Prediction Of MiRNA-disease Association Based On Graph Convolutional Neural Networ

There are various diseases in the world,such as diabetes,cancer,and heart disease.Cancer alone has an annual incidence of 3.93 million people,posing a serious threat to human health and safety.Mi RNAs are small non-coding RNAs that can affect m RNA expression by controlling RNA splicing or translation inhibition.Through extensive experimentation,it has been proven that mi RNAs play important roles in many biological processes such as cell proliferation,development,differentiation,metabolism,aging,apoptosis,signal transduction,and viral infection.There is a complex association between mi RNA dysregulation and the occurrence and development of complex diseases,which is important for the diagnosis and ultimate cure of diseases such as cardiovascular disease,malignant tumors,mental illness,and diabetes.So far,many mi RNAs have been confirmed to be associated with cancer.Identifying mi RNA-disease associations is of great significance for the diagnosis and ultimate cure of related diseases.Therefore,researchers have conducted a large number of experiments in the hope of finding better methods to predict the relationship between mi RNAs and diseases in the future.Graph Convolutional Neural Networks(GCN)is a deep learning model specialized in handling graph-structured data,which can extract features and perform classification by utilizing the adjacency relationships between nodes and the similarities between node features.It has good robustness and high scalability,and has demonstrated good predictive performance in link prediction.However,there are still some problems that need to be addressed.In order to better predict the potential association between mi RNAs and diseases,this dissertation makes targeted improvements and optimizations to the method based on GCN,and applies it to the prediction of potential associations between mi RNAs and diseases.The specific research contents are as follows:(1)In order to address the problem that GCN has difficulty assigning different weights to different node features,a method is proposed to introduce layer attention mechanism.Firstly,a heterogeneous network is constructed using known associations between mi RNAs and diseases,as well as known mi RNA functional similarities and disease semantic similarities.Then,new mi RNA-disease associations are obtained through encoding and decoding operations.Meanwhile,layer attention mechanism is introduced in the encoding process.After the introduction of layer attention mechanism,all useful structural information from multiple graph convolutional layers can be integrated to build a graph attention network,which assigns different weights to different node features to improve the accuracy of predictions.(2)In response to the problem of the unsatisfactory performance of learning potential feature representations for mi RNA-disease inductive matrix completion models,we propose to introduce adversarial training methods into the neural inductive matrix completion model.First,we define the discriminator and generator networks,where the discriminator judges the authenticity of the feature matrix and the generator generates pseudo-feature matrices.Then,we use binary cross-entropy as the loss function and define an adversarial loss function to train the discriminator and generator.Next,we train the discriminator and generator networks,aiming to accurately distinguish between generated pseudo-data and real data through training the discriminator network,and to deceive the discriminator as much as possible to make the generated pseudo-data more similar to the real data through training the generator network.Finally,we apply the model to mi RNA-disease association prediction,effectively improving the accuracy of prediction.(3)To address the sparsity and incompleteness of existing datasets,matrix completion algorithms were applied to a graph autoencoder to predict potential mi RNA-disease associations.The matrix completion algorithm was used to reconstruct the mi RNA-disease association matrix by integrating multiple data sources into a mi RNA-disease bipartite graph that includes mi RNA nodes and disease nodes.In the mi RNA-disease association dataset,known associations were considered as positive links between mi RNA nodes and disease nodes.To maintain sample balance,an equal number of negative samples were constructed by randomly selecting the same number of links from unknown associations and adding them to the mi RNA-disease bipartite graph.Then,all positive links were labeled as 1 and all negative links were labeled as 0 for subsequent model training.Finally,a graph neural network-based encoder was used to generate embeddings for mi RNA and disease nodes,a bilinear decoder was used to reconstruct the links in the bipartite graph,and a cross-entropy loss function was used for end-to-end training of the entire model.Extensive experiments have shown that the three improved models have excellent predictive performance,providing new insights for the prevention,diagnosis,and treatment of some diseases.