Identifying Complementary Therapeutic Relationships for Drug Repurposing via Spectral-Spatial Graph Contrastive Learning.

Drug repurposing facilitates the discovery of novel therapies by identifying alternative indications for clinically approved drugs. However, many existing deep learning approaches rely heavily on the assumption that drug similarity implies therapeutic similarity, where drug similarity is usually derived from chemical-structure related and other biomedical similarity profiles. This restricts their ability to identify repurposing opportunities where drugs are therapeutically relevant yet function ally diverse. Several successful repurposing cases (e.g., metformin in cancer therapy) reveal the importance of modeling both interaction and dissimilarity attributes in drug behavior. To uncover such "metformin-like" discoveries, we propose MPGCL, a unified spectral-spatial graph contrastive learning framework for identifying repurposed drugs that form complementary therapeutic relationships with existing treatments. MPGCL employs a mid-pass spectral filter to extract dissimilarity signals and a low-pass spatial filter for interaction-aware features. A contrastive learning objective further enhances the distinction between drugs with different chemical or biological similarity profiles, enabling more expressive and informative representations. Extensive experiments on real-world drug repurposing benchmarks demonstrate that MPGCL consistently outperforms existing methods. Case studies on breast cancer further validate the model's ability to uncover promising drug candidates with dissimilarity attributes in drug combination therapy. These findings suggest that MPGCL offers a robust framework for discovering novel therapeutic relationships beyond conventional similarity guided paradigms.