We apply RLXF across five diverse protein classes to demonstrate its generalizability and effectiveness at generating optimized sequences by learning functional constraints beyond those captured during pre-training

Experimental validation reveals the RLXF-aligned model generates a higher fraction of functional sequences, a greater number of sequences more fluorescent than CreiLOV, and the brightest oxygen-independent fluorescent protein variant reported to date

We align ESM-2 to experimental fluorescence data from the CreiLOV flavin-binding fluorescent protein. The aligned model learns to prioritize mutations that enhance fluorescence, many of which are missed by the base model

RLXF follows a two-phase strategy inspired by RLHF. Supervised Fine-Tuning initializes the model in the right region of sequence space. Proximal Policy Optimization directly aligns sequence generation with feedback from a reward function like a sequence-function predictor

We are excited in the @philromero.bsky.social lab to share our new preprint introducing RLXF for the functional alignment of protein language models (pLMs) with experimentally derived notions of biomolecular function!

Favorite foods! Tandoori chicken and chili momo's: everestkitchen.ca. Onigiri! www.onigiriya.ca. Pho: www.viethouserestaurant.com.