Genome-wide DNA Methylation Prediction Based On Deep Learning

As one of the main contents of epigenetics,DNA methylation is an important mechanism in the regulation of gene expression,which is closely related to many biological processes such as cell differentiation,embryonic development and immune regulation.Previous studies have shown that suppressor genes of complex diseases would be methylated and lost expression.Therefore,methylation status can be used to detect the specificity of tumor suppressor gene.The next generation sequencing technology can realize locus-specific genome-wide methylation measurement,but the sequencing method is relatively expensive.The prediction of DNA methylation by computational methods has become a hot spot in bioinformatics research and an important supplement to experimental methods.Although some traditional machine learning methods have been used to predict DNA methylation status,traditional methods are difficult to effectively extract feature information,and the prediction accuracy needs to be improved.This inspires us to use deep learning model to study this project.Deep learning technology has been proved to be a powerful automatic feature extraction technology.Based on multi-layer structure,it can effectively extract highly complex and important nonlinear features.In this paper,we use deep learning algorithm to construct the prediction model for DNA methylation status.The main contents of this paper are as follows:(1)Deep neural network(DNN)model is used to predict DNA methylation status.Based on the developed DNN prediction model,we systematically compare the importance of the six feature combinations of DNA methylation and then analyzed their effects on the prediction performance.(2)Residual network(ResNet)and Factorization-Machine based Neural Network(DeepFM)models are used to predict DNA methylation status.Using all the feature information of DNA methylation,we carry out a lot of experiments to select the optimal parameters to improve the prediction accuracy of DNA methylation status.(3)Abstract features extracted from deep learning model are used to predict DNA methylation status.Firstly,ResNet and DNN models are used as feature extractors,and some popular machine learning models are used as classifiers.The prediction results show that eXtreme Gradient Boosting(XGBoost)model has the best prediction performance.Then,the features extracted from the two deep learning models are combined with the raw features to further improve the prediction accuracy of DNA methylation status.Finally,we combine the abstract feature data extracted by DNN and the raw feature data to make the prediction accuracy of DNA methylation status reach 92.37%,which is better than the existing prediction methods based on deep learning.