Front Bioinform. 2026 Jul 1;6:1873526. doi: 10.3389/fbinf.2026.1873526. eCollection 2026.
ABSTRACT
INTRODUCTION: Identifying potential miRNA-disease associations is essential for clarifying the molecular basis of complex diseases and accelerating the discovery of diagnostic biomarkers and therapeutic targets. However, the performance of existing computational methods is often limited by sparse biological interaction networks, highly imbalanced disease distributions, and the small number of experimentally validated associations.
METHODS: To address these challenges, we propose MMAG, a novel framework that formulates miRNA-disease association prediction as a meta-conditional distribution alignment problem on multi-scale biological graphs. MMAG integrates three complementary components. First, a multi-scale representation learning module captures hierarchical biological information from local topological connectivity, mesoscopic functional organization, and global spectral structure. Second, a meta-learning strategy models each disease as an individual task, enabling the model to learn disease-specific prototype representations from support samples and adapt effectively to few-shot settings. Third, a conditional adversarial alignment mechanism reduces feature distribution discrepancies across diseases with different data scales, thereby enhancing cross-task knowledge transfer and generalization.
RESULTS: Extensive experiments demonstrate that MMAG consistently outperforms several state-of-the-art methods under few-shot, long-tailed, and cross-dataset transfer scenarios.
DISCUSSION: These results indicate that MMAG provides an effective and scalable solution for miRNA-disease association prediction and offers a promising strategy for broader biological network inference tasks.
PMID:42460394 | PMC:PMC13368795 | DOI:10.3389/fbinf.2026.1873526