MAGMDA: a multi-order adaptive graph-based miRNA-disease association prediction model.

backgroundMicroRNAs are key biomarkers for human diseases; however, experimentally identifying miRNA-disease associations was costly and inefficient. To improve the robustness and interpretability of computational models, we developed a multi-order adaptive graph-based miRNA-disease association model (MAGMDA).

resultsMAGMDA introduced an adaptive moment-order selection mechanism and a dynamic threshold network, which replaced the fixed moment orders and manually defined thresholds commonly used in traditional high-order statistical modeling with learnable components. By integrating numerically stable high-order moment computation with cross-order attention aggregation, MAGMDA enhanced model robustness and interpretability while maintaining computational efficiency. Five-fold cross-validation on the HMDD v2.0 dataset showed that MAGMDA achieved an AUC of 93.58% and an AUPR of 93.48%. Across multiple evaluation metrics, MAGMDA outperformed representative existing methods, exhibiting particularly consistent advantages on ranking-related metrics such as the area under the receiver operating characteristic curve (AUC) and the area under the precision–recall curve (AUPR), while maintaining smaller performance variations across different cross-validation folds, indicating stable predictive ability under different data partitions. Mechanistic diagnostic analysis further verified the stable utilization of the adaptive modules across different training folds. To demonstrate biological utility, we applied MAGMDA to hepatocellular carcinoma research. Based on the model predictions and analysis of the TCGA-LIHC cohort, we constructed a seven-gene prognostic signature, which was successfully validated in independent datasets and showed strong prognostic stratification ability, highlighting its potential for clinical translation.

conclusionsMAGMDA effectively improved miRNA-disease association prediction through an adaptive multi-order moment modeling framework and, supported by systematic biological validation, provided a powerful tool for translational research from computational discovery to clinical application. The MAGMDA framework and data resources are publicly available at https://github.com/zhangclbio/MAGMDA .