Predicting circRNA-Disease Associations by Using Multi-Biomolecular Networks Based on Variational Graph Auto-Encoder with Attention Mechanism

CircRNA-disease association (CDA) can provide a new direction for the treatment of diseases. However, traditional biological experiment is time-consuming and expensive, this urges us to propose the reliable computational model to predict the associations between circRNAs and diseases. And there is existing more and more evidence indicates that the combination of multi-biomolecular information can improve the prediction accuracy. We propose a novel computational model for CDA prediction named MBCDA, we collect the multi-biomolecular information including circRNA, disease, miRNA and lncRNA based on 6 databases, and construct three heterogeneous network among them, then the multi-heads graph attention networks are applied to these three networks to extract the features of circRNAs and diseases from different views, the obtained features are put into variational graph auto-encoder (VGAE) network to learn the latent distributions of the nodes, a fully connected neural network is adopted to further process the output of VGAE and uses sigmoid function to obtain the predicted probabilities of circRNA-disease pairs. As a result, MBCDA achieved the values of AUC and AUPR under 5-fold cross-validation of 0.893 and 0.887. MBCDA was applied to the analysis of the top-25 predicted associations between circRNAs and diseases, these experimental results show that our proposed MBCDA is a powerful computational model for CDA prediction.