pretty_name: Provisional ligand–pocket collection
Provisional — do not use
This dataset is provisional and should not be used by anyone at this point. Licensing for one of the source datasets remains unresolved. No dataset license is assigned by this repository. Included source-license notes are provenance, not a grant of rights to the combined collection.
This is the final canonical heavy-atom collection and the exact packed runtime
cache used by FSBFPocket. Preparation/training/sampling code and definitions are
in https://github.com/MurrellGroup/FSBFPocket . The corrected annotation schema is
pli-interaction-evidence-v2-all-scopes: ligand–ligand, ligand–protein, and
protein–protein interactions, symmetric order-invariant ring-plane tests, and
explicit uncertainty. These are geometric annotations, not measured bonds.
The collection contains 1,240,225 records: 597,749 GEOM-Drugs conformers, 500,000 ProFSADB fragments, 94,381 BindingNet docked examples and 48,095 PLINDER examples including peptides. The manifest distinguishes origins and ligand classes. Seven confirmed PLINDER defects are retained for provenance but excluded by the training admission policy in the code.
Use the GitHub download command to verify and extract the artifacts. The full portable collection includes provenance, original structures, primary heavy-atom arrays and rich annotation evidence. The packed cache contains all inputs and split assignments needed for training; original source downloads are not required to train. Older annotation files inside the portable archive are provenance only: the reader and training use the corrected v2 annotations.