Evolutionary multi-objective molecule optimization in implicit chemical space

Abstract

Abstract Optimization techniques play a pivotal role in advancing molecular optimization, prompting the development of numerous generative methods tailored to efficiently design optimized molecules derived from existing lead compounds. However, these methodologies often encounter difficulties in generating diverse, novel, and high-quality molecules when addressing multi-property tasks. Consequently, efficiently searching for diverse optimized candidates that simultaneously satisfy multiple properties remains a significant challenge in molecule optimization. To address this problem, we propose a multi-objective molecule optimization framework (MOMO). MOMO employs a specially designed Pareto dominance-based multi-property evaluation strategy at the molecular sequence level, specifically designed to guide the evolutionary search in a latent molecular space to optimize multiple molecular properties. A comparative analysis of MOMO with extant state-of-the-art baselines across three multi-property molecule optimization tasks reveals that MOMO markedly outperforms them all. These results suggest the efficacy of the proposed MOMO framework for simultaneous optimization of multiple properties in molecule optimization.