Sergey Ovchinnikov

I'm excited to share that I'll be joining @MITBiology as an Asst Prof. in Jan 2024! Come join us! 🤓🧪🖥️🧬

@jvarga92 @compchemm I was thinking, you can get an "artificial" boost by using newer sequence databases for either MSA generation or training a new protein-language-model. Beyond simply the classic "did they train on newer structures in the PDB".

@sokrypton @compchemm It should be not the same MSA but the same data pipeline they were trained on, no? Af3's own data pipeline though runs longer, improved results or showed complementary results for us, both for peptides and nanobodies (or at least reduced the number of seeds necessary).

Open Models - beat AlphaFold3 "AlphaFold3 is no longer the best model" is a scope error — AF3's claim is joint generality across "proteins, nucleic acids, ligands, ions and modified residues".

@alphabreacher thanks for the feedback! I'm working on adding support to share resources and marketplace for trading.

@sokrypton 3/n it's addictive as hell, the thing to look at is that when playing 2v2, resources and troops aren't shared, which seems unreasonable, one can also try to breed the sheeps that are roaming around for more meat. p.s. i can't stop playing this

🍹Long weekend Project: Since Claude Fable is banned for Science, I thought it might be fun to see if it can be used for something less scientific. 😎 Introducing Age of Epochs! ⚔️ An attempted reproduction of Age of Empires II in Javascript. ageofepochs.com

@paularambles Still more than what I made a couple years ago at Harvard as an independent fellow. 🙃

$65k to help build the future of biology in san francisco the first experiment is whether you can survive on that

@josearkanos No science is allowed on fable.

@sokrypton No biologist can use fable?

@shevsea Time to change my research direction! 🤣

@sokrypton Sky is the limit

@shevsea Are you saying I might actually be able to publish Age of Epoch at PNAS? 😅

@sokrypton Complex battlefields favor strong soldiers over large armies in social animal warfare | PNAS share.google/fZy4SqHUpI4yuc…

@TensorTwerker It's live generated via JavaScript via procedural processes that adjusts depending on what is happening (music will change upon attack for example) 😅

@sokrypton loving the soothing music, playlist?

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 build ai to build ai to fight other ai building ai?🤠

Any feedback/requests are welcome! If you see something strange, save the game json and share over github for debugging! github.com/sokrypton/aoe Other features: - Experimental Multiplayer support via PeerJS - Mobile/tablet support

I hope everyone had a great 4th of July, and we've got a great seminar for you this week from @jamesproney! Join us this Wednesday, July 8th 2026 at 7pm EDT in Room 181, Building 68, @MIT "Energy-Based Models of Protein Structure, Dynamics, and Folding" bpdmc.org

Good morning ICML@🇰🇷， Finally to announce our first small step towards a real drug discovery engine (aka OpenDDE), an open source all-atom foundation model for biomolecular co-folding. Meanwhile, we are in the process of training and extending other functional components towards a complete drug discovery engine, that is why it is a preview version. GitHub: lnkd.in/gBgwvNMj Project page: lnkd.in/g7t8WW55 Structure prediction is the core of AI driven drug discovery. Or to say, a well-trained trunk-encoder+diffusion-decoder is the foundation of a complete drug discovery engine. So prior to a complete engine, we first build a good co-folding module. OpenDDE introduces atomic latent reasoning over biomolecular tokens, an extensible folding-centered foundation for drug discovery workflows, and scaling laws across data, model size, and test time compute. Across antibody antigen benchmarks, OpenDDE achieves 70.0% success on FoldBench-AB, 51.0% on PXMeter-AB, and 66.4% on our new 2026ARK-AB benchmark for low-homology AbAg interfaces. We are not competing with other baselines. To us, drug discovery is for human good, more than just being competitive landscapes. For Aureka, OpenDDE is part of a larger AI-native drug discovery infrastructure spanning foundation models, compute, molecular design, high throughput functional validation, closed loop optimization, and asset data room generation. OpenDDE is not a complete drug discovery engine yet. It has flaws, but it forms a foundation layer. By connecting biomolecular modeling with physical wet lab validation, we hope to make therapeutic discovery more scalable, reproducible, and accessible. Read more about Aureka and OpenDDE in lnkd.in/gF_NUWda #AI #DrugDiscovery #OpenSource #ICML

Come join us at the first New England Computational Biology Symposium Oct 1-2 at MSR New England! Keynotes will include @sokrypton @CarolineUhler Marc Vidal and G. V. Shivashanakar. Abstract deadline is August 14, early registration ends Sept 11.

@MartinPacesa @miangoar I wonder why they called it openDDE, unless they had some insider info of what isoDDE actually did, since methods are not described anywhere (?) 🤔

@miangoar I would not call it anything related to IsoDDE, just another AF3 clone

🧵1/4 A lab has just released OpenDDE-preview, the open-source (Apache 2.0) version of DeepMind's IsoDDE, introduced a few months ago and interpreted as an "AlphaFold 4"-like model. At least on paper, it looks like one of the strongest co-folding models available today.

We used FlashPPI2 to predict PPIs across hundreds of millions of proteins across 130K genomes on SeqHub, surfacing these interactions in the context of a protein search. A big thank you to Steinegger Lab, Milot Mirdita, @sacdallago, EBI et al. for making the high-confidence AlphaFold complexes openly available!

@RolandDunbrack @biorxivpreprint Nvm, once I click on any link I'm getting:

@RolandDunbrack @biorxivpreprint Works for me. I think you've been reading too many preprints, they think you are an AI agent 🤪

I can't get to @biorxivpreprint at all. Two browsers. Private window or not. Anyone else having this problem?

I gave a workshop on AlphaFold & related programs @ the University of Utah. Part 1 covers protein structure prediction before/after AlphaFold & deep learning neural networks. Part 2 covers the math behind the scores. Part 3 covers PyMOL & the webservers. @rolanddunbrack" target="_blank" rel="nofollow noopener">youtube.com/@rolanddunbrack

@DrLeucine @chrisbarber @JohnJumperSci @GabriCorso @MartinPacesa @ferruz_noelia @DdelAlamo @HannesStaerk @MoAlQuraishi @pranamanam @KevinKaichuang Well now that he is free from the golden handcuffs of deepmind.... 😎

@chrisbarber @JohnJumperSci @sokrypton @GabriCorso @MartinPacesa @ferruz_noelia @DdelAlamo @HannesStaerk @MoAlQuraishi @pranamanam @KevinKaichuang @JohnJumperSci posted twice in the last 2 years.

I asked a few people: "who are the best people to follow on X for AI + bio?" Protein, molecule & biologics design: @JohnJumperSci @sokrypton @GabriCorso @MartinPacesa @ferruz_noelia @DdelAlamo @HannesStaerk @MoAlQuraishi @pranamanam @KevinKaichuang AI / foundation models for bio & genomics: @BrianHie @owl_posting @exnx @btnaughton @anshulkundaje @AnimaAnandkumar @nmboffi @andrewwhite01 @JCoolScience Biosecurity & pandemic preparedness: @jtmonrad @JanikaSchmitt @ClaireQureshi @JassiPannuMD @T_Inglesby @JacobSwett @OliviaHelens @nanransohoff @incredutility AI security, policy & what to expect: @lucafrighetti @dylanscandinaro @logangraham @peterwildeford Builders, writers & investors: @NikoMcCarty @p_maverick_b @erika_alden_d @dr_alphalyrae Thanks to @arturszalata, @gottapatchemall and @JonasSandbrink for suggestions