This page provides about 1200 protein-ligand binding datasets within the BindingDB collection that are designed to be useful for parameterizing or testing protein-ligand modeling codes. Each set has roughly 10-50 congeneric small molecules with a range of affinities for a single protein target, and at least one complex in each series has a structure in the PDB, as a basis for modeling the rest of the series.

Each row in the following table corresponds to one dataset and provides the article(s) from which the data were drawn, a freely downloadable SDfile with the compounds and data, and a link to preview the compounds on-line. No effort has been made to establish appropriate protonation or tautomer states of the compounds. The 3D SDfiles on this page provide conformations for the free molecules generated with the program Vconf.

These validation sets were regenerated January 2020 to take advantage of the increases in data holdings of both BindingDB and the PDB. There are now more data sets than before, and it is unlikely that any present set exactly matches any prior set. All of the prior sets are still available for view and download as a zip file here.

The present validation sets were generated as follows. We first identified every PDB protein-ligand cocrystal structure whose protein and ligand (HET group) exactly match a measurement in BindingDB. Then, for each of these PDBs, we we recovered all the other compounds in BindingDB with binding data for the protein and having a maximum common substructure (MCSS) with the HET group that encompassed at least 0.6 of the non-hydrogen atoms in the HET group, and that also had a Tanimoto similarity of at least 0.6 with the HET group. If these filters left five or fewer ligands, the set was discarded. This procedure generated an initial collection of validation sets. This initial collection was then condensed by merging any two sets if they had the same protein and if either HET ligand was the same as one of the non-HET BindingDB ligands in the other. MCSS was computed with the OpenEye API, and Tanimoto similarity indices were computed with JChem fingerprints.