Human miRNA–disease Association Prediction Via Residual GraphSAGE With Nonlinear Adaptive Feature Fusion and Triplet Contrastive Learning

Abstract

MicroRNAs (miRNAs) play pivotal roles in cellular regulation, and their dysregulation is closely linked to a wide spectrum of human diseases; thus, accurate miRNA–disease association prediction is critical for guiding experimental validation and therapeutic development. In this work, we propose RGFMDA, an innovative framework designed to predict miRNA-disease associations more effectively. RGFMDA employs a residual graph sampling and aggregation network to enhance information localization within miRNA and disease networks. It also features a nonlinear integration of features and a global context integration module that synergistically combine feature interactions and oversee global dependencies. Additionally, the framework uses triplet contrastive learning to refine the distinction between associated and non-associated miRNA-disease pairs, enhancing the accuracy of predictions. On the HMDD v2.0 benchmark, RGFMDA achieved an AUC of 0.9524, surpassing existing approaches whose reported AUC values range from approximately 0.916 to 0.942. On the HMDD v3.2 dataset, RGFMDA further improved performance with an AUC of 0.9604, exceeding state-of-the-art models that demonstrate AUCs between roughly 0.912 and 0.953. Case studies involving lung, esophageal, breast, and colorectal cancers have further confirmed the efficacy of RGFMDA. In summary, RGFMDA represents a robust and reliable computational tool for uncovering novel miRNA–disease associations, thereby facilitating future biological discovery and therapeutic development.