PertiNet: Interaction-Aware Multimodal Modeling of Protein–Protein Interactions

Protein–protein interaction (PPI) prediction requires models that combine sequence, structure and functional evidence for proteins excluded from training. We developed PertiNet, which represents these inputs as separate modality tokens, models their interactions and assigns each token a protein-pair-specific weight. On the fixed protein-disjoint Bernett benchmark, PertiNet using sequence and structure achieved AUROC, AUPRC, balanced accuracy and MCC values of 0.71, 0.72, 0.66 and 0.33, respectively. Adding Gene Ontology (GO) annotations increased these values to 0.72, 0.73, 0.67 and 0.35, compared with published TUnA values of 0.70, 0.69, 0.65, and 0.30. Within-study Bernett controls showed that complete fusion improved on concat–MLP with identical sequence and structural inputs, using either the original sequence representation or ESM-2. Matched RBP400 ablations examined the contributions of modalities and fusion components, and Dset_186_72_PDB164 provided a complementary evaluation of residue-level interface localization. These results support integrating structural and functional context with sequence representations through interactions among modality tokens.