| MiRNA-disease associations prediction relied on the traditional clinical trials in the past few decades.Trials have consumed much time and cost because of diverse miRNAs and diseases.Scientists have employed statistical strategies or deep learning methods to predict the potential miRNA-disease associations in advance with the continuous development of biological data,which boosts the success rate of experiments and guides the direction of biological experimental research effectively.Based on the method of fusing meta-paths and graph neural network,this thesis focuses on the topic of miRNA-disease association prediction.The main research contents are as follows:(1)Predicting miRNA-disease associations by multiple meta-paths fusion graph embedding model.The traditional meta-path-based miRNA-disease association prediction methods often only consider neighbor nodes,ignoring the influence of other associated nodes,which results in one-sided extraction of node features.So,this thesis proposes a Multiple Meta-paths fusion Graph embedding model to predict MiRNA-Disease Associations,M2 GMDA.A miRNA-disease heterogeneous networks has been constructed,which expresses miRNA and disease-related information as meta-paths.A single meta-path has been encoded on average,and the attention mechanism has been used to adaptively calculate the node aggregation characteristics of multiple meta-paths under different meta-path instances and meta-path types.The graph embedding learns the characteristics of associated nodes.M2 GMDA has achieved an AUC value of 91.82% in 5-fold cross validation.(2)Predicting miRNA-disease associations by a combined embedding model.M2GMDA adopts mean encoder on the intermediate nodes of the meta path,however,the nodes contain uneven information,which makes the information in the feature extraction stage insufficient.The node-centric methods often ignore the implicit semantic information between pairs of nodes.In response to the above questions,this thesis proposes a Combined Embedding model to predict MiRNA-Disease Associations,CEMDA.The Gated Recurrent Unit has been used to obtain the node features containing the meta-path information,and the important node information on a single meta-path has been extracted.The Multi-Layer Perceptron has been used to capture the implicit association information between the pairs seriously.CEMDA has attained an average AUC value of 92.03% in 5-fold cross validation.(3)Predicting miRNA-disease associations by heterogeneous graph convolutional network model with enhanced layer.M2GMDA and CEMDA have been affected by the noise of the original similarity matrix data.To solve this problem,the thesis proposes a Heterogeneous Graph Convolutional Network model with enhanced layer to predict MiRNA-Disease Associations,HGCNMDA.The original features of nodes have been optimized by the Random Walk with Restart,and an enhanced layer has been added to the hidden layer of the graph convolutional network.Similar information between nodes in the feature space has been retained and a multi-head attention mechanism has been introduced to calculate the influence of neighboring nodes on the target node.HGCNMDA has reached an average AUC value of 93.01% in 5-fold cross validation.In this thesis,there are 27 figures,12 tables and 100 references. |
