Grid-Based Evolutionary Algorithm for Multi-Objective Molecule Generation Enhanced by Reinforcement Learning

Meijia Song, Lian Shen, Ziyun Zou, Yinghui Jiang, Jia heming, Juan Liu, Xiangrong Liu

Keywords: molecular generation, evolutionary algorithm, reinforcement learning, multi-objective optimization

Abstract: Fragment-based drug discovery (FBDD) is limited by the need to construct and maintain static fragment libraries. To overcome these challenges, we propose a novel evolutionary framework. Our method starts with sample molecules that are fragmented and fed into a policy-decoupled architecture. This architecture utilizes reinforcement-learning-guided crossover and mutation operators to recombine and modify fragments. This approach expands the latent fragment space without relying on predefined libraries. By employing a grid-based fragment-masked crossover, our method enables combinatorial explorations and extends beyond conventional fragmentation patterns. In comparative experiments, our method outperforms recent state-of-the-art methods on most PMO benchmarks and target-protein docking tasks. Additionally, it achieves a low average synthetic accessibility (SA) score and maintains a structural novelty rate above 90%.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 6162