Fig. 2

Classifier performance from different perspectives. a: Results (as Area Under ROC Curve) of RF classifiers trained on different subsets of training data, compared to EnhancerFinder results, a method developed by Erwin et al., based on SVM and using 4-mers and histone modifications along with TF-binding, DHS and evolutionary conservation. b: Number of important features selected with Boruta framework for classifier trained on 4-mers and Tier1&2 histone modifications. First row presents how many features where selected important by either heart or brain classifier, whereas second row presents how many of those features were important only for one of tissue-specific classifier c: Predictions by non-specific (trained on both heart and brain) classifier for genomic windows labeled as non-specific DHS, compared to non-DHS. Scores returned by classifier, can be interpreted as “probability of being enhancer”. d: Predictions for tissue-specific (fetal heart and fetal brain) DHS and non-DHS, returned by tissue-specific classifiers