Fig. 2. Phase separation degeneracy explains variability of single-molecule conformations.

a In the considered 2.5 Mb wide locus chr21:34.6Mb-37.1 Mb of human HCT116 cells, the model median distance matrix in the globule separated state compares well against imaging data³⁷ (top, the Pearson and genomic-distance corrected correlations are, respectively, r = 0.95 and r′ = 0.84). The probability of a domain boundary in 3D conformations along the locus (middle, model-experiment correlation r = 0.79) and the corresponding separation score (model-experiment r = 0.85) also match. The location of ChIP-seq CTCF (orange circles) and cohesin (RAD21, blue circles) sites⁵⁸, and the intensity of the model four types of specific binding sites along the locus (bottom, as in Fig. 1b) are also shown. The vertical dotted lines are drawn to help comparing the panels. b Consistently, an all-against-all comparison of single-cell imaged³⁷ (top) and model predicted (bottom) 3D structures by the RMSD method shows that all imaged conformations statistically map onto model single molecules in the globule state (bottom, see text and Supplementary Figs. 5b, 6). The varying TAD-like domains result from the intrinsic conformational degeneracy of such a thermodynamic state. c The degree of variability of single-molecules is measured by the genomic-distance corrected correlation, r′, of pairs of distance matrices. The distribution of r′ between pairs of imaged structures (blue, average r′ = 0.27) is statistically not distinguishable from the r′ distribution between imaged and model distance matrices (dark gray, two-sided Mann–Whitney p value = 0.19). The model-model r′ distribution is in red and in light gray a control. d The model and experimental average boundary strength (error bars s.d.), and e the gyration radius distributions are also not distinguishable (two-sided Mann–Whitney p value = 0.40). Overall, the polymer globule phase separated state of the model returns single molecule structures with features consistent with both single-cell imaging and bulk Hi-C (Supplementary Fig. 4a) data.