CSDPCDA: A miRNA-Mediated Cross-Semantic Regulatory Network Framework for Predicting circRNA-Disease Associations.

CircRNAs are emerging regulators of human diseases, often functioning through miRNA-mediated regulatory networks. However, experimentally validated circRNA-disease associations remain limited, hindering systematic investigation of circRNA functions and disease mechanisms. Here, we propose CSDPCDA, a miRNA-mediated cross-semantic regulatory network framework for predicting circRNA-disease associations by integrating multi-layer molecular information. CSDPCDA constructs a heterogeneous regulatory network incorporating circRNA-disease associations, circRNA-miRNA interactions, miRNA-disease associations, and molecular similarity information. Multiple biologically meaningful meta-paths are modeled to capture diverse regulatory patterns, and a cross-semantic attention mechanism is employed to integrate complementary molecular representations for association prediction. Comprehensive evaluations on the circR2Disease and circRNADisease datasets showed that CSDPCDA obtained higher AUC, AUPR, and F1-score than the representative existing methods under the same experimental settings. Ablation analyses demonstrated that incorporating miRNA-mediated regulatory information consistently improved the predictive performance across different datasets, while integrating complementary meta-path information further enhanced the characterization of complex molecular regulatory relationships. Moreover, case studies of breast cancer, colorectal cancer, and gastric cancer showed that many of the top-ranked predictions were consistent with previously reported disease-associated circRNAs, providing literature-based evidence for their potential biological relevance. CSDPCDA provides an interpretable framework for prioritizing potential disease-associated circRNAs and facilitating further biological investigation.