CROWN: Curated Repository of Well-Resolved Noncovalent Interactions

The usefulness of structural data sets for machine-learning models of protein–ligand recognition depends not only on their size, but also on how structural imperfections are handled. Existing resources typically trade structural rigor for scale. Recent data sets such as SPINDR improve large-scale structural data through systematic preprocessing, but restrictive inclusion criteria still discard much of the chemical and biological complexity represented in the Protein Data Bank (PDB). Here, we present Curated Repository Of Well-resolved Noncovalent interactions (CROWN), a machine-learning–ready data set built around a repair-first curation philosophy: structural inconsistencies are corrected whenever this can be done without compromising experimental evidence at the binding interface, while filtering is reserved for cases that cannot be repaired conservatively. Starting directly from deposited PDB structures, CROWN resolves alternate conformers and connectivity errors, reconstructs complete interaction environments, repairs missing or nonstandard regions outside the binding pocket, assigns protonation states at physiological pH and applies constrained energy minimization using flat-bottomed restraints to relieve local strain while preserving experimentally supported geometry. This strategy yields 193,374 protein–ligand complexes spanning 81,606 PDB entries and 18,516 UniProt accessions, representing more than a 5-fold increase in both complex count and protein diversity relative to SPINDR while retaining a uniform structural representation. Rather than organizing the data set around sparsely available binding-affinity measurements, CROWN adopts a geometry-centric design in which the three-dimensional arrangement of atoms at the binding interface provides the primary source of information. On CASF-2016, CROWN-trained knowledge-based potentials achieved the highest scoring and ranking performance among five training resources, while matching PDBbind- and HiQBind-trained models in docking and screening and consistently outperforming SPINDR-trained counterparts. Together with multiaxis similarity annotations and a companion web interface, CROWN provides a scalable resource for scoring-function development, pose prediction, binder generation, ligand-conditioned protein design, and other structure-based machine-learning applications.