Structure-Guided Reinforcement Learning for High-Affinity Antibody Design

Antibodies enable precise targeted therapies for immune-related diseases through specific recognition and binding to pathogens or inflammatory factors. While recent advances in deep learning have enabled structure-based antibody design with promising accuracy, existing generative models primarily mimic natural antibody distributions without explicit optimization toward binding affinity. To address this gap, we propose EvoAb, a reinforcement learning framework that combines a pretrained antibody co-design model with thermodynamics-grounded reward modeling for high-affinity antibody design. We introduce Multi-round Mutation Policy Optimization (M2PO), an iterative algorithm that integrates affinity-weighted sequence learning with distribution anchoring, progressively enhancing binding affinity while preserving structural plausibility. By leveraging structure-aware reward signals, EvoAb enables efficient \textit{in silico} directed evolution without expensive physics-based calculations or wet-lab experiments. Extensive experiments demonstrate that EvoAb achieves state-of-the-art binding affinity optimization, reducing mean from 2.89 to 2.45 kcal/mol and increasing stabilizing mutation rates from 13% to 19%. Cross-validation with FoldX confirms that our optimization yields physically meaningful improvements, highlighting EvoAb's potential for accelerating therapeutic antibody development.