Fig. 2

Comparison of TAD callers on downsampled and simulated single-cell Hi-C based on data from IMR90 [14]. Panels a and b show the average results of 20 independent downsamplings in each chromosome. a The (log2) change in the number of predicted TAD-like domains. b The similarity of TAD-like domains, as inferred by AMI and WS, between the raw data and the downsampled data. c Workflow of the single-cell Hi-C simulation. From left to right, the panels represent the normalized Hi-C contact matrix of chr18:50–55 Mb for GM12878 ensemble Hi-C from Rao’s data [18], an ensemble of 100 modeled 3D structures of this region, and the 3D structure modeled from the simulated ensemble Hi-C from model #100. Each dot in the right panel represents a 10 kb-length particle, and the dots with same color belong to the same predicted TAD-like domain ensemble. d Similarities of predicted single-cell TAD-like domains between different thresholds and predictors. e An example of the simulated data. The upper and lower parts of the heatmap represent the simulated reference and single-cell Hi-C data from model #13, D = 500. Predicted TADs are shown in sawtooth. AMIs between TAD-like domains predicted by deTOKI and IS on the two datasets are 0.873 and 0.660, respectively. f Classification based on deTOKI-predicted TAD-like domains of models on chr18:50–55 Mb and chr18:10–15 Mb, mimicking two single cells. Each dot represents a model, D = 500. g Number of misclassifications, using predicted TAD-like domains. *P < 0.05, **P < 0.001, NS: not significant, two-sided Mann-Whitney U test