Evaluating generative models for structure-based drug discovery from de novo design to lead optimization

Advances in protein structure prediction have created opportunities for structure-based drug design, yet translating structural knowledge into actionable small-molecule ligands remains difficult. Existing benchmarks often use simplified datasets and computational surrogates that poorly capture the heterogeneity and constraints of real-world drug discovery. We introduce MolGenBench, a comprehensive benchmark for de novo molecular generation and hit-to-lead optimization built on a large, structurally diverse dataset. Evaluation of 17 leading methods reveals persistent limitations, including weak virtual-screening performance, limited coverage of target-specific bioactive space, frequent high-risk structural motifs, poor target awareness, and reduced generalization to unseen proteins. Although some models improve three-dimensional conformation generation, they still struggle to produce active molecules reliably. MolGenBench provides standardized data, metrics, and analyses that connect structural prediction with practical molecular design and guide the development of more effective structure-based generative models.