PPISHES-an enhanced physicochemical approach for predicting protein interaction sites using graph neural networks

Accurate prediction of protein interaction sites is crucial for understanding biological processes. Many existing methods capture structural, evolutionary, and sequence features; however, they overlook important physicochemical properties, limiting their performance. We propose Prediction of protein interaction sites based on solvent accessible surface area, hydrogen-bonding propensity, and electrostatic potential sites, an enhanced model that incorporates the three physicochemical features to improve site prediction for both obligate and non-obligate complexes. Feature ablation and other analyses identified key features to improve model performance. The model achieved up to 42.8% and 29.3% improvements in Area Under the Precision-Recall Curve for Test_315 and Test_71, respectively. The model also outperforms current state-of-the-art methods across other key metrics, such as Recall, Area Under the Curve, and Matthews Correlation Coefficient.