Heterogeneous Graph Contrastive Learning with Graph Diffusion for Drug Repositioning.

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• Drug repositioning offers a cost-effective alternative to traditional drug development.
• Accurately modeling complex drug-disease relationships is a significant challenge.
• Existing methods struggle to capture both local and global feature representations effectively.

Purpose of the Study:

• To introduce HGCL-DR, a novel heterogeneous graph contrastive learning framework for improved drug repositioning.
• To effectively integrate global and local feature representations for enhanced drug-disease relationship modeling.
• To validate the framework's performance and practical utility in identifying novel drug candidates.

Main Methods:

• Developed an improved heterogeneous graph contrastive learning approach for drug-disease relationships.
• Employed a bidirectional graph convolutional network with subgraph generation for local feature extraction.
• Utilized graph diffusion for long-range dependencies and contrastive learning for global feature extraction.

Main Results:

• HGCL-DR consistently outperformed state-of-the-art baselines across four benchmark datasets.
• Achieved superior performance in AUPR, AUROC, and F1-score metrics.
• Case studies demonstrated practical utility in identifying potential drug candidates for Alzheimer's disease and breast neoplasms.

Conclusions:

• HGCL-DR is an effective computational approach for drug repositioning.
• The framework successfully integrates global and local features for robust drug-disease modeling.
• The proposed components significantly contribute to the overall performance and utility of the model.