Robust and comprehensive meta-path modeling with multi-graph learning for miRNA-disease association prediction

MicroRNA (miRNA)-disease association prediction is important for biomarker discovery and therapeutic target prioritization. However, existing methods have not adequately captured heterogeneous biological semantics while maintaining robustness under incomplete or perturbed graphs. Therefore, this study aims to develop a robust graph-learning framework that more effectively captures heterogeneous biological semantics. We propose the Semantic Hierarchical Integration for miRNA-Disease Association prediction (SHiMDA) on contextual heterogeneous biomedical graphs. SHiMDA constructs independent views for direct associations and biologically meaningful meta-paths, and performs segment-wise message passing to explicitly encode the internal structure of meta-path instances rather than collapsing each path into a coarse source-target shortcut. A confidence-weighted ranking loss and over-smoothing regularization further reduce supervision bias and representation homogenization under graph perturbations. Extensive experiments on benchmark datasets show that SHiMDA achieves strong predictive performance in basic and contextual settings and consistently outperforms representative state-of-the-art baselines under adversarial edge and feature perturbations. By explicitly modeling segment-wise meta-path semantics and introducing robustness-oriented optimization, SHiMDA improves the reliability of miRNA-disease association prediction under incomplete or perturbed graphs. This work enhances the reliability of computational miRNA-disease association analysis, facilitating more effective biomarker discovery and therapeutic target prioritization in precision medicine.