Generative Agents for the Continuous Evolution of Target-Binding Proteins

ABSTRACT

Directed evolution in the laboratory is constrained by physical limits on library size and mutation rates. We present an entirely *in silico* framework utilizing Generative Agents that simulate the continuous evolution of target-binding proteins. Using a physics-informed reinforcement learning environment, these agents introduce mutations, evaluate binding affinities using surrogate oracles, and actively select for stability and specificity across thousands of simulated generations per hour. This agent-driven evolutionary algorithm discovers highly potent binders that bypass the developmental biases favoring simple structures, uncovering complex molecular assemblies previously inaccessible to standard generative models.