Research On Prediction Of MiRNA-disease Associations Based On Heterogeneous Networks

A growing body of research has shown that abnormal expression of miRNAs is associated with complex human diseases.Predicting the associations between miRNAs and diseases is important for the diagnosis,prevention,and treatment of diseases.Early biological experimental methods were time-consuming and expensive.With the development of computer technology,constructing efficient and accurate miRNAdisease associations prediction models has gradually become a research hotspot.Based on heterogeneous graphs and graph neural networks,this thesis studies the association prediction algorithms for miRNAs and diseases.The main work is as follows:(1)Traditional graph convolution networks only focus on the first order neighbors of nodes.Although multi-layer stacking can be used to expand the receptive field,deep graph convolution often presents an over-smoothing problem.Therefore,this thesis proposes a personalized Page Rank and graph convolutional network-based prediction model for miRNA-disease associations,abbreviated as PRGCNMDA.The model constructs miRNA-disease heterogeneous network,introduces an improved personalized Page Rank method to score and encode the local neighborhood of each root node,and uses transfer probability to preserve locality.The model obtains a larger neighborhood of the central node through an improved graph convolutional network,thereby aggregating richer node features.At the same time,it effectively solves the over-smoothing problem of the deep network.The AUC value of PRGCNMDA using the 5-fold cross-validation in the HMDD v2.0 dataset is 93.24%.(2)The original graph attention network uses an attention mechanism to aggregate the characteristic information of adjacent nodes to calculate the weights of different nodes.Because graph attention networks focus on all their neighbor nodes,which consumes a large amount of computing resources,and not all neighboring nodes have high correlation with the central node.Aggregating these low correlation neighbor features may weaken the weight of the central node.To solve these problems,this thesis proposes a prediction model for associations between miRNAs and diseases based on neighbor selection graph attention networks,abbreviated as NSAMDA.The model uses an improved neighbor selection mechanism to select the k most important neighbor nodes of the central node,and then combines them with a graph attention network to obtain the embedding of miRNA and disease nodes.The 5-fold cross-validation results show that NSAMDA obtain 93.69% of the AUC value in the HMDD v2.0 dataset.(3)Both PRGCNMDA and NSAMDA methods focus on aggregating local features,which often ignore global information in heterogeneous graphs.Therefore,this thesis proposes a new prediction model for miRNA-disease associations based on deep adaptive propagation neural networks(ADPMDA).The model stacks the embedding vectors of k propagation layers to obtain the global features of the nodes,and introduces a trainable projection vector to measure the weight of each propagation layer,thereby adaptively adjusting the network embedding for deep propagation.ADPMDA can effectively and adaptively aggregate local and global information of nodes,and obtain 94.75% of AUC value in the HMDD v3.0 dataset.