Learning Universal Knowledge Graph Embedding for Predicting Biomedical Pairwise Interactions.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Predicting biomedical interactions is vital for biological process understanding and drug discovery.
• Graph neural networks (GNNs) excel with ample data but struggle in low-data scenarios due to labeling costs and distribution shifts.
• Self-supervised learning (SSL) shows promise for pre-training GNNs to improve generalization.

Purpose of the Study:

• To introduce LukePi, a novel self-supervised pre-training framework for GNNs on biomedical knowledge graphs (BKGs).
• To enhance node representations by integrating topology-based and semantics-based self-supervised tasks.
• To improve GNN performance in low-data and distribution-shift settings for biomedical interaction prediction.

Main Methods:

• LukePi employs two self-supervised tasks: node degree classification (topology-based) and edge recovery (semantics-based).
• Node degree classification predicts node degree from its local graph structure.
• Edge recovery infers candidate edge type and existence using semantic information within the BKG.

Main Results:

• LukePi significantly outperforms 22 baseline models on synthetic lethality and drug-target interaction prediction.
• The framework demonstrates superior performance in both low-data and distribution-shift scenarios.
• The integrated self-supervised tasks effectively capture rich BKG information, enhancing node representations.

Conclusions:

• Self-supervised pre-training with LukePi is a powerful strategy for GNNs in sparse biomedical data settings.
• LukePi enhances GNN generalizability and predictive power for critical biomedical link prediction tasks.
• The proposed framework offers a robust solution for challenges in biomedical knowledge graph analysis.