Extended Data Fig. 6: Fifteen-state ChromHMM model.

a, Schematic of the ChromHMM strategy applied in this study. b, Heatmaps showing the maximum Pearson’s correlation of each state in the full model (y-axis) with its best matching state in each simpler model (x-axis). The median correlation of all 24 states is shown in the plots on top of the heatmaps. c, Classification of the k-means clustering of the emission probabilities from all the models. The optimal number of states was defined by the smallest value of k that showed a ratio equal to or higher than 95% (orange line) of the maximum clusters’ separation (red line). SS, sum of squares. d, The emission probabilities for each chromatin mark in each state, as defined by ChromHMM, for both replicates. e, Spearman’s correlation of emission probabilities from ChromHMM models derived from two biological replicates, colour-coded by state (left) or by modification (right). f, Comparison of the ChromHMM model reported here with previously published ChromHMM models. Horizontal white bars indicate chromatin states identified in our study that did not have a clear counterpart in those studies. g, Similarity between replicates from the same tissue-stage (n = 66), from the same tissue any stage (n = 702), or from any tissue any stage (n = 8,646). Similarity measured as pairwise binary distance. Two-sided Mann–Whitney test. h, Enrichment of each mark in state 11 (permissive) relative to state 15 (no signal, genomic background). The ChromHMM emission probability for H3K36me3 in state 11 is >30-fold higher than genomic background. i, Enrichment of chromatin states relative to annotated genes. Gene annotations were not considered during model training or genome segmentation.