MGPPI: multiscale graph neural networks for explainable protein-protein interaction prediction

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in structural biology

Background:

• Protein-protein interactions (PPIs) are crucial for biological processes, impacting cancer diagnosis and drug development.
• Existing computational PPI prediction methods struggle with extracting comprehensive structural information and lack interpretability.
• There is a need for accurate, interpretable computational models for PPI prediction to advance biomedical applications.

Purpose of the Study:

• To introduce MGPPI, a Multiscale Graph Convolutional Neural Network, for enhanced protein-protein interaction prediction.
• To improve the extraction of local and global protein structural information using a multiscale module and multiple convolutional layers.
• To enhance model interpretability by introducing Gradient Weighted interaction Activation Mapping (Grad-WAM) for identifying key binding residue sites.

Main Methods:

• Developed MGPPI, a Multiscale Graph Convolutional Neural Network model incorporating a multiscale module and multiple convolutional layers.
• Implemented Grad-WAM, a novel visual explanation method, to highlight critical residue sites involved in protein interactions.
• Evaluated MGPPI performance against state-of-the-art methods on diverse datasets, including multi-species data and cancer patient survival datasets.

Main Results:

• MGPPI significantly outperformed existing methods in PPI prediction, demonstrating strong generalization capabilities across species.
• The Grad-WAM method successfully identified key binding sites for SARS-COV-2 spike protein and human ACE2 receptor.
• Grad-WAM-identified residues showed potential as biomarkers for predicting patient survival in various cancer types.

Conclusions:

• MGPPI offers a highly accurate and interpretable approach to PPI prediction, overcoming limitations of previous methods.
• The model aids in identifying potential drug targets and guiding personalized cancer therapies.
• Grad-WAM provides valuable insights into protein binding mechanisms and clinical relevance, supporting biomarker discovery.