A Graph-based Approach to Predicting Protein-Protein Interactions

Accurate identification of protein-protein interactions (PPIs) is fundamental for understanding cellular mechanisms and facilitating drug discovery. Although high-throughput experimental methods have expanded the known interactome, they remain resourceintensive and prone to noise. Consequently, computational approaches integrating functional annotations such as Gene Ontology (GO) have emerged as necessary complements. However, the current state-of-the-art methods often process these annotations as unstructured text sequences, neglecting the explicit topological structure of the underlying interaction network. In this study, we introduce a graph-based framework that leverages Graph Neural Networks (GNNs) to encode both the semantic attributes and structural connectivity of proteins. We evaluated this approach on standard fixed-split benchmarks and updated interactome datasets for Homo sapiens and Saccharomyces cerevisiae. Performance is evaluated across five independent random seeds per configuration. On the standard STRING v11.0 benchmark, our GATv2 model attains a mean Area Under the Receiver Operating Characteristic (AUROC) of 0.976 on Homo sapiens, while on Saccharomyces cerevisiae the GCN and GATv2 encoders reach statistically indistinguishable mean AUROCs of 0.971 and 0.968 respectively; these results closely match the TransformerGO baseline of 0.974 and 0.961 on the two organisms, respectively, while explicitly incorporating PPI network topology into the prediction framework. Differences are assessed through a two-stage statistical procedure (Shapiro-Wilk normality, ANOVA or Friedman omnibus, Bonferroni-corrected post-hoc). Beyond predictive performance, we provide an explainability analysis combining attention coefficients, GNNExplainer subgraphs, and Integrated Gradients attributions across twelve experimental configurations, showing that the framework’s predictions are traceable to specific edges, neighbours, and GO sub-vocabularies. These findings indicate that graph-based architectures provide a competitive and interpretable alternative for PPI prediction, while explicitly integrating functional annotations with interaction-network topology.