Research On Disease-related CircRNA Prediction Methods Based On Deep Learnin

Noncoding RNAs(nc RNAs)involved in gene transcription and post-transcriptional regulation contribute to many human diseases.Circular RNA(circRNA),as an important member of nc RNA,is an endogenous non-coding RNA with a covalently closed loop structure.With the development of high-throughput RNA-seq technology and bioinformatics methods,more and more studies have discovered the biological functions of circRNAs in multiple signaling pathways related to tumorigenesis,immunity and metabolism.circRNAs affect the disease process by participating in the regulation of proliferation,apoptosis,invasion,migration and cancer metastasis.Therefore,circRNA can not only serve as an immune modulator,but also as a potential biomarker,providing potential opportunities for the treatment of various stages of disease.It is the most traditional method to determine related circRNAs and diseases through biological experiments.Although the results are reliable,a large number of biological experiments have the problems of large consumables,high time cost and high failure rate.To improve research efficiency,scholars have proposed efficient computational methods for predicting potential circRNA-disease associations.Although the existing calculation methods are effective,there are still some problems that need to be solved,such as the inability to accurately represent the underlying features,the inability to grasp the complex structure of the data,the lack of rich biological information features,and the low accuracy of model predictions.In response to the above problems,this paper proposes two computing models for mining disease-related circRNAs based on deep learning algorithms.The main research contents are as follows:(1)A convolutional neural network-based circRNA-disease relationship prediction method(DMFCNNCD)is proposed.Aiming at the problem of insufficient circRNA-disease association information,DMFCNNCD uses projection layers to automatically learn potential representations of circRNAs and diseases,reducing the sparsity of circRNA-disease associations.Second,the DMFCNNCD model calculates the functional similarity of circRNAs,and simultaneously uses two methods to calculate the semantic similarity of diseases.In addition,a more informative training unit is constructed by integrating new circRNA-disease association information,circRNA functional similarity and disease semantic similarity.DMFCNNCD can model nonlinear associations with the help of multi-layer neural networks to grasp more complex data structures.Finally,the AUC value of the five-fold cross-validation(5CV)experiment reached 0.9343,the AUC of the leave-one-out validation(LOOCV)reached 0.9590,and the case study results proved the effectiveness of the DMFCNNCD model.(2)A generative adversarial network model based on adaptive spatial features(ASFGANCD)is proposed.ASFGANCD first fuses the functional similarity,Gaussian kernel similarity of circRNA and the semantic similarity and Gaussian kernel similarity of diseases through adaptive spatial features.Then use the extracted new features and circRNA sequence information to form new training samples to train the generative adversarial network,extract more representative features,and finally use the extreme learning machine to predict the circRNA-disease association.ASFGANCD fuses circRNA and disease information,and adds circRNA biological sequence information as an auxiliary,constructs training samples with richer content,and extracts more representative data information in circRNA-disease association through confrontational neural network,so as to improve the prediction accuracy of the model.The AUC value of the ASFGANCD prediction model in the five-fold cross-validation(5CV)experiment reached 0.9404,the AUC value of the leave-one-out validation(LOOCV)reached 0.9503,and the results of the case study all proved the effectiveness of the model.