Struct2Graph: a graph attention network for structure based predictions of protein-protein interactions.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Accurate analysis of protein-protein interactions (PPIs) is crucial for understanding cellular signaling and protein functions.
• Existing computational methods for PPI prediction often rely on protein sequences, limiting their ability to capture 3D structural information.
• Deep learning advancements offer potential to complement experimental PPI analysis.

Purpose of the Study:

• To develop a novel method for predicting PPIs directly from 3D protein structural data.
• To address the limitations of sequence-based methods in accounting for protein 3D organization.
• To identify key residues involved in protein-protein complex formation.

Main Methods:

• Development of Struct2Graph, a graph attention network utilizing 3D structural data of proteins.
• Application of a mutual attention mechanism for unsupervised knowledge selection and identification of interaction sites.
• Evaluation of the method on balanced and unbalanced datasets for PPI prediction accuracy.

Main Results:

• Struct2Graph achieved 98.89% accuracy on a balanced dataset and 99.42% average accuracy on an unbalanced dataset (1:10 ratio).
• The method successfully identified interacting residues with 30% sensitivity, 89% specificity, and 87% accuracy.
• The attention mechanism demonstrated an ability to distinguish disease-causing residue variations from benign ones.

Conclusions:

• Struct2Graph is a pioneering 3D-structure-based graph attention network for PPI prediction.
• The learned low-dimensional feature embeddings from graph structures outperform other machine learning classifiers.
• The mutual attention mechanism provides valuable insights into protein interaction sites and disease-related variations.