One thing that helped with was implementing a simulated gameplay. Have Claude run 100s of games, get feedback and update ai.js. It is still doing some silly things... but getting better. Perhaps a future direction would be to let multiple LLM compete with eachother by building better ai.js?🤠 (3/4)

I wanted to understand how much information is lost when tokenizing protein structures using bio2token. Claude code whipped up this viz amazingly well

Is there a way to bulk download the MSAs from AlphaFold DB? cc @ebi.embl.org

I'm super excited to announce the first preprint of my PhD, together with Chenxi Ou and @sokrypton.org! ML has revolutionized protein modeling, but crucial challenges remain. For example, we can't reliably predict complicated protein structures without MSAs, which limits what we can design.

My first full contribution from my time in @opig.stats.ox.ac.uk is now out! Together with @fspoendlin.bsky.social (and with contributions from King Ifashe), we created FlAbDab and FTCRDab: two large-scale, open molecular dynamics datasets to study flexibility in immune receptors.

ANARCII is great github.com/oxpig/ANARCII - it installed and ran the first time I tried, and has a python API. Much easier than previous things. Thanks @opig.stats.ox.ac.uk !

GitHub - oxpig/ANARCII: A language model suite for numbering antigen receptor sequences.

A language model suite for numbering antigen receptor sequences. - oxpig/ANARCII

github.com

One cool thing about BoltzGen is the flexibility of the conditioning information you can provide (e.g. "constrain the structure of these residues and also these but let their relative placement vary"). I'm really curious to see what the community does with it! github.com/HannesStark/...

GitHub - HannesStark/boltzgen

Contribute to HannesStark/boltzgen development by creating an account on GitHub.

github.com

I’ve been testing BoltzGen a bit recently and while I haven’t done any experimental testing yet, the quality of the software is very clear. It installs, runs, logs everything, has tons of options. Very excited to test out the designs irl!

Hannes Stark@hannes-stark.bsky.social · 9mo ago

Excited to release BoltzGen which brings SOTA folding performance to binder design! The best part of this project is collaborating with a broad network of leading wetlabs that test BoltzGen at an unprecedented scale, showing success on many novel targets and pushing the model to its limits!

TCR-seq q: collaborator has ~10% of clones assigned to alleles IMGT says are "ORFs" like TRAJ44*01, TRBJ1-6*01 (this is 10x in mice). What can cause this? How likely are they productive TCRs? Maybe @jamieheather.bsky.social @victorgreiff.bsky.social @10xgenomics.bsky.social have ideas? 🙏

Newbie antibody design q: if I take a mouse IgG1 antibody and express its VH/VL on a human IgG1 constant framework, how likely is it to still bind the target? Does that kind of thing usually work?

I like how the deepseek r1 tech report includes an "ideas that didn't work" section (although it is admittedly very short). I wish more papers had that

Anyone receive the AF3 weights yet? (Applied on Mon, have not heard anything)

Missed this when it came out this summer - high doses of MHC II neoags in a peptide vaccine can be counterproductive due to a population of FOXP3− CD4+ regulatory cells that *kill* cDC1s. Big implications for peptide cancer vaccines if true www.nature.com/articles/s41...

Neoantigen-specific cytotoxic Tr1 CD4 T cells suppress cancer immunotherapy - Nature

Type 1 regulatory T cells (Tr1) represent a major obstacle that compromises naturally occurring and therapeutically induced tumour-specific immunity.

nature.com

This changed how I think about infectious disease - highly recommend www.annualreviews.org/content/jour...

I moved to Cambridge MA! If you're around and would like to meet up let me know. Looking forward to getting to know the community here

Say I have a deep learning model that runs on pairs of things M(a,b), and I want to distill it into three models X,Y,Z where Z(X(a), Y(b)) approximates M(a,b), and hopefully Z can be shallow and fast. Is that problem studied? What should I google?

Anyone know of a nice alphafold multimer batch inference script? I am doing an all against all run and would like to reuse MSAs etc. Running on an academic HPC cluster

Most memorable thing I've learned from #singlecellgenomicsday so far is from 2022... LIVE-SEQ does a "cytoplasmic biopsy" to profile RNA from a living cell without destroying it www.nature.com/articles/s41...

High quality benchmark of AlphaFold-Multimer and other tools on Ab/Ag complex structure prediction. Lots of room for improvement but the trope that AF can't do antibodies is not really right anymore after the late 2022 models update (v 2.3.0) www.biorxiv.org/content/10.1...

I love the simplicity of this idea - 1992 paper showing how to use FFTs to do rigid body molecular docking efficiently www.pnas.org/doi/abs/10.1...

Ok going to try using this. Who should I follow?