Fig. 4

Imputation performance and clustering of scM&T-seq mouse ESCs [11] based on genome wide methylation profiles. a Prediction performance on test set for imputing CpG methylation states in terms of AUC. Higher values correspond to better imputation performance. Each colored boxplot indicates the performance using 10 random splits of the data in training and test sets; due to high computational costs, DeepCpG was trained only once and the boxplots denote the variability across ten random sub-samplings of the test set. b Example promoter regions with the predicted methylation profiles for three developmental genes: Myc, Esrrb, and Nog. Each colored profile corresponds to the average methylation pattern of the cells assigned to each sub-population, in our case Melissa identified K=3 clusters