Struct2Graph

Struct2Graph predicts protein-protein interactions from three-dimensional folded protein structures using graph attention networks to analyze spatial organization.

Key Features:

• Graph Convolutional Networks (GCNs): Employs two GCNs with weight sharing to extract geometric features from query protein pairs.
• Mutual Attention Network: Integrates a mutual attention mechanism with an unsupervised knowledge selection process to enhance feature extraction and highlight potential interaction sites.
• 3D-Structure-Based Analysis: Directly utilizes structural data of folded protein globules to capture spatial configurations beyond sequence information.
• High-Dimensional Feature Embedding: Produces high-dimensional embeddings for protein pair representations used for interaction prediction.
• Predictive Performance: Reports 98.89% accuracy on a balanced dataset and a 99.42% five-fold cross-validation average accuracy on an unbalanced 1:10 positive:negative dataset.
• Residue Identification Metrics: Identifies residues likely contributing to complex formation with reported sensitivity 30%, specificity 89%, and accuracy 87%.
• Disease-Variant Preference: Shows preference for disease-causing residue variations over benign polymorphisms.

Scientific Applications:

• Protein-Protein Interaction Prediction: Prediction of PPIs using 3D structural information.
• Interacting Residue Identification: Identification of residues that likely contribute to protein-protein complex formation.
• Multi-Ligand Competition Analysis: Applicable to proteins with multiple ligands competing for the same binding area.
• Dynamic Adhesion Interaction Analysis: Applicable to dynamic protein-protein adhesion interactions.
• Disease Mechanism Interpretation: Prioritization of disease-causing residue variations relative to benign polymorphisms for studying disease mechanisms.

Methodology:

Uses a graph attention network architecture comprising two weight-sharing graph convolutional networks and a mutual attention network with an unsupervised knowledge selection process applied to folded 3D protein structures to produce high-dimensional embeddings; model evaluation included five-fold cross-validation on an unbalanced dataset.