TraNce: Type-aware hypergraph neural network with biological mediators for drug repositioning

Computational drug repositioning (DR) accelerates pharmaceutical discovery by identifying novel therapeutic indications for approved drugs. Recently, to capture higher-order biological interactions, hypergraph-based DR approaches are proposed. Nevertheless, existing studies exclusively rely on drug and disease similarity networks, or drug-disease association network, to construct hypergraph structures, while neglecting other diverse biological mediators. Furthermore, conventional hypergraph neural networks (HGNNs) encounter the intractable type-agnostic aggregation challenge, wherein heterogeneous entity representations are indiscriminately fused during two-stage message passing process, causing semantic boundary dissolution. To this end, we propose a novel type-aware HGNN model for DR, namely TraNce, which incorporates additional biological entity types into the constructed hypergraph and tackles semantic confusion through devising type-aware message propagation and aggregation schemes. Specifically, we retrieve the biological entities maintaining direct interaction relationships with both drugs and diseases from biomedical knowledge graphs, then discover the bridging intermediates, including genes, phenotypes and pathways, which are further conceptualized as hyperedges linking relevant drug and disease nodes. Additionally, for achieving type-aware feature integration, we separate the propagated messages according to node and hyperedge types, then apply an adaptive element-wise gating mechanism for integrating type-specific message embeddings. Extensive experiments across three large-scale datasets under 10-fold cross-validation confirm the model effectiveness, with TraNce achieving up to 2.17% AUC and 2.69% AUPR gains over the strongest baselines. Source code and datasets are available at https://github.com/OleCui/paper_TraNce.