Figure 2

Heart enhancers can be identified accurately using data from different cellular contexts. A random forest classifier was trained to distinguish E11.5 heart enhancers from the genomic background and from enhancers active in other tissues. Each classifier used H3K4me1 and H3K27ac patterns from different sets of cellular contexts as features (Figure 1). (A) In five-fold cross validation, the classifiers based on data from other stages of heart development accurately identified E11.5 heart enhancers; these classifiers achieved ROC AUCs of 0.96 vs. the genomic background and 0.85 vs. other enhancers. (B) The classifiers that used data from non-heart tissues as features performed well (AUCs of 0.91 and 0.72), but were worse than the developmental-stage-based classifiers. (C) When mapped between species, the histone marks from mouse heart development were also able to identify human developmental enhancers better than random (AUCs of 0.87 and 0.82). Note that the results in (C) should not be directly compared to the Stage and Tissue results, because they are based on different sets of enhancers. As expected, distinguishing heart enhancers from the genomic background was easier than from non-heart enhancers in each scenario.