TADDM-WMHHGCN: a weighted multi-head hypergraph convolutional network for predicting triple associations among drugs, diseases, and miRNAs

Associations among drugs, diseases, and microRNAs (miRNAs) are important for elucidating disease mechanisms and supporting drug repositioning. Several computational models predict pairwise drug-disease, drug-miRNA, or miRNA-disease associations. However, few models have addressed associations involving all three entity types, and no model has specifically focused on triple associations among drugs, diseases, and miRNAs (TADDMs). We developed TADDM-WMHHGCN, a weighted multi-head hypergraph convolutional network for predicting TADDMs. Candidate TADDMs were constructed by integrating validated drug-disease, drug-miRNA, and miRNA-disease pairwise associations and were represented using a three-dimensional (3D) adjacency matrix. Raw drug, disease, and miRNA features were derived from this matrix as 3D Gaussian interaction profile (GIP) kernel matrices. We then constructed a weighted heterogeneous hypergraph from the adjacency and GIP kernel matrices. After linear transformation, the raw features were passed through a multi-head hypergraph convolutional network to obtain high-level representations, and a multilayer perceptron decoder generated the predictions. In internal cross-validation on four datasets containing different categories of unlabeled triples of drugs, diseases, and miRNAs, the model achieved AUROC and AUPR values above 0.94 for every dataset. TADDM-WMHHGCN outperformed several existing triple-association prediction models across most evaluation metrics, demonstrating its potential as a tool in the research on the associations among drugs, diseases, and miRNAs.