researchsquare.comSource publication:
Loading content…
Single-cell 3D genome data are so sparse that domain boundaries are hard to call reliably, and positional disagreement across cells blocks comparison of boundary frequency and strength. CellRoB couples single-cell boundary detection with cross-cell alignment that minimizes an insulation-score loss, improving boundary accuracy and robustness to contact dropout on mouse brain Dip-C and human GM12878 scMicro-C data from 40 kb down to 5 kb. On multimodal mouse embryo data, boundary-strength profiles split the EPI cells that the RNA embedding left as one group.
Single-cell 3D genome studies can now give each cell a set of domain boundaries comparable across cells, instead of pooling cells into pseudo-bulk contact maps. Single-cell contact maps are extremely sparse, weak boundary signals may go undetected, and positional uncertainty makes it hard to tell biological heterogeneity from technical variability; even when imputation or representation learning shares information, downstream boundary positions remain inconsistent. On mouse brain Dip-C data at 40 kb, CellRoB-predicted boundaries showed stronger enrichment of CTCF and two other boundary-associated factors than alternative methods, the highest Jaccard index and measure of concordance between nearest-neighbor cell pairs, and, with DeDoc2, the highest reproducibility when 5% of contacts were randomly removed to build replicate datasets.
On high-resolution single-cell data with million-level contacts per cell, boundary counts do not explode as bins get finer. Boundary counts from HiCS, DeDoc2 and DeTOKI rose exponentially as bin size decreased, whereas CellRoB and Higashi did not show this pattern; Higashi is constrained by a predefined TAD-length parameter. On human GM12878 scMicro-C data binned at 40, 30, 20, 10 and 5 kb, CellRoB and Higashi achieved the highest measure of concordance, and CellRoB showed stronger enrichment of boundary-associated biological features than all other predictors even at 5 kb.
Rare boundaries present in only a small fraction of cells can be defined and analyzed separately while retaining cell-type information. Aggregating single-cell contact maps into pseudo-bulk strengthens recurrent signals but obscures low-frequency boundaries, while analyzing boundary calls independently makes it hard to tell whether nearby calls are one shared boundary or distinct features. Grouping aligned boundaries by cross-cell occurrence into rare, weakly conserved and strongly conserved classes gave distinct epigenetic enrichment patterns, with rare boundaries showing limited enrichment for CTCF, H3K4me3 and HK genes; UMAP embeddings from each class separated cell types to some degree, and genes near rare boundaries were enriched for nervous system development and neuronal differentiation.
Boundary strength itself carries cell-state information and can separate subpopulations without boundary gain or loss. In the HiRES mouse embryo data, the RNA-derived embedding showed epiblast and primitive streak (EPI) cells as a single homogeneous population, and binary boundary presence or absence also failed to subdivide them. Embeddings built from CellRoB boundary features recapitulated the neuroectodermal and mesenchymal trajectories and further split EPI into two subpopulations; their differentially expressed genes overlapped little with the highly variable genes used in the transcriptomic embedding, with Subtype 1 upregulated genes enriched for neural tube closure and Subtype 2 for spermatogenesis and limb development.
The demonstrated scope covers mouse brain Dip-C data at 40 kb, human GM12878 scMicro-C data binned from 40 kb to 5 kb, and mouse embryo HiRES data that jointly profiles chromatin conformation and gene expression; the intended reader is a researcher who needs comparable structural features from sparse single-cell contact maps.
The precise relationship of the two EPI subtypes to subsequent lineages remains to be established, and the suggestion that Subtype 1 may represent an early neural precursor state is consistent with, not proven by, the presence of neural tube cells in the data. Rare boundaries show weaker enrichment for canonical boundary-associated factors such as CTCF and H3K4me3, and the biological meaning of this warrants further investigation. The method depends on input coverage and preprocessing choices including imputation, bin size and insulation-window size; sharing information across cells may attenuate genuine structural differences in sparsely sampled populations, and the peak-region representation may merge nearby but biologically distinct boundaries.