The Research On Deep Graph Learning For MiRNA-Disease Association Prediction

As a class of endogenous non-coding RNAs,miRNAs are considered as a crucial signal to identify the occurrence and aggravation of human diseases.Precise prediction of the correlation between miRNAs and diseases is conductive to diagnosis for diseases and understanding of complex regulatory mechanism of diseases.Due to the high-cost and time-consuming issue of biological experiments,the study of building computation models for predicting the associations has become a significant topic.Recently,deep learning on graphs has shown great advantages.This paper is devoted to the study of these methods and their application in the associations prediction between miRNAs and diseases.The specific work is as follows:(1)We proposed a prediction model based on metapath2vec and graph autoencoder algorithm,called M2VGAE-MDA.First,a bipartite graph is constructed with known associations between miRNAs and diseases.Metapath2vec is then used to extract the structural features of nodes,which were later integrated with the functional and semantic features.Finally,a graph autoencoder is used to aggregate the neighbor information of nodes to obtain low-dimensional representations.Experiments show that the accuracy of the M2VGAE-MDA is higher than that of traditional models.Compared with the deep learning models,M2VGAE-MDA uses fewer parameters and training time to obtain comparable accuracy.(2)We proposed a prediction model based on neural graph collaborative filtering,called NGCF-MDA.Based on the bipartite graph,the node neighbor information is propagate in a collaborative filtering way.By combining self-connection information,first-order neighbor information and interaction information,the generated miRNA embeddings that regulate the same diseases tend to be similar.So do the disease embeddings regulated by the same miRNAs.Experiments in which diseases are grouped by sparsity show that NGCF-MDA helps to improve the performance of nodes with insufficient associations.In the glioblastoma experiment,49 of the top 50 predicted relevant miRNAs are confirmed in relevant databases,indicating NGCF-MDA can learn the complex regulatory mechanisms of miRNAs.These results proved that NGCFMDA can be used as an effective computational tool to assist biological experiments.