Fig. 3

Comparison of TAD callers with Tan’s data [54]. a The radar plot shows the modularity and structure entropy of predicted TAD-like domains in chr1 in 32 cells (15 GM12878 cells and 17 PBMC cells). The cells were ordered by the modularity and structure entropy of deTOKI’s predictions. b,c Genome-wide distribution of ChIP-seq peaks of structural proteins (CTCF and RAD21) and histone marks (H3K4me3 and H3K36me3) flanking the single-cell TAD-like domain boundaries for the 16 GM12878 cells are shown in b and c, respectively. The shadows represent 95% confidential intervals, as calculated by bootstrap. The y-axis represents the mean number of peaks per bin with the same distance to the predicted TAD-like domain boundaries (MNPPB). The enrichment p values are calculated by the permutation test (n = 10,000). d deTOKI-predicted TAD-like domain boundaries match the boundaries predicted by 3D modeling [54]. The example shows the original matrix of radii of gyration for chr10, cell #11 of GM12878. The deTOKI-predicted allelic TAD-like domain boundaries are marked with vertical blue lines. e The AMI and WS between the 3D-modeled hierarchical TAD-like domains (as defined in [54]) and predicted allelic single-cell TAD-like domains in GM12878. f Significance levels of Pearson’s correlation coefficients between the number of contacts and the number of predicted TAD-like domains by deTOKI, deDoc, IS, and scHiCluster in each chromosome of 150 mESCs [55]. The threshold “P value = 0.01” is indicated by the horizontal blue line. g An example of mini-TAD-like domains predicted by deTOKI and IS. The mini-TAD-like domains in the circles are zoomed in as embedded sub-plots. The color codes for the four TAD predictors are all identical to those in a. *P < 0.05, **P < 0.001, NS: not significant, two-sided Wilcoxon rank-sum test