HeMDAP: Heterogeneous Graph Self-Supervised Learning for MiRNA-Disease Association Prediction

MiRNA-disease associations (MDAs) are particularly insightful for revealing the complex pathology of human diseases. Traditional experimental methods for MDA studies are costly, time-consuming, and low-throughput. Thus, many machine learning and deep learning-based methods have been proposed to predict MDAs. However, due to the limitations of supervised learning, these methods often fall short in predictive performance. To address this issue, we propose a novel MDA prediction method, HeMDAP, which is based on graph contrastive learning using meta-path view and network structure view of heterogeneous graph. The main innovation of HeMDAP lies in designing two complementary graph representations, network structure view and meta-path view, as well as introducing self-supervised contrastive learning and supervised contrastive learning to effectively optimize node embeddings. Additionally, we utilize knowledge-aware enhancement to further improve embedding quality. This multi-view learning and multi-task training strategy can more comprehensively capture the complex relationships among miRNAs, genes, and diseases. Experimental results on public datasets show that HeMDAP outperforms all existing methods in terms of prediction accuracy. In five-fold cross-validation, HeMDAP achieved an AUC of 94.92% and an AUPR of 95.07%. These results demonstrate the effectiveness and superiority of our proposed method in the MDA prediction task.