Evaluation of the Complementarity of Three Major Databases of Protein–Ligand Binding Affinity Data

Abstract

Protein–ligand binding affinity data underpin a variety of tasks at the early stage of drug discovery. Public databases including ChEMBL, BindingDB, and PDBbind curate such data with distinct focuses, yet their data overlap and complementary attributes remain insufficiently quantified. In this work, we systematically cross-mapped protein–ligand binding data in PDBbind (version 2024) against data records sourced from ChEMBL (version 35) and BindingDB (version 2025.08). Our analysis revealed that merely 22.1% and 23.9% of PDBbind’s annotated protein–ligand complexes possess matching binding data records in ChEMBL and BindingDB, respectively, confirming that the PDBbind dataset is not a subset of the other two larger repositories. This limited overlap arises from fundamental differences in scope: ChEMBL and BindingDB center on drug discovery datasets, whereas the data entries in PDBbind encompass a broader chemical and biological landscape. Our analysis further uncovered substantial data redundancy in ChEMBL and BindingDB, where 40%–50% of protein–ligand interaction pairs are associated with multiple, discrepant binding affinity values. Currently, no single resource achieves a comprehensive coverage of available protein–ligand binding data, mandating researchers to adopt resource-selection strategies customized to their goals. While ChEMBL and BindingDB host extensive bioactivity archives, PDBbind directly links three-dimensional complex structures with carefully curated binding affinity data. This unique value makes it best suited for structure-based studies.