Adaptive Topology-Semantic Fusion Contrastive Learning for Drug Repositioning

Drug repositioning offers an efficient alternative to de novo drug development; however, accurately identifying potential drug-disease associations remains challenging. Existing methods do not fully exploit higher-order collaborative information at the drug-drug and disease-disease levels, and their feature fusion strategies are often static, lacking adaptability to node-specific differences. To address these limitations, we propose ATSFCL-DR, an adaptive topology-semantic fusion contrastive learning framework for drug repositioning. The method constructs a topology graph and a semantic graph, and introduces an adaptive fusion strategy to integrate node representations learned from these two views, thereby enabling more effective use of higher-order information. In addition, the framework designs a multiview fusion contrastive learning mechanism to enhance the consistency and discriminative ability of cross-view representations. To mitigate the noise introduced during higher-order neighborhood aggregation, the framework further incorporates a layer-level contrastive learning objective to improve the robustness of representation learning. Experiments using 10 repetitions of 10-fold cross-validation on three public benchmark data sets show that ATSFCL-DR outperforms representative baseline methods, achieving an average AUROC of 0.9093 and an average AUPR of 0.5362. Furthermore, by predicting previously unknown drug-disease associations on the Gdata set and validating them against authoritative databases, ATSFCL-DR identifies pramipexole as a promising candidate drug for Alzheimer's disease. In summary, ATSFCL-DR provides an effective approach for drug-disease association prediction and candidate drug screening by adaptively integrating topological and semantic information within a contrastive learning framework.