Fig. 2

Performance of enhancer classifiers in murine ESC and across different cell types and species. A Precision-recall curves for CRUP-EP (light orange lines) and REPTILE (light blue lines) trained on an mESC sample (mESC ⁺) and tested on ten randomly sampled independent test sets. The curves for the best performances are highlighted in darker colors (area under the curve AUC-PR: CRUP-EP =0.95, REPTILE =0.94). Additionally, the performance results of different ChromHMM segmentations for the same ten test sets are depicted (gray shapes). B ChromHMM emission probabilities for mESC using five and eight chromatin states, ranging from 0 (white) to 1 (dark blue). C CRUP-EP was trained on and applied to samples from different cell types and species (human hepatocytes (a–c), mESC (d), mouse adipocytes (e–h), mouse fibroblasts (i, j), mouse hepatocytes (k, l)). The result can be summarized in a 12×12 heatmap where each entry is shaded according to the computed AUC-PR (in percent). The origin of the training data can be found in the rows and the origin of the test sets in the columns. The diagonal shows the performance results on an independent test set within one sample. For instance, training and applying CRUP-EP in mESC ⁺ (highlighted in red) led to an AUC-PR = 0.93 based on the whole test set. D CRUP-EP was trained on samples from different cell types and species (see C) and applied to mESC ⁺. Shown are the number of predicted enhancers which are shared between all classifiers (“consensus”, orange) and which remain after excluding this consensus set (“without consensus,” blue). Additionally, mean probabilities are displayed for both classes, showing that all enhancer calls yield higher probabilities within the consensus set