To study large-scale chromatin structure we first calculated ChromHMM (15 states model) state-frequencies identified in 200bp genome-wide bins across 127 epigenomes. Then we averaged state frequencies over the 2Mb genomic regions, thus defining 1458 long vector for each state. The unsupervised clustering of a 15×1458 matrix (using Pearson correlation as a similarity measure and complete linkage) revealed 11 distinct genomic clusters enriched in different subsets of chromatin states (Fig. 5d, top heat map). Clusters had different sizes, with the smallest one (c1) containing only 27 bins, while the largest cluster (c9), occupied predominantly by a ‘Quiescent’ state for all epigenomes, had 377 bins. For each 2Mb bin in each cluster we calculated average gene density, Lamin B1 signal (see section 4 above), overlap with different cytogenetic bands (Fig. 5d, bottom, that displays also average levels across each cluster). We also show chromosomal locations of the clusters as well as distributions of CpG island frequency across the 2Mb bins in each cluster (Extended Data 5d).
