A Data-Driven Modular Framework for Predicting Single-Cell DNA Methylation Landscapes

Abstract

DNA methylation can now be measured at single cell resolution due to recent technological advancements. Current techniques, however, are hampered by insufficient CpG coverage, therefore methods to forecast missing methylation states are essential for genome – wide analysis. In this project, we use DeepCpG, a deep learning – based computational technique to predict methylation status. Here, we evaluate DeepCpG on single – cell methylation data from five cell types generated using alternative sequencing protocols. Compared to other methods DeepCpG yields more accurate results. Furthermore, we demonstrate that the model parameters can be understood, by revealing how methylation variability is influenced by sequence composition.