Multi-View Contrastive Learning for miRNA-Disease Associations

Wei Xu^*

• Guangxi Normal University, Guilin, China

MiRNA-disease associations (MDAs) are particularly insightful for revealing the complex pathology of human diseases. Traditional experimental methods are expensive, time-consuming, as well as low throughput. Thus,many machine learning-based methods have been proposed to predict MDAs, but these often fall short in predictive performance due to limitations of supervised learning. To address this issue, we introduce a novel MDA prediction method named MVCL-MDA, based on heterogeneous graph meta-path views and network structure view for graph contrastive learning. This heterogeneous graph consists of miRNAs, genes, and diseases, constructed from six different databases. In MVCL-MDA, miRNAs and diseases are first encoded through Graph Convolutional Networks (GCN) in the meta-path view, while in the network structure view, nodes are encoded through node and type-level cascade attention mechanisms. Subsequently, a contrastive learning method is applied to optimize the embeddings of diseases and miRNAs based on these two types of node embeddings. Finally, the embeddings of miRNAs and diseases are concatenated and inputted into classifier to obtain their association scores. The results on publicly available datasets show that MVCL-MDA outperformed all existing methods in terms of prediction accuracy. It achieved an AUC of 94.43% and AUPR of 94.71% in a five-fold cross-validation.