Predicting MiRNA-disease Associations Through Deep Sparse Autoencoder

MicroRNA(miRNA)is a series of endogenous non-coding RNAs with regulatory functions that play an important role in the growth and development of organisms Many biological experimental studies have proved that miRNA is closely allied to human diseases,include the generation,prevalence,transmission,diagnosis and treatment of diseases.Therefore,it is important to identify disease-associated miRNAs for understanding the pathogenesis of complex diseases.However,biological experiments to verify disease-associated miRNAs is often expensive and time-consuming.Therefore,the development of effective models and methods to predict disease-associated miRNA,and,can provide alternative disease-associated miRNAs for biologists,enhance the purpose of biological experiments and reduce the cost of biological experiments.Predicted the association between miRNA and disease is a valuable topic.In this paper,a neural network model based on sparse autoencoder was proposed to predict miRNA-disease association from two aspects of miRNA and disease similarity calculation and miRNA and disease correlation autoencoder neural network model.The results show that the prediction performance of the neural network model of sparse autoencoder is superior to the previous methods.The main work includes:(1)In the study miRNA and disease similarity network,this paper improved the calculation method of disease semantic similarity matrix.Compared with the single disease semantic similarity method,the improved disease semantic similarity matrix is more reasonable and reliable.(2)Aiming at the study on miRNA and disease associated neural network model of the autoencoder level,this paper proposed the miRNA-disease association prediction model of deep sparse autoencoder(DSAEMDA for short).Add sparsity to the exist autoencoder to form a new computational framework which named DSAEMDA(deep sparse autoencoder miRNA-disease association).First,we obtain two disease semantic similarity matrices by calculating disease semantic similarity in two ways,and obtained miRNA functional similarity matrices by calculated functional similarity.Bound to the Gaussian interaction spectrum kernel similarity matrix of the disease and miRNA,respectively.Then,in terms of feature extraction,we embedded integrated miRNA and disease data into high-dimensional space to extracted high-dimensional expression of disease and miRNA.Finally,we trained our deep sparse autoencoder use proven miRNA-disease association data.In addition,the reconstruction error can be used to predict the correlation value of certain disease-associated miRNA.Experimental results show that DSAEMDA is superior to other methods in AUC and can effectively predict disease-related miRNA.And the ability to identify potential disease-related candidate miRNAs.