MoDeMol: A Decomposition-Based Evolutionary Framework for Multi-Objective Molecular Optimization

Lead compound multi-objective optimization requires balancing mutually restrictive drug-like properties, yet conventional latent generative and evolutionary optimization methods suffer from structural-function misalignment, population structural redundancy and low search efficiency under limited evaluation budgets. This work proposes a decomposition-based evolutionary framework built on a score-informed latent space (MoDeMol), with four core innovations: (1) Contrastive fine-tuned Transformer-VAE latent space supervised by predicted dual-target activity scores to improve the alignment between latent spatial proximity and predicted target-activity profiles. (2) Hybrid population initialization combining lead latent perturbation and high-quality molecular bank screening to balance local scaffold retention and global high predicted-activity region exploration. (3) Adaptive resource allocation mechanism for dynamic subproblem computation scheduling to boost optimization efficiency when the property evaluation budget is limited. (4) Dynamic Morgan fingerprint similarity penalization to suppress structural redundancy and maintain diverse molecular scaffolds in the population. Extensive benchmarks on single-target GSK3β and dual-target GSK3β/JNK3 kinase optimization tasks demonstrate that MoDeMol outperforms MolFinder, JANUS and MOMO in success rate, Pareto hypervolume and qualified molecular yield. Quantitative latent-space analyses, qualitative visualization, scaffold tracking, and ablation experiments provide complementary evidence for the contributions of the proposed components. The proposed framework addresses representation alignment, structural redundancy, and search-resource allocation within a computational multi-target lead-optimization workflow under limited evaluation budgets.