←
Jan Domanski907 posts
Jan Domanski avatar

Jan Domanski

@jandom_random

Forbes 30U30, ex-CTO @Labstep, ex @DEShawResearch, ex @charmtherapeutx. Currently building frontier open-source co-folding models at @open_fold

◷ Joined March 2016▧ 121 Photos and videos
Lambda avatar
⟳ Jan Domanski retweeted
Lambda@LambdaAPI·Nov 13

TurboPairFormer: open-source GPU kernels that deliver up to 1.73× faster OpenFold3 GPU computation per optimizer step than the original Triton backend, with more accurate gradients and reproducible results in fixed-input tests. Triangle attention scales cubically, so doubling the training crop makes it ~8× more expensive. Existing backends each trade away speed, precision, or reproducibility. Researchers from Lambda, Stevens Institute of Technology, and the OpenFold team built TurboPairFormer to address all three on NVIDIA Hopper GPUs: • Up to 1.73× faster than OpenFold3 Triton • Lowest mean Q/K and pair-bias gradient errors vs an FP64 reference • Bitwise-identical gradients across all four types (Q, K, value, pair-bias) in all 600 fixed-input cases Supports H100 and H200. On H200, 1.54× faster than OpenFold3 Triton on 768 pixel images (14.67 vs 22.53 sec/step). Read the blog for the full method and results lambda.com/blog/faster-more-…

Post attachment
Lucas Nivon 🇺🇸 avatar
⟳ Jan Domanski retweeted
Lucas Nivon 🇺🇸@lucas_nivon·Nov 12

We keep seeing the frontier of general purpose LLM moving forward -- and we were thrilled to see the @claudeai is using @open_fold for their AAV capsid work.

Post attachment

Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.

0:19
Jan Domanski avatar
Jan Domanski@jandom_random·Nov 6

Bare with me as I'm trying to learn how to use twitter/X... been absolutely overwhelmed with responses!

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

Hey if this goes super viral, maybe somebody will thrown in $5m and we’ll make you an open-source next-generation inhibitor 🚀

You're welcome sir. THANK YOU! More from Jan the man here:

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

All this takes time, time that patients usually don't have. We're yet to see an explosion of AI-enabled drugs – I helped make one such molecule, in trials now, so i'm pretty AI pilled but almost can't stand the hype around it. Show me the drugs! :D It's just the start...

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

And the model is just the start: it's just a prediction, not the truth. You end up validating everything in the lab, initially just in vial, then a petri dish, then an animal, and then a human. You have to design compound after compound, make sure it's safe and it works.

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

Most of the resistance mutants occur in the pocket where the drug binds but not all. Biology is complicated and it's not immediately clear how these "far away" mutations have an effect on the drug.

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

It's a steep climb: when Kevin and I started predicting his form of ALK, it wasn't giving us reliable results – these models generalize poorly outside what they've seen in training. Luckily, the "healthy" version of the ALK is reasonably well predicted, so we started with that.

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

We're not drug-makers ourselves, we're tool builders for the drug makers. Models like @open_fold can help us understand, at atomistic level of detail, how a drug and its target interact, and so we can understand how to fix/patch a drug when a mutation does happens.

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

Our mission at @open_fold and @openmsf is to build open-source tools (models, programs) to help bring these advances faster to patients. Fully open: the model checkpoints, the dataset, the training code. All of it.

Jan Domanski avatar
Jan Domanski@jandom_random·Nov 5

Huge thanks for the shout-out @kevinace , i'm blown away how much of the nuance you've managed to capture – making drugs is as complicated as making space ships, maybe more :-) It takes an absolute village! (And lots of time and money – too much if you ask me)

So get this...a stranger DM'd me on X and offered to help design my backup cancer drug. For free! His name is Jan Domanski @jandom_random. He runs science at @open_fold, an open-source AI model that predicts how proteins and drugs physically fit together. My cancer is driven by one broken protein called ALK. Lorlatinib (the daily pill I take) jams itself into it and shuts it down. It's working. But resistance isn't an if, it's a when. The tumor mutates the pocket and the drug stops fitting. So Jan started modeling it on his own time, burning his own tokens. He ran AI models against ALK and lorlatinib, mapping where the known escape mutations land. And what did he find? Most of them hit right in the drug's pocket, swapping small pieces for bigger ones that physically block the drug. That's the good kind of bad news. A pocket problem is something you can design around. But the normal route to a new drug is ~2 years and ~$4-5M just to be ready for human testing. Jan's dream is getting a medicine to one patient in ~1 month. I'm not taking a homemade molecule while lorlatinib works. But if it ever fails....shit...I've got a guy! I started posting actively on X just a few weeks ago and it has brought me together with the most amazing, generous people!

Post attachmentPost attachment
Jan Domanski avatar
Jan Domanski@jandom_random·Oct 31

It was fun to be 6 hours delayed with Amtrak… let’s call it a cultural experience! This is more of a useful speed to me! Just over 300km, same distance as NYC to Boston. For 60 EUR, in 2.5h 😹 someone is getting screwed hard!

Post attachment
Lucas Meyer 🎙️ avatar
⟳ Jan Domanski retweeted
Lucas Meyer 🎙️@meyer_lucas·Oct 30

Patrick Stewart in the final scene on Star Trek: The Next Generation. It’s extremely simple, but reveals so much of his character if you watched. Somehow his sense of duty and sensitivity, along with the admiration of his crew shines through. RIP Captain.

1:16
Michael Farris, Jr. avatar
⟳ Jan Domanski retweeted
Michael Farris, Jr.@mokepf7·Oct 29

Never forget Patrick Stewart delivering one of the most iconic Trek moments of all time. THERE ARE. FOUR. LIGHTS.

2:10
Lambda avatar
⟳ Jan Domanski retweeted
Lambda@LambdaAPI·Oct 27

We’re at the AI Drug Discovery Development Summit in Boston, Oct 27–29. Come find us at booth #8. On Oct 28 at 9:30 a.m. ET, @jianwen_xie and @jandom_random of @open_fold take the main stage to talk about open-source model training for drug discovery. First up, fine-tuning OpenFold3 and training it on synthetic data. Then a look at how Lambda’s infrastructure speeds up that fine-tuning, with a deep dive into TurboPairFormer, a fast, stable GPU kernel built for OpenFold3 training.

Post attachment
Jan Domanski avatar
Jan Domanski@jandom_random·Oct 20

Oh yes @open_fold !! Thanks for the mention @kevinace

Six days ago I replied to @bryan_johnson about my cancer vaccine. That reply put my story in front of a lot shit ton of new people. Since then: ~1M impressions, 5.5K profile visits, ~550 replies, and 60+ DM's and emails. The part I didn't expect: a lot of it wasn't well-wishes. It was real help. Jan Domański, a scientist at OpenFold, offered to model my cancer's resistance mutations with AI, for free. Dennis Watson at Travera offered immune testing to track how my vaccine is working. @miriamgonp, a fellow patient, is building her own vaccine path, and we're comparing notes. Thank you, Bryan, and thank you to everyone who reached out. I read every message. If you're someone who wants to help or better yet if I can help you in any way, my DMs are open.

Post attachment