Learning to Place Structured Waters: Crystallographic Evidence for Solvent-Aware Drug Design
Abstract
Water molecules often complete protein--ligand interactions, but their positions are missing or inconsistently retained in structure-based drug design. We ask whether waters placed from a solvent-free protein structure recover crystallographic evidence that was never used for placement. ProSEEK-Water ranks and places high-confidence structured hydration sites. On a blind coordinate benchmark of 38 protein chains and 414 targets, it reaches 0.643 macro recall at 2K/1Å, compared with 0.370 for HydraProt and 0.118 for GalaxyWater-CNN under matched inputs and evaluation. A separate crystallographic cohort contains 33 usable cases from a 40-chain candidate manifest. Across these cases, 82.2\% of placed waters coincide with strong omit density, compared with 4.5\% for geometrically plausible random placements. The result establishes a validation route from water placement to reflection-space evidence and, ultimately, solvent-aware drug-design hypotheses.