While writing my long-form articles, I often stumble across really nice links that I’d like to share more broadly. I like how Construction Physics puts the explicit listing of such links behind a paywall, but keeps the actual synthesis of those links freely available. I’d like to follow the same model, feels very aligned-incentives-y.
I’ll separate them into ‘technical’ and ‘causal’ sections. Technical articles are more scientific in nature (usually biology or ML related), while casual articles are ones that require no domain knowledge to understand (usually non-science blog posts, interesting podcasts, or pop-sci videos).
Technical
BindCraft: one-shot design of functional protein binders
Comments: I’m admittedly a bit late to reading this paper (which came out a
few weeks ago), but finally went through it. The reported success rates here are extraordinary, everything from 10%-100% success rate in nanomolar-level one-shot binder design, which is several order of magnitudes higher than RFDiffusion. But the design process is also super complex, relying on a mixture of Alphafold2 noising + the soluble version of ProteinMPNN + some PyRosetta magic. I wonder where most of the magic is coming from….in the eyes of the creator, it seems to come from the reliance of Alphafold2 to ‘generate’ the amino acid sequences and not ProteinMPNN.
Live-cell imaging in the era of too many microscopes
Comments: Good overview of microscopy techniques! Working on a piece right now over one particular type of it, and this was a decent primer of the topic.
Comments: High level point is this: ML models are improving at propagating known functions but cannot reliably predict truly novel functions. This feels like a result not too dissimilar from the BELKA study…there is something fundamentally non-transferable (for now) about the learnings from one domain of biomolecules to another domain. Why is this? I’m not sure…
AlphaFold predictions of fold-switched conformations are driven by structure memorization
Comments: This is a similar case as previously: training set is everything. Context: people discovered in 2022/2023 that alternative conformations for proteins can be generated by Alphafold2 by messing with the MSA + brute forcing generation. This seems to be the case for Alphafold3 also. But, according to this paper, both of their conformation predictions are likely driven by memorization than actually learning the physics, meaning that O.O.D is a much more significant concern…
