Graph Neural Network for Protein-Protein Interaction Prediction: A Comparative Study

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Proteins are essential macromolecules driving biological functions through protein-protein interactions (PPIs).
• Understanding PPIs is key to deciphering cellular mechanisms, including immune responses.
• Computational methods offer efficient alternatives to experimental techniques for predicting PPIs.

Purpose of the Study:

• To conduct a comparative analysis of various graph neural network models for predicting protein-protein interactions.
• To evaluate the efficacy of different network architectures in PPI prediction using protein sequence data.

Main Methods:

• Comparative study of five network models: Neural Networks (NN), Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), Hyperbolic Neural Networks (HNN), and Hyperbolic Graph Convolutions (HGCN).
• Utilized protein sequence information as input for all models.
• Evaluated models on fourteen diverse PPI datasets.

Main Results:

• All analyzed graph neural network models were capable of predicting protein-protein interactions.
• Hyperbolic graph neural networks (HNN and HGCN) generally outperformed other models.
• The superior performance of hyperbolic models was consistently observed across the tested protein-related datasets.

Conclusions:

• Hyperbolic graph neural networks show significant promise for accurate protein-protein interaction prediction.
• This study highlights the potential of hyperbolic deep learning architectures in bioinformatics.
• The findings suggest that hyperbolic geometry may better represent complex biological interaction networks.