Intra-Inter Graph Representation Learning for Protein-Protein Binding Sites Prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in structural biology

Background:

• Graph neural networks (GNNs) are increasingly used for analyzing complex, irregular data like proteins.
• Representing proteins as graphs is a natural approach for computational analysis.
• Existing methods for protein-protein binding site prediction using GNNs often fail to capture comprehensive interaction information.

Purpose of the Study:

• To develop an advanced graph representation learning framework for accurate protein-protein binding site prediction.
• To improve upon existing methods by better integrating information within and between protein structures.
• To introduce the Intra-Inter Graph Representation Learning (IIGRL) model.

Main Methods:

• Employed graph neural networks for modeling protein structures.
• Implemented intra-graph learning by maximizing mutual information between local node representations and global graph summaries.
• Developed inter-graph learning by fusing ligand and receptor protein graphs to capture inter-protein residue affinities.

Main Results:

• The IIGRL model demonstrated superior performance compared to state-of-the-art methods on multiple benchmark datasets.
• The intra-graph learning component effectively encoded global protein information into node representations.
• The inter-graph learning component successfully captured crucial inter-protein interactions, enhancing residue pair discrimination.

Conclusions:

• The proposed IIGRL model offers a significant advancement in protein-protein binding site prediction.
• Capturing both intra- and inter-protein information through graph representation learning is key to improving prediction accuracy.
• IIGRL provides a powerful new tool for structural biology and drug discovery research.