I do my best to make protein models Which error matters depends on context: in a cell, in a process, or averaged over no context at all. and Composable, documented, and runnable by someone who did not write it..
Composable JAX tooling for protein modelling, tested at the bench. Microbiology PhD candidate at MIT, Ovchinnikov Lab.
A functional JAX/Equinox reimplementation of LigandMPNN. Reproduces the PyTorch reference to ≥0.999 Pearson across all five decoding paths and runs 8–61× faster on a single structure, by trading eager dispatch for jit, vmap and scan. A sample/score API with no model objects to wire up.
Sequential Monte Carlo over protein sequence space in JAX. Fitness functions compose arbitrarily, so ProteinMPNN likelihood, ESM likelihood and codon adaptation index can be optimised together — selecting for folding and expression at the same time rather than in sequence.
Deciding which computation or experiment to run next, under a budget. Once you can run a thousand experiments, the scarce resource stops being execution and becomes the choice.
Most people who model proteins have not run the assay. Most people running the assay do not write the kernels. Two things I have found there:
If you work on the computational side
A There are two hard problems in computer science: cache invalidation, naming things, and off-by-one errors. as written is never the protocol. Every wet-lab method carries context that did not survive being written down, and the idiosyncrasies are where your assumptions die. Automation's real value is not throughput — it collapses those idiosyncrasies into explicit settings, and lets you record far more of what actually happened.
If you work at the bench
A library is only as good as its design. When the computational side does not understand the assay's constraints, its dynamic range, or what would actually be useful, it produces tools for nobody in particular.
Part of the alpha is that we actually use the tools rather than just implementing the math — so we see where the problems crop up and where the bottlenecks are.
Relax selection and a population spreads across the landscape; tighten it and the population collapses onto whatever it was covering. That schedule is the same object in this browser, in proteinsmc, and in a continuous evolution experiment.
step 0 · β = 0.00 · mean ESS 600 · mean entropy 2.98 · 0% resolved (uniform, no information yet)
Schematic: the landscape here is synthetic, not a protein. A version running against real per-position likelihoods from aminx is in progress.
Allostery, coevolution, quiescence entry, phage genomics.
Leftfield first. Also techno and house.