Associate Professor at @FredHutch
Studying cytokines, (auto)antibodies, and cancer immunotherapy
Founder: @SimchaTx, @Seranova, @StippleBio, and @AriaxBio
T cells do the heavy lifting when it comes to the anti-tumor action of checkpoint immunotherapy. But do antibodies play a role too?
Thatโs the question @yile_dai looked to answer in his thesis work, out today in @Nature! ๐งตbelow
https://t.co/VgTByRKRCX
@DdelAlamo And of course opus and fable refused this immediately. Astra happily tried, but blew up its own repo to 30,000+ LoC with all sorts of nonsense complexity.
@DdelAlamo Fine tuning soluble MPNN right now, so yes. And I've never trained a model before, just wanted to see if Neon would tackle a hobby horse of mine for a while.
The compute credits in the Gemini Ultra sub are insane. The $100 plan gets you 1000 colab credits/month. That's ~55 H100 or 100 RTX6000 Pro (G4) hours.
The colab CLI let your agent to use these credits like Modal... no notebooks required!
Come for the Neon, stay for the GPUs.
@iskander The agents seem to have a pretty good understanding of how to use the Google Colab CLI. After a few minutes of tinkering, Neon had it figured out and created a skill file so that we could essentially use the compute credits like we do for Modal. No human thought required.
This version (v.1.0.8) also adds BindCraft2-style speedups using cross-trajectory AF2 validation caching.
BC2 is fantastic, but for some targets/epitopes, OG BindCraft continues to give compelling designs. For miniproteins, it's worth trying both.
TMol is awesome. I updated FreeBindCraft to add it as a 3rd scoring engine besides PyRosetta and classic freebindcraft (openMM-relax).
TMol restores rosetta fastrelax and missing interface metrics. It's also faster than both PyRosetta and OpenMM.
https://t.co/L4jH7WnmFP
I promised more on TMol in the RFD4-Proteina thread. Now with the paper on arxiv here we go! ๐งต
Rosetta-style all-atom energies, GPU batching, and gradients in PyTorch. A way to put molecular energetics directly into your model, combining deep learning with physics. 1/n
@rodydavis@hqmank Auto approvals could use work, still way too many approval requests imo. Little things like displaying usage vs limits without having to click through settings. Easy controls for opting out of data use.
@AjasjaLjubetic@KevinKaichuang You make it sound like getting a (Py)Rosetta license is easy. Takes months of dealing with a university tech transfer office even if you can afford the fees. Rosetta Commons does not make this process friendly at any level.