Predicting miRNA-Disease Associations Based on Adaptive Neighborhood Propagation and Feature Spatial Recombination

MicroRNAs (miRNAs) are critical regulators in biological processes such as cell proliferation, differentiation, and apoptosis, with their aberrant expression strongly linked to a range of complex diseases. Because traditional experimental methods for predicting miRNA-disease associations (MDAs) are both time-intensive and costly, computational models offer an efficient alternative. Graph neural networks (GNNs) have shown promise in MDAs prediction. However, existing models often suffer from limitations, including inadequate neighborhood information aggregation, inflexible propagation schemes, and an imbalance between global and local information. To address these issues, this paper presents a novel GNN framework, APKAGN, designed for predicting miRNA-disease associations. APKAGN enhances performance through three innovative mechanisms: 1) Adaptive local propagation, leveraging a gated recursion module to dynamically adjust propagation depth while employing residual connections to preserve multi-scale features. 2) Multi-subspace global aggregation, capturing global topology information via multi-dimensional projection and density-aware KNN selection. 3) Dynamic feature fusion, integrating local and global representations using an attention-based gating mechanism. Evaluated on the updated HMDD v3.2 dataset across multiple independent random seeds, APKAGN achieved an outstanding average AUC of 95.09%, an accuracy of 88.23%, and an F1-score of 88.34%, outperforming seven state-of-the-art baseline models. Case studies on lymphoma, prostate, and breast tumors further demonstrated the predictive performance of the proposed model, with 26, 25, and 27 of the top 30 predicted miRNAs validated in the dbDEMC and miR2Disease databases, respectively. By leveraging adaptive propagation and dynamic KNN mechanisms, APKAGN significantly enhances the accuracy of MDAs prediction, offering a powerful tool for investigating disease mechanisms and identifying biomarkers.