sk

@billytcl @shae_mcl Should look for who is the depositors

@shae_mcl The unique thing about PDB that sets it apart from other bio datasets is that these structures are verified and can be thought of as ground truth. We are nowhere close to that for vast majority of omic types. We are still at the frontier of measuring things in bio.

It’s estimated that the Protein Data Bank (PDB) cost around $13B to create. Alphafold was only possible because of it. If we want ML to solve biology, we should be funding the creation of databases and the development of new assay technologies. ML is nothing without data.

repo now available on deepwiki: deepwiki.com/aqlaboratory/g…

@MolBioMike I have recently read a bit about the structure and interaction of rbx1 and its partners during & after the competition. I came to the conclusion & similar to a postdoc who left a comment on one of posts that targeting rbx1 as a monomer is the worst-case' biological scenario.

My methods and a post-mortem of the RBX1 competition. Enjoy! What 322 Designs Taught Us About Binding RBX1 open.substack.com/pub/mikeminson…

@TensorTwerker Interesting I use protenix msa

pip install msa-nim

@MartinPacesa @miangoar

@MartinPacesa @miangoar

AI can now design antibodies that bind with atomic precision, but not ones that cells can produce. Our preprint closes this gap, delivering a structural principle, an AI-guided rescue pipeline, and adalimumab variants with 20-100x in vivo potency. biorxiv.org/content/10.648…

Every time I tell AI utopianists that biology is too complex for AI to "solve", they cite the success of AlphaFold. No, AlphaFold did not "solve" protein folding. It gets broad structures correct ~70-88% of the time (depending on evaluation), enabling useful but flawed statistical guesses. True "solving" would require ~99.9%+ accuracy, practically zero meaningful edge cases, and high confidence across fine details like side chains and conformations. Even then, this is just one narrow slice of the complexities of proteomics. The persistent gap between the "AlphaFold solved protein folding" claim and reality is a perfect example of AI overhype in biology.

@TensorTwerker Before you proceed

writing some nextflow pipelines and the inference is crazy fast

@outsource_ This guy Qwen3.6-27B-UD-Q4_K_XL(1.13GB) can generate high quality draft than Qwen3-1.7B-UD-Q4_K_XL(1.34GB) with just 210MB extra size will test tomorrow and share results. thank you once again and there are smart people than me.

@outsource_ Conclusion: Dynamic Quants from Unsloth Better I assume I can get easily 180-200t/s with the dynamic quants. These two guys Qwen3-1.7B-UD-Q4_K_XL(1.13GB), Qwen3.6-27B-UD-Q4_K_XL with Qwen3.6-27B-UD-Q4_K_XL in ik_llama.cpp

My 4090 went from 26 -> 154 tok/s Qwen 3.6 27B🤯 Same GPU. Same Q4_K_M . No FP8, no extra quant. The unlock: ik_llama.cpp + speculative decoding using Qwen3-1.7B as the draft model. 85% acceptance rate. Full config + benchmarks 👇🏻

@outsource_ OMG had 91.5% acceptance

@design_proteins @seqdesign Keep the lengths something similar to bhardwaj rfpeptide and try some macrocycles.

@seqdesign Here's our highest ipSAE design from Protein Hunter: ELAKEAVENKDEKLMDEAISVAFTDKEKFL binds right at the MYC-MAX interface it would be great to test this out in the lab and see if it actually binds and inhibits dimerization and downstream signalling

let's design a peptide to inhibit MYC-MAX with AI 🧵 MYC is the most-cited oncogene in human cancer and is deregulated in ~70% of tumors Forty years of small-molecule chemistry has produced exactly zero approved direct MYC inhibitors

And yes—our 10 designs? They're now shipping for expression & BLI affinity testing. 🚀

Despite the silence from one corner, the response elsewhere has been incredible. Several companies + academic groups reached out🤝

If I had $10M–$1B? I'd go all-in on one modality. One of these; minibinders, cyclic peptides, nanobodies Pick one. Master it. Skip the "let's do everything" trap. Our designs were quietly excluded from the recent RBX1 competition—no feedback, no documented reason nothing! 👇

@skewbed @iotcoi That's 3.5 not 3.6 here: huggingface.co/z-lab/models

@iotcoi Did you train your own DFlash model? I don't see one for Qwen3.6-27B here: #supported-models" target="_blank" rel="nofollow noopener">github.com/z-lab/dflash#s…

Qwen3.6-27B-FP8 + Dflash + DDTree, 256k context, 10 agents ~200 tokens/sec max decode 136t/s average on a single tiny GB10 GPU at 49W power

@ChrisHayduk Cool! anything on speed comparison across different protein lengths could be very useful

A couple of months ago, I announced that I was partway through implementing a simple, readable AlphaFold2 in pure PyTorch, inspired by @karpathy's minGPT. Today, I'm happy to share minAlphaFold2 - the completion of that project. Repo link: github.com/ChrisHayduk/mi…