

17:43 Advantages of building a company in India

31:30 Policy prescriptions for India

35:39 Questions on vaccine design

50:55 What does PopVax do?

01:01:58 The role of machine learning in vaccine design

01:12:07 The (conservative) culture of vaccinology

01:26:57 Hiring in India

01:46:52 How fundraising for an Indian vaccine design startup is coming along

01:57:36 How is PopVax so good at designing vaccines?

02:02:07 Pet theories on immune mechanisms

02:09:07 mRNA beyond infectious diseases

02:12:38 What would you do with $100 million dollars?

Transcript

[00:00:00] Preview

[00:00:00] Soham: I think fundamentally, there's a problem of cowardice in India, right? In our elites. Our elites talk a big game. Increasingly, they talk a big geopolitical game as well. But when it comes to actually doing difficult and risky things, we tend to shrink away from it

I think it's an important distinction because there's lots of immunologists.

Not that many vaccinologists, right? People actually practically working on the problem of better vaccines, fewer than you would think. Lots of, like, people working on, like, basic science immunology, systems immunology, increasingly. People actually working on practical, like, nuts and bolts vaccine design, not so many.

We can find parts of the design space that maybe haven't been tested before by other people that in combination you get these really good results.

[00:00:42] Abhi: Because other people don't have like, don't have like the knobs to tune on those because they're outsourcing it all?

[00:00:47] Soham: Yes, they don't have the knobs to tune and they don't have the wherewithal to test all this different stuff.

[00:00:50] Abhi: Gotcha.

[00:00:51] Soham: Right? And I have all the knobs, right? All the knobs are in my control. And so, and we find lots of wacky stuff, right? We find that there are certain lipid nanoparticle formulations that work better for certain vaccine designs. Mm hmm. Do I know why? No. Is my team trying to figure out why? Yeah, totally.

And that's what we're doing at PopVax to some extent. It's 10 times cheaper for us to do a lot of the wet lab work. It's 10 times cheaper for us to do animal work.

And so that means for the same dollar, I can do either 10 programs or I can potentially go 10 times as fast on the same program, right? And eventually, we will just leapfrog what a company like Moderna is able to do because, you know, they can't keep losing the billions of dollars that they're losing forever.

[00:01:31] Introduction

---

[00:01:31] Abhi: I'm incredibly excited to talk with Soham Sankaran, the founder and CEO of Popvax, an mRNA vaccine development startup that he has ran for the last three and a half years. Soham has an interesting background because it's not at all in biology. He has a bachelor's in computer science from Yale, got partway through a PhD in robotics from Cornell, dropped out to start a software for robotics company, and then decided to pivot to biotech during the height of the COVID pandemic.

Most interesting of all is that unlike almost every other R&D biotech company today, PopVax is not based in the U.S.. Or Europe or East Asia, it is based in India, which as any biotech venture capitalist would tell you is an extremely unlikely place for novel biotech research to be conducted. Yet the company is incredibly successful, having won multiple large research contracts and having their first upcoming U.S. based clinical trial fully conducted and sponsored by the NIH. Congratulations.

[00:02:23] Soham: Thank you.

[00:02:24] Abhi: Today we'll be discussing what has stymied biology research in India, how machine learning is changing vaccine design and questions over PopVax's recent success in vaccine development. Thank you for coming on to the show Soham.

[00:02:36] Soham: Very excited. Thank you for having me.

[00:02:38] Why is there such little biotech research in India?

---

[00:02:38] Abhi: So, like, first question is to start things off. you've mentioned in the past that one of the key things preventing further biotech research in India is an inability to be comfortable with technical risk.

Do you think this is cultural? Something that is entirely built into the psyche of India? Is it downstream of a thousand other other things? And then, is it at all fixable?

[00:02:54] Soham: You know about the, the chaebols in, in South Korea?

Yeah. Yeah. Right. And like, you know, the zaibatsu in, in Japan and so on and so forth. Right. They are these sort of big conglomerates. They do a bunch of different things, right? They're like Samsung's and so on and so forth. Yes, as you said. But, what's, I think, particularly interesting about them is like, they're somewhat similar to businesses that we have in India, right?

So the Tata group that you may be familiar with, right? Reliance, which is also a big conglomerate that started in, textiles, then went into petroleum and now does a bunch of all other things, including, they're the biggest telecom operator in India with Jio and so on and so forth. And there are some emerging groups like the Adani group in India, that are also trying to do a whole bunch of different businesses, right?

So there's some parallels here. Right. And I think the government in India has pushed a strategy of these sort of national champion conglomerates, which are, you know, have, you know, government sanction in some sense to, to do risky things. Right. And just sort of build infrastructure and so on and so forth and, you know, build the country.

Right. And I think that that parallels explicitly the strategy that, that South Korea took, that Japan took, and, you know, to some extent Taiwan as well. Right. The distinction, and this is a very critical distinction, is that those companies, and increasingly in China as well, have invested a lot of money in R&D, right?

So they started by doing more copycat type things, you might say, right? Volume manufacturing or sort of, you know, infrastructure projects that are similar to what happened in the West. But then, you know, companies like Samsung have like a big chemical businesses that do specialty chemicals. They have semiconductor businesses where they're the cutting edge of, you know, memory or the cutting edge of other semiconductor processes.

And you see that in China as well. Increasingly, Chinese companies are investing in new R&D and they're the cutting edge of battery technology and, you know, electric cars and so on and so forth in India. These conglomerates, which have been around for a very long time, and also, you know, historically, predate many of, you know, the expansion of many of these East Asian businesses, right?

Like the Tata group has been around since Pre independence, it was huge pre independence, right? haven't invested in new technologies in the same way. and they haven't sort of done, they haven't done, this long term R&D that you would need in order to be at the forefront of some technology globally.

Right? And as a result of that, those companies largely sell domestically within the Indian market, right? Whereas these other conglomerates that I'm talking about in East Asia, they sell globally, right? They export their cars, they export their batteries, their semiconductors, and so forth. So I think that is, you know, one huge problem that we have in India, which is that these big companies that have the resources, billions of dollars in cash, don't invest in R&D, right?

And yes, I think that is a broader cultural problem. That's not specific just to those businesses. The government doesn't really invest in R&D in India. We invest a tiny, tiny fraction of our GDP, you know, compared to not just rich countries, but also developing country competitors. there's just, you know, fewer dollars available for whether it's bio R&D or other R&D, right?

And then I think there's beyond that an unwillingness even among venture capitalists, you know, who should be taking sort of technical risks in India to invest in businesses that are technically risky or to underwrite technical risk in part because they don't have experts that understand it. Right.

Almost no venture fund in India, like a, you know, sort of generalist venture fund has even like a biotech partner or biotech arm, for example, right, where they can evaluate biotech businesses. And that's just one example, right? There are all these other deep tech, hot tech businesses where they're sort of not really able to evaluate.

[00:06:21] Abhi: I remember you had one of the, like a crazy line and, like a document I read of yours where like some guy closely associated with the Serum Institute of India has invested more money into his own Bollywood production company than in any early stage R&D. Yeah,

[00:06:33] Soham: that's maybe a bit hyperbolic, but the guy who owns the Serum Institute of India that runs the Serum Institute of India has invested a bunch of money in, in, a Bollywood production company, like a hundred million dollars very recently.

Karan Johar's production company, he's a very famous director. and, they invest, yeah, not all that much money into early stage discovery R&D. So, you know, it may in fact be more. Is it like

[00:06:57] Abhi: And instinctively, is it fair to say, like, if he took that a hundred million dollars and said, put it into early stage R&D, it would be money well spent, or is there like some other fundamental problem that's going on where like money is not allocated.

Like even if the money is allocated, it will not be used particularly well.

[00:07:13] Soham: Within that company or within...

[00:07:15] Abhi: Within like, like all of R&D and in India,

[00:07:17] Soham: I think there's like more than a hundred million dollars of spare capacity, you know, in terms of things that could be done projects that are almost ready to go, that no one is able to find money for.

So I think, yes, I think a hundred million dollars in R&D would be extremely well spent and I think PopVax is a good example of this, right? We, have spent not. More than, I think, I think it's like, you know, on the order of 15 million dollars, right? Since, since we started three and a half years ago.

Plus minus. I think that investment would have been a good investment for many Indian pharmaceutical companies, right? And we have this Indian pharmaceutical business, right? Companies that export even to the U.S. Or export to Europe. But it's all generics and biosimilars, right? So they, they make copies of drugs that are either off patent or about to go off patent.

And they make them very cheap at high volume and, you know, in some cases very high quality. And this is good for public health and it's good for the world, but they don't invest in R&D. Right? R&D would be the next obvious step, one would think, right? And in East Asian countries, again, it has been the next step, right?

Companies start doing copycat stuff. They start doing generics or biosimilars in biotech. And then they wake up and they go like, oh, we could do, you know, we could do better. We could sort of do something that's, that's novel. And that's what you're seeing in Chinese biotech today, right? Companies haven't done that.

Lots of companies had the 15 million that PopVax has cost so far. Sure, they didn't have me or some of the talent that we have, and maybe they wouldn't have been able to attract exactly the same people, but maybe they would have been able to get people that were, you know, somewhat at the same level, and maybe they would have been able to do so even easier than I was, because I was a nobody who didn't know any biology, right?

But none of them did spend that money in practice, and so they didn't, you know, have the assets that we have and the platform that we have. Yeah.

[00:08:56] Abhi: I, like, like, like America has this America, like uniquely American trait of having like billionaires who are very interested in research and specifically biology research.

Like you have Jeff Bezos funding Altos labs. You have Patrick Collison funding Arc Institute and like Brian Armstrong funding NewLimit. Is there, like, any Indian billionaire who's, like, willing to, like, like, be in, like, play in that same space of funding, like, very high risk, high reward biology research?

And even if there is, do you think, like, their money would be actually, like, well spent in India versus, like, anywhere else?

[00:09:25] Soham: So, Kris Gopalakrishnan, who's one of the Infosys guys, has put a bunch of money into, into neuro research and brain research at IIT Madras, I think, and also in some institutes in Bangalore.

So I think he's maybe one good example of this. And then Azeem Premji, from Wipro runs, is sort of the person behind Azeem Premji University and they do some research and so on. But not, you know, anything close to the scale of what you see in the US. And also, I have talked to billionaires in India who should be interested in funding this kind of research, and without naming any specific names, they've been quite dismissive of the possibility that good research could be done in India at any price, and that's their excuse for not spending the money.

Like, the

[00:10:06] Abhi: Talent doesn't exist here, so why would they spend the money?

[00:10:09] Soham: Which

[00:10:09] Abhi: isn't true, right?

It's not like like, empirically,

[00:10:10] Soham: It's not. Empirically, it's not, right? And then on top of that, they I think, you know, they're sort of a... a minimum level of philanthropy that U.S. Billionaires engage in, even when they're criticized, they generally engage in like some minimum level that's like reasonably high, even if it's not for, you know, out of the goodness of their hearts, it's because they have interest, right?

Indian billionaires, especially ones who are multi generational, right, it's their father's money, or their father's father's money, or so on and so forth, right? I think tend to be uninteresting, and uninterested, and incurious, and so they're not fundamentally interested in funding your research. And they, you know, find convenient excuses not to do it, right?

I think fundamentally, there's a problem of cowardice in India, right? In our elites. Our elites talk a big game. Increasingly, they talk a big geopolitical game as well. But when it comes to actually doing difficult and risky things, we tend to shrink away from it, right? And so the business community in India has I've historically been very interested in finding ways to put up regulatory barriers to prevent access to the Indian market and ways for, you know, their monopoly or monopoly esque businesses to make more and more money off Indian consumers, and essentially no time on how to make globally competitive products that they can export to the rest of the world.

And this is, again, in direct contrast to what you see in many other countries. And I think it comes from the same fundamental problem that makes them also not interested in investing in R&D, you know, just at a fundamental level, they don't, this is not interesting to them, right? They would like to find a way not to have to do this hard thing if they really could avoid it, not have to think about it, right?

And so as long as there's, you know, sort of government support for enacting these barriers that will prevent, you know, foreign businesses from coming in and doing business in India, and they have enough of a market locally, even though I think from a strictly financial perspective, it obviously makes a ton of sense for them to sell stuff to bigger markets, right?

India is big, but like the U.S. is bigger, right? They just don't do it.

[00:12:18] Abhi: I mean, like you mentioned earlier, like, like the chaebols in Korea and like you see of this phenomenon also play out, like in China, Japan, Taiwan of like being very focused on manufacturing and then like pivoting to R&D once they get enough money.

Well, why do you think that didn't happen in India or like, do you think it's like underway and we're just like not seeing the fruits of it yet?

[00:12:36] Soham: Oh,

[00:12:36] Abhi: because they're

[00:12:37] Soham: cowards. Like, it's like, there's a, there's a, it is a cultural problem, right? Like, I think like the, the people, the same sorts of people, like the third, fourth generation Samsung people who are like, oh, let's invest.

Roughly speaking, like, all of our free cash flow into semiconductor, right, 20 years ago, the people in India who are sort of in the same positions did not do that, are not interested in doing that, aren't doing it now, right, for whatever the equivalent is. I'll give you an example of this. Nandan Nilakani, who's not a, a sort of a dynastic billionaire, but who's one of the Infosys billionaires, again, one of the other Infosys billionaires, gave the speech, maybe he said this multiple times.

But he said like, Oh, India shouldn't train its own foundational, like LLMs. So it shouldn't train its own foundation models. it's too expensive. We just shouldn't do it. And we should instead focus on specific applications, like making better call center chatbots.

[00:13:30] Abhi: That's bizarre. That's such a, that's such a like pessimistic view of the world.

Or pessimistic view of the nation that he's in.

[00:13:37] Soham: Exactly. And it's also like, we are a country that knows how to do cutting edge research for less money, but we shouldn't make a nonsensical virtue out of frugality, right? Sometimes you just got to invest the money, right? And DeepSeek just came out with their reasoning model, which they've made for like almost no money, it looks like, right?

[00:13:52] Abhi: And they also said this thing about like,

they were one of the, the, the CEO of whatever hedge fund is behind DeepSeek, like was asked, like, where'd you get the talent to build DeepSeek? And he said something along the lines of like, we did not brain drain like the U.S. We developed the talent in house.

[00:14:08] Soham: It's like a hedge fund, right? It's like a hedge fund that like decided to do

[00:14:11] Abhi: Yeah,

[00:14:11] Soham: so clearly they had smart people

[00:14:12] Abhi: to begin with. But like still, they didn't have access to like the OpenAI like employees or Anthropic employees. They developed it in house.

[00:14:18] Soham: It's like lots of prop funds and hedge funds have big offices in India.

Like Optiver has a huge office in like in Maharashtra, right? And I think in Bombay actually. Like you could totally pull some of the best talent in the country in if you paid them enough money to, build a, you know, new AI research shop, I think, right. If, if that's what you want it to do. And I think there's lots of areas where actually Indian scientists have made substantial strides in the past, right.

That were far ahead of where they should have been given the investment. Like we had really good nuclear reactors for less money. We had really good rockets, like all of this kind of stuff where if ISRO, which is the Indian Space Research Organization, which so pre SpaceX, ISRO had the lowest like per pound launch capacity, of anywhere in the planet, right?

So we should have been like global leaders in commercial space, and ISRO was just heavily under invested in. . So, you know, the program stagnated, and they never worked on reusable or sort of, you know, these, these sort of new innovations that would have reduced the cost of the launches further, and made it more frequent to be able to launch on.

They just got, you know, whooshed past by Elon. But there was no, like, fundamental reason. The rocket scientists were good, and the computational folks were good. Those two things could have been put together. It's just no one thought to do so. And now we have a commercial space industry that's flowering in India a little bit, right?

But it's so far behind what it could have been if people were willing to make those investments when there was still substantial. Like today, if you go, like, you need to make a reusable rocket that can sort of land, right? There's a proof of principle you can do that. And so it's a bit less technically risky.

I think the problem is, when people can't imagine, you know, something for the first time, right? When they haven't seen it actually done, and it's not a copycat, right? It's not just, we'll make it cheaper. The willingness and the wherewithal to underwrite that kind of technical risk, like first in the world products or first in the world R&D exists nowhere in India, right?

And it should, it absolutely should. We are one example. Many other people are, you know, I think shining examples of sort of, you know, even with less money, people are able to do this, but there's this sort of constant bottleneck on the ecosystem of like, you know, the people who have the money saying that there's not the talent right over and over and over again to convince themselves, I think that they're not just being lazy and sort of fearful.

[00:16:31] Abhi: And so there's like a huge bottleneck placed on the people who are actually like, like, courageous enough to think like this is possible. And then they just have barriers put in their way.

[00:16:38] Soham: And then they just leave, right?

[00:16:39] Abhi: Yeah, that's true.

[00:16:40] Soham: They just come to the U.S. And like, suddenly when you get on the plane and you reach here, you're in Boston or San Francisco and you're like, I'm going to do first in the world thing.

Even if it's the exact same team, now you can raise money suddenly, right? Which is unfortunate. And I think, you know, we've been, people, investors have told me, like, oh, your technology is really good. It's, you know, our preclinical data for influenza and for COVID is just about as good, if not better, than what anyone else has, right?

And people have been like, why don't you just move the whole kind of company to Boston, right? Like you're just so much easier to fund. If you, you know, it'd be easier for us to write the check if you just move the whole thing. And it's not because the talent that we have is bad today. It's just because of sort of, that's just what they understand.

[00:17:16] Abhi: Yeah.

[00:17:18] Soham: So I think if I'm a US investor and I don't fully understand that India has the talent, like I can excuse that to some degree. That's something we have to educate them about. If I'm, you know, a rich person in India or a powerful person in India, and I don't see around me the wonderful talent that is being crushed by the system, then I'm blind.

Right. And I think that's, you know, that's what's going on. Unfortunately.

[00:17:38] Advantages of building a company in India

---

[00:17:38] Abhi: On the topic of like why you incorporated in India to begin with instead of the United States, I think it's interesting to consider like the comparative advantages that certain places in the world have for certain types of research.

Like America has like Harvard's and MIT's of the world are able to create the craziest looking antibodies. China has incredibly low, like has access to like looser drug approval guidelines. I think it's like a guilty or like innocent until proven guilty drug approval and, so they can move a little bit faster, with that context, what are some advantages of operating in India specifically?

[00:18:10] Soham: So I think there's a lot of talent in India. it's not well recognized, I think, how much wonderful wet lab talent there is in India and how much wonderful computational talent there is in India. and. The cost of operations, is much lower. So, you know, 5x or 10x lower.

So we can do, we can hire people and we can do wet lab experiments and especially in vivo, like animal experiments. Thousands of animal experiments, and we've tested like a thousand plus novel vaccine constructs in vivo now. Much more cheaply, five or ten times more cheaply than you can in the U.S. or, or you can in, in, you know, in places that have similar economic structure, right? And so I think what that means is, if you look at, pharmaceutical program, right? These programs cost billions and billions of dollars, right? In some cases, a vaccine program, you know, could cost a billion dollars and still fail quite badly, right.

If we could reduce the cost structure of these programs, we could make it possible to test way more things up front then you would be able to basically then you would have the bandwidth to in a typical US pharma context. And then I think the other interesting thing that we can do in India is like what we've done is we have end to end you know all of our concept to clinical production in one roof, right?

And we are able to access relatively quickly talent for all those pieces, right? And so that means we can kind of jointly optimize end to end over the whole pipeline of things we're doing, right from concept to making it possible to produce, right? And that's something that big pharma companies can do in the U.S., but that small startups are often not able to do. And that means they often encounter translational bottlenecks in between. So for example, you know, I know companies, vaccine companies that have waited one, two years to get access to a CDMO, like a contract manufacturer, to be able to go into phase one in the first place.

And then that gap can kill the company, right? Someone will get ahead of you or you just run out of money. We have the ability, because it was relatively cheap to build that facility in India to do small scale manufacturing, to just, you know, be able to do that end to end, not be delayed by what external people are doing.

And, you know, the same thing for analytical operation, for animal house operation, you know, for, you know, for our teams working on, all kinds of, of, of assays that you might typically outsource in, you know, immunology assays or high throughput assays that you might outsource in a in a U.S. pharma context, right? And I think having all that under one roof, being able to scale that up relatively cheaply, and then just being able to hire more and more people, more hands, so to speak, quite reliably in a way that I think you just don't have access to that number of people that can do wet lab work necessarily in any, company in the U.S. because you have so much competition also with other innovative companies that are doing interesting works in the Boston area, right? Gives us the ability to just test way more things way faster be able to advance them into clinic more quickly than you would with the same structure, in a US company. So it's I think it's cost and just amount of talent. There's a number of people available who can do the wet lab work that then sort of you know compound into this broader advantage of being able to test more things

[00:21:11] Abhi: Like end to end.

[00:21:12] Soham: End to end.

Yeah, test them more quickly and then move them into clinic.

[00:21:14] Abhi: If you look at places like, like China or like like emerging biotech markets, like outside the United States, do you see kind of like the same parallel? I'm not super familiar with, drug developers in China. Do you, do you see like the similar, like,

[00:21:28] Soham: I think that's what they're doing in China.

Like,

so what I've heard is, and this is hearsay, but I think this is what is happening. Yeah. You probably heard that like this year it was like, or in 2024, I think 30 percent of assets licensed by big pharma were from China and was previously 50

[00:21:42] Abhi: percent of like oncologic INDs or something like that. And like

[00:21:45] Soham: Previously, it was zero, right?

Big pharma was not buying anything from China. I think what has happened is like some version of like, I'm a professor at Harvard and I read a paper and I'm like, I found this cool new cancer target. I'm going to make an antibody and I go to like, you know, Third Rock Ventures or whatever. And I get, I start a company to do it.

And I'm like, I announced my board of directors and my SAB and like, you know, I get some office space in Cambridge and like about a year has passed. And by this point, the Chinese company that read my paper, that read my preprint has an asset in clinic. Right. and I am screwed. And I think that is what it's like in China because they have the ability to, to effect that pipeline end to end, as we are trying to do as well.

You know, they can take a new target or a new idea and put it into their engine, right? Which is like unfashionable in the U S to have the whole engine right now. You, you have all of these like new, like labless ventures that are being spun up. They have the engine, they plug in the thing. And what I found is maybe this is just me as a non biologist, not being a good manager, but I think it's more general, right?

When you are in the lab with the people who are doing the work, you can spot issues in the data, or you can spot issues in protocol. If not me, my team can spot them and correct them in real time. And that compounds. So if you like go, go to WuXi and you're like, hey man, like do this assay for me. And then a month later you get the readout and you're like, oh, this is all trash.

Which happens, I think, quite frequently, right? That's a month lost, whereas I'm going to figure that out in two days, right? And so I'm going to correct, and I'm going to be suddenly now a month ahead of you, right? I think that's what's happening in China, where they, like, co located with their labs, right?

They understand, you know, in some sense, developability, right? Like, much more cleanly, because they know what's working and what's not, and they're able to, like, discard ideas or molecules that aren't working.

[00:23:25] Abhi: Like ballpark figure like if you had if PopVax had started in america, how much more of a multiple would you need to like run the exact same experiments we're running right now?

[00:23:34] Soham: I think it would have been like between 5 and 10x more money.

[00:23:36] Abhi: Okay.

[00:23:37] Soham: And like maybe now we have a lot of programs now and we're able to do them very very cheaply I think it would be. You know, instead of the company being like single digit million dollars per year to operate because we're 70 people, right?

We're 70, like mostly wet lab scientists, mostly folks with at least a master's, mostly with PhDs, right? And and you know, the requisite amount of, of, of sort of animal work. We do a lot more animal work than most vaccine companies. We have thousands and thousands of, constructs now that we've tested in vivo, and we've used, I think, probably 10,000 plus animals, actually more, more than that.

And so just, just the cage capacity, the animal handling, and sort of, you know, having the team alone I think would balloon the cost up 10x from where we are. And you can see that when we look at other vaccine companies, that, you know, are publicly disclosing their financials that, you know, went public or whatever.

You see that, which is at the sort of late preclinical stage, just before phase one, some of these companies are spending 50 to 100 million dollars a year.

[00:24:38] Abhi: Mm hmm. Yeah. And I think, like, going back, sorry, going back a little bit to, like, the cultural problems in India that have, like, prevented this sort of, like, high risk, high reward R&D work.

The UK is also going through a bit of this recognition of his research stagnation and has formed ARIA to combat it. and for like context for listeners, ARIA is a kind of a parallel to America's DARPA, so a funding vehicle for extremely high risk, high reward scientific or engineering projects. Is there like a similar initiative in India is like, I know like ISRO maybe was at some point.

Is there anything like that today?

[00:25:11] Soham: No, not that I'm aware of. So, they have recently reorganized the scientific funding structure of the various different scientific funding organizations that were in the country. I think they've rolled it into one organization, which they're calling the National Research Fund or the National Research Foundation.

But there is no, DARPA style agency or ARPA style agency, that has these kinds of specific program project mandates that they then execute. There is a, an organization in biology called the BIRAC, which funds sort of industry research. but it's really nothing like a DARPA or ARPA style organization at all.

It's really focused on present day incremental work and the funding amounts are very small.

[00:26:00] Abhi: I mean, that lends well to like the next question. Yeah.

[00:26:03] Policy prescriptions for India

---

[00:26:03] Abhi: What are your policy prescriptions to make India a global leader in biotech?

And maybe like, like R&D in general.

[00:26:11] Soham: So yeah, so the government should spend much more money on R&D. We should spend like as much as Israel does in terms of GDP percentage, like single digit percentage of the GDP at a minimum, like on R&D, right? I think this is the future of the Indian economy. We have to build products that no one else in the world can build.

Like clearly the world is becoming more protectionist, right? Just, you know, manufacturing something that everybody else can manufacture that maybe automation will take over in a few years is not going to be sufficient. We have to make things that nobody else has because that's when people are forced to purchase them, essentially, right?

And we have the talent, I believe, in India and Indian origin talent that we can import to do this, which brings me to my next policy prescription. So, are you familiar with the Thousand Talents program, the Chinese program?

[00:26:51] Abhi: Oh, yeah, like

bring, bring talent from the United States back to China. Exactly.

[00:26:55] Soham: And some of those people aren't even Chinese.

Like they're like just, I think some of those are just white people that just want to, you know, come to China for some arbitrary amount of money. So, we should have a Thousand Talents Program for India where we have India, especially Indian PIs, but also other people who are really good, who want to come back to India or want to come to India and work on what they're working on.

And we should give them a large amount of money over, you know, four or five years to do this and build labs, you know, in India to attract the best students globally to do this. And this is the ideal time to be doing this, because I don't know if you've noticed, but anti immigrant sentiment is building very quickly here in the U.S. and Canada and other places, right? In Canada, they've now cut, I think, student visas in half or more, right? So you're not going to be able to go to Canada as easily to do your PhD work, whether you're from India or from somewhere else, right? In the U.S. Who knows what's going to happen, but I don't think the environment is necessarily going to be conducive to more people whether it's starting here or sort of coming in here on H1Bs.

Maybe it will, maybe it won't be. So now is the time, if the Indian government is smart about it, to basically say, we're going to, be very open to scientific talent, Indian scientific talent, Indian origin scientific talent, but also scientific talent from all over the world that wants to work in a democratic, free country, which has a wonderful legacy of scientific research, come in and build, you know, your scientific legacy here in India.

And here's the money to do it. And so if I was in the Indian government, the Thousand Talents Program equivalent is the number one thing that I would do to change the trajectory of science research in India. I

[00:28:21] Abhi: mean, like the Thousand Talents, like the existence of the Thousand Talents Program sounds like an implicit recognition by the Chinese government that like a lot of talent from the country is going elsewhere to work in, work in areas that clearly don't exist in China.

[00:28:32] Soham: Which has happened in India.

[00:28:33] Abhi: Sure. Yeah,

like exactly. And like, it feels like, If, if China has explicitly recognized it, is there like some parallel to India, like recognizing that like graduates of like IITs and like talented people in general are just leaving India? Has there been any like national policy recognition by India that like attempts to stop the brain drain or do they just like not, not to attempt to engage with it?

[00:28:53] Soham: I'm not aware. So the government has not put up barriers, substantial barriers. They may. They may not. I worry that if they recognize it, it would be to try and just put up authoritarian barriers where they'd be like, you just can't leave now, which I think would suck. You know, you have the Singaporean version of this where, like, they have to do three years of, you know, You know, working in Singapore, like some years of working in Singapore, or something like that, even if you're from outside of Singapore, you know, if you, if you go get a degree at NUS with some money from the Singaporean government or whatever.

I, I think that, there has been no substantial recognition of this in a positive way, at least, like, no program has been set up to do this, which you would expect. You know, India's great advantage globally is that our, you know, immigrants from India are all over the world in extremely powerful positions in all fields, right?

And those people ideally could be used to, you know, form a core of an intellectual workforce that can help improve things for us, you know, in India, not to say that, you know, Indian faculty or folks doing research in India shouldn't be funded. They should also be funded substantially. But I think, you know, them having access to better colleagues and better students who would be attracted by researchers coming from top universities across the world and coming to India, I think would change the game, right?

But I haven't seen any movement yet towards anything that looks like this.

[00:30:13] Questions on vaccine design

---

[00:30:13] Abhi: Yeah.

Okay.

Hopefully it happens.

I think like outside of these, like those discussions of like, you know, Indian cultural problems and like policy recommendations, I have a lot of questions about like vaccine design in general. So you have antigens which are the molecular targets that vaccines train the immune system to recognize. But you also have immunogens which are what you actually put in the vaccine to generate that immune response, correct?

Sometimes they're the same thing. Sometimes they're not. And that's led to some really interesting approaches in vaccine design. Could you walk us through how our understanding of this relationship between antigens and immunogens have evolved since the birth of vaccinology?

[00:30:49] Soham: Yeah. So, I think the, the critical distinction between an immunogen is that the immunogen is sort of what goes in and elicits the response that you want. And I'm going to talk really in the context of antibodies, because that's what we work on at PopVax, and I think that's sort of simpler than talking about T cells.

But basically, the immunogen in some sense is what goes in, that elicits the antibodies that you want, that then, you know, whether in the short term or in the long term, form the core of the immune response against the actual pathogen, which has the antigens, right? And I think what's critical to understand is, you can have, you know, an antigen that is on the pathogen, right?

That isn't necessarily the right immunogen to use in your vaccine, right? Or you can have sort of a native form of the antigen that is on the pathogen that, when modified, would be the right immunogen, but when you put it in its native form, either doesn't elicit the antibody response you want, or the immune response you want, or even worse, causes problems, right?

And a good example of this is in, in RSV and respiratory syntactical virus where like, there's the RSV fusion protein or F protein. That's what's used in a lot of RSV vaccines as kind of the, the main, antigen, from which then immunogens are designed. And it turns out when you introduce the RSVF protein in its native form, right, when you in some sense take the antigen as it is, right, and you use that as your immunogen, it elicits antibodies that actually for people who have not yet been exposed to RSV, cause antibody dependent enhancement (ADE) of the disease, where it elicits antibodies that which, you know, bind to the pathogen in some way, but actually inhibit the ability of your immune response to clear the pathogen, right, whether that's inhibiting other antibodies or, or sort of in other ways, right?

And

[00:32:46] Abhi: Is that, is that just because like the antibodies generated by the RSV immunogen are just like not good enough?

[00:32:53] Soham: So typically, when you see ADE, you'll see something like the antibody binds, it binds really tightly, but it doesn't neutralize. So it doesn't prevent the you know, your, your virus from entering the cells.

And what it actually does is it prevents other, better, neutralizing antibodies from binding and preventing. Right, and so it like, it pulls away the surface that they could use to stop the pathogen, but it itself doesn't stop the pathogen. Or they're very weakly neutralizing. Right, and if you have overproduction of these antibodies, because that's the response that the body remembers to make.

Right, You know, memory B cells, you know, of this particular type of antibody, then you crowd out better antibodies that might actually be able to neutralize or stop the pathogen, right? and so in the case of the RSV fusion protein, the antigen is the RSV fusion protein because, you know, the antigen is what's on the pathogen.

But the immunogen that works, now we have several approved RSV vaccines, is a modified version of this, where you lock it in its pre fusion conformation. by doing two mutations, or a few mutations, rather. and it's similar to the mutations that have been made in the spike protein for COVID 19 that were used in those vaccines, right?

those were the S2P mutations. and actually invented by the same guy, Jason McClellan, right? who's a structural biologist who's pioneered this work, where basically he figured out, and this is kind of intuitive, if you have your, protein that, that binds with the cell receptor. So it's a cell entry protein for, a virus.

And there's a whole class of these called class one fusion proteins. And, you know, RSV is a class one fusion protein and COVID is a class one fusion protein and so on. Right. they exist in a pre fusion and a post fusion conformation. So they're in a, there's one shape that they take before they bind with the cell, right, before they bind with the receptor, and there's another shape they take after they've already bound with and the pathogen has entered the cell.

Right. That second conformation is kind of useless to elicit antibodies, because it's already

[00:34:41] Abhi: inside. It's already in,

[00:34:42] Soham: yeah. So what you want is you want to lock the protein in the conformation, that it, you know, that it is before it binds, so that you get antibodies to that conformation that may be prevented from binding, either by directly blocking binding with a receptor or by blocking it from changing shape.

Right? but the pathogen is crafty, and so it's not going to show you that conformation natively a lot. And so, pathogens where that is the case, if you show the native version, you typically get bad responses. And in this case, in the case of RSV, actively harmful responses. And so here, the immunogen that you want to use is the RSVF protein in this particular conformation, right?

but the antigen natively is not in that conformation all the time, or most of the time, and so it doesn't give you the antibodies you want.

[00:35:26] Abhi: And typically this is something like people only discover after like structural characterization.

[00:35:30] Soham: Well, in this case, people discovered it because they injected RSV vaccines into people.

It's just like, okay, it

[00:35:35] Abhi: didn't work.

[00:35:36] Soham: No, it, it made it worse. And so that's why, like, I think this was in the eighties, like a bunch of like RSV vaccine trials and then it all went horribly wrong. and then we just didn't have RSV vaccines until now, like in the last two years, there've been new RSV vaccines approved for the first time.

And all of them use this prefusion conformation version of the RcFusion protein.

[00:35:59] Abhi: Does a similar problem exist with like, like, like the COVID 19 virus as well?

[00:36:04] Soham: So all of the vaccines that I'm aware of use this prefusion conformation. that, that are recombinant, that use only the spike antigen. There's also, the, there's also the inactivated vaccine from Bharat Biotech in India that I don't think uses this, to my knowledge.

And it still kind of works, but it's not as good. So I think in COVID, it doesn't actually cause ADE. So COVID doesn't seem to have this ADE problem, but it just makes it better. And it produces a sort of more neutralizing response to lock it into this conformation.

[00:36:36] Abhi: That feels like, the.

The realization that there are different conformations you need to be aware of when designing immunogens feels like, like a, like a step level change in your ability to design good immunogens in the first place. Is there like another one on the horizon that you see that like, in terms of like, like a new bit of vaccinology that we're stumbling across that leads us to design better immunogens, or is it kind of like fuzzy right now?

[00:37:00] Soham: So I, yeah. Okay. So to go back to this antigen immunogen difference and then answer your question, right? So an antigen is anything. you know, that that sort of binds to an antibody, right? That's that's what the definition is, right? And then an immunogen is something that elicits, in the case of antibodies, that, that antibody, right?

So the immunogen for that antibody is the immunogen that would hypothetically elicit that antibody, right? I think what's changing now in vaccine design and what we work on a lot at PopVax is an extension of, of what we've just talked about with RSV, where you really want to elicit specific types of antibodies, right?

That do specific things, that have specific functionality. They, you want them to neutralize, right? You want them to neutralize broadly, right? Right? so, in the case of hepatitis C, which is one of the pathogens we work on, the pathogen mutates in the host. And so it presents as a quasi species, which means that If you try to knock down just one, you know, iteration of the quasi species, right, one particular version of the pathogen, maybe that's not sufficient because there are all these different mutated versions in the body, right?

So you need to, to get rid of hep C, you probably need, you know, broadly neutralizing antibodies. Antibodies that are able to go up against a whole bunch of mutated versions of the pathogen. and so I think what's changing in vaccinology now, is we're asking the question for the first time, can we elicit specific classes of antibodies, right?

that do these specific things and not other antibodies that are bad, right? So we find like a set of antibodies to elicit target antibodies. And increasingly, next generation sequencing techniques, like, you know, for example, doing single cell B cell sequencing allows us to go and identify specific antibody clones, right?

Like specific B cells that have a specific antibody. in patients who have been able to clear the pathogen or have had good outcomes, right? And then check the functionality. Is it broad neutralizing? You know, does it have a effector function? Is it is it good in some way and then try to actually elicit them, right?

And this is something that started in HIV vaccinology because there are these people in HIV called elite controllers who are able to control the progression of the of the disease. It doesn't go to, you know, it doesn't go from HIV to AIDS, without having antivirals, right?

[00:39:10] Abhi: Just like innately.

[00:39:11] Soham: They're like, they're just able to do this, right?

And so it was believed for a long time that, in fact, maybe true, that they are able to elicit these. you know, particular classes of, of broadly neutralizing very potent HIV antibodies. and at least in some cases for elite controllers, that does seem to be the reason, right? But they're a tiny, tiny fraction of the population.

And for a long time, HIV vaccinologists were like, let's try to get people to elicit these specific antibodies, right? And then we'll have a vaccine for HIV. The problem is these are extremely somatically hypermutated. So they're like 8, 10, 12, whatever, mutations away from the germline, naive version of the B cell to get to this antibody.

And it's a very torturous path to be able to elicit them. But the other problem is like fundamentally they believed, oh, if I could have an immunogen that binds really well, right? to this antibody, then I can elicit that antibody, right? And so they did a lot of like yeast display and stuff to try and find immunogens, like versions of the epitope, which is this like specific segment that is immunogenic and elicits the antibody, right?

Or elicits the immune response. they're like, they've tried various ways to, you know, you know, to get these epitopes to bind better with their antibody of interest and they're like, oh, and then when we put this in humans, the best binding one is most likely to be able to elicit the antibody. As it turns out, very expensively, we found out that's not true, right?

It turns out that binding is not a sort of complete representation of elicitation. In fact, they're quite different processes, right?

[00:40:35] Abhi: Sorry, going back quickly, going back to the HIV thing. Yeah. I still don't quite understand, like, why, why. Why can't we just like replicate this specific antibody, like the ability to, like an immunogen that causes this specific type of antibody to appear.

Why can't we just, why can't we just do that?

[00:40:51] Soham: Well, we don't know what it is.

[00:40:52] Abhi: Oh, like, it's just like, why don't we know what it is? Well,

[00:40:57] Soham: in HIV's case in particular, HIV mutates on the host too. That's true.

[00:41:00] Abhi: Okay, so like, even

if you could replicate it, like, you have this group of people who can resist the progression to AIDS really well, but each of the antibodies they develop that resist that progression are kind of unique from person to person.

[00:41:10] Soham: They're not, some of them aren't, some of them seem to be more broadly applicable. Okay. But it's, what you're hitting at is a very important problem, It's sort of not intuitive to understand, which is just because we have the antibody, we, most of the time, we don't know what elicited it.

[00:41:25] Abhi: Gotcha.

[00:41:25] Soham: Right? Especially in these pathogens where there's a lot of mutation going on, right?

and we don't specifically know what the immune system saw. What shape or conformation of the antigen that it saw from the pathogen that allowed it to elicit that antibody and sometimes maybe you just get lucky right and you like you see some cryptic pocket that's not usually exposed, but in this case in this patient it was and some B cells saw it and it really worked well. So that B cell replicated and then you got this right but it happens in so few cases that it's not a repeatable process.

[00:41:57] Abhi: I'm, I'm curious, has it been like elucidated, like what's unique about the, is there anything unique about these people or, is it just like they, they got lucky?

[00:42:02] Soham: There are lots of different hypotheses and I don't, I'm not an HIV expert, but I think it's like, it's a very hard field because there are lots of things that I think are simultaneously true.

And there's no one specific answer to the question of, like, what makes these elite controllers elite controllers. Yeah. And so what happened is basically, we, is, I think this exact thing, which was like, oh, we have the antibody, now it should be easy. And then, like, 20 years later, no HIV vaccine, it's not that easy, right?

[00:42:30] Abhi: Yeah.

[00:42:31] Soham: And in part because it was this dogma became like sort of, oh, if we make it bind really well, it'll elicit the antibody. And I think the answer is that's not true, right? And so we have to explicitly in some sense, you know, model the elicitation process. And so what we're trying to do at PopVax, and I think what is kind of the future of vaccinology as a field is, I have a set of antibodies that I want to elicit.

I need to be able to build a machine learning model or build some kind of, some kind of model one way or the other, right, that allows me to go from that set of antibodies to the immunogen that then elicits them, right? And the only way to do that is to collect a bunch of data about how libraries of differently designed immunogens elicit antibodies, right?

And go front and back and then use that as a way to actually design your immunogens.

[00:43:13] Abhi: Yeah, I mean, like, instinctively, like, before this conversation, I would have assumed, like, immunogen equals, like, something that, like, can, like, that binds really well. Yeah, yeah.

[00:43:21] Soham: But that's the antigen, right? And,

[00:43:23] Abhi: like, I, I guess.

The the part that i've been a little bit confused about like we've talked a little bit about like what is the what's like chemically going on that causes the elicitation process? Is that not well understood?

[00:43:34] Soham: It's it's I think it's sort of mechanistically understood to some degree. But there's so many different sort of there are different things that can happen in that process, so many different combinations of the of your You know CDRs or your variable reasons in the antibody that It's, it's not something that can be, I think, explicitly modeled from theory alone, right?

It's also, so if you want to go really wild about it, have you heard of Jerne's network theory?

[00:43:57] Abhi: I have not, no.

[00:43:57] Soham: Okay, so Jerne, and it, it, Jerne's network theory, by the way, is totally true, despite the fact that it sounds insane. So you have these cascades of immune responses. So I have my immunogen, it goes into the, the body, it elicits, like, it goes, you know, binds with some B cell, and then maybe there's some mutation and the B cell replicates.

It's closer to binding now with this antibody after the mutation. it turns out those antibodies can also elicit antibodies. Okay, like there's like

[00:44:21] Abhi: cross immune, like antibodies can talk to other antibodies. No,

[00:44:24] Soham: antibodies can elicit other antibodies. So the antibody goes and binds with a naive B cell, it elicits another antibody against the antibody, right?

and then those antibodies can elicit further antibodies. And then, but those antibodies, right, sort of the anti idiotypic antibodies they're called, are now mimics, structural mimics, of the, original immunogen. So you can do things, crazy things, like you can inject an animal, and you can go seek out these anti or idiotypic antibodies, and you can use them as an immunogen to elicit the immune response that you want.

[00:44:59] Abhi: That's bizarre.

[00:45:00] Soham: Right? And so it's a net, it's like this cascade of this elicits that, elicits this, and it's not just happening in one dimension because a single immunogen is going to list a whole bunch of different antibodies, right? And so it's, and then they're all, there's all this crosstalk going on, and somewhere at the end of that is my mature antibody that I want.

[00:45:17] Abhi: That actually like, like, neutralizes whatever pathogen you have. Okay.

[00:45:22] Soham: So it's, it's, it's not, it's not obvious to me that there's like a simple mechanistic way to simulate this out.

[00:45:28] What does PopVax do?

---

[00:45:28] Abhi: I haven't actually given an opportunity for you to like give the full overview of what PopVax is and like how you guys function. So feel free to just give me an overview.

[00:45:36] Soham: Yeah, sounds good. So Popvax works on broadly protective vaccines. We develop broadly protective vaccines, both against pathogens where there are existing vaccines, but we think they're not broad enough. So an example of this is COVID, whereas I've talked about a few times now, you know, existing vaccines didn't cover a bunch of the new variants that emerged pretty rapidly.

Influenza, which is another key indication for us where, you know, there, there are these seasonal vaccines that people take each year, but oftentimes those vaccines, aren't as effective as they could be because eight months ago, the prediction of what strain was going to be dominant was incorrect, or that strain has since mutated in such a way that it's no longer effective.

And of course, these seasonal vaccines don't offer protection against potential pandemic influenza like H5N1, which is now spreading in cows and other animals in the US, right? We also work on vaccines against, pathogens where there are no existing vaccines. An example of this is Hep C, which I mentioned, which is also sort of a broadly protective case, where, you know, there, as I said, it mutates a lot in the, in the human host.

And so you end up having to protect against not just one, pathogen, but like a whole quasi species of this pathogen to get an effective vaccine. And also, you know, for, for other pathogens like strep A, where, existing vaccine design approaches created some kind of, issue. So we've talked about antibody dependent enhancement.

In the case of strep A, existing strep A vaccines, you know, a few decades ago, caused autoimmunity because there was an antigen that they were using, in their vaccine that, that actually elicitated antibodies against a human protein, right? So, all of these are, pathogens where some kind of precision approach is needed, right?

Whether you're getting, you know, trying to elicit antibodies for broader protection or antibodies that are avoiding some kind of, response that would be bad with antibody dependent enhancement or it's, you know, it's some kind of autoimmunity. Right? And the way we're organized essentially is we do all of this stuff in house.

So we have our own mRNA platform that is able to, display immunogens in this repeating form on things that are similar to virus like particles that's encoded in RNA. So basically what you inject is a standard mRNA lipid nanoparticle without having to do fancy manufacturing, right? We design our own lipids for lipid nanoparticles.

So our own novel ionizable lipids. We have a library of about 400 of these now that that our formulation and delivery teams have worked on and we've tested now thousands of different lipid nanoparticle formulations, you know, many in vitro and, and, and even hundreds in vivo, that allow us to be able to basically pinpoint what the right lipid that we've designed and what the right formulation is to deliver these mRNA that encode these vaccine immunogens, And we've taken a very similar empirical approach to optimizing these as we have our immunogen designs, which I've talked about.

And then, of course, we design our immunogens, which we've been talking about quite a bit. But basically, we use, you know, precision immunogen design, to try and elicit specific classes of antibodies we're interested in, avoid antibodies that we're not interested in, or that, you know, could cause, you know, cause problems, in particular for, for broad protection that is also potent at the same time.

And, and we use this mRNA encoded, you know, virus like particle structure to display these immunogens in a way that massively increases the magnitude of the antibody response elicited and also gives us some additional breadth in terms of the VLP structure also helps with breadth. So we are about 70 people, based in our Hyderabad facility.

And we actually go all the way up to GMP production. So we, we make, clinical doses, we're making them for our first, COVID 19 phase one trial. And so we can really go from concept all the way to clinical dose production in this one facility with our, you know, end to end team. And I think that's also quite different from, from the way a lot of biotech startups operate in the U S today, where they contract out a lot of different work, exactly.

Whereas we kind of do all of this in house.

[00:49:27] Abhi: Like how, like PopVax does precision immunogen design, I think it's good to give something background context on how it actually works. So you have have like mRNA, in a, a lipid nanoparticle that encodes for a virus-like particle (VLP) that has a linker that is then attached to the immunogen, right?

So you have this like, like shield of immunogens wrapped around the, the virus-like particle. Does, do you use machine learning at every step of the design of the Virus-Like Particle, the linker? And the immunogen, do you focus on one over the other? Yeah, like walk me through how, like, a workflow goes.

[00:49:57] Soham: Yeah, so for our COVID vaccine, we have, this sort of self assembling protein, that we attach via a design linker to our immunogen of interest, which we also sort of, you know, redesigned to be more compatible with that. and, those, you know, basically together form a single protein that we encode in mRNA.

And that is, when it goes in the cell, it translates into this, you know, single protein all linked together. And then the self assembling proteins, you know, sort of join together, they self assemble. And then what you get is sort of something that looks like a virus like particle that displays the immunogen repeatedly.

We started by not using a whole ton of machine learning in this process. We used existing known self assembling protein. We used existing known linkers. We used, you know, sort of more manual sort of Rosetta based modifications to the, the immunogen, to try and present it better on this, this sort of VLP like structure.

And we just did a bunch of combinatorial optimization in vivo of what worked best, right? And so the advantage was, we just did a whole bunch of testing and we were willing to try, I think, as I was saying earlier, like a whole bunch of things and see what worked best, weren't really too wedded to what would work best with, you know, what we got is what worked best.

and, and then that, we,

[00:51:14] Abhi: Sorry, with the primary measure of efficacy being neutralizing antibodies?

[00:51:18] Soham: In this case, the primary measure of efficacy being two things. One is sort of neutralization magnitude, sort of, depth where we, like, how much better is it in neutralizing than the existing BioNTech Moderna vaccine sequences, which we use as comparators, right?

And so our vaccine, for wild type, when we did all these optimizations, and displayed it in this way, turned out to be, like 55X better than just the RBD immunogen at the same mRNA dose in terms of neutralizing antibody data against wild type. And 22x better compared to, the BioNTech sequence, basically sort of the existing wild type mRNA spike COVID vaccine.

And, and that was, you know, our, with our own experiments, you know, in vivo. So we didn't have the whole, the actual BioNTech vaccine, but we made the same sequence and we encapsulated in the same lipid nanoparticle so forth. Right. So a reasonable facsimile. And so that was one measure. And the other measure was breadth.

So are we able to, you know, as I said at the beginning, we were funded to do this vaccine, and I started the company, because, the existing COVID vaccines were failing as soon as there was a sort of new variant, right? And in India, we had the emergence of this Delta variant, which emerged in Maharashtra, where I was in Pune, where I'd moved my software for a robotics company, right?

And there was no real attempt by the Indian vaccine companies to make a Delta specific variant anytime. That would have helped, right? So we had a lot of people essentially die as a result of that. And so a big goal for me was, can we make a vaccine that has broader protection against variants? So the other big measure for us was, we had a whole bunch of variant pseudoviruses, where we had both existing real variants that had emerged, and we mutated these pseudoviruses to, you know, present versions of the spike protein that didn't exist in nature, but could very well, right?

And some of this was based on a data set from a guy called Jesse Bloom, who's at, Fred Hutch, who does deep mutational scanning, which I'm sure you're familiar with, right? So, like, basically, you know, mutate the, you know,

[00:53:10] Abhi: The single and double substitutions

[00:53:13] Soham: Yeah. Exactly. Exactly. All singles and then a couple of doubles.

Like some distribution of doubles, not all. Where, you know, it's again, a pseudo virus system where he mutates the spike protein. And then he figures out which of these are still stable and functional and then which of these are now able to escape, you know, antibodies that were elicited to the original vaccines.

And so we use that, not deep mutational scanning, but we sort of used, you know, versions of the spike protein from his assays that had shown this kind of escape potential. Okay. In our pseudoviruses, so not the live virus, so very safe, right, they are not functional virions. But, we use those as a, as a measure of how much breadth we're able to get.

[00:53:53] Abhi: You have these like two measures of efficacy of whatever you're designing.

And you mentioned that, when you first started off, you started off with like well known scaffolds, well known linkers, well known immunogens. Like where do you, where do you go from there on that?

[00:54:06] Soham: So from there, now what we're doing is we're using different, sort of self assembling proteins, some of which are de novo generated by our team.

We're using different linkers, some of which are de novo generated by, by machine learning. and then we're using versions of the immunogen, which are, in the way that I was describing before these precision immunogens where we either eliminate epitopes and then fill that in using machine learning or we are scaffolding only a single epitope that's intended to elicit a single type of antibody but we're still able to use this mRNA encoded VLP structure to substantially boost the neutralizing antibody response.

And this is particularly important for epitopes, which might be naturally subdominant, right? So, there's this notion of immunodominance. If you have a whole bunch of antigens in a pathogen, some of them are going to elicit a big response, and some of them less so. And some of that is due to T cell help, and there are other reasons for it.

But, let's just accept that as fact, right? Sometimes the best antibodies are elicited by epitopes which are not immunodominant, which sucks for you as a vaccine designer, you know, if you, if you're trying to sort of elicit specifically to that. So one way to deal with that is you eliminate the other immunodominant epitopes that you don't want, right?

And so we just don't show those, right, if we don't want to list antibodies to them. The other way is using something like what we're doing, which is this sort of VLP approach, which boosts the immune response even to these subdominant epitopes.

[00:55:24] Abhi: Gotcha.

Right. Is it, is it immediately clear like, which part of this, like three step redesigning or like, like three areas to redesign?

Is it clear like which one has the highest impact? Is it clearly the immunogen or is like the linker and the scaffold actually.

[00:55:38] Soham: So the immunogen gives you specificity. Okay, so like by redesigning the immunogen to make it better and better. We're able to get closer to eliciting the specific kinds of antibodies we want. By changing the linker and the self assembling protein, we're able to get basically a bigger magnitude of response.

[00:55:53] Abhi: Gotcha.

All right. Yeah, that makes sense. Historically, it sounds like, like, the initial step is, like, redesigning the immunogen, and then, like, once you've got a good immunogen, you start working on everything else. Or is it, like, not as cleanly segmented?

[00:56:09] Soham: No, it's not segmented like that at all. So, like, we do everything in parallel, basically.

[00:56:11] Abhi: Okay. Okay.

[00:56:13] Soham: Because we initially didn't redesign the immunogen that much at all. We started with this new, the presentation methodology. And then we started optimizing the immunogen.

[00:56:20] The role of machine learning in vaccine design

---

[00:56:20] Abhi: On the topic of like how you actually do this in practice, like PopVax is a biotech company, but also they're like, you guys have like a machine learning team and you're trying to use these protein foundation models for the purposes of, immunogen design.

Could you walk me through how like useful are these? Like, well, it's like AlphaFold, RFdiffusion. Are they revolutionary, somewhat useful or not useful at all?

[00:56:41] Soham: So, so we use, I think the way to describe what we do at PopVax is basically, we take existing protein design models. So as a diffusion models or protein language models that can spit out new proteins, that we basically condition on the task of scaffolding epitopes, which are, you know, the pieces of your antigen that, that, you know, that are specifically binding to the antibody, that therefore, you know, in some fashion, elicit the antibody, right?

And we use them to generate libraries of these scaffolds, that we can then test either in vivo, so in mouse models, or in organoid models, which are based on human cells, to see what antibodies they elicit, right, and then use that to inform a feedback loop where basically we can rank these immunogens, these design immunogens, based on how close they got to eliciting the antibodies we wanted to elicit.

and then use that, using RL or DPO or, you know, basically alignment and fine tuning methods like this to make the design models spit out things that are closer to what we want, that are closer to being able to elicit the antibodies that we want. Right.

[00:57:48] Abhi: Gotcha.

[00:57:49] Soham: So like we, as a base case, we use the existing models.

We don't train our own models on PDB, at least not yet. Right. but then we're using this, this dataset, this new dataset that we have basically, where we map elicited antibodies for the first time, which no one else is really doing to create a feedback loop that is able to get us closer and closer to eliciting the antibodies we want to elicit, right?

And so that's how we use these models. And I think in that sense, the models are very good at spitting out proteins that, you know, have stable structure and that, you know, we can thus use as, as test cases. and I think the models are amenable to being fine tuned and to being aligned. And so we're not the only person doing this in some context to the other.

People are doing this for antibodies. People are doing this for, you know, better performing enzymes, but data sets with hundreds or, you know, certainly thousands of data points can actually be used to move these models towards giving you structures that are closer to what you want. Right. and so I think in that sense that are useful and we've seen that they're useful in this way.

Structure prediction models, we use a number of those. Basically, once we generate these designs from these, from these models. and when we use sort of more manual methods of design, like Rosetta, we use the structure models to predict whether we're getting something close to what we want it to get.

Right. And we can use multiple of these different structure prediction models, as sort of orthogonal measures a little bit. They're not that orthogonal cause they're all trained on pretty similar data sets, but like somewhat orthogonal measures of like, you know, these two models are agreeing. Probably this is a protein that I would want to produce.

Whereas, you know, when the prediction is completely different from what the design model was designing in terms of structure, then we typically discard. So we use that as part of our down selection pipeline in silico.

[00:59:22] Abhi: Is it, when it comes to, like, machine learning actually used for this, is it clear, like whether like structure models are really good at this, language models are even better ?

[00:59:30] Soham: We mostly use structure based diffusion models. PLMs we've had sort of less good results with, but maybe that'll change. I mean, I think that there are new models that people are dropping, you know, very frequently. I think what's nice about what we're doing is we can use, other people's quote unquote sort of protein design foundation models.

We can fine tune them using a kind of straightforward, you know, set of methodologies and we can sort of use what's best or, you know, depending on what the licensing terms and so on and so forth are these models. In future we may, you know, develop our own sort of immunology foundation models, but I think we'll need a much larger scale of data than we currently have to do that.

[01:00:07] Abhi: Yeah. I mean, I'm not sure how much like information you're able to share about this, but like the actual design process, like. Like given like AlphaFold2, how do you use AlphaFold2 to redesign the immunogen?

[01:00:16] Soham: So like AlphaFold2 we use as a, you know, as a, as a kind of checkpoint where basically if we're designing, a protein that's intended to be in a certain structure, we use AlphaFold2 to predict out the sequence once we've generated the sequence, right?

So, if you're using a, a structure model that spits out a backbone, the backbone, then goes into what people call a, inverse folding model or like, you know, something like protein MPNN. Yeah, exactly. yeah. Some of these models are, you know, give you a sequence as well, right? And then we, we can predict out basically using Alphafold2, using some structural prediction model.

Is it conforming to the structure that we expected? And that if, if it's not, right, we, if we generate tens of thousands of options, we'll eliminate the ones that don't fold in the way that we would expect.

[01:01:00] Abhi: So I bet, I guess, like, how do you, like using the structure prediction model, what are you actually trying to optimize the end result for?

[01:01:08] Soham: So the structure prediction model, again, we're basically just trying to eliminate bad designs that we don't want to test in vivo.

[01:01:14] Abhi: Okay, and like how do you define a bad design ?

Like low pLDDT?

[01:01:18] Soham: Yeah low pLDDT is one example,.

[01:01:21] Abhi: So there's like

in silico metrics you're using to decide.

[01:01:24] Soham: Yeah. So we start with, you know, tens of thousands of designs and then we downselect down to hundreds that we would actually test in vivo. Yeah.

[01:01:30] Abhi: And so then like how do you take into account, like the data that you've tested in vivo, how does that get like fed back into this like redesign process?

[01:01:38] Soham: So we can create a preference ranking basically, right? So you got a set of antibodies, want to elicit, which of the immunogens got closest to eliciting the antibodies we want to elicit, and we can do this with function, which we were talking about earlier. So like, which of them, you know, neutralize the best, which of them neutralized the most breadth, right?

And then you can create rankings and then you can use different versions of these rankings to do DPO or do fine tuning basically.

[01:01:58] Abhi: I'm curious. Has it like, have you like, have you tried Boltz-1? Has it been like a big step up from AlphaFold 2?

[01:02:04] Soham: I, I'm not sure that we've tried it extensively enough to have an answer to that question.

Gotcha. Okay. Okay.

Structure prediction is not such a big bottleneck

[01:02:15] Abhi: So what actually is the bottleneck then?

[01:02:17] Soham: Yeah, I think we, it'd be nice to have... so the, the design models themselves are not necessarily so great at scaffolding the epitopes that we want. It's, they're still not able to produce the structures that we want and the conformations that we want reliably for some of the problems.

And so I think making those better and we're, we're, you know, we're doing our best also to try and fine tune these models to get better at those problems, even in silico before we get to the in vitro and vivo data, right, is more critical for us, whereas I think certainly structural prediction models aren't ideal.

But for the size of proteins that we're using, you know, we don't think that's sort of a big bottleneck to us being able to test these designs, in vitro and vivo, because even if we throw away some designs that are, you know, maybe actually good, things that are predicted with high confidence of these models tend to actually be what, what you would expect them to be.

[01:03:07] Abhi: Is it typically clear, like at the end of these, like design sessions, is it, do you have a good sense of like, what do you want the end antibody to resemble?

[01:03:17] Soham: Say again, please.

[01:03:17] Abhi: Like, you're, you're, you're designing an immunogen, you want to elicit some antibody response. Yeah. And you want this antibody response to be like good for whatever you're trying to protect against.

[01:03:28] Soham: Yeah.

[01:03:30] Abhi: Do you have like a gold standard antibody that you're comparing it against or are you purely looking at the immune response of whatever you've injected this into?

[01:03:38] Soham: So we can do both, right? So, and that's a good question. So, we can look at the functional response, right? Which is to say we can, inject into mice, any specific design.

We take the serum out of the mouse and then we can use that serum in a neutralization assay, in an effector function assay to see the functionality of the antibodies elicited. Are they able to prevent the pathogen from entering the cell? Are they able to clear infected cells? Right? So that is one thing, but we also have, say in the case of Hep C, there are these antibodies, which are clearance associated antibodies, sequenced out of people who are able to clear the pathogen.

And, and it's an important distinction here between HIV and Hep C, because in Hep C, it's like 30 plus percent of people clear the pathogen straight away. So it's not rare, right, compared to HIV, where it's like some fraction of a percentage.

[01:04:24] Abhi: So there's like some similarity in antibodies amongst 30 percent.

[01:04:29] Soham: That remains to be seen, people, I mean, lots of people have done sequencing at a small scale, not a ton of people have done it at a large scale in a diverse dataset, which is one thing we're trying to fix, but hypothetically, you know, you, you can get out a set of target antibodies, and we do have a set, some, a small set, at least, of target antibodies that we're quite sure will help clear the pathogen, based on that research that other people have done, right?

For those antibodies, we actually can do sequencing of the B cells that are elicited, right? and we can figure out how close we got in terms of sequence and predicted structure to those antibodies. Now, in a mouse model, those aren't going to be exact, right? but in an organoid model, we can get closer to what would actually be elicited in a human.

[01:05:08] Abhi: Is it like an insane idea to like, like hallucinate an antibody that binds to some like virus and then work backwards from the hallucinated antibody to design the immunogen that would cause that hallucinated antibody to exist. Does that does that at all make sense?

[01:05:25] Soham: It's not crazy. I just, I think the, the, I don't think you can do that without a new data set, right?

Because there's this sort of black box of elicitation, which is very complex, right? Even though it has simple constituent components because of, you know, all of this interaction, it becomes non trivial, right? and I think we don't currently have a dataset that allows you to, to, to make that jump, right?

People are doing de novo, you know, kind of hallucination design of antibodies intended to neutralize .... there was a paper recently on neutralizing, snake venom, right? You know, neutralize viruses. Like, I think there are ways to do that with existing datasets. Immunogen design in particular, I think, is a problem that cannot be tackled with these existing datasets in and of itself, right?

[01:06:11] Abhi: Do you think it's just like a scale question? Like, people haven't collected enough data yet?

[01:06:15] Soham: People haven't collected any data.

[01:06:16] Abhi: Oh, okay.

[01:06:18] Soham: I think, if you ask me, like, who has data that maps from immunogen design to what antibodies were elicited, I think the answer is me.

[01:06:29] The (conservative) culture of vaccinology

---

[01:06:29] Abhi: I mean,

I guess that kind of kind of raises a question.

Why, why don't vaccine companies do more precision immunogen design, like using these sorts of models?

[01:06:38] Soham: I'm glad you asked. I think the answer is vaccine companies are very conservative, and vaccinologists are very conservative. and so. There are a few things we do differently, which I think really, they should all be doing, right?

One is we test a lot of immunogens in vivo. We've tested, as I said, a thousand plus immunogens in vivo, or, you know, something in that order of magnitude for, for our novel vaccine constructs. We have characterized the response to all of those immunogens, and we use that to sort of decide what we want to advance into, into clinic, right?

A typical vaccine program, as far as I understand from people in big pharma, really doesn't test more than a dozen or a couple of dozens of these in vivo, of these different sort of design candidates to the extent that they design it all, right? Before they move forward into clinics, so they go to clinic pretty quickly in big pharma because they have a lot of money. But then often these designs don't do what you want them to do because you haven't really optimized them to and so they fail.

[01:07:31] Abhi: So, it's not necessarily that they're designing only 10 because they have like such a significantly higher hit rate than like anyone else.

[01:07:37] Soham: Oh, no. And you know, one of the reasons big pharma has pulled back from vaccine investment is every big pharma company, almost all of them have some like crazy failure story where they invested a billion dollars in a vaccine program and it went, it went just super wrong, right? Like Sanofi has this Dengue vaccine program that we were talking about before this, where similar to RSV, when you take their vaccine and you haven't been exposed to dengue before, it actually makes your case of dengue worse, right?

And they invested like a billion dollars into that program, which is effectively money set on fire, right? And had they been smarter about precision immunogen design. I think they wouldn't have made that screw up.

[01:08:15] Abhi: Is, do you think, like prior, like, like pre 2021 precise immunogen design was just like off the table for anyone?

Like, like prior to the release of AlphaFold 2? What are options for people to do precision immunogen design before?

[01:08:28] Soham: Not in the way that we're doing it, but certainly there are ways to sort of rational design using Rosetta and tools like that. And I think there are ways to sort of more comprehensively characterize the outcome of, you know, of, of injecting your immunogen into it's a, whether it's a transgenic mouse model or an organoid model or something closer to humans, I think that was available with a little bit of work.

Right. And I think they could have, I think not just them, but I think lots of people could have done better about advancing ways to characterize those responses comprehensively to figure out whether the response you're getting is actually the response you want.

[01:09:02] Abhi: Yeah.

[01:09:02] Soham: But I think, someone recently, you know, someone who has spent many years in big pharma was telling me, look, the vaccines aren't sexy.

And so preclinically big pharma doesn't invest enough money in like the cool new technologies for, for vaccines. Right. And, you know, the people who are the most ambitious about working on new and interesting stuff often leave vaccines, whether it's within a big pharma company or broadly in the field, wouldn't you rather work in a field like cancer immunotherapy, where there's so much more money and so much more application of the latest methods, right?

Whether it's like single cell transcriptomics, or it's, you know, these machine learning models for protein design or whatever, right? Rather than like a backwater, like vaccines, right? And so I think as a result, you end up with like fairly conservative people doing like a small number of tests, not really comprehensively characterizing the immune response, like a systems immunology perspective either.

And also just often getting it wrong about what they should be designing for in the first place, right? So like we've had this legacy of HSV2 vaccines trying to design for neutralization. Where it turns out, and there's a lot of good data that shows this, neutralization is not sufficient, in order to, to be able to clear HSV 2, or to prevent, HSV 2 infection.

You need effector function. which is something that people have known now for a little while.

[01:10:22] Abhi: Sorry, could you, like, like, walk me through the, like, the rationale for that?

[01:10:25] Soham: For what?

[01:10:25] Abhi: For, like, why you need effector function.

[01:10:28] Soham: Empirically.

[01:10:29] Abhi: Okay. Just like, that's just like,

[01:10:30] Soham: I don't know why,

But like empirically you need a factor function, like you, you need to be able to have these, like, you know, some kind of ADCC (Antibody-dependent cellular cytotoxicity) or, you know, ADCP (Antibody-Dependent Cellular Phagocytosis) or something like this, you know, to be able to, to, to clear, HSV2 or to prevent HSV2 infection.

And what happens is, what happened is, I think, people assumed that, that neutralization would be okay, would be sufficient. And so there are a bunch of in clinic HSV2 vaccine programs that were conditioned on like, oh, we got a good neutralization response, and it just turned out not to be good enough, right?

And with tuberculosis, we have a TB vaccine program that we want to start, right? People have been focused on T cell responses for a very long time, and there are good theoretical reasons, like it's an intracellular pathogen and so like maybe antibodies won't do anything, right? But we actually do know that there are antibodies that, that have, you know, substantial effect on, on the bacteria, right?

We know that there are people who have profiled T cell responses from people who are able to clear TB and found that actually the T cell response that matters most, are like helper T cells that essentially help the elicitation of antibodies, right? But nobody really seems to have a TB vaccinology program based on antibody elicitation, funnily enough, right?

And when we propose this to people, like, old hats in the room say like, oh, we know it's T cell response. Well, if you know it's T cell response, why 30 years? Yeah, why don't they work? Exactly.

[01:11:58] Abhi: You've mentioned before that, like, immunologists are like, Like not very trusting of like empirical data, how like, like, like, do you, do you think that like the distrust is like a little bit rational in their eyes?

Have they been burnt before a lot on like,

[01:12:12] Soham: So I don't think I said that. Let me actually rephrase what you said. What I think is that they're not sufficiently empirical in that I think they're not willing to try enough stuff and then just sort of like systematically figure out what works and what doesn't.

So like, I think what's really worked for me as a non biologist, like as a computer scientist is like, I don't a priori have any pet theories about what works and what doesn't, right? Like, I'm very driven by like, okay, what are the possibilities of what could work here, right? And for that, you do need to know some immunology and, you know, I asked the right people, right?

And then I'm just like, okay, let's make designs that are going to optimize all these different things separately. And let's see what works best. So if you're an immunologist and you know, effector function is a thing. Okay. Right? Or if you're an immunologist or a vaccinologist working on TB and you know that antibodies are a thing, why wouldn't you just have another, a design that focuses on antibody elicitation also, right?

Why wouldn't you want to like systematically just create a matrix of all the things that could matter, and then make designs that could optimize in those directions, and then see what works and double down on that, right? I think that's sort of the recipe for what we've done at PopVax that works super well, right?

And that, you know, strangely enough, I think a lot of people who are immunologists or vaccinologists aren't doing. I think they come in with some theory about how this vaccine is going to work. and then they make a very expensive, multi year, maybe even in some cases multi decade bet on that theory being correct, without having any backup options, or without even really comprehensively comparing it to other alternative hypotheses that might be the case.

[01:13:41] Abhi: Instinctively the reason, the reason why they would do that is because they have so few shots on goal that they need to like have some very strong prior.

[01:13:49] Soham: But preclinically, they have lots of shots on goal.

[01:13:51] Abhi: That's fair.

Yeah.

[01:13:52] Soham: And even phase one, frankly, they have lots of shots on goal. So like , in phase one, right?

I think you would still be able to distinguish between a strategy that might work much better in phase one, because you could take that, you know, you could, you could see that, it's eliciting effector function. Maybe that's good. Right? Like maybe that, that's going to do something. Whereas it's neutralizing, but like, you know, it's maybe weakly neutralizing or sort of not giving us, complete clearance of the pathogen, if you passively administer those antibodies into the animal model, right?

And so there are all sorts of things you can do if you have some early stage data, which isn't terribly expensive. Whereas instead, these folks are jumping straight into phase two, phase three of hundreds and hundreds of millions of dollars. And then that money is all set on fire.

[01:14:38] Abhi: Is that just because like, overconfidence, like what's causing that like clearly like mistaken step?

[01:14:47] Soham: I don't know. I don't, I'm not these folks and they're very smart people who know more than me. And so it's possible they have very good reasons for what they're doing. But in practice, I think what we see is, you know, there's like a, there's a lot of resistance to trying out things quickly, right?

And in the field, and I don't see why, you know, in addition to whatever theories they have, they wouldn't be benefited by just trying out more things quickly.

[01:15:11] Abhi: Like orthogonal theories, just like to have backups.

[01:15:14] Soham: Yeah, yeah, and then maybe you can combine that, right? Like, you can combine the immunogen that gives you really good neutralization, good effector function, and then you put that in.

Whatever the mechanism is by which it eventually works, it's probably more likely to work.

[01:15:27] Abhi: Instinctively. I imagine that like. I kind of, like, assumed, like, the focus on broad protection, like, goes without saying. Like, every vaccine company is interested in broad protection versus just one variant.

[01:15:36] Soham: But they're not.

[01:15:37] Abhi: Yeah, like, that's interesting. Why, like, like, you know that the virus is going to mutate.

Right. Like, why wouldn't you try to, like, design for, like, highly conserved epitopes?

[01:15:47] Soham: Well, so, designing for highly conserved epitopes and actually achieving broad protection are very much not the same thing.

[01:15:52] Abhi: Oh,

okay. Right. Well, why is that?

[01:15:54] Soham: It turns out just a naive strategy of designing for conserved epitopes like doesn't work, right?

Okay. And so because this has been hard, I think traditional vaccine companies have shied away from it where, you know, what you can do reliably is you can do high multivalency, for example, or you can show as in the case of pneumococcal vaccines, you know, you can show 20 serotypes, right? And, and thus get a response against all those and at least reliably protect against those, right?

But in the case of influenza, we don't even do that. So influenza vaccine platforms, typically use egg based production. That's what most influenza vaccines are made in. And, you know, the problem there basically is, it's difficult enough to even optimize two or three or four strains in the eggs, which is up to four valence is what seasonal influenza vaccines have been, and, it's, the strains that are recommended are often not the strains that work well in eggs, so you have to find some close by strain that, that produces well in the egg.

Scaling that up to high multivalency is very hard. So even broadly protective influenza vaccines haven't really been made, right, that are these highly multivalent vaccines like PCV. But yeah, actually designing vaccines for broad protection using conserved epitopes requires the kind of precision design strategy that we've been talking about that we do at PopVax because naive approaches to it just don't work.

[01:17:15] Abhi: My naive assumption of like, you take your virus, you align it to the nearby virus families, find the conserved epitopes, turn that into your immunogen. Why, why, why doesn't that work? Why is like the not just focusing on the most conserved epitopes, not a good strategy for design?

[01:17:29] Soham: It's just not sufficient, right? So if you look at beta coronaviruses, so like SARS, MERS, you know, these, these pathogens, right? You take the conserved epitopes, there are a few conserved epitopes, and you just make a vaccine. That has those epitopes and I know because we've done it and you inject it, you get something that like doesn't neutralize, and, is no good for protection.

[01:17:50] Abhi: So then what actually causes broad protection? Like, what, like, what, yeah, like, what causes it?

[01:17:55] Soham: So it's not sufficient to just look at conserved epitopes. What you have to do is you have to look at conserved epitopes and then find a way to redesign the, or display that epitope, you know, design the immunogen, right, in a way that actually elicits that neutralizing or effective function or a response that you want.

Natively just displaying the conserved epitopes is not good enough, right? And because nobody has developed the kind of pipeline that we're developing, or that maybe a few others are working on, up till now, and in some sense the technology hasn't really been available up till now, to do this kind of feedback loop, where you have, you know, a whole bunch of computational designs, and you figure out what they elicit.

And then you use that to fine tune your model, right? People have been trying to do this manually, where they, you know, Rosetta, to like manually design the display of the epitope. And that just doesn't work. You're not able to elicit the antibodies that you want, right? And so in theory, oh look, there are these conserved regions, let's elicit something.

In T cell responses it kind of works, so you can get a T cell response against these conserved epitopes. But typically a T cell response is not sufficient to actually give you a protective response against a pathogen. In fact, maybe one exception, none of the infectious disease vaccines that we know work, right, that are licensed operate primarily on the basis of a T cell response. Eventually, it's the antibodies, which, you know, in some sense are early enough to be able to clear the pathogen. A T cell response only like a, say, a CD8 T cell response, definitionally only work once your cell has been infected and then it can clear the cell, which in many ways is too late.

It's the antibodies that can prevent the pathogen from actually entering the cell in the first place.

[01:19:26] Abhi: When, I imagine like the average immunologist does think a lot about conserved epitopes and like comes at it from a very structural biology perspective. Does PopVax ever consider like structural biology, like when you guys are redesigning stuff, or is it very like hands off, black box approach?

[01:19:41] Soham: So like vaccinologists do think about epitopes. Yeah, yeah, yeah. but no, I think it's an important distinction because there's lots of immunologists.

Not that many vaccinologists, right? People actually practically working on the problem of better vaccines, fewer than you would think. Lots of, like, people working on, like, basic science immunology, systems immunology, increasingly. People actually working on practical, like, nuts and bolts vaccine design, not so many, right?

Especially ones with innovative strategies, right? The kind of next, sorry, can you repeat what was the second half of your question?

[01:20:10] Abhi: Like, are the, PopVax is coming at it a very like, like, like black box.

[01:20:15] Soham: Yeah. So we use a lot of structural biology.

[01:20:17] Abhi: Really?

[01:20:17] Soham: So the guy who runs our immunogen design team, co-runs it with me , as a structural biologist by training, not originally machine learning person.

And so, you know, yeah, as I said, our initial approaches were very kind of Rosetta -y structural biology approaches to designing this display mechanism for immunogens, and then redesigning again, using a structural biology approach. How, the, the RBD immunogen in particular for SARS CoV 2, to try and sort of,

to, to sort of better display certain epitopes on it to, to kind of close off certain other epitopes and try and get some of these broader antibody responses we want. And that was all very structural biology influenced. We do a lot of molecular dynamics simulations and stuff, as I told you. The black box approaches we've been able to do more recently, A, because the machine learning models for protein design have gotten better and B, because we've had the money to be able to actually scale up data collection to be able to, to build this feedback loop, right?

But when we started, it was a lot of kind of more manual structural biology approaches.

[01:21:12] Hiring in India

---

[01:21:12] Abhi: And kind of like on this topic of like, creating this whole pipeline sounds incredibly challenging, like meeting GMP standards, building something in India. And as I mentioned earlier, India is very much not a place where a lot of biotech research companies come from.

And as a result, I imagine hiring people is incredibly challenging. But at the same time, the IITs of India almost certainly produce a number of really high quality engineers and scientists. How have you, how have you, like PopVax in general, tailored your interviews to discover these like diamonds in the rough?

And is there anything counterintuitive you've learned about identifying good talent that can do good biology research?

[01:21:48] Soham: So I think there's a lot of good talent in India. Let me say this straight out, right? Just in case anyone is confused about it, there is a lot of good biology research talent in India.

And if you are watching this, you, whoever you are, should fund those people. But, I think what, what we've learned is there is a lot of talent. There's a lot of noise. So there are a lot of people who have impressive looking CVs, whose actual research work is less impressive than they describe it to be, or in the worst case, they haven't really participated in or fully understood the research work as deeply as they're portraying, because there's a lot of top down hierarchical management of PhD students and of, you know, any researcher in India, I think more so than there is in the US or in other places.

So you will encounter researchers who have done cool looking work who like really don't understand the work, right? Because it's the idea of their advisor, their advisor told them what to do, et cetera, et cetera, right? What we've also found is people have impressive looking CVs who can't do wet lab work, right?

And so what we ended up doing to try and screen is we, we have a multi part interview process where we initially start with a casual kind of conversation where we have them, you know, they usually, when they're on the job market, folks in India and in the biosciences, they have some kind of presentation, which they show to people.

We tend to disrupt that a little bit, which is to say we ask a lot of specific questions, and then we sort of, you know, we don't let them proceed in their, their kind of memorized flow of what they intend to do, but rather dig deep into one or two specific things and see how deep they can really get in explaining what and why they did it.

I think a lot of people then sort of are not able really to explain. And so we, we kind of eliminate them from our, from our process. But even, you know, even from that initial screen, I think, we were able to eliminate 80 percent of people that we talked to, right? So we can, we have a large number of initial candidates that we interview for any position and maybe even hundreds, right?

But we can cut those down pretty quickly with like a 15 or 20 minute conversation, right? And we have a large enough team now that we can kind of split the load across people interviewing, right? Then after that, we typically do another, kind of deep dive into some specific piece of work that they've done.

And then we have them come into our lab, and actually do wet lab work. So, especially people who are on our wet lab teams, we, we have all of them, no matter how senior, we make them come and do some assay, right? That we give them the protocol for, and we kind of walk them through the reagents. For people who are in analytical, who do HPLC work, we tell them, look, here's the molecule you have to identify, tell us what reagents you want and we'll procure them for you so you can run this on our machine, right?

[01:24:28] Abhi: I've had friends in the U.S. In wet labs wish they could do something like that, but they always say like it's too expensive to like scale up across many different people. Is it just like the cost is so low in India that it's possible to do stuff like that?

[01:24:40] Soham: Yeah, like the cost of, you know, the cost of, of, of getting somebody in to do this is it's not that high, and people seem willing to do it. You know, we pay for their travel and so on and so forth. But we don't compensate them for their time, which is something we tell them up front, and, you know, they seem willing to do it.

And we, you know, the reagents I think are cheaper than the cost of hiring someone wrong.

[01:25:03] Abhi: Yeah, that's fair.

[01:25:04] Soham: So we've already eliminated most people in the pipeline by the time, by the time we get them to this point. and we find that people are actually excited by the activity, too. Like, you know, many of them have not, they're in some kind of interview process where they're talking to a lot of people, but they not have the opportunity to actually get into the lab and show their skills off.

So the people we want at PopVax are people who are excited to be in the wet lab, right? Who, you know, they have ideas and they can, you know, they can come up with, new designs or they can come up with new approaches to our assays and so on and so forth, but they also want to go and execute those things.

They don't just want to talk about them. And we find that those people are jumping, you know, at the opportunity to do interesting work in a well equipped wet lab and to talk to people about the work they've done. When people come in and they're reluctant to jump into the wet lab, we typically, even if they're senior, even if they have like impressive CVs, we typically take that as a negative signal.

We say, we don't want to hire this person. Right?

[01:25:56] Abhi: Like you've mentioned in the past, like there's like this, the talent in India is like largely wasting away at companies that are not doing R&Dwork. Yes. Is there, are there, like, like particular institutions that like if you see on someone's resume, you're like, oh, they're probably like bored out of their minds there and are very talented.

They should come work at PopVax.

[01:26:14] Soham: Yeah, like if there are people who are at IISc or NCBS or CCMB which are all like, you know, central government, you know major institutions that do basic and translational science work and then we see that they're at a contract research organization and there are a bunch of those in Hyderabad and in Bangalore, where they're just doing the same assay over and over and over and over and over again, or they are at a contract research organization for synthesis where they're just doing chemical synthesis, the client tells them what to do, they don't have any input into the design process. And we see somebody from one of these top institutions, then yes, I think that's a good signal to us that maybe we can pry them away. If we offer them interesting work. And I think that's another advantage for us in India. Hiring is hard, right?

Hiring is hard everywhere. I think canonically it's a hard problem, right? I don't know how it is for you folks at Dyno or like, you know, your colleagues, but, yeah. I think because we are one of the few people in India doing innovative R&Din the biosciences, that's also translational enough that it's really close to being applied.

Like we're in clinic pretty soon from, you know, not, not immediately, but like we'll be in clinic next quarter. Right. We, we have this pipeline of six different new vaccines. We want to take to clinic in the next two years. And so that means if you come in and you do something that has a chance of affecting something that'll be injected into somebody not too long from now.

Right. What we find is as a result, we can motivate some of the best people in the country, to come work with us. Whereas in other places in the world, they would have many options to choose from. And so I think in that way, we can agglomerate a lot of the best talent, right. And become a beacon for the best talent to come and collect.

And not just within India, but talent from outside of India that are Indian origin, but want to return to India, whether for family reasons or, you know.

[01:27:50] Abhi: I was, I was going to ask that as like PopVax have their own version of that, like Chinese policy to bring minds back home, where you guys like do reach out, to a really promising scientist like abroad and like try to get them to come back.

[01:28:00] Soham: Yes. So I spent a lot of time on LinkedIn going like, hey man, like, would you ever consider coming back to India? And you know, the, the two folks you met today from PopVax, my colleagues, Maunish and Darshit are both people, who grew up in India.

Went to undergrad in India, did grad school, did their PhDs outside of India here in the U.S. and Canada. and then, you know, I convinced them, and, I think to some extent they convinced themselves. But also, I was able to convince them and our team and our work was able to convince them to come back to India and work with us at PopVax.

And we have, you know, but, a dozen more people who fit that description who are people who we have convinced to come back to India from other places to work with us, mostly people who wanted to come back to India already had a bias towards that, but that we were able to show, look, we're doing good enough work that you wouldn't waste away if you came and worked here, you would actually be doing very exciting things. And so that has been a big advantage for us, because again, as I said, not too many people in India offer that opportunity. So if you're somebody with a bias to coming back to India, and you want to work on bio, you should work at PopVax, and you should email me at soham@popvax.com.

But, but also I think, you know, they, we are kind of an obvious option, right? As more people hear about us, and as I do more things like this, I think more and more people will reach out to us as they already continue to. So when I was in Boston last time in December and I'll be back in Boston tomorrow.

I had a meet up with about 15 people who you know, on LinkedIn, had reached out and said, I'm interested in coming and potentially working at PopVax. And these are people working at extremely well funded biotech startups in Boston, at Big Pharma, Merck, you know, like Sanofi, et cetera, et cetera. and so I think that, that is a very interesting opportunity that we have that other people can have if they come and start those companies, but they haven't yet.

Right. And so that gives us the ability to recruit maybe a more, substantial density of talent than if we were in an ecosystem, like in Cambridge, where there's so many different competing companies, right?

[01:29:53] Abhi: Like, at least in China now, it feels like over the last few years, there's been a vibe shift of like smart students wanting to stay in China and just like work at like, like Baidu, Deepseek, whatever, interesting companies are in China.

Do you think a similar like that, similar thing like that is happening in India currently, or we're still like a few years away?

[01:30:08] Soham: think this in a limited way. in software, you know, in Bangalore, to some extent, in bio, it's still a lot of people wanting to go abroad because they're just better resourced, better resourced labs, better resourced companies, and they get to do more interesting work.

But I think people would love to come back to India, if given the right opportunity. I think lots of people feel that they're culturally out of place in other countries. Increasingly, they feel there's anti immigrant sentiment. In Canada, I know people who feel that there's rising anti Indian sentiment in particular in the last few years.

So I think lots of people would love to come back, given the opportunity, but they don't feel the right opportunities exist, especially in bio. So I think that's something that in software has changed a little bit. I think in bio too, if we can do a good job of bringing together a coalition of companies who are doing interesting work like this, I think it will change.

[01:30:54] Abhi: I'm curious, outside of like PopVax, do you see any other like really like great R&D effort going on within bio or outside of bio? In India.

[01:31:03] Soham: There are a couple of very interesting antimicrobial efforts like in, in antibiotics. There's a company called Bugworks run by a guy called Anand Anandkumar in Bangalore, which has raised a bunch of funding is doing interesting work in novel antibiotics. There was recently an antibiotic approved in the US. A novel antibiotic which was actually developed in India by a company called Orchid Pharma. Which is like a Chennai based company, boostrapped, that has been doing, and now they're bought, but they're doing really interesting work for many years on new antibiotic development. A company called Wockhardt, which is, an Indian generics player, also has developed a new antibiotic.

So there's some interesting work in that field. In novel biologics, less so, there is, a company called Immunoact, and then another company called Immuneal, both of which are doing CAR-T. And, and largely they've been focused on the, like, cost angles to make it cheaper for folks in India and other rest of world geographies, as they're called, ROW.

But ImmunoAct actually has an interesting technology where it's sort of a more immunized, sorry, more humanized, kind of de immunized, cell that they're using than, than existing CAR T. And so it has a cleaner immune profile, they claim. And I think some of the data shows than existing approved, CAR T therapies.

And so I think that's an interesting, that's actually a novel product. I think it's maybe licensed off NIH, but I think it's a novel product that they've developed and they've got approved in India. But I don't know that they have the intention of bringing it to the U.S. Or bring it to, to, to richer countries, right?

So, I think there's like some beginnings of interesting stuff happening. But I also wish these companies were more ambitious about saying, okay, my product is going to be best in the world. And so therefore I'm going to get it approved, not just in India, but in countries where there's a lot more money to be made, because that can fuel that money will fuel your R&D engine in the future, right?

Cause that's, you know, if you want to make money off biologics, you need to be approved in the U S right?

[01:32:51] Abhi: Yeah. It's like, and that's actually kind of related to a question I was just about to ask. PopVax is starting your phase one trials that are going to be ran by the FDA using

[01:32:59] Soham: No no. NIH.

[01:33:02] Abhi: And that does lead to a question like why focus on appealing like purely to like US based markets or like US based like federal regulatory agencies. Is this, is that just like more valuable for you guys?

Does India have issues with approval of novel biologics or something else entirely?

[01:33:19] Soham: So I think, a few different things. One is obviously it gives us a lot of credibility to get phase one data in the U S. I think people still ask questions about the credibility of clinical trial data in India, which is unfortunate.

And there are some reasons for it, but we want to be able to say, look, you know, this data is unimpeachable. And if NIH generates the data, it's pretty unimpeachable, right? So that was a very good opportunity for us. And you can't really beat free, right. You know? But I think the other reason is, yes. the Indian regulator is undergoing, I think, a process of evolution.

I think at the moment they are harder to get approval, for new phase one trials in India than it is in the U.S. And often they have more, they object in ways that, you know, they ask for animal data that takes a long time to generate, that isn't necessarily consistent with what current sort of modern standards would say.

So a good example of this is, we've not done NHP studies, you know, primate studies for our vaccine. That's because FDA told us we didn't have to. But to do the phase one in India, it's almost certain that we would have to do NHP studies. Even though it is extremely difficult now and expensive to do NHP studies in India because of their sort of other animal regulation issues that have made that difficult.

And so in fact, it's actually easier for us to get a slot in an NHP lab in the U.S. In terms of just time than it is to get one in India.

[01:34:32] Abhi: Is this unique to like kind of like the novel modality you guys are approaching or like even like small molecule drugs?

[01:34:37] Soham: I think it's worse for the novel modalities. It's like biologics.

[01:34:40] Abhi: Really?

[01:34:40] Soham: It's much worse. And so there's a lot of education of the regulator that has to happen because they haven't seen this movie before. Whereas the FDA has seen this movie before. They've seen the sequel that you haven't seen yet because they've, you know, they've previewed all this data. And so they can actually, I think, be very astute.

And that's the other reason the FDA, I think, is very valuable is we did a pre IND with FDA, where we basically wrote them a bunch of questions and they give us a bunch of really detailed answers. And in those answers, implicitly, I think was the accreted knowledge of seeing a lot of mRNA trials and really understanding where we should be looking out for safety risk, which is what you care about in phase one.

Whereas in India, there have been, there's been one mRNA vaccine that has gone through trials, right? Which is quite different from ours in many ways. And, the regulator just doesn't have that knowledge yet. It will, they will. I mean, they're smart people, right? So I think what we hope for, and I think this is one of my other policy prescriptions for India is, like China has done, we were talking about this briefly earlier, China has made it much easier for people to do Phase I's.

Maybe made it too easy for drugs to be approved, some people would say, but I think Phase I's should be really easy to get approved. Australia has this really cool regulatory framework where they don't even, you don't have to apply to the regulator, you can like, have these sort of, decentralized, ethics committees basically make the call for a phase one.

I think some system like that in India would be a huge boost to biotech research and would really marry our, similar to China, like marry our ability to do really cheap trials, with the ability to do really fast trials, especially the early stage, and develop data for assets that we could then, you know, whether it's actually get them approved in the U. S. or market them to, you know, to, to U.S. big pharma companies for collaboration and so on and so forth, as Chinese companies are doing. If the Indian government doesn't want to do that across India, they could do that in a in an SEZ, or they could do that in like some kind of special zone where every like, you know, historically, the problem in India has been, and I empathize with this, has been, you don't want people from other countries coming in and using our people as cheap guinea pigs.

And I feel this, I don't think we should be doing that, right? I don't think that's what we're trying to do, right? We're an Indian company trying to develop new drugs that will help India, you know, develop new vaccines that will be valuable for infectious diseases in India, right? But, I understand the concern.

And so I think therefore having some kind of special zone maybe where when you go in, you basically sign a, a series of documents that explain that, you know, you have maybe more education than average. You have sort of more ability to understand what's going on than the average person in India. And so therefore you're, you know, kind of a de risked population in which to, to, to do this, right?

[01:37:09] Abhi: Do these, like, do these SEZ's exist? Within India for any other industry?

[01:37:12] Soham: There have been talk about doing, for example, medical SEZs for medical tourism, where the laws are somewhat different, would be somewhat different in terms of qualifications and so on.

I don't think it's happened yet, but I think that the framework certainly exists to do something like this, where there are SEZs for, there are SEZs that have very different regulations, like GIFT city in India for, for financial, you know, and, and, and sort of, you know, foreign exchange purposes. So there's no reason that you couldn't do this as well, right?

And so I think that if, if you did that, you could build in India, a biotech research ecosystem end to end where it's like, we can do the early stage work, you know, the discovery, we can do the phase ones, right. And, you know, potentially even later stage trials where you can generate. all this really good data, and the Indian population is quite diverse, so in a quite diverse population of whether these, you know, novel drugs that you're designing work, right?

And that becomes a very then compelling case to be able to take those cost structure advantages and use them to just get faster and faster at developing the best biologics, right? And that's what we're doing at PopVax to some extent. It's 10 times cheaper for us to do a lot of the wet lab work. It's 10 times cheaper for us to do animal work.

And so that means for the same dollar, I can do either 10 programs or I can potentially go 10 times as fast on the same program, right? And eventually, we will just leapfrog what a company like Moderna is able to do because, you know, they can't keep losing the billions of dollars that they're losing forever.

If I'm just way more efficient, I think I'm a better bet. I'm a better investment, in terms of being able to generate good outcomes for for new vaccines and from a philanthropic perspective as well, right? You know, we work with our philanthropic partners.

[01:38:48] Abhi: Like dollars are better well spent here.

[01:38:49] Soham: Exactly. Yeah, dollars per life saved. I think, you know, if you even if you assume we have the same hit rate as Madonna, why would you give them 10 X the number of dollars to develop the same drug? And in fact, I think we will have a better hit rate, right? So, like, you know, I could be wrong about that, but I don't think I'm wrong about the cost structure.

[01:39:05] Abhi: Mm hmm.

Yeah, I mean, that feels like empirically true as the last three and a half years I've demonstrated.

[01:39:09] Soham: And I think in particular, right, why Big Pharma is pulling back from vaccine investment. We talked about this briefly. One of the reasons is it's a billion dollar program and it won't generate as much money as a cancer immunotherapy or a GLP 1 because it's only dosed once and it, you know, it costs like a hundred dollars or something, right?

Which is why, by the way, anti vax people who say like, you know, vaccines, oh, pharma companies just trying to make money off vaccines, pharma companies don't want to make vaccines. they, they want to not invest in vaccines. But I think with our cost structure, if I can make a program cost one fourth or one fifth as much, we'll use, I still end up spending a lot of money on late stage clinical trials in the U.S. or whatever. But maybe you do that after you've already got approval in another country for a completely novel vaccine. And so now that's de risked to get that next round of capital to do that. I think I can do a program at one fourth the cost, right? As a big pharma company can do that. And if I'm also better at getting a more efficacious product through the door, if my hit rate is higher, right?

I think that combination makes vaccines not just investable, but a highly, highly profitable endeavor. Because the best vaccines still make, you know, two plus, in some cases, eight billion dollars a year, right? And so if we can be a company that makes the next generation of the best vaccines, then and just because we have a repeatable engine to be able to do this cheaper and faster, you know, we can be a pharma company that's the size of any global big pharma company, but then because we have the same amount of money as they do, and we have access to this cheaper R&D base, I can suddenly do 10 times the amount of R&D.

And if this keeps compounding, you know, eventually it'll be a, I think a huge win condition for the world and for us in India and biotech R&D.

[01:40:44] How fundraising for an Indian vaccine design startup is coming along

---

[01:40:44] Abhi: On the subject of like funding at all, I know that, when, PopVax, first started, I think you guys primarily survived off of the Gates Foundation money.

I think historically, as far as I can tell, PopVax hasn't ever received money from an Indian organization.

[01:41:01] Soham: No.

[01:41:02] Abhi: Is that, yeah, like, why has there been, like, such hesitation, like, over three, three and a half years?

[01:41:07] Soham: We haven't looked that hard recently, so I can't say recently what the, the outcome would have been.

The government funding available to us was too small.

[01:41:15] Abhi: Gotcha.

[01:41:16] Soham: So it was, like, 50,000, sorry, 50 lakh rupees, which would be the equivalent of somewhere between, like, somewhere in the 60,000 dollar range. It was sort of the prototypical grant that you get at the early stage from BIRAC, which is this organization I'm talking about.

It's not nothing for, you know, doing work in India, and it's good that they give that out. But yeah, it wouldn't have moved the needle for us. Beyond that, it was difficult to find funding from the government that didn't require that we were at a TRL, like a Technology Readiness Level that was much higher than we were actually at.

Right. So like, which is basically like, you need to be in phase one or you need to be sort of like already very close to phase one, which we weren't three years ago. Right. Private funding in India, we did try and we were very disappointed. Like we heard a lot of people tell us either it was impossible. or we had people tell us that, you know, the government should be funding this.

I heard a lot from like billionaires. I was like, you, you live here, like, you know, you know that the government doesn't fund this. Or at least not as much as governments in other countries do. And you know, perhaps there's an opportunity for that to change. And maybe there are people working on this.

You know, I know there are lots of smart people who have tried to push policy in the direction of funding more R&D, but it hasn't happened yet, right? And then Indian venture capitalists, we didn't, again, pursue very seriously, but the early conversations we had just to understand, whether this would be worth pursuing, were that because they did not have biotech folks, they would not be able to sort of, you know, assess the technology in a way that would allow them to lead a big round.

I think if we had been smaller and we needed less money when we went out to raise, it would have been possible to get from some small funds that now think about biotech, like small amounts of money, like a million, two million. But you know what we need is more than that. So when we when we started raising recently, it's a Series A that's much larger than that.

And so that, I don't think, is yet something that these funds are like the funds that have the money in India are able to do. So, yeah, we just. You know, we tried, we, we didn't get money from those people. We got money from other people. So most of our money, in the very beginning was Gates money, for this, for our platform and for our next generation COVID vaccine, where the intention was basically, make it such that it's able to be broader and it's able to protect against, multiple different variants that might emerge rather than sort of becoming useless after a few mutations, right?

And also to try and make the vaccine more thermostable so that it could be distributed in developing countries, right? Then after that, the next big chunk of money was from Vitalik Buterin, who's the Ethereum cryptocurrency co founder, who's very interested in biosecurity, and in sort of open source, public health work, where like, how can we make these things really accessible in a way that people can not only buy them, but also sort of develop further on top of this.

And so our COVID vaccine, which is funded largely by his group, Balvi, is actually open source. So we're going to put out all the information about how to make it. We're going to allow people to, to make changes to it and, and sort of make new versions of it without enforcing our intellectual property on that specific vaccine.

Though we'll be able to use our IP for other things, other vaccines that we're producing, right? And so that was also a weird funding mechanism that I said, you know,

[01:44:31] Abhi: I mean, alongside like the, like how like the financial barriers in india are just so much lower for vaccine development compared to everywhere else. You have a few articles published on your Substack about some of the results you guys have had over the past year or so and they feel like clearly extraordinary when you look at them.

[01:44:46] Soham: Thank you.

[01:44:47] Abhi: You have an influenza vaccine that has better IgG and HAI titers compared to your competitors.

[01:44:51] Soham: 250x better IgG titer is the headline result for H1.

[01:44:55] Abhi: Which is insane and those antibodies still remain effective against mutated influenza strains?

[01:44:59] Soham: Yeah, so this is pre clinically, so this is all in mice to be clear, but like, yeah, so we, we found our influenza result is that, our version 3 influenza seasonal vaccine construct, it's not actually version 3, it's many more versions than that, but sort of, you know, major version 3, is 250x better in terms of H1N1 IgG titer elicitation and also way better for, for H3N2, those two of the key seasonal strains in the vaccine, like more than 100x better, but also is able to, yes, as you said, elicit robust IgG titer against H5, H5N1, even though it does not contain, the H5 antigen.

[01:45:29] Abhi: Yeah.

And so like, like looking at this on face value, you've made something clearly that people want, and it seems clearly better in every capacity compared to your competitors. Yet, you've told me that fundraising is a continuous challenge. Is this just because like vaccines are like a, like a hard deal for anyone trying to work in this space, or do you think there's like some level of like, like VC education or like government education that needs to be done?

[01:45:50] Soham: So two things. One I'll say is since we spoke last, the last week of, of investor meetings have been, have, have, I think we've turned a corner, I think, I think in part because of this data, right? I think we will be able to, you know, I think it's now materially easier for us to raise money. And it's because, you know, as it should be, because we have really good data.

And we also won this 2 million BARDA award, as part of their Patch Forward Prize, in part based on this influenza vaccine data. And so BARDA, which is the U.S. Biomedical, sort of the biodefense agency in some sense, but in HHS, is, you know, both a potential funder of future influenza programs for us and a potential customer of pandemic influenza vaccines.

And so I think that creates a setup in which people are much more willing to take seriously the idea that, actually, these guys might be in the future, one of the best influenza vaccines, right? But I also think we are at the intersection of three different things that people don't understand. One is, we're an Indian company doing novel biotech R&D.

It doesn't exist.

[01:46:46] Abhi: Yeah, I talked to multiple biotech VCs in preparation for this, and none of them had helpful advice to give because they've never met, like, a biotech founder or biotech, like, interesting biotech company that came from India.

[01:46:57] Soham: How, how skeptical were they that it's possible?

[01:46:59] Abhi: Like, they were all actually like pretty curious.

[01:47:02] Soham: Oh, really?

[01:47:02] Abhi: Yeah, like none of them like wrote it off immediately. They were all like, yeah, like I've never thought about it, but I've never met one.

[01:47:07] Soham: Yeah, but they've never had one, right? So clearly I'm not wrong that like almost no one does this. Yeah, yeah. the second thing I think that, and so I think part of that is like, is there, is there a talent gap where you just can't find the talent in India?

And the second question is like, is the data reliable? Because the China had all these data issues and we have data issues in India and so on. And I think the answer to all those questions, it can be done. You just have to be very careful about how you do it, right? The second issue is, yes, vaccines, people for many years have not invested as much in vaccines. COVID was this big spike in vaccine investment. but it felt, it's like a bit of a sugar rush. Like after that, it's like the, the fall has been precipitous. and right now, Moderna, is being hammered for being an infectious disease, Pfizer is being hammered for being an infectious disease company. And I think that's because people think historically, these vaccine investment cycles are very long and, you know, probability of failure is high is the typical thought process. And obviously we think we're different in that regard, right?

But that's also some education that has to be done and some understanding of what the new cost structure looks like. And the fact that we think we can be much more likely to succeed than existing vaccine design approaches. The, the third thing that is a. you know, as an issue is that the biotech funding market as a whole, as you likely know, has been a bit bad, right?

[01:48:16] Abhi: It does seem to be getting a little bit better recently.

[01:48:18] Soham: A little bit of a new dawn, it does seem like. But you know, for two years, it was quite, it was like really as bad as anyone had ever seen it, right? I think, and public biotech stocks are still not doing so hot, right? In many ways.

So I think the intersection of all those three things made it a little bit difficult for us to fundraise. But I think the main barrier that has now been crossed is, in the very beginning, right? If you met me when I started this company, I think a lot of people believe that the probability of success was not high, right?

Because I didn't know any biology, right? And, you know, I was starting this extremely technical biology company in a part of the world where people didn't know a priori that there was really good bio talent, right? So that was, you know, the initial piece of skepticism. I think now I know enough biology and I have a wonderful enough team around me that people are not as skeptical, that I am the right leader for this and that my team is the right team for this, right?

The other big thing I think was, like, can you show that your vaccines are actually much better than what existing vaccines are, you know, are able to do in mice? And I think that's something where, starting with our COVID data, like, you know, about a year and a half ago, and now with this influenza data, we've been able to show repeatedly, oh, we can do this much better, at least preclinically, than, you know, then companies that are much better funded than us that have spent much more time on this problem than us.

And then I think the U.S. Government coming in, right, so like NIH, as part of Project NextGen, being willing to do our COVID vaccine trial, and then BARDA now giving us this award shows that I think, you know, in some sense, the people that know best about this kind of stuff, the scientists that are, you know, I think the alignment of their motivation is very clearly, like, they want better vaccines, right?

If these people are putting some support behind this, then I think that has sort of woken up, a lot of investors and also partners, like, you know, commercial partners, that have reached out to us, like, big, big pharma companies who are like, oh, this is real now, like, clearly, like very smart people have vetted this and have spent some time thinking about whether these would be valuable sort of new vaccines or new vaccine designs for, for the United States, which is the market which makes the most money and the answer has come out: yes. Right. And so I think that was the last piece and sort of giving us the credibility needed for us to say, like, look, this whole thing may seem super weird to you, right? Like the construction of this company and where the money has come from and like who started it and where it is. It may all seem super strange, but the data can't lie, right, you know, at least preclinically, the data is quite repeatable and we can show it to people.

Right. Whether it translates to humans, we'll have to see, but, you know, we're really doing something that's quite different here. And lots of smart people, you know, in the U S government, who are, you know, motivated to get better vaccines are saying that, hey, PopVax is maybe one path to better vaccines for, for Americans.

Right. And I think that has been the, the, the validation that has, you know, made other people jump in and be like, oh, okay, this might be a real thing. Let's jump in and maybe fund or maybe partner or whatever.

[01:51:08] Abhi: Okay. So JPM was a big mental update.

[01:51:11] Soham: Yeah. Yeah. Well we'll see, I think still vaccine funding is such a small subset, like, okay, if I'm a cancer immunotherapy company and we may do cancer, you know, immunotherapy with many of the same ideas that we're doing for vaccines today.

Right. This idea of an immunology foundation model, or at least immunology models that can model elicitation, is also, you can see, sort of intuitively valuable for cancer immunotherapy, right? We're not doing that as our main thing today, right? If I'm at JPM and I'm a cancer immunotherapy company, then I'm the belle of the ball and everybody loves me, right?

Or if I'm a GLP 1 company or, you know, I'm doing something in metabolism, right? With vaccines, there's a small number of investors and partners. It's not all the big pharma companies. It's not all the investors. But I think the good thing is those people, the people who do vaccines, they really understand vaccines.

You know, there are people who've worked on big vaccine programs, who've worked on vaccine research. And if you're committed to vaccines, then typically these funders have built actually big teams, at least, you know, large enough teams, that really understand vaccines scientifically and understand the market.

And so we can have very, you know, high level, sorry, I don't mean high level, I mean rather detailed, you know, and sort of very granular conversations with them about the data and about our approaches, and they really understand those things, right? And so, I'm happy to work with those people, and we'll see which of those people we end up working with in the long term, right?

But it's certainly not most biotech investors, right? Like, most biotech investors. Like, we went to a party at JPM, which one of, you know, with, with one fund that is interested in biotech that invited us to it. And, you know, my colleague Manish was asking people like, hey, you know, would you be interested in talking to us?

We work on vaccines. Just by the by, right? While getting drinks. And, you know, they would be very interested when we said machine learning. And then as soon as we said for vaccines, they were like, Oh, we don't care. Our investment thesis is that we don't invest in vaccines, yeah. Or we're like, we were burned investing in this vaccine company, so we just won't do it again.

[01:53:10] Abhi: Is it because like, like largely because they don't understand the cost advantage of like the fact that you guys are in India? Like, like the terms are a little bit different.

[01:53:18] Soham: The conversation doesn't even go that far. It's like immediate right now.

[01:53:21] Abhi: Okay.

[01:53:23] Soham: I think if we show we can take programs to like phase two, phase three at that cost structure advantage.

Then I think the game will open up. But right now we need true believers. I think in like vaccines, we want to work with people. We believe in vaccines. We want to work with people who believe in vaccines. And I think one big learning for me is like, I don't want to spend my life right now convincing people who really don't believe in vaccines to believe in vaccines.

And then back us because they won't be good partners for us in the long run. Right. Tomorrow, let the like, you know, the johnny come lately is, you know, when in the future when I need way more money, you know, we'd be happy to welcome them in. But at this early moment, because we've not taken any equity capital, right?

So like, I don't have any investors on my board or anything, right? The people that I want to partner with, I want to partner with people who really care about vaccines. And they don't have to be people who have invested in vaccines before, but people like understand that vaccines are a valuable modality.

Both from a public health perspective and a financial perspective, and that there is a way, and people have made lots of money off vaccines, right? You know, there's a company called Vaxcyte right now, that you may have heard of, that is doing a competitive pneumococcal vaccine to Pfizer's 20 valent vaccine, they have a 31 valent vaccine.

And that, just on the basis of phase 2 data for that vaccine, they're at, like, a 10 billion valuation in the public markets, right? Because there's a clear market for that vaccine. They made something that's clearly better. I think what gets lost in the fog of this, is that vaccinology, because, you know, because it's hard and because people have taken these outdated approaches, lots of vaccine companies have failed, not because of the structure of the market being bad, but because it's a hard problem and their approach didn't work.

[01:54:58] Abhi: Yeah. Do you think they're approaching it perhaps in like a fundamentally wrong way?

[01:55:02] Soham: In a fundamentally wrong way, and these are, you know, they, they took too long in some sense to go validate their hypotheses, right? Whereas we want to be able to validate both preclinically and in clinic lots of our hypotheses very quickly.

[01:55:15] How is PopVax so good at designing vaccines?

---

[01:55:15] Abhi: And I think like, maybe, like, relatedly to how, like, other people are doing it incorrectly, and you guys have clearly, like, created something that's better than a lot of other people.

[01:55:22] Soham: Well, to be clear, we don't know if we're doing it correctly yet, until we see efficacy in humans.

[01:55:25] Abhi: Sure, sure.

Like, mouse, mouse efficacy seems to be quite good.

And on the, like, there's, like, enormously better immune response, like, to, like, 250x, like you just mentioned.

[01:55:33] Soham: It's much higher.

[01:55:34] Abhi: Yeah, yeah, yeah. What, what does PopVax do differently for your vaccine design? Whatever you do, why isn't, why isn't that there's, is it like specific to this, like precise immunogen design?

Is there something else?

[01:55:44] Soham: It's like, like, I think we take an engineering approach to the problem.

[01:55:47] Abhi: Okay.

[01:55:47] Soham: Right. Like I said, we do this, you know, mRNA encoded display on these self assembling particles. We do, this precision immunogen design. We do design of our own lipid nanoparticles. What we find is when we co design all those things together and we can just have a bigger combinatorial space that we can test empirically, which is our secret sauce, quote unquote, not very secret, right?

We can find parts of the design space that maybe haven't been tested before by other people that in combination you get these really good results.

[01:56:15] Abhi: Because other people don't have like, don't have like the knobs to tune on those because they're outsourcing it all?

[01:56:20] Soham: Yes, they don't have the knobs to tune and they don't have the wherewithal to test all this different stuff.

[01:56:24] Abhi: Gotcha.

[01:56:24] Soham: Right? And I have all the knobs, right? All the knobs are in my control. And so, and we find lots of wacky stuff, right? We find that there are certain lipid nanoparticle formulations that work better for certain vaccine designs. Mm hmm. Do I know why? No. Is my team trying to figure out why? Yeah, totally.

But does it matter? Right? It, as long as I can find the right set of knobs to tune that give me something that's better. And then, and critically, I can translate that into humans as well. And we'll have to figure out some translational process, using these organoid models and transgenic mice, potentially, which we're doing now.

And using human data we collect in phase ones, and we want to do a lot of phase ones as well, to be able to make that sort of translational gap narrower between the the models we're using now on actual humans. I don't need to, you know, I don't need to, mechanistically explain everything exactly that's going on, right?

I, because within the design space, there are going to be lots of little tweaks, which we don't know exactly why they're better. Right. But we will be able to see in practice that they are better. I think antibody design is kind of the same. I don't know. It's probably similar for you guys with AAV's as well.

[01:57:30] Abhi: I feel like it may be the case for all like AI biotechs were like they were primarily built by like people who have an engineering mindset as well. And they're, they are happy treating the problem as a black box. Like, yeah.

[01:57:40] Soham: And I would like to, once we have something that's better and I advanced that, at the same time, I want to figure out the mechanistic insight if I can use that to inform the next design.

But, I think we have to be extremely rigorous about being empirical and testing, you know, the, the biggest design space that I think will be useful for us, which is much bigger than other vaccine companies have done. But I don't think we have to be perfect about mechanism.

[01:58:04] Abhi: Yeah.

[01:58:04] Soham: I think lots of drugs, people like don't really know the mechanism like they say they do.

But I think a lot of like a lot of drugs like the mechanism is not fully understood right or until later. Yeah.

[01:58:14] Abhi: For, for vaccine. Is that true for vaccines as well? Or is vaccine. Like typically, vaccinologists really want the mechanism of action elucidated before.

[01:58:20] Soham: No, so like there are lots of vaccines like, HPV.

It's like

[01:58:24] Abhi: that mechanism is not fully like,

[01:58:26] Soham: Yeah. So like there's no, what's called a correlative of protection where like, if you get this number up, then you get protection. We don't know what that is for HPV vaccines, and the reason is very interesting. It's because HPV vaccines are too good, so for the strains in which HPV vaccines do work, and there's some space to be explored in making them broader, right?

But HPV vaccines, in the strains that they are designed for, right, in the genotypes rather, they're so effective, even actually in one dose. Even though they're originally a three dose regimen, that you don't get a gradient that you need to be able to say like, oh, you know, it's, it's only partially effective here.

It's more effective here. It's more effective here. And so it's actually, it's this variable that gives you the nice linearity, you know, between that variable and effectiveness.

[01:59:10] Abhi: It's just like a step change.

[01:59:11] Soham: It's just like, it works. And so because of that, we don't really know, whereas in the case of flu, Because the vaccines are often shitty, we do actually know that hemagglutination inhabitation titer and pseudovirus neutralization titer both actually independently correlate with protection.

[01:59:27] Abhi: Interesting.

[01:59:28] Soham: Right? Isn't that cool? So like, we don't really know. Is it effector function? Is it neutral? People think it's neutralization. It could be lots of different things. Even in influenza vaccines, effector function is not usually used as a correlate of protection. Good evidence that it actually matters quite a bit for final disease phenotype and severity, right?

[01:59:45] Pet theories on immune mechanisms

---

[01:59:45] Abhi: You mentioned, like, I think we at some point discussed about how, like, immunologists don't seem to, like, sufficiently update strongly enough on new, on new outcoming data.

[01:59:54] Soham: They do not, yes.

[01:59:55] Abhi: Are there, like, specific, specific theories about immune, like immune mechanisms for some disease that you like strongly buy into, like, like perhaps PopVax is not working on.

[02:00:08] Soham: Sure.

[02:00:08] Abhi: Like you would want like people to be aware that like this particular.

[02:00:11] Soham: So like TB, right?

Tuberculosis, which we do intend to work on.

[02:00:14] Abhi: Okay, yeah.

[02:00:15] Soham: We think antibodies play a big role. If you look at Babak Javid's work at UCSF, he's shown that there are functional antibodies against TB, that, that you can use to potentially clear the pathogen, right? Okay. Nobody's working on vaccines to elicit those antibodies for TB and it drives me insane.

[02:00:31] Abhi: What do they typically, like what do they typically work on?

[02:00:33] Soham: T cells.

[02:00:33] Abhi: T cells, okay.

[02:00:35] Soham: Yeah, 30 years of T cell vaccines for, for TB. And it's like, it's

[02:00:39] Abhi: Never worked, so.

[02:00:40] Soham: Yeah, like there's a study, the Gates Foundation and a bunch of others spending like almost 500, like 300, 500 million dollars, some crazy amount on a multi country study for this one, this, you know, relatively new TB vaccine.

And by new, I mean, it was designed 20 years ago. And, the hope is that it'll be 50 percent efficacious. And lots of people off the record have told me things like, oh, we don't think it's going to work, but at least it'll set up the trial infrastructure for the next shot. I'm like, what if there is no next shot?

I think in TB in particular is a wonderful illustration of this problem where I think there should be a portfolio of a bunch of different candidates that are trying to do different things, elicit different kinds of immune response, and we should do a whole bunch of phase 1s. Right? And see what seems to work well in phase one, and then make decisions about what to put 500 million into.

And by we, I mean the whole community, right? The scientific community, the public health community. In fact, what is happening is a bunch of candidates designed like 20 years ago, which happen to have phase one data are being rushed into these extraordinarily expensive phase two, phase threes, instead of testing a broad portfolio of strategies upfront.

[02:01:46] Abhi: Why do you think, I guess like we have like mildly touched on this in the past, but it is surprising that like over 30 years, no one like mentally updates, like maybe this doesn't work. Is there just like, you know, like what's the cultural problem that's causing this?

[02:01:58] Soham: I think the argument that they would make is like there's no good empirical evidence that antibodies would be sufficient or important.

And they're not wrong about that, but you know, like what you've been trying for a long time hasn't been working.

[02:02:09] Abhi: I'm curious, like why do you, why do you think antibodies are like the answer to this?

[02:02:13] Soham: Well, I, you know, I speak not primarily my own opinion here, but, you know, research that, I think two branches of research that have sort of convinced me that there's something to be investigated.

I'm not saying it's going to work, but we should try it, right, is my main point. One, we know for, we know that there are antibodies because people have done this really, like, Babak has done this work. We know there are antibodies, against TB, which you can get even from, from patients, right? That do something.

They're functional. They're able to like, you know, do something to, to the pathogen, right? Which is not something that people thought was possible for a long time because it's an intracellular pathogen, right? In many ways. We also know, that, all the, as I said, all vaccines basically that are currently approved that actually are efficacious seem to work primarily using an antibody based mechanism.

Why would TB be any different? Right? And again, intracellular pathogen, I get it. But maybe, if you have an antibody response, you can stop the pathogen before it's able to invade the cells and take up residence. Maybe, you know, there are some antigens which which end up on the cell surface, even off the, you know, even off the, the cell that's infected with the pathogen, right?

Maybe there, there's another mechanism, which is the TB could be latent, could be inside the cells, but it doesn't matter. Because if you have the antibodies, when it tries to come out and replicate, something kills it, right? Exactly. And that's good enough, right? So, I think it should be investigated, right?

But I think there's, there's a certain, and you know, I'm not an expert in this field. I haven't been in this field for 30 years, but it doesn't seem like anyone is working on this particular vaccine design approach. There may be a few people, right? And so it seems to me that that's under invested, whereas the approach that hasn't worked for a very long time is very over invested.

And you know, maybe, maybe a good analogy to this is like all this like Alzheimer's plaque stuff, right? Like, it seems like there are these, you know, these, these dogma type things that happen in scientific fields. And then people just don't update for a very long time, right? Even if there's no good evidence that your approach is working, they're deathly afraid of other people getting resources to try a different approach.

Maybe because they think it won't work, but perhaps more insidiously, maybe some people are afraid that it will.

[02:04:33] Abhi: It's, it's interesting that, because like my, my initial assumption with like why Alzheimer's research has been stagnant for so long is like, almost they had too much money to play around with.

And so like they like they had this hypothesis. This hypothesis had money already behind it. So they just kept pushing money toward, like they were happy to pursue the hypothesis because there was money there. And so I assumed like because there's not that much money in vaccinology, there'd be more emphasis on like creativity or like in just trying things.

It seems like that's still not the case.

[02:05:00] Soham: I think, like, kind of the way I model this, and I could be wrong, is, like, in fields where there's, like, a ton of money, you do end up with more creativity because there's some, like, marginal money that gets just thrown at random stuff. That's not the case in vaccinology.

There's less money, and so it just, it, it becomes really, like, focused on, I think the money also gets deployed in big chunks, rather than, like, okay, HIV vaccines, so crazy over invested in, okay? The drugs work really well, now there's this new thing that is like a long acting antiviral that works for like 365 days or something and like 99 percent efficacy.

HIV is the hardest vaccine design problem. But there's like a hundred times more investment, maybe more, maybe it's a thousand times in HIV vaccinology, there are whole international organizations, like IAVI, the International AIDS Vaccine Initiative, that exist only basically to make HIV vaccines, right, That are invested in hundreds of millions up to billions of dollars, whereas we're not investing that money in hep C and TB vaccines, in novel hep C and TB vaccines, even though, even hep C, forget TB, which kills a million plus people a year, right?

I love vaccinology, like I work on vaccinology, but it's just such a hard problem, and there's so many easier problems that we still haven't solved. Why don't we work our way up to it? Right? And I think at one point, there was a good argument. There were so many people dying of HIV. But people haven't updated to the basic reality that the pathogen just doesn't kill quite as many people anymore.

[02:06:21] Abhi: Yeah.

[02:06:22] Soham: Whereas like, these other pathogens, they really do.

[02:06:24] mRNA beyond infectious diseases

---

[02:06:24] Abhi: That makes sense. And kind of like, beyond infectious diseases as a whole, I remember like, during like, the heyday of Moderna, they had this whole pitch about like, we'll use this, like mRNA technology for a bunch of things beyond infectious diseases like

[02:06:37] Soham: And they're pivoting to be more of a cancer company now yeah .

[02:06:39] Abhi: And like given PopVax's platform capabilities, do you think about potentially expanding into these adjacent areas?

Partially because of the commercial viability or also partially because like the raw capability you think mRNA has.

[02:06:50] Soham: So, look, we're always going to be a vaccine company. We're not going to give up on vaccines because that's why I started the company. That's what we care about. And today we're a vaccine company. And I can say that with pride. We're a vaccine company, right? In five years, if things are going well, we should not just be a vaccine company.

We want to be a company that harnesses the full power of mRNA as a platform to test lots of different designs. Right, and protein design that is informed by in vivo immunology data or informed by at least, you know, extremely high quality in vitro immunology data to be able to design the best immunotherapies, whether that's vaccines for infectious diseases, vaccines for cancer, you know, mRNA encoded cancer protein therapeutics, et cetera. So I think cancer is the next obvious thing for us. I think autoimmune also is another space where autoimmune vaccines, you know, immunotherapies for autoimmune conditions, where we have a bunch of ideas, right?

[02:07:45] Abhi: Do you think Moderna's relative lack of success in those areas has more to do with Moderna rather than for there to be vaccines for cancer?

[02:07:57] Soham: I think Moderna is an exceptional company in many ways, and we stand on their shoulders. I think Moderna is not good at protein design.

[02:08:03] Abhi: Gotcha.

[02:08:03] Soham: And I think a lot of Moderna's problems come from Moderna being bad at protein design. So recently they had this RSVH MPV vaccine safety issue, if I'm, if I'm not mistaken.

And I think a lot of, you know, again, it's like antibody dependent enhancement of disease. I think a lot of these companies are not natively protein design companies. Moderna is a mRNA delivery company. Their job, in some sense, was to get mRNA as a delivery platform to work. And then they just put a bunch of, you know, antigens that are what you would expect into their vaccine programs, and a bunch of protein replacement therapies, and a bunch of, now, antibodies that they're trying.

And then, you know, T cell cancer vaccines, which had been a concept in peptides, but now encoded into mRNA, leveraging, like, the simplicity of manufacturing of that platform, right? But they're not really design companies.

[02:08:50] Abhi: So in a certain sense, they got kind of lucky that COVID 19 was relatively easy to design for?

[02:08:56] Soham: Well, yeah, and they didn't, it wasn't their design, right? It was Jason McClellan's design out of NIH, and then they got sued, and they had to pay him a, I mean, they'd pay NIH a bunch of money because they used that design. So yeah, they got, I think we all got lucky. We got lucky, but two things. One is that COVID is an easy pathogen to make vaccines for.

Lots of different types of vaccines work for COVID. Which is why it's kind of ironic that the big vaccine companies, your GSK, Sanofi, Merck, which are the biggest vaccine companies, all failed at making COVID vaccines, which maybe tells you a little bit about where those companies are and their ability to develop novel products, right?

But, we'd also, and by we, I really mean, again, Jason McClellan, and his group, had done structural biology work on SARS one and MERS, and they had invented this stabilization, in the pre fusion conformation for the spike protein. And so that was directly applied by Moderna, by BioNTech, Pfizer, and by Novavax to, to their vaccines.

[02:09:56] What would you do with $100 million dollars?

---

[02:09:56] Abhi: Gotcha.

And I think like this may be like the last question I have. If you were given $100 million for PopVax to spend on like vaccine development, maybe like arbitrary basic research you want done. What would you do with that money?

[02:10:12] Soham: Like a hundred million philanthropic dollars.

[02:10:14] Abhi: A hundred million philanthropic dollars and yeah, equity free.

You can do no strings attached. You can do whatever you want with that.

[02:10:18] Soham: That's great. I already have this plan. So it's, we call it the million lives mission where we want to save a million lives per year with vaccines, which we can do with, if we have successful vaccines against three new, three pathogens where there are no existing vaccines, HCV, TB, and strep A.

And philanthropic dollars are great for this because, especially TB is a pathogen where lots of people think you can't make long term profits. HCV, I think, will actually be profitable, but, you know, it's a non trivial development process, something we're already working on, but speeding that up would be great.

And strep A is another pathogen where we think there's a market, but there's also a lot of developing world impact, right? And these pathogens kill lots of people, right? Two million, something like that, two million plus people per year. If we make vaccines that are reasonably effective and that are reasonably well distributed across the world, we can save maybe a million lives per year with these vaccines.

So what I would do with that money is I would scale up this feedback loop approach that we have, this machine learning feedback loop, to design libraries of immunogens to elicit specific antibodies that we know to work against these pathogens. In the case of TB, we'd have to do a more basic science effort.

To find, you know, these antibodies in the first place. So we, there's some very interesting study designs where people have done, where you can basically go to places like Bombay, where I grew up or in South Africa, where there's lots of active TB and you go to people whose houses have somebody with active TB, but who don't themselves have active TB.

And then you sequence out antibodies from those people, for example. And you can, those people are often people have been exposed to TB, but somehow have protected themselves against it.

[02:11:47] Abhi: Are there currently no like international efforts to collect that sort of data?

[02:11:50] Soham: For antibodies, not so much. So for hep C, there are some people doing this, but again, I think not a sufficient scale, right?

And so I would scale up those efforts as well. Yeah, I think, you know, one very valuable public good would make like a public antibody database, like a PDB esque database, that is then classified by disease phenotype and exposure to disease. So like, this is the antibody of somebody who like, had very mild disease or like asymptomatic, but was surrounded by people who had this pathogen.

As an example.

[02:12:21] Abhi: Yeah, I mean like, this cleanly feels like a huge public good to have. Has anyone tried to push for it and it just doesn't like work out in practice or just no one has really tried to take it to the next level?

[02:12:31] Soham: I don't think at scale anybody has tried to do that. NIH had a grant, program, I think like a couple of years ago for something that resembled this, but at scale nobody's really done this, I'm quite sure.

[02:12:41] Abhi: Okay.

[02:12:42] Soham: Okay. Thank you. So it should be done. I proposed it as an FRO a while ago. This feels very FRO y. But like, yeah, yeah, so maybe, maybe the FRO people will listen to this and think it's a good idea now. But, but I, yeah, a hundred million dollars, you know, this is what I would do with that.

And so I'd basically do the effort to sequence out the antibodies for TB and Hep C, I would, you know, design for the antibodies I want to elicit. And, in the case of strep A, we want to, design for, the ability to basically prevent the pathogen from getting a grip, you know, in your, you know, in your nose or throat or, basically just, you know, prevent it from, from adhering to, to the cells in the first place.

So it can never really colonize, right? So, So, you know, we would basically design, to elicit antibodies that do those things. optimize in the feedback loop way that I told you and like really scale up the data points we could get using these, especially organoid models, which use human immune cells.

To be able to get maybe even tens of thousands of data points for each of these pathogens that go from immunogen design to elicited antibodies and, you know, both what is the sequence and what is the functionality, and use that to optimize to basically design vaccines to elicit the antibodies we want, which hopefully then we could take into humans, do phase one, phase two.

[02:14:05] Abhi: And the hope of this, like a data collection of people who have latent TB and have like potentially useful antibodies is that you have like a good, like you have a good idea of like what you want your elicited antibody to look like.

[02:14:19] Soham: Well, I mean, for TB, we would have to find what those would be, but for hep C, we have some idea of what we want them to look like.

[02:14:23] Abhi: Okay, but like the collection effort is so you can like for for TB, you can actually learn.

[02:14:27] Soham: Yes, we can learn that.

Yes, exactly right.

[02:14:28] Abhi: Yeah. All right.

[02:14:30] Soham: I think in 100 million, you could get those three vaccines through phase one.

[02:14:33] Abhi: Yeah, I mean, like, that sounds reasonable enough. especially in India.

[02:14:36] Soham: Yeah. So if you want to save 1 million lives per year and you have 100 million dollars

[02:14:43] Abhi: People will reach out to you. I think those are about all the questions I have.

Thank you so much for coming on today

[02:14:50] Soham: Thank you for doing this I really enjoyed it and I'm very excited to hear more podcasts that you do with much smarter people than me.

Introduction

In my first (real) podcast episode, I talk with Corin and Ari Wagen, two brothers who I met through my writing. They are building something super cool: a molecular simulation company called Rowan (which recently got into the Nat Friedman AI grant program). We discuss neural network potentials (NNP’s), whether dynamics are useful at all, the role of computational chemistry in drug design, what the future of the field looks like for molecular simulation, and a lot more. Also, for almost every paper/result discussed here, I attach the reference in the transcripts below!

If you work in molecular simulation, I recommend trying out their tool at rowansci.com. I’m not a chemist and cannot vouch for the tool personally, but I can vouch for how much I’d trust Corin and Ari to build something useful. Not a paid sponsorship, not anything I have an investment in, my opinion and no one elses, etc, etc, I just genuinely want their startup to succeed.

Finally, if you enjoyed this podcast and would want me to make more, please subscribe and consider upgrading to a paid subscription! Studio time is expensive, and, while Rowan generously agreed to cover the cost of filming, I don’t want to make that be a barrier to interviewing people. I have many ML-biology scientists I’d want to interview, specifically in the realm of cryo-EM, antibody engineering, and radiology, and it’s much easier to do that if filming is covered by this blog. I’ll probably still film stuff anyway, since I’m mentally bucketing this blog as ‘expensive hobby’, but money still helps!

Jargon explanation

This podcast is really meant to be consumed by people at least vaguely familiar with the molecular dynamics (MD) field. If you’re confused, I have written up a primer to MD here, which may be useful. Corin’s blog here is also incredible for understanding deeper nuances of the area. Rowan’s blog is also quite good, especially this one: the role of quantum chemistry in drug discovery.

Here is some breakdown of jargon used in the episode:

• ‘Molecular dynamics’ and ‘molecular simulation’: Technically, dynamics refers to time-dependent simulations, whereas simulation in general can also be time-independent. I use these interchangeably, partially because I’m not an expert, but also partially because even people in the field sometimes use them interchangeably.

• ‘Levels of theory’: This is equivalent to saying, ‘how close are we simulating the true nature of reality?’. In this case, the ‘truth’ is the solution of the Schrodinger equation, which is usually intractable to solve. Higher levels of theory means you’re closer to the equation (more accurate, including in quantum mechanics), lower levels means you’re further from it (potentially less accurate, usually only classical mechanics). We sometimes refer to ‘density functional theory’ or ‘coupled cluster theory’, both of these are at the higher end of the theory spectrum.

• ‘Tim’s tweet’: This is in reference to Timothy Duignan, a minor celebrity in the world of neural network potentials. He wrote what is likely the most famous thread in this niche field, where he showed a crystal nucleation event using a neural network potential that had never before seen nucleation events.

• ‘Periodic systems and molecular systems’: Rowan has a good article on this. But, just to recap:

  • Molecular [systems] are exactly what they sound like—isolated molecules or groups of molecules surrounded by a vacuum (or a dielectric field). This is good for studying small molecules, clusters, or even larger biomolecules….

  • However, most physically relevant materials are so large as to be effectively infinite relative to the molecular scale. Cutting out a chunk of these materials and modeling them with molecular calculation introduces significant edge effects. To solve this problem, we [model them as a periodic system].. Materials can be modeled using a single unit cell, where the molecule or group of molecules “sees” itself tiled infinitely in all dimensions.

Timestamps:

00:00 Introduction

01:19 Divide between classical and quantum simulation

03:48 What are NNP's actually learning?

06:02 What will NNP's fail on?

08:08 Short range and long range interactions in NNP's

10:23 Emergent behavior in NNP's

16:58 Enhanced sampling

18:16 Cultural distinctions in NNP's for life-sciences and material sciences

21:13 Gap between simulation and real-life

36:18 Benchmarking in NNP's

41:49 Is molecular dynamics actually useful?

53:14 Solvent effects

55:17 Quantum effects in large biomolecules

57:03 The legacy of DESRES and Anton

01:02:27 Unique value add of simulation data

01:06:34 NNP's in material science

01:13:57 The road to building NNP's

01:21:13 Building the SolidWorks of molecular simulation

01:30:05 Simulation workflows

01:41:06 The role of computational chemistry

01:44:06 The future of NNP's

01:51:23 Selling to scientists

02:01:41 What would you spend 200 million on?

Transcript:

[00:00:00] Introduction

Abhi: Today I'll be talking to Corin and Ari Wagen, two brothers who are co founders of Rowan, a quantum chemistry simulation startup.

Of note, Rowan was recently accepted into the Nat Friedman AI grant program. Congratulations. Past that, I believe Corin is one of the most interesting thinkers in the intersection of molecular dynamics and machine learning today. He also runs an incredible scientific blog, which I'll attach in the description of this video.

Thank you both for being on the show today.

Corin: Thanks for having us.

Abhi: So first question, just to set the tone for the rest of this podcast, give me a high level overview of what molecular dynamics is and what neural network potentials are.

Corin: So molecular dynamics, is a way that we can study the dynamics, how they evolve over time of molecules.

So a lot of calculations focus on taking a static molecule and, asking some question about it. Molecular dynamics also lets us sort of time integrate equations of motion. Basically, we can make videos of molecules moving around and learn things from them. Neural network potentials are a way to accurately predict a lot of things, but most relevantly for this energies and forces of molecules.

So it lets us get accuracy that's much closer to the truth to quantum mechanics. at a fraction of the cost that usually takes. It's a more accurate replacement for traditional force fields.

[00:01:19] Divide between classical and quantum simulation

Abhi: One, immediate question I had when I was learning about this field a few months ago was this, divide between classical mechanics and quantum mechanics. You wrote this post a while back called the two cultures of atomistic simulations. I'm curious, could you just recapitulate that for the audience?

Corin: Just briefly, it's useful to have some historical context for the field.

So the field of computational chemistry is about 100 years old. and back in the old early days of Heisenberg, people, did calculations on two atom molecules on pen and paper, with the advent of computers, I'd say in the seventies, the field grew in two different directions. So some people wanted to scale up the very rigorous, very physics based approaches, the quantum mechanics, to larger and larger systems to the limits of the hardware.

And so this grew into quantum chemistry, where now you can model up to a couple hundred atoms, with, very good accuracy, derived from first principles with all these layered approximations. And then the other half of the field said, Let's simulate the stuff we really care about, like DNA, like proteins, like these complex biological systems.

And let's work backwards from how fast things need to be, and invent a theory that's fast enough to do this. And this is essentially molecular mechanics. So the early, CHARMM and AMBER work, you're essentially using polynomials to fit quantum mechanics, and it you know, it works astonishingly well.

Like you could model protein movement. You can model like antibody, like motion sort of solution structure around things. and this was just, I think blew the field open in the late seventies and early eighties. And so now what you have is you have these two opposing paradigms were what we'll call classical molecular dynamics, more of the biological side of simulation has results that you can model the things you care about, but you get the wrong answers because the theory is wrong and the quantum mechanics side of things.

You get very accurate results, but on things which are less immediately relevant. And so I think a huge open challenge is how do we now, 50 years later, start to try to bridge this gap and, bring accurate simulations to the things we care about, which it seems like for the first time we're maybe finally able to do.

Abhi: And that's like the ultimate goal of the neural network potential is to get this like nice Pareto optimal frontier of both very fast and very accurate.

Corin: Yeah, I think that's exactly right. And there's, this is a goal a lot of people have had. So this is, the quantum computing people talk in very similar terms, I think neural network potentials are there.

Definitely right now look like the closest and by far the most promising way to do that. that's what we, and I'd say a majority of the field is really excited about.

[00:03:48] What are NNP's actually learning?

Abhi: Just to give some background context, a lot of neural network potentials are based on this. or like train on this approximation of the Schrodinger equation, density functional theory.

What I've always found a little bit interesting is what these neural network potentials are actually learning about physics. Is it, do you believe it's learning some lower dimensional manifold of the results of the Schrodinger equation? Do you think it's something else entirely?

I guess some other contexts is for protein structure models. The prevailing theory is that they're doing some sort of fuzzy homology search and then local energy minimization on top of that for neural network potentials, which what's actually going on inside there?

Corin: I think we don't really know, like we're just seeing some of the interpretability work come out on ESM2, using sparse auto encoders to try to understand what actual like features in the feature space correspond to.

And I don't think anyone's done anything like that on neural network potentials, although hopefully they will. But I think in trying to understand how it's possible, like how is it even possible to speed things up so much? I think the scope of things we care about in the context of life sciences, even just molecules that can exist on earth, is so much more restricted than the scope of all possible molecules. So quantum mechanics is almost too good. Like people do benchmarks where you put random elements in random places in space, and then you score a quantum mechanics methods based on how well they do relative to very high level, like non approximated methods.

So you can score approximations this way. It's the mindless benchmarking. And you can say, we're okay being bad at, a beryllium here, a radon here, a technetium here, and, a krypton here. we don't need to be good at that. If we just learn that 15 elements that are in the human body, in ways that would not immediately explode in contact with our atmosphere, we have such a low dimensional slice of like chemistry that we need to learn that we can give our models an inductive bias in that direction.

Abhi: Like there is no free lunch here. You are going to be like failing on some weird out of distribution space, but you're fine with that.

Corin: Yeah. I think maybe what the field hasn't appreciated and what I didn't appreciate until I started, this is just how vast the starting distribution for like chemistry, the field is for like any combination of atoms that like, even taking like a tiny slice of that well encompasses like everything we could care about.

[00:06:02] What will NNP's fail on?

Abhi: Another question I had was after AlphaFold2 was released, you saw this rush of papers claiming that AlphaFold2 failed on this like weird branch of kinases or globular proteins. There's a lot of immediate pessimism after some really interesting result from the field pops up and whether that pessimism is like actually, like real or not.

it just, it happens consistently every time. What do you think that will be for machine learned force fields?

Ari: I think the thing people are pointing to is these long range interactions.

I have one charged particle and it's 20 angstroms away from another charge particle and the model has a cutoff radius of 10 angstroms. And so the particles don't see each other. And It's as if they were, infinity angstroms apart in the energy of the model returns. And you're like, look, these neural network potentials aren't good for anything.

And I think people are working on a lot of different, charge handling schemes. but I think a question to ask whenever people have this pessimism is is this case that you're pointing out is like a, known failure, something that we're going to be trying to model? Does it matter and do we need to remedy it?

And I think with charge handling, that's still like a very open question.

Corin: Yeah, I think that's exactly right. And I think there's, it's going to be really important to figure out like, to what extent is failure predictable or non predictable, right? Because something that works 80 percent of the time is very useful if you know which 80 percent of the time.

So if bad for globular proteins, that you can just not use it for globular proteins. If there's this stochastic hallucination problem, I think that will be a much bigger issue. And we're seeing this, Ari's done a lot more benchmarking than me on the state of the art models, but there's some we see are really good for confirmations, there's some we see are really bad at that, some for thermochemistry.

And I think it'll be, we're very used to the approximations we already have in computational chemistry. We have an intuitive sense for what would be good or what would be bad. And we'll need to build up the exact same intuition for neural network potentials. which will just take time and practice and hard work.

[00:08:08] Short range and long range interactions in NNP's

Abhi: A lot of these neural network potentials, like Ari mentioned, are largely based on modeling the short range interactions between atoms and long range interactions, like electrostatics, are left to purely physical equations that like go through the usual, classical mechanics process.

Is it clear when, when purely modeling short range interactions and like deferring everything, deferring physics to long range interactions will fail. Or is that also unknown?

Corin: I think it's a super huge open question. I think it's one of the biggest sort of architectural puzzles facing the field.

And you can get people with very strong opinions that are like directly in conflict with one another, all of whom seem very smart. So there's there's a body of work that says message passing is all you need, like scale fixes this problem, like just scale it up more. yeah, at the limit of low data, you can't learn the long range things because they require more data to learn, but like just 10x the scale and it will all be fine.

There's another field like body of work that says it's too much to throw all the physics out. We should mix the easy physics back in that will make things much more robust, much more stable. Then there's another body of work that says the architectures are all wrong, like chemistry is less local than we think.

We need to mix like descriptors across coarse grained link scales. And I think it's fundamentally really hard to answer this until we get just better. Like we, we need to figure this out experimentally. I don't think we can armchair solve this problem.

Abhi: Yeah, that makes sense. I like, do you personally have a bet that you're making?

Corin: I can say that with our current generation, we're trying the message passing is all you need thing, because that's, that's the same thing that Meta's FAIR-Chem team has done. There's a lot of like toy systems where you can show that it can have problems, but if you keep like it seems fine for everything that matters and there's cases you can find where people have tried to add in very fancier solutions and it's just worse.

So it seems like the default option right now is just try just building a regular graph, and then we'll see, I think we'll learn a lot from this either way once we're finished benchmarking our current model and maybe update for the next generation.

[00:10:23] Emergent behavior in NNP's

Abhi: One thing I've often seen, and I think many people have seen, you often see this emergent behavior in these large general models, like the zero shot linguistic capabilities in GPT3 and more relevant for the biology world, this, protein conformation generation ability of Alphafold2.

Is there some analog to that in the neural potential world?

Corin: One of the things that's different about these models is that they're not super generative in a sense. So the most basic use cases, you take a cloud of atoms and maybe some like metadata, like charge and spin, and then you return like an energy and the forces of which are the derivatives of the energy.

So it's very like you very restricted output schema. Like you're, doing a simulation and it's, you don't want to be surprised by the energy. Like you don't want an unexpected energy. Like you. The ideal thing is you get exactly the energy you would have gotten from running that like reference level thing that you're trained against.

I think what will be super interesting and where I do think we might be surprised is, like there, there's a lot of work on multi head outputs or what happens inside the model? And can we stitch this together? Can we combine all of the sort of weights and like representation that we get from these like really large, really accurate simulation methods and can we do unexpected things with that and that's like a vague, that's not like a specific proposal but there's you know people show with like language transformers that you take like a math model and a Japanese model and then it can do math in Japanese like what's the analog for that in chemistry?

Like what's that gonna look like? I'm not sure but it seems like if you can train a model to always predict structure to energy, you've learned something pretty fundamental. Like that in some sense is the most fundamental relationship in chemistry and it seems like that should be transferable. Like some amount of internal intuition should be transferable to other tasks.

Abhi: I'm not sure if I'm like reading this correctly, but there was like that paper by Unke and also like Tim's crystallization work.

I guess for like context for people who haven't seen like Tim's tweet, they observed, nucleation, crystal, like crystal nucleation events. using neural network potentials that had never seen nucleation events before. And I think fairly someone in the responses replied that the nucleation event itself, shouldn't have happened in real life, but it's cool that like, structure arose where there was no structure at all.

Do you, that is also potentially an example of emergent behavior?

Ari: Definitely really cool. And I think, one thing that's just I naively exciting about this is you could train a model and it could work on multiple phases, right? If you think if Tim trained that model on, liquid and solid crystal data, it seems like, okay, the model is able to make these phase transitions.

And so you can start to see a path towards you know, a foundation model for atomistic simulation, or something that can handle, like, charge and phases. And, hopefully eventually radicals and transition metals. And, this is one of the things people are thinking about is, how do we, expand the coverage of these models to work on all sorts of, chemistries.

Which is I think a very different research problem than like, how do we scale these models to work on really big systems? And they're both like very promising and interesting research questions.

Corin: I think it teaches you something really fundamental about how information flows too, which is Even if you might say that the event that, dissolved salt shouldn't form solid salt under those conditions, like the description of the event wasn't quite right, it shows that like training on liquids teaches you something about solids.

And that I think is like really cool because that's the premise of a large pre-trained model in the first place, right? Which is that like when we dump all this data in, like somehow the data from other domains is making my domain better. Because otherwise, like why not just train a separate model for every protein or for every task?

Why, like why train a large language model? Why not train a code model and a math model and a translation model? Like we have this idea that somehow in language space, like you get better at language as one unified thing and that training on math somehow makes my code better even if it's not like as direct and I think we're we're seeing this from this as well like training on one phase makes my other phase like there's information transfer there, which is really cool.

Abhi: Do chemists naively bake this information flow into their mental models of like liquids and solids or for them like these domains are separate.

Corin: Ooh. If you think in inspecting the internal states of models is hard, chemists are way harder. But I, I think, yes, I think there's some, there's a language of chemistry, like you think in terms of structures and drawings and like you have a sort of, I don't know, a specific like ontology or metaphysics as a field.

And I think that does transfer between phases pretty well.

I think this really gets the nature of what is interpolation and what is extrapolation?

Which is such like a fundamental question out like just not even in chemistry just like in general. Like I think Conway's Game of Life. There's the cellular automata and you can make, it's like a, each cell only understands the positions of its neighbors, but you can build these like massive emergent systems that display this like complex, like Turing Complete behavior.

And it's, is it that interpolation or extrapolation? Like you're interpolating in like rules space, but you're like extrapolating in like outcome space.

Abhi: Yeah.

Corin: And I think that's like the analogy that makes sense to me here because it, turns out like, it seems like crystallization in water should be like an extrapolation in rule space.

But it seems from the model, like it's an interpolation in like rule space. And then the rules are more fundamental than it seems. it seems like there should be a separate sort of like physical behavior governing like dissolved sodium chloride and like solid sodium chloride. But at least superficially, it looks like it, it learns the both of them just fine.

And so I think this is like a, yeah, one way to think about neural network potentials is conservatively, like just assume they can learn with enough data in a given area, like they can learn like the rules of that data very well. They can learn the chemistry that's contained within that domain.

And then you can run simulations, see like how much. How extensible are those? Like in some sense, like how, much ground does that actually cover in output space?

[00:16:58] Enhanced sampling

Abhi: On the topic of enhanced sampling, I feel like I see relatively few papers bringing ML into the picture.

And it feels like enhanced sampling is one of those very, magical areas where you're very much supposed to, know what you're supposed to be doing before you ever touch it. Is, do you see ML, poking its way into, unphysical, like, modifying the system in such a way that you get to the spots that are interesting.

Corin: I think, yeah, a hundred percent. Hannes and Bowen have done some work on this. I think a bit, like John Chodera has tweeted about this a few times. It's gonna happen. I think the reason why this is so like finicky to get right is that to generate steps, like you want to generate a distribution that's like consistent with the Boltzmann distribution.

Otherwise all your like, binding affinity integrals or whatever property integrals you get are wrong. If you just, you can imagine a lot of ways to enhance sampling where you just, shuffle things around, like a grid search, but those don't, reproduce the correct answer. you, you need it, you need, the right Boltzmann acceptance criterion in, the StatMech terms.

And getting the ML to, rigorously reproduce the, right physical limit is, It's tough, like you have it or you don't. So there, there has to be some way to bake in the like verifiable correctness that a lot of other schemes have.

[00:18:16] Cultural distinctions in NNP's for life-sciences and material sciences

Abhi: I noticed that there's a lot of distinction between neural network potentials for small molecules, neural network potentials for proteins, and neural network potentials for materials.

Why is there such a strong distinction between each of these fields? Is it just kind of the limitation of scaling up and like different fields, different areas need to make different inductive biases or is there, it's like some more cultural difference.

Ari: I think it's, a product of mostly the age of the field is like neural networks, deep neural networks are like relatively young.

And a lot of the focus originally was on language than images. Now, we're starting to see people working on these geometric libraries for graph neural networks. And so I think, what we're seeing is like the problem that people chose to focus on first. And there haven't been very many iteration cycles in these models.

People have gotten out at most, maybe five generations of models. maybe a few more, but there, if you pick up a problem space, you say I'm a materials researcher, I'm going to build a model for materials. You try to build one model for materials and then get it out in the world, you start testing it on things and you think about how do I make this better?

And probably you're going to, not expand your scope super aggressively until you figure out, the area that you really care about. And so I think, like an early computational chemistry and in the neural network potential space, we're seeing some people care about scaling quickly and they're trying to figure out how do I make these work on proteins?

There are people who care about maybe materials discovery and property prediction. and then there are some people who care about, really accurately reproducing DFT results in trying to replace quantum chemistry. And I think like our hope at Rowan is to, start with replacing quantum chemistry, like DFT methods where we can.

And once we've done that, and we're satisfied with it, we'll start working on these scale challenges. And so I think you'll see that, a lot of people do start to work on these, foundation models for atomistic simulation. I just think it's we're too early.

Corin: It's sort like the advent of television, however, in maps and new medium onto whatever they understand best in the old medium. People used to just basically take plays with no modification and then just film them. And then it took, it took a while for people to realize you could do like dramatic zooms, like you didn't need to emote as much, you could, add special effects, stuff like this. Yeah, I guess it's like a snapshot of whatever problem you thought was most relevant at the time the new technology is dropping. And then iterate from there.

[00:21:13] Gap between simulation and real-life

Abhi: Slightly returning to the idea of training these neural network potentials in the first place, the whole concept is you take, usually density functional theory trajectories, take the forces and energies from those and train a model to simulate those instead of relying on the physical equations themselves.

What I find interesting is that for most of the molecular dynamics field, ground truth datasets are basically impossible to actually gather at all. You're relying entirely on pure in silico measurements. What do people in the field think about this gap between simulation and real life behavior?

Is there a gap?

Corin: Yeah, there's always, there's a huge gap right now because the simulations don't work. So like even the best periodic DFT simulations of water, maybe not the best, but that once people often use like PBE water, probably the most common periodic functional is like a solid at like room temperature, like you, if you see a PBE simulation of water, it's usually heated to 80 degrees or so, just cause that way it's liquid.

Which is, it's just one of these things that you, I think once you're in the field for a while, you take that for granted and then you stare, you step back and you're like, hey, that's bothersome, isn't it? The foundational solvent for all life is not really modeled very well here, is it?

And, I think there's...

Abhi: So we can't model the boiling point of water?

Corin: You can show that if you layer on enough approximations, there have been a few papers about this in the last couple years, that you like, get better, but boiling point is actually pretty hard, like it's, I don't know, there's a lot of molecules involved, there's like solid gas, or liquid gas interfaces, like it's, a highly emergent property that results from like very, small inaccuracies in energy, and I, yeah, I think the intuition that a lot of people in simulation have, or at least that we have, I don't, need to speak for a field, is that like we can see microscopically the ways in which we're wrong.

So we can run our ultra high level quantum calculations, we can run the stuff we usually use for production, and we can see this is where we're wrong and by how much. We can compare, and then we can see macroscopically that our predictions are like inaccurate. We're pretty sure with neural network potentials we can fix the microscopic predictions.

The experiment that we're running, as with everyone else, is does that fix the macroscopic predictions? And it's very logical. The answer should be yes, but it's not. It by no means guaranteed.

Abhi: Are there like hypotheses as to why that may not be the case?

Corin: This somewhat ties to like the

There's so many ways you can imagine this not being true. So one thing that density functional theory and most quantum chemistry ignores is nuclear quantum effects.

So like hydrogen tunneling and such. Now I think most people's intuition is that this isn't, outside of like certain enzymatic processes that are like pretty circumscribed, this isn't a super big thing. You know it probably affects the kinetics of proton transfer in water a bit, but like we already know that like you can replace most of the hydrogens in your blood with deuterium or like a surprising amount and it doesn't really affect you that much.

So like it seems like biological models shouldn't be incredibly sensitive to the exact kinetics of like H atom transfer, but like maybe this is wrong, maybe these there's on the scale of a whole protein solvated in water, like even being a little bit off on these things matters a lot.

Yeah, I don't know.

Abhi: On the topic of boiling water, what do you think is like off there lately? We can't measure a system large enough or is there some like minor, do we need to go even deeper than density functional theory to actually model boiling?

Corin: So there's actually two ways we do DFT and this is like a not super well appreciated, I think. So there's molecular systems and there's periodic systems. So you're trying to describe the electron density and the electronic structure of a system. And for isolated molecules, that looks like putting basis functions, like describing the density in an atom centered way.

For periodic systems where your system is actually infinite, like a slab of metal or like a box of water molecules, that ends up not working super well. And so often people do these like plain wave, they use like a Fourier basis to describe electron density. And there's a lot of downstream things that you do.

So there's so many approximations that you do in density functional theory, and those approximations end up shaking down differently between molecular and periodic DFT. I think one of the consequences is that a lot of the most accurate methods from molecular DFT that we can super rigorously verify against ultra high level calculations don't exist in periodic DFT.

And a lot of the high level calculations don't exist in periodic DFT either. So the functionals, you can't do like electronic exchange, like quantum exchange, for instance, very well. and that ends up like being pretty important for a lot of things. And I don't know, we know when we run these functionals on molecular systems, we can benchmark, you're like, yeah, the water sticks together like 20 percent too much.

And so when you do periodic systems and you maybe it just imagine it's about the same, like you stick together 20 or 40 percent too much that throws the bulk boiling point off by a lot.

Abhi: If we like went even like more accurate, even more slower, like coupled cluster (theory) like do we, are we then able to model boiling?

Or even then, like there's like potential issues that start showing?

Corin: I think so. I think it is, So people have been able to dial in the accuracy. I can probably find the reference for this and you do approach the correct boiling point with a technique. So it doesn't seem like there's something like fundamentally massive that we don't understand here.

I think it just shows that, density functional theory in the sort of like life sciences is viewed as like an ab initio high accuracy method. But within the like world of high accuracy simulation, density functional theory is actually like the plebeian, like the dirty stuff for losers. those people are all working on these hyper orbital optimized like wave function methods that work on 12 atoms and you know from the theory point of view they're absolutely right. Like I think it just goes to show that solving the electronic structure problem is like Really hard, you know it the exact solution is O of n factorial which is terrible Yeah, like it's like the three body problem, but worse is all this quantum through space stuff there's hundreds of electrons and like it's it's, just hard.

Abhi: Are there, like papers showing that lately right now, everyone that uses DFT for, training data, but like potentially higher quality data is what you want to care about rather than the scale of data. Do you imagine like in 5, 10 years, people are going to go beyond DFT to something even more high accuracy or DFT genuinely is like sufficient for a lot of things.

Ari: I don't know. So one interesting model to talk about here is like the ANI model that was fine-tuned on a coupled cluster data set. It, performs surprisingly well on benchmarks against coupled clustered data like today. so I think like maybe we'll see people try to replicate that. but one thing that I would bet on first is that people are going to, ditch periodic DFT for generating training data.

Because the best methods or like the best DFT methods are only implemented for molecular systems. And so I think one challenge is figuring out, can I train a model, a neural network potential that will, work on periodic systems, but it's only been trained on molecular systems.

And I think that's a big question, but if you can get it to work, then in theory, you know, you'd be able to model periodic systems with higher accuracy than any DFT functional that's implemented for periodic systems would let you.

Abhi: Do you think you'll see this pattern of people like, like starting out really coarse scaled, and then bootstrapping their way up to have like smaller, like lower n data, but higher quality data. Do you think that's the future?

Corin: I think it is right. Like we're trying to solve such hard problems here. Like simulation is just really hard to do well.

I think every source of information you can get is valuable. so probably pre-training at a, ton of stuff at a lower level of theory to initialize your weights and biases is probably a good move. People do denoising as an auxiliary task. So FAIR-Chem has done this, that seems to be good.

Mixing in different like levels, like multi fidelity learning seems to be good where you can do it. even adding experimental data. So like crystal structures, we know the forces are zero, like that seems to be good. Yeah, I think it's, I, don't exactly know how much each of these things will contribute, but like any mix of like more tools to throw the problem, more sources of truth, I think is, super, super valuable.

And people even do you know, you can do backprop through a whole simulation. So if you have an experimental tautomer ratio, for instance, you can back prop through a simulation and like train to get the right answer, like over all of the molecular dynamics steps.

Abhi: Given a final end state.

Corin: Given a final end state. Yeah. Or like a different energy. The problem is that you have this sort of dimensionality problem because you have one thing and there's like one experimental result in so many states,. That seems like probably not sufficient to do it like a whole model from scratch.

Abhi: It feels like a very RL problem where you have this, like you have a final end reward and nothing else.

Corin: Yeah. And like the number of steps is more than chess. So it's tough.

Abhi: Yeah. That makes sense.

I think like like my initial suspect, like suspicion when I first entered the field is that, like obviously people are maybe using these in-silico DFT measurements for a lot of things, but there has to be like some sort of physical measurement being pushed through the window also.

And I assumed it was going to be like NMR, nuclear magnetic resonance imaging, because it feels like that's the only way you can actually measure dynamic movement of molecules. But it seems I have never seen that actually used in a paper. Is there a reason why?

Corin: The NMR timescale is so long, like you can get clever and try to view fast processes, but you're looking at about a microsecond just because the spin states are so long lived.

So like you can see two different species if they're like, and you can do the like pulse things to look at the kinetics, but they have to be pretty long lived species. So if you think what a microsecond is. So that's 10 to the negative 6 seconds, right? And the usual time scale for a simulation step is 10 to the minus 15 seconds.

So you've got nine orders of magnitude still, like there's just a lot of room at the bottom. Yeah. I think so the faster spectroscopy method you can do or something like multi dimensional infrared spectroscopy, which gets you down to about, yeah, 10 to the minus 11 I'd say seconds. So that's like much closer.

It's like the timescale of bond vibration are much closer to that. and so that helps, but again, it's difficult. Like you can't do 2D IR and make like a map of what a protein or biomolecule looks like and how it's moving. You can probe very specific things like a complex lifetime, but it like the, I think the value of the data and the difficulty to acquire each measurement just makes it, tough.

Like really crystallography for all we complain about it actually works really well. Like robots can look at crystal trays. You can grow a lot of crystal structures. And a lot of these like fancier spectroscopy methods require like grad student years. And that is an expensive currency.

Abhi: So like with x-ray crystallography, I'm surprised it gives you a sense of dynamics at all.

Do you mean that it gives you a sense of dynamics or it gives you a sense of something related to dynamics?

Corin: It doesn't give you a sense of dynamics at all. Obviously it's a static structure. I think what it does do is a crystal is something that's usually in a ground state, right?

Or modulo thermal and packing effects.

Abhi: So that's like a zero energy thing.

Corin: Yeah. And so zero forces, zero forces. And that lets you, that's like some piece of experimental truth that should be useful. It'd be like this thing, whatever it is, like a local minimum. And I don't think you can extrapolate the whole relationship from that.

But if I. It's clearly telling you something and then something is experimental and should be very useful.

Abhi: Is it, is like that fact ever used in these neural network potential papers? Like it's like a, as a possible like end state.

Corin: I, can you think of any, I haven't seen it be used, but it seems like it should be ultimately.

Ari: Yeah. I haven't seen many papers that are trying to like fit to experimental measurements. It seems like, you ought to be able to add them as tasks. They're in the benchmarks. Good benchmarks are on experimental sorts of data. And especially with benchmarking, these periodic neural network potentials, a lot of the benchmarks are, experimental properties.

I think my hope for the field is that sort of, as we work on building these models, we benchmark and train if we can on starting with small system properties and working our way up to these like bulk and larger molecule properties.

Abhi: Do you think like using these sorts of like real life data sets are like higher hanging fruit and it's not really worth engaging in it until we like really speed up the in-silico measurements?

Corin: It feels there's a lot of obvious things that we could do right now. And it like, I think it depends on how crucial like the experimental data ends up being. Cause it might be like, we've taken a pretty pessimistic view, I think, of the state of the art so far here. But, Schrodinger, binding affinity prediction, docking.

Yeah, they don't work perfectly, but, they work, they clearly provide value. It's a huge and great company. They, everybody uses it. And you can say, that seems to work pretty well. Yeah, exact numbers are off, r squared is not quite there, boiling point of water, yadda, yadda, some of the proteins need constraints.

But it's not like it totally doesn't work. You can say, look, we can make all of the forces way more accurate with training to DFT. We can do high quality DFT, this seems like verifiably very good data. Maybe that is enough to, maybe it doesn't get you R squared of 1.000, but it doesn't take that, maybe that's just pragmatically ends up getting you 90, 95 percent of the way there, and it doesn't take that much experimental data to fix this. I think that's we haven't tried the obvious strategy enough to know that it fails. And so I think it's putting the cart before the horse to start like, going for the higher hanging fruit.

Abhi: Yeah, that makes sense. Investing millions of dollars into better crystallization, better electron detectors.

Corin: Computational data is, yeah, it takes money to run the computers, but you just click run, you get your AWS credits and then you're, set.

[00:36:18] Benchmarking in NNP's

Abhi: I'm curious, how are neural network potentials benchmarked in practice?

Do you, is you, talked a little bit about this, you have this like potential end state you, we can back prop through that. And is one of the goals for neural network potentials is to recapitulate that end state, or is there some hope that it can also follow along the trajectory and match the trajectory that's like from DFT exactly?

Ari: A lot of the benchmarking work that I've done and I've seen is on, less exciting things than that. It's does this recreate energy and forces from DFT. Does this, sometimes like one of the benchmarks that people are using now is this like SRME, like thermodynamic stability benchmarks?

If I run an MD video, not does it recreate trajectories, but is it stable is what the benchmark tries to measure.

Abhi: By stability, does that mean like the atoms stay in the same place or like forces don't explode? What's the measure of stability here? I don't actually know how it's implemented.

Corin: At a qualitative level, early neural network potentials often looked good on things that looked like the training set. But if you run an MD simulation, they get out of distribution and then start returning random numbers. And, physically, the simulation, explodes. it's like a grenade went off in the computer.

And so people, one of the benchmarks people have developed is okay, make sure it doesn't do that. so that's pretty crude.

Ari: There are, in a lot of papers, like side by sides of we ran some system in here's one potential energy surface, here's another potential energy surface.

There's no like benchmark number that they give you. It's just two pictures side by side. And you're like, you got some of the wells and peaks, And others of them very wrong. And I think like more benchmarks like that, can we recreate potential energy surfaces? Maybe not exact trajectories. though that would be again, really cool.

I think like MD at least, at like room temperature or higher, like it's very chaotic. And so I think what I would be more interested in is hey, can we accurately reproduce potential energy surfaces? And if we could, benchmark that well for systems. I think that would be like a really interesting and useful benchmark.

Abhi: I remember I used to be really into computer vision when I was in college and there was this meme of people like getting 0.01% better on CIFAR-10, a common benchmark used in computer vision. Is there some like analog to that with benchmarking in molecular dynamics where people fit themselves really well on toy problems, but like the gain doesn't actually matter all that much?

Corin: There've been, I think, different generations of benchmarks.

So QM9 was a big one. There was like nine atom molecules with a bunch of different properties that saw a ton of use back in the day. I think people have realized, or maybe just people got too good at it or like it didn't end up being incredibly useful. You see it from time to time. I think it's early days.

I don't think there's like a uniform set of benchmarks yet. There's ML benchmarks. Maybe you're actually a better person to talk about this, because you've spent more time here.

Ari: I think for molecular neural network potentials, there's still a lot of alpha in benchmarking. It's a hole that I'm trying to work on filling in my free time.

Yeah. I think for like periodic models that are working on materials, the Materials Project has tried to do a good job building data sets and also benchmarking. And so the like sort of MatBench discovery leaderboard, people like post about OrbV2 beat MatBench Discovery, and then a few days later, Open Materials 2024 from FAIR-Chem, topped OrbV2 just slightly.

And they're like, our model is at the top of MatBench discovery. I think like one thing that I haven't seen yet is does a model that tops MatBench Discovery turn into shareholder value somehow? And I think, not yet the jury's out. We'll see. But I think trying to figure out, what are the benchmarks you need that really tell you, is this thing going to be useful or important to find?

And I think at least MatBench Discovery is an attempt to do that.

Abhi: Do you, can you explain what MatBench is?

Ari: Yeah, it's a collection of, benchmarks for periodic systems. And, so there's this nice table on the website that shows like the models in the rows and the benchmarks in the these questions like, can you predict energies of systems correctly?

Can you reproduce forces? They recently added the first, MD related benchmark, which is this thermodynamic stability one. And then they have a way to calculate a total score for the model at some, weighted average of those other benchmarks and you can sort them and they have some compliance criteria that the models have to meet.

I don't know exactly what those compliance criteria are. yeah, but it's just that it's a collection of pretty standard benchmarks, but it's a way to at least have some way of knowing, when someone publishes a new paper, is this better or worse than the last paper that was published in this field.

[00:41:49] Is molecular dynamics actually useful?

Abhi: On the topic of these models actually producing shareholder value. you may, you've made this point in the past about how a lot of these molecular systems are studying microscopic properties in hopes that they translate to macroscopic properties. Are there, it's surprising to me that at least I haven't seen any neural network potential paper try to poke at this problem of am I recapitulating the macroscopic property?

Or are there?

Corin: I think the papers exist. there's nice work from, some folks out of Cambridge essentially showing that you can get like a hydration free energy really well. So like how strongly is this molecule solvated by water that you can learn that really well with a good neural network potential that came out like maybe a year ago.

I thought that pretty great. If you get the details right, the outcome is right too. And I was really excited to see that. I think people are working on it with binding affinity with these big, like free energy perturbation, like protein ligand interaction questions. The problem ends up being that things are still too slow.

So you have to do various approximate methods. It's like end state correction. It's not a hundred percent clear that like existing models are capable of describing protein ligand interactions super accurately yet. So there's a lot of asterisks and the results I think overall are unclear.

There was a recent paper from Exscienta that argued basically it's not better. Like it, it's about the same as just refitting the torsions in the small molecule force field. I think one of the things that's, challenging about this field is, the questions you're asking, macroscopic benchmarking, like, how do we check that we're better at the things we care about, are like, they're exactly the right questions, right?

It's, I think the, very logical thing to ask. We're rebuilding, a legacy. drug and material science tech stack not from the ground up, but we're having to port pieces over. It's like porting things into CUDA. Like you need everything to work. And we're still like over the past couple of years, like frantically as a community building the infrastructure to like, how do we actually run FEP?

Like, how do we get like these protein predictions, like the melting temperatures, the helicity, all this stuff. Like, how do we do this with our own neural network potentials? How do we scale it? How do we, there's just a lot of sort of practical work that I think will and is, very actively being done. But it's, there have been, like, two papers this year which show you can use NNPs for full proteins, and they're, like, some of the first two that actually have done it in a useful way.

So it's, just very, early, I think.

Abhi: I think one interesting point I've come across as I like study this area is this disagreement as to whether dynamics at all are useful and you instead just want to like sample the distribution of possible dynamic states.

Corin: So dynamics is like clearly useful when you need timing information, which is perhaps like an obvious point, but one that should be said, if I want to, like if I have some kinase and it has like an open and closed loop confirmation, if I want to study how long it takes me to get from open to close, like what sort of the kinetics of that are you like, you really need dynamics there because that is like a dynamical question.

It's time bound. I think oftentimes people use molecular dynamics, not because they care about the time evolution of a system, but much more because they just want an efficient way to sample different states. So you're trying to take some statistical mechanic, like average, you want to get ergodicity, like some sort of like unbiased sampling and like MD is just a super robust way to do that.

But in that case, you might imagine that there's like much, much more efficient ways to sample than MD because your time steps are obviously very, correlated with one another. So the information per frame is pretty low.

Abhi: I think especially a lot of like pure computational ML people are very pessimistic about molecular dynamics as a field. Like from their point of view, they can do conformational sampling with AlphaFold. AlphaFold also seemingly has a sense of, flexibility on par with molecular dynamics.

You can do like in docking with DiffDock. I imagine there are some areas for which MD has been like genuinely outstripped by the capabilities in ML, but I'm also reasonably positive there has to be some areas for which like molecular dynamics is still going to be important and will continue to be important for the future.

I would love to hear what your thoughts are on the situation.

Ari: I think one thing that's really important to just say is that like MD today relies on force fields, which are these polynomial approximations of quantum chemistry. And so what people think of as MD is very different than what the hope for like neural network potential MD is going to be.

And so MD today is useful for some tasks, right? Like I think the one that comes to mind is like free energy perturbation, which generates, binding affinities, it seems a lot better than docking to me, though, also more expensive and not without it's failure cases. but taking a step back, like the most intuitive way to run a computation or to run a simulation is to simulate what's actually happening, right?

And, it's very small and it happens very quickly, but like the proteins and the little small molecule drugs in our bodies when we take pills, they do actually move around. They play out videos that happen over time and they interact and I think modeling that faithfully is always going to be useful.

Abhi: It feels like it's I do get the instinctive vibe to try and model things like faithfully to what's actually going on. I am curious, do you think the future of neural network, but we don't currently appreciate the value of neural network potentials because the current state of MD is just like really bad and just like intractable to do anything useful with.

Do you imagine there'll be like new use cases that spawn as a result of having NNP's are actually like fast, reliable, and able to be scaled up.

Corin: Yeah, I do. I think so there's like use cases in like model like. So here's an easy one is like covalent reactivity, right? So like modeling covalent docking and like the reactivity of covalent enzymes. Usually like force fields can't model reactions, quantum chemistry can't model large systems. Covalent inhibitors are reactive things that react with large systems, so that's like pretty intractable with state of the art methods. There's various sort of ways you can get around that, but they don't work super well. That's something where like you do, that's that's useful, like to actually model that, like covalent inhibitors are awesome.

They're widely used these days. you can look at the KRas work, And that's, to actually be able to model the covalent inhibition seems very important. And I think too, like echoing what Ari said earlier, there's some intuition that like if you want your like DiffDock or whatever to be accurate to actually get like super useful accuracy like you have to be learning chemistry implicitly somehow, cause we know that binding one molecule that has a hydroxyl versus one that has like a more strongly hydrogen bond donating group or like a less, hydrogen bond acceptor or like a different high stacking preference.

Like those all matter, like those demonstrably matter. And so maybe you can learn all of this in some implicit roundabout way with like tokenized language models. Like you put your ESM2 in, you put your tokenizer, like the recent QuickBind paper, you learn some sort of like interaction matrix there.

But it seems like you're just reinventing the things you want a neural network potential to do in a very like just a strange way a little bit like you to get the whatever accuracy you want like at some point like these are the sorts of modifications and like structure activity relationships that people really care about and like at least in small molecule space it seems like you need to know the chemistry and maybe the most parsimonious seeming way to do that is to teach a model chemistry and then model the process.

And, it might turn out that there's some DNA encoded library way to backdoor all of this, but that, it seems less likely to me.

Abhi: Yeah. I think that's actually a really interesting way of framing the whole problem that like neither side is implicitly discovering secret knowledge about the system.

The system is the system and you might as well like fit, like simulate what's actually going on inside of it rather than focusing only on static structures and hoping to learn accidentally what's going on.

Corin: Yeah. And I think this is problem specific too. So there's some cases where there is some secret knowledge you need to know.

It's like, you've written it before about toxicity prediction. That is a case where you're not simulating the liver, Rowan in 10 years, you won't have an atom per atom, like map of the liver. And then you like draw your molecule in and like just play and see what happens.

Like there are cases where you're learning some sort of groping around some large elephant of something and trying to like divine useful patterns there. And I think that's that's just a very different beast, but for the specific and like extremely important problem of binding two known things together in which we know that matters a lot.

We know we can't do it, and that's like a direct simulation problem in some sense.

Abhi: On this topic of like physically modeling the liver, do you not think like we'll ever get to the point with like incredibly hyper course grained models that like understand the many, many body problem that's going on.

And we can simulate like entire cells, entire organs. Is that at all like a genuine possibility in the next 10 years? Or it's just like in the realm of sci fi for the large part?

Corin: 10 years feels super ambitious there. I think chemistry is one like way to see the world. Med chemists like think in terms of atoms, like you can see the med chemists in a talk by the people who look bored for all of the slides with like histology on them. And then when they see any structure of a molecule, they perk up. That's a med chem phenotype. I don't know that it's the right way to tackle all problems. I think there's enough problems that it's like incredibly interesting, but maybe a sort of like phenotypic approach to some of these like high order problems is just more it's just better.

I think that's a, you, you've read about the Recursion play as well, this, almost coarse graining over cells is like a crude way to put that. And I think that's, even with like antibodies, like I think I'm certain there's like useful stuff we could do if we could do an atom per atom, like model of an antibody.

Like I'm sure we'd learn surprising things, but it might also be the case that we have so much inductive bias, we have so much like evolutionary information around antibodies that like it's not super crucial, to do like a rigorous atomistic simulation in the same way that like the, the models we have are going to be much more effective.

I think small molecules is such a such an unconstrained design space. Like you're literally positioning every atom that like needing to be closer to the metal of physics is just unavoidable. And I think the more atom per atom we go, the more true that will be, so in the like non canonical amino acid direction as well.

[00:53:14] Solvent effects

Abhi: Solvent effects is one of the things like a lot of pure ML models completely ignore. It pretends that, there's it doesn't exist at all.

It's not a factor to take into account. How important are like diverse solvent effects when you're dealing with these structures? Is it like the, universal solvent that's learned by like DiffDock, is that kind of sufficient for a lot of things? Or are there like a vast variety of solvents that actually exist within the body?

Corin: I think, to a first approximation, pH 7.4 water with some electrolyte background. a good proxy on the scale of proteins and ligands. I think obviously, with membranes, you, can, you, there's, going to be exceptions to that rule, on the scale of cells, it's not pH 7.4 water. We all know that, but on the scale of an individual protein as something right next to it, I think that's a pretty good model. I think where solvent effects become super, super important is in like the, reactivity, like crystallization process side of things. And then in material science as well, when you, just deal in much more diverse environments, That makes sense.

Abhi: Yeah. I think, imagine if you're going to like the thousands of Kelvin's, then like it, the, simulation actually becomes important. It's really hard to train a model for that.

Corin: Yeah. Or, like these, I know a super big problem is like predicting solubility under various conditions like for crystallizing out an active drug product.

And there you can easily have a blend of three different solvents, right? Or like even inside a battery you get the same thing. So you have some like ethanol, you have some carbonate, you have some water. and then this is like it, modeling solvent becomes like a factorial problem of complexity. But, the body, yeah, sure there's, cancer cells are slightly more acidic, but it's pretty similar.

Abhi: I didn't actually know that. That's interesting.

Corin: Yeah. There's some interesting work where you can design like acid released, like essentially payloads, that will, I think in theory is somewhat selectively be activated in the presence of cancer cells. I think it works less well than ADCs to do the same thing.

Like an antibody is more selective than a pH sensitive group, but it's a cool idea.

[00:55:17] Quantum effects in large biomolecules

Abhi: I'm, curious, do you think there's like interesting quantum effects going on inside of antibodies, like, large biomolecules in general that most people are just ignoring because it's too hard to study?

Ari: This is like really unexplored territory. I don't know how we would discover that those effects are happening, without new tools.

Abhi: Like, that Microsoft, A, AI2BMD or BMD2AI paper

The, context, it's a ab initio neural network potential paper that's specifically trained on fragmented proteins is able to scale up extraordinarily well, able to recapitulate the true dynamics of what's going on in proteins. As far as I can tell, correct me if I'm wrong, it's like the first way to actually study really large biomolecules using quantum, like quantum level accuracy. Do you think that paper is going to unlock a lot of interesting things?

Corin: I hope so. I really hope so. I think it's There's, there's like a divide here between there's known unknowns and unknown unknowns. We know we don't know how to do protein ligand binding affinity super well in a lot of cases. When we go to the realm of antibody dynamics, like maybe there's experimental data here that I don't know much about.

I'm not an antibody expert at all, but I think we just really don't know what we're gonna find. And that's I think, terrifically exciting as a basic research question. As like a startup person, I don't really see that as like a market that we're thinking about as like we're gonna solve antibody problems, but like I think there's just a humility you have to have no, we don't know and it like it could be anything.

[00:57:03] The legacy of DESRES and Anton

Abhi: This kind of leads well into my question on like the legacy of molecular dynamics. One of the, one of the big pushes in the field was D.E. Shaw's Research Anton, which is a, a supercomputer with thousands of custom built ASIC's, built by hardware engineers, paid extraordinary amounts of money.

The supercomputer led to papers with incredible titles, like 20 Microseconds of Molecular Dynamics Simulation Before Lunch. And for people who are not in the field, 20 microseconds is an immense amount of time for a dynamic simulation to be run. Yet, the company hasn't ever released a drug. It's largely been papers.

They have a therapeutic arm right now, no drugs has come out of this. What do you think is the legacy of DESRES (D.E. Shaw Research) and Anton?

Corin: I think, so it is worth maybe mentioning the partnership with Relay here that they have. Relay Therapeutics, a company in Cambridge, a lot of respect for them.

Pat Walters is there. Like a lot of great folks are at Relay. One of the original ideas is using DESRES and Anton to discover like allosteric sites for known targets that had proved like resistant, to previous treatment. It's very tough to know, I think, how much Anton actually helps.

To the extent that Relay gets drugs successfully approved on the market, what percent of the credit does Anton and MD and DESRES get? I think it's yeah, I think, there's probably a handful of people in the world who know the answer to that question, none of which are in this recording studio.

Yeah, I do, I do think, part of the issue, if you step back and look at the whole field, like, where would we expect molecular dynamics to be useful? We want simulations because ultimately they should be faster than experiments. Like we should be able to iterate quickly in the computer.

That's how other fields use simulation. If you look at aerodynamics, like you simulate a bunch of wings and flaps and then you don't have to make as many in the machine shop. that's, clearly useful. Like you have this R&D spend, you have this search problem, like a design and simulation problem, and you can quickly like funnel down the list of things you need to actually try in real life because the simulation is like of a sufficient fidelity.

I just don't like we're just not there like I think people don't really think about this because we take it for granted that you have to try everything in the lab if you want reliable data but that's where we want MD to be useful. Like that's where sort of the MD shaped hole is I think it's in like this, hit finding, hit to lead optimization part, at least in small molecule land of that drug design workflow.

And so we, we talk about prioritization. We talk about gaining insight. We talk about like some screening, but like the, at the end of the day, we're a Boston based company. If you wander around Kendall Square, buildings and buildings of people just manually doing the search that like abstractly you'd wish MD were able to do.

And I think that's, whether or not the impact of MD is like 0% or 2%, like it's not where it seems like abstractly like it should be. Like it's not doing the things that we'd want it to do and so trying to get there, like I think if MD were already like some fantastic workhorse in simulation, that'd be great for the field, but there wouldn't be a need for us for what we're doing like that.

Our company wouldn't exist.

Abhi: Like what's weird is that there clearly are successes of MD. Like Elliot Hershberg and Bruce Booth have written about Nimbus Therapuetics' partnership with Schrodinger. They supplied basically what became blockbuster drugs and they got zero money out of it, but like they produced the drugs. Do you think that was like a fluke that they were able to do those three drugs so well, but none of their therapeutic arms have like really worked out beyond that?

Corin: I think it's too early to tell what the Schrodinger therapeutic arms right because those are pretty recent.

Abhi: That's true.

Corin: So like the MALT1 one, there's the covid. There's something else I think jury's still out on those ones just because it's only a couple years old. I think you have to believe either like one of two things looking at. So I think that the Nimbus, the deal that I have in my head is the TYK2 inhibitor that they, sold to Takeda, right?

And I think the price was 6 billion. And do you remember how much?

Ari: I think Schrodinger got under 1 billion of that. I think, yeah, I think it was under 400.

Abhi: I thought they got like nothing. (I am wrong about this)

Corin: I think it was a one or 200 million.

Abhi: Okay.

Corin: And so there's there's two conclusions here. One is the value creation is low or the value capture is low.

And I guess, I think you, my, my hypothesis or like my gut, I don't know if it's even a hypothesis, it's just like that the value creation is lower than it seems that you're not bearing that much of the risk with the tools, like you're not, you don't just hand someone a drug on a platter, like you work with their experimental teams, but you still need, like Nimbus is a real company, like they have really smart people there who are like laboring in the trenches to build the drug and like some combination of the premium you carry for the risk and all the experimental work you still have to do even with the simulation makes it that, I imagine that split reflects value in some fair way.

[01:02:27] Unique value add of simulation data

Abhi: Yeah.

I guess also when I think of simulation, I also think of the potential to understand parts of your system that would be intractable to understand in real life. Is there such a thing like that in chemistry or like protein design, molecular design where you need simulation to understand something that is genuinely impossible to understand without it?

Ari: I think the easy example here is something like reaction mechanisms. They just happen way too fast to study with an electron microscope, you can't point an electron microscope at a reaction, but it's concerted. And yeah, I think that's where these tools get a lot of use right now and where they're really valued.

I think for, bigger problems, that there are these like predictive accuracy thresholds they have to pass to be really valuable. And like maybe MD was accurate enough for a few proteins, but not accurate enough for, any protein off the shelf. What else do you think?

Corin: Yeah. I think just the role of insight in general is it's very tough to quantify, like it's tough to put a dollar sign on like insight, how much value does insight provide your organization, but, fundamentally, atoms are really small, chemistry happens really fast, and it's pretty easy to live, weeks, months, years, as someone who works in like the world of atoms without really getting any like tangible like window into what's happening.

And I think this is why people like docking so much. You can read all these papers arguing about how docking is not, it's not useful, docking is information theory like negligible. As someone like told to me like, from an information theory perspective, docking is useless. But at the end of the day, like if you talk to med chemists and you ask what do you like? They're like I love being able to see how my compound might or might not fit into the pocket like to just get a sense of how big it is and how it might fit in 3D is really useful for me. Even if the numbers don't mean anything, I just derive a lot of satisfaction, and it helps me brainstorm just to see how it might fit.

It helps me generate ideas. And I think, that's, I think that's just, useful for people. And yeah, like, reaction modeling, covalent inhibitors, like dynamics, like watching how things move, like protein pockets. I think there's, this is it's not directly impacting the bottom line, but just building, really valuable, tools that help scientists think and build intuition better is underrated.

Abhi: I think, where my mind immediately leaps to, I know you guys interviewed a lot of scientists while like you were building up Rowan. How much of this like med chemists, intuition that oh, these tools are like really helping me understand what's going on. How much of that is like not real because they, like seeing that something might happen, even if it doesn't actually match up with what's actually happening.

How much of that, is just, I don't want to say the word cope, but somewhat cope?

Corin: Yeah, it might be cope. I think there is some sense, if your tool generates random pictures, I think it, people will suss that out. I think there is, med chemists actually are pretty quick to learn to distrust computational tools.

It doesn't take much. Many people you talk to have a well adapted immune system for not believing things on computers, which is I think probably rational from their point of view.

I do think like being conceptually useful and being correct are not always the same and if you're sufficiently correlated with being correct, people will perceive it as useful even if it's not. it's a good model for something, but that doesn't mean it's you're on the right path of searching for capital T truth in this case.

Yeah, I don't really, and people are just susceptible to pretty pictures too, so if you like, give them something nice for their slide deck, they'll probably like it, because it makes them feel like they're acting more, like they're more rational in their job, as opposed to just screening things randomly, which is, a very effective strategy and one that's widely employed.

[01:06:34] NNP's in material science

Abhi: I have very little background in the material science applications of this. I would love to like on the topic of catalyst design and areas within materials, material science, I would love to hear what, is the use case of NNPs there?

Corin: Yeah. In simulation people talk about material science as a monolith, but it's eight tiny fields hiding inside a trench coat, I think. So there's some things which are like, you can essentially use the same models you would for drug design. Like you're modeling organic drug like molecules, but in different contexts.

So some of these like redox flow batteries, battery electrolytes, maybe you have different salts and more phosphates floating around, fewer amino acids, but like it, it's very similar problems on some level, like it's solution modeling, it's molecular dynamics, polymer properties, like these thermostats, like you're, modeling distribution of systems, other things like solar, upconversion, like these, processes end up like being very, different.

I think material science writ large... Like for drug design, there's like a playbook. Maybe there's a few playbooks. There's an antibody playbook. There's a small molecule playbook. Maybe you guys are writing an AAV playbook. I don't think playbooks exist in the same way in material science, like it's much more like everyone has slightly different research problems and solves them in a slightly different way.

And so I think trying to port a bunch of like very specialized workflow tools hasn't worked quite so well in material science. like some people are designing OLEDs, some people are designing like new inks. Some people are designing like electrooptic materials. And you need to be versatile.

You need to be able to be generally useful. And I think it takes at a very base level though, if you have a model that understands chemistry, all of these are on some level chemical problems. And so you can be useful. It just, the solutions need to be a lot more adaptable.

Abhi: I think when I like the few stories I've read about, research in material science and specifically like the creation of blue LEDs and like ongoing work and, semiconductors, it feels very much like a field where it's like a, you try a billion things until something works is that, and it feels like chemistry, like drug design is a little bit more rational and how things are designed.

Is that like a fair distinction?

Corin: I think in both fields, there's moments of rationality and there's moments of sometimes the most rational thing to do is just screen a million things. Like I think that's part of where the rationality comes in is like figuring out when and how to screen a million things, because you can imagine blue LEDs, there's some like band gap you're tuning.

Like you can imagine how you change the molecule that will change it. But there's also there's packing effects, there's stability and you, want to get in close enough and then you're like, all right, now we screen. And I think it's the same with drug design. You don't, maybe there it's we often start with a high throughput screen. We start with a DNA encoded library, some fragment soaking something. And we, bake in the randomness up front. And then once we have a hit, the intuition is that we can use our medicinal chemists, computational tools, all of this sort of intuition and skill to rationally get somewhere.

But I think like one of the best and worst things about chemistry as a field is it sits between being able to be fully understood and being unintelligible. Like that there's at all levels, this mix of you need to understand things, but you can't understand everything.

And that's part of what I like about it.

Abhi: if you, go back to the Japanese salaryman, who created the blue LED in the first place and you gave him like material science neural network potentials, like would there be any real benefit? Could he like, can you do anything interesting with it?

Or even that's is a little bit challenging.

Corin: Can you, I don't actually know what the, I've read a tiny bit about like band gap effects and LEDs. Our coworker Jonathan put together a blog post on that, but I don't know, I don't actually know what the like challenge he was solving was.

Abhi: The challenge was almost like, in like deposition of one material over another, you need to get it like perfectly exactly right.

Otherwise certain things wouldn't work. There was also this there's the band gap problem. He needed to have, I forget the details, but he needed to have this weird structure to ensure that like the electrons actually flowed correctly. And it was just, he, I think the video I watched over, it says something like he worked like 15 hour days just working with the, like the deposition machine, just like trying a bunch of different things.

And because he built it himself, he could try different things. Yeah. it was like, oh, and that feels like a very macroscopic thing. He's like doing it and measuring what happens after.

Corin: Yeah.

Abhi: Could a neural network potential do anything in that vein?

Corin: To say no to an open ended question like that feels like rude, but that doesn't seem like something I would choose a neural network potential I don't want to be one of these tech people who wanders into science and is, proclaims that simulation will solve everything. If you look at fields like chip design, what's the breakdown of total R&D like expenditure versus like simulation expenditure. I think it's somewhere in the neighborhood of 5%, like you're 19:1 in favor of actually doing things in real life. And I think that's reality is complicated. Like we should have humility as people working in simulation that like, we're not going to get everything and where you need to touch grass.

Like you need to actually do experiments and find stuff. And if your simulation is not useful, then stop wasting money and just go do the experiments. You know I think there is that being said right, simulation is a lot cheaper. So whenever you can do simulation it's much nicer too, but there's a lot of problems that I think will remain experimental for some time

Abhi: That makes sense

Ari: With the blue LEDs.

I might have watched the same video.

Abhi: Was it the Veritasium one?

Ari: Yeah. I think that, there may be a way to use some of these materials models to, predict the, stability or relative energies of these different, crystal structures and figure out, at least what ratio of these atoms do you need, But I have no idea how that translates into if you get an answer from the computer, oh, like this crystal structure looks like it might be viable.

How do you like use the deposition machine to make that specific crystal structure? I think it's, still going to be the same long arduous process. And so it's really hard to say with any confidence that, oh yeah, there's real alpha there. But I, think that there's, a hope. Yeah.

Abhi: I imagine if like the Holy Grail is achieved.

You can scale up as much as you want, stuff like that does become possible, but it's also just so far away that it's hard to imagine.

Corin: And yeah, I think I'm happy thinking like one or two steps away from where we are today. I think, you get into like pop sci territory when you start trying to think like four steps away from where we are today.

Like in the future neural network potentials will like, design you a custom drug and then fold your sheets for you and stuff. like I don't, know, like it's, our company will have either succeeded or failed long before that comes to fruition.

[01:13:57] The road to building NNP's

Abhi: When I talked about Rowan at the very beginning, I described you guys as a quantum chemistry simulation startup, which is you guys are building a front end for actually doing quantum chemistry, but you guys have also pivoted to working on your own neural network potentials entirely. Is there a reason you guys went through that pivot?

Ari: Well, in the early days of Rowan, we were. exactly this. We wanted to build a web platform to help people run their computational chemistry workflows. We had no intention of training neural network potentials. We didn't even know what they were. at least I didn't know what they were. And, at some point last fall, the Isayev group at Carnegie Mellon released their like AIMNet2 model, which is like a successor to the ANI models.

And we put it on Rowan and it got a lot of use, from our users in industry and academia. And we were honestly really impressed by its performance. And I think, to this day it remains one of the like leading neural network potentials. And as we were thinking, what would the most useful version of this look like for our users?

We realized that there wasn't anyone who is building it. And I think that's, the moment where you sort of look around and you say, who's going to solve this problem? And you're the only one there standing ready to solve the problem. And we added a third co-founder to Rowan to lead the effort.

Abhi: Has there been like new challenges associated with trying to like, like previously, I imagine that Rowan, the main focus was not scientific. It was very UI UX based. And now you guys are moving into a more like pure science direction.

Corin: Actually, I think that's a fair assumption based on what we've shipped so far, but we, from the start, we're trying to do applied R& D.

I think to give a little motivation on what the sort of backstory of Rowan was. So I did an experimental chemistry PhD. I happened to idiosyncratically have some simulation experience and software development experience. So I was able to do like a lot of simulation in support of my own research.

Which was like, it was very powerful. It felt like the future of chemistry. It felt great. But it wasn't really scalable to anyone else in my research group or department. Like at the things I was doing, I couldn't really help other people to do. Because the solution of go learn to program for three years and then come back is not a practical one for most graduate students.

And so the core mission of Rowan was like, we should build like the tools that scientists should have. The future of chemistry of molecular design of working with molecules and materials should be like simulation seems like it should be a part of that. Like if you watch science fiction movies they're simulating things in computers like we simulate things in other fields. This should be a bigger part of the day to day workflow than it is and if you work back from like building tools not just for computational scientists but from every scientist, like it can't take a week to run.

And so we were doing quantum chemistry. It was too slow. We were like frantically trying all these ideas to make it faster. So we wrote our own code from scratch. We were trying out all these approximations. We were like fiddling with all these knobs and levers. All of like none of which are even worth speaking about because they all failed and it turns out like just trying to do the same thing, but find a way to make a two orders of magnitude faster in a hundred year old scientific field is like really hard.

There's a lot of 20 percent improvements. And so then when we saw the like neural network potentials and we put them on our site and saw that was the first thing we'd done that people actually liked, it was like a light bulb moment. This is what we've been waiting for. We were just totally on the wrong tech tree.

Abhi: I think a lot of people right now are like looking at tools they worked, with during their PhD and now they're trying to build companies that like help solve some of the problems of those tools. But it seems like they're running into this issue where the incumbents don't really care that much and they're happy to use the existing tool sets and Rowan feels like it's a bet on a brand new like way of working with these tools entirely. Do you think the types of people who are trying to modernize old existing or existing tools will succeed?

Ari: You have an adoption curve that you, I don't know, people think about this in startups, you have like your, early adopters and your innovators who are willing to try new things, who are willing to learn, maybe a new way of thinking about the place of simulation and design tools in their workflows.

And I think those people, are willing to try tools like ours and tools like these other companies are building. Over time people retire and new PhDs graduate and start working in industry. And I think that is going to be the biggest driver of like tooling shift at these companies is as people, start using new tools maybe for their research, maybe their early career, and they don't already have, a preferred tool, they're willing to try things and they don't have this like preexisting mindset and bias. And so I think that a lot of it is going to be very slow.

Abhi: Corin, you've talked many times about how terrible the tools you worked with during your PhD were. I feel like when scientists talk about how terrible their tools are, it's vague.

It's never specific, oh, this, specific functionality was impossible to do. What was specifically hard about these tools?

Corin: Like if you want to run just a sample calculation. So I have a molecule, I want to optimize it and figure out what shape it's going to be.

It's like a very simple computational chemistry task. To actually do that requires like you, you draw the molecule out. Okay. That seems like a sane step of the workflow. You write an input file with it where you like memorize all these like little cryptic, like acronyms to describe how you want it to be done.

You have to put in a lot of nonstandard ones to get it to work like robustly and with like state of the art things. You, transfer all these files you've created to a remote server, somewhere. You then, invoke this massive Fortran executable. It, runs, maybe it runs out of memory, maybe it doesn't, maybe it leaks memory, maybe it doesn't.

It then returns you, a hundred megabyte text file. And then you, grep through it and try to figure out where your answer lives in the text files. You memorize the phrases that you search for, and then you like cut and paste it out and try to make sense of the results. And so you can get really good at this.

So skilled computational chemists have an army of like awk scripts that they like selectively deploy it like the right moments. But it's, there's a lot of incantations and it's like a very steppy process. So like just the whole thing you described there, like even if the calculation itself takes, 15 seconds.

It's, you're, in for a 10 to 30 minute process. And that's like once you know what you're doing by the time you get everything sorted out and analyzed.

If you're trying to mentor a younger graduate student, you're trying to bring them under your wing.

You're like, all right, let's like, let's before you step into the lab, let's see if this molecule is even going to be the right shape at all. And then you give them like a 30 point stack of like instructions for how to do this. The adoption is just very, there's a lot of friction to use this sort of startup term.

And then trying to think about a lot of clever things like, okay, what if we want to scan through like a bunch of docked poses and evaluate the strain of all the docked poses where you do that automatically in a high throughput way for every different ligand and then extract the results and make a plot.

Like it just, becomes very burdensome.

[01:21:13] Building the SolidWorks of molecular simulation

Abhi: And actually on this whole idea of like the role of simulation in research workflows, you have this really amazing blog posts that discuss how molecular simulation software should try to emulate SolidWorks, specifically that it is easy to use, doesn't attempt to replace actually building things in the real world and assists human intuition instead of replacing it.

In many ways, Rowan is trying to build the SolidWorks in molecular simulation. What do you view as the biggest blocker to actually doing that?

Corin: One is that like our model of reality is not incredibly accurate.

So sometimes SolidWorks, you model a machine and there's not a lot of intuition, but if you cut a block of metal, it will look like the block of metal you cut. It's like deterministic. Like you can say this will fit together like this. And there's this uncertainty and there's this fuzziness around how we model things and geometries and properties.

And this is what we're trying to address with neural network potentials, like a scaling up accurate simulation to a useful pace and robustness and what we've been talking about. I think the other pieces on the like user exposure and the human computer interaction piece is very hard because there's I think generations of experimental chemists who have learned several things about simulation.

One of them is this is a thing that's for experts. It's very complicated. And if you do it wrong, people get mad at you. So there's like a learned helplessness. Another one is like this is a thing that other papers or other people in my organization do. I can tell it doesn't work. I don't know why they have jobs.

But it's not for me and I don't trust anything that comes out of it. There's maybe another couple of archetypes. there's some people are this seems like it should be cool, but it's like out of my pay grade. Like I can't do that. Like I, I chose a different, like character path in my like scientific journey.

And like I've, those fields are forever closed to me. Like maybe I could have done simulation, but like maybe if I switch jobs, I'll try to pick something up. I, think part of the, exciting thing about Rowan and also a challenging thing is that we're trying to build tools that like many scientists can use, people who haven't traditionally done simulation.

And we can solve, there's like a product, like an engineering, like a robustness, like a workflow, like packaging this all, like making it simple and understandable and robust, so it gives you high quality results. But then there's also just like a really basic education piece, like, you, you may have never done simulation before.

It's going to be good at these things. It's not going to be good at these things. here it maybe is fine, but you should like really double check it carefully. Here's about how long it will take. there's we're trying to inculcate like a new behavior. That's like before you just run into the lab to make something, maybe spend ten minutes checking if it's obviously going to be a stupid idea before you commit three weeks to it. And that's like a, it's asking for a change in behavior in a way that's I think harder than I expected because I'm an early adopter. Like I was always excited about simulation and we have people who use Rowan who are like that, but part of growth will look like us being able to woo people who are on the fence and were more naturally skeptical.

Abhi: And I imagine like your first user base are going to be PhDs in chemistry.

Corin: That's right.

Abhi: Who do you imagine like the second wave will be like pure machine learning people, will it be structural biologists? I imagine structural biologists are also like probably part of the first group Yeah, who's who is the second group?

Ari: I think it's folks who are working at small companies Whether they're in the like adjacent areas of material science where they're working, you know on biotech problems I think it's you know, maybe your fresh graduates. It's your company that you can't afford one of these legacy licenses and they're looking for a solution in their company.

Corin: Yeah. Sorry, just to clarify, you said something about the first group as being like chemists, people with PhDs in chemistry? Yeah. I think you actually are underselling the magnitude of the gap here. there's, if you look at an org, like a top 20 pharma org, and you ask like how many people in this organization really use Schrodinger or a comparable tool, I think I have it from a couple of sources in different ways that the number is usually about 40 or 50.

Abhi: Across the entire org?

Corin: Across the entire org, which is…

Abhi: It's astonishing, right?

Corin: It is astonishing. Like it's a power user tool. And I think that's because it's not a tool for somebody who's a chemist. It's a tool for somebody who has a PhD in using it, essentially. That's the thing. And so even like the, I don't want to call it a fight, but one of the challenges we face is you have a PhD in just a different area of chemistry that isn't simulation.

Like you're clearly very smart. You're clearly a motivated person. Like how do we get you to find utility and to be able to benefit from what we've built? Because we know that. for people who can use simulation. We know simulation is valuable, but there's there's just what about the other a thousand chemistry PhDs there?

Like that is like just a blue ocean, I think for us and for practically everybody, of people that need great tooling.

Abhi: I like, I imagine like, speaking as someone who mainly works in like machine learning.

Corin: Yeah.

Abhi: I would be terrified of using any molecular simulation tool exactly for the reasons that you suggested that there's so much going on there that I don't truly genuinely understand.

Do you think there are like, like in my head, everything in the world of molecular simulation is an art and none of it can be distilled down to like my level, like my like common denominator level. But do you think there is like this world in which there are actually a lot of workflows that I'm just not aware of that could be commoditized and fit into my non chemist head?

Corin: There's like a scientific maturity you need to have as a workflow to be push button, like where you run a couple of checks, you're like, yeah, this will be pretty good and then you just blindly trust the output answer. We're trying to get as much stuff there as we can.

A lot of our workflows at Rowan currently work like that. And we've avoided adding ones that don't work like that just because it feels cruel. But I, I think there is a challenge of for you to use a chemistry tool you need to understand chemistry and we're okay with that I think for now being an entry requirement, like Rowan can't, and probably for a long time, won't be able to teach people chemistry from scratch.

Maybe we're trying to work with classes, so maybe we can be involved in the education process. But if you don't understand anything about chemistry or like what a PKA is or what a reaction is like, this is going to be tough. Like we're, talking different languages. I think we feel like the people we, Oh, it too.

the people who we really should work for is the people who understand chemistry, but know nothing about simulation. that feels like that's the thing we need to be able to solve. If you're a great chemist, you've never touched anything about a computer, we should be able to meet you where you are and be like, here's how we can be useful.

Here's how we can save you time. You can trust us that we'll not lie to you. We won't talk down to you or try to sell you snake oil. Like here's what we can model and here's where this will be useful. And that, that I think is like something we're still. We're still iterating on, but I think we can succeed at.

Abhi: Yeah, it's instinctively a little bit surprising that, you need to know simulation, like you need to like have done your PhD with simulation software, even if you're a chemist, actually use it because it feels like it's not, it feels for example, almost like every biologist feels like they know how to code, like to some very minor degree.

Why is there such a, like a strong gap in simulation? Is it Like a cultural problem or the tools is just really that bad all of the above?

Corin: I think it's a mix of both. I think the cultural piece is I don't know, it's probably it's rated highly and I think it's properly rated. Like there, there's, I don't know, there's like fields have cultures.

I think organic chemistry in particular is my home field is I think the among the oldest of scientific fields and the modern practice of science, like the, foundation of the modern, research university comes from Friedrich Wohler, who was an organic chemist, in the 1830s. and there's, a rich tradition.

People have been doing this research that came out in the 1990s is considered like new in chemistry. Like it's such a vibe shift to come over to ML and be, Attention Is All You Need was 2017. And like functionals from density functionals from 2010, like still aren't in some software packages.

Cause they're too recent. there's, I don't know, like people. there's a it's like an art, like there's a craft to organic chemistry in some way and there's like this is what my advisor did who had his advisor before him and his advisor before him and then, you know, that guy was a Nazi and his advisor before him like it's, it's, it's suspicious of like tools based innovation and it's like very, very difficult to break into that I think.

I don't know and I think biology is just sort of more the wild west.

[01:30:05] Simulation workflows

Abhi: I'm curious, like what, what workflows do you imagine will forever remain in the realm of art?

Corin: So one, this is on the very, the very chemical side, but like reaction prediction and like figuring out how a reaction occurs, which is what mechanism will have.

So if you have some. reactant and you have some product and you're trying to understand how they transform, like actually understanding the mechanism by which they interconvert and like, how many molecules are involved? What is the geometry like? What's a reasonable possibility? What's not a reasonable possibility?

How do we do this search? Like it's a very open ended and creative problem. I did my PhD mainly in this, so I'm biased. I think it's a really fun sort of molecular Sherlock Holmes type thing. You have to mix in computation experiments and all the right ways. Like it's very hard. That to me is like an easy example of you can do automatic transition state finding things and they work if you have the right, if you know exactly what you're looking for, they can be effective, but the problem of like, how's this going to happen?

That is it's artistic and it's I don't think there's such a thing as a systematic solution there.

Abhi: You don't think like a search, like a search algorithm could figure it out?

Corin: Like a network search? Yeah, but there's so many it blows up really fast because you can involve multiple equivalents of things.

So you can have dimers, you can have higher order things, you can have solvent involvement, you can and maybe these are long tail things, like maybe I'm, being super nitpicky here. It's it's just one of these like network combination things that like feels like it's going to be really difficult to get right.

Yeah. And it's just, I think it's also something where you need to mix in inductive biases from experiment. Yeah. probably for a while.

Abhi: That makes sense.

Corin: And so that's not easy to black box. Like you need to, have that in. I think FEP is an example of something that's difficult depending on who you talk to impossible or very difficult to like black box today, like to make just like a push button workflow.

There's no, it really depends on who you talk to. I've got strong opinions on both sides here, but that feels like something that probably is solvable, like that you should be able to do that in a push button way. And it just. isn't quite there yet in most cases.

Abhi: My actually, I always assumed free energy differences, at least, or feel like they're pretty push button, right?

You just swap out molecules. Am I, missing some major nuance there?

Ari: So with FEP, to save on simulation costs, cause these are incredibly expensive simulations to run. You want them to happen overnight. There's all of this crazy. statistics looking work to me, where they're running, the protein once and they have molecule A and molecule B and FEP will like move on the slider of like how similar is like you, you model both at once and you model it at like different points along the spectrum between the model. So it's we're modeling 30 percent A 70 percent B here. And you will like often do this with a whole series of molecules and you have them ordered by how similar they are. you go from A to B and B to C and C to D. And so there's all of this like weird statistics work that goes into it.

And then often to the, proteins are too big to be modeled overnight. And so people just start chopping off random parts of the protein to get the system to run. And you chop off part of the protein, and then you start running it and make sure it's stable. And if you chopped off the wrong part, you have to try chopping off a different part.

And, you like the theory of oh, you, run a video of protein molecule solvent, compute free energy. It sounds really great, but to make it a tractable problem for the cost of compute today and the architectures, blah, blah, blah. There's all of this like work that's done by hand and there's a lot of guess and check that goes into it.

Abhi: Yeah. So it's like setting up the system is challenging, setting up the like state by which you transform one, the system to another state, when the system to another system is challenging and that's really challenging to automate.

Corin: Yeah. I think so. I think there's a lot of vibes based analysis, and there's just a paper out yesterday showing that like the input pose ends up mattering actually, like you hope it doesn't, but it does.

So you have to make sure that you get the pose right. You worry about proton transfer if you have potentially basic sites and acidic sites where you get proton transfer from the ligand to the protein. And it like, it's just like changing between charge states is hard. these are all things that you're like, yeah, you, could solve that.

One is you could just solve it if you could run it faster. If you could get to more time, we wouldn't care so much about the input pose. Otherwise if you run it for too long, sometimes the protein starts to fall apart and unfold or the ligand drifts out, so you have to make sure it doesn't do that.

I think these are all like, you should solve these, maybe neural network potentials will solve these, maybe other advances will, but it's just finicky.

Abhi: So like you've, discussed the things that will probably remain art for quite some time. What about the things that you think are lower hanging fruit than probably most chemists actually realize? And it's like fully within the realm of their capability.

If they were handed a good enough tool,

Corin: I think one of the most basic yet underrated sort of things to do is just to understand the confirmation, the shape of a molecule that drives so much of how it behaves. That's such a foundational thing. We take it for granted for everyday objects, but you can be working on a molecule and just not even know what shape it is.

Like rigorously, I don't know, for an entire project, I've definitely done that before. Yeah, I think this is. something that's like pretty easy to predict, pretty useful, like immediately intuitive. And especially as we get to these larger molecules, like macrocycles, these beyond rule of five things that can target PPIs, like interesting peptide systems, actually understanding the shape and the trends, how substitution affects like confirmation and properties are like, it's really hard.

Like it's not obvious at all. And it's like an outstanding challenge in the field, like understanding, like what shape is this even going to be? And if I switch this ancillary group here, will this totally remap like the overall, like confirmation of my macro cycle? And this is useful, like this, matters a lot and is like an easy computational problem more or less.

Abhi: This is going to betray how little I know about chemistry, but my initial impression of a lot of molecules is that they're incredibly flexible. So how much do you really gain from knowing like a few states?

Corin: Yeah. So some molecules truly are exceptionally flexible, like a massive fatty acid that is just like a snake of carbons.

That is intrinsically disordered small molecules, so to speak, and there I think it's not incredibly helpful because it's just gonna wiggle around no matter what. oftentimes there are, there's a lot of accessible states on the energy landscape, but if you're, say, 1 percent or .1 percent of the distribution, if you're a couple of kcals per mole up in energy, then if you bind in that confirmation, you're disfavoring the binding pose because you have to distort to get into the bound position. So it's like a really common way to gain some efficacy in a small molecule drug to make like to freeze something to lock it in the bound pose. so you can, using, this is one of the things people use quantum chemistry for now is you can create some map of all the different potential poses, like what the bound one is, where, how much higher in energy it is, and how much you're losing by not being in the ground state pose.

And then you just engineer a molecule that has the right pose. And that's like a, again, that's one of those things that like experts, people with PhDs do, but regular people don't do that they could and should be doing like it. It's and it's so intuitive, like it's so like directly maps onto things that are like intelligible.

Abhi: With small molecules, is there, a, like a strong suspicion that once you actually introduce it to the body, it's going to dramatically change?

Corin: I think there's so many fewer degrees of freedom that, I think it's much less likely, there's small molecules in some ways are just simpler. like you, you care about if it will be protonated or not, but you're not going to get like mass electrolyte effects or some sort of you're much less likely to get complexes.

Like you can always stick to albumin or you can binds to proteins. But I think it's the confirmations in water and in the body often are quite similar into a first approximation, pretty much the same.

Abhi: Over the topic of structure optimization, you've mentioned in a prior blog post that you could optimize the structure of azithromycin, a common antibiotic in five minutes using an open source neural network potential method.

Whereas it would take nine hours using a DFT based method, both of them and having similar accuracy at the very end. What's the ultimate payoff of this for a chemist? Is it like just much faster iteration time, less spent, less time spent in lead optimization. Something else I'm not thinking of.

Corin: I think this makes a ton of sense in the context of a, SolidWorks type vision, like structure optimization is like, just like ground zero for anything else you do.

If you're going to dock it, if you're going to try to figure out how it reacts or something, you need to start with the right structure in the first place. At the margin, like how fast something runs, I guess it affects your cloud spend a little bit, but it's not a huge deal. I think what matters is when you can hit this sort of like order of magnitude changes that engender a shift in behavior.

So a calculation that takes one month to run is of no use to anyone outside academia, a calculation that takes overnight to run like nine hours is different. That's a I'll check back on this tomorrow, and we can talk about this in group meeting next week. A calculation that like runs while you get a cup of coffee or use the bathroom is like a, oh, I can do this and understand it today.

And then the goal is like something that like just responds like intuitively, like a real time, like you draw your structure and you instantly get the right thing. I think the utility to the end user from like an insight and from a, design tools perspective, like increases exponentially as you like decrease the time it takes to run something.

And we do see this like with some of our users at Rowan, where they'll. They'll be able to just sit and like experiment, they do their design. Simulate, think, design, simulate, think, cycle, back and forth with the computer. When the, property they're optimizing over only takes, 30 seconds to run, I think people do this a decent amount with redox potential predictions.

That's more on the synthesis side of things, but trying to figure out how easy it will be to add or remove electrons to something. There, we like, we have a good solution that runs in virtually no time. And so you, if you want a molecule that has a specific redox, try it, try something else. Try something else.

Try something else. And this this starts to be like what you hope the future looks like in more and more areas. Like it'd be great if we could do that with, drug binding affinity as well, where you're like, oh, that didn't bind so well. Oh, what if we added something here?

Like that would be an amazing future for drug discovery, but it's, difficult to get there.

[01:41:06] The role of computational chemistry

Abhi: Yeah, that makes sense. Vaguely, you feel like this like culture shift of like, waiting for a month to calculate something so you just ignore it entirely versus like it takes a few hours to calculate something so you actually do seriously investigate it.

Do you think something like that happened with the rise of Schrodinger? and are there like lessons to be learned as to like how, drugs, were developed and like how they, that, that changed? I imagine it's like it's the answer to this question is known by like very few people but I'm curious whether you have a like an insider thoughts on that.

Corin: Don't know if I have like secret knowledge here. I think there, there was a shift, like you can look at the start of Vertex, you can look at the dawn of Schrodinger as like computers being useful in drug discovery for the first time that like, you can go back to the 50s and 60s and I don't, computers weren't useful for much in the simulation realm back then.

Then there started to be these like both internal tools and external tools sort of things becoming mass market, right? People built the computational teams or like new companies like Vertex built around computation. And then this, you start to have an expert who's like a scientist who like uses computation to support various things.

And I've heard the role of the computational chemist right now often described as like just a, like a helper. Like you support the med chemist. Sometimes it looks like do modeling. Sometimes it looks like doing data processing, like building little like ML models on the data. I had some guys say like, I'll do whatever is useful.

Sometimes that looks like getting coffee. Like it's a, it's very much a sort of role where you're trying a lot of things out. And I think we've, we have seen that this is now a part of like virtually every organization, like every top pharma company uses Schrodinger pretty much like almost every team involves a computational person, but then the like utility of that role has plateaued like a little bit, not maybe fully plateaued, but it's tailed off.

Like it's not like more and more people are becoming computational within these organizations.

Abhi: Does the average, like a computational chemist, think they are useful at like the average pharmaceutical company? Or are they more like, I hope someday I could be useful.

Corin: I think they are useful and I think they're useful, but I think there's a lot of…I don't have an axe to grind with these people at all.

I think the world of them and I think they're often very humble. It's like they understand that their models are flawed and that they, if they want to remain trusted and useful, which they, by and large do, like you, you're, honest about that. You're like, hey, like this is what the modeling predicts.

We think this will be good. There is often attention where perception of computation often lags reality. So like experimentalists are always maybe unjustifiably skeptical and computational scientists are always maybe a little bit too optimistic. So there's some sort of like dialectic there, I do think people have a clear sense of their role and are happy to be useful.

[01:44:06] The future of NNP's

Abhi: That makes sense.

Corin: We think of things, much more from the small molecule point of view. Cause that's, it's much more chemical. And a lot of the arguments I've, we've talked about have been around how this, what we're building will be useful for chemistry.

One bear case for Rowan would just be like, trivially, everything becomes gene editing and we never need to think about atoms ever again. Do you have a, do you have a take here?

Abhi: If you look at all of, like human biochemical processes as like a flowchart.

I think you're the one who said that small molecules are the equivalent of cutting out, one box to another box in terms of interactions. And the addition of proteins or genetic elements are like adding in a new box entirely. They do feel like they're playing in different areas. just like beyond, like the, flawed analogy itself.

There is also this like, small molecules are small. They can slip into places where like larger things just cannot reach. I do think there is this world where. small molecules do seem to be getting larger. A lot of, proteomics based drugs seem to be tending towards smaller. Like you're going from antibodies to nanobodies.

There may be a happy medium there somewhere where everything is like macrocycles. but who, who, knows for sure? I, think a lot of revolutions in therapeutics are very much like non iterative. They just appear out of nowhere. and that may very well be the case again here.

Corin: Yeah, it's interesting to think about. I think one of the things that I've been dwelling on a lot the past few weeks is like the, one of the big differences I think between proteins and biologics based approaches and like small molecules, it's just like the, almost like the information density per unit area is so much higher.

And I see a lot of the trends. With unnatural amino acids, with macro cycles, with, everything is getting more complicated. And it makes sense that as we, we want to go beyond 20 amino acids. We want to, access, we want to be able to turn more knobs essentially when we're optimizing something.

And this is I guess even the large molecules, the design starts to feel less evolutionary and more like a small molecule problem. But that doesn't necessarily mean that the small molecule tools are going to be the right answer for everything.

Abhi: That's an interesting way of phrasing it. I think traditionally I very much think of, I think a lot of like protein design is like trying to obey the laws of human physiology, a fair bit more than small molecules.

I think you've, described molecules as zero day exploits. I think proteins are very much like you're, we're trying to fit the same like binding pockets that like nature already has something that binds to it. I think the advent of non canonical amino acids does change the game a fair bit.

And I think I am probably the person least well suited to opine on like where that actually leads us. But it does feel like it's heading in interesting directions. There are like protein modeling papers that are trying to account for the, the existence of non canonical amino acids. I think it's still very much early days.

But I do feel like those are, that is one of the areas where I feel like, dynamics are really the only thing you have. Because you don't have this, decades long historical collection of, non canonical amino acids.

Corin: Yeah, I think that's right. I guess falling back on physics is like a decent, When nothing else works, hopefully physics will work.

Even it's maybe not the best tool for every job, it's at least it's reliable.

Abhi: When, like general AI companies are developing like their own foundation models, like you have Anthropic with Claude, OpenAI with ChatGPT, they're all hosted on their own platform.

There's very little interconnection between any one, one bot and another bot. Do you imagine a similar phenomenon will pop up with neural network potentials? You'll have this community of like open source neural network potentials created by some well meaning academic. And you'll also have these gamut of startups that are developing their own neural network potentials and no one will want to play nicely with one another.

Ari: I think there's going to be a lot of variance startup to startup. I think we're already starting to see this. Startups like Orbital Materials have been open sourcing their NNPs with really permissive licenses. And so I think like it, I could imagine a world where, a startup decides, we do train neural network potential sometimes, but we've decided that, it's not a core part of our strategy.

It doesn't help us build power as a business. And so we're gonna open source that work. And I think that this is what, a company like Meta FAIR Chem is doing too..

Abhi: I didn't know, Facebook had a, like AI, like neuro network potential research group.

Ari: They do, and I think it's because that there's a story, they're working on materials for their new glasses. They need to be able to model materials really well. Maybe somehow they're like, we should do basic research on materials in some part of the Meta organization. And, that's one myth for how the, FAIR-Chem people started training NNPs.

I don't know if it's a true myth. But, they're open sourcing their NNPs so far to they're saying, this is a thing that we want for our business, but it's not a core part of, helping connect people, which is, Meta's goal, the NNPs are completely tangential to that. And so they're happy to open source them.

I think there are also some startups who are happy to open source them. really going after the, we're going to be like a model building and architecture company and I would be so surprised if those people open source their models unless they think they can build some sort of great open source business around that Databricks or something.

I don't know. I think it would be hard. Yeah, so I would imagine a fractured future but still high quality open source models.

Abhi: I think there is it's like a genuine possibility that a lot of these protein foundation companies become like maybe winner take all situations where one protein model is genuinely good enough to model the full universe of all possible proteins.

Do you think that'll ever be the case in the world of simulation?

Corin: You always have some tradeoff between like specificity and generality, where you can, we've talked a bit about speeding up like inference here and you can imagine would you take a 500 billion parameter model that does all elements and all spin states and all like confirmations? Or, would you rather have a 20 million parameter model that's just really great at amino acids. There's definitely some applications for which like a much smaller, much faster model that's good at specific things would be advantageous.

I think at the limit, a lot of papers right now are like, just fine tuned on a single protein. And it seems like being able to quickly, not have to retrain a whole new model to modify your system. Like it seems like you want some amount of transferability, to say that the protein people get one model, the OLED people get a different model that doesn't seem ridiculous to me. And maybe you can distill the massive model down in some way. I don't know. AI people are great at their sneaky tricks.

[01:51:23] Selling to scientists

Abhi: Switching away from like the scientific discussion for a bit, I feel like I've heard multiple times that scientists are very often terrible customers because their needs are often so hyper specific and they also have the least money to actually give you to satisfy those needs.

if you agree with this, are you often trying to convince scientists that Rowan is worth it? Or do you try and target more executive level people first, or do you disagree with the concept entirely?

Corin: I think there's definitely some truth there. Like science is a, tough field. There's a lot of details to get right.

Like it's not, there's plenty of like horizontal SAS plays that are, I think are easier and like simpler. There's a bigger TAM, et cetera, et cetera. Like you, I think we really like what we do. I really like scientists. I like working with scientists. It makes it easy to get up in the morning and want to go to user meetings and scheme about how to make it better.

And I do think there is like it is a little bit cope that scientists have no money to spend because so much money is spent on science. Like this is like just the ballpark site, right? Like 200-250 billion spent on drug design per year. Oh, I can't make money in this. I'm not saying that it's easy to start a scientific software business, but it's not that nobody cares about this.

It's not that no money flows through this. there should be a way if you're doing good work that matters to people to make a great business here and I think it's, every business is a bad business for some reason or another, this one has its challenges, but there's, definitely a way to win.

Ari: Yeah, I think selling into a big organization, you really need to get a lot of people on your side to, to, close a big deal. And that. I don't think that we've really done this in a way that, I'm, dreaming of doing yet, but I think, you want the users to love your product, to be willing to use it.

And you also want, these, executive strategic decision makers to understand, how this is important, be on board with whatever, spend, they're going to commit to this tool. and I think that. With any company, there are all these communication challenges that are often underappreciated by really technical thinkers, where if you're an engineer, you might think I just need to make my product better and then I'll have a great business.

But your product is only as good as the people who use it. And probably the people who can pay for, those, guys to purchase tools.

Abhi: Have you guys seen the, like that Spring Discovery tweet thread about, Spring Discovery is like a high throughput screening platform for looking at brightfields images and the company's shutting down like after a decade.

Corin: Yes, I did see this.

Abhi: Yeah, I like, he had this like, really interesting tweet thread about how, like scientists love the software. It's deployed to five out of, I think 20 of the big pharma companies, a bunch of leading academic groups, Broad Institute, UCSF, University of Toronto.

But they just didn't make enough money to stay afloat. I think there's this, like the hard part of developing scientific software is like capturing the value that you're actually bringing the scientists. Do you think there's like a big failure mode that a lot of scientific software companies, especially in the simulation space run into with regards to this.

And are there like ways to alleviate it?

Corin: Yeah. I think so two failure modes that we've seen a lot, at least we've thought about a lot with Rowan is like, one is that you labor in ignominy and perish in obscurity, that like you can do the world's greatest work behind the closed walls of your company and in your platform, but if you're not, you need to be able to get other people excited about it, and you need to be able to communicate to the world what you've done, and you're like, you can't just do that when you're knocking on doors asking for checks, like it's too late at that point, like you need to be like legibly exciting to other people such that they like are trust you and want to work with you.

I think another issue is just like it's trying to connect what you do to like real outcomes, you know because the bottom line like what makes people money what creates shareholder value is, particularly in drug discovery so there's so many sort of RL steps away from what you do in like early stage R& D, like we make it faster for scientists to screen compounds or do brightfield imaging, which helps improve their capabilities, which will help increase our odds of not failing a clinical trials, which will increase our like, it's a, it's difficult to put numbers on that and to really...

Abhi: you're very far away from the money.

Corin: You're very far. That's a much better way to say it. Yeah. And so yeah. You have to be object level really good at doing the actual thing that you do and that is like what gets you in the door, right? That's how you get it. And then I think you have to be really good at being honest with yourself and like finding a way to justify the value that you're creating and to actually like to be able to say that in the negotiating table.

I think a lot of people fail at that and it's tough to really peer inside and figure out is it, who's wrong? Because if I think I'm providing you a million dollars of value and you're paying me 50k a year like, are you wrong about my value or am I wrong about my value because something's not right there. And I think that's super case by case. Something being nice something that scientists like to use is like necessary, but not sufficient to build a great business.

Abhi: How do you convince people that like what you've built is useful? I think like Rowan's an incredibly aesthetic piece of software, but how do you actually connect that to saying this will provide to you tangible value that's worth how much we're charging you?

Ari: I think you have to connect the simulation or whatever tool to a problem that someone is currently facing or at least cares about. Maybe it's not a current problem they're facing, but one thing that we've had a lot of success with is there are people who are trying to tune redox potentials of their molecules.

And they're saying, I really care about this, property of my molecules. How can I design molecules where this property is different? If we say look, you can draw the molecule, press this button. And it'll tell you that property, with decent accuracy, then that's really useful. People are like, Oh, like I'll start using this in my workflow.

If there are too many steps, you have to run this workflow and then you have to do some statistics and then you have to go to this other tool and do something. Then, the values not connecting all the way to the thing that the end users actually caring about, they don't know if the values in the software or the statistics or whatever other step you have them do.

And so I think it's a much harder thing to communicate about. And so I think we think about is like, okay, what are the problems people have and think about and how do we work to make simulation actually solve that problem all the way. And

Corin: It's something relatively few other companies are doing.

I think of trying to go the extra mile of just like at the instant in your life where you have this question that you could answer with simulation, we want there to be something in Rowan that's like answer blank question. You click a button and then it gives you the answer, like having a lot of empathy for the user there, which I think maybe comes from not being a computationalist by training, but there's, yeah, the best and worst thing about selling to scientists I think overall it's good.

It's they're very data driven. So we predict things and at least to the extent that we've done sales with success, it's like, you want to predict this thing? Let's try predicting it. The predictions are good. And they're like, Oh, the predictions are good. All right. Then, that's like a, you, can't really hide much in that.

Abhi: I guess my initial thought is, they look at the results, and they're like, oh, what if it's wrong in some other area? They're, like, they're so data driven, that they're unwilling to accept anything that falls under their normal distribution of, good, software to use.

Ari: I think a lot of people when they're testing some new tool, they resort to the standard known test cases that they've been trained on. So if this is like a new language model, maybe one of the test cases that people throw at it is they'll ask it how many R's are in the word strawberry.

And it's now like this informal benchmark that whenever a new model comes out, people go and immediately the first thing they ask it is how many R's are in the word strawberry. And then if the model says two, they're like, this is a horrible model. If it says three, they ask it another question. And I think when people are using our software for the first time, they do a similar thing.

Abhi: There's a set of sanity checks that they've informally created for themselves.

Ari: Exactly. And so maybe if they're computing redox potentials, they'll they remember from a textbook, I've memorized the redox potential of benzene, they draw benzene, and then they compare the, number, maybe they don't, memorize, exactly that.

But I think that they have these built in test cases, and if you fail on one of those test cases, they'll discount your software immediately and you have to pass those first, basic test cases. And then once you've done that, you have this baseline of trust that then you're building on and you can, work on actually okay, let's evaluate a data set.

And, go from there.

Corin: I think this is true for everything though. this is, there's some stat about Uber that like the first, one or two rides you have, like dictates how you think of the app, like whether you churn or not, like pretty quickly. If you open the app for the first time and it's no cars within 20 minutes, you're like, this is a piece of crap, I'm not using this.

And it's, I think it's the same with Facebook, like the number of, friends you have in your first week pretty strongly dictates whether or not you stick or not, I think it's just human nature, I don't want to waste time on something that's obviously bad. And that's very rational.

[02:01:41] What would you spend 200 million on?

Abhi: I think like the, like one of the, like the last questions I have is what do you think is the current bottleneck to making better neural network potentials outright? Is it data set quality, model size, data set diversity, or something else entirely?

And when you're answering this question, assume you've been given like 200 million by an anonymous donor to like push, push the state of the art as much as possible, like what could you get?

Ari: I think that there's some architecture questions that we would want to run experiments on. These are things like message passing. How much message passing do you do? Do you need to strictly enforce SO3 equivariance and do you enforce it at the beginning and end of your model? Do you enforce it at every layer or can you throw it out altogether? I think you want to ask those questions as you try to scale the models to, to figure out, if I'm going to start training this on more and more data, which one will actually scale and, more data means less error, better model. I think that there is a lot of data set generation work that remains to be done. I personally think that like we should generate a metadynamics data with range separated hybrids on all the systems we care about, and we should train on that.

Everyone's got their own opinions in this. I think, if we can find a way to do multi fidelity learning to use lower quality data sets, that'd be super valuable. And then I think as we try to scale to bigger and bigger systems, you walk into these questions about coarse graining. So can I coarse grain out hydrogens, residues, solvents?

Those are like the easy, coarse graining, beach heads maybe. And then I think, when you're thinking about using these neural network potentials for molecular dynamics. So when you start walking into these sampling questions, if I'm trying to actually now, run a whole video on a system, am I just running normal MD?

Am I going to run metadynamics? Can we do some sort of like Monte Carlo step generation and acceptance criteria that will recreate my potential energy surface? And I think that, there are going to be a number of great research teams that sort of start in the next decade and we'll work on answering all of these questions.

And if someone gave me a big pile of money, I would just start now, let's try to, get definitive answers to each of those questions.

Abhi: On that note, I'd like to ask you your opinion afterwards. Do you think the current resources like ATLAS and MD Repo are like they're tending towards a good direction or you think there's some fundamental, like failure point?

Corin: ATLAS and MD Repo, they're both on MD, right? Like just regular, like Amber or something like that. Yeah. I think, it's just debatable exactly what the quality of a ginormous MD repo is. If we have skepticism, which we do about like force field quality, like it's like a lot of these early QM datasets where they generate a ton of data at a pretty poor level of theory.

Like it's great as an ML exercise to see if you can fit the data. It doesn't move the needle that much in terms of ultimate quality because you can learn something on inaccurate data. It's still not going to match experiment very well. So yeah, I think there's definitely value, but I'd much rather just do the MD properly.

I think that is a self serving and easy to say answer it to not have to do it. But that's my feeling.

Abhi: As of right now, there's no real data set out there that you think approaches like the level of quality and quantity that you want?

Corin: Yeah, I think that's right. I think to underscore something Ari mentioned briefly , I wish our ML co founder were here because he could make this point much better, but there's not a GPT quality architecture for molecules or I think like other like 3D graph problems yet. We can't just dump in a ton of data and expect like the H100's to all just go brrr and it to like magically get better. Like I think it's possible, but there's still some sort of like scale architecture problems, which are not fully worked through yet.

And I think the only way to do that is to try scaling and to figure out what works. And I think it's very likely that things will function okay, but like...

Abhi: It hasn't been tested yet.

Corin: Yeah, we can't just like trivially spend 200 million dollars on CPU time, generate like a crazy amount of like terabytes of data, and then just like press play.

Abhi: They are like consortia that are trying to like brrr and try to develop huge amounts of data. Do you think there's a very genuine possibility that the datasets they end up with are not useful?

Corin: I, this is one of my big concerns with the materials science sort of NNP world is that, so Materials Project.

It's a really cool, so it's like a consortium type thing. Like they, they have a standard set of theory. They like have worked to put together like a big database. They have huge collections of data. And then there's also some papers showing that like actually the settings they chose were a little too cheap.

And you get these like sneaky errors here. Like here's all these issues. Some of the predictions are fine. Some aren't fine. And so then you're like... I guess you have like very highly correlated error. For better or worse in sort of molecular land, it's everyone for themselves. And some people do really good jobs and some people do really bad jobs, but there's not the same sort of single point of failure.

And so I don't know what the future should be. we generate our own data sets internally just cause so we can have tight control over that. I know some companies do that. Some don't do that. I think it's...

Abhi: It's the final results will dictate who is actually the winner.

Corin: Yeah, ultimately, we don't, we're not doing this because we think it's just nifty, and we're trying to waste investor money.

We want to make an impact on real things, and everyone has a different tactics of how to get there, but we're all trying to get to the same place. it'll be fun over the next year or two to see how it all shakes down.

Abhi: And, for you, Corin, the 200 million, would you also focus on just, generating high quality data, or do you think there's somewhere else?

Corin: Yeah, I think I agree with everything Ari said, and to frame it a little bit differently: there's this massive gap right now between how fast MD is and how fast neural network potentials are and how fast they need to be to do everything MD does.

So a lot of the things we do today are like QM, like DFT, but way faster. And that's super cool. But a lot of the really awesome things we want to do are like MD, but way more accurate. And you can say maybe we're four orders of magnitude off in speed. That might, I think that's probably about right.

And there's maybe eight different ways we could imagine getting an order of magnitude speed increase. And so there's some sort of problem of we just need to get like half of them to work. Okay. And that's I don't, it's not like blue sky research. It's not like figure out like a cure for cancer.

It's like an applied research, ML engineering, like an algorithm optimization problem, like some combination of figure out how to get our accuracy, like a bit higher with more data, figure out how to scale to really large graphs. And then it's just make it be faster. And I think that's, like just, I don't know, just spin up some little teams. Like a 2018 OpenAI. Right.

Abhi: Yeah, that's the dream.

Corin: That's what we're all working for.

Abhi: Yeah. Well, thank you so much for coming on the show and going through three hours of talking. Yeah, thank you so much.

Corin: Thanks for having us.

​