Guiding Generative Protein Language Models with Reinforcement Learning

Abstract

Protein language models (pLMs) have demonstrated success at generating functional proteins across vast sequence spaces but lack the ability to design high-fitness variants on demand. Here, we iteratively guide pLMs toward user-defined objectives by applying reinforcement learning (RL). We demonstrate that RL can steer pLMs toward various protein properties, such as topologies or binding affinities, in a few iterations through long evolutionary trajectories. We apply our framework to the design of epidermal growth factor receptor (EGFR) binders, achieving a 26-fold increase in binding affinity in two iterations.