MAGIN-GO: Protein function prediction based on dual graph neural networks and gene ontology structure

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Biology

Background:

• Accurate protein function prediction is crucial for biological research.
• Deep learning, especially Graph Neural Networks (GNNs), shows promise but faces limitations in feature representation and capturing long-range dependencies.
• Existing GNN methods struggle to integrate diverse data sources like sequences and interaction networks effectively.

Purpose of the Study:

• To develop an advanced method, MAGIN-GO, for improved protein function prediction.
• To address limitations of traditional GNNs in feature representation and data integration.
• To leverage multi-source protein information, including sequence, protein-protein interactions (PPI), and Gene Ontology (GO) annotations.

Main Methods:

• MAGIN-GO combines Graph Isomorphism Network (GIN), Graph Convolutional Network (GCN), and Graph Convolutional Self-Attention Network (GMSA).
• It integrates protein sequence features with PPI graph node features.
• Pre-trained GO term embeddings are incorporated into a multi-label classification framework.

Main Results:

• MAGIN-GO significantly outperforms existing methods on the UniProtKB/Swiss-Prot dataset.
• Achieved superior Area Under Precision-Recall (AUPR) scores: 0.569 (MF), 0.434 (BP), 0.754 (CC).
• Demonstrated high performance with Fmax, Smin, and AUC scores across all GO domains.

Conclusions:

• MAGIN-GO offers a robust and effective approach for protein function prediction.
• The integration of multi-source data and advanced GNN architectures enhances predictive accuracy.
• The method shows significant potential for advancing protein research and understanding biological activities.