Prediction Of Association Between MiRNAs And Disease Based On Graph Attribute Embedding

MicroRNA(miRNA)is an endogenous,miniature,non-coding RNA.miRNAs not only play a variety of regulatory roles in biological processes such as cell growth,reproduction,differentiation,and apoptosis,but also participate in the abnormal expression of many complex diseases,resulting in the loss of the immune system.Therefore,the identification of miRNAs can help to study a variety of complex diseases,and thereby to understand the pathological mechanism and pathogenesis,and promote the development of bioclinical medicine.The comprehensive exploration of miRNA-disease potential relationships has also laid a solid foundation for life science research.In recent years,researchers have successfully discovered the associations of some miRNAs,but there are still many unexplored relationships.Most of the traditional machine learning methods can only study the shallow relationship between them,and can not accurately predict the potential association of miRNA diseases.Moreover,the laboratory methods are limited by the long experiment cycle,high requirements for environmental equipment,and expensive prices,so it is difficult to carry out outside the laboratory environment.Therefore,this thesis aims to study the deep relationship between miRNAs and diseases and has done the following work:(1)Collect a dataset of association relationships between miRNAs and diseases.According to the description methods of the two datasets,miRNA and disease,using the Medical Subject Headings(Me SH)and MISIM numerical calculation methods to calculate the disease semantic similarity and miRNA functional similarity respectively,build a similarity network,and use the three different methods we proposed.The methods CFM1,CFM2,and CFM3 construct three numerical feature matrices.(2)To solve the problem that miRNA and disease network do not combine attribute features in embedded learning,this thesis proposes a prediction model that combines role embedding(Role2vec)method with deep neural networks(DNN)to learn and extract the potential relationship features of miRNA and disease.Based on the role embedding method,the feature vector is mapped into the role space to learn the embedding of each node,and then the numerically expressed feature matrix is input into the deep neural network for training.Finally,a 50 fold cross validation method is used to prevent data from over fitting.In order to further highlight the efficiency of the prediction model proposed in this thesis,the proposed role embedding depth neural network model is compared with different types of feature extraction methods(HOPE,LINE,SDNE)and different classifiers(SVM).The experimental results show that the proposed model has better prediction performance.(3)Aiming at the problem that miRNA and disease can not keep the filter low pass in graph attribute embedding learning,and that the filter and weight affect the experimental performance.This thesis proposes a prediction model that combines Discrete Wavelet Transform(DWT)method with Autoencoder(AE)to learn and extract the potential relationship features of miRNA disease.Based on the method of discrete wavelet transform,the feature vector is filtered by high/low pass to obtain high/low frequency components,and more feature information in low frequency is learned.Then the feature matrix that has been numerically expressed is input into the encoder for training,and the original data is restored finally.In order to further highlight the advantages of the prediction model proposed in this thesis,the proposed discrete wavelet transform self encoder model is compared with different feature extraction methods(FFT,HHT),different feature matrix processing methods(CFM1,CFM2,CFM3)and different parameters.The experimental results show that the proposed model has better performance.In the future work,I hope it can become an effective complementary means for genomics research.