Enhancer classes display differential motif content that is predictive. (A) Cartoon of STARR-seq highlighting that a genomic-wide library of candidate fragments is tested for enhancer activity in a constant sequence environment (for details, see Arnold et al. 2013). (B) Definition of nonoverlapping enhancer subsets for subsequent parts of the analysis (schematic). (LOOCV) Leave-one-out cross-validation (see Methods for details). (C) Heatmap showing motif enrichments in four enhancer classes compared with negative control regions (neg). Shown are only motifs with significant enrichments in at least one of the four enhancer classes (FDR adjusted P-value ≤0.01 and fold enrichment ≥2); matrix cells with nonsignificant enrichment values (FDR adjusted P-value ≤0.01) are shown in white. (D) Receiver operating characteristic (ROC) plot for binary enhancer classification of all four enhancer classes versus the negative (dark colors) and positive (light colors) control sets using LOOCV. ([AUC] Area under the ROC curve.) Controls (gray) were performed by randomizing the sequences’ assignments to the enhancer or control groups (see Yáñez-Cuna et al. 2012).
