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

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Drug Discovery

Background:

• Protein-protein interactions (PPIs) are crucial for biological processes.
• Experimental identification of PPI sites is resource-intensive.
• Current computational methods struggle with 3D structure, limited data, and performance variance.

Purpose of the Study:

• To develop an advanced computational method for accurate protein-protein interaction site (PPIS) prediction.
• To address limitations of existing methods, including inadequate structural feature modeling and overfitting.
• To improve the robustness and stability of PPIS prediction.

Main Methods:

• Introduced GeoPPIS, a geometry-aware approach utilizing a geometric graph learning module.
• Employed a two-stage hybrid transfer learning strategy to mitigate overfitting.
• Implemented density-based clustering and ensemble learning for post-prediction refinement.

Main Results:

• GeoPPIS achieved state-of-the-art performance on two benchmark datasets (>370 proteins).
• Achieved high prediction accuracy (0.857 and 0.860) compared to 11 baseline models.
• Demonstrated superior modeling of 3D structural features and reduced prediction variance.

Conclusions:

• GeoPPIS offers a reliable in silico tool for robust protein-protein interaction site prediction.
• Incorporating geometric information enhances prediction accuracy and stability.
• The method facilitates advancements in precision medicine and drug development.