Paper. Young MD, Behjati S. SoupX removes ambient RNA contamination from droplet-based single-cell RNA sequencing data. GigaScience. 2020;9(12):giaa151. https://doi.org/10.1093/gigascience/giaa151 (original), licensed CC BY 4.0. Below is a short summary in our own words; the full text is reproduced under the paper's licence.
SoupX removes ambient RNA contamination from droplet-based single-cell RNA sequencing data¶
What this covers¶
A method paper in computational genomics: how to detect and remove "ambient RNA" contamination — cell-free mRNA from the surrounding solution that gets sequenced alongside a cell's own transcripts in droplet-based single-cell RNA-seq — and a tool, SoupX, that does it.
The question¶
Droplet scRNA-seq (10X Chromium, Drop-seq, etc.) assumes each droplet's counts come from one captured cell. In practice every droplet also picks up free-floating mRNA already present in the cell suspension — lysed cells, secreted transcripts — so the observed counts are a mixture of genuine cellular expression and this "soup." The authors set out to show that this contamination is not a rare artefact but a pervasive, experiment-specific background, and that it can be large enough to distort the biological conclusions drawn from the data: a cell can appear to express a gene it never transcribed. The question was how to quantify and strip out that background without needing specialised experimental controls, so that any existing dataset could be corrected before standard downstream analysis.
The approach¶
SoupX works in three steps. First, it estimates the ambient expression profile directly from droplets with very few UMIs, which are assumed to contain no cell and so represent pure soup. Second, it estimates a contamination fraction for each cell (or channel): the proportion of that cell's UMIs attributable to the background rather than to its own transcription. This is the hard step, and it rests on finding genes/cells where true expression can be assumed to be zero — for example, haemoglobin genes in non-erythroid cells, or marker genes of one cluster in cells belonging to other clusters — so that any counts for that gene in that cell must be contamination. Where such a gene set isn't known in advance, SoupX offers an automated procedure that uses many cluster marker genes to independently estimate the contamination fraction and takes the most consistent (most frequent) value as the true one, reasoning that wrong estimates scatter while correct ones agree. Third, given the background profile and the contamination fraction, it subtracts the estimated background contribution gene-by-gene from each cell's observed counts (estimated via a multinomial likelihood rather than simple subtraction, to cope with sparse counts), producing a corrected count matrix that can be dropped into any standard downstream pipeline unchanged.
The method was validated using "species-mixing" experiments (human and mouse cells run together), where any mouse transcript appearing in a human cell's droplet is unambiguous contamination, providing ground truth against which to check the estimates.
What it found¶
Contamination was present in every dataset examined. In the species-mixing controls, roughly 1% of transcripts in a droplet were cross-species contamination at minimum, with the true total contamination (including same-species background) around 6% in the PBMC dataset tested; applying SoupX reduced cross-species (false) signal by a factor of 2 to an order of magnitude while leaving correct-species expression essentially unchanged. The background profile closely tracked the average expression profile across all cells in a channel (Pearson correlation 0.71-0.96 across datasets, median 0.86), supporting the assumption that contamination is roughly a uniform sample of the cell population rather than coming from one source. Contamination fraction was fairly constant across cells within a channel but varied substantially between experiments, consistent with it being a property of sample preparation rather than of individual cells.
Applying correction changed biological interpretation in concrete ways: in PBMC data it sharpened marker gene specificity and revealed additional markers not detected in uncorrected data; in a kidney tumour dataset it removed apparent collagen-gene expression from immune cell clusters that would otherwise have suggested (incorrectly) that those cells were tissue-resident, and it increased cross-sample mixing entropy, indicating that some of the apparent batch effect was actually background contamination; in fetal liver data it resolved cells that appeared, before correction, to co-express erythroid and non-erythroid markers in a way that looked like multiplets but was in fact ambient contamination.
Limits and context¶
The method assumes a single, roughly constant contamination fraction per cell or channel; it does not model contamination that varies gene-by-gene within a cell in ways unrelated to overall background abundance. The authors note that distinguishing contamination from genuine unexpected expression still requires judgement — they give an example (GYPA expression in a macrophage population) where SoupX's output supported treating the signal as genuine rather than contamination, but this depended on the user interpreting the estimate, not on the algorithm deciding automatically. Accurate correction is easiest when the user can supply biologically justified genes/cells known to have zero true expression; the automated alternative is offered for when this isn't available but is presented as a fallback, not the preferred route. The paper distinguishes SoupX from contemporaneous alternatives it discusses (SoupOrCell, CellBender, DecontX), arguing those either require mixed-genotype designs, carry a heavier computational cost and change the output format, or depend more heavily on clustering accuracy — but it reports no head-to-head quantitative benchmark against them in this paper.
Citation¶
Young MD, Behjati S. "SoupX removes ambient RNA contamination from droplet-based single-cell RNA sequencing data." GigaScience, 9(12):giaa151, 2020. https://doi.org/10.1093/gigascience/giaa151. Open access (CC BY 4.0); code at https://github.com/constantAmateur/SoupX.