MHNNMDA: multi-stage hypergraph neural network for predicting miRNA-disease association types.

MicroRNAs (miRNAs) are short non-coding RNAs that play crucial regulatory roles in biological processes and are closely implicated in human diseases. Current computational methods predominantly focus on binary miRNA-disease association prediction, suffering from two principal limitations: the inability to discriminate specific association types and the reliance on simple graph structures that fail to represent complex higher-order biological relationships. To address these challenges, we propose a multi-stage hypergraph neural network for predicting miRNA-disease association types, termed MHNNMDA, which sequentially integrates hypergraph construction, dual-attention feature learning, and hypergraph convolutional propagation. Specifically, MHNNMDA integrates multi-source biological data to construct similarity networks for miRNAs and diseases, and further transforms these networks into hypergraph structures to capture higher-order group interactions. MHNNMDA incorporates a dual-attention hypergraph neural network with node-level and hyperedge-level attention mechanisms, enabling adaptive feature aggregation and dynamic weighting of high-order semantic relationships. Subsequently, a hypergraph convolutional network propagates and refines node embeddings to infer multiple association types. Comprehensive experiments demonstrate that MHNNMDA consistently outperforms some state-of-the-art methods across multiple benchmark datasets, validating its effectiveness in predicting association types and modeling complex biological interdependencies.