Hyperbolic graph contrastive learning for drug repositioning over heterogeneous biological networks.

Drug repositioning has become an important computational strategy for identifying new therapeutic uses of existing drugs, especially when conventional drug discovery remains costly and time-consuming. However, accurate drug-disease association prediction remains challenged by the sparsity of verified associations and by the limited ability of Euclidean models to capture the hierarchical structure of heterogeneous biological networks. To address these issues, we propose HGCLDR, a hyperbolic graph contrastive learning framework for drug repositioning. HGCLDR first enhances the sparse drug-disease bipartite graph by incorporating similarity-derived structural information. It then introduces a protein-mediated two-hop topological projection module to infer biologically meaningful high-order drug-disease relations from drug-protein and protein-disease associations, while an adaptive denoising strategy is used to reduce hub-driven and propagation-induced noise. In addition, a layered sampling strategy is designed to construct semantically consistent yet structurally diverse contrastive views from observed associations and projected relations. Based on these views, drug and disease nodes are embedded on the Lorentz manifold and optimized through a hyperbolic graph contrastive learning framework, enabling the model to better capture the hierarchical and non-Euclidean characteristics of biological networks. Across the three benchmark datasets, HGCLDR attains the highest AUROC, AUPR, Accuracy, and Recall values; it also obtains the highest F1-score on the B- and C-datasets and a comparable F1-score on the F-dataset. Further evidence from ablation studies, case analyses, molecular docking, and a representative molecular dynamics simulation supports the effectiveness and biological relevance of the proposed framework.