Study On The Prediction Of MiRNA-Disease Associations Based On Graph Convolution Neural Network

microRNAs(miRNAs)are a class of non-coding RNAs that play a crucial role in gene regulation within cells.Experimental evidence has demonstrated that abnormal expression of miRNAs can lead to the occurrence of various complex diseases.Therefore,investigating the predictive association between miRNAs and diseases holds paramount significance for the advancement of biomedical research.However,traditional biological experiments face several bottlenecks such as high costs,lengthy experimental cycles,and low efficiency.In order to further improve the efficiency of the study,researchers used the similarity based method to predict the associations between miRNAs and diseases in the early stage,and achieved good results.In recent years,the development of machine learning and deep learning techniques has provided novel insights to researchers.Nonetheless,most existing methods only focus on the direct neighboring node features of miRNAs and disease nodes while struggling to capture high-order neighborhood node characteristics.This thesis leverages the theory of graph convolutional neural networks in deep learning to investigate the associations between miRNAs and diseases from two distinct perspectives: clustering and classification.A structural deep clustering network model(SDCNMDA)is proposed to predict the associations between miRNAs and diseases.SDCNMDA integrates the structural information of miRNAs and diseases into the deep clustering process,aiming to address the classification task using clustering methods.Specifically,the model incorporates the integrated similarity of miRNAs and diseases into an autoencoder,and subsequently propagates the obtained novel features from the autoencoder to the graph convolutional layer through a propagation operator.The model is supervised through a dualsupervision mechanism.On the HMDD v2.0 and HMDD v3.0 datasets,SDCNMDA achieved average AUCs of 93.23% and 94.58%,respectively,using a five-fold crossvalidation strategy.To further evaluate the model’s performance,case studies are conducted on three diseases: breast neoplasms,lung neoplasms,and lymphoma neoplasms.The experiments substantiated that among the top 50 miRNAs associated with breast neoplasms,lung neoplasms,and lymphoma neoplasms,48,46,and 46 miRNAs,respectively,are validated by the db DEMC or miR2 Disease databases.Additionally,a high-order graph convolutional network model(MIXHOPMDA)is proposed to predict the association between miRNAs and diseases by leveraging mixed high-order neighborhood node information.Firstly,the model established a heterogeneous bipartite graph of miRNAs and diseases,ensuring the projection of miRNA and disease feature vectors into the same vector space.Subsequently,the Delta operator is employed in the graph convolutional network to compute different powers of the adjacency matrix,enabling the learning of intrinsic node features of high-order neighboring nodes.Then,the obtained feature matrix is utilized to construct a score matrix through fully connected layers.Finally,the association probability between miRNAs and diseases is determined based on the scores.On the HMDD v2.0 and HMDD v3.0 datasets,MIXHOPMDA achieved average AUCs of 93.36% and 94.63%,respectively,using five-fold cross-validation strategy.Case studies are conducted on three diseases: esophageal neoplasms,colon neoplasms,and lymphoma neoplasms.The research findings revealed that among the top 50 miRNAs associated with esophageal neoplasms,colon neoplasms,and lymphoma neoplasms,47,47,and 48 miRNAs,respectively,are validated by the db DEMC or miR2 Disease databases.This thesis contains 20 figures,13 tables and 84 references.