Hypergraph Learning with Hyperedge Gating and Multiscale Topology Feature Learning for Predicting Disease-Related circRNAs.

As circular RNAs (circRNAs) play a crucial role in the onset and progression of diseases, identification of disease-related circRNAs can provide deep insights into the pathogenesis of complex diseases. Constructing graph learning-based models is a widely used advanced approach to infer the associations between circRNAs and diseases. The biological correlations among multiple circRNAs, diseases, and miRNAs are complex and diverse. However, existing methods cannot sufficiently represent and encode those correlations. In addition, each circRNA (disease) node has its specific semantic within a meta-path, which is often neglected by the previous methods. To address these limitations, we propose a circRNA-disease association prediction model, HSCDA, which can encode the biological correlations among multiple circRNAs and diseases, learn the semantics of meta-path neighbors, and integrate the multiscale topology structure within a heterogeneous network. First, we construct a hypergraph comprising circRNA and disease nodes to uncover the intricate biological correlations among multiple circRNA and disease nodes. The hypergraph has learnable hyperedges so that HSCDA can dynamically refine the hypergraph topology to reflect multiperspective correlations. Second, we propose hypergraph convolution with a hyperedge gating mechanism (HGHG). HGHG adaptively estimates the hyperedge importance by leveraging the contributions of circRNA, disease, and miRNA nodes contained within each hyperedge. Third, a semantic-enhanced graph convolutional network (SGCN) is proposed to integrate the diverse semantics between circRNA and disease nodes within each meta-path. SGCN is able to generate semantic representations for each circRNA and disease node, derived from integrating the semantic features of meta-path neighbors and the node features. Finally, a multiscale topology transformer (MST) is introduced to effectively learn the structural characteristics of nodes by analyzing their multiscale neighbor topologies in a heterogeneous network. MST can adaptively form topological features by integrating information across various scales and encoding the close relationships among them. Experimental results demonstrate that HSCDA outperforms six comparing methods in predicting circRNA-disease associations. Ablation studies further demonstrate the effectiveness of HGHG, SGCN, and MST, while case studies on three diseases highlight HSCDA's ability in discovering potential circRNA candidates related to the diseases.