We then annotate each of the ~3.5M DNase peaks with the chromatin states they overlap with in each of the 111 Roadmap reference epigenomes, using the core 15-state chromatin state model, and focusing on states TssA, TssAFlnk, and TssBiv for promoters, and EnhG, Enh, and EnhBiv for enhancers, and state BivFlnk (flanking bivalent Enh/Tss) for ambiguous regions. Out of these, ~2.5M regions are called as either enhancer or promoter across any of the 111 Roadmap reference epigenomes. Note that because DNase data is not available for all Roadmap epigenomes, the set of regulatory regions defined may exclude DNase regions active in cell types for which DNase was not profiled (Fig. 2g). Although most regions are undisputedly called exclusively promoter or enhancer, there are ~530k regions that needed further study to decide whether they should be called promoters, enhancers, or both (‘dyadic’). We arbitrate on these regions by first clustering them (using the methods in the following section) with an expected cluster size of 10,000 regions, and then for each cluster calculating (a) the mean posterior probabilities for promoter and enhancer
