| miRNA is a little non-coding RNA that assumes a significant part in the existence exercises of organic entities.Research in the past ten years has shown that the regulation of miRNAs also profoundly affects all stages of diseases.People can understand the pathogenesis of diseases at the molecular level by studying the effects of miRNAs on diseases,in order to track down better approaches to treat sicknesses.In the past,studies on the interaction between miRNAs and diseases depended on biological experiments.However,due to the shortcomings of biological experiments and the rapid development of bioinformatics,the use of bioinformatics methods to find miRNAs potentially associated with diseases has become a new trend.Today,computational methods for predicting miRNAs potentially associated with disease can be mainly divided into two categories: network-based methods and machine learning-based methods.The core of these two methods is to use different methods to score unknown miRNA-disease association items.The higher the score,the more likely the miRNA is related to the disease.However,some previous methods have shortcomings such as poor generalization of miRNA/disease feature vectors 、poor method performance and inability to predict isolated miRNAs/diseases.In this paper,the above shortcomings are improved,and based on some knowledge in deep learning,two computational models are proposed to predict the pathogenic miRNAs of diseases,the details of two methods are as follows:(1)EBPMDA,a method for predicting disease-associated miRNAs based on word embeddings and BP neural network.The method first calculates the miRNA/disease comprehensive similarity,and then uses the k-means algorithm to cluster the unknown miRNA-disease association,and selects negative samples equally from each cluster of the clustering results.Then use the improved word embedding method in this paper to extract miRNA/disease feature vectors,and finally use the feature vectors and labels corresponding to positive and negative samples to train a 5-layer BP neural network and use the BP neural network to obtain the association score of unknown samples.Under the experimental conditions of 5-fold cross-validation,the AUC and AUPR values of EBPMDA are 0.9299 and 0.9266,respectively,which can be higher than the AUC and AUPR values of some advanced algorithms by more than 3.3% and 3.4%,respectively.And further case studies were conducted on three cancers.Among the top50 predicted miRNAs,50(colon cancer),50(stomach cancer),and 50(pancreatic cancer)were experimentally verified.(2)ESABPMDA,a method for predicting miRNAs potentially associated with diseases based on word embeddings and stacked autoencoders.This method improves the selection method of negative samples on the basis of EBPMDA,and selects negative samples proportionally from the results of k-means clustering.Then,the word embedding method is still used to obtain the miRNA/disease feature vector.But before training the BP neural network,the initial values of the parameters in the BP neural network are obtained through the stack auto-encoder,and finally the BP neural network is trained and the prediction results are obtained.Under the condition of 5-fold crossvalidation experiment,the AUC value and AUPR value of ESAPPMDA were 0.9357 and 0.9326,respectively.Besides,case studies for gastric cancer,lung cancer,and breast cancer were also performed,and among the top 50 predicted miRNAs,50(stomach cancer),50(lung cancer),and 50(breast cancer)were experimentally validated. |
