Heterogeneous graph contrastive learning with gradient balance for drug repositioning.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Drug repositioning accelerates drug discovery by finding new uses for existing drugs.
• Label sparsity in drug-disease association (DDA) prediction is a key challenge.
• Graph contrastive learning (GCL) offers a promising approach to enhance DDA prediction by generating self-supervised signals.

Purpose of the Study:

• To propose a novel heterogeneous graph contrastive learning method with gradient balance (GCGB) for improved DDA prediction.
• To address limitations in existing GCL methods regarding augmented view generation and task optimization imbalance.

Main Methods:

• Introduced a fusion view integrating drug/disease similarity networks and a heterogeneous biomedical network.
• Designed inter-view contrastive learning tasks to contrast fusion, semantic, and interaction views.
• Implemented adaptive gradient balancing to optimize auxiliary and main tasks.

Main Results:

• The proposed GCGB method effectively captures higher-order interaction semantics.
• Gradient balancing improves optimization and guides parameter updates toward the main DDA prediction task.
• Experiments on three benchmarks demonstrated the effectiveness of GCGB.

Conclusions:

• GCGB offers a robust framework for drug repositioning and DDA prediction.
• The method successfully addresses challenges in view generation and task optimization.
• GCGB shows significant potential for accelerating the identification of novel therapeutic indications.