ResSAGELink: Residual GraphSAGE with Layer-wise Jumping Knowledge for Robust Drug Repurposing.

INTRODUCTION: Drug repositioning offers a cost-effective alternative to traditional drug discovery. However, computational methods often struggle to capture multi-scale structural patterns and frequently suffer from information loss in deep architectures. METHODS: We propose ResSAGELink, a novel heterogeneous graph neural network that integrates residual connections and jumping-knowledge strategies for drug-disease association prediction. The model effectively captures both topological structure and node features while mitigating oversmoothing in deep networks. RESULTS: Extensive evaluations on five benchmark datasets show that ResSAGELink outperforms seven state-of-the-art baselines, achieving an average AUROC of 0.946 and AUPRC of 0.601. Ablation studies confirm substantial improvements resulting from the addition of residual connections and the integration of jumping knowledge. DISCUSSION: Case studies on five representative drugs reveal that 80% of predictions are validated by biomedical databases, demonstrating practical utility. The model's inductive learning capability also enables predictions for novel compounds not present in the training data. CONCLUSIONS: ResSAGELink provides a reliable computational framework for discovering new therapeutic applications, supporting efficient drug repurposing with strong performance and biological plausibility.