RL finetuning for protein binder design is being hailed as the next obvious step. But I ran the numbers, and something uncomfortable is happening under the hood. 🧵
introducing tubestack: bringing substack’s zen to youtube
a quiet, local desktop app to explore youtube without visual pollution. download here for free on mac & windows: https://t.co/Y336LBNnGM
(typography taken from the pretty blogs of @AnthropicAI, @OpenAI, @X, @Substack)
7/ Takeaway:
Main takeaway: reward gains are useful, but they shouldn’t be read alone.
For RL protein design, composition controls, off-target scores, specificity margins, and independent checks make the biological claim much clearer.
Full writeup:
https://t.co/iXohq4Koih
We set out to audit the RL results for IDiom, a protein language model for designing intrinsically disordered regions.
The question was simple:
Did RL learn specific localization, or did it learn sequences that score broadly well for related compartments?
The results 🧵
Protein design has been dominated by diffusions due to a "structure-first" perspective. What about intrinsically disordered proteins? We scale language-based design using the modern RL stack and our model IDiom.
Paper: https://t.co/mW0uMUBwZu
Try it: https://t.co/azcGCdqc4n
6/7 This is where things got interesting.
Stress-granule RL sequences scored highly for stress granule, but also scored highly for P-body.
So the target reward improved, but ProtGPS itself didn’t cleanly certify fine-grained specificity.
Exactly - biology is a fundamentally different domain than text and scaling laws do not apply cleanly
~All the tasks you want an LLM to do are contained in the text data itself. For biology, NONE of the tasks you want the model to do are contained in the sequence data itself.
"The Bitter Lesson has fully arrived in sequence biology and protein structure. Evo 2, AlphaFold 2 and 3, ProGen3, RFdiffusion".
This sentence has some issues IMO. 1/
Excited to share our new preprint:
“Computational design of membrane fusion proteins”
Huge thanks to all collaborators, co-authors, @KingLabIPD, and everyone at @UWproteindesign who contributed to this work.
Preprint:
https://t.co/5TDujafHJ9