Figure 4

Enhancer predictions across cell types. (a) Barplots of Receiver Operating Characteristic (ROC) area under the curve (AUC) values for (i) cross cell-type predictions using models trained using data from a cell type other than the one being evaluated (blue, yellow, green, red, dark blue bars), (ii) a combined model trained on all other cell types excluding the evaluated cell type (gray bars), and (iii) 5-fold cross validation within the cell being under evaluation for comparison purposes (white bars). Combined models effectively predicted enhancers across 5 cell types, typically outperforming the models built from a single cell type. (b) Overlap of PEAS enhancer predictions with ChromHMM, and FANTOM enhancer definitions (top left) and with ChromHMM, and p300 binding enhancer definitions (bottom right) in CD4+ T cells. PEAS predictions significantly overlap with all three definitions. (c) Schema of our analyses to predict enhancers in EndoC-βH1 cell line using combined models by integrating data from GM12878, PBMC, CD14+, CD4+ T, and islet cells. (d) Precision Recall (PRC) curves for EndoC-βH1 enhancer predictions using (i) EndoC-βH1 data with 5-fold cross validation (blue) (ii) combined model (orange) and (iii) models trained only from an islet sample since that is the most relevant cell type among the five studied (red). Note that combined models perform similarly to islet-specific models. (e) Example region around IRX5 and CRNDE genes that show PEAS enhancer predictions based on combined models (marked with yellow bars). Note that the three PEAS predicted enhancers in this region are not enhancers in the cell types used for model training, however they are enhancers in EndoC-βH1. ChromHMM states in EndoC-βH1 are not used in model building, however shown in this figure to evaluate PEAS predictions. Shown ChromHMM states include enhancers (Yellow), promoters (Red), transcribed regions (Green), Insulators (Blue), Polycomb Repressed (Dark Gray), and Other regions (Light Gray).