MiRNA-disease Associations Prediction Research Based On Graph Autoencoders And Collaborative Training

Micro RNA(miRNA)is an endogenous non-coding RNA molecule with a length of about 22 nucleotides.Many studies have shown that miRNA plays an important role in the development of human complex diseases.Therefore,predicting the association between miRNA and disease is helpful to the accurate diagnosis and effective treatment of the disease.Because traditional biological experiments are expensive and time-consuming,many computational models based on biological data have been proposed to predict the association between miRNA and disease.In this paper,a deep learning model MDAGAC based on graph autoencoder and collaborative training is proposed to predict the association between miRNA and disease in an end-to-end manner.The experimental results show that,compared with the other ten advanced models,the prediction performance of the model MDAGAC proposed in this paper has been greatly improved.The main contents of this paper are summarized as follows:First of all,disease integration similarity and miRNA integration similarity are obtained by calculating disease semantic similarity,miRNA functional similarity and Gaussian interaction profile kernel similarity for disease and miRNA.Secondly,based on disease integration similarity and miRNA integration similarity,construct the similarity graphs of miRNA and disease.Each miRNA or disease is represented as a node on the graph,and the similarity matrix can be regarded as the adjacency matrix of the miRNA graph or disease graph.After that,the effect of label propagation is improved through graph autoencoder and collaborative training.This model established two graph autoencoder on miRNA graph and disease graph,and conducted collaborative training on these two graph autoencoders.The graph autoencoder of the miRNA graph and the disease graph can reconstruct the score matrix through the initial association matrix,which is equivalent to propagating the label on the graph.The prediction probability of miRNA-disease association can be obtained from the score matrix.Finally,we evaluate the performance of the method through 5-fold crossvalidation,and the results show that the model MDAGAC achieves an average AUC and standard deviation of 0.9603 ±0.0030 in 5-fold cross-validation.Compared with the previous methods,MDAGAC has the following advantages:(1)MDAGAC implements label propagation through graph autoencoders and cooperatively trains two graph autoencoders,which improves the ability to learn feature representations from miRNA graph and disease graph.(2)MDAGAC is an end-to-end deep learning framework.By reconstructing the association matrix,the graph autoencoders model with collaborative training improves the robustness and accuracy of prediction.(3)The experimental results show that MDAGAC is superior to several existing methods to predict miRNA-disease association.In addition,it provides a general method for the prediction tasks of other biological entities.