Structure-Based Drug Design

The field uses protein structures to predict how small molecules fit into binding pockets and to guide the search for compounds with useful binding properties. Approaches include molecular docking and physics-based scoring, equivariant neural networks, contrastive pocket–ligand retrieval, and diffusion, flow-matching, or autoregressive models that generate molecules in a pocket. Methods are evaluated on benchmarks such as CrossDocked and PDBbind, as well as virtual-screening enrichment tests and experimental studies of specific targets. A central challenge is that better predicted docking scores do not always yield accurate affinity, selectivity, stable geometry, or experimentally validated leads, particularly for flexible or shallow pockets.