Deng, S.; Wan, X.; Mu, Z.; HUANG, S.-Y.; Yan, C.

Accurate prediction of protein-protein interactions (PPIs) is essential for understanding cellular mechanisms and disease. However, existing deep learning models often fail to generalize to novel proteins with low sequence similarity to training data. Here, we present a mechanism-aware deep learning framework to enhance generalization by incorporating residue-level interaction modeling. Our model integrates a dual-attention-based PPI prediction module with an inter-protein contact predictor based on protein language model-embedded geometric graphs. Joint training on structure-informed PPIs from the Protein Data Bank (PDB) enables the model to predict both interactions and the key residue pairs that mediate them. The model generalizes well to unseen proteins across species in the high-quality interactome dataset (HINT). Fine-tuning on Homo sapiens PPI data from HINT further improves the performance. Compared to D-SCRIPT, Topsy-Turvy, and TT3D, our model, referred to as PLMDA-PPI, achieves superior accuracy and robustness under strict sequence dissimilarity settings. This work provides a generalizable framework for biomolecular interaction prediction through the integration of structural and non-structural interaction data.