Hi-C is an assay that counts how often two pieces of DNA touch inside the nucleus. It is usually run on tissue that mixes many cell types. The resulting contact map is therefore an average, and differences in cell-type proportions between samples can masquerade as differences in DNA folding. Reference-free deconvolution splits bulk maps into per-cell-type components without needing pure reference profiles. THUNDER does this with two rounds of non-negative matrix factorization, and its authors report that it breaks down at fine resolution such as 5 kb, because the matrices become too sparse. That limit matters because enhancer–promoter loops, the contacts that link disease variants to genes, are invisible at the coarse 25–40 kb bins where the method works (HiView; corrected gene proximity map).
We propose to estimate cell-type structure where data are plentiful and refine it where data are scarce. A coarse bin-pair is a pair of large genomic windows. Each cell type's coarse contact count is learned at that scale. The model then asks only how that count is split among the fine 5 kb sub-pairs inside it. By default, every cell type shares the same split, and a cell type is allowed to deviate only where the bulk counts across samples demand it. Because contact counts add exactly across cells, the model can work on integer counts and assign each observed bulk count to latent per-cell-type counts. The claim is that this coarse-to-fine tying recovers cell-type-specific fine loops from tens of bulk samples. Neither single-resolution deconvolution nor a naive "fix proportions, then regress" pipeline can do that.
Core Claim
Coarse resolution is enough to estimate cell-type proportions. It is not enough to see loops, and at fine resolution there are too few counts per bin-pair to estimate a free profile for each cell type. A hierarchical count model makes fine profiles identifiable from realistic cohort sizes. In that model, fine profiles are a coarse profile times a split that is shrunk toward a cell-type-shared split. The novelty is the cross-scale parameter tying inside the deconvolution itself. Existing work handles scale or mixture, never both.
How This Relates to Prior Work
Three lines of work motivate the gap, and each handles only one half of it. THUNDER handles mixtures at a single resolution: it selects informative bin-pairs and factorizes them, and it is our main competitor and the source of the sparsity failure we target. MrTADFinder varies a resolution parameter to find nested domains in a pure sample. It shows that boundary biology is scale-dependent, for example in how H3K4me3 behaves across boundary scales, but it ignores composition. We borrow its distance-aware background model. RefHiC-SR imputes fine contacts from a panel of 30 reference maps. Its outputs are still bulk maps, and the file notes a risk of false loops when the target cell type differs from the panel. That constrains us to avoid external references for fine structure.
The method borrows two ideas. From Count Bridges, which deconvolves 1D transcript counts by forcing latent single-cell counts to sum exactly to the bulk counts, we take that exact-sum principle and extend it to 2D bin-pairs and across scales. We do not use its generative bridge process. From HiCNAtra, which normalizes contact maps with a Poisson regression on GC content, mappability, and copy number, we take the form of the bias terms. HiCNAtra assumes clonal cell lines. We restrict the main study to brain tissue, where copy-number changes are rare, and treat tumors as an explicit extension.
Model
Let $y_{s,p}$ be the bulk count for sample $s$ at fine bin-pair $p$, whose coarse parent is $P$. Let $d_s$ be sequencing depth, $b_p$ a known bias and distance-decay factor, $\pi_{s,k}$ the proportion of cell type $k$, $C_{k,P}$ the coarse contact rate of type $k$, and $w_{k,p}$ the fraction of $C_{k,P}$ falling in sub-pair $p$:
$$ y_{s,p} \sim \text{Poisson}\Big(d_s\, b_p \sum_k \pi_{s,k}\, C_{k,P}\, w_{k,p}\Big), \qquad w_{k,\cdot\mid P} \sim \text{Dirichlet}\big(\tau\, \bar w_{\cdot\mid P}\big). $$
Here $\bar w$ is the split shared by all cell types, estimated from pooled bulk data. The concentration $\tau$ sets how strongly each type is pulled toward that shared split, and it is learned per coarse region. Fitting uses expectation–maximization. In each round, every bulk count is assigned to cell types and sub-pairs in proportion to their current rates. Proportions and coarse rates are initialized from THUNDER at 40 kb. A cell-type-specific loop is called where the posterior of $w_{k,p}$ departs from $\bar w$ beyond a distance-matched background, using a SnapHiC-style local enrichment test.
Because each fine split only has to explain deviation from a shared pattern, the number of free parameters per cell type grows with the number of real cell-type-specific loops, not with the number of fine bin-pairs. That is the source of the expected gain at low sample counts.
Research Program
We will build pseudo-bulk mixtures from single-cell Hi-C with cell-type labels. The sources are mouse cortex Droplet Hi-C, with over 40,000 cells across 20 cell types, and the human prefrontal cortex single-nucleus methyl-3C data analyzed by SnapHiC, with 2,869 cells across 14 types. Ground-truth fine loops come from pooling all cells of a type and calling loops at 5 and 10 kb.
Each simulated cohort varies three things: the number of bulk samples (10, 25, 50, 100), depth per sample, and how unequal the proportions are, including minority types at 2–5%. Cell-line mixtures of GM12878, HAP1, and HeLa, the THUNDER setting, serve as a sanity check with very distinct components. The real-data test is the adult temporal cortex bulk Hi-C that THUNDER analyzed. There we test whether neuron- and oligodendrocyte-specific 5 kb loops are enriched near cell-type-specific enhancers and marker genes, compared with loops shared across types. The real-data result is supporting evidence only, because it has no ground truth.
Baselines
The naive reference point is to call loops on the averaged bulk map and assign every loop to every cell type. This gives a floor for cell-type loop precision and shows how much masking the mixture causes.
The strongest existing method is THUNDER run directly at 5 and 10 kb, as well as at its native resolution with coarse profiles upsampled. We use the authors' informative bin-pair selection, give it the same samples, and give it the same number of components chosen by the same criterion. For proportion accuracy, the file reports that THUNDER reduced mean absolute deviation by 23% to 36% relative to standard NMF and TOAST on simulated brain mixtures. No absolute error value and no fine-resolution loop recovery number for any bulk Hi-C deconvolution method was found on Lacuna, so we will establish THUNDER's numbers ourselves. For context on the single-cell route, SnapHiC captured over 91% of bulk loops using 742 mouse ES cells. Beating THUNDER means higher per-cell-type loop area under the precision–recall curve at 5 kb, with proportion error no worse than THUNDER's own coarse estimate.
The key ablation is a two-stage pipeline. It takes THUNDER's coarse proportions, holds them fixed, and solves a non-negative Poisson regression independently at every fine bin-pair with no cross-scale tying. A second ablation keeps the hierarchy but drops the shrinkage by setting $\tau$ near zero. A third replaces the Poisson count likelihood with least squares on normalized maps. All variants share the same bias correction, the same loop caller, and a fixed hyperparameter search budget chosen on held-out chromosomes.
Why not just…?
Why not just fix proportions from coarse THUNDER and regress each fine bin-pair? This is the real threat. With many samples and well-separated proportions it may match us, because each per-bin-pair regression is then well-posed. Our claim is limited to the regime where it should fail: few samples, collinear proportions, and minority cell types. In that regime a regression with $K$ unknowns per bin-pair on counts of zero or one is badly underdetermined. The two-stage ablation is designed to test exactly this. We will report the sample count at which its advantage disappears. If that count is around 25 samples, the contribution shrinks to a recipe, and we will say so.
Why not just super-resolve the bulk map first with RefHiC-SR, then deconvolve? Super-resolution produces a sharper bulk average, not cell-type maps. Its reference panel was built from other cell types, and the file notes that transfer to dissimilar types risks false loops. We will run this pipeline as an extra comparison and score false-positive loops in minority types specifically.
Why not just profile single cells with Droplet Hi-C? We use single-cell data as ground truth precisely because it is the gold standard. SnapHiC's result that HiCCUPS finds 2 loops in 75 cells shows that fine loops still require pooling many cells per type. Existing cohorts with clinical or post-mortem tissue are overwhelmingly bulk. The method targets that archive, not new studies.
Why not just smooth the fine bulk maps before NMF? Smoothing by random walk with restart, as SnapHiC does for single cells, borrows strength from neighboring bin-pairs regardless of cell type. That blurs exactly the cell-type-specific loops we want to separate. We include smoothed-input THUNDER as a baseline.
Decisive Checks
The proposal succeeds if, at 5 kb with 25 or fewer bulk samples, the hierarchical model gives clearly higher per-cell-type loop precision–recall than both THUNDER and the two-stage ablation for cell types under 10% abundance, on both mouse and human cortex mixtures. It fails if the two-stage ablation matches it across sample counts. In that case cross-scale tying adds nothing beyond coarse proportions.
A second check is calibration. The model should recover close to zero cell-type-specific loops when a mixture is built from two random halves of the same cell type. A third check is whether the learned $\tau$ is low in regions holding known cell-type-specific enhancers and high in regions with stable domains. That pattern would show the shrinkage tracks biology rather than noise.
References
- THUNDER: A reference-free deconvolution method to infer cell type proportions from bulk Hi-C data
- HiView: an integrative genome browser to leverage Hi-C results for the interpretation of GWAS variants
- The corrected gene proximity map for analyzing the 3D genome organization using Hi-C data
- MrTADFinder: A network modularity based approach to identify topologically associating domains in multiple resolutions
- Reference panel-guided super-resolution inference of Hi-C data
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Identification and utilization of copy number information for correcting Hi-C contact map of cancer cell lines
- Droplet Hi-C enables scalable, single-cell profiling of chromatin architecture in heterogeneous tissues
- SnapHiC: a computational pipeline to map chromatin contacts from single cell Hi-C data