A deep belief network system for prediction of DNA methylation

A Deep Belief Network architecture is proposed for prediction of DNA methylation characteristics across genetic regions. The proposed system uses an image analogous visualisation of DNA methylation features through an efficient mapping model. Implementation of this method has resulted in an accurate classification of DNA methylation for multiple CpG regions identified in cancer cell lines and has been designed to address variability in patterns found in a given human cell, regardless of their function or disease state. The proposed method is compared to time-tested supervised learning algorithms that include Support Vector Machine and Random Forest classifiers and has been validated using data from cancer cell lines. Using documented features, it achieves differentiation of DNA methylation states, while predicting distinct features with an average value of sensitivity 92%, specificity 99%, accuracy 95% and Matthew's Correlation Coefficient 0.91. The feature set coupled with the deep learning model makes the system efficient for DNA methylation prediction, while being independent of the data set used.