Research On The Function Prediction Of MicroRNA Based On Network Structure And Graph Neural Network

Micro RNAs(miRNAs)are a class of short,endogenous single-stranded non-coding RNA with a length of 21-25 nt,which are highly conserved in evolution.As an indispensable component of cells,miRNAs play important roles in multiple complex biological processes.A large number of studies have shown that miRNAs,as a class of important transcriptional regulatory factors,are widely involved in the gene regulation process of human diseases.Considering that it is time-consuming and expensive to explore disease-related miRNAs through biological experiments,it is necessary to develop reliable and accurate computational models to predict the functions of miRNAs in diseases and provide valuable candidate miRNAs for biological experiments and pathological studies of diseases.In this thesis,we mainly focus on the function prediction of miRNAs,including correlation prediction of miRNA-disease,the prediction of multiple association types of miRNA-disease,and the identification of coding and non-coding cancer driver.In the prediction of miRNA-disease association,a novel algorithm(BRWHNHA)based on hybrid recommendation algorithm and unbalanced bi-random walk algorithm is proposed to predict potential miRNAdisease association.BRWHNHA takes full advantage of the topological information in the disease similarity network,miRNA similarity network and the experimentally verified miRNA-disease binary association network to predict disease-related miRNAs.First,the disease-disease similarity network and miRNA-miRNA similarity network are constructed by integrated semantic similarity and Gaussian kernel similarity of diseases,functional similarity and Gaussian kernel similarity of miRNAs,respectively.Then,a hybrid recommendation algorithm is used to calculate the probability transition matrix of miRNA-disease binary network.Finally,a weighted miRNA-disease heterogeneous network is constructed and an unbalanced bi-random walk is adopted to predict potential miRNAdisease associations.Through the comparison between BRWHNHA algorithm and several algorithms on 22 diseases by five-fold cross-validation,the comparison results demonstrated that BRWHNHA improves the prediction performance of miRNA-disease association.In addition,case studies of lung cancer and prostate cancer both prove that the BRWHNHA algorithm is an effective tool for predicting miRNAs associated with disease.The type of miRNA-disease association is classified in HMDD database based on experimental evidence.There are up to five types of association between an miRNA and the same disease,including genetics type,epigenetics type,circulating miRNAs type,miRNA tissue expression type and miRNA-target interaction type.Therefore,the pathogenesis of disease cannot be fully understood unless the type of association between miRNAs and diseases,rather than just the association status,are known.Several algorithms have been proposed to predict miRNA-disease association types based on existing association type data by restricted Boltzmann machines,label propagation theories and tensor completion algorithms.However,the non-linear characteristics of the miRNA-disease network with multiple association types are ignored in all models.In this thesis,an algorithm(m DLinker)based on multi-layer heterogeneous graph embedding is proposed to predict miRNA-disease association types.Firstly,an attributed multi-layer heterogeneous network is constructed,in which the attributes of miRNA nodes and disease nodes are defined as the representation vectors embedded through Node2 vec algorithm on miRNA similarity network and disease similarity network,respectively.Then,the embedding representation of miRNA and disease are studied for each association type by an attributed multi-layer heterogeneous network embedding algorithm.Finally,the machine learning algorithm is used to predict miRNA-disease association and related type.The performance of m DLinker algorithm is better than the state-of-the-art two latest algorithms by comprehensive experimental comparison.In addition,two case studies further confirmed that m DLinker algorithm can accurately predict miRNA-disease association and its association types.Cancer drivers play an important role in regulating cell growth,cell cycle and DNA replication,including mutated and unmutated coding genes,non-coding RNA miRNAs,et al.MiRNAs are widely involved in many biological regulatory mechanisms and coordinate with many regulatory factors.An algorithm(HWVote Rank)is proposed to identify multiple cancer drivers based on the co-expression network constructed by the gene expression profiles and regulatory mechanism among miRNA and coding genes in this thesis.In HWVote Rank algorithm,cancer drivers are considered to be key nodes in the co-expression network,and a novel voting algorithm based on heterogeneous network is developed to identify coding and non-coding drivers of cancer.Compared with two popular algorithms that can identify both coding and non-coding cancer drivers,and three key node identification algorithms on breast cancer dataset,HWVote Rank algorithm has a better ability to identify cancer driver genes,especially non-coding drivers of miRNA.The HWVote Rank algorithm has also been applied to identify miRNA drivers in epitheliomesenchymal transformation and cancer typing drivers.Through literature studies,it is found that the HWVote Rank algorithm can be efficiently used to identify drivers in different conditions.