Fig. 2

Predicting CTCF loops from ChIP-seq datasets. a Model of CTCF loops implied by CTCF motif orientation in loop anchors. RAD21: cohesin subunit. b Percentage of motifs for which motif orientation is concordant with loop anchor orientation (as depicted in a), for motifs in different score categories in MCF-7 and K562 cells. c Schematic illustration of CTCF loop prediction algorithm. Triangles represent CTCF ChIP-seq peaks, with color and size reflecting binding strength (peak scores). Black arrows represent CTCF motifs, with orientation indicated by arrow. Predicted loops are represented by double-headed arrows, with color reflecting predicted loop strength. Briefly, the loop prediction algorithm scans the genome from left to right connecting start anchors to end anchors based on motif orientation (1^st iteration, light green horizontal arrows). New loop sets are begun when new start anchors are encountered beyond end anchors (loop 3). Subsequently, weaker peaks (light green triangles) are filtered out and the genome is rescanned, identifying stronger, longer loops (loop 4, dark green arrow). This process is repeated with progressively higher peak thresholds up to a maximum threshold. d Example of CTCF ChIP-seq peak-based predicted CTCF loops (below) compared to experimentally determined CTCF ChIA-PET loops (above, data from two replicate experiments shown) in MCF-7 cells. Darker lines indicate stronger (predicted) loops. e Comparison of percentage of correctly predicted ChIA-PET loops with 1000 random control datasets in which motif orientation within ChIP-seq peaks is randomly permuted, for K562 and MCF-7 cell lines. Grey density plots depict the distributions of the random control percentages. Black diamonds indicate the actual percentages. f Comparison of percentage of correctly predicted Hi-C loops with 1000 random control datasets in which motif orientation within ChIP-seq peaks is randomly permuted, for 7 cell lines