Prediction Of CircRNA-Disease Associations Based On Deep Learning

Circular RNAs(circRNAs)are non-coding RNAs with multiple biological functions,which are of great research value in the diagnosis,treatment and prognosis of diseases.However,traditional biological experimental methods to mine circRNA-disease associations are expensive and time-consuming,while using computational approaches to predict circRNA-disease associations not only has high prediction accuracy,but also can save a lot of resources and time required by traditional biological experiments.Therefore,in this paper,we propose two novel models for circRNA-disease association prediction based on deep learning techniques.The main work is as follows:(1)To address the inherent sparsity of raw network data that leads to poor prediction performance,this paper proposes a circRNA-disease association prediction model based on a two-stage fusion implementation of graph autoencoder(GIS-CDA).In the first step,circRNA(disease)integration similarity is calculated.In the second step,the graph autoencoder acquires the low-dimensional representations of circRNA(disease);in the third step,the learned features are input to the inductive matrix complement to fill in the missing values of the known association matrix and reconstruct the original association matrix;in the fourth step,the important features in the feature matrix are focused through a self-attentive mechanism to reduce the dependence on other biological information.Finally,model optimization is performed using the Adam optimizer.The results of the five-fold cross-validation show that GIS-CDA exhibits the best performance(AUROC value of 0.9303 and AUPR value of0.2271)compared with the classical model.In addition,the case study verifies that the GIS-CDA model has a good use value.(2)To address the problem of insufficient local graph network structure information leading to low prediction accuracy,this paper constructs a deep learning model based on graph network structure and bilinear representation based on graph node features for circRNA-disease association prediction(CDA-DGRL).In the first step,circRNA(disease)integrated similarity is calculated to construct circRNA-disease heterogeneous network;in the second step,circRNA(disease)node features are obtained using sparse autoencoder;in the third step,circRNA-disease heterogeneous network and node features are input to graph convolutional neural network to capture the local graph network structure;in the fourth step,the local graph network structure is captured using node2 vec is used to depth-first sample the circRNA-disease heterogeneous network to capture the global graph network structure,thus alleviating the problem of sparse raw data;in the fifth step,the local and global graph network structures are combined and input to a hyper-random tree classifier to identify potential circRNA-disease association relationships.The results of the five-fold cross-validation implemented on the circR2 Disease dataset show that CDA-DGRL exhibits the best performance(AUROC value of 0.9866 and AUPR value of 0.9897)compared to similar existing state-of-the-art models.Moreover,Compared with other classifiers based on machine learning,the hyper-random tree classifier used in this model performs better in terms of prediction performance.Meanwhile,the case study validates the effectiveness of the CDA-DGRL model.