Cell Composition in Bulk ATAC-seq: Why a Differentially Accessible Peak May Record Which Cells Were There

Zetobit · Bioinformatics Insight Series

Epigenomics · Study Design

Cell Composition in Bulk ATAC-seq: Why a Differentially Accessible Peak May Record Which Cells Were There

A bulk chromatin accessibility measurement is a weighted average over cell types, and the weights are free to move. When they do, the result is a peak list that reads as regulation and reports a census.

Kanna Nandakumar, PhD · Zetobit, LLC

A bulk ATAC-seq experiment on treated and control tissue returns four thousand differentially accessible peaks. They are enriched for coherent transcription factor motifs, they map to plausible pathways, and the effect sizes are large. The natural reading is that the treatment changed chromatin. There is a second reading that produces a numerically identical result: the treatment changed which cells were in the sample. Nothing in the peak list distinguishes them, because the assay never measured the thing that separates them.

The arithmetic has two free parameters

What a bulk ATAC library measures at a given region is a mixture. Each cell type in the tissue contributes fragments in proportion to how abundant it is and how accessible that region is within it:

signal at peak j  =  Σ_k  ( proportion of cell type k )  ×  ( accessibility of peak j in cell type k )

A differential test compares that quantity between conditions. The left-hand side can move because the second factor changed, which is regulation, or because the first factor changed, which is composition. One equation, two unknowns, and the reported output names only the second. This is not a subtle statistical point — it is visible in the model — and yet the convention for reporting bulk accessibility results has no field for it.

Two boundaries are worth drawing before going further. This is not the batch-confounding problem covered earlier in this series: batch effects are technical, they arise from a flawed design, and in principle a better design removes them. A composition shift is real biology, faithfully measured, in a study that may be designed perfectly. Nor is it the spatial deconvolution problem, where proportions are the declared object of estimation and the reference set is the thing under scrutiny. Here nobody is estimating proportions at all. The composition term is absorbed silently into a per-peak effect size and reported in the vocabulary of gene regulation.

Chromatin accessibility is unusually exposed to this

Accessibility is arguably the most cell-type-discriminating genomic layer routinely measured. That is precisely why it is used: single-cell chromatin profiling separates closely related populations because their regulatory landscapes differ sharply,¹⁰ and atlas-scale surveys of human tissues resolve open chromatin into modules that are largely cell-type-restricted.⁵ A peak set from a heterogeneous tissue is therefore dominated by elements that are open in some populations and closed in others — which is exactly the configuration in which a change of mixture produces a large change in bulk signal.

The property that makes ATAC-seq a good instrument for telling cell types apart is the same property that makes it maximally sensitive to a change in which cell types are present.

The field already ran this experiment on the adjacent assay

The most informative evidence does not come from ATAC-seq. It comes from DNA methylation, where the same mixture arithmetic applies and where the question was settled a decade ago in a way that should have travelled further than it did.¹

The setup was a literature reporting age-related methylation changes in whole blood. Using a reference-based deconvolution approach,² the authors established three things in sequence. First, the measurement is largely a cell-identity measurement: 63.5% of CpGs on the array differed across sorted blood cell types at P < 0.05. Second, blood composition changes monotonically with age, consistent with thymic involution — so the exposure of interest and the confounder move together by construction. Third, of the CpGs reported in the literature as age-associated, 86.7% showed significant differences across cell types.

Then the decisive test. In flow-sorted CD4+ T cells and monocytes — where composition cannot vary because it is held fixed by the sorting — no probes reached significance at a 5% false discovery rate for age. The effect that had generated a body of literature did not survive holding cell type constant.

One further result is the one I would put in front of any client, because it shows that the failure is not merely a false-positive rate. The authors ran functional enrichment on the reported age-associated sites, then reran it after removing sites associated with composition. Before filtering, 10 of the top 20 enriched categories were clearly immune-related and 3 were developmental. After filtering, 9 of the top 20 were developmental and only 4 immune. The biological story changed identity. A confounded analysis had produced a coherent, interpretable, publishable narrative about immune function — which is exactly the narrative that a shift in immune cell proportions would write. The confounder does not just add noise. It supplies its own interpretation, and the interpretation is internally consistent.

The obvious correction is not a correction

The reflex is to estimate the proportions and add them as covariates. That was tested in the same work. Adjusted estimates moved toward zero, but at this level of confounding the naive regression could not be shown to yield unbiased estimates, and the authors explicitly recommended against relying on regression adjustment. Removing unwanted variation with a factor-based method^8,9 attenuated the association considerably further — but still not completely.¹

The reason is geometric rather than algorithmic. When the condition of interest and the composition vector are strongly collinear, they occupy overlapping space in the design matrix and no post-hoc procedure can attribute variance between them. This is the same structure as the confounded-batch case, arriving from a different source: a correction applied to a confounded design does not recover the truth, and it can move the answer confidently in either direction. There is also a cost in the other direction — where confounding is minimal, adjusting for estimated proportions adds variance and loses power.

Where this bites hardest: pharmacodynamics

In a treatment study the composition shift is very often the mechanism of action. An immunomodulator that expands a lymphocyte subset, a cytotoxic agent that depletes proliferating cells, a compound that recruits myeloid cells into a tumour — each changes the mixture by design. Run bulk ATAC-seq before and after, and the readout is guaranteed to move, guaranteed to clear significance, and, unless someone intervenes, guaranteed to be described as chromatin remodelling.

The uncomfortable corollary is that the artefact scales with efficacy. A drug that works produces a larger composition change than one that does not, so the confounded signal is strongest in exactly the samples that look most encouraging. A pharmacodynamic biomarker built this way will replicate cleanly in the discovery cohort, hold up in a validation cohort drawn from the same tissue, and still not mean what the report says it means. It will be a good marker — of cell counts.

The same structure appears in disease cohorts. Alzheimer's brain tissue shows robust cell-type proportion differences, including loss of specific neuronal subpopulations,⁶ so any bulk brain accessibility comparison between cases and controls begins with a composition difference already present in the samples. A peak that looks like a neuronal regulatory programme switching off is the expected signature of neurons being less abundant.

Deconvolution helps, and is not the same as a solution

Reference-based tools for chromatin accessibility now exist and are worth running. EPIC-ATAC was built from 564 sorted ATAC-seq profiles into cell-type-specific marker peaks and reference profiles, and estimates malignant and non-malignant fractions from bulk tumour and PBMC samples.³ Cellformer goes further, separating whole-genome cell-type-specific accessibility from bulk, applied across 191 brain samples.⁴ Both make the invisible term measurable, which is a large improvement over assuming it away.

Two limits should be stated alongside. Reference completeness bounds the answer: proportions are estimated over the categories the reference contains, so a population absent from the reference is not reported as missing — it is redistributed into whatever resembles it. And measuring the confounder does not de-confound a collinear design. Knowing that treated samples contain 40% more of a cell type tells you the effect is confounded; it does not tell you what the remaining signal would have been at equal proportions. The value of deconvolution is diagnostic before it is corrective.

What actually happened	What bulk DA reports	What would separate it
A population expands or contracts	Many DA peaks, coherent motif enrichment, large effect sizes	Cell counts or deconvolved proportions on the same samples; sorted or single-cell comparison
Chromatin is remodelled within a cell type	The same output, indistinguishable on its face	Effect persists in sorted cells or in per-cell-type pseudobulk at matched proportions
Disease depletes a cell type	A lineage programme appearing to shut down	Proportions treated as a primary endpoint rather than a nuisance term
Dissociation or nuclei yield differs by arm	A composition shift with no counterpart in the living tissue	Both arms processed together; report input and yield per sample
Composition and condition are collinear	A confidently significant, uninterpretable result	Nothing post-hoc. This one is a design change

The first two rows produce output that is identical in form. Everything that distinguishes them comes from outside the bulk experiment — which is why the distinction has to be planned for rather than discovered in the results.

Bulk accessibility at a region is the sum over cell types of abundance times per-cell accessibility. A change in the first factor and a change in the second produce the same pileup, the same log fold change and the same adjusted p-value. Only measurements taken at cell-type resolution — counts, sorting, or single-cell profiling — distinguish which factor moved.

What to do about it

1. Measure composition rather than assuming it. Flow counts on the same specimens, or deconvolution against a matched reference, on every sample in both arms.
2. Report proportions as a result. In most treatment and disease studies the composition shift is a finding worth stating, not a nuisance to regress away.
3. Check whether the hit list is a marker list. If the DA peaks are enriched for lineage-defining elements of a population you expect to move, that is the diagnostic, and it is cheap to run.
4. Watch the coherence of the effect. Composition shifts move whole cell-type programmes together; targeted regulation is usually patchier. Suggestive rather than decisive, but informative early.
5. Do not lean on regression adjustment under strong confounding — and do not add proportion covariates reflexively when confounding is minimal, because that costs power.
6. Confirm in a cell-type-resolved subset. Sorting, or single-nucleus profiling on a few samples per arm, converts an untestable claim into a testable one.
7. Aggregate deliberately when you do go single-cell. Per-cell-type pseudobulk at matched proportions is the comparison that answers the regulation question.⁷
8. Design the arms to be processed identically and together — dissociation protocol, nuclei yield, operator, day. A composition difference created at the bench is indistinguishable from one created in the patient.
9. For pharmacodynamic endpoints, pre-specify which claim you are making. Composition change and regulatory change are different endpoints with different mechanisms, and the assay alone cannot arbitrate between them after the fact.

The coordinate carries the claim

A differential accessibility result is reported as a genomic location. That format is doing quiet work: a coordinate implies that something happened at that place in the genome, in cells, to DNA. It is a mechanistic claim encoded in a data format. But the arithmetic behind the number has two free parameters, and the coordinate names only one of them. A row reading chr7:1,234,000–1,234,500, log2FC 1.8, FDR 0.001 is entirely accurate about the signal and silent about the mixture that produced it.

None of this argues against bulk ATAC-seq, which remains the practical choice for cohorts of any size and gives up much less than the single-cell alternative in most respects. The composition question is answerable: references now exist for blood and for the common tumour and brain settings, flow counts are inexpensive where the tissue permits them, and a single-cell subset used as a control is a modest addition to a study budget. What is not defensible is leaving the question unasked, because the confounded answer arrives looking like the answer you wanted — coherent, mechanistic, and enriched for exactly the pathways the change in cell counts would predict.

The chromatin was accessible in the cells that happened to be in the tube. A bulk differential test asks which cells those were, and answers in the language of regulation.

References

1. Jaffe AE, Irizarry RA. Accounting for cellular heterogeneity is critical in epigenome-wide association studies. Genome Biology. 2014;15(2):R31. doi:10.1186/gb-2014-15-2-r31
2. Houseman EA, Accomando WP, Koestler DC, et al. DNA methylation arrays as surrogate measures of cell mixture distribution. BMC Bioinformatics. 2012;13:86. doi:10.1186/1471-2105-13-86
3. Gabriel AAG, Racle J, Falquet M, Jandus C, Gfeller D. Robust estimation of cancer and immune cell-type proportions from bulk tumor ATAC-Seq data. eLife. 2024;13:RP94833. doi:10.7554/eLife.94833
4. Berson A, et al. Whole genome deconvolution unveils Alzheimer's resilient epigenetic signature. Nature Communications. 2023;14:4947. doi:10.1038/s41467-023-40611-4
5. Zhang K, Hocker JD, Miller M, et al. A single-cell atlas of chromatin accessibility in the human genome. Cell. 2021;184(24):5985–6001. doi:10.1016/j.cell.2021.10.024
6. Consens ME, et al. Bulk and single-nucleus transcriptomics highlight intra-telencephalic and somatostatin neurons in Alzheimer's disease. Frontiers in Molecular Neuroscience. 2022;15:903175. doi:10.3389/fnmol.2022.903175
7. Teo AYY, Squair JW, Courtine G, Skinnider MA. Best practices for differential accessibility analysis in single-cell epigenomics. Nature Communications. 2024;15:8805. doi:10.1038/s41467-024-53089-5
8. Gagnon-Bartsch JA, Speed TP. Using control genes to correct for unwanted variation in microarray data. Biostatistics. 2012;13(3):539–552. doi:10.1093/biostatistics/kxr034
9. Leek JT, Storey JD. Capturing heterogeneity in gene expression studies by surrogate variable analysis. PLoS Genetics. 2007;3(9):1724–1735. doi:10.1371/journal.pgen.0030161
10. Buenrostro JD, Wu B, Litzenburger UM, et al. Single-cell chromatin accessibility reveals principles of regulatory variation. Nature. 2015;523(7561):486–490. doi:10.1038/nature14590

Zetobit, LLC · Bioinformatics consulting · Lexington, KY · zetobit.com
Related in this series: Batch Confounding in Public Data, Spatial Deconvolution, Sparsity in Single-Cell ATAC-seq.