Research On MiRNA-disease Association Prediction Based On Random Walk And Convolutional Neural Network

The identification of disease-related microRNAs(disease miRNAs)helps us to understand the etiology and pathogenesis of the diseases.It has been found that miRNA regulates the expression of target gene(mRNA)and performs its biological function.However,the number of target genes verified by biological experiments is relatively less.Therefore,many methods based on target genes to predict disease-related miRNA are difficult to achieve ideal prediction results.At present,most methods for predicting disease-related miRNAs are based on such a biological premise that miRNAs with similar functions are usually associated with similar diseases and vice versa.These methods make use of miRNA and disease-related information to construct bilayer heterogeneous network of miRNA-disease,and predict the potential disease miRNAs by integrating various information in the network.However,these methods do not consider the influence of miRNA family information on calculating miRNA functional similarity.Since miRNAs from the same family usually participate in the process of disease occurrence and development together,it is necessary to incorporate family information into miRNA functional similarity.In addition,these prediction methods are based on shallow models,and it is difficult to mine the hidden,complex and nonlinear deep features between miRNAs and diseases.This will affect the prediction performance of these methods to some extent.In view of the shortcomings mentioned above,we propose two kinds of prediction methods: the first is miRNA-disease association prediction based on random walk with restart,and the second is miRNA-disease association prediction based on the dual convolutional neural network.miRNA-disease association prediction based on random walk with restart,according to the semantic and phenotypic information of diseases,the similarity of diseases is calculated and the disease network is constructed.Functionally similar miRNAs are often involved in regulating similar diseases and vice versa.According to the diseases and family information associated with miRNAs,the similarity of miRNA function is calculated,and the miRNA network is constructed.According to the association information between miRNAs and diseases,the miRNA-disease association network is constructed.MiRNA-disease bilayer heterogeneous networks is obtained by integrating these three networks.On this basis,we propose a prediction method of random walk with restart.For the diseases with some known related miRNAs,the network nodes are divided into labeled nodes and unlabeled nodes,and the transition matrices are established for the two categories of nodes.We assign different transfer probability weights to different categories of nodes.In addition,we control the walker's walking range by restarting probability,which is helpful for reducing the negative effect of noisy data.In the prediction process,the random walk method makes full use of miRNA similarity,known miRNA-disease associations,disease similarity and topology information of the heterogeneous network,we also consider the importance of different network layer information.Our approach has achieved excellent results in terms of ROC-AUC and PR-AUC for 15 human diseases,in addition,the results of case analysis of breast cancer,colorectal cancer,and lung cancer further corroborate the ability of the method to discover potential miRNAs associated with diseases.miRNA-disease association prediction based on the dual convolutional neural network,the traditional association prediction methods are based on miRNA and disease initial feature information(similarity and association feature).There is no representation learning of miRNA and disease initial feature information to mine the non-linear data features implied in the initial information.In view of the fact that the prediction method based on shallow model is difficult to fully mine the nonlinear relationship between miRNA and disease features,we propose a prediction framework based on deep learning and an association prediction method based on dual convolutional neural network.This method not only integrates the initial feature information of miRNAs and diseases,but also captures the topological features of miRNA and disease network.By combining the biological premise of miRNAs and diseases,we construct the feature embedding layer according to the initial features and the topological features of the network,and represent the data information of the feature embedding layer by convolution neural network.The prediction framework is divided into left and right modules.The left module focuses on the initial feature information of miRNAs and diseases,and obtains the deeper feature representation by the representation learning.The right module focuses on the network topology information of miRNAs and diseases,and uses the prediction framework to express the features to learn the hidden nonlinear relationship in the topology information.Finally,the prediction results of the left and right modules are combined to get the miRNA-disease association score according to a certain strategy.Compared with the other four methods,the proposed method is much better than other methods in evaluating the recall rate of the top k candidate results,ROC-AUC and PR-AUC.In addition,case analysis of breast cancer,colorectal cancer and lung cancer further validates the predictive performance of the method.