HCLAMCMI: Prediction of circRNA-miRNA Interactions Based on Hypergraph Contrastive Learning and an Attention Mechanism.

Abstract

Circular RNA (circRNA)-microRNA (miRNA) interactions (CMIs) play important roles in regulating gene expression, cell proliferation, and tumorigenesis. Accurate identification of CMIs is critical for understanding disease pathogenesis and for advancing diagnostic and therapeutic strategies. However, conventional biological experiments are time-consuming and labor-intensive, and existing computational models, although effective, still provide suboptimal circRNA and miRNA representations. Here, we propose HCLAMCMI, a computational model for the CMI prediction. Three types of raw features of circRNAs and miRNAs were extracted from the adjacency matrix, similarity matrix, and heterogeneous network comprising circRNAs, miRNAs, and diseases. Hypergraphs were then constructed from two complementary views to capture high-order relational information. These hypergraphs were processed by using hypergraph convolutional networks, contrastive learning, and a channel attention mechanism to generate high-level feature representations. The features were subsequently refined through two-layer fully connected neural networks, and interaction scores were obtained by the inner product to construct the recommendation matrix. HCLAMCMI was evaluated on two benchmark CMI data sets, achieving AUC and AUPR values above 0.98 on training data sets and approximately 0.97 on independent test data sets, consistently outperforming all existing models. Additional analyses confirmed the rationality of its architecture and highlighted the advantages of integrating hypergraph-based learning with attention mechanisms.