Fig. 3

Prediction performance using cross-validation. a ROC plot for different models to predict chromatin looping interactions. The sensitivity (y-axis) is shown against the false discovery rate (1 – specificity, x-axis) for thresholds of the prediction score. Curves show averages of 10-fold cross-validation experiments. The models “Dist”, “Orientation”, and “Motif” contain only a single feature as indicated and all three genomic features are combined in the model “Dist+Orientation+Motif”. The models “RAD21”, CTCF”, “ZNF143”, “STAT1”, “EP300”, and “POLR2A”, contain the genomic features and the ChIP-seq correlation of the indicated factor. The model “all_TF” contains the genomic features and correlation of all indicated TFs. The model “across_TFs” contains the genomic features and a single correlation feature across the six ChIP-seq datasets as described in the main text. b PRC plot of precision against the recall for different prediction models. Color code as in (a). c Values of the area under the ROC (top) and PRC curves (bottom) as prediction performance. Error bars indicate standard deviation in 10-fold cross-validation experiments. d Example region on chromosome 11 in the genome browser showing: human genes, RAD21 ChIP-seq data in GM12878, CTCF motifs, CTCF motif pairs with that interact according to Hi-C or ChIA-PET data (green arcs) and predicted chromatin loops from RAD21 ChIPseq data using 7C (blue arcs). e Prediction performance of 7C as auPRC values for models with 124 TF ChIP-seq data sets from ENCODE. Error bars as in (c). The dotted horizontal line shows prediction performance of only the combined genomic features