RWRGDR: Random Walk and GraphSAGE-based Framework for Enhanced Drug Repositioning

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Traditional drug development is costly and slow.
• Computational methods for drug-disease correlations are gaining traction.
• Existing methods often underutilize network and drug-disease association data.

Purpose of the Study:

• To propose the RWRGDR framework for unsupervised feature learning to identify potential drug-disease interactions.
• To leverage Graph Neural Networks (GNN) and Random Walk with Restart (RWR) for enhanced prediction.
• To improve drug repositioning strategies.

Main Methods:

• Utilized GraphSAGE for low-dimensional representation encoding.
• Employed Graph Attention Networks (GAT) for neighbor weighting.
• Integrated Random Walk with Restart (RWR) for global network perspective.
• Fused local features and long-range dependencies for superior predictions.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.84 and Area Under the Precision-Recall Curve (AUPRC) of 0.91.
• Demonstrated highly competitive performance, surpassing previous techniques.
• Case studies validated the practical applicability and reliability of the RWRGDR framework.

Conclusions:

• Comprehensive network exploration enhances understanding of complex interactions for optimized predictions.
• The RWRGDR model excels in prioritizing highly ranked minority positives, indicated by a superior AUPRC.
• RWRGDR presents a viable drug repositioning strategy, particularly for emerging diseases with unmet needs.