Associations Prediction Of MiRNAs And Diseases Based On Graph Neural Network

Micro RNAs(miRNAs)are one class of endogeneous,microscopic,non-coding RNAs.Recent studies have shown that the development and evolution of lots of human complex diseases are influenced by the mutations or abnormal expression of miRNAs.Therefore,the identification of potentially disease-related miRNAs can help healthcare professionals understand the pathological mechanisms of diseases and facilitate the development of clinical medicine.However,traditional biological wet-lab methods suffer from long lead time,low efficiency and high costs.Therefore,there is an urgent need for efficient computational models to predict potential associations between miRNAs and diseases.To explore the complex interaction information between miRNA molecules and diseases,this thesis proposes three computational models to predict potential miRNA-disease associations from the perspectives of node classification and link prediction based on the idea of graph neural network,as follows:An associations prediction algorithm of miRNAs and diseases based on fully connected graph convolutional network is proposed.This model treats miRNA-disease pairs as nodes in the fully connected graph and classifies the nodes in the graph to achieve the associations prediction effect.For the purpuose of fully propagating information in the fully connected graph,this model balances the similarity between miRNAs and the similarity between diseases as the weights of edges.Meanwhile,the principal component analysis algorithm is utilized to obtain the low-dimensional and highdiscriminational features of miRNAs and diseases respectively,which are concatentated to form a miRNA-disease pair feature matrix.Finally,a two-layers graph convolutional network is used to predict potential associations between miRNAs and diseases.5-fold cross validation results show that this model achieves an average AUC value of 92.85% on data1 dataset.An associations prediction algorithm of miRNAs and diseases based on heterogeneous graph auto-encoder model is proposed.This model solves the problem of predicting associations between miRNAs and diseases as a miRNA-disease heterogeneous graph,with the aim of predicting whether there are association edges between miRNA nodes and disease nodes.This model designs a graph neural network-based encoder structure to synchronously aggragate the heterogeneous neighbourhood information of nodes in the heterogeneous graph in the manner of aggregator functions,and fuse it into the initial representation of the nodes.By stacking this encoder,the structural information of the heterogeneous graph can be explored in depth to generate effective feature embeddings of miRNA nodes and disease nodes.Finally,a bilinear decoder is utilized to reconstruct the associations between miRNA and disease nodes.This model achieves an average AUC value of 93.56% on data1 dataset based on 5-fold cross-validation.An associations prediction algorithm of miRNAs and diseases based on heterogeneous graph attention network is proposed.This model addresses the miRNA-disease associations prediction problem from the perspective of link prediction.Firstly,this model designs a transformation matrix to project the heterogeneous nodes of the graph to the same vector space.Secondly,this model adaptively distinguishes the degree of influence of each neighbour node on the central node based on attention mechanism and aggregates the heterogeneous neighbour information.At the same time,the attention mechanism is extended to multi-heads to explore the structural features of different vector subspaces.Afterwards,neighbour aggregation features are fused with node attribute features as feature representations of miRNA nodes and disease nodes.Finally,a fully connected layer is used to predict potential associations between nodes.Under5-fold cross-validation,this model achieves average AUC values of 93.52% and 94.82%on data1 and data2 datasets,respectively.This thesis contains 13 figures,18 tables and 87 references.