Transformer-based multi-modal representation learning and hybrid interaction modeling for miRNA–disease association prediction

MicroRNAs (miRNAs) play critical roles in gene regulation and are closely associated with various human diseases. Identifying miRNA–disease associations is essential for understanding disease mechanisms and developing effective therapeutic strategies. However, experimental validation is costly and time-consuming, making computational approaches increasingly important. In this study, we propose THIMDA, a novel framework that integrates multi-modal representation learning with hybrid interaction modeling for miRNA–disease association prediction. Specifically, miRNA expression profiles are encoded using a TabTransformer to capture complex feature interactions, while disease representations are constructed using SapBERT based on MeSH ontology to incorporate rich semantic information. These heterogeneous representations are projected into a shared latent space and further modeled through a hybrid interaction module that combines bilinear transformation and nonlinear learning to capture complementary interaction patterns. Experimental results demonstrate that THIMDA achieves strong predictive performance, with AUC scores of 0.9569 and 0.9486 under global and local leave-one-out cross-validation, respectively. Case studies on breast and lung cancer further demonstrate its ability to identify biologically relevant miRNAs. THIMDA provides an effective framework for miRNA–disease association prediction by integrating multi-modal representations with complementary interaction modeling. The proposed approach may facilitate the prioritization of candidate miRNA–disease associations and provide useful computational support for subsequent biological investigation.