Prediction Of MiRNA And Disease Associations Based On Heterogeneous Networks

MicroRNA(miRNA)is a kind of non-coding RNA with a length of about 22 nt,which is closely related to the occurrence of many diseases.However,it is costly and timeconsuming to verify the association of miRNA-diseases through biological wet experiments.Therefore,it is meaningful to design effective miRNA-disease prediction methods to provide reliable miRNA candidates.Most of the previous prediction methods only use the similarity and association data related to miRNAs and diseases,but do not take into account the family and cluster information of miRNAs.However,similar diseases are usually related to miRNAs that belong to the same family or clusters.In this paper,three miRNA-disease association prediction methods are proposed.1.A prediction method of miRNA-disease associations based on fully-connected autoencodersIn this study,we proposed a miRNA-disease association prediction method based on fully-connected autoencoders,FMDA.By integrating the similarity and association information between miRNA and disease nodes and the node attributes of miRNAs,we can learn the topological representations of miRNA nodes and disease nodes and the attribute representations of miRNA nodes to predict potential miRNA-disease associations.Firstly,we construct a miRNA-disease heterogeneous graph with node attributes based on miRNA similarity,disease similarity,miRNA-disease associations,and family and cluster attributes of miRNAs.We extract the topological embedding of miRNA nodes and disease nodes on the miRNA-disease heterogeneous graph.s The topology representations of miRNAs,the node attribute representations of miRNAs,and the topology representations of diseases are obtained respectively through three fullyconnected autoencoders.Then the fusion representations of miRNA nodes are obtained through another autoencoder.Finally,the scores of miRNA-disease associations are given by Light GBM(Light Gradient Boosting Machine).By testing and comparing several advanced miRNA-disease association prediction methods on the public data sets and doing case studies,it is proved that FMDA can recommend reliable candidate miRNA.2.A prediction method of miRNA-disease associations based on graph convolution autoencodersIn this study,a miRNA-disease association prediction method,GCMDA,which is based on graph convolution autoencoders and a feature-level attention mechanism is proposed,which fully integrates the topological features of miRNA and disease nodes,the family and cluster attributes of miRNAs to predict potential disease-related miRNA candidates.We constructed a miRNA-disease heterogeneous graph deeply integrating miRNA family and cluster attributes by constructing the miRNA family attribute graph and the miRNA cluster attribute graph and integrating them into the miRNA-disease similar association heterogeneity graph.The topological features of miRNAs and diseases,the family features of miRNAs,and the cluster features of miRNAs are learned through three graph convolution autoencoders.The three kinds of features related to miRNAs are adaptively fused through the feature-level attention mechanism.The experimental results show that GCMDA has better performance than other advanced miRNA-disease prediction methods.The results of the case studies confirmed GCMDA's ability to identify miRNA candidates associated with diseases.3.A prediction method of miRNA-disease associations based on the relation-aware generate adversarial networkIn this study,we propose a miRNA-disease association prediction method,RGMDA,which is based on the relation-aware generative adversarial network.The relationship awareness strategy is used in both the generator and the discriminator to discriminate the different relationships between nodes in the miRNA-disease heterogeneous graph.The relationship-aware generator can generate feature vectors of potential nodes that have a similar or related relationship with real miRNA or disease nodes,while the relationshipaware discriminator can score whether there is a specified relationship between the two nodes.In the comparative experiment,RGMDA shows better performance than other advanced methods in terms of AUC and AUPR.The results of case studies on three typical diseases verify that RGMDA can identify potential miRNA-disease associations.