Prediction of Potential miRNA–Disease Associations Through a Novel Unsupervised Deep Learning Framework with Variational Autoencoder

Abstract

The important role of microRNAs (miRNAs) in the formation, development, diagnosis, and treatment of diseases has attracted much attention among researchers recently. In this study, we present an unsupervised deep learning model of the variational autoencoder for MiRNA–disease association prediction (VAEMDA). Through combining the integrated miRNA similarity and the integrated disease similarity with known miRNA–disease associations, respectively, we constructed two spliced matrices. These matrices were applied to train the variational autoencoder (VAE), respectively. The final predicted association scores between miRNAs and diseases were obtained by integrating the scores from the two trained VAE models. Unlike previous models, VAEMDA can avoid noise introduced by the random selection of negative samples and reveal associations between miRNAs and diseases from the perspective of data distribution. Compared with previous methods, VAEMDA obtained higher area under the receiver operating characteristics curves (AUCs) of 0.9118, 0.8652, and 0.9091 ± 0.0065 in global leave-one-out cross validation (LOOCV), local LOOCV, and five-fold cross validation, respectively. Further, the AUCs of VAEMDA were 0.8250 and 0.8237 in global leave-one-disease-out cross validation (LODOCV), and local LODOCV, respectively. In three different types of case studies on three important diseases, the results showed that most of the top 50 potentially associated miRNAs were verified by databases and the literature.

PMDFI: Predicting miRNA–Disease Associations Based on High-Order Feature Interaction

“…In 10-fold cross-validation procedure, this method achieved the AUC of up to 0.8884. Zhang et al (2019) proposed an unsupervised deep learning method implemented by variational autoencoder. The method combines miRNA similarity and disease similarity with identified associations to get two spliced matrices as the input of variational autoencoder, and then obtains the association scores of miRNA and disease.…”

MSFSP: A Novel miRNA–Disease Association Prediction Model by Federating Multiple-Similarities Fusion and Space Projection

“…Inspired by the successful application of machine learning methods in the field of bioinformatics, many researchers used supervised machine learning methods to predict a miRNAdisease association (Chen et al, 2015a(Chen et al, ,b, 2017a(Chen et al, , 2018d(Chen et al, ,f, 2019aLuo et al, 2017a;Xuan et al, 2018Xuan et al, , 2019bWang C.-C. et al, 2019;Wang L. et al, 2019;Zhang L. et al, 2019;Zhao et al, 2019), but which need negative samples for training. Because it is hard to obtain the experimentally verified less-known miRNA-disease associations and negative samples, some semi-supervised learning approaches (such as regularized least squares) with remarkable prediction results were proposed (Chen and Huang, 2017;Chen et al, 2017cPeng et al, 2017b;Xu et al, 2019).…”

Degree-Based Similarity Indexes for Identifying Potential miRNA-Disease Associations

“…Chen et al presented an unsupervised deep learning model for variational autoencoders for MDAs prediction (VAEMDA). VAEMDA respectively constructed two spliced matrices to train the variational autoencoder (VAE), and then obtained the final predicted association scores between miRNAs and diseases by integrating the scores from the two trained VAE models [36]. The most significant advantage of VAEMDA is to avoid noise introduced by the random selection of negative samples.…”