Ambient RNA in Single-Cell Data: Why a Cell Can Appear to Express a Gene It Never Transcribed

Zetobit · Bioinformatics Insight Series Single-Cell Genomics

Bioinformatics Insight Series

Ambient RNA in Single-Cell Data: Why a Cell Can Appear to Express a Gene It Never Transcribed

Dissociation ruptures cells, and their mRNA goes into the suspension. Every droplet that captures a cell also captures a sample of that free-floating pool, and those molecules are counted against the cell's barcode. The result is not random noise. It is a systematic addition, weighted toward whatever the most abundant cells in the tissue were transcribing, applied to every cell in the experiment — which means a marker gene appears at low level in clusters that never expressed it, and looks exactly like real low-level expression.

Kanna Nandakumar, PhD · Zetobit, LLC

The single-cell sub-series in this collection has covered four ways the analysis can mislead: doublets fabricating intermediate cell types, resolution parameters deciding how many cell types exist, annotation forcing novel populations into known labels, and integration erasing the biology it was meant to preserve. Each is a property of the analysis. This one is a property of the input, and it is upstream of all four.

Droplet-based protocols assume that the RNA in a droplet came from the cell in that droplet. The assumption is close enough to true that the whole field works, and false enough that it distorts results in ways that are easy to read as biology. Any cell-free RNA in the input solution is captured alongside the cell,¹ and the contamination is present even in datasets of the highest quality.²

What the soup is, and how much of it there is

The mechanism is unglamorous. Tissue dissociation — enzymatic, mechanical, or both — lyses some fraction of cells. Their transcripts enter the suspension. When the suspension is partitioned into droplets, each droplet receives a small volume of that solution along with its cell, and the ambient molecules are barcoded and counted as though they were endogenous.¹

The magnitude is the first surprise. The SoupX authors describe contamination as ubiquitous, varying between experiments from 2% to 50%, with around 10% reasonably common, and note there is no way to know in advance what it will be — solid tumors and low-viability samples run higher.³ The DecontX evaluation on species-mixing data found median contamination of 1.09% for human and 2.75% for mouse cells but a range spanning 0.43% to 45.09%, which is why per-cell estimation is necessary rather than a single global figure.⁴

Two properties make this consequential rather than merely a background subtraction problem.

The soup is not a random sample of the transcriptome. It reflects what the lysed cells were expressing, so its composition is dominated by the most abundant and most transcriptionally active cell types in the tissue. In brain single-nucleus data, where neurons outnumber other cell types and carry more transcripts, the ambient profile is predominantly neuronal — and the analysis showed that all glia are contaminated with neuronal ambient RNA.⁵ The contamination has a direction, and the direction points from the abundant population toward the rare one.

It is experiment-specific. Because the soup's profile depends on which cells lysed in that particular sample, it varies between channels and produces a batch effect.³ That connects directly to the integration piece in this sub-series: a technical signal that differs systematically between samples is precisely what integration algorithms try to remove, and precisely what they can over-remove.

Rare cell types are the most exposed, because the ambient profile is built from the abundant ones. The population you most want to characterize is the one most likely to be described in someone else's transcripts.

How it reaches the results

The failure is legible in a specific way: a marker gene turns up at low level in clusters that do not express it, at a level proportional to each cluster's share of the ambient pool rather than to its biology.

The clean demonstration is hemoglobin. Applying SoupX to kidney tumor data removed hemoglobin expression from all cells except red blood cells — the authors note this particular case is unlikely to be biologically misinterpreted, but that in other cases misattribution of expression to cell types that do not truly express a gene could lead to false conclusions.¹ Hemoglobin is the control case precisely because everyone knows a T cell does not make it. The problem is every gene where that prior is weaker.

The same reanalysis found that correction changed cluster membership for many cells and produced two fewer clusters at identical clustering parameters, while the expression ratio of marker genes between their own cluster and all other cells increased systematically.¹ Ambient RNA blunts the very contrast that clustering and annotation depend on.

Figure 1The shape of ambient contamination in a marker gene. Values here are schematic, drawn to illustrate the pattern rather than taken from a specific dataset; the hemoglobin case in kidney tumor data is the published example of the same structure.¹ The diagnostic feature is not that expression is high somewhere wrong, but that it is low and nonzero everywhere.

Where it becomes a biological conclusion

Two findings take this past the level of a preprocessing nuisance.

Differential expression and pathway enrichment inherit it

Because ambient transcripts are added to cells that did not express them, they can appear among differentially expressed genes and subsequently drive identification of significant biological pathways in unexpected cell subpopulations.⁶ Before correction, ambient RNA could make it seem that certain genes or pathways were active in the wrong cell types; after correction the false signals were reduced and analysis highlighted pathways specific to the right populations.⁷ A pathway result is several inferential steps downstream of the count matrix, and nothing in those steps flags a transcript that arrived from the suspension.

A whole annotated cell type can be an artifact

This is the strongest result in the literature, and it is worth stating carefully. Previous single-nucleus brain studies described "immature oligodendrocytes," marked by higher expression of many neuronal genes that were never functionally validated. On reanalysis, approximately 80% — 46 of 58 — of the immature oligodendrocyte markers overlapped with the top 200 most abundant ambient RNA markers.⁵

The cells sat between oligodendrocyte progenitors and mature oligodendrocytes in pseudotime, which is exactly where a transitional population should sit — and they were removed by ambient RNA cleaning. Crucially, their UMI counts were comparable to OPCs and mature oligodendrocytes, so they were not doublets.⁵ The doublet piece in this sub-series covers cells that blend two real transcriptomes; this is a different failure that produces a similar-looking artifact, and the UMI count is what distinguishes them.

The reanalysis also worked in the other direction. After removing the contamination, a genuinely rare transitional population — committed oligodendrocyte progenitors, marked by GPR17, BCAS1, and FYN — became detectable across datasets, at roughly 0.04% of cells in one adult cohort.⁵ Ambient RNA had simultaneously manufactured a false population and masked a real one.

Where ambient RNA lands in a single-cell analysis

Stage	What it does	Why it isn't caught
Count matrix	Adds transcripts the cell never made, proportional to soup composition	Counts are real UMIs from real molecules; no QC metric distinguishes source
Clustering	Blunts between-cluster contrast; changes membership and cluster number1	Output is a clean embedding either way
Annotation	Marker genes appear outside their lineage, weakening label confidence	Low-level expression is biologically plausible for most genes
Differential expression	Ambient transcripts surface as DEGs in the wrong populations6	Statistics are computed correctly on the counts provided
Trajectory inference	Contaminated cells occupy intermediate positions, mimicking transitions5	A smooth trajectory is the expected shape of the answer
Integration	Sample-specific soup composition behaves as a batch effect3	Correcting it and correcting biology look identical on a UMAP

Effects cited are from the studies referenced; the stage-by-stage mapping is the transferable point.

What correction can and cannot do

Several tools exist — SoupX, DecontX, CellBender, and more recent entrants — and they work by different routes. SoupX estimates the ambient profile from empty droplets and subtracts a contamination fraction.¹ CellBender uses a deep generative model that outputs an imputed profile rather than raw counts with background subtracted.¹ DecontX relies on clustering to estimate the background without requiring empty-droplet counts, which means its results depend heavily on the accuracy of the clustering supplied.¹

That last dependency is worth pausing on, because it creates a circularity the sub-series has described before in another form: clustering is affected by ambient RNA, and a correction method that takes clustering as input is estimating the contamination from a partition the contamination helped produce.

Three limitations deserve stating plainly.

Correction is incomplete. In the brain analysis, CellBender substantially reduced contamination but enrichment remained significant; a small subcluster of highly contaminated cells persisted and had to be identified and removed manually before the ambient signature disappeared.⁵ The tool got most of it. Most is not all, and the residue was concentrated rather than diffuse.

Methods disagree on the same data. Practitioners report substantially different contamination estimates from different tools on identical input — one comparison found SoupX estimating a mean around 1% per cell while DecontX estimated a much wider range on the same dataset.⁸ An estimate that varies by method is a parameter, not a measurement.

Over-correction is a real failure mode. Recent benchmarking found that no single method performs best across all datasets and metrics, and that some methods are more prone to over-correction in the absence of ambient RNA than others.⁹ A separate benchmark applying a count-integrity criterion — checking for gene-cell positions where corrected values exceed raw counts — found that certain tools do not merely denoise but restructure the count matrix, with one replacing the large majority of counts with values derived rather than observed.¹⁰ Removing transcripts a cell did not make is the goal; inventing ones it did not have is a different problem.

What is genuinely reassuring

The counterpoint matters, because a piece that stops at the failures implies single-cell data cannot be trusted, which is not the finding.

Correction demonstrably improves biological interpretation where contamination is present. Benchmarking found that several methods improve the quality of cellular neighborhoods in contaminated datasets while having negligible impact when applied to negative controls, and that marker gene expression becomes more focal — expressed by a more selective set of cells — relative to uncorrected counts.⁹ In the brain datasets, ambient removal consistently increased correlation with a physically sorted ground-truth dataset, indicating better reproducibility rather than merely different numbers.⁵

There is also a wet-lab lever. Physical sorting of nuclei by flow cytometry removed extra-nuclear ambient RNA, and sorting to deplete the dominant cell type before droplet capture removed its contamination from the remaining populations.⁵ Where the abundant population is known and dispensable, the problem can be addressed before sequencing rather than modelled afterward.

And the magnitude is sample-dependent in a way that is partly predictable: high-viability, cleanly dissociated samples carry less soup than necrotic tissue or solid tumors.³ This is not a uniform tax on all single-cell work.

What this means for a pipeline

Measure the contamination fraction, always

Even if you decline to apply a correction, you should know how contaminated the data are³ — it is a per-sample QC metric on the same footing as median genes per cell or mitochondrial percentage, and it varies enough between experiments that a project-level assumption is not safe.

Use known-negative genes as an internal check

The most reliable diagnostic is a gene that certain cells cannot express. Hemoglobin in non-erythroid populations, immunoglobulins in non-B populations, insulin and glucagon in non-islet pancreatic cells, and mitochondrial transcripts in single-nucleus data all serve;¹¹ Xist works for mixed-sex datasets, with contamination estimated from male cells.¹¹ These give a direct read on contamination without trusting a model.

Report the tool, version, and estimated fraction

Given that methods disagree,⁸ a corrected matrix is not reproducible without knowing which correction produced it. This is the same requirement the series has made for benchmark versions, splice-predictor thresholds, and germline filtering strategies: the parameter is part of the result.

Check whether correction inflated any counts

The count-integrity check is cheap and diagnostic: no corrected value should exceed the raw count for that gene in that cell.¹⁰ A method that produces values above the observed counts is doing something other than subtracting background.

Treat rare and intermediate populations with extra scrutiny

Rare cell types are most exposed, because the soup is built from the abundant ones.⁵ Before reporting a novel intermediate or transitional population, check whether its markers overlap the ambient profile, and check whether its UMI counts distinguish it from a doublet. The oligodendrocyte case is the template: the finding survived neither test.⁵

Order the preprocessing deliberately

Ambient correction, doublet detection, and clustering interact. Methods that consume clustering are affected by contamination-driven clustering,¹ and the sub-series has already covered how resolution choice determines cluster count. Fix and document an order rather than letting it emerge from whichever tutorial was open.

The shape of the problem

This series keeps returning to measurements that are correct about something other than what the reader assumes. A CH variant is genuinely somatic, from the wrong tissue. An FFPE artifact is a real chemical change, from the cassette rather than the patient. A deconvolution proportion is a real model output over a reference that lacked the cell type actually present.

Ambient RNA is the single-cell instance. Every UMI in the matrix corresponds to a real molecule that was really captured and really sequenced. The barcode records which droplet it landed in. What no part of the pipeline records is which cell transcribed it — and for a technology whose entire premise is the word single, that is the assumption everything downstream is built on.

The tools exist, the benchmarks exist, and the diagnostic genes are free. What is missing is the habit of treating contamination fraction as something you measure rather than something you hope is small.

Related in this series: the doublet piece covers the other way a droplet produces a transcriptome no single cell had — and UMI count is what separates the two artifacts; the clustering resolution and cell-type annotation pieces cover the steps ambient RNA degrades; the batch integration piece covers what happens when sample-specific soup composition is treated as a batch effect; and the spatial deconvolution piece covers the same "signal attributed to the wrong source" structure at tissue scale.

References

1. Young MD, Behjati S. SoupX removes ambient RNA contamination from droplet-based single-cell RNA sequencing data. GigaScience. 2020;9(12):giaa151. https://academic.oup.com/gigascience/article/9/12/giaa151/6049831
2. Young MD, Behjati S. SoupX removes ambient RNA contamination from droplet based single-cell RNA sequencing data (preprint version). bioRxiv. 303727. https://www.biorxiv.org/content/10.1101/303727v2.full
3. SoupX package documentation — contamination range, batch-effect behaviour, and the recommendation to quantify contamination regardless of correction. https://github.com/constantAmateur/SoupX (tool documentation, cited for the 2–50% range and practical guidance)
4. Yang S, Corbett SE, Koga Y, et al. Decontamination of ambient RNA in single-cell RNA-seq with DecontX. Genome Biology. 2020;21:57. Preprint with per-droplet contamination figures: https://www.biorxiv.org/content/10.1101/704015.full.pdf
5. Neuronal ambient RNA contamination causes misinterpreted and masked cell types in brain single-nuclei datasets. Neuron / PMC deposit. https://pmc.ncbi.nlm.nih.gov/articles/PMC9789184/ (preprint version: bioRxiv 2022.03.09.483658)
6. Understanding and mitigating the impact of ambient mRNA contamination in single-cell RNA-sequencing analysis. PLOS ONE. 2025;20(9):e0332440. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0332440
7. Summary coverage of ref. 6, including the pre- and post-correction pathway comparison. RNA-Seq Blog. https://www.rna-seqblog.com/understanding-and-mitigating-the-impact-of-ambient-mrna-contamination-in-single-cell-rna-sequencing-analysis/ (secondary source; cite ref. 6 for the primary finding)
8. Practitioner comparison of SoupX and DecontX contamination estimates on identical input. Biostars discussion. https://www.biostars.org/p/9595094/ (community source, cited as an illustration of method disagreement rather than as a formal benchmark)
9. Benchmarking computational decontamination of ambient RNA. bioRxiv. 2026. https://www.biorxiv.org/content/10.64898/2026.01.13.699237v1.full (preprint)
10. Benchmarking ambient RNA removal across droplet and well-plate platforms reveals artificial count generation as a critical failure mode of scAR and CellClear. bioRxiv. 2026. https://www.biorxiv.org/content/10.64898/2026.04.08.717130v1 (preprint)
11. Problems with ambient RNA. In: Multi-Sample Single-Cell Analyses with Bioconductor (OSCA). https://bioconductor.org/books/3.13/OSCA.multisample/ambient-problems.html