Geometry-Aware Protein–Protein Binding Site Prediction Using Geometric Learning and Pretraining Strategies

Accurate prediction of protein-protein interaction sites is of great importance for understanding complicated biological processes and developing targeted drugs. However, conventional experimental measurements for the identification of binding sites are time-consuming and labor intensive, while the current computational solutions have their drawbacks including inadequate modeling of three-dimensional structural features, overfitting susceptibility due to limited training data, and high variance in model performance. Motivated by these issues, we propose <b>GeoPPIS</b>, an advanced <b>GEO</b>metry-aware approach that achieves state-of-the-art performance in the <b>P</b>rotein-<b>P</b>rotein <b>I</b>nteraction <b>S</b>ite prediction task. Particularly, a geometric graph learning module is introduced to capture the three-dimensional structural characteristics of proteins, which contributes to a more comprehensive protein representation compared to conventional sequence or graph-based approaches. Besides, a two-stage hybrid transfer learning strategy is devised to alleviate the problem of overfitting to limited data. In addition, we also introduced two postprediction strategies including a density-based clustering module and an ensemble learning module to reduce high variance and achieve stable prediction. Extensive evaluations on two benchmark data sets that involve more than 370 proteins and 78,000 sites show that GeoPPIS achieved state-of-the-art performance among 11 baseline models with accuracy of 0.857 and 0.860. This research provides a feasible way of incorporating the geometric information on protein, offering a reliable in silico tool for robust PPIS prediction, ultimately facilitating advancements in precision medicine and drug developments.