Fig. 4

BoostMe error across chromatin states as a function of distance to the nearest non-missing CpG. Absolute prediction error was measured as the difference between the predicted and actual methylation (beta) values for (a) a holdout test set of 5 million CpGs and (b) a holdout set of ~ 1.3 million CpGs in high-variance regions that met BoostMe criteria for training and testing. Gray shaded areas indicate confidence intervals for each smoothed line created using a generalized additive model. Marginal histograms display the distribution of predicted CpGs across chromatin states within the range of the graph