Graph Representation Learning Based on Specific Subgraphs for Biomedical Interaction Prediction.

Area of Science:

• Biomedical Informatics
• Network Biology
• Machine Learning

Background:

• Discovering biomedical entity associations is crucial for identifying disease biomarkers and drug targets.
• Graph representation learning (GRL) shows promise for predicting interactions in biomedical networks.
• Current GRL methods aggregate neighbor features equally and lack transparency in higher-order feature integration.

Purpose of the Study:

• To propose a novel graph representation learning framework, MGRS, for enhanced biomedical interaction prediction.
• To address limitations in current GRL methods regarding feature aggregation and transparency.
• To improve the accuracy and robustness of predicting interactions within biomedical networks.

Main Methods:

• Developed a multi-order graph neural network based on reconstructed specific subgraphs (MGRS).
• Implemented a multi-order graph aggregation module (MOGA) for integrating multi-hop neighbor features.
• Introduced a subgraph selection module (SGSM) for reconstructing specific subgraphs with adaptive edge weights.

Main Results:

• MGRS demonstrated superior performance compared to state-of-the-art baselines on four public biomedical networks.
• The framework effectively learns node representations by integrating multi-hop neighbor features.
• SGSM enabled exploration of feature dependencies and learning subgraph-based representations.

Conclusions:

• The proposed MGRS framework significantly improves biomedical interaction prediction.
• MGRS offers a more robust and transparent approach to GRL in biomedical networks.
• This method facilitates the discovery of network biomarkers and potential drug targets.