Dual-granularity feature learning reveals static and dynamic mechanisms of ncRNA-drug interactions.
Accurately identifying interactions between non-coding RNAs (ncRNAs) and drugs is crucial for elucidating drug mechanisms and advancing drug repositioning. Existing deep learning-based methods for ncRNA-drug association prediction typically leverage multi-source biological information to learn informative representations, thereby alleviating the cold-start issue arising from the scarcity of experimentally validated associations. However, the direct fusion of multi-source features often introduces feature redundancy, semantic conflicts, and high-dimensional feature sparsity, thereby hindering model convergence and degrading predictive performance. To address these challenges, this study proposes DC-MetaMG, a deep learning framework based on a causal disentanglement strategy that models association responses as the synergistic interplay between two complementary mechanisms: static binding and dynamic regulation. By characterizing the principal patterns underlying differential responses between drugs and ncRNAs across multiple dimensions, the proposed framework effectively suppresses spurious associations arising from irrelevant feature interactions and mitigates the adverse effects introduced during model training with multi-source biomedical information, thereby improving the reliability of association prediction. Experiments conducted on lncRNA-drug and miRNA-drug datasets demonstrate the superior performance of the proposed model. Under five-fold cross-validation, the proposed model achieves an average AUC improvement of 0.41% over competing methods. Under the cold evaluation experiments, the AUC, ACC, and AUPR metrics show average increases of 1.22%, 0.51%, and 0.35%, respectively. Case studies further demonstrate the interpretability of the model. Furthermore, visualization analyses confirm the effective characterization of the two underlying association mechanisms, highlighting the model's ability to disentangle and integrate multi-source biological information, thereby demonstrating its suitability for training scenarios involving complex biological information.