Hi-C is a sequencing assay that measures which stretches of DNA touch each other inside the nucleus. It is the main tool for finding which distant gene a disease-linked noncoding variant actually regulates. Most Hi-C is still run on bulk tissue, a blend of millions of nuclei from many cell types. Suppose a researcher compares diseased and healthy brain and finds that an enhancer–promoter "loop" is stronger in disease. The usual reading is that disease rewired the genome. Often the true cause is different: the diseased samples contain more of the cell type in which that loop already exists. The same trap has been documented in expression genetics. When expression was modeled as varying with tumor purity, roughly 46% of standard expression–genotype associations turned out to be false positives in simulation, and only about 15% of 57,000 breast-cancer associations belonged to cancer cells (Cancer eQTLs … by modeling variation in tumor purity). Nobody has built the equivalent correction for chromatin contacts.
We propose a differential-contact test for bulk Hi-C that does two jobs. First, it removes changes that are explained by shifts in cell-type proportions. Second, it assigns each remaining change to the cell type in which it happens. The test combines three inputs: per-loop contact counts from bulk Hi-C, per-sample cell-type proportions, and a single-nucleus contact atlas. The atlas supplies two things, the candidate loops worth testing and each cell type's baseline contact level at those loops. Testing at the loop level rather than across all millions of bin-pairs is also what makes 5 kb resolution workable. Sparsity at 5 kb is the bottleneck the bulk Hi-C deconvolution method THUNDER reports.
Core Claim
For a fixed list of candidate loops, cell-type-resolved differential contact can be estimated from bulk Hi-C with a count model in which cell types mix linearly in contact-rate space. The claim is that this model controls false discoveries when composition is correlated with condition, and that it assigns real changes to the correct cell type. The specific way it should beat simply adjusting for proportions is attribution: knowing that a loop gained contact in microglia rather than oligodendrocytes. That distinction carries real stakes. Sorted brain Hi-C placed Alzheimer's inherited risk in microglia but its epigenetic changes in oligodendrocytes (Neuronal and glial 3D chromatin architecture), and a NeuN-negative "glia" pool merges the two.
The model works as follows. Let $y_{il}$ be the read count for loop $l$ in sample $i$, and let $s_i$ be that sample's depth. Let $p_{ik}$ be the proportion of cell type $k$ in sample $i$, $c_{kl}$ the baseline contact rate of loop $l$ in cell type $k$ taken from the single-nucleus atlas, $g_i$ the condition indicator, and $\beta_{kl}$ the condition effect on that loop within cell type $k$. We fit
$$y_{il} \sim \mathrm{NegBin}\Big(s_i \sum_k p_{ik}\, c_{kl}\, e^{\beta_{kl} g_i},\ \phi_l\Big),$$
where $\phi_l$ is a per-loop overdispersion parameter. The test is whether $\beta_{kl}=0$. Mixing is linear because bulk contact counts are sums over nuclei. The exponent keeps each cell type's rate positive.
Three features make the problem identifiable. Fixing the $c_{kl}$ from the atlas removes most free parameters. Proportions vary across samples. Effects are shrunk across loops with an empirical-Bayes prior. Even so, attribution is only possible for cell types whose proportion varies enough, or whose baseline contact $c_{kl}$ is large enough at that loop. We will report this as a per-cell-type power map rather than hide it.
Relation to Prior Work
Two papers supply the method. The purity-interaction eQTL model (Geeleher et al.) provides the core move: let a genetic or condition effect depend on composition so that it can be extrapolated to a pure compartment. Its own limits are that it works on one-dimensional expression, splits tissue into only two parts (cancer versus everything else), and uses a linear purity term. We carry the idea to pairwise contact counts with many cell types and an explicit mixing equation. SnapHiC (SnapHiC) calls loops from sparse single-cell Hi-C by smoothing each cell's contacts with random-walk imputation and testing across cells. On 2,869 human prefrontal cortex nuclei it found 817 to 27,379 loops per cell type across 14 cell types. We use those loop calls and their per-type contact levels as $c_{kl}$. SnapHiC itself cannot read bulk libraries or compare cohorts.
THUNDER constrains and motivates the design (THUNDER). It uses non-negative matrix factorization without a reference to estimate proportions and cell-type contact profiles from bulk Hi-C. It cut proportion error by 23–36% relative to standard factorization and TOAST, but it reports sparsity problems at 5 kb and has no differential test. We treat THUNDER as one source of $p_{ik}$ and as a baseline. We do not compete on proportion estimation itself.
UniversalEPI (UniversalEPI) predicts cell-type contact maps from DNA sequence plus accessibility data. It reaches Spearman correlations above 0.90 on unseen cell lines, but it assumes pure input. It is the main alternative that requires no Hi-C at all, and we address it directly below. Finally, count-respecting generative deconvolution (Count Bridges) shows that respecting integer counts beats continuous deconvolution such as CIBERSORTx and MuSiC on transcripts. That result motivates our negative binomial likelihood. We do not adopt a full generative bridge, because our goal is hypothesis tests with calibrated error rates, not sampled single-cell matrices.
Research Program
We will build the estimator in three parts. Each part feeds the next, and each is evaluated separately.
Candidate loops and baselines. Candidate loops and their per-type baselines $c_{kl}$ come from SnapHiC run on cortex sn-m3C-seq data, which profiles contacts and DNA methylation in the same nucleus. We pool counts within each loop's 5–10 kb anchor footprint, which turns millions of sparse bin-pairs into tens of thousands of informative features.
Proportion estimates. Per-sample proportions $p_{ik}$ come from three sources, which we compare. The first is reference-based regression of bulk loop counts onto the $c_{kl}$ matrix. The second is THUNDER. The third is paired bulk RNA when it exists.
Robustness. We will measure how much error in the proportions leaks into the effect estimates $\beta$. We will also test a version where $c_{kl}$ is learned jointly with $\beta$, rather than fixed, to see whether the atlas is actually necessary.
Baselines
The naive reference point is a standard per-loop negative binomial model that uses only condition and depth. This is what most bulk differential-loop analyses do in effect. Any composition-correlated design should inflate its false-discovery rate, and the size of that inflation sets the stakes. The closest published figure is the roughly 46% versus 3% false-discovery gap that Geeleher et al. report for expression. No reported result was found on Lacuna for differential chromatin contact calling under composition confounding.
The strongest simple competitor is the same model with proportions added as plain covariates. This removes confounding without attributing changes to cell types. A reviewer will rightly ask whether it already captures most of the benefit.
The strongest existing method is THUNDER followed by a comparison of its per-component contact profiles between conditions. THUNDER's best reported benchmark is a 23–36% reduction in proportion mean absolute deviation compared with standard factorization and TOAST. It reports no differential-contact numbers.
All baselines will see identical loop counts, identical proportion estimates where they use proportions, the same overdispersion fitting, and the same false-discovery procedure (Benjamini–Hochberg). Hyperparameter search budgets will be equal.
Two ablations isolate the key ingredient, which is the fixed per-type baselines from the atlas. The first estimates $c_{kl}$ freely. The second uses atlas loops but sets uniform $c_{kl}$.
Beating the competitors means two things. First, the method must hold nominal false-discovery control where the naive test does not. Second, among true changes, it must assign the correct cell type substantially more often than THUNDER-based comparison does, while matching covariate adjustment on overall detection power.
Why not just…?
Why not just sequence single-cell Hi-C in every sample? Single-cell contact data are mostly zeros. In 75 mouse embryonic stem cells, the standard bulk loop caller HiCCUPS found 2 loops where SnapHiC found 1,219 (SnapHiC). Cohort-scale single-cell Hi-C with enough nuclei per rare type, per person, is far costlier than bulk, and it cannot be run on archived bulk libraries. We use single-cell data once, as a shared atlas, rather than per sample.
Why not just sort nuclei before Hi-C? Sorting works, but in practice it produces coarse two-way splits, such as NeuN-positive versus NeuN-negative. That split merges microglia and oligodendrocytes, which is exactly the distinction behind the opposing Alzheimer's findings (Hu et al.). Sorting also needs suitable fresh tissue. We will use sorted data as ground truth rather than as a competitor.
Why not just add proportions as covariates? This is the real threat. If the scientific question is only "is there a cell-intrinsic change somewhere," covariates may be enough. The proposal survives only if attribution adds value. The planted-effect experiments below measure exactly this. If attribution accuracy is no better than assigning each loop to the cell type with the largest $c_{kl}$, we will report that the simpler method suffices.
Why not just predict cell-type contacts with UniversalEPI from per-sample single-cell ATAC? This route needs per-sample single-cell accessibility, which is itself costly. It can also only detect contact changes that are mirrored in accessibility; architectural changes at loci whose accessibility stays the same are invisible to it. We include it as an extra comparator on the semi-synthetic data, using pseudobulked accessibility for each cell type, and report the loops where the two methods disagree.
Proposed Experiments
Simulated cortex mixtures. Starting from the 2,869 prefrontal cortex nuclei, we build synthetic bulk samples by resampling nuclei under controlled proportions. We then plant condition effects on chosen loops in chosen cell types, and vary three factors: cohort size (20 to 100 samples), the correlation between composition and condition, and sequencing depth. We hold out half the nuclei when estimating $c_{kl}$ so that the atlas and the "bulk" do not share cells. The metrics are false-discovery rate, power for each cell type, and the fraction of discovered changes assigned to the correct cell type.
Real mixing with real biology. This is the decisive semi-synthetic check. We mix reads from the THP-1 macrophage activation Hi-C time course with GM12878 Hi-C from ENCODE at known ratios. Macrophage proportion is made to rise with activation time. True differential loops are defined from the pure THP-1 series. The method should recover those loops, attribute them to the macrophage component, and reject loops that only reflect the changing GM12878 share. The naive test should not achieve this.
Negative control. We run the method on the unmixed THP-1 time course, where composition is constant. Here the method should agree with the naive test, which shows it does not invent attributions.
Real tissue. In the adult temporal cortex bulk Hi-C analyzed by THUNDER, we will check whether loops attributed to neurons and to glia overlap the neuron-specific and glia-specific interactions found by sorted Hi-C (Hu et al.) more than random loops matched for distance do.
We will judge the proposal by whether attribution beats covariate adjustment in the THP-1 mixture experiment. If it does not, the contribution reduces to a calibrated confounding correction, and we will present it as only that.
References
- Cancer expression quantitative trait loci (eQTLs) can be determined from heterogeneous tumor gene expression data by modeling variation in tumor purity
- Neuronal and glial 3D chromatin architecture illustrates cellular etiology of brain disorders
- SnapHiC: a computational pipeline to map chromatin contacts from single cell Hi-C data
- THUNDER: A reference-free deconvolution method to infer cell type proportions from bulk Hi-C data
- UniversalEPI: robust prediction of cell type-specific and differential chromatin interactions from DNA sequence and chromatin accessibility
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data