Abstract
The prediction of microRNA-disease associations (MDAs) is crucial for understanding disease mechanisms and biomarker discovery. While graph neural networks have emerged as promising tools for MDA prediction, existing methods face critical limitations: (1) data leakage caused by improper use of Gaussian interaction profile (GIP) kernel similarity during feature construction, (2) self-validation loops in calculating miRNA functional similarity using known MDA data, and (3) information bottlenecks in conventional graph neural network (GNN) architectures that flatten heterogeneous relationships and employ over-simplified decoders. To address these challenges, we propose CoupleMDA, a metapath-guided heterogeneous graph learning framework coupling structural and semantic features. The model constructs a biological heterogeneous network using independent data sources to eliminate feature-target space coupling. Our framework implements a two-stage encoding strategy: (1) relational graph convolutional networks (RGCN) for pre-encoding and (2) metapath-guided semantic aggregation for secondary encoding. During decoding, common metapaths between node pairs structurally guide feature pooling, mitigating information bottlenecks. The comprehensive evaluation shows that CoupleMDA achieves a 2-5% performance improvement over the current state-of-the-art baseline methods in the heterogeneous graph link prediction task. Ablation studies confirm the necessity of each proposed component, while case analyses reveal the framework's capability to recover cancer-related miRNA-disease associations through biologically interpretable metapaths.