Research On CircRNA-disease Association Prediction Based On Graph Representation Learning And Multi-source Data

In the wave of the information age,bioinformatics has developed rapidly,and research on the interaction between circular RNA(circRNA)and diseases has also been rapidly advanced.circRNA is a type of non-coding RNA with a covalently closed loop structure that can participate in regulating various biological activities within an organism.Complex diseases,especially malignant diseases,are often related to the dysfunction of a group of biological genes or specific biological pathways.circRNA may be a regulatory factor or carcinogen in the development process of complex diseases.Therefore,by analyzing the association between circRNA and disease,we can explore the role of circRNA in disease and provide assistance for practical medical research.This article focuses on how to predict disease-related candidate circRNAs in the heterogeneous network composed of circRNA and diseases.The research content includes the following two major parts:(1)In order to solve the problem that existing computational methods do not fully utilize the rich topological feature information contained in the circRNA and disease similarity networks and effectively combine circRNA and disease spatial information,this paper proposes a circRNA-disease potential association prediction method GGAECDA based on graph representation learning and graph autoencoder.The model first focuses on the topological relevance of similar nodes in the similarity network and uses a Graph Attention Network(GAT)to aggregate low-order neighbor features while also using the Random Walk with Restart(RWR)to obtain global topological features.The final features of circRNA and diseases are then obtained by fusion.Finally,two Graph Autoencoders(GAE)are constructed to simulate the label propagation process,and a co-training method is applied to combine information from both disease space and circRNA space.Experiments such as five-fold cross-validation and case studies further verify the effectiveness and reliability of the GGAECDA model in predicting potential circRNA-disease associations.(2)In order to solve the problem of missing multi-source information and effectively mine richer potential features in circRNA-disease association prediction,this paper further proposes a dual-view circRNA-disease potential association prediction method MGATCDA based on multi-source data and graph attention network.Firstly,this paper collects interaction information between circRNA,mi RNA and diseases based on the Circ R2 Disease database to obtain a three-layer biological regulatory network of circRNA-mi RNA-disease,which injects multi-source data information into the MGATCDA model.Then,in order to better mine deeper information about circRNA and diseases,a multi-source topological network view and feature view are designed.In the multi-source topological network view,the model uses the RWR algorithm to explore and obtain global topological network information of the circRNA-mi RNA-disease association network;in the feature view,deep autoencoders are used to extract low-dimensional features of circRNA and diseases respectively.GATv2 with more expressive power is further used to assign higher attention coefficients to important features.Finally,MGATCDA uses random forest as a classifier to predict potential circRNA-disease associations.In five-fold cross-validation,MGATCDA has an AUC value of0.9465 and its performance is superior to that of GGAECDA model.The results of case studies further verify the ability of the proposed method in discovering potential circRNAs for human complex diseases.