Fig. 2

Evaluation of TAD calling methods. a Average ChIP-seq signal at TAD boundaries and surrounding regions (± 10 bins) (from left to right, CTCF, SMC3, and RAD21). b Proportions of Hi-C signal variability explained by the called TADs (measured by TAD-adjR²) at a different genomic distance between two interacting loci (average TAD-adjR²: OnTAD 0.33, Arrowhead 0.26, DomainCaller 0.26, rGMAP 0.23, TopDom 0.08, and TADtree 0.06). c–e Reproducibility of TAD boundaries (Jaccard index): c between two biological replicates (GM12878, 10 kb), d between resolutions 5 kb vs 10 kb) and 10 kb vs 25 kb, and e across different down-sampled sequencing depths (GM12878, original vs 1/4, 1/8, 1/16, and 1/32 of the original sequencing depth, raw data was used). Note: TADtree was not included in d, because it has difficulty handling data with 5-kb resolution due to its large memory consumption. It also has difficulty for chr1–3 at 10-kb resolution either. Thus, these three chromosomes were excluded for all TAD callers in all comparisons. All p values are calculated based on two-sided t test