DRGCL: Drug Repositioning via Semantic-Enriched Graph Contrastive Learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Drug repositioning accelerates the discovery of new drug indications, reducing development costs and timelines.
• Computational methods, particularly graph neural networks (GNNs), are increasingly vital for identifying drug repositioning candidates.
• Existing GNNs often overlook graph semantic information, potentially leading to inconsistencies between global topology and local semantics.

Purpose of the Study:

• To propose a novel drug repositioning model, DRGCL, that addresses the limitations of existing GNNs by incorporating semantic information.
• To enhance the modeling of drug-disease associations by integrating topological and semantic graph features using contrastive learning.

Main Methods:

• Constructed a topology graph from known drug-disease associations.
• Developed a semantic graph by selecting top-similar neighbors based on drug/disease similarity, preserving rich semantic details.
• Employed graph contrastive learning to align embeddings across different spaces, improving feature representation.

Main Results:

• DRGCL demonstrated superior performance compared to state-of-the-art methods across four benchmark datasets.
• Achieved an 11.92% higher average Area Under the Precision-Recall Curve (AUPRC) than the second-best method.
• Case studies confirmed the reliability and effectiveness of the DRGCL model.

Conclusions:

• DRGCL effectively integrates topological and semantic information for improved drug repositioning predictions.
• The proposed graph contrastive learning approach enhances the accuracy and reliability of computational drug repositioning.
• DRGCL offers a promising advancement in identifying novel therapeutic indications for existing drugs.