Fig. 4

Impacts of sequence and epigenetic features on DNA methylation heterogeneity. a Differences between the observed and predicted DNA methylation heterogeneity scores in human 8-cell stage embryos. Promoters were classified into 5 classes based on their DNA methylation states in the zygote stage. The size of each solid circle represents the number of promoters. Each gray box shows the 1st and 3rd quartiles of the predicted heterogeneity scores in each class. b Bar plot of the importance of sequence and epigenetic features for DNA methylation heterogeneity at the 8-cell stage, as determined by random forest analysis. The features include CpG ratio, chromatin accessibility, and H3K4me3 and H3K27me3 levels at the 8-cell stage. c Box plots of the differences in CpG ratio between different categories of promoters in the S1 and S2 classes. The promoters in the S1 and S2 classes were divided into three categories: higher-heterogeneity gene promoters (HHGs; with observed heterogeneity higher than the 3rd quantile of the predicted heterogeneity score of the class), model-predictable heterogeneity gene promoters (MHGs), and lower-heterogeneity gene promoters (LHGs; with observed heterogeneity lower than the 1st quantile of the predicted heterogeneity score of the class). Student’s t test was performed for comparisons between adjacent categories (***p value < 0.001). d Estimated parameters for different groups of promoters with distinct CpG ratios during the first three cell cycles of human embryogenesis. The promoters were grouped into three groups with high (≥ 0.6), moderate (between 0.4 and 0.6), or low (< 0.4) CpG ratios. The lines link the mean values of the parameters for each cell cycle. The gray shaded areas represent the 95% confidence intervals around the mean values. Red indicates the parameter u, blue indicates the parameter d, and yellow indicates the parameter p. e Box plots of the differences in the mean squared error between the original prediction obtained using uniform sets of u, d, and p for all promoters and three conditional predictions based separately on CpG ratio grouping, chromatin accessibility grouping, and H3K4me3 signal grouping. The mean squared error was calculated as the average squared difference between the predicted methylation heterogeneity score and the observed score. Student’s t test was performed for comparisons between the original prediction and each of the conditional predictions (***p value < 0.001; **p value < 0.01)