Graph Neural Networks with Attention Mechanisms for Predicting Protein-Protein Interactions in Drug Discovery

Predicting protein-protein interactions (PPIs) is fundamental to understanding cellular processes, elucidating disease mechanisms, and accelerating rational drug discovery. Although high-throughput experimental techniques have mapped extensive interactomes, they remain constrained by high false-positive rates, substantial financial costs, and labor-intensive workflows. In response, computational approaches, particularly deep learning on molecular graphs, have emerged as viable alternatives. In this study, we propose a Hierarchical Relational Graph Attention Network (HR-GAT) tailored for high-accuracy PPI prediction in virtual drug screening pipelines. Our architecture combines residue-level structural graphs with pre-trained protein language model representations, deploying multi-head self-attention to capture both local physicochemical interfaces and long-range topological constraints. Evaluated across standard benchmark datasets (including STRING and BioGRID) alongside a rigorously non-redundant cross-species dataset, HR-GAT demonstrates superior predictive performance, achieving an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.948 and an Area Under the Precision-Recall Curve (AUC-PR) of 0.932, outperforming traditional sequence-based models and conventional graph convolutional baselines. Furthermore, interpretability analyses reveal that attention weights correlate significantly with known biochemical binding motifs and catalytic hotspots, notably within oncogenic targets such as the MDM2-p53 and KRAS interactomes. These results establish our attention-guided graph framework as an accurate, interpretable predictive engine for biomedical data analytics and target-based therapeutic discovery.