Introduction
Back in early May, I flew to North Carolina to interview Alex Rubinsteyn (left) and Benjamin Vincent (right). Together, they run the Personalized Immunotherapy Research Lab (PIRL) at UNC-Chapel Hill. It would be difficult for me to emphasize the degree to which Alex and Benjamin are in a league of their own amongst almost any other researcher I have ever met, demonstrating a degree of translational ambition, scientific insight, clarity of thought, and just genuine kindness that one rarely sees. And each time I stumble across a person like this, I try my best to get them in front of a camera. This is the first interview I’ve ever done where I traveled specifically for a guest(s), and I’m very, very glad I did it.
We’re in an interesting moment with cancer vaccines. Between Sid Sijbrandij’s ‘founder mode’ journey on his cancer (which involved cancer vaccines), and the dog cancer vaccine story, a lot of eyes are being directed here. As luck would have it, Alex and Benjamin have spent a fairly high fraction of their career in this exact field, and have contributed to some of the most foundation work in it: MHCflurry 2.0, OpenVax, LENS, and plenty of other work. In fact, Alex actually helped run a neoantigen cancer vaccine trial a few years back! So obviously, they’d make for great conversation here. We talked for three hours about this area, eventually stretching beyond cancer vaccines and discussing the grander field of personalized immunotherapy. This is comfortably the longest episode I’ve ever recorded, and I suspect we could have gone for an hour longer.
Finally: this is a complicated subject, and we really get into the details. Because of it, it may be worth skimming a companion cancer vaccine essay I released a month back. Many subjects from the article are discussed + expanded upon in the podcast!
Timestamps:
[00:00:26] Introduction
[00:01:26] What cancer vaccines are, and the “easy living drug” dream
[00:04:34] TAAs vs neoantigens — and why the field switched
[00:08:18] MHC, HLA, and neoantigens, defined
[00:12:57] The 100,000 MHCs per cell and the information-theoretic problem
[00:15:28] Why a cancer can’t just drop its MHC
[00:20:18] Immunopeptidomics: why predicting antigens isn’t knowing them
[00:28:57] How many targets you actually need, and the HLA-loss problem
[00:30:22] Can a tumor mutate its own MHC?
[00:31:34] What are tumor-associated antigens and cancer-testis antigens?[00:34:09] The thymus, AIRE, and why T cells were never trained against CTAs
[00:37:35] Why “normal” proteins is still an unfinished reference problem
[00:41:05] The short list of T-cell killing success stories
[00:43:20] Did the old vaccine literature just need Keytruda?
[00:47:31] Why not cut out the middleman and engineer the T cells?
[00:50:09] Autologous vs affinity-enhanced TCRs, and a database-matched middle path
[00:54:23] Can we TCR-T and CAR-T in-vivo?
[01:02:08] Why whole-tumor and lysate vaccines never worked
[01:04:46] Immunodominance, or why a great vaccine can still do nothing
[01:10:38] Can you predict immunogenicity with ML?
[01:12:24] What you tell the FDA for hyperpersonalized drugs
[01:14:14] How secret is the computational pipeline?
[01:17:40] Why the field desperately needs real benchmarking
[01:21:57] Moderna and BioNTech: are the success stories real?
[01:30:04] Tumor mutational burden, and scraping the bottom for glioblastoma
[01:36:46] If cancer is so heterogeneous, why does any of this work?
[01:38:40] The hyper-optimistic case: metastatic disease and antigen spreading
[01:43:00] Antigenic drift, driver variants, and the 2030s adaptive vaccine
[01:48:02] Why cell therapy at all, instead of antibodies?
[01:52:43] Neoantigen vaccines as a categorical departure for the FDA
[02:00:03] The antigen-selection safety logic, and why it’s flawed
[02:04:42] Running investigator initiated trials (which exist in the US!)
[02:19:40] The tragedy of the commons in cancer vaccine trials
[02:22:38] How Ben and Alex met (on Twitter!)
[02:27:49] Founder-mode oncology and the rise of concierge cancer care
[02:31:50] How good is concierge oncology, really? The Sid index case
[02:40:03] Why old precision oncology was useless, and why now is different
[02:43:07] LLMs, and patients advocating for their own testing
[02:47:30] The molecular-testing trial that should exist
[02:48:50] Single-cell long-read as the one true assay
[02:49:56] What would you do with $100M equity-free?
[02:53:27] The automated box: tumor in, RNA therapeutic out
[02:55:58] Why identifying targets is the easy part
Transcript
[00:00:26] Introduction
Abhi: Today I’m in beautiful North Carolina, sitting with Alex Rubinsteyn and Ben Vincent, the two PIs who together run the Personalized Immunotherapy Lab at UNC Chapel Hill. They are among the few academics I know who operate at the intersection of cancer, machine learning, and aggressive clinical translation, and I learn something new almost every time I talk to them. Today we’re going to talk about cancer vaccines, neoantigen identification, the cancer care of the ultra-wealthy, and a lot more. Thank you both for coming onto the podcast.
Ben: Yeah, thank you for having us. It’s fantastic.
Abhi: So the obvious first question is that cancer vaccines are a very zeitgeisty subject right now, what with Rosie’s dog and Sid, the founder of GitLab, and his whole cancer journey. I don’t think I quite naively understood what cancer vaccines even are until I talked to both of you. Just for background context for the rest of this episode, could you walk me through what exactly cancer vaccines are and their historical progression over the last two or so decades?
[00:01:26] What cancer vaccines are, and the “easy living drug” dream
Alex: Sure. It’s an exciting therapy. It completely nerd-sniped me out of computer science and into biology. The basic idea is that you want to reprogram the immune system to recognize a cancer and distinguish between cancer cells and healthy cells, and that reprogramming happens by giving examples of, “Okay, this is something that’s bad. Hey, immune system, if you see this bad thing, you should kill it.” And if that works, then it’s a sort of easy living drug. Your immune system now recognizes cancer as distinct from healthy, it kills it on sight, that keeps happening, and you don’t need ongoing treatment. And it works through a mechanism that we already know is effective and safe.
We can clear viruses, we can clear pathogens of various sorts, and so we just treat the cancer as a pathogen, and it’s gone. It’s a very nice idea, and I think people recognized the potential of that general trend way before we had the knowledge and technological components to realize it. So it’s been kind of a long road to figure out what you need to do to start making that actually clinically effective. Do you agree with that description?
Ben: I do agree with that, and I think theoretically, as a medicine, it’s almost mythologically compelling, because all cancer cells are different than healthy normal cells — otherwise they wouldn’t be cancer, and they can’t be cancer. So if you can train the immune system to recognize those differences, then every cancer carries within itself, in its genome or in its transcriptome, the seeds of its own destruction. This kind of antigen-directed immunotherapy may be a general solution for cancers across cancer types and clinical contexts.
Abhi: And so the typical construction of a cancer vaccine is that you want to include some fragments of whatever makes a cancer a cancer, alongside an adjuvant that alerts the immune system to go to these antigens, pick them up, present them to the T cells, and then mount an immune response to the cancer. The obvious question is, what should these antigens actually be? You have a really good presentation about this, Alex, where you talk about the actual cancer vaccine pipeline, and it seems like there are two directions you can take. You can find proteins that are typically expressed by cancer cells, or expressed outside their usual place, or you can find what are called neoantigens — antigens uniquely presented by cancer cells. I think the cancer field pre-2014 was very focused on TAAs, or tumor-associated antigens. I’d like to get your take on what this field looks like. If you look at where most cancer research is today, it seems very focused on neoantigens, but you make some pretty interesting arguments that TAAs are actually very useful and we should be paying attention to them more.
[00:04:34] TAAs vs neoantigens — and why the field switched
Alex: Yeah. Maybe one thing to start with is, why did you have a progression where we focus on one thing and then switch focus to the other thing? I think it’s not really about the intrinsic properties of those possible targets. It’s really a technological shift. The Illumina HiSeq sent everyone looking for cancer mutations, and once we could find them, a couple of labs really quickly started figuring out, “Oh, we could take this whole framework of existing cancer vaccines and load them with mutations. That sounds better, it sounds more tumor-specific.” And so everyone said, “Okay, great, now we’re going to do mutational targeting in our vaccines.”
But if you zoom out and you don’t really care exactly how you found the thing, the range of mutations is bigger than what you could easily see with short-read sequencing. Neoantigens are actually a really big category. We mostly look at just a couple of types of those. And then the expressional weirdness of a tumor is also really, really big. Tumors do all kinds of weird stuff. They reactivate endogenous retroviruses. They splice incorrectly. They express silenced reproductive-associated genes all the time. They express intergenic regions of the genome that aren’t really protein-coding genes but sometimes create weird repetitive elements.
So there’s a lot of ways that a cancer is bizarre, and I think all you really care about, from the point of view of killing that cancer cell, is having something distinguishing that’s tumor-specific, which you know is presented — that the cancer cells are all going to be making and presenting that target. And so when you get into people trying to operationalize cancer vaccines, they treat these as really different beasts. They’re like, “Oh, are we going to do a TAA vaccine or a neoantigen vaccine?” And I think it actually doesn’t really matter. You just want distinctive tumor surface targets. Which, from the point of view of vaccines — we haven’t talked about the immunology. We should.
Abhi: Yes.
Alex: Because I’m going to start saying MHC presentation, and some of the audience is going to feel lost. But what you want is things on the MHC of a tumor cell that are not on the MHC of any healthy tissue that you care about. That can be an expressional target. It could be a protein like PRAME that’s really reproductive in function. It can be a virus — if you have a viral origin for your cancer, HPV is a great target. A lot of the cancer vaccine success stories are with HPV. And it could also be mutational in a really broad sense. It could be a splicing error, it could be all sorts of stuff. So I think as we move out of the exploratory tech-build phase of cancer vaccines towards clinically effective ones, all of this is just going to collapse into: find all the targets that are actionable in the sample, then prioritize those as a unified set, pick the best ones, and go after those.
Ben: It’s probably important to say that it’s highly likely these will be largely individual, person to person. So if you have all pancreatic cancer patients whose cancers have a certain, say, RAS mutation, which is very common, only a minority of those people will have the right HLA to even potentially present a certain mutant RAS peptide. And if they have the right HLA to present it, the protein actually has to be processed and presented at enough copies of the peptide-MHC on the cell surface to be recognized by T cells.
Abhi: So maybe we do talk about the immunology and define a little bit of the background.
[00:08:18] MHC, HLA, and neoantigens, defined
Abhi: Let’s define the words MHC, HLA, and perhaps neoantigen.
Alex: Okay. Ben, you should do it. I can feel myself wanting to go deep in the weeds. I feel like Ben is good at staying on task.
Ben: So, cells of the body all express what’s called class I MHC. MHC stands for major histocompatibility complex. HLA stands for human leukocyte antigen, and HLA is the MHC of humans. So you’ll hear us use those terms interchangeably, but really HLA is the MHC in humans. Those are molecules that bind to short peptides and present those peptides on the cell surface to then be recognized by T cells. There are actually two classes of classical MHCs. Class I presents short peptide epitopes to CD8-positive T cells, which are the cytotoxic T cells, or the soldier cells of the immune system. Class II MHC presents slightly longer peptides to CD4-positive T cells, classically known as helper T cells — but there are some reports now in cancer that those can cause direct cytotoxicity just like CD8s. Immunology is very complicated; that’s part of what makes it so fun. But in order for a peptide to be presented to a T cell, it has to be derived from a protein that is within the cell, that’s then processed, chopped up into small pieces. Then those small pieces get loaded onto the MHC molecules, and the peptide-MHC complexes get shuttled to the cell surface. There are complicated ways that happens for both class I and class II, but that’s the basic principle. So T cells can only see cancer-specific aberrant peptides if those peptides are actually bound to and presented by MHC molecules on the cell surface — or in humans, HLA molecules on the cell surface.
Abhi: And there’s also this heterogeneity in how the MHC is structured from ethnicity to ethnicity, and I think there’s even subdivisions within that.
Ben: Yeah, this is super important. The MHC, or HLA, loci in humans are the most polymorphic of all the germline-encoded regions, and there are huge differences in HLA allelic distributions by ethnicity. How many are there now, Alex — twenty thousand plus?
Alex: You know, I’ve been curating pan-species, so I’m up to like sixty thousand, but a lot of that is human. There’s a lot of human alleles — like maybe twenty thousand human alleles known. Every time you find a new remote mountain village, you’ll probably find a new HLA allele.
Ben: And this actually really matters a lot, because what the T cells are recognizing is not the peptide epitope by itself. It’s the pair of the peptide epitope and the MHC molecule that presents it. That’s one of the main reasons why the same mutations in different patients will be or won’t be presented — if those patients have different MHC allele haplotypes.
Abhi: So in an absolute ideal setting, you have this cancer cell that has some sort of MHC on the surface. It’s expressing the aberrant chopped-up peptides that it’s creating. A T cell wanders by. The T cell has gone through a self-selection process, so it’s looking for things that are foreign to the body. It looks into the MHC, sees something is off, and then it either instructs the cell to kill itself, or it just kills the cell.
Alex: That’s roughly correct. There are a few things I want to say about that. One is that the more familiar place where we see this act is in viral infection. Why do we have this entire papers-please surveillance state where T cells are going around getting samplings of what cells are making and then killing them if it’s wrong? It’s really, I think, evolutionarily geared toward viruses. A bunch of weird proteins show up in a cell, and then some T cells show up, and they kill it. So this is kind of the clearance phase of a virus. There’s only so far you get with just neutralizing antibodies. Beyond that, you need other mechanisms, and so some of those mechanisms are the second adaptive bit of the immune system that can sample the interior of a cell and then kill the offenders.
Abhi: Mm.
[00:12:57] The 100,000 MHCs per cell and the information-theoretic problem
Alex: So that should be at least a starting mental model for, “Oh, how do we even kill the cancer cells?” Well, you get them cleared the same way that virally infected cells get cleared. That’s one thing. Another thing is, numbers-wise, it’s useful to think about how many MHCs are in a typical cell. It varies a lot, but it’s around 100,000.
Abhi: Now that you’ve brought that up — I kind of assumed it was one, but it makes sense that it was many.
Alex: Well, it serves an information-theoretic purpose. There’s an infinite diversity of stuff going on inside a cell, and you want to get a really succinct kind of histogram of all that activity. If every single subsequence of every protein were on the surface of the cell, that’s just way too much — you can’t lock onto a signal. If it was just one single MHC, first of all, how would a T cell even find it? You’d have to scan the entire surface looking for the one MHC. So you need a bunch of them, so that when a T cell comes by there’s a high probability there’s an MHC for it to look at. And then you need restricted presentation — only very specific subsequences of each protein get onto those 100,000 molecules.
Because if every subsequence of a protein could get on there, even if it’s some actin peptide but a different one every time, you can’t lock onto, “Okay, wait, is this the weird one?” The information is just diffusing. So you need it to be really repetitive. Every protein is represented by a few exemplars. You achieve that through a few things. One is that MHCs are a kind of delightfully easy machine learning problem — which peptides they present is really easy to predict. The biology of it is very restrictive in sequence space. They have strong preferences, like position two needs to always be a valine, things like that. And then the antigen processing also kicks out a lot of things that could encounter MHC. So it’s this 100,000 molecules on the cell surface, of which maybe there are ten or twenty thousand distinct peptide sequences, and that’s tractable for T cells as a distributed machine learning algorithm to process. Does that make sense?
[00:15:28] Why a cancer can’t just drop its MHC
Abhi: It does make sense. An instinctive question I had when you first explained this to me is — if these 100,000 MHC molecules are how the T cells are looking into the interior of a particular tumor or healthy cell, why doesn’t the cancer cell just drop its MHC?
Alex: So they do sometimes. There are a variety of mechanisms that keep cells from doing this. For reasoning about that, you should think more about viruses, because viruses often try to hijack and subvert and repress MHC. But there’s a whole other compartment of the immune system called NK cells, natural killer cells, and their job is roughly to make sure that cells are presenting something on MHC. So it’s not a perfect pressure, because cancers do eventually find their way to dropping MHCs — at least certain cancers in some settings do. But there’s some evolutionary barrier there. If you just naively do it, there’s a whole other cell type that comes and kills you.
Abhi: Gotcha.
Alex: There’s also the fact that the MHC locus is pretty dense, and it has some things in there that are not related to presentation to T cells. It has other genes — RNA polymerase genes and things — that if you drop them, you also suffer a hit in fitness. So if you look at cancer sequencing data, especially if someone gets checkpoint blockade and you know that the T cells were doing something but then there’s an escape clone that is the recurrence, often there’s a loss of part of the MHC locus — maybe a particularly important allele that is presenting something the T cells really relied on to know that that cancer was a cancer. That kind of stuff — chopping off a bit of a chromosome to lose an allele or two — is easier than losing both copies of the full MHC locus.
Abhi: Mm-hmm.
Alex: It still does happen, but it seems to also open up other vulnerabilities. One of the craziest things — and I like this story, and don’t understand it — is that microsatellite-unstable colorectal cancer often ends up with some or complete MHC loss and then is really sensitive to checkpoint blockade, totally breaking the entire mechanistic diagram. There are some papers that came out saying, “Oh, gamma delta T cells,” which you normally don’t think about, pick up the slack and do it. But the details of that are not worked out. The thing you know is that just losing MHC does not make a cancer invulnerable. And even cancers that — there are a few transmissible cancers, and in their transmissible state they lose MHC, but through repression. There’s a nasty dog STD cancer, and when it’s in a new host, it has to start making MHC again, for some not-worked-out reason related to fitness. It’s just not trivial to completely get rid of it.
Abhi: Is it fair to say that if I was a cancer cell and I wanted to evade the immune system, the obvious thing I would do would be to continue presenting normal-looking peptides onto the MHC? Is it just that the act of becoming cancerous is so orthogonal to that, that that just doesn’t really happen?
[00:20:18] Immunopeptidomics: why predicting antigens isn’t knowing them
Ben: That’s actually one of the things that makes cancer cells harder to see than virally infected cells.
Abhi: Okay.
Ben: In virally infected cells, where they’re pumping out huge amounts of the viral proteins, there’s a much higher density of viral-peptide-specific MHCs on the cell surface. In cancer, the vast majority of the MHCs are presenting self-peptides. So the aberrant peptides are hidden amongst the sea of self-peptides, and it makes it more difficult. But the other thing — we know there’s selection that works along that axis, because you can follow tumor subclonality developing over time, and you can see certain subclones drop after immunotherapy or under selection pressure. And then the cancer that grows out — maybe there was a mutation in a passenger gene that it didn’t really need for fitness, but there was T cell selection pressure through that, and the cancer will just drop either expression or lose the genetic material, so it’s no longer vulnerable to those T cell populations.
Abhi: Interesting.
Ben: Which, incidentally, is one of the reasons that Alex and I think we really need to be doing immunopeptidomics validation of peptide presentation. Because if we’re giving vaccines to peptide antigens that are predicted but not actually presented on the tumor cell surface, we can be eliciting immune responses that are completely useless for controlling the cancer. They can be real and measurable by immunological assays but do nothing clinically.
Abhi: And just to define immunopeptidomics — it feels like there are two options on the table if you want to learn what peptides your cancer cell is actually expressing on the MHC. One, you can do whole-exome sequencing or whole-genome sequencing of the cancer, identify where the mutations are, and then slide a window over that region to create all the possible peptides. This is a partial answer to the question.
Alex: That’s candidate sourcing. There are other ways you could do the candidate sourcing. It doesn’t tell you what’s really there, though.
Abhi: Yeah, yeah.
Alex: I guess if your informatics got good enough, you could start squishing the — in my mind, the generation of the candidates is really different from validation, because the candidates are so low-value. The predictions are so bad. But if you got really good at the predictions, then maybe you could start to conflate those two more.
Abhi: From looking at the clinical trials that have been run with cancer vaccines, they do seem to be sourced directly from the genome, and they don’t actually tend to do this — to actually define the immunopeptidomics, you lift off the MHC, run it through mass spec, and identify all of the peptides that are on that region.
[00:21:31] Validating what’s on the tumor
Alex: Yeah. Okay, so let’s talk about how you’d validate, and then I’ll tell you why no one does it. If you wanted to know what’s actually on a tumor cell, you have a few options. You don’t have to do mass spec. Maybe I’ll start with the other alternative that’s even rarer, but a few companies have wandered into it. You could look at tumor-infiltrating lymphocytes. So you look at T cells in the tumor, you sequence their T cell receptors, and then you do some work to figure out what they recognize. You could also look at their expressional state when they’re in the tumor. There are certain signatures related to, “Hey, I found my target, I’m going to kill it now.”
And so you focus on those clones, look at their T cell receptors, and then take your whole candidate set — you do your typical thing, exome sequencing and RNA sequencing, run that through MuTect and Strelka, and then run the predicted mutational sequences through NetMHCpan. So you take all that stuff, which is a really long list, almost none of which is on the tumor, and then you do some work to figure out which of the T cells in the tumor look like they’re dividing or trying to kill stuff or just generally activated.
Do they recognize any of those targets? And if you find a T cell clone that was in the tumor that recognizes a predicted target, that’s really strong evidence that the tumor was probably making it. It could be that you screwed up the matching of the T cell receptor to the antigen — maybe it’s cross-reactive, needs a little bit more work to get confident — but that’s already way more validation than these just-predictive searches give you. That’s one way. A totally different way, if you have abundant frozen tumor tissue — and this is step one of why no one does it, because all of the logistical machinery of a hospital goes against it. They don’t want to have freezers full of stuff. They want room-temperature formalin-fixed, just a little bit of tissue. They don’t want frozen chunks of tumor in freezers from every single patient.
But if you have a lot of frozen tissue, you can hook into mass spectrometry to detect some of these peptides. There are different ways to do it with different sensitivities, but the basic idea is that you pull the MHCs off the cells, then you change the pH to make the peptides pop off, and then you put them through a column so that you get this gradient separation, and then you run it through — often something like an Orbitrap to do two stages. You figure out the mass-to-charge of the peptide, then you fragment it, then you look at the mass-to-charge of all the fragments, and you can informatically figure out which peptides were coming off the tumor.
So this is a lot of work, and it also requires a sample that’s pretty much never available. And if you do it the way I described, it also has a lot of sensitivity problems — you actually can’t see a lot of the peptides for a variety of technical reasons we can talk about if you want. So if you really want to know what was there, you have to do a whole extra thing. Naively, one of the problems is that mass spectrometry has severe biases in sequence space. So if you wanted to know where to point the acquisition of this instrument and what to expect from the peptides — not from this computational figuring-out from lots of spectra — you could synthesize some of the peptides and then try to figure it out. But that’s still not quite right; it gives you a little more information.
So there’s a trick. You make the peptides, but in a way that they’re heavy-isotope-labeled, and you drip them in with the peptides from the tumor, so you know exactly when to acquire. And then you also run them separately, so you know everything about, “Okay, this one doesn’t really fly so well, but at this moment, coming off the column, you’re going to see these peaks.” So you combine the information about the reference peptide with the fact that it’s running with your sample.
Abhi: Mm-hmm.
Alex: The reference peptides cost like a thousand dollars each to make and take like a month, right? So it’s this whole extra thing. Now you want to check a hundred candidate peptides — it’s a hundred thousand dollars, and it’s an extra month before you do your mass spec, which is a complicated, fancy assay, before you start making your vaccine itself, which takes another two months. So I wonder — is that enough to explain why no one does it?
Abhi: Okay, so maybe if you were trying to be maximally accurate, you would need to go through this whole isotope-labeling thing that’s super expensive and consumes a lot of time. Alternatively, it seems like what you get out of the mass spec is perhaps not all of the peptides, but at least the peptides you do get and you’re confident about. And maybe — is that enough?
Alex: Right. You can run in different modes, and there are ways that have more sensitivity. So you could theoretically try to do mass spec in a way that’s like, “We might only catch a quarter of them, but those are still actionable.”
Abhi: Yeah.
Alex: I think people did try that a bit. Michal Bassani-Sternberg, who’s a great mass spec immunopeptidomics researcher, published a lot of work on this, trying to read off mutational targets from melanoma samples. Because melanoma — there are big cutaneous chunks of tumor, you can get it fresh, you can re-biopsy if you need — so it’s a more accessible tumor type. She did a study in 2016 where she tried to find these neoantigen targets on twenty-five samples, and she found them on three of the patients. Coincidentally, what she did find a ton of is mine and Ben’s favorite targets, which are cancer-testis antigens. We like them because you can look in public data and know that they’re real. So she did that in 2016. I don’t know if she’s done a repeat of it. I think with a little bit of update on the methods and newer instruments — because there are now fancier, more sensitive instruments — you might get that up to finding a peptide in like ten of the patients. But it’s still limiting. Without doing the really annoying targeted version where you can quantify and get down to really low abundance, you end up sometimes just having nothing.
Abhi: And in that case, you might as well just opt for the whole-exome sequencing approach.
Alex: Well, you still do the whole-exome sequencing.
Abhi: I mean as the confirmatory.
Alex: Yeah. You have to do that in order to even know what you’re looking for.
Abhi: That’s fair.
Alex: You can never get away from it. If you want mutational targets, you have to do some kind of sequencing. It doesn’t have to be whole-exome sequencing — if you did just single-cell long-read RNA sequencing, we have a method that might work for that. But you still need something that’s nucleic acid from the tumor generating a candidate list for you.
[00:28:57] How many targets you actually need, and the HLA-loss problem
Ben: Yeah. One of my favorite papers in this space, Ehx et al. from Claude Perreault’s lab in leukemia from 2021 — they went very deep and included using a genomic reference against expressional antigens that aren’t mutations, but are part of the dark proteome. Their findings were dominated by tumor-associated antigens. It was a really, really well-done work. And maybe the answer to your question of “is it enough” is conditioned on how many you need to make a good vaccine. You need at least one that’s actually presented by the tumor cells, right? But you probably need more than one, because of the potential of HLA loss. So if there are two copies of the HLA locus on chromosome 6, and it’s relatively easy for the tumor to genetically lose one, then ideally you would make a vaccine that has epitopes that span HLAs on both chromosomes. Then you’re creating selection pressure that would cause the tumor to have to drop both and incur the fitness cost therein, rather than just dropping one of them.
Alex: And ideally spread across different alleles, because then when the loss happens, there’ll probably just be a piece of this one and a piece of that one.
[00:30:22] Can a tumor mutate its own MHC?
Abhi: Maybe this is incorrect, but it feels like a lot of cancer vaccine papers kind of implicitly assume, “This patient has this particular MHC allele,” and they just pick the neoantigens using that as a reference point. But are the tumors able to modify their own MHC?
Alex: Like mutations in the MHC coding regions?
Abhi: Yes.
Alex: I think people have found that, but it’s really uncommon. When you think about tumor evolution, you’ve got to think about the diversifying mechanism — by what means can tumors have a big population that then evolution acts on? There are some that have infinite point mutations, but more commonly they can diversify by just losing chunks of chromosomes.
Abhi: Yeah, makes sense.
Alex: So in a cancer that’s driven by some mutational process to introduce tons and tons of point mutations, that one will stumble on, “Oh yeah, you should disrupt residue 72 of the binding groove.” But that’s a hard trick to pull off for a cancer that’s not point-modifying everything in the genome.
[00:31:34] What are tumor-associated antigens and cancer-testis antigens?
Abhi: We’ve mentioned the word TAA. We’ve also mentioned the word CTA — cancer-associated antigen. I would love to get your explanation of that, because I think it’s a very fun and clever way to approach cancer vaccines.
Alex: Okay. TAA, I think, is older terminology. This is people looking at cancers like 30, 40 years ago and finding that there are some proteins that just come up a lot. It doesn’t actually say anything about the specificity of it. Often they’re an overexpressed protein — “Oh yeah, cancers will just make a lot of this.” And there are different therapeutic approaches toward the proteins that come up a lot in cancer samples, some of which were vaccines, and there’s a lot of other therapeutics that people pursued also. If you wanted a subset of TAAs that had really nice tumor specificity — so you feel no worry in control-F deleting every cell that makes that protein — then you get to something that’s roughly cancer-testis antigens, plus a few other things. Cancer-testis antigens are lineage-restricted to some stage in reproduction, and most of them are related to spermatogenesis. They’re like a strange moment in spermatogonial stem cell development where they’re not yet pumping out tons of protosperm, but they’re about to, and they need this one protein. There are also placental proteins — placenta is super weird, it does all kinds of stuff that no other tissue in our body does. There are things like bits of the motor protein of a sperm that other cells in our body don’t need.
So it’s not strictly accurate to say these are not in healthy tissue, because if you’re a guy and it’s the motor protein of a sperm, a really effective cancer therapy would also probably make you infertile. So in some ways they’re disposable tissues, or tissues you can warn against. You don’t want to give a cytotoxic therapy against placental proteins to someone who’s pregnant. But you can navigate around the reactivity to make it not toxic.
Abhi: Importantly, the fun part about CTAs is that there’s no T cell repertoire — T cells have not been selected against them.
[00:34:09] The thymus, AIRE, and why T cells were never trained against CTAs
Alex: Well, some of them have. This is an open question for us, and we went looking. Most of them are not expressed in the thymus, which is to my surprise, because they’re in the genome. We don’t talk about the thymus, but it’s an education — it’s a finishing school for T cells, and they go there to figure out what not to kill. Most cancer-testis antigens don’t show up in the thymus, so T cells are not educated to not kill them. There are a few that do, and some of them are really high-value. There’s this one I would love to make a therapy against, FATE1, because every Ewing sarcoma makes FATE1 — it’s required by its oncogenic fusion. But that is the one we saw the highest thymic expression for. So the therapy as a vaccine might work, but it’s less likely than some other ones where the thymus is just really silent.
Ben: This is a really important point, though, because the T cells are negatively selected against selfness not at the level of the complete genome, but at the level of which proteins are expressed in medullary thymic epithelial cells, under control of this really cool protein called AIRE that causes the mTECs — medullary thymic epithelial cells — to be able to broadly express genes from across the genome. But it’s not completely uniform.
Alex: Without dying, which is weird.
Ben: Without dying.
Alex: Those proteins, when you put them all together in one cell, should just — it’s not any one cell type anymore, it’s just a jumble. You’re a little bit liver and a little bit brain and a little bit sperm. And then somehow they don’t die. And thymus cancers are very rare.
Ben: I mean, they exist, but — this is crazy biology that I think has not yet been unlocked.
Abhi: I never naively thought about how the selection training happens, but it makes natural sense that there’s a cell in charge of it. It is strange, though, that they’re able to express every single protein.
Alex: Just super surprising.
Ben: And so a large number of the cancer-testis antigens are not expressed in the mTEC. By the way, the mTEC carefully curated RNA-sequencing dataset we use also came from the Perreault lab — super thankful for that. The famous cancer-testis antigens, of which T cells have been discovered by multiple groups over many years, are all nearly negligibly expressed in the medullary thymic epithelial cells.
Abhi: Has that stuff been mapped out? Do you know every possible CTA that exists?
Ben: No.
Alex: No. I spent like a month this year doing a new curation of CTAs against different data sources, and I think the criteria have just changed over time. The available technology changed, so there are some studies that’ll find like two thousand of them. But if you really do care about — “I just don’t want any brain expression of this, I don’t want to make a T cell that goes into the brain and finds its target there” — then you lose a lot of that two thousand. The brain makes a lot of weird proteins. If you’re really cautious about cardiac expression, because there have been CTA-targeted T cell therapies that cross-react with the heart and they kill people — then you lose some of that candidate set. So maybe it’s not right to say you could just go and find a bunch of new ones. I think the work of finding really actionable ones is still a little bit open.
[00:37:35] Why “normal” proteins is still an unfinished reference problem
Abhi: Why is it not as simple as: this particular cell in the thymus contains the full universe of possible CTAs, and now I’ll just cross-reference that with the transcript of every other cell type in the body?
Alex: Well, there are good data sources now. The Human Protein Atlas lets you get many different tissue types. One problem is that “normal” encompasses a lot of different cell behavior. So it’s not enough to just — it’s like you got a brain sample, great, well, you need to treat the cortex differently from the cerebellum.
Alex: And then you start these subdivisions, and someone else goes and does the single-cell sequencing and finds that actually there are five different subtypes in this one area, and then you need to do the normalized TPMs within each subtype, because you do get rare cell types that, if you kill them, you give someone diabetes. If you had some broad expression profiling of the pancreas, you would miss the very small number of cells that are pretty vulnerable to getting killed. So the work of defining “normal” is not exactly finished, even just at the level of how much of each gene is made. And then there’s a whole other repeat of that entire unfinished project of doing long-read sequencing, because you don’t know all the isoforms they make.
Abhi: Oh, yeah.
Alex: So if you wanted to capture the real isoform diversity, you’d need to go back and redo GTEx and the Human Protein Atlas and all that work with long-read sequencing, which I think some people are starting to do, but it’s just an unfinished spot in our reference databases.
Ben: No — just another level beyond that, the subcellular location of proteins matters to how efficiently they’re processed and their peptides presented. Take, for example, a leukemia antigen that’s a self-antigen called cathepsin G. Cathepsin G is essentially in granules that are not normally exposed to the cytosol, so it’s not normally presented on HLAs from normal healthy bone marrow hematopoietic stem cells at a very high level anyway. But some leukemias actually express this and get it into the cytosol, apparently, and then it’s expressed at ridiculously high levels, and it’s a target for therapeutics. So it’s in the genome, it’s in the self-transcriptome, it’s expressed in medullary thymic epithelial cells. By all rights it should be a dangerous self-target. In fact, it’s a very fruitful target in at least myeloid leukemias because of this subcellular distribution of the protein and relative likelihood it’ll get processed and presented.
Abhi: Maybe a naïve question — shouldn’t you expect the adaptive immune system to clean it up, then?
Alex: Well — the adaptive immune system flops all the time. Yes, it kills a lot of cancers; probably in my lifetime I’ve had some infinite number of cancers that would-be deviated cells that got cleaned up. But there are limits to how much it can figure out.
Ben: That’s actually a really good question, though. Because it’s expressed in the medullary thymic epithelial cells, the T cell precursor frequency is very low, so you want to target it with something like a TCR-T — a T cell that you exogenously engineer to have a T cell receptor that will bind it, that you’ve discovered some other way — or an antibody, or something like that.
[00:41:05] The short list of T-cell killing success stories
Alex: But also, if you look at — we should talk about what the success stories are for T cell-mediated killing, and it’s a pretty short list. If you just look at when we’ve gotten T cells to really eradicate established cancers, a lot of them are engineered cell therapies, where you skip the vaccine part and then you stick a receptor that you know recognizes your target directly into T cells. That’s mostly for cancer-testis antigens — PRAME, NY-ESO-1, MAGE-A4, these reproductive proteins — also for viral proteins, so HPV. A few mutational targets, but not a lot, and usually the ones where you had good mass spec evidence that they’re presented. So if you look at those success stories — if someone has a PRAME-positive melanoma, which a lot of them are PRAME-positive, there’s this great TCR-T from Immatics, an engineered T cell receptor therapy that seems to work pretty well.
Abhi: It’s not yet approved?
Alex: — I think they’re having a Phase 3 trial right now. But I’ve heard from doctors who’ve used it on patients that you get impressive responses. Not always durable ones, because it’s one target, but impressive regressions of tumors.
Abhi: Mm-hmm.
Alex: Those tumors had PRAME all over their surface to begin with. And PRAME — when we look at the mTEC RNA-seq, the expression is like 0.05 TPM, right? So there’s not a lot. Maybe somehow there’s a lot of tolerance being induced there, but probably there’s not. So the capacity to kill the cancer mediated by this one PRAME-derived peptide was in the patient’s immune system the whole time. But immune systems just try to do a lot. Just because there’s one good target in a cancer does not mean that when the cancer cells are imploding and draining into some lymph node, the immune system will figure out, “Aha, this is the good target, we’re going to go after it.”
Abhi: I also imagine — though you vaguely mentioned this — the adaptive immune system is also dealing with the fact that a cancer is actively suppressing it.
Alex: Yes.
[00:43:20] Did the old vaccine literature just need Keytruda?
Abhi: I do wonder how much of the pre-Keytruda cancer vaccine literature was actually hampered by Keytruda not existing — and if we went back to it, there would actually be a lot of really rich stuff there if you just combined it with Keytruda.
Ben: Hopefully that’s true, but the PD-L1/PD-1 axis is not the only axis of immunosuppression. One of the things I think about a lot — we’re talking now about how to personalize vaccines or maybe TCR-T therapies. I think the reversal of immunosuppression and immunoevasion mechanisms activated by the cancer will also have to be equally personalized for every patient. For some people, Keytruda or similar is going to be the right thing to give with a vaccine. For others, it might be a thing that eliminates regulatory T cells, or myeloid-derived suppressor cells, or tumor-associated macrophages, or cancer-associated fibroblasts, or the cancer’s ability to upregulate molecules that kill the T cells that come to kill the cancer. There are so many different ways the cancer can fight back against immune recognition and killing that I think we’re fooling ourselves if we think there’s a one-size-fits-all solution.
Alex: That makes sense. But my hesitation around, “Did all the old cancer vaccine literature not work just because we didn’t have Keytruda” — there are a few things there. One is that I have a personal annoyance at the degree to which cancer vaccines hide behind checkpoint blockade. If I just add it to every single trial and then get essentially the expected historical response rate to checkpoint blockade, then everyone looks at it and says, “Aha, the cancer vaccine worked.” So yeah, maybe if you had Keytruda it would work, but —
Ben: They measure vaccine-induced immune responses.
Alex: They do in the trials, yeah, we can talk about that. But if that were the issue, it wouldn’t be uniformly the issue. In those trials you would see a few exemplars where the cancer didn’t figure out, “Oh, you should make some PD-L1,” and then those people would have had strong monotherapy responses to the vaccine. When you look for monotherapy responses to vaccine, even in a small fraction of a trial, it is such a short list. It’s like HPV vaccines in really early-stage cancers — there you can really point to, “Oh yeah, there was a really early-stage or precancerous condition that’s HPV-positive, the vaccine could clear it sometimes.” There’s a BioNTech melanoma trial in which they had one complete response. To me, as someone who works on cancer vaccines, I think BioNTech did not see it as a success. It’s a low response rate, they had a checkpoint blockade, it only boosted it a little bit.
So I think they kind of shelved that fixed-antigen vaccine. But I saw their Phase 1 trial and thought, “Holy shit, they did it.” It’s only one patient, but no one else has done that. They didn’t need some combination that obfuscates which agent’s doing the work. They gave the patient a vaccine, they got like fifteen percent of their circulating CD8 T cells recognizing antigens from the vaccine, and then their melanoma disappeared. That is a setting in which you can go look at the other patients and see, “Okay, who else got like five-plus percent circulating T cells responding to the vaccine but didn’t get a complete response? Let’s look at their cancer and see whether checkpoint blockade was the missing part.” I think that’s why in the subsequent trial they added a checkpoint blockade agent, and it didn’t work as well as they hoped, and then they threw a small fit over it, I think. But maybe the answer’s not easy. That’s where you need to go look for what is the extra thing that’s missing. You need to have a strong success signal even in just one patient to then figure out, “Okay, the potential’s here, so why didn’t it work for the other people?”
[00:47:31] Why not cut out the middleman and engineer the T cells?
Abhi: We can come back to the actual neoantigen or CTA part in a bit, but one thing you mentioned — maybe we just cut out the middleman here and engineer the T cells against the particular antigen we want to mount an attack against. Why doesn’t everyone — is it just expensive, is it logistically complicated, is it —
Alex: It’s so expensive and so complicated. We had that thought, so we’ve gotten into doing TCR-T engineering preclinically in the lab. We’ve been working with a sort of rebooted personalized TCR-T company. We’ve tried to make broader contacts in that space. There are a lot of reasons why that’s not the current go-to solution, but the simplest one is just how logistically complicated and expensive cell therapy manufacturing is. It is really quite a slog. You can see it with CAR T, because CAR T felt like a miracle, and then the deployment of CAR T is really muted. Not that many people get access to it, even though it’s a single construct — you’re just trying to get this one construct into people’s T cells and give it back to them, and it’s still almost a civilizational challenge to scale it up. So imagine you’re trying to do that with a set of different receptors per person, and you have to find them first. It’s really hard.
Ben: Yeah. It’s going to happen, though. It’s part of what we need to cure otherwise incurable cancers, I think.
Alex: I mean, Ben and I are working on that — we are actively working on faster ways to do in vivo TCR therapies. But there are technical challenges there, and in the meantime there’s a middle period where we have hints of success from the vaccines, and they’re logistically much easier. So probably it’s reasonable to allocate like eighty percent of the effort there while you try to bring up the technology for the higher-impact, more effective therapies.
Abhi: I imagine the concern — and maybe this isn’t a consistent concern with both CAR T and especially TCR — is that sometimes you get these weird cross-reactivity things you couldn’t have predicted in advance. One of the things I stumbled across was the MAGE-A3 cross-reactivity... In some sense, it feels like if you cut out the middleman, you also cut out some safety features, in that you might design something that is just very toxic to the patient. Is that just something you have to live with?
[00:50:09] Autologous vs affinity-enhanced TCRs, and a database-matched middle path
Ben: Yes. Or — well, one nuance there is where do you source the T cell receptors that you’re going to deploy as TCR-Ts? Because if you source them from the individual patient that you’re then going to dose, then all of the TCRs you discover that are antigen-specific have been through thymic selection in that individual. So in that individual, they’ve been at least safe enough not to cause fulminant autoimmune disease before you gave them back to the person.
Abhi: But I thought you were engineering those.
Ben: Well, that’s — your other option is to pick them allogeneically, and then you could ex vivo engineer them for better affinity. And it’s the affinity engineering that I think introduces more potential cross-reactivity danger, without an effective screen, than sourcing them autologously.
Alex: And there’s a preprint that just came out that’s great. They’re looking at that A3A TCR that I think killed two patients — I think they dosed them on the same first day of the trial, which is crazy. They looked at the affinity-enhancing changes that were made to the TCR and then rolled them back one at a time to see, can you get rid of the titin cross-reactivity, so it’d be cancer-specific but not heart-specific? And they can. They got rid of two mutations — it was surprisingly in not the loop you’d expect, but the second loop, in CDR2. And they were able to make a TCR that was higher affinity than the parental TCR that had been taken from a donor, but was not reactive with the heart. So that work of affinity enhancement without introducing cross-reactivities is really, really delicate, and Ben and I don’t want to engage in it at all, because it seems very stressful. But there’s an opposite extreme that Ben was talking about. You can source the TCRs from the patient. There are some challenges there — you’re adding a bunch of extra time, the needle-to-needle turnaround time here is getting kind of long, because you have to do TCR discovery in the patient before you put the T cell receptors back into their T cells. But that has a much nicer safety profile; the safety logic is better. And then Ben and I have also talked about a kind of intermediate possibility — if you could do really large-scale T cell receptor discovery on shared targets, like cancer-testis antigens, you could have a big database of T cell receptors that have specificity to a CTA target, and you know the HLA type of the person they came from. And you can match people the way you do as tissue donors — “Oh, you have a four-out-of-six locus match, and this TCR will hopefully be safe for you.” So that would be an intermediate amount of risk. You could still potentially personalize, but because it’s not heavily affinity-enhanced, you’re less likely to introduce a crazy cross-reactivity.
Abhi: Has any of this reached — I know there are some CAR-Ts that are approved. Has any of the TCR stuff been approved, like personalized TCRs?
Alex: So Adaptimmune — they got one approval and one... did they — they got a breakthrough designation after their approval. But this comes back to how hard these things are to manufacture, and how expensive. Adaptimmune made two TCR-Ts that work really well in rare sarcomas for which you really need therapies, and they hit CTA targets and get a high response rate — they’re great. They then went totally bankrupt and had to sell all their assets to some other company, because the manufacturing build-out just ate all of their money, and their expected reimbursement for these rare sarcomas was kind of low, so they just ran out of cash. But they got it there. There’s an approved therapy, and probably Immatics will get theirs approved. The ones going after that category of targets look, as a class, reasonable. Probably there will be more approvals than that as time goes on, but it’s really slow to build them out.
[00:54:23] Can we TCR-T and CAR-T in-vivo?
Abhi: I imagine the obvious ideal-world solution to this is just to manufacture them inside the person. How good is in vivo TCR-T today?
Alex: I mean, it doesn’t exist.
Abhi: Okay.
Alex: I think we’re — we’re doing it in our lab to some degree.
Abhi: In vivo CAR-T?
Alex: In vivo CAR-T exists. So Capstan has an in vivo CAR idea. Okay, so we should talk about what in vivo CAR T is. In vivo CAR T is some version of, we’re going to skip cell and gene therapy manufacturing, and instead inject the patient with something that will, in some way, get the receptor into their T cells, either permanently or temporarily, but the CAR is going to be in patient T cells.
Abhi: Mm-hmm.
Alex: And so the manufacturing profile for that is comparatively simple. At the worst it’s like a viral vector; the easiest versions are an RNA. That’s a bubble right now — there are a lot of players that have entered that space, and more rushing in, and some of those have gotten to the clinical stage. There are big acquisitions happening, and they seem to work. They have maybe different properties depending on how you do it. So Capstan is doing an RNA-based one, and we’re kind of inspired by them to do RNA-based TCR-Ts. And then there’s a Chinese company, MagicRNA, that has something loosely similar to Capstan, but they’ve got clinical data way faster. They use theirs actually for autoimmunity — they nuked B cells temporarily to get rid of lupus, and it seems to work.
Abhi: Interesting.
Alex: So there is clinical proof there. And then there are a bunch of riskier but more effective-for-cancer versions of this that are actually integrating in some way and creating stable CARs. The RNA’s got a burst of expression and then it goes away, whereas if you have some sort of viral vector you can make actual CAR Ts, and then they stick around.
Ben: A few important considerations there. One of the limitations of most of the in vivo methods is you can’t use lymphodepletion first, because if you deplete all the T cells you could then be in vivo engineering, it would be counterproductive. But the dogma goes back to the early days of cellular therapy, that one must have lymphodepletion for cellular therapy to work. I don’t necessarily believe that’s true. I think one must have means of the cells in the cell therapy expanding and persisting and doing their job. Lymphodepletion is one way to effect that, that we knew thirty years ago. There are other ways we can achieve that same purpose now, but that’s a hypothesis. And there are some companies working on ex vivo scaffold technology, where you have a scaffold impregnated with T cells, and then you give lymphodepletion and implant the scaffold back in, so the T cells release. So you can get the best of both worlds — the manufacturing advantages of an in vivo solution, but still be able to give lymphodepletion. The degree to which we need lymphodepletion is really important to which of those strategies works.
Alex: But if it turns out you can do it without lymphodepletion, then that is the ultimate patient win. Because lymphodepletion is hella toxic, very dangerous.
Abhi: And that’s literally sifting the T cells out of you?
Ben: No, no. Lymphodepletion’s chemotherapy.
Abhi: Oh, okay.
Ben: So, preconditioning chemotherapy. A person has cells manufactured. Before those cells are given to that person, they get a highly lymphodepleting chemotherapy regimen, so that the majority of their lymphocytes are gone. And then the T cells are infused so that there’ll be — what’s commonly taught is — immunological space. And what immunological space is, is some combination of physical space in the secondary lymphoid tissues, availability of antigen-presenting cells, and availability of cytokines and chemokines that the endothelium may make because there aren’t a lot of T cells around.
Alex: Your body has a set point for how many T cells it wants. So if you get rid of all of them, it starts cranking out all these signals for T cells to grow, and that’s when you infuse CAR T cells.
Ben: And so the infusion expands more in vivo because there’s this physiological space for it to move into. But the lymphodepleting chemotherapy is actually a fairly intensive chemotherapy regimen. It means you often can’t treat people who are older with serious comorbidities, or even when you do, there are risks of infection and bleeding, other complications, organ dysfunction —
Alex: People just die of it. Look at our TCR-T trials — it’s just toxic. You look at the mortalities, and it’ll be, “So-and-so, the CAR T didn’t work, died of the cancer,” and then there’s the number that died of their lymphodepletion regimen.
Abhi: So intuitively I’ve always thought of CAR T, in vivo or otherwise, as a very toxic therapy, but it sounds like the true toxicity is actually the lymphodepletion stuff. The actual CAR is —
Ben: No, there’s both.
Abhi: Okay.
Ben: They’re both dangerous. The CAR T toxicities are inflammatory toxicities, like —
Abhi: Like creating cross-reactivity.
Ben: Well — that’s actually a good point. The cells can have on-target but off-tumor toxicity, if you have an antibody that can cross-react. The common toxicities are things like cytokine release syndrome, which is an inflammatory injury that in its most severe form can cause vasodistributive shock. Or something called ICANS — immune cell-associated neurotoxicity syndrome — which we don’t really understand the pathophysiology of.
Alex: Your immune system freaks out. That’s it.
Abhi: They’re all sort of shorthand for your whole immune system going, “Ahh.”
Alex: Yeah.
Abhi: In vivo CAR T clearly is doing something. And is it as cut-and-dried as: because in vivo CAR T works, in vivo TCR T should also work — the tech tree is already completed, you just repeat the exact same process?
Ben: Part of the tech tree.
Abhi: Okay, what’s missing?
Alex: Well, first of all, we don’t yet know if in vivo CAR T really works. The autoimmune setting is much easier than cancer. If you want to just clear your B cells for a week, that’s different than wanting to actually eradicate every single cell. So that’s one limitation. Then the receptor is really different. A CAR has its own signaling built in, so how much of the T cell receptor do you need? Do you need to get rid of the other T cell receptor? There’s a lot of technical nuance to, “Okay, I’m adding a second T cell receptor to all my T cells — are some of the T cell receptors they already have kind of dangerous for me to be expanding? Are they going to compete for CD3, so the one I’m adding is not going to work as well?” So there’s a bit more to figure out. The tech tree’s not as built out as you’d like, but the fact that in vivo CAR T is having a bonanza and everyone’s going to enter trials this year — that’s useful. You learn a lot about delivery platforms and how to get the constructs into cells.
[01:02:08] Why whole-tumor and lysate vaccines never worked
Abhi: On the subject of things that are not necessarily cancer vaccines but are strange — I stumbled across this line of research called GVAX, where you harvest tumor cells from a patient, genetically engineer them so that they pump out a particular protein that dendritic cells really perk up at, and then infuse them back into the patient. It doesn’t seem to work.
Alex: Yeah.
Abhi: But to me, that feels like you’re throwing the full book at the immune system, saying, “This is the tumor, pay attention to it.” How come that doesn’t seem to do anything?
Alex: Well, whole-tumor stuff in general has never really worked. It was a nice idea, and there were tumor lysate vaccines of every imaginable flavor. I think this gets into the issue of the information problem you’re asking the immune system to solve. It was not able to lock onto the signal when the tumor was alive in the patient’s body. And probably that has to do with how hard it is to sift out what’s distinguishing. You wanted to learn a very small number of distinctive peptide-MHC complexes that, upon being presented by a cell, mean you should kill that cell. There aren’t that many of them that are distinctive, and ninety-nine point nine percent of the ones in each tumor cell are actually self-antigen that you shouldn’t be killing.
And you need the immune system to really confidently figure it out and then pump out a huge effector population to do the killing, because they have to go all over the body and extravasate into every little niche and look for metastases. So it might immunologically work a tiny bit — you might get a little bit of responsiveness to some tumor-specific thing. That’s sort of the general story with tumor lysate: you get a little whisper of, “Oh, this looks interesting, maybe we should be a little more cautious about this peptide-MHC.” What you don’t get is the really confident thing you get with a viral infection, where these nine amino acids from the spike protein — if any cell’s making them, kill them, and raise a giant army and make sure we’re all going out doing this job.
Ben: There’s just a ceiling to non-antigen-specific interventions, for that reason — for a basic information-processing reason. And to overcome that, we’re going to have to get tightly focused on the very best specific tumor targets.
[01:04:46] Immunodominance, or why a great vaccine can still do nothing
Abhi: One of the things I found very interesting about deciding what neoantigen you want to put into the vaccine is the concept of immunodominance. I would love for you guys to talk about that.
Alex: It’s a big extra challenge that I wish we did not have. You could have a vaccine where, let’s say, you’re doing a pretty good job with your selection and half the stuff you put in the vaccine is actually on the tumor — that would be much better than people currently do. So you encode them all as mRNA, and you have some mRNA construct with twenty targets, of which ten are real and could be the determinants of tumor killing. And then you give the patient the vaccine, and they get a sky-high immune response — you use a really good spleen-targeting IV platform, twenty percent of T cells recognize a target — it’s awesome, but it’s just one target, and it wasn’t one of the ones that are actually on the tumor, so the vaccine does nothing.
Abhi: Yeah.
Alex: That is sort of what happens right now, especially with the mRNA platforms. But in general this happens with all vaccines. The immune system has a kind of multi-tiered competitive dynamic that tries to concentrate the immune response on just a few things, which I think also has some sort of evolved information-processing role. The immune system will not just respond to everything you throw at it, because it has a budget. There’s a certain number of T cells, and it’s trying to get a concentrated response. There are a few places where this happens. One place is that the antigen-presenting cells take up your vaccine but may not make every single thing in it. There’s also a kind of local neighborhood in a lymph node where there’s competition for, “Oh, this T cell clone’s winning, and it has a certain target, and it’s suppressing the other ones.” And there are other competitive dynamics just in circulation, and probably even more I’m not thinking of right now — Ben’s probably got a bunch more — but immune cells compete.
Abhi: Yeah.
Alex: And if you —
Ben: From the perspective of the individual T cell bearing a single T cell receptor that’s probably reactive to at most one of the antigens in the vaccine, it will compete with all of its brothers and sisters around bearing different TCRs in the setting of general stimulation. The neoantigen vaccines were originally designed with the sense of, “Let’s give as many epitopes as we can in the vaccine, because we know our predictions are imperfect, so if we just get one or two that are good, that’s a win.” I think we know now that that might not be a win, because the wins you get may prevent the elicitation of better ones.
Abhi: This is maybe a — there’s no known answer to this — but what’s the evolutionary point of this? It seems to me the immune system is almost playing a lottery game, like, “I’m not going to uniformly respond to this threat, I’m going to pick one of the potential responses and just move forward with that.” But that feels obviously worse than starting multifaceted and then slowly affinity-maturing your way to the real threat.
Ben: It’s a historical co-evolution, I think, between, say, pathogens — because for most of human history we haven’t generally been old enough to have to worry about cancer.
Abhi: Sure.
Ben: So maybe specifically viral pathogens, the MHC locus, and then stochastic T cell receptor generation in T cells.
Alex: I think there’s also — first of all, antibodies do a better job at breadth than T cells. If you look at antibody responses, you get dominant clones, but you also get a lot of — they get the second chance at diversification. Somatic hypermutation can take them on more complicated paths. Your T cell compartment’s floating around with a fixed diversity. Every single T cell is essentially a random classifier, and you’re just re-weighting those classifiers. So I think if you tried to write an algorithm like that, you’d come up with something that’s trying to concentrate on a smaller number. It’s like it’s L1-norm penalized, not L2-norm penalized. To get that kind of learning algorithm to work, you are going to get a little bit of spikiness in which clones get to win. I also think there are diverse T cell responses to viral infection. Whenever you want to understand the evolutionary rationale, looking at viruses is more informative. You do get T cell responses to — someone gets a COVID infection, and usually there are T cell responses to like two peptides from the nucleocapsid and one from the spike and one from the E protein. So it’s not all one peptide-MHC, but there’s some effort to keep it from being every possible peptide-MHC.
Ben: You can think about immunodominance at the population level or at the individual level. At the individual level, it’s the mix of the pathogen’s genome with the individual’s TCRs and MHC allele haplotype. And the set of possible pathogen-specific responses is much larger than the actual set that ever obtains in any individual. And then there’s the population layer Alex was alluding to — which proteins in the pathogen, or protein subunits, are more likely across the population to elicit responses.
[01:10:38] Can you predict immunogenicity with ML?
Abhi: For any arbitrary neoantigen-MHC pair, is it at all possible for me to design an ML model to predict how immunogenic and immunodominant it’s going to be?
Alex: That’s possible. We have grad students working on it. I think we’re not the only ones — other people, MS companies, sort of claim to do this.
Abhi: I guess one argument for why it’s impossible is that everyone’s T cell repertoire is so different that the competitive dynamics are going to be different from person to person.
Alex: I think there are probably at least two layers here. One is, could you model some intrinsic properties of T cell targets that are more likely to make them the winners, versus what is the intrinsic stochasticity of that competition? I think you probably can attack the first layer. Really dumb things we know about immunogenicity: when a peptide is in the MHC, its ends are usually buried and the middle’s sticking out, and if you have bigger residues in the middle, T cells can see that more easily. It’s a really naive ‘80s-style cheminformatics-type feature, but it does roughly hold. So if you had two different peptide-MHCs, one of which was all glycines in the middle, one of which has really big side chains, then the one with the bigger peptide is more likely to win the competition. It’s not guaranteed to, because then there’s a whole extra layer of just stochastic competition.
[01:12:24] What you tell the FDA for hyperpersonalized drugs
Abhi: There have been a few neoantigen cancer vaccines that have gone through Phase 1 and one or two Phase 2 trials. Are they required to share how they’ve selected these neoantigens from patient to patient, or is it proprietary?
Ben: They are required to register with the FDA how it’s done. They’re not required to publish or share it.
Alex: And the FDA had a funny moment. When we were starting trials at Mount Sinai, they accepted very high-level descriptions of the pipelines. I just wrote up some text that was like, “We predict mutations, and then we...” — I just wrote up a blog-post-level summary of everything without too much technical detail, and that went into all the FDA documents. They were fine with that. And then there was a moment where they’d accepted enough of these and hired someone who really understood the informatics, and they went back and asked every single person running any of these trials to fill out these very detailed workbooks about what you’re doing at every single step — which versions, how are you running it. So they realized there are technical details they should care about. I thought it was interesting, because it seemed like at first there wasn’t a full appreciation of how much of a computational drug this is. They really cared so much about peptide stability — “We’re going to make these peptides, how are you storing them, how are you validating they’re stable at six months for any redosing?” They cared about the chemical composition of the vaccine, and it seemed like they just didn’t know how to think about the fact that they’re different every time. And then at some moment they were like, “Wait, this is just a Python program that picks these drugs for every patient.” And then they went back and really wanted to know what’s going into your pipeline.
[01:14:14] How secret is the computational pipeline?
Abhi: How secret is this pipeline, in the sense of — if some representative from Moderna and BioNTech was here to share with you exactly how they’re creating this particular cancer vaccine, how they’re selecting the neoantigens, how useful would that be to you, versus you already suspect, “This is what they’re doing”?
Ben: We’ve asked them in various ways, and they don’t want to — I mean, broad strokes, yep, it’s very similar. And the software that comes out of our lab is all open source.
Abhi: Sure.
Ben: So others can look at it and use it if they want.
Alex: And I know that some of the stuff I wrote at Mount Sinai that we’ve kept developing at Pearl, openVax — that’s pretty open. And I know it’s been used beyond just Mount Sinai, or at least components of it have been. And then I think there’s a desire to say there’s some proprietary edge — “Our algorithm’s super good.” I usually view that quite skeptically, unless they really have a good data source that would explain why their algorithm is better than usual. Almost all of them are something equivalent to a variant caller like MuTect and Strelka, with a thing sort of like NetMHCpan or MHCflurry predicting MHC features. The times where something surprising comes out is because they’re investing heavily in data generation and then modeling some biology — the edge is in understanding the biology better, not in composing existing tools or making machine learning models of existing data better. So you really need to have an insight on what we’re missing about the biology and then model that.
Abhi: I was expecting you to say, “Oh, they probably have a better peptide-MHC binding model.” What do you mean by there’s some aspect of biology?
Alex: Okay, so I’m working on MHCflurry 3 right now. This is off public data, but I’ve been combining it in non-obvious ways, really thinking about how antigen processing works in different types of tumor cells and how it works in the antigen-presenting cells. There’s a little bit of a mismatch — your vaccine goes to antigen-presenting cells, and then they try to recruit T cells that are going to go kill the tumor, and the way the vaccine’s chopped up might be different than the way the source protein’s chopped up in the tumor. So trying to account for those differences — thinking through the mechanistic chain of how I get a killer-nanobot army to really go after the cancer, at a detailed level. If you then realize you have a gap — “Oh, we should do more mass spec of this sort, or we need to mass spec the tumor versus the APCs from the same people” — so generating data around wherever you perceive gaps in knowledge potentially adds quite a bit. Whereas there’s a flood of these MHC-peptide affinity models — like ten of them come out every day — and it’s very easy, it’s the most predictable part of biology, and none of them really have any advantage over any others. It doesn’t matter which one you use, because they’re not addressing — they all work well, and then the rest of the biology is totally uncharacterized.
[01:17:40] Why the field desperately needs real benchmarking
Ben: Yeah — sorry, go ahead. No, just, if you break down the problem of antigen identification, the first step is finding the tumor-specific variants and whether they’re expressed. There are tons of tools to do that, and the tools are actually fairly easy to benchmark — you can benchmark DNA and RNA experiments pretty easily. But then after that, you have to figure out, of the RNAs, which ones actually get translated into protein, and then how the protein gets processed. And that is very, very hard to actually benchmark in any reasonable way. So you don’t really know how great the tools are, or your tool is. And then the next step is predicting loading onto the MHC and the MHC getting to the cell surface. And peptide-MHC binding prediction is part of that, but not the whole story. You can benchmark that end-to-end with immunopeptidomics, but you have to really do the expensive version of immunopeptidomics, so people won’t do that either. So one of the things our field desperately, desperately needs is true benchmarking, so that all of us can improve our tools, and/or know which tools to actually use.
Alex: Benchmarking at these intermediate steps.
Ben: Yes, and the end-to-end process.
Alex: And I think some of the intermediate steps are invisible, so they’re hard to benchmark. If you look at the vibe-coded cancer vaccine stuff, their understanding of what’s going on is like, “Oh, you put the mutations in a binding affinity predictor and it tells you what’s on the tumor.” And like 0.1% of those are going to be on the tumor. So if you wanted to get closer to what is actually on the tumor, you have to start thinking about stuff that’s not modeled well and also hard to measure. How is the proteasome — the main protein-chopping machinery in a cell — chopping up a source protein with a mutation in the tumor? Did you look at the state of the proteasome in the tumor — which subunits are being expressed — and in the antigen-presenting cells, which may have an immunoproteasome that’s got different subunits?
Okay, now we’ve figured out how it’s been chopped up, and some of the pieces line up and some don’t. It turned out the tumor is deficient in TAP, so it’s not getting into the endoplasmic reticulum at all, so you really can only look at proteins that are localized to the ER and you have to ignore the ones in the cytosol. Or no, it has TAP, but it’s missing ERAP2, so some of the trimming as you’re loading is not — that kind of stuff really does affect you. We look at experiments of mass spec of the MHC ligands with these different deficiencies, and they get totally different sets of peptides. And people, if they even knew to think about it, are usually overwhelmed, and they just go back to, “Well, I hope that NetMHCpan does a good job.”
Ben: Or whether tumors are surrounded by sufficient concentrations of interferon gamma changes what their antigen presentation is.
Alex: Interferon gamma totally changes all those sentences.
Ben: So — but this is, again, why I’ll make yet another plug for immunopeptidomics, because it doesn’t matter if only 1% of your predictions are real if you can use immunopeptidomics to actually find the real ones. If you can use all the predictions to get you into a zone where you can make a reference library for immunopeptidomics data analysis, which you need to do, then you use immunopeptidomics to find the real ones. It’s hard —
Alex: Yeah.
Ben: — but then you’re actually good.
Abhi: You don’t need to understand the intermediate biology.
Ben: Yeah. And I guess, Alex and I have had this discussion, but eventually we’re going to understand the intermediate biology well enough to make those predictions.
Abhi: Mm-hmm.
Ben: But until then, we need to know the truth to work from, to then go back and study that stuff.
Abhi: Okay. So clearly the process of identifying peptides within the MHC is incredibly complicated. 0.1% of the candidate space is actually there at the scene. Okay.
Alex: Don’t quote me on that.
Abhi: Sure, some —
Alex: We’re recording it, but — a small amount.
Abhi: A small amount.
Alex: A small amount. The exact amount varies depending how you do it.
[01:21:57] Moderna and BioNTech: are the success stories real?
Abhi: Yeah. Despite it all, though, it seems like there are at least two seemingly success stories of neoantigen vaccines — one by Moderna, one by BioNTech. I think the BioNTech is for pancreatic cancer, Moderna is for melanoma. What’s going on there? How are they able to solve such a seemingly impossible problem?
Alex: Right. There are two different directions to go there. One is to talk about how close this is to really working, and the other is, do we currently have the working manifestation? My hunch is that we do not currently have the working manifestation. So I don’t think that either the BioNTech pancreatic cancer vaccine or the Moderna melanoma vaccine are it, in the sense that they will get clinical responses you could attribute to the vaccine at a rate high enough that they care about commercially. They might get an approval — it really depends on exactly how the Moderna Phase 3 trial plays out.
But what’s useful about them is they give us additional hints of success. In my mind, the really strong success cases are not those. They are: we can do TCR-Ts against cancer-testis antigens, and that seems to work pretty well. In a smaller subset, we can do TIL therapies and TCRs against mutational antigens, and that also sometimes works. So T cells can kill cancer cells based on tumor-specific MHC-presented targets — that’s definitely true. We also, for the shared targets, have a few success stories — viral ones and CTAs, with BioNTech’s FixVac, where the vaccine itself actually has a clinical effect separate from any other therapy. You could just point to it and say, “The vaccine did that.” Small minority of patients, or very early-stage cancer, but that still works. So you’re in the neighborhood of success.
And then you get to the trials that have currently been running, and they do have some hints that they might be working. They also have some complicated other things to figure out — they’re all combinations, there’s a lot of agents active, only the Moderna trials are randomized, the BioNTech one is not randomized. So you have to start peeling apart pretty complicated chains of causality. But I’m going to be optimistic and just say, okay, could we construct a story by which they’re working? The BioNTech vaccines have maybe 20 targets, and in a subset of patients they do seem to get pretty strong immune responses, and those immune responses are concentrated in the patients who are still alive from their 16-patient trial.
It’s definitely possible that some of those patients are alive because of the strong immune response lining up with a lucky choice of antigen that is actually on the tumor. I don’t think all 20 antigens in the vaccine were on the tumor; probably most of them weren’t. But for a subset of the patients, you get an alignment of strong immune response and tumor-presented mutational target, and that could have clinical benefit that’s manifesting in people with a really nasty cancer not having that cancer now. So that feels like a hint of success for which we want stronger evidence. The Moderna story is kind of complicated, because — it’s hard to say for sure, because neither company wants to run big comparisons of their platforms; no one wants to lose. But Moderna seems to have a less immunogenic vaccine. They have to inject a much higher dose, so they don’t seem to get the same kind of big circulating T cell responses. On the other hand, they have more antigens — they have thirty-four antigens. So they might be picking out more targets that are actually on the tumor.
Abhi: That also means they have more opportunity to lose, right? Because of immunodominance.
Alex: Because of immunodominance, yeah. This is a place where immunodominance is not a total curse. You do sometimes get, in one patient there’s one response, but in another there were five — and what they mean by “response” changes quite a bit. So there’s a way you could imagine the Moderna trial also having clinical benefit for a subset of patients. To me it’s slightly less clear than the BioNTech one, because BioNTech has this necessary precondition for success — these huge T cell expansions.
So given the background that a bunch of other T cell therapies and vaccines have had clinical benefit, and here’s a mutation-targeting vaccine that gets strong T cell responses, it doesn’t seem crazy to me to say that some of those patients are benefiting. But it’s probably not all eight that are still disease-free at this point. And I don’t know exactly how many — it could be one, it could be three. So the job now is to take everything we know and all the successes we’ve observed, and start maxing out these orthogonal categories, because you need the vaccine to succeed in multiple ways. We know that it can sometimes, so now you just need to start dialing it all to eleven.
Abhi: I’d like to hear from your clinical experience, Ben. For me, PDAC just feels like — okay, you get this, you die. Maybe the patient is fortunate enough to get it resected, it’s early-stage enough that that’s fine. But does it almost always come back? Eight patients had an immune response, seven of those patients never got recurrence. Does that sometimes happen, or is it extraordinary enough that you feel like —
Ben: With all caveats that I’m a cell therapy clinician and not a solid tumor oncologist, much less a pancreatic cancer doctor — yeah, I think it can happen.
Abhi: Okay.
Ben: That, say, people have a treatment course that doesn’t include the vaccine, they could still be living without pancreatic cancer if they had complete resections.
Alex: And I was trying to be optimistic, in that I do think this stuff is starting to work, and that we’re in the neighborhood of making mutation-targeting vaccines that work. On the other hand, it’s a 16-patient trial. The eight patients with stronger immune responses have many favorable baseline characteristics — where in their pancreas their tumors were, the degree of lymph node involvement, the size of the tumor, other factors. If you go through that supplemental table of all the characteristics of the patients, every time you look at something that could make a difference in their outcome, it usually is concentrated in the patients who were in the eight with immune responses, who are now perceived as having received clinical benefit from this vaccine. So there’s a lot of confounding there, and a very small sample size. I don’t think this is the trial you can point to and say, “Oh yeah, we got neoantigen vaccines working, look at the pancreatic cancer trial.” But when you line it up with the other evidence, it feels like you can’t totally dismiss it. Which is better than the entire past generation of neoantigen vaccines.
[01:30:04] Tumor mutational burden, and scraping the bottom for glioblastoma
Abhi: One thing I was trying to tease out, with the help of Claude, and I wasn’t really getting anywhere — pancreatic cancer and melanoma are on two different sides of the spectrum with regard to tumor mutational burden. Should that weigh into this at all when you’re deciding how well neoantigens actually work in this type of cancer, or is it irrelevant to the problem?
Ben: I think if you have more predicted neoantigens to choose from and features to prioritize with, maybe you’re more likely to have a good one that gets into your vaccine, because it’s actually there and it’s actually presented. So it does seem like there’s a higher prior for melanoma to respond over pancreatic cancer.
Alex: And my experience in doing some of the vaccine designs in a couple different trials is that when you have a high mutational load, everything at the top looks great. “Oh, there’s like 700 RNA reads spanning this mutation, and the predicted affinity is single-digit nanomolar, and everything just lines up. The clonal fraction of this mutation is 100%.” They all kind of look like that. Smoker’s bladder cancer looks like that, melanoma often looks like that. When you go into low-mutational-burden cancers, you’re really scraping the bottom. We’re trying to fill 10-peptide vaccines for glioblastoma, and when you get to peptide number eight, you’re really like, “Oh, should we lower the minimum number of reads?” You start twiddling with the filters to include everything, because one of them had 12 RNA reads and the other one’s got three. So you definitely end up with worse candidates in the vaccine. Pancreatic cancer is not going to be as bad as glioblastoma, but it’s closer to that, where you’re not able to prioritize things that all look really promising. You have to set your thresholds kind of low in order to fill out your vaccines.
Abhi: The BioNTech one, I think, failed in CRC and smooth muscle — I forget exactly what type of muscle cancer. Does that update you one direction or another?
Alex: I thought they also had a melanoma one, right?
Abhi: They also had a melanoma one that failed, but it was not resection, I think.
Alex: Yeah. I mean, until really recently I’ve been kind of a neoantigen pessimist. We ran all these trials at Mount Sinai, and then I was like, “Oh wait, these things don’t work, we should either fix them or find a different therapy.” Ben and I, through a research journey, found cell therapy, and we were like, “Oh, we should really focus on cell therapy for all characterized antigens. We should always do mass spec, because the predictions are useless.” And really recently I’ve kind of realized that, oh, you actually probably could get the machine learning good enough. We do have a lot of new data sources now, you can think through the biology better. I’m feeling optimistic that some of the stuff we’re building is going to help make the in silico predictions possibly good enough, combined with the fact that BioNTech figured out a great vaccine platform — and there are a couple other ones, like maybe the Elicio vaccine’s good, it seemed like Gritstone had an okay one. So the platform question is also a little bit more settled, in that there are vaccine platforms that look decent. So if we make the informatics better and make the vaccine platform good, all this stuff could work. My hunch is that the trials we’re starting, at the dawn of this era — these designs are locked down, once they have the platform submitted they don’t want to change it — so the design decisions here, informatically, are probably pretty primitive. I don’t think they’ve — I mean, I might be wrong, someone from BioNTech might call you up and be like, “This is total bullshit, we did like $20 million in mass spec to make a better model.” And I know they bought Neon, I think, because Neon was a vaccine company that had a lot of mass spec data. So I’m sure that’s on someone’s mind there. But I don’t think these trials have really integrated the advances in mass spec profiling and data generation, or in deep learning, so they probably are mostly putting noise into the vaccines.
[01:34:36] What a rational pharma would build — and the Gritstone/Neon graveyard
Abhi: If I was a rational pharma company, what I would probably do is set up a pipeline of cryopreserved frozen tissue to mass spec and train a really good model for what is actually surfaced on a tumor cell, given that this is its genome. You don’t think that’s what they’ve done, and that they’ve actually just picked the naive, maybe OpenVax-looking approaches?
Alex: I think the stuff they were starting before all that data was generated was kind of naive — the way we built it for OpenVax, and the way lots of other trials have run. But at the same era, there were people, and really two companies, who bet much of their strategy around doing what you’re saying — Gritstone Oncology, who I think now doesn’t exist, and Neon, who got sold to BioNTech quite cheaply. Their idea was that they’d generate a bunch of mass spec data and figure out what’s really on the cell surface. It was more nuanced and technical than you’re saying — they weren’t all different tumors, and they weren’t looking for mutations, they weren’t doing the weird fancy kind of mass spec Ben and I are talking about, so they had to calibrate more off of self-antigens. But they did generate a bunch of data, and they did build models. We had a sponsored research agreement with Neon to build machine learning models on top of that data, and I think that probably was moving them in a better direction as far as the quality of the predictions. However, you really need all the things to work.
Abhi: Yeah.
Alex: So my sense is that Neon made some progress in the predictions using mass spec data, but they were using a really weak vaccine. And Gritstone did some stuff with mass spec, but I never worked with them; I don’t know how it ended up. Their vaccine was better than Neon’s, but still, I don’t think they hit the whole checklist of stuff.
Ben: I thought Gritstone was not doing personalized vaccines.
Alex: No, they had one.
Abhi: They did.
Alex: They had a shared-data personalized one — GRANITE and SLATE — and they both failed, and they went bankrupt.
Abhi: Good names.
Alex: Yeah.
[01:36:46] If cancer is so heterogeneous, why does any of this work?
Abhi: If I was interviewed on the street and someone asked me, “Is a cancer vaccine possible?”, I would probably say no, because within a solid tumor there are so many heterogeneous subpopulations of tumors, and each one is going to have a completely different MHC presentation. I know for pancreatic cancer it’s a very clonal cancer, many of them are going to share neoantigens. But I would expect for something like melanoma it would be highly heterogeneous. How come it works?
Ben: Yeah. So now you’re getting another design point, beyond my soapbox about immunopeptidomics. I also have a soapbox about: you should cover all the tumor subclones with your vaccine epitopes. You can measure that if you have a sample and you do single-cell RNA sequencing, and you’ve identified the antigens and know what their coding transcripts are.
Abhi: But you’re going to measure them for the sample that you picked.
Ben: That’s right.
Alex: And then somewhere else is going to be kind of different.
Ben: Yeah. We can’t see all of it, but — so then let’s overlay that with what’s the right clinical context for a vaccine. It’s probably not huge-burden metastatic disease, but these vaccines that seem to be working better are given in the adjuvant setting after surgical resections.
Alex: I strongly — I think that’s become a meme that the whole vaccine industry drafts off of, and I think that actually, when vaccines work, they’ll work even with established diverse disease, because your targets are great. And it’s literally that one patient in the BioNTech melanoma FixVac trial that gives me hope. If you have a good target, you chose well so that it spans clonal diversity — they all still have to make it for some reason — and your vaccine’s really good, and it’s really presented, I think you could get —
Ben: Don’t you think you need multiple targets, then?
Alex: Yeah, I think they’ll escape from the one target.
Abhi: So each of the neoantigens you put in the vaccine is a good one, and so you cover the full breadth of all possible.
Alex: Yeah.
[01:38:40] The hyper-optimistic case: metastatic disease and antigen spreading
Abhi: I guess in your hyper-optimistic case, you think a patient with very metastatic stage four cancer, in the cancer vaccine of the 2030s, it should work in that case?
Alex: I think it’s not crazy to imagine that. There are a few things that work in your benefit. Let’s say you have a great platform, and some large fraction of your immune system is now mobilized against the cancer. So all over the body you’re getting these encounters with T cells, and they’re cytotoxically hunting down the tumor cells. And then hopefully — this is a hopeful thing — you will reveal more of the tumor antigen to the immune system. So you get some antigen spreading all over the body, so different draining lymph nodes of these encounters are then eliciting additional anti-tumor immune responses. That’s a way by which you can go from, “Twenty percent of my T cells are going around killing cancer.” They alone are not covering the full diversity of the presented tumor antigen, and that vaccine-induced burst of T cells — twenty percent’s your peak, and they’re declining — that is not sufficient on its own to do it. But coming up behind them you have additional waves of T cells being primed from the killing that that first wave is doing, and they’re also going to cover more of the diversity. But I think you need a big burst. It can’t be a subtle — a little bit of immunogenicity doesn’t do that. You need a really big burst of cytotoxicity.
Abhi: I did see people poking at the possibility of polyclonal responses. Where does that come from? My impression is, you give the T cell all the information you know about the tumor, it goes off and performs the killing of those particular tumors expressing those neoantigens, but then it stops after that, because it just doesn’t realize that all these other things are also worth killing. Where do you get this extra diversity from? Why do you get the spreading?
Alex: I mean, it doesn’t always happen, and figuring out the context of what makes it happen most efficiently is important. But the basic idea is that when a T cell kills a tumor cell, all that debris, with the cytokine context of the killing, goes to the nearest lymph node. So now you have a bunch of antigen-presenting cells that are like, “Oh wait, T cells were killing cells that were making this.” And the “this” is just all the debris of the cancer.
Abhi: But don’t you go back to the previous problem of, most of this debris is self?
Alex: Yeah, you do. It’s hard, it’s a hundred percent — there’s some adjuvant effect from whatever you give with the vaccine as well. But —
Ben: So maybe your dendritic cells know.
Alex: But the immune system always has this problem, and it’s not like it’s always succeeding or always failing — the context matters a lot. The debris draining in the context of a bunch of T cells that killed it, that are secreting all the cytokines — the general signaling of a bunch of cytotoxic T cells showed up here and killed like ten million cells — that pushes this imperfect system further toward, “Okay, figure it out, find the thing. Something here was dangerous, we need to elicit more responses.”
Abhi: I could buy that. I hope the future you’re describing comes to pass. When I read these clinical trial readouts, I got initially very excited, and then I saw, “Oh, this is resection.” In an ideal case, you are able to do this in the metastatic.
Alex: Where they’re positioning is just an easier problem, and it probably makes sense if you don’t yet know the magnitude of impact you can have. But it also makes it much harder to know when you’re having any impact at all.
Ben: Yeah. And clinical trials have to be bigger and run longer so that you can see differences in groups and recurrence. But if you have widely metastatic disease and then elicit a CR in two months.
I mean, that’s a case report in the New England Journal of Medicine, right?
[01:43:00] Antigenic drift, driver variants, and the 2030s adaptive vaccine
Abhi: On one hand, I would expect cancer vaccines not to work because there are just so many subpopulations. On a similar note, I would expect there to be almost antigenic drift over the months it takes to actually manufacture these things. Is that practically speaking not really a concern, because there’s no selective pressure to pull away from these particular neoantigens?
Ben: It’s definitely a concern. We’ve seen that very thing happen in mouse models, so we presume it happens in humans as well. And that’s one of the limitations of the earlier approach of trying to find antigens by just finding antigen-specific T cells, without doing immunopeptidomics at the time you’re going to treat. What you’re seeing then is just an immunological snapshot of what the T cell populations have seen of the tumor in the past. They exist because they were elicited before. That doesn’t necessarily mean that the tumor clones and subclones that are there now are still expressing and presenting the targets of the T cells that are there trying to fight them. It’s evidence that there may have been some antigen exposure of the T cells at some point in the past; you don’t know how current that is.
Abhi: Yeah.
Ben: So, again, immunopeptidomics as soon as possible before treating, I think, is how you mitigate that as best we can now.
Alex: But your point that it’s a dynamic system — your snapshot at time of surgery, even if it takes five months, is worse than if it takes one month. So it does push you to shorten the timelines. On the other hand, maybe it’s just a fundamental problem — we got everything else figured out, and then you have to figure out this additional problem that it’s really a diverse moving target. So there’s an interaction between that drift and just the baseline diversity. You’re not going to cover the entire cancer, and this one subclone’s increasing over time. So you generally have to have a story for why you’re going to get coverage over lots of subclones, or everything you’re targeting is somehow mechanistically linked to its being a cancer. There are a few cases where you can have that — if you’re targeting the driver variant, it’s hard for the cancer to lose the driver variant.
Abhi: Sure.
Alex: So it makes KRAS and p53 mutations really appealing, but it’s not useful for most people — they’re not presented on most MHCs. So then you have to figure out what else is there that’s closer to linked mechanistically. But for all your passenger-y things, you do want some sense of diversity of coverage, and also some mechanism by which you hope other antigens get picked up. You could also do this adaptively. The 2030s version is not going to be a single surgery snapshot followed by a four-month manufacturing process that you just hope works. Probably it’s going to be endlessly, continuously adjusted. You pick up any cell-free DNA with a hint of a mutation, it enters the set of candidates. If you pick up cell-free RNA — sometimes there’s cell-free RNA — and it’s got a hint of expressing a mutation, that goes into your log of, “Here’s what I think may be in the body now.” Maybe you have other minimally invasive ways of keeping tabs on the cancer — extracellular vesicles can sometimes come off the cancer, and you look at their MHCs. So that kind of continuous monitoring and then adjustment — and if it’s a personalized therapeutic, adjusting it to be new is not that hard. This is much harder with fixed-composition matter drugs: “Oh, the inhibitor for that receptor stopped working — which other small molecule things exist?” You’re limited by someone having spent 20 years making a compound for you. But if you’re making the drugs from scratch anyway, then that adjustment could be much more dynamic.
Abhi: Have any of the liquid biopsy companies explored this sort of stuff — how the cell-free DNA actually points you?
Alex: There’s definitely interest. I don’t know which of those interests I can talk about. But they definitely are interested in this, because it’s a very natural use case for their product. If you had a therapeutic that you had to keep adjusting based on a cell-free tumor DNA assay, then it’s a natural match for them. I think they want to be in that business.
Ben: And one of the technical challenges is you really need whole-genome sequencing for that to work. You don’t want to be limited to some set of mutations you’ve observed in the initial tumor biopsy and only see those forever into the future. You want to see those changing, but you also want the capacity to see new ones arise if they do.
[01:48:02] Why cell therapy at all, instead of antibodies?
Abhi: That makes sense. This is a side question, but something I just recently thought about. Why even do cell therapy at all ? Why not instead just infuse polyclonal antibodies that are geared to hit the specific neoantigen?
Alex: Yeah. If you could make them.
Abhi: Yeah.
Alex: So this is a thing. There are TCR T-cell engagers. The way they work is you have an affinity-matured T cell receptor that recognizes that peptide-MHC target, and it’s a bispecific, and the other side is an anti-CD3 Fab or something like that. So it grabs any passing T cell and pulls it in and gets it to start signaling. Any T cell is converted to a tumor killer. It’s a really nice idea. In the lab, we have not found them to be as effective as a cell therapy. The T cell receptor signaling in a cell seems much stronger than what’s induced by something grabbing onto CD3 and forcing proximity. There are also TCR-mimetic bispecifics, and those are just straight-up antibodies — they make antibodies against the peptide-MHC and do the same anti-CD3 trick to induce cytotoxicity.
Ben: So it’s the making of peptide-MHC antibodies that are highly specific for the peptide-MHC and not cross-reactive — that’s the real challenge there.
Abhi: Got it.
Ben: But if you could make those at will, then sure, you would want to give them. And you could give them as antibody-drug conjugates if you wanted, or whatever modality.
Alex: Radioligand therapies —
Ben: That would be fantastic if one could make those things quickly and reliably. And I think we’re —
Alex: We’re entering — we’re not there yet, but we’re entering a world in which you could theoretically make antibodies on demand for peptide-MHCs.
Abhi: That’s what I was going to ask — how real is that field?
Alex: People are doing that. A hundred percent, people are doing it. There are challenges specifically with peptide-MHC targets. It’s much easier to do this when you have either a really predictable structure or a true crystal structure from a protein that’s not such a delicate little target area. The peptide-MHC complex is weird, so there are very few of them in the PDB, and the AlphaFold prediction of that structure gets a lot of the side chains wrong. Because you’re targeting something tiny — it’s a nine-amino-acid peptide, and three amino acids are kind of buried in the binding groove, and you’re trying to recognize the pattern of the side chains sticking out of the groove, and it has some flexibility, so it moves around a little bit. And if you run those through AlphaFold or Boltz or something, they don’t really come out looking the way they do when you know the structure. And I think that makes it harder to get something that’s a really specific binder for just residue four — there’s a valine-to-arginine mutation at residue four sticking out of the binding groove, and the way it’s going to be recognized by a TCR or a high-quality TCR mimetic — it’s hard to just de novo design those binders right now. To a degree that’s a data challenge or machine learning challenge, but it’s at the frontier of what you can do with BindCraft or whatever. Those are pretty hard to make.
Ben: Yeah. The other piece of it — and this is super exciting, by the way, so thank you for bringing it up — the other question is, how are you actually going to translate this, presuming you could make your binders well, in a way that the FDA is going to allow you to give it to a person without some kind of extremely detailed toxicity screen? How are you going to prove that it’s not just specific for the peptide-MHC you’re looking for, but that it doesn’t cross-react against some other important protein in the body that it would be exposed to, such that it could cause toxicity?
Abhi: Well, I would hope they wouldn’t ask questions here if they’re not going to ask questions for the neoantigen vaccines or the TCR-Ts.
Ben: So I’m talking about binders that are generated completely novel — de novo binders.
[01:52:43] Neoantigen vaccines as a categorical departure for the FDA
Alex: I think maybe we should come back to this, because I do think soluble therapeutics are an interesting direction to go in. But maybe one thing we haven’t talked about that’s the most important thing about neoantigen vaccines is that they’re a complete departure from how the FDA has ever interacted with any therapeutic ever. They’re categorically different from all other medicine. And I think that’s really important, because they are a computational process that designs a therapeutic from scratch, and then you make it for the first time ever and inject it into a person after fairly minimal testing to make sure that you made what you thought you made and that it’s not contaminated with something. But all of the usual logic about drug development is totally out of view in this process. And that is much more AI-shaped than anything else in medical development. So when I see all this interest in making better AlphaFold-type models like IsoDDE — “this is going to rapidly accelerate medical progress” — I’m kind of skeptical, because every single compound that comes out of that still enters this really, really slow process. So maybe better things enter the pipeline, hopefully, but they’ll still take ten to twenty years and a few billion dollars to develop. Whereas our scrappy little trials at Mount Sinai treated sixty-ish patients with hundreds of different compounds that we made totally computationally — if you treat every single peptide you put in a vaccine as a different therapeutic, we’re outcompeting GSK. And I think if you can start fitting more therapeutics into that shape, where your safety prior is really strong, so you have a reason to believe — and this is the issue with, can you do peptide-MHC binding bispecifics, radioligand therapies or something? The safety logic’s not there, so it can’t be neoantigen-vaccine-shaped to the FDA, because they’ll ask some really reasonable questions like, “Won’t this bind to some unexpected protein and then just melt people’s eyes?” And you won’t be able to answer, “No, of course not,” because you don’t know. So the thing you need is the strong safety priors that have gone into fully personalized therapeutics, which are essentially neoantigen vaccines and also neoantigen-targeting TCR-T, which one company really did.
Ben: Maybe the n-of-1 gene therapies are developing along that line.
Alex: They’re in that direction, but if you look at the amount of work they do for that single patient, it’s much more than for a personalized vaccine. It’s still —
Abhi: Like Baby KJ.
Alex: Yeah, Baby KJ. Baby KJ is dozens of people hustling for months to put together this package.
Abhi: But now that they’ve done it for one patient, does that not make it easier to do for every other patient, or is it the same slog over and over again?
Alex: I think it can improve it. There’s another example I like a lot, which is the antisense oligo therapies that are similar in shape to the Baby KJ case. There’s — what’s the name of the nonprofit? They have a podcast.
Abhi: Are you talking about the phage folks now?
Alex: No, the phage folks are also a great example of this kind of thing, but they haven’t done as much. There’s a nonprofit that does antisense oligo therapies for germline defects, and their development timeline’s also kind of long — it takes them maybe a year to make the therapeutic, but it is for a different mutation every time.
Abhi: Okay.
Alex: So you have to find those examples and try to stitch together what makes it safe and fast for them to do those one-offs. But the n-of-1 therapeutics are still, in terms of speed and testing burden, a tier above — a tier worse than — neoantigen therapeutics, where you really have it computational, to synthesis, to administration.
Abhi: For the neoantigen TCR-Ts and vaccines, what data package did Moderna and BioNTech need to present to the FDA to be allowed to go through the trial at all?
Ben: That’s another really interesting point. Before COVID and the SARS-CoV-2 vaccines, we didn’t really know that RNA vaccines of any kind would be safe, or effective, in humans. And vaccines to that point in cancer had largely been peptide-adjuvant vaccines and dendritic cell vaccines, and the RNA vaccines were being developed, but it seemed —
Alex: BioNTech, yeah.
Both BioNTech and Moderna had their cancer discovery —
Ben: No, I know. No, no, no, I know that. But that they were —
It wasn’t wildly dissimilar in the process.
Alex: Right. You’re saying it gave us something prior.
Ben: Yeah.
Alex: Like, they’d run small RNA vaccine trials.
Ben: Yeah. So if you look at it, peptide vaccines predated personalization with KRAS and p53 [?] , right? And then the dendritic cell vaccine — I mean, the first neoantigen vaccine, I think that clinical report was the dendritic cell vaccine in 2015. And then in 2017 you had the back-to-back Nature articles with a peptide vaccine and an RNA vaccine, which became the BioNTech vaccine. So they were co-developing, but I guess I feel like the zeitgeist in the world was that peptide vaccines have been safe, have been tried for many years, and that’s the way forward for therapeutic neoantigen vaccines. And then the success of the RNA vaccines has, at least in my view, totally flipped it —
Alex: Yeah.
Ben: — not my opinion of it, but just what I think the general view of the field is.
Alex: But I think it also matters to the FDA, because now, if you are delivering a therapeutic where at least the platform has been in several billion people, they don’t ask as many questions. The data package is not, “How do you know that your SM-102 ionizable lipid is safe?” They’re like, “Well, okay — every human on earth has now been exposed to this ionizable lipid, so probably it’s fine.”
Abhi: Well, they perhaps may not have any questions about the mRNA itself, but they may have plenty of questions about the fact that each patient is getting a new set of neoantigens. How do you know those neoantigens don’t cross-react with something else in the body?
Ben: Yeah. So the logic — when we submitted our neoantigen FixVac to the FDA, it was like a 78-page white paper and a ridiculously long, multi-sheet spreadsheet document that essentially detailed all the selection pieces, with parameterizations and tools and so on, so they can see all of the computational logic end to end. But really, I think what they care about for safety in antigen selection is, you give them evidence that this is a thing that is expressed and/or predicted to be presented, or known to be presented, by tumor cells, that’s not also on healthy cells, such that raising T cells against it would be dangerous. So that’s the key safety logic.
[02:00:03] The antigen-selection safety logic, and why it’s flawed
Alex: And that safety logic is flawed.
Abhi: It sounds quite flawed.
Alex: Yeah, it’s quite flawed. Okay, so I’ll tell you two things that make it quite bad. And this is the thing you have to worry about now that we’re closer to a regime of these vaccines really working, so you have to revisit. The original safety logic was just hiding under the cover of the fact that peptide vaccines suck — the immune responses are anemic, you almost never hurt someone. We have a paper we’re working on from a glioblastoma vaccine at Mount Sinai, where we’re not sure if one of the patients, who died but didn’t have cancer when they died, died from a reactivity to a vaccine peptide. I think this might be the only case in which anyone has ever managed to potentially be hurt by a peptide vaccine. Five other patients are cancer-free like seven years later, so cancer trials are rough. But that case had really special characteristics — they were getting regular injections in the neck-draining lymph nodes, so we’d really tried to max out immune responses in a way that peptide vaccines usually don’t. Usually you get injections in your arms, and the immune response is actually pretty wimpy, so it doesn’t matter what you say about safety — you can inject anything and I think you’ll be fine. I would feel comfortable taking any peptide vaccine with an adjuvant that’s one of the ones that have been considered safe. When you get into vaccines that are two-plus orders of magnitude more immunogenic, you do have to worry about what this is really going to cross-react with. And if you look at the famous MAGE/titin cross-reactivity, the amino acid sequence difference between the MAGE epitope and the titin epitope is pretty big.
Abhi: Mm-hmm.
Alex: It is not a one-amino-acid difference, it’s a four-amino-acid difference out of nine. It’s a pretty different peptide. And the safety logic of “one amino acid is different in this peptide, so of course the immune system would never mistake this peptide for a healthy one” is actually kind of weak. T cell receptors contact a few hot spots on the peptide. You’re hoping that in that limited view of the peptide they don’t mix it up, but they might. They’re not looking at the whole sequence.
Ben: Yeah. So this is why — thousands of people have been treated with neoantigen vaccines now. Seeing the safety signal in the large reports will be really helpful to help us understand, is there a risk of these catastrophic tail outcomes, and if so, how frequent are they?
Alex: Yeah. And I think the larger BioNTech trials will probably be informative for how often that goes wrong.
Abhi: Sorry — the MAGE cardiotoxicity thing, if I remember correctly, was a TCR one, right? Yeah. Instinctively, I’m thinking, if you were using an mRNA construct that is only eliciting your existing T cell repertoire, you can naturally worry a lot less about off-target effects.
Alex: Yes.
Abhi: Is that clean logic?
Alex: Yeah, it’s pretty clean, but a lot of these things — when I came into this as a computer scientist, I treated these as rules, and I think they really work out more as re-weightings. There are a lot of naive T cells that you don’t want to turn on. If you turned a random naive T cell into 10% of your circulating T cells, you’re not guaranteed that the thymus did all the stuff you hope it did.
Abhi: That’s just like — I mean, the existence of acquired autoimmune disease is an example.
Alex: Right. And all the immune-related adverse events under anti-CTLA-4 in particular, but checkpoint blockade in general — all those T cells were in your body, and it just required some extra mechanism to keep them from trying to eat your colon or whatever they’re doing.
Ben: Yeah. So the question is, how well does your vaccine discriminate in stimulating only the peptides you want — T cell responses against only the peptides you want — versus ones that are T cells bearing cross-reactive T cell receptors that weren’t problematic when they were low in frequency, but could become problematic, causing autoimmune disease, if they were higher in frequency?
Abhi: That makes sense.
Alex: Like, how often does that set of events happen in any individual?
Ben: I think we don’t know.
Alex: Yeah, we don’t know. I think this is to be learned from the newer vaccine trials, plus extra research to be done.
[02:04:42] Running investigator initiated trials (which exist in the US!)
Abhi: The second-to-last thing I want to talk about — maybe it’ll be the second-to-fifth thing I ask. You’ve been referring to your work at Mount Sinai, and also your work with —
Alex: I realized, as I was — “Oh, I didn’t really frame this in terms of —”
Abhi: Yeah. And now you’ve left Mount Sinai, you’re now at UNC, and you guys also have this consulting company for high-net-worth individuals for dealing with their cancer care. I would love to go through each one of these steps and talk about what you did at Mount Sinai, what you’re doing today, and how you also play into this.
Alex: Okay. So first of all, the reason I talk about the clinical trials work is, it’s mostly at Mount Sinai. There are some other trials and single-patient INDs I’ve worked with or been involved in, but the trials that I really got to play a foundational role in — writing the software, helping write the protocols, helping run the trials — were all at Mount Sinai. It’s not entirely of my own desiring, but when we got to UNC, UNC is just a much more centralized institution. Mount Sinai is a bit chaotic. And in that chaos, there are degrees of freedom that don’t exist in a larger institution, or a more logically assembled one. UNC makes more sense, and there are downsides to that. One thing that is nice in the chaos of my last job — which is now a long time ago, like six-plus years ago — is that Nina Bhardwaj got to build out a wonderful operation, and I worked with her that whole time. She has a vaccine and cell therapy lab. She took over a big chunk of lab space on the fifth floor of the Hess building and stamped out three GMP suites and can just run a ton of trials. She has a weekly meeting, it’s wonderful, there are tons of immunotherapy trials, and all the trials have their, “Oh, we need more of this or that,” or, “We need to reschedule these patients, we’re going to start a new arm where we’re going to add FLT3 ligand.” And if you have an idea for a trial, you just show up there and figure out how to make it work. Having your own GMP facility with your own regulatory people, your own research nurses — I haven’t figured out how to do that at UNC. I’m sure other people have some way to spin up trials, but I haven’t had luck with that.
Abhi: Does that exist outside of Mount Sinai?
Alex: There are productive investigator-initiated trial groups, but it’s the kind of thing I hear more about people going to China for.
Abhi: I thought — I’ve always heard IITs are a China phenomenon.
Alex: They’re not. There are IITs all over the US.
Abhi: That’s interesting.
Alex: I don’t know if there’s the same sort of fully integrated operations everywhere, but IITs happen everywhere. But having a big university that aligns its mission around “we’re going to have logically planned how we do all this stuff” actually makes it a lot harder.
Ben: Yeah. I think Dana-Farber has their program. Wash U has their program.
Alex: Yeah, all these early vaccine trials were all IITs at different institutions.
Abhi: I imagine your view of these places is positive — they’re universally a good thing to have, versus maybe the funding going into single-PI mouse experiments.
Alex: Say that again, sorry.
Abhi: The counterfactual of not having this GMP facility is that the money goes elsewhere, into non-human translational research. Is that a better use of money for cancer vaccines specifically?
Alex: I think that trials are really high-value. I also think that the cancer vaccine field has a pretty long history of trials that don’t tell you a lot. So I don’t know how to balance those two things. Being able to spin up a trial to test an idea is really, really valuable in a way that decades of mouse work won’t really resolve a question. I also think it requires some incentive structure, or push, toward, “Okay, but what is the question you’re answering with this trial?” — which cancer vaccines historically have been pretty bad at positioning their trials to really answer any question.
Abhi: Do you think IITs in general are not often run particularly well?
Ben: I think the IITs I’ve seen have been run actually quite well. They’re usually run on a shoestring budget relative to clinical trials.
Abhi: Yeah, that was my hesitating “no” — I think they run pretty well, but they’re not very well-funded.
Alex: Interesting.
Abhi: So I guess then, is the fact that people are not learning that much from this simply a fact of the underlying biology being very complicated — if you’re doing small trials, you’re just simply not going to learn that much?
Ben: Well, every group is incentivized to test just their thing, whatever their thing is. Is it better informatics for antigen identification? Is it a better vaccine formulation? Is it this clinical context or that clinical context? And so you run a small trial with a bunch of variables fixed and no comparison group, and you can have 300 of those and not necessarily know how to put them all together to figure out the actual best way to do this thing we’re calling therapeutic neoantigen vaccination.
Alex: Yeah. I think this is a funding/incentive-structure problem, where no one — if you make your trial twice as expensive but it answers a question well, you’re not necessarily going to get rewarded for that in any way, and you’ve now spent twice as much money.
Abhi: Yeah.
Alex: You’ll go from $5 million to $10 million. Your paper is not necessarily much more publishable, nor will anyone give you money in the future to do it again. So I think the fact that they’re kind of underfunded, and there’s not — there’s this notion of the trialist, and they just put the thing in the clinic, but it’s not integrated with a research question — maybe is part of why this stuff doesn’t quite all connect.
Ben: To answer your question from before, though — I think you also need mouse models, because there are too many variables in the therapeutic strategy to test them all in humans. There aren’t enough cancer patients to test everything you might want to test, like different formulation benchmarking, prime-and-boost strategies, combination therapy strategies, TCR-Ts in the mix before or after vaccination. And in mice, which are not a perfect surrogate for humans — but if we’re honest, a ton of relevant human immunology was first worked out in mouse models and then successfully translated to people.
Not everything, but a lot of it. And you have the opportunity to explore much more of the combinatorial therapeutic space, to get some intuition about what’s the best trial I can write. So if I can winnow my space 95% by doing some careful mouse experiments for a year and then have a much better trial on the other end, I think that’s better than doing another less informative Phase 1 with 12 patients in it. Maybe I need that, maybe that’s wrong, but I think if we were really working together as a whole field, we’d have this supportive ecosystem of mirroring clinical trials with deep immunological investigation and mechanistic understanding, feeding highly directed bets around human clinical trials with comparison groups in them. But that’s not where we are as a community.
Alex: And I think there’s one thing that maybe misfires in the trial context, which is that trials always are written and conceive of themselves as on the road to approval. So you’re supposed to do all your research in mice, which are okay immune surrogates, and then once you’ve made your thing, now you’re testing safety.
Abhi: Yeah.
Alex: And all of the real research happens in that Phase 1 trial that’s nominally supposed to be just for safety. And then you kind of hide the fact that you’re looking for signals that actually tell you about the biology in there. And then you do your Phase 2 trial, which academically is really tough to fund, but no one’s going to fund it as a Phase 3. So really, the IITs are all Phase 1 trials that, in our ritual of medicine, are only supposed to test safety and feasibility — could you do it? And so there’s not a strong reason to instead write the trials as, “We’re going to do a 12-patient trial, but we’re going to do two different antigen prioritization algorithms within each patient, and we’re going to do mass spec on their tumor samples and see, while we’re making the vaccine, could we find evidence of anything. Or after we vaccinate them, we’ll vaccinate them with antigens from two different sources and see which one’s got stronger immune responses. Or we’ll alternate — three doses of one vaccine and three doses of another one, and we’ll also do the reverse group, and we’ll look at immune responses.” People don’t want to write trials that way. They’re complicated, more expensive. Those would give you really scientifically valuable information. But if you’re just supposed to be testing safety for a fixed product that’s supposed to eventually get approved, you just do it and see if you have any adverse events.
Abhi: I would imagine IITs are given a bit more flexibility in this regard, as opposed to pharma companies, where they have some budget they want to put into this program before they just kill it. For academics, I imagine the sky is kind of the limit with regard to —
Ben: That’s true, but how are you going to raise ten million dollars as an academic?
Alex: That’s true, yeah. The budget’s not there to do the interesting trials, and then no one rewards you if you do it. I think that’s — and then in the broader context of trials, they’re not thought of as meant to answer the questions that they could. So you just end up with boring trials.
Ben: So you could design them that way as an academic, and I think you could send a grant application to the NIH for doing that, but even the big NIH grants, with the exception of some special huge ones, don’t have enough funding as part of it to actually do the thing. So I think they would support it and be excited about it. If you could do it for a tenth of the cost, then they probably would support more of it. But —
Abhi: Do you imagine this is a problem in China as well? Or is there some mechanism to help ensure there’s a community-wide effort to ensure that, whatever modality you’re exploring, you work on it across labs?
Alex: All my knowledge of the Chinese trial landscape is from reading people’s Substacks and stuff, so I don’t have enough knowledge to answer that. But I haven’t read about super interesting trial designs. They just sort of hit on good therapeutic hunches and then test them very directly, faster. That’s what I know.
Abhi: Mm-hmm.
Alex: But it’s possible there’s also some long tail of — what is the state of cancer vaccine research in China? I have no clue.
Abhi: Was there any attempt, while you were at Mount Sinai, to try to get people to work toward a common direction together? Or are the incentives so misaligned that it’s hard to convince people to combine their hypothesis with your hypothesis?
Alex: I think what we worked on was in some ways self-contained and had units of funding. So, test out the peptide vaccines with OpenVax — we had some funding from Pisces — and then if we add this checkpoint blockade drug, we get a sponsor that’s making that drug. We were doing glioblastoma, and we got some money from the tumor-treating-fields company. So it was really just this — it was the OpenVax group and Nina’s vaccine and cell therapy lab, plus wherever we could scrape money for the trial. There was no bigger effort there. I think now there’s a little bit more interest in, can you start gluing people together and testing this in a more systematic way.
Abhi: I actually was not aware of this pipeline of pharma — multiple pharmas — giving you some money to run this clinical trial and create some hypothesis for the for-profit companies to then develop further.
Alex: It’s a huge thing. That’s why there are IITs in the US, I think. The simpler-sounding way to do an investigator-initiated trial is to apply for a clinical R01, and that happens to some degree. But — and Ben, you can correct me, because my whole experience is essentially one hospital network in New York and a little bit of UNC — my understanding is that it’s dwarfed by sponsored research. You’re going to run a trial, but it uses this also-ran PD-L1 agent, and then we’re going to pay for a lot of the trial.
Abhi: Gotcha. So I imagine the problem with UNC is you don’t have this GMP manufacturing facility on site.
Alex: We do.
Abhi: You don’t have clinical trial people.
Alex: Oh, you do.
Abhi: Okay.
Alex: But it’s centralized.
Abhi: Okay. What’s the difference between centralized versus —
Alex: There are a lot of different people across many different compartments of a big bureaucracy that don’t have any reason to feel urgency around the stuff that you want to do. They have other plans, other initiatives and programs, so there’s a lot of magic to everyone being in one room once a week and it’s all kind of lined up. The exact same capabilities, even if they’re much more scaled up and in some ways more logical-sounding, if they’re diffuse and everyone’s not in the same room together trying to do the same thing, it just doesn’t work out. It slows it down so much, I think, is the real thing.
Abhi: It’s interesting, because — okay, so it sounds like even though everyone was in the same room at Mount Sinai, there still wasn’t the capability to really work together in common research directions, because the money was allocated to test one specific thing per group.
Alex: Mount Sinai is a good example of being able to start lots of trials, and they often have an interesting hypothesis — like, if we add FLT3 ligand to this vaccine and we do intratumoral injection, will we see some tumor regression? So you could test that. But if you wanted to do something that’s a little less traditional, like really maximize the scientific knowledge from the Phase 1 trial, no one’s going to fund it. There’s no way to double the cost of the trial to get more scientifically valuable things, other than, “Does this combination of therapeutics have some strong hint of working together?”
[02:19:40] The tragedy of the commons in cancer vaccine trials
Abhi: Why is there, like, a tragedy of the commons here, where everyone wants this knowledge but no one is willing to pony up the money to actually get it? Because it sounds like the knowledge you would get from a direct search for scientific knowledge would be useful for everyone.
Alex: Okay, so to make it more specific, let’s just talk about cancer vaccines as a trial community — people starting trials — because I don’t want to talk about medicine in general, I feel like I don’t know enough to say it. In the cancer vaccine field, in the ‘80s and ‘90s people were running TERT vaccines and these TAA vaccines, and then they were like, “Oh yeah, but CTAs might be a little better,” so running CTA vaccines. And they all kind of come out with the same thing, which is, “We got some immune responses by some somewhat discordant weak threshold of immunogenicity.” And then they run more of them, and more of them, and more of them. And they’re all Phase 1.
But once in a while they hit, and they can show that we get immune responses, and they find a company that thinks they should try to take it to market. Then they run some Phase 3s and they fail hard. None of these did anything, right? But the IIT world was just running a gazillion of these trials that are all, by design, going to get a couple immune responses. So the question is, how do you intercept that behavior and instead make it an optimization process? So rather than everyone stepping forward with the exact same information — “these vaccines are safe, their manufacturing is feasible, we got immune responses” — and there must be a thousand trials that had that conclusion — how could we have instead had a process by which they start with, “We know they’re not good enough yet, we want to test if this is an improvement”? We want to start doing coordinate ascent in each trial. I think that’s a funder problem.
Abhi: That feels very focused-research-organization.
Ben: Yeah. It’s not like we’re talking about something that’s unrecognized. I think everyone generally would agree that’s a problem, and say, “Had we had enough money, that’s what we would have done” — optimizations across the path.
Alex: Yeah, it’s not like this is secret knowledge that Ben and I stumbled on. I think you ask anyone in the field, they’d be like, “Well, yeah, but we never knew if poly-IC was actually better than CpG.” And then you ask them, “Why didn’t you do the trial?” And they’re like, “Well, you could do the trial with the CpG sponsor, you could do the NIH trial, but they want a poly-ICLC because they think it’s generally safe. No one would have funded the comparison, so we just didn’t do it.”
Abhi: Okay. That makes sense. And that was your Mount Sinai days, your PhD?
Alex: Yeah, Mount Sinai days.
Abhi: Now you’re — what was it, Pathfinder?
[02:22:38] How Ben and Alex met (on Twitter)
Alex: Yeah. Okay, so there’s a little bit of a journey here. I met Ben through Twitter.
Abhi: Okay.
Alex: And we dug up that first Twitter DM, and it was me asking about the trial that Ben was working on starting. So maybe I could let you talk about Pandevac for a second.
Ben: Yeah. So Pandevac was what we were working on and have submitted to the FDA. The idea of the trial would be to adapt the vaccine if the tumor evolves over therapy. It was going to be a trial in squamous tumors of the lung and the head and neck. We were really excited about it. We got initial favorable approval of the IND and submitted the informatics and all that. And we’ve been stuck for a few years on various manufacturing aspects and — probably more than we should go into here — but we haven’t treated anyone.
Abhi: Is it because of the aforementioned GMP manufacturing problems with —
Ben: Well, it’s too complicated. I feel like we shouldn’t discuss it —
Abhi: We could show you some email chains after this recording.
Alex: Yeah. But I think the diffusion-of-responsibility problem is really acute in this one.
Ben: Yeah. I’ll take it on myself. Skill issue.
Alex: Skill issue.
Ben: Skill issue for me. I’ll also say Alex’s experience at Mount Sinai was inspirational and, in general, a very good thing, even though they couldn’t solve all the problems. But regardless, Alex saw that we had this registered with clinicaltrials.gov and reached out to me.
Alex: Yeah. And when we first started talking, I was curious about the trial, how they’re going to design it. And then once we realized we were on the same page about, “This is a really promising direction, this is — algorithmic medicine will eventually be the future of how you make medicine, however, none of this shit works right now” — having both of those thoughts is rare enough that we really started talking quite a bit. And then eventually we came to this idea of, “Oh, I should do a job talk around how to improve cancer vaccines,” because I had thoughts from the Mount Sinai experience where those trials really were, at best, unclear. I think the glioblastoma trial we ran may have benefited people, but it’s unclear if the antigens we put in mattered.
The other two trials I was on, I don’t know if they benefited anyone, and there’s a fourth one — so we treated a bunch of people, and I didn’t feel like I could point to anyone and say, “The part that I owned, the antigen selection, had a benefit to this person.” And I had thoughts about how we might improve that. So in talking with Ben, he was like, “Oh, you should turn this into a job talk, this is just a list of things that are a good research program.” So I went on the academic job market, which is a thing I’d never had any intention of doing, and then ended up getting a job at UNC.
And, to the shock and horror of the genetics department, I just showed up and said, “Okay, I’m co-running a lab with Ben.” Which I think they were kind of in denial about for a while, and eventually were just like — still are. They don’t like it. They want a clear understanding of what each lab’s specialty is, which grants they’re going for, who owns which grants. It’s just weird for them. So we started working together, year and a half or so of disruption from COVID. We tried working on COVID vaccines. We made the only bad COVID vaccine. Literally every COVID vaccine worked except for our T-cell vaccine, which elicited very strong immune responses in mice, and they all died when you exposed them to COVID.
Ben: Yeah. Actively antagonistic. In fairness to us, in hindsight, when we started, we had no idea that the RNA vaccines would work. So, like everyone else, we were —
Alex: We threw our hat in the ring, but everyone else succeeded to some degree. Literally every viral vector, the inactivated vaccines — everything worked. We made a peptide vaccine using all the computational principles you use in cancer vaccines, and we made a thing that elicits T-cell responses and does not in any way protect mice from dying from COVID. And I was like, “Huh, this does feel related to how well the cancer vaccines work.” So we started on this program of, “We’re going to optimize the vaccine formulation.” We’ve done a lot of mouse work on making more immunogenic vaccines. We started making better informatic tools. We got into long-read sequencing as an alternative source of antigens. We then got into single-cell sequencing. We found that single-cell long-read is pretty magical — it’s become my favorite modality. We started thinking about cell therapies as the upper bound on what a vaccine could get you — it should be somewhere below a cell therapy, so why don’t we make some of the cell therapies directly? And then along the way, we’ve also started intersecting more with this thing I think you’re curious about, which is the world of hyper-concierge oncology.
Abhi: Yes.
[02:27:49] Founder-mode oncology and the rise of concierge cancer care
Ben: Yep.
Alex: So there are a few places where I’ve intersected with this. There are various research programs at Mount Sinai that had individual sponsors, sort of in the interest of developing things that might help someone in their family.
Abhi: Mm-hmm.
Alex: That’s closer to the more traditional “name a building after yourself” philanthropy, but these were more focused. You kind of knew the guy, you knew what the cancer type was, you knew why they wanted this research done. But that was still pretty distant for me. There’s also a vaccine nonprofit, the Jaime Leandro Foundation, and they essentially exist to facilitate access to cancer vaccines. It was based on the idea that it’s a complicated situation right now where you don’t know when these work, for whom they work, whatever. But there are cancer patients who are past standard of care and would like to try a therapy that at the very least is likely to be safe.
Abhi: Yeah.
Alex: And there’s some chance they might benefit. So the Jaime Leandro Foundation has this whole review board for the individual vaccines. I spent some time working on their review board, just out of curiosity, because they’re doing this thing parallel to what I had done. I learned a bit about that, and I helped with some vaccine designs. That has some overlap with — there was a program at the Rare Cancer Research Foundation that also had various patient-focused research programs. The Rare Cancer Research Foundation deals with a variety of rare cancers, but some of them come with funding, because it’s the rare cancer of a person who wants their cancer to be better understood. And through that, Ben and I started getting a little bit deeper into the world of what I think is now kind of called founder-mode oncology.
Abhi: Sure, yeah.
Alex: So there was a company that then spun out of the RCRF work. The main person doing that is Willy Hoos, and he started a company, Pathfinder Oncology, and Ben and I are minor co-founders in that. It was really started at a cafe table with me, Ben, and Willy. We’ve kept our labs and have an academic focus, but through working as consultants for Pathfinder, we kind of nudged the scientific direction. We helped Pathfinder resuscitate this personalized TCR-T company and reboot it as something that can deal better with the costs of cell therapy manufacturing. If the people you’re manufacturing for can pay the large cost, then it’s less insane to do it than if you want to bring it to market. And generally we just understood a bit more about patients who have the resources to try to pay for R&D in a way that normally doesn’t happen. Usually you’re at the recipient end of a societal program — we collectively fund medical research, and then individual bits of research get boxed up as IP, and then investors try to invest in turning that into medicine, and 20 years later that pops out as a thing your insurance will reimburse for. If someone doesn’t have any options from that entire pipeline but they do happen to have a billion dollars — it’s kind of surprising to me this hasn’t happened more in the past, but now the thought is increasingly occurring to patients that, “If I have my options, maybe I can pay for some new ones.”
[02:31:50] How good is concierge oncology, really? The Sid index case
Abhi: If I’m a billionaire with a rare cancer, and maybe I exhaust standard of care and I look toward these personalized health concierge services for oncology — how good is it really? How much better is it than just staying on standard of care and seeing how long I can get there?
Alex: Yeah, it’s a tough question, because it’s a really heterogeneous mix of situations. I think Sid is a good index situation, because he’s so public about this. You can talk to him, you can talk about him by name, and he has osteosarc.com with all of his scans and genomic data. So this is someone where privacy is not a concern — he wants his situation to be understood. And I think you could see the possibilities and the limits of this approach from the really complicated journey he went on. He does not have good standard-of-care options. I don’t know if you agree with this, but if you stay on the NCCN-guideline course, I don’t think his sarcoma would go away.
Ben: Yeah, it eventually would recur, and more aggressively, and more tumors, more places.
Alex: But he’s tried a lot of stuff, and most of it, it’s really hard to say whether it did anything. And then there are a few things that might have done quite a bit, but it was through a pretty elaborate hedged strategy that required — I think it employed the services of all three of the concierge oncology companies, one of which he started, as well as a lot of personal effort and people he’s hired just personally. So that kind of thing points at the possibility of a better outcome than you would normally get, at tremendous financial and time cost. You need to fly all over the world and hire a ton of people and try a lot of stuff, some of which may be somewhat harmful, some of which may be neutral. And it becomes a huge endeavor. Most patients, even if they could figure out how to do that, don’t want to do that. So that could also be seen as a kind of upper bound, to bracket the amount of benefit you could get if you’re a billionaire employing the services of someone to try to recreate this personalized therapy flow.
Abhi: I’ve never gotten an oncologist’s take on his actual osteosarcoma journey. He did have a cancer vaccine arc. He also had an experimental FAP drug arc, and maybe a bunch of other things in between. Is it clear — can you attribute causality? Or can you attribute most of the variance in the situation to this final thing, or was it the cancer vaccine, or something else?
Ben: Yeah, I don’t think you can attribute causality. And this is really an interesting point to think about in general — how much, and what exactly, can we learn from n-of-1 experiences? In evidence-based medicine generally, randomized controlled trials are the king; we want to be informed by them whenever we can. But at the limit, where every cancer is a rare cancer and everything is an n-of-1, it’s impossible to do a randomized controlled trial for every drug or every combination you might consider. It’s mathematically impossible — there aren’t enough patients, there’s not enough time, and there aren’t enough drugs that you would want.
But even if there were — so you have to think, what can I learn from an n-of-1 context? And if you look at how an n-of-1 trial would actually be designed, it would be for some condition where you could randomize yourself to receive different interventions, with washout periods in between, and see how much better you get from each one. And that’s actually better knowledge for you than randomized-controlled-trial knowledge. Let’s say you have a chronic condition, like chronic headaches, and you want to figure out which of five medicines are best for your headaches. You randomize yourself to take them for two weeks at a time, with two-week washouts, and keep careful logs of frequency and intensity of your headaches.
At the end of that, if you do it right, you’ve gained some information about which medicine works best for you for headaches. In the cancer context, there’s a path dependency to everything that happens one after another, so you can’t really stop and wash out after individual drugs, because the cancer is continuing to evolve as your therapeutic strategy is changing over time. So that option isn’t open to you. But on the other side, in advanced — say, stage four metastatic disease that’s exhausted all lines of standard of care — the likelihood that a person would get a deep response that lasts at all, much less for some period of time, by doing nothing at all, is minuscule. Maybe it’s not zero — maybe it’s one in a million that you could get a deep response by doing nothing.
So if you’ve taken some action that’s taking you from no response to response, and the likelihood of having a response without taking any action is close to zero, then I think you could argue you have learned something about you and something you responded to. But what you can’t do is parse it — if you’ve taken seven interventions in an intervention span, you can’t look back and say, “Oh, I know it was this one, or the combination of those two or those three, that did the job.” So I think we have to think carefully about how to learn from these n-of-1 experiences, in terms of what monitoring to do — cell-free DNA and RNA, and imaging, and other clinical monitoring.
Can we set these up such that we learn the absolute most from them? And if more people are doing this — if there’s not one founder-mode-on-cancer, but a hundred or a thousand — then how do we learn at a population level from those hundred to a thousand experiences, such that we learn more and more over time? I think the way it was best put to me, in thinking about why do this at all, is from Mark Laabs, the founder of the Rare Cancer Research Foundation and himself a cancer patient.
He said, “Developing curative therapies for really hard cancers — we don’t know how to do it, and it’s happening too slowly. And it’s not happening for all the possible cancers out there at equal rates” — some are studied more than others because of incidence numbers and so on. So he said, “We have to let wealthy individuals who are both deeply savvy and informed and risk-tolerant lend the risk capital, and the risk to their own bodies, to teach humanity how to cure otherwise incurable cancers.” I remember meeting with him some years ago, and that inspired me — what he said. It rang true then, and it still rings true now, that if we can learn somehow from Sid’s story how to do this better, and others like him, and then bring that — once we know how to do it, scale it and make it cheaper and deploy it more widely — that will have been a great thing to have done.
[02:40:03] Why old precision oncology was useless, and why now is different
Abhi: Some of these personalized oncology companies seem like they’ve been around for at minimum a decade — I think some of them are even from the early 2000s. Have they ended up bringing a lot of knowledge to oncologists that otherwise would have just never been explored?
Alex: Which ones are you thinking of? What category of company?
Abhi: There’s this one company — I think it’s called something like Personalized Cancer Care.
Alex: Oh, I don’t know what that is.
Abhi: It has a very old-fashioned website. I just know that they’ve been around at least longer than — Sid, I think, was the most maximalist, but there have been others before him.
Alex: I mean, as an industry, the “we’ll try to tailor a cancer therapy to you” thing feels like it couldn’t have meaningfully existed until pretty recently.
Abhi: Okay.
Alex: I don’t know what they did, but I know what they could have done, because the options —
Abhi: Oh, I think you’re the one who told me they convene tumor boards, and that’s all they really do.
Alex: Oh, yeah. That’s a separate thing.
Abhi: Yeah, yeah.
Alex: Tumor panels.
Ben: Yeah.
Alex: The Genki website company. They’re great, though.
Ben: They’re awesome.
Alex: They put together really well-constructed, informed tumor panels. But in this kind of thing — what are the degrees of freedom available to the oncologist on those tumor panels? Often pretty limited. And what molecular profiling could they go off of? Also pretty limited. It’s really quite different now versus 10 years ago. I saw some of these — there were like three different programs called the personalized-cancer-somethings at Mount Sinai, and they really didn’t get anything done. They all had some version of, “Well, look at your genome sequencing, and then we will tell you to take” — I don’t know, ivermectin or something. Drug repurposing, drug sensitivity prediction, things like that. They were just useless. And they might still be useless, I don’t know. But they definitely were useless when the drugs were bad, when you didn’t have really precisely targeted drugs. There’s something quite different now, in that the molecular profiling is a bit more sophisticated, but then the armament of drugs is getting way better really, really quickly. So you went from having mostly dirty small-molecule-type drugs, where you’re hoping you could pick up on some statistical signal of how they’re well-matched to this cancer, to: well, we look at the single-cell RNA sequencing, and here’s the tumor cluster, it’s making 200 times more of this receptor than any other cell population, and when we look in our catalog of possible drugs, there is an antibody-drug conjugate for that receptor, and also a radioligand therapy, and also a T-cell engager. So we’ve got to pick which one of those three we want to use. That is a really different situation than has previously existed.
[02:43:07] LLMs, and patients advocating for their own testing
Abhi: I guess another thing is, there’s this whole Rosie’s-dog thing about designing a cancer vaccine — and one can go back and forth as to whether the vaccine did anything versus the doggy Keytruda. But it did seem to me like that world could not have existed without access to LLMs to teach people about all these things. I’m curious to get your perspective — your patients, do you find that they’re opting for stranger treatment because they’re operating off of the 2026 literature?
Ben: No. The difference is — I’ll give you an example of what an intervention might be. Let’s say you have a tumor type, metastatic disease, treatable by chemotherapy maybe, but it doesn’t usually get, say, HER2 testing.
Abhi: Sure.
Ben: But some molecular studies get done on a broad panel, and it lights up for HER2. Well, now there are multiple clinical products that can target HER2. So then that person can go to their treating oncologist and say, “I know I have cancer X that doesn’t usually get treated with HER2 therapy, but look, I have this study showing that my tumor is positive for HER2.” And that can be done in a pretty clean way. You could have an RUO screening study that surfaces possibilities, and then the sample goes to a clinical lab for a CLIA test. So you’re still working on your approved pathological test that can guide therapy. And then the oncologist and the patient can, in shared decision-making, figure out if they want to do that or not. But the product is actually more information for the person.
Abhi: Well, I guess that’s what I’m talking about. Maybe not every oncologist is able to stay as on top of the literature as the LLM is now able to. Do you imagine patients are able to better advocate for their own care in this post-LLM world in oncology?
Ben: I think they are now, and more will be very, very soon, for that exact reason. But you’ve still got to make the testing available.
Abhi: Yeah.
Ben: And then it gets back to, who’s going to pay for that? Because there are a ton of second- and third-order effects of gating all medical diagnostics and care decisions by manifests of what the insurance company will support.
Abhi: Mm-hmm.
Ben: And one of those is that not everything is widely available, and we have to work within that system of constraints.
Abhi: That’s fair.
Alex: But I do think there’s eventually going to have to be a moment in which testing gets broader, because — pharma’s actually done a really good job. I was surprised at the rate at which useful-seeming drugs were all coming online at the same time. And that’s despite all the flaws people talk about — clinical trials are hard to start, it’s really expensive, whatever. We still have a ton of new drugs. A bunch of those are probably going to be therapeutically beneficial. And to know if they’re useful for you, you would, as a cancer patient, want to get at the very least RNA sequencing, which is surprisingly hard to get, right?
Ben: DNA, RNA sequencing, and a basic proteomics panel of your tissue slides.
Alex: But even the Tempus RNA-seq is kind of seen as exotic.
Abhi: Really?
Alex: Yeah.
Abhi: I kind of assumed there’s a —
Alex: No. Not everywhere. It depends on the hospital, depends on the indication. Some indications, they always get Tempus exome and RNA, but often, if you’re doing genomics, you’ll get an in-house panel that the hospital can reimburse for — a targeted panel of like 500 genes. MSK does MSK-IMPACT, Stanford does their thing. They don’t do RNA, and then you need to do some special extra thing to get the information about even the hint of a targetable protein.
Ben: Yeah.
Alex: So there’s going to be increasing pressure toward surfacing these targets.
Yeah.
Create that candidate list in a higher-throughput way, and then do some validation in a higher-throughput way. You can’t be limited by, “Oh, sorry, the path lab has no antibody for this thing that might save your life.”
[02:47:30] The molecular-testing trial that should exist
Ben: So — you were talking before about this being a problem suitable for a focused research organization. This is really a problem. Imagine a clinical trial where you take 200 cancer patients — it’s a big basket trial — and you randomize them into a bunch of molecular studies for screening, CLIA validation of hits, and return that information to the patient and the treating clinician, versus standard of care to 100 patients. And you just follow them out five years.
Abhi: Yeah.
Ben: And then you answer your question. You see if it does or doesn’t matter.
Abhi: Yeah.
Ben: Does this additional information actually lead to more effective therapeutic decisions for the patients and the clinicians, or has standard of care just been fine all along?
Abhi: Do you have a suspicion on which way it’ll go?
Ben: Oh, I definitely think more acting on more molecular testing.
Alex: I think 10 years ago it wouldn’t have made a big difference, and there were trials like this. There was one big trial trying to do precision oncology with a really limited set of therapeutics, and they had stuff like BRAF inhibition with subclonal BRAF mutations ending up not mattering in some cancer. So they’re like, “Oh, precision oncology doesn’t really work” — in 2015. I think if you did it now, it would be dramatically different.
[02:48:50] Single-cell long-read as the one true assay
Abhi: I remember I was looking through Sid’s website, and he had — “Oh, I did single-cell RNA-seq,” and I was like, “Oh, that’s normal, right?” And then I asked the oncologist on our team, and she was like, “No, no one ever does that.”
Alex: Sid has beautiful data. He’s got a bunch of single-cell long-read RNA sequencing, and him doing that, and a few other Pathfinder clients doing it, completely won me over that as the one true assay. It’s really all you need. You get the T cell receptors of the TILs, you can see the mutations in these long transcripts. You just had to do one assay, that’s the one assay, and it’s almost never done. A handful of billionaire cancer patients have done it on their samples.
Abhi: There’s no diagnostic test that requires that, right?
Alex: No, no.
Ben: No. So if you wanted to return it to clinicians in the classical fashion, you would have to get that whole thing CLIA-validated in some lab, with no actual biomarker case for funding it.
Alex: Yeah, just for permission.
Ben: Just for the fun of it, but —
[02:49:56] What would you do with $100M equity-free?
Abhi: We’re almost at time, but I want to hear a short paragraph from each one of you. If you were given $100 million, what would you do to push this field forward as fast as you possibly could? Go first.
Alex: Okay. I really want to do within-patient controls in clinical trials. I want to run clinical trials that give you maximal information about your design decisions, and clinical trials cost a lot, so you could spend $100 million on that. The boring version of that is things that involve vaccine platforms — immune responses, that’s not the really interesting part. I think the more interesting part is doing targeted mass spec validation, quantitative immunopeptidomics, on all the things that you put in the vaccine from different selection methods. So you take your algorithmic signals — that currently you just bake in a couple things based on your priors with no real validation — and you fill the vaccine from multiple ranked lists. Take two ranking methods, whatever those in silico signals are, fill a vaccine with both of them, and then you have the patients — they’re going to have a resection, you can get a lot of tumor tissue — and then spend a bunch of money making the heavy-labeled peptides and confirm which of those antigens are actually in the tumor. Ideally also do it on some antigen-presenting cells from the patients, and work through the whole chain of causality: I vaccinated them, their immune system presented these peptides, their tumors also presented these peptides, and then they had strong ex vivo responses against these particular epitopes — and that’s why the vaccine didn’t work, because the two immunodominant ones were not the four that lined up between the APCs and the tumor. Because they’re trials, they would burn through a lot of money that way, but it would teach you stuff that hundreds to thousands of trials that don’t do that do not teach you.
Abhi: I have questions, but I’ll ask you them later and put them in the text of this video. What about you?
Ben: So I don’t disagree with Alex at all, but, assuming I have my own $100 million — I would set up a mouse translational system where the animals are essentially wild and are carcinogen-exposed, but I also had complete control over experimental design. And then I would create a co-clinical trial context for testing complex immunotherapy strategies and optimize vaccination, TCR-T, immune checkpoint inhibition combinations — but in ways where you have all access to tissue and so on and so forth, and can understand mechanisms.
Abhi: This feels like something you could get a grant for. It feels very NIH-shaped. What’s the —
Ben: I haven’t tried.
Abhi: Okay.
Ben: And maybe. But it would cost a lot more than $2.5 million over five years for one R01. And I don’t honestly know what I would learn. I have 20 hypotheses about what therapies and therapeutic strategies would work best and why, and I could test them all in parallel if I had hundreds of mice being enrolled every month. But if I have to just do one track at a time, then it almost becomes not worth it.
Abhi: That makes sense.
[02:53:27] The automated box: tumor in, RNA therapeutic out
Alex: Okay, so I came up with a different one. The one I gave you is more the one I’ve been incubating in conversations for a long time — “Oh yeah, we should profile everything deeply, figure out causality.” But this conversation about the founder-mode cancer trial — that’s not limited to cancer vaccines. You do high-throughput profiling on — take half the patients, they get standard of care, do high-throughput profiling on the other half — and ideally you could structure it so it’s not just a suggestion to the treating oncologist, because that’s the way to make it acceptable to the existing system. But maybe you have an actual shared platform where you make therapeutics for them, like RNA-encoded therapeutics against a wide variety of targets. So, single platform, informatically selected target and therapeutic, and then see if you can make the box — literally, tumor tissue goes in, profiling happens, target selection happens, therapeutic selection or design down into an RNA sequence all happens as an entirely encapsulated computational process, and what comes out is synthesized RNA.
Abhi: Yeah.
Alex: I think if I had $100 million, I would want to try that.
Ben: Yeah. I’d want you to try that too.
Alex: But it wouldn’t inform the cancer vaccine stuff I’ve been talking about.
Ben: But the thing is, I don’t think we know enough about how to overcome suppression and evasion in the tumor microenvironment to do that in a personalized way, and I think we’re just going to bump up against the ceiling of cures until we can personalize those things just as much as we can personalize TCRs or vaccines.
Alex: Yeah. It’s not like I haven’t deeply considered it — it just sounds exciting. Can you take the founder-mode spirit and put it into such a streamlined, automated fashion that it’s all one process?
Abhi: In some sense, it feels like that is about to happen with the for-profit companies that are spinning up, like Valius. That’s going to be —
Alex: No.
Abhi: Is that not going to be a natural experiment they’re basically running? People who —
Alex: I mean — we’re in the Pathfinder camp, but Ed Larkin, the head of Valius, I like him a lot. I’ve talked to him a bunch, and I’ve also interacted with the Private Health people, who are the slightly less high-tech but still in the same space. And I don’t think anyone has some real secret edge there as far as how the analysis happens, and it’s very manual.
[02:55:58] Why identifying targets is the easy part
Ben: I think the key point to your $100 million dream, though, Alex, was not just identifying targets. Identifying targets is hard, but that’s really only the first step.
Alex: Yeah.
Ben: You have to be able to action the targets in a way that’s beneficial and non-toxic for the person. And throwing a bunch of genomics data through some computational processes to predict a set of possible targets — that’s doable now. But actually figuring out how to action all the targets in good ways — cracking that is, you know —
Alex: Yeah. I think what currently happens in this hyper-concierge context is that you facilitate access to a bunch of high-throughput profiling, you get all the data, you run a bunch of tools on it, someone’s doing that, and then they’ve got their Jupyter notebooks full of analyses, and they’re making figures, and a few weeks later they have a slide deck. That slide deck goes to the client, who then talks to their doctor, who’s skeptical, and then eventually you talk to the doctor, and then you make a new slide deck for the doctor. And eventually you’re nudging a decision toward, “Yeah, you should try that TRP2 ADC, because there’s a lot of TRP2 expression.” And it takes a few months, and maybe you get them access to something they could have gotten access to before. What I’m talking about is, why don’t we try the extreme form of that, in which all of that is being automated?
Abhi: So the patient plus the platform is all the agency that goes into managing the cancer. There is no institutional presence watching over you.
Alex: Yeah. You’re like a last line of defense against the cancer. You take someone who didn’t have a great option, and then you input tumor sequencing, or whatever profiling, and then try to push it through further than just “find a good target” — to “we know there’s this antibody for that target, and we can conjugate that to an anti-CD3 domain, and that could be a bispecific that’s a T-cell engager for the target on your tumor, and we know we have a shared way to encode that, we can make an mRNA encoding of it.” There have been a few trials of mRNA-encoded bispecifics, and so that’s going to be the therapeutic this box gives you. So you skip the “me making figures in Jupyter and then assembling the PowerPoint to convince a chain of people to ask for access to something.” You just say, “We find the targets, we make the therapeutic, here’s the therapeutic.”
Abhi: I kind of assumed that it at least has therapeutics.
Alex: It is not. We tried to put one inside Pathfinder, through this TCR-T thing, but making TCR-Ts is really slow. The version of this that is “data comes in, therapeutics come out” does not exist yet.
Abhi: Okay.
Ben: That’s a big thing.
And ideally, what you’d have at the end of the day is your box would give you, “Here are your seven options. Here’s everything I can know or predict about potential efficacy and toxicity profiles.” So you go and speak with your clinician, you have that goals-and-values, deep discussion and explanation back and forth, and then you make a plan. But you have it all there for you relatively easily, broken down by, “Here are three things you can get off-label. Here’s one clinical trial. Here are five things that the box can actually make for you, if you want that.”
Alex: Right. And one of those things is, the expression value is lower but it’s very tumor-specific, and this one’s HER2, which might have cardiac toxicity risk. So you’re right, there’s a slightly more complicated package to deliver, but the key thing that does not at all exist right now is an end-to-end process that goes from data to therapeutics.
Ben: Yeah. And personally, I would really love to practice that. I wish that box existed and it would drop in my clinic, so a patient would come, and that’s what my interaction would be.
Abhi: I feel like I could ask questions for another few hours. You guys are incredibly fascinating. But unfortunately, I do have to catch my flight out of North Carolina. Thank you so much for coming onto the podcast.
Ben: Thank you for coming.
Alex: Yeah, thank you for having us. This is awesome.
