RLXF Feeds Wet-Lab Measurements Back into ESM-2 to Outdo Evolution on Protein Design
bravo_abad · x · 2026-09-13
Blalock and coauthors propose RLXF, using reinforcement learning to feed experimental measurements back into the protein language model ESM-2, steering generation away from "what evolution would choose" toward "what performs better in the lab":
- Case study: the light-sensing protein CreiLOV was never evolved to be bright. At position 43, ESM-2 prefers the natural cysteine, even though experiments showed swapping in alanine makes CreiLOV brighter.
- The mismatch between model preference and measured results is used as the RL signal to shift generation away from evolutionary bias.
- Notably, simply stacking the best individual mutations performed poorly, highlighting complex combination effects; the RL-guided sequence generation route matters more than finding a few good mutations.
More from Research
- AIR: Effective-Learning-Rate Scheduler Beats Cosine on 128B-Token MoE Pretraining — stochasticchasm · 2026-10-30
- CMU's Gym-Anything Auto-Builds 200+ Agent Environments from Any Software, Selected by GDP — wellecks · 2026-10-30
- NVIDIA researcher: train robot policies with privileged sim info, improving Pi 0.5 — wellecks · 2026-10-30
- Leveraging Simulator Privileged Information: Three Methods to Improve Robot Models — wellecks · 2026-10-30
- CMU Auto-Builds 25 Robot Simulations with Agents; Frontier Model Solves 0 of 25 Test Tasks — wellecks · 2026-10-30
- DeepMind's f_charton shares his take on the current state of AI and mathematics in Sorbonne talk — f_charton · 2026-10-30