Uncertainty-Aware Deep Learning Framework for Protein-Protein Interaction Inhibitor Prediction

Protein-protein interactions (PPIs) are attractive therapeutic targets but are often considered undruggable due to their large and flat interfaces. Since fewer than 0.02\% of the disease-associated PPIs in the human interactome have a known inhibitor, most targets therefore lie outside the training distribution, where models become overconfident and produce false-positive rates in large-scale screening. This study presents an uncertainty-aware framework for PPI inhibitor prediction in which proteins and compounds are represented by pretrained embeddings integrated with knowledge graph embeddings and physicochemical descriptors, and cross-attention blocks explicitly model protein-protein and drug-target relationships. Against nine baselines, the framework generalizes best to unseen PPI targets. Deep Ensembles quantify predictive uncertainty, substantially reducing the overconfidence. Cross-attention weights align with true interface residues, providing mechanistic evidence for predictions on characterized targets. By incorporating predictive uncertainty into ZINC library screening, we identified novel candidates for KRAS-SOS1 inhibition that docked as well as a known inhibitor.