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

Introduction: Drug development is expensive and time-consuming. Advanced computational methods that mine drug-disease correlations are becoming increasingly popular and are gradually replacing traditional biological experiments. However, most existing techniques rely primarily on network information. They do not fully leverage integration details and rarely capitalize on drugdisease associations. This study proposes the RWRGDR framework, which uses Graph Neural Networks (GNN) for unsupervised feature learning to identify potential drug-disease interactions. The Random Walk with Restart (RWR) algorithm serves as a complementary mechanism to enhance prediction performance. Methods: The GraphSAGE algorithm first encodes low-dimensional representations, leveraging GAT for multi-head attention to weight the significance of neighbors. The RWR algorithm then captures the global network perspective from a given target node, complementing the initial embeddings with global topological descriptors. This convex integration fuses local features and long-range dependencies, ultimately leading to superior downstream predictions. Results: Our model, based on the Multilayer Perceptron (MLP) classifier, achieved outstanding performance, with Area Under the Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC) values of 0.84 and 0.91, respectively. This performance is highly competitive, surpassing previous techniques. Case studies validate its practical applicability. Discussion: Comprehensive network exploration facilitates an in-depth understanding of complex interactions and extracts meaningful insights required for optimized predictions. Despite a relatively lower AUC, our model outperformed prior methods in AUPRC, highlighting its ability to prioritize highly ranked minority positives. Conclusion: RWRGDR represents a potentially reliable drug repositioning strategy, as demonstrated through case studies, indicating its practical significance, particularly for emerging conditions with no recognized treatments.