🧬 biology

Identifying Robust Single-Cell 3D Genome Domain Boundaries with CellRoB

CellRoB is a computational framework that overcomes the sparsity of single-cell 3D genome data to robustly identify both common and rare TAD boundaries, thereby revealing biologically significant structural heterogeneity and enabling the resolution of cell subpopulations that are indistinguishable by transcriptomic data alone.

Original authors: Yusen Ye, Shiji Liu, Lin Gao

Published 2026-10-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yusen Ye, Shiji Liu, Lin Gao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside every cell of your body, the DNA is not a loose, tangled string but a tightly folded structure, organized much like a library where books are arranged on specific shelves. This three-dimensional arrangement is crucial because it determines which genes are turned on or off, effectively acting as the cell's instruction manual. Scientists have long known that this genome is divided into distinct neighborhoods called domains, separated by invisible fences known as boundaries. These fences ensure that the instructions for one part of the cell do not accidentally mix with those of another. However, when researchers look at a single cell rather than a crowd of them, the picture becomes incredibly blurry. The data is so sparse and full of gaps that it is nearly impossible to see where these fences actually stand, making it difficult to understand how individual cells differ from one another or how they decide to become specific types of tissue.

A team of researchers at Xidian University has developed a new computational tool called CellRoB to solve this problem of invisibility. Instead of trying to force a clear picture out of the noisy data, the researchers built a method that respects the natural variation between cells while still finding the common patterns. They started by filling in the missing pieces of the genetic contact maps for individual cells, using information from similar cells to make educated guesses about where the connections should be. Once they had a clearer view of each cell's internal structure, they identified the potential boundaries. The real innovation came next: they aligned these boundaries across thousands of cells, not by forcing them to match perfectly, but by finding the most likely position where a boundary exists for a group of cells. This approach allowed them to distinguish between boundaries that are present in almost every cell, those that appear only in a few, and those that are unique to a single cell.

The researchers tested their method on data from mouse brain cells and human cells, comparing their results against existing techniques. They found that CellRoB could identify boundaries with much greater accuracy and consistency, even when the data was missing a large number of contacts or when the resolution was extremely high, down to 5,000 base pairs. Unlike other methods that either missed rare boundaries or created too many false ones as the data became more detailed, CellRoB remained stable. It successfully identified boundaries that were shared by most cells, which are often associated with the cell's basic structural proteins, but it also uncovered a hidden layer of variation: boundaries that appeared in only a small fraction of the cells. These rare boundaries were not just noise; they carried specific information about the cell's identity and were linked to genes involved in specialized functions, such as the development of the nervous system.

Perhaps the most striking discovery came when the team applied CellRoB to a dataset of mouse embryos that included both genetic structure and gene expression data. In this dataset, a group of cells known as epiblast and primitive streak cells appeared as a single, uniform group when analyzed by their gene activity alone. However, when the researchers mapped these cells based on the strength of their domain boundaries, the group split into two distinct subpopulations. These two groups were not separated by the presence or absence of boundaries, but by subtle differences in how strong those boundaries were. The cells in one subpopulation showed genetic signatures suggesting they were heading toward becoming neural tissue, while the other group seemed biased toward non-neural lineages. This finding suggests that the strength of the genome's structural fences is a critical, yet previously overlooked, factor in determining a cell's fate. By revealing these low-frequency structural variations and quantifying the strength of the genome's architecture, CellRoB provides a new way to see the hidden diversity within living cells, offering a clearer view of how complex organisms develop from a single starting point.

Technical Summary: Identifying Robust Single-Cell 3D Genome Domain Boundaries with CellRoB

Problem Statement
Single-cell 3D genome technologies offer the potential to investigate chromatin organization beyond population averages, revealing structural heterogeneity. However, the extreme sparsity and technical noise inherent in single-cell contact maps (e.g., scHi-C, Dip-C, scMicro-C) obscure topologically associating domain (TAD) boundaries. This sparsity leads to two primary challenges:

Detection Reliability: Weak boundary signals are often undetected, and positional uncertainty complicates comparisons across cells.
Comparability: Even when boundaries are detected, positional discrepancies between cells hinder systematic comparison of boundary occurrence and strength.
Existing computational approaches (e.g., scHiCluster, Higashi, DeTOKI, DeDoc2) often focus on contact imputation or domain detection but fail to establish a consistent boundary representation across cells. Furthermore, aggregating data into pseudo-bulk maps strengthens recurrent signals but obscures low-frequency boundaries shared by only a small fraction of cells, limiting the characterization of rare structural variations.

Methodology: CellRoB Framework
CellRoB is a computational framework designed to robustly identify and compare single-cell chromatin domain boundaries while preserving cell-specific variation. The pipeline consists of four main stages:

Contact Imputation:
- The method first utilizes Higashi to generate a cell-cell similarity matrix ( $W$ ) based on 500 kb binned contact maps.
- For each cell, only the top $k=15$ nearest neighbors are retained.
- The raw 40 kb contact map ( $H_i^{raw}$ ) is imputed using a weighted average of the raw maps of its neighbors:
  $H_i^{imp} = \frac{H_i^{raw} + \sum_{j=1}^N W_{ij} H_j^{raw}}{1 + \sum_{j=1}^N W_{ij}}$
Initial Single-Cell Boundary Detection:
- Inspired by density-based clustering, CellRoB identifies boundary candidates in individual cells using two indices for each genomic bin $i$ $i$ :
  - Insulation Strength ( $\rho(i)$ ): Calculated as the normalized difference between intra-domain and inter-domain contact sums within a window ( $w=15$ bins at 40 kb).
  - Minimum Distance ( $\delta(i)$ ): The distance to the nearest bin with higher insulation strength within a range $M$ (20 Mb).
- These indices are normalized and combined into a score $\eta(i) = \rho'(i)\delta'(i)$ .
- A reflection point $r$ is determined to balance high scores with low rank. Bins satisfying $\eta(i) > \alpha \eta(r)$ (where $\alpha$ scales with bin size) are selected as initial boundaries.
Cross-Cell Alignment:
- To address positional discrepancies, raw boundary signals from all cells are aggregated and smoothed using a uniform kernel.
- Local valleys in the smoothed signal partition the genomic space into groups.
- For each group, a consensus bin is determined by minimizing the squared insulation-score loss across all cells that have a boundary within that group. This aligns boundaries to a common coordinate system without forcing a binary "present/absent" decision for every cell.
Conservation Classification:
- Aligned boundaries are categorized based on their occurrence frequency ( $f_i$ $f_{i}$ ) across the cell population:
  - Rare: $1\% \le f_i < 10\%$
  - Weakly Conserved: $10\% \le f_i < 50\%$
  - Strongly Conserved: $f_i \ge 50\%$

Key Results

Performance on Dip-C Data: Benchmarked on mouse brain Dip-C data (chromosome 6, 40 kb), CellRoB demonstrated superior performance compared to HiCS, Higashi, DeTOKI, and DeDoc2.
- Biological Validity: Predicted boundaries showed the highest enrichment for canonical factors (CTCF, H3K4me3, housekeeping genes).
- Consistency: CellRoB achieved the highest Jaccard Index (JI) and Measure of Concordance (MoC) when comparing boundaries between structurally similar cells.
- Robustness: It maintained the highest F1 scores across noise levels (5% to 40% contact dropout) and showed high reproducibility in replicate datasets.
High-Resolution Performance (scMicro-C): Evaluated on human GM12878 scMicro-C data across resolutions from 40 kb down to 5 kb.
- Unlike other methods (HiCS, DeDoc2, DeTOKI) where the number of predicted boundaries increased exponentially as bin size decreased, CellRoB remained stable.
- It achieved the highest MoC and strongest enrichment of biological features even at the 5 kb resolution.
- GO enrichment analysis of genes associated with different boundary classes in G1-phase cells revealed distinct functional profiles: rare boundaries linked to nervous system development, while strongly conserved boundaries linked to cell division.
Characterization of Rare Boundaries:
- CellRoB successfully detected boundaries present in only a small subset of cells, which are typically missed in pooled analyses.
- Rare boundaries carried distinct cell-type information (e.g., enrichment for neuronal differentiation genes) and exhibited different epigenetic signatures compared to conserved boundaries.
Resolving Developmental Subpopulations (HiRES Dataset):
- Applied to multimodal mouse embryo data (Hi-C + RNA-seq), CellRoB-derived boundary strength profiles resolved two subpopulations within the Epiblast and Primitive Streak (EPI) group that were indistinguishable in RNA-derived embeddings.
- Differential analysis revealed that these subtypes were distinguished not by the presence/absence of boundaries (binary features) but by quantitative variations in boundary strength (insulation scores).
- Subtype 1 showed enrichment for neural tube closure genes, while Subtype 2 showed enrichment for non-neural developmental processes, suggesting distinct developmental trajectories.

Significance and Claims
The paper claims that CellRoB provides a robust framework for linking reliable single-cell boundary identification to the discovery of low-frequency structural variation. Its primary contributions are:

Improved Comparability: By modeling recurrent boundary regions and aligning them via insulation-score minimization, it enables systematic comparison of boundary occurrence and strength across cells.
Recovery of Rare Features: It enables the definition and characterization of rare boundaries that are obscured by population averaging, revealing biologically informative structural heterogeneity.
Quantitative Structural Features: The study demonstrates that quantitative boundary strength, rather than just binary boundary presence, can serve as a structural feature to resolve cellular heterogeneity and developmental states that transcriptomic data alone cannot separate.

The authors note that while CellRoB improves consistency, it depends on input coverage and preprocessing choices (imputation, bin size). They acknowledge that rare-boundary classification relies on frequency thresholds and that independent experimental validation is needed to establish the physical significance of inferred boundaries, as no definitive gold standard exists for single-cell boundary annotations.

Technical Summary: Identifying Robust Single-Cell 3D Genome Domain Boundaries with CellRoB

More like this