Fig. 2

Evaluation of the seven methods for defining active and non-active CRMs. A Performance of LR models trained on the pairs of positive and negative sets defined by the seven methods (Table 1) in the six cell lines using the four (CA, H3K4me1, H3K4me3, and H3K27ac) or three epigenetic marks (omitting H3K27ac) as the features. Models trained on the CRM+TF⁺/Non-CRM sets or on the CRM+TF⁺/CRM+S⁻ sets outperformed significantly those trained the other pairs of datasets when three epigenetic marks (omitting H3K27ac) were used as the features; **p<0.001 (Wilcoxon signed-rank test). B Boxplot of expression levels of genes closest to the positive and negative sets defined by each method in the six cell lines. ^∆∆p<4.64×10⁻⁵⁶, **p<2.2×10⁻³⁰² (Mann-Whitney U test). C Boxplot of weights (coefficients) of the four epigenetic marks in the LR models for discriminating the positive and negative sets defined by each method in the six cell lines. **p<0.001 (Wilcoxon signed-rank test), the weights of the marks are significantly higher than those on the other three marks. D Performance of the seven types of classifiers trained on the CRM+TF⁺/Non-CRM sets or on the CRM+TF⁺/CRM+S⁻ sets in the six cell lines using the four epigenetic marks as the features