Fig. 6

Impact of scHi-C normalization and de-noising methods on the compartment and domain detection for IMR90. Comparison of detected TADs and A/B compartments between bulk IMR90 Hi-C data and the aggregated single-cell Hi-C data after normalization or de-noising using Kim2020 data set with the known IMR90 cell type label. The numbers after the red squares at the left bottom or right upper corner of each contact matrix represent the minimum interaction frequency for the reddest locus pair. The black arrow highlights one example region that keeps showing over-imputation artifacts by Higashi across all five cell types compared to the bulk Hi-C data as a gold standard. First row: The insulation scores [36] that trace the TAD boundaries are depicted on the contact matrices with gray lines corresponding to bulk Hi-C data and purple for BandNorm, blue for scVI-3D, green for Higashi, and yellow for scHiCluster. Second row: A/B compartments are detected using the eigenvector of the correlation map of bulk (upper right triangles) or aggregated (lower left triangles) Hi-C matrices, values of which are displayed above each correlation matrix