Pre-training graph neural networks for link prediction in biomedical networks

Area of Science:

• Biomedical informatics
• Network science
• Machine learning

Background:

• Biomedical networks model interactions crucial for disease understanding and drug discovery.
• Graph neural networks (GNNs) are used for link prediction but struggle with integrating diverse data sources for feature extraction.
• Effective feature integration is challenging for various biomedical link prediction tasks.

Purpose of the Study:

• To propose a novel framework, PT-GNN, for integrating diverse data sources for link prediction in biomedical networks.
• To enhance feature extraction for nodes using deep learning and graph convolutional networks (GCNs).
• To leverage pre-trained node features for improved performance and reduced training time in downstream tasks.

Main Methods:

• Developed deep learning methods (CNN, GCN) to learn node features from sequence and structure data.
• Proposed a GCN-based encoder to refine node features by modeling network dependencies.
• Implemented a pre-training strategy using graph reconstruction tasks for node features.
• Applied the framework to synthetic lethality (SL) and drug-target interaction (DTI) prediction tasks.

Main Results:

• PT-GNN demonstrated superior performance on both SL and DTI prediction tasks compared to state-of-the-art methods.
• Pre-trained features from PT-GNN improved the performance of existing models.
• The framework reduced the training time for downstream link prediction tasks.
• Experimental validation confirmed the effectiveness of the proposed PT-GNN framework.

Conclusions:

• PT-GNN offers an effective approach for integrating multi-modal data in biomedical networks for link prediction.
• The pre-training strategy significantly enhances prediction accuracy and computational efficiency.
• This framework advances the application of GNNs in biomedical research, particularly for SL and DTI prediction.