Fig. 2: Twins delivers meaningful embedding distances for a range of different biological settings.

a Mean performance and separation index for independent Twins networks on three datasets; T-cell differentiation, ΔNIPBL and CTCF degron. b Normalised Hi-C maps at 10 kb resolution generated from ΔNIPBL versus control hepatocytes were split into regions measuring 2.56 Mb in size, and the resulting images were paired by genomic location. Twins embedding distances for conditions (orange) and replicates (blue) are plotted across test chromosome 2. Data are presented as mean values ± the 95% confidence interval. c As in (b) for neural progenitor cells under CTCF degron versus control conditions. d As in (b) for thymocytes at CD4 SP versus DP stages of differentiation. e Regions are categorised by the density of ChIP-seq peaks. For ΔNipbl data, regions are split into high and low (n = 2063, 13,154, respectively) and Twins scores are calculated for the two sets. Regions with high levels of RAD21 binding have higher Twins scores by two-sided t-test (p = 1.4e − 191). For Smc3 binding, regions are split into high and low (n = 1813, 13,404, respectively). Regions with high levels of Smc3 binding have higher Twins scores by two-sided t-test (p = 2.9e − 230). f For the CTCF degron data, regions are split into those with high and low CTCF binding (n = 4192, 10,043, respectively) and those with high and low H3K27me3 binding (n = 1312, 12,923, respectively). Regions with high levels of CTCF binding had significantly higher Twins scores by t-test (p = 2.3e − 37). but regions with high levels of H3K27me3 binding did not have significantly different Twins scores by t-test (p = 0.053). g The results of the T-cell network applied to an enhancer deletion in DN2 thymocytes. The Twins network is able to identify the region containing the enhancer deletion as having a differential chromatin conformation.