Research On Disease-related CircRNA Prediction Method Based On Graph Neural Networ

Accumulating studies have shown that circRNAs play essential roles in the physiological and pathological pathways of humans.The study of disease-related circRNAs can help people to understand the pathogenesis of diseases,which is conducive to the prevention and treatment of complex diseases.Biological verification experiments require a lot of manpower,material resources and time,and the computational method has the advantages of low cost,high efficiency and orientation,which makes it develop rapidly in this field.Although the current computational models have achieved some results,there are still some urgent problems to be solved,such as low prediction accuracy of the models,inaccurate representation of potential features,singular data sources and similarity metrics,etc.In view of the above problems,GCNMFCDA and VGAECDA computational models are proposed.And a circRNAdisease-Micro RNA(miRNA)multivariate heterogeneous network is constructed.(1)A prediction model GCNMFCDA based on Graph Convolution Network(GCN)and Matrix Factorization(MF)was constructed.Firstly,GCNMFCDA utilizes circRNA functional similarity and Gaussian kernel similarity to form the initial features of circRNA.Similarly,the disease semantic similarity and Gaussian kernel similarity are utilized to constitute the initial features of the disease.Then,the model applies GCN to learn potential embeddings of circRNA and disease based on initial node features.Finally,the inner product between circRNA embedding and disease embedding is introduced to construct the new score matrix based on matrix factorization.(2)A circRNA-disease-miRNA heterogeneous network was constructed.Considering that circRNA can co-regulate the generation of complex diseases as miRNA sponges,the biological data of miRNAs were introduced into circRNA-disease association information.First,the detailed information and association of circRNA,miRNA and disease were downloaded from multiple datasets.Then,data preprocessing operations were performed to remove redundant information.The naming methods of circRNA,miRNA and diseases were collected and unified.Finally,the heterogeneous network was created and visualized based on graph neural network technology.On the basis of this network,more biological macromolecular information can be introduced,so as to form a more detailed and larger-scale disease regulatory network.(3)A prediction model based on Variational Graph Auto-Encoders was constructed(VGAECDA).Firstly,miRNA integration similarity is constituted by miRNA functional similarity,miRNA gene expression profile similarity and miRNA Gaussian kernel similarity.Similarly,the circRNA integrated similarity is constituted by circRNA functional similarity,circRNA sequence similarity and circRNA Gaussian kernel similarity.The disease semantic similarity,disease Gaussian kernel similarity and BioBERT pre-trained model provide disease information to form disease integrated similarity.Next,the variational graph autoencoder is employed to integrate the features and topological information of circRNA and disease nodes,resulting in latent feature vectors.Finally,XGBoost is used as a predictor to obtain the predicted scores of potential associations.The experimental results show that GCNMFCDA and VGAECDA prediction models have a large improvement in performance metrics such as accuracy,precision,F1-score and AUC value.Furthermore,the case studies confirm their ability to predict disease-related circRNA to certain extent.