Multiview Convolutional Neural Networks for Predicting the Association between miRNA and Diseases

The existing computational models for predicting the association between microRNA(miRNA) and diseases are affected by issues such as network sparsity caused by a single known association and difficulty in distinguishing the importance of features extracted from different networks. The thesis is based on graph convolutional neural networks and effective channel attention mechanisms model named MDA-GCNECA for predicting miRNA-disease associations. The model first utilizes Gaussian interaction kernel similarity for data filling to reduce data sparsity, and inputs each similarity data into the graph convolutional network module for feature embedding. Each embedding is considered as a channel, ultimately forming multiple channels of miRNA and disease mixed data; Next, the effective channel attention module is used to obtain the weights of each channel, and the channel features and attention weights are multiplied before nonlinear activation to obtain the updated feature map; Finally, an information aggregator is used to aggregate the embeddings output from each channel and input them into the classifier to obtain the corresponding prediction score. Through experimental comparative analysis of real data from liver cirrhosis and esophageal cancer, the results showed that the proposed model had a 100% accuracy in predicting the top 15 miRNAs associated with the disease with the highest scores.