Datasets, checkpoints and container for the manuscript "Multimodal Protein Foundation Models Reveal Dataset-Driven Separability Regimes in Allosteric Site Prediction"

Filename Description allosteric_pockets_kinsite.h5 HDF5 file containing reconstructed allosteric binding pocket structures for the KinSite dataset. Each pocket comprises the 100 residues closest to the annotated allosteric ligand and is formatted for OneProt pocket embedding extraction. competitive_pockets_kinsite.h5 HDF5 file containing reconstructed orthosteric (ATP-competitive) binding pocket structures for the KinSite dataset. Pockets were generated using the same protocol as the corresponding allosteric pockets and are compatible with the OneProt framework. KinSite_splits.zip Train, validation, and test splits for the KinSite dataset generated using MMseqs2 clustering at 30% sequence identity to prevent homologous proteins from appearing across data partitions. PPI-site_splits.zip Train, validation, and test splits for the PPI-Site dataset generated using MMseqs2 clustering at 30% sequence identity and used throughout all downstream experiments. AlloDiverse_splits.zip Train, validation, and test splits for the AlloDiverse benchmark dataset (ASD + PPI-Site + KinSite) used in the multimodal allosteric site prediction experiments. ASD_original_split.zip Original train, validation, and test partitions of the Allosteric Site Database (ASD) used during dataset preparation and construction of the AlloDiverse benchmark. PPI-site_KinSite_original.zip Original processed PPI-Site and KinSite datasets prior to MMseqs2 sequence clustering and train/validation/test splitting. These files are provided to facilitate reconstruction of the benchmark datasets and preprocessing pipeline. binding_pockets_allosteric.h5 HDF5 file containing reconstructed allosteric binding pocket structures derived from the PPI-Site dataset. Pockets were generated by selecting the 100 residues closest to the annotated ligand centroid. binding_pockets_orthosteric_competitive.h5 HDF5 file containing reconstructed orthosteric (competitive) binding pocket structures derived from the PPI-Site dataset using the same fixed-size pocket construction procedure as for allosteric sites. ASD_binding_pockets.h5 HDF5 file containing reconstructed allosteric binding pockets generated from residue-level annotations in the Allosteric Site Database (ASD). Residues were clustered using DBSCAN in three-dimensional space, and fixed-size pockets were constructed from the 100 residues nearest to each cluster centroid. oneprot_container.sif Apptainer/Singularity container image providing the complete software environment required to reproduce the computational workflow described in the manuscript, including preprocessing, multimodal embedding extraction, model evaluation, and analysis scripts. checkpoints.zip Archive containing all pretrained OneProt downstream model checkpoints used