Pocket and ligand embeddings for 72,199 protein–ligand complexes from the Protein Data Bank

This work presents a retrieval resource covering 72,199 protein–ligand complexes from the Protein Data Bank. Each complex carries a 1,280-dimensional pocket embedding, obtained by mean-pooling final-layer ESM-2 representations over binding-site residues within 8 Å of the ligand and computed per chain rather than over a concatenation of all polymer chains; a 2,048-bit RDKit Morgan fingerprint of the ligand, taken from the idealized Chemical Component Dictionary entry and therefore shared by complexes binding the same component (26,619 distinct fingerprints across the 72,199 rows); and HDF5 coordinates, per-residue heavy-atom centroids for the pocket and per-atom for the ligand, with residue and element identifiers. A residue-to-token mapping makes each pooled embedding reproducible from the released files. FAISS indices are provided over the two modalities separately, because they answer different questions: pocket-only retrieval recovers a same-UniProt complex at rank 1 for 81.2% of queries, whereas ligand-only retrieval is the stronger signal for chemotype search.