GATSDCD: Prediction of circRNA-Disease Associations Based on Singular Value Decomposition and Graph Attention Network

With the deepening of research, we can find that circular RNAs (circRNAs) have important effects on many human physiological and pathological pathways. Studying the association of circRNAs with diseases not only helps to study biological processes, but also provides new directions for the diagnosis and treatment of diseases. However, it is relatively inefficient to verify the association of circRNAs with diseases only by biotechnology. This paper proposed a computational method GATSDCD based on graph attention network (GAT) and neural network (NN) to predict associations between circRNAs-diseases. In GATSDCD, it combined similarity features and semantic features of circRNAs and diseases as raw features. Then, we denoised the original features using singular value matrix decomposition to better represent circRNAs and diseases. Further, using the obtained circRNA and disease features as node attributes, a graph attention network was used to construct feature vectors in subgraphs to extract deep embedded features. Finally, a neural network was applied to make predictions about potential associations. The experimental results showed that the GATSDCD model outperforms existing methods in multiple aspects, and is an effective method to identify circRNA-disease associations. Case study also demonstrated that GATSDCD can effectively identify circRNAs associated with gastric and breast cancers.