A Heterogeneous Graph Convolutional Network-Based Deep Learning Model to Identify miRNA-Disease Association

MiRNAs are proved to be implicated in human diseases. The disease-related miRNAs are expected to be novel bio-marks for disease therapy and drug development. This work develops a Heterogeneous Graph Convolutional Network-based deep learning model, namely HGCNMDA, to perform a MiRNA-Disease Association prediction task. We construct a three-layer heterogeneous network consisting of a miRNA, a disease, and a gene layer. Then we prepare two kinds of attributes for every node in the network and refine the nodes in the network into several node types according to their attributes. After that, a heterogeneous graph convolutional network is employed to learn feature representations for miRNAs, diseases, and genes with finer-grained node type and edge type information. Finally, the miRNA-disease associations are recovered by the inner product of the miRNA features and disease features. The experimental results on the human miRNA-disease association dataset show that the HGCNMDA achieves better performance in AUC values than other five state-of-the-art models.