MUS-HGFC: Inferring miRNA-disease associations with multi-scale hypergraph representations

MicroRNAs (miRNAs) are pivotal post-transcriptional regulators, and their dysregulation is a hallmark of numerous diseases, particularly cancer. Computational prediction of miRNA-disease associations (MDAs) is crucial for identifying biomarkers and therapeutic targets. While graph neural networks (GNNs) have shown promise in this domain, they are often limited by an oversimplified pairwise modeling of complex biological interactions. To address these challenges, we propose MUS-HGFC, a novel computational framework for MDA prediction. Our method introduces a leakage-free miRNA similarity measure, constructed from an experimentally validated miRNA-target gene network, to capture intrinsic functional characteristics independent of disease association data. Furthermore, we model the complex regulatory landscape using a hypergraph structure, which naturally represents higher-order relationships among biological entities. MUS-HGFC employs multi-scale hypergraph convolutional layers to learn comprehensive node representations by integrating both local and global topological information. Extensive benchmarking experiments demonstrate that MUS-HGFC achieves superior performance compared to state-of-the-art methods, including those based on simple graphs and conventional similarity measures. Case studies further confirm the model's practical utility, with a high validation rate for top-ranked predictions. Our work provides a robust framework that effectively captures the higher-order complexity of biological networks, offering a powerful tool for the discovery of novel miRNA-disease associations. The source code is available https://github.com/wuyuanwuhuii/MUS-HGFC.