Paper. Cargnelli, C. B., Nielsen, J. V. & Madsen, J. G. S. (2026). Benchmarking computational decontamination of ambient RNA. bioRxiv preprint, posted April 1, 2026. https://doi.org/10.64898/2026.01.13.699237 (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.
Benchmarking computational decontamination of ambient RNA¶
What this covers¶
An independent benchmark of computational methods for removing "ambient RNA" contamination from single-cell and single-nucleus RNA-sequencing data, speaking to the practice of sxRNA-seq preprocessing in genomics.
The question¶
Single-cell and single-nucleus RNA-seq data are contaminated by ambient RNA: molecules released into solution, chiefly by cell lysis during harsh sample preparation, that get captured in droplets alongside a cell's own transcripts and are misattributed to it. This distorts expression measurements, confounds cell-type classification, and biases downstream conclusions. Several computational decontamination tools exist, but no independent study had evaluated them side by side on comparable data — leaving developers without a shared standard and users without a basis for choosing a tool, or for judging whether one is warranted at all. The paper also asks a second question: since decontamination tools are now run routinely inside standard pipelines (e.g. nf-core), does applying "correction" to data that is not actually contaminated introduce its own distortion?
The approach¶
The authors benchmarked seven methods — CellBender, DecontX, FastCAR, scAR, scCDC, SoupX, and CellClear — run with default parameters, across four dataset categories chosen to supply different kinds of ground truth: three species-mixing experiments of increasing complexity (from a simple human-mouse mixture to human islets xenografted into mouse kidney), where reads aligning to the "wrong" species establish genuine contamination; a strain-mixing experiment using genotype-specific SNPs between mouse strains; synthetic datasets built by sampling from negative-binomial models fitted to real cell lines with a specified ambient fraction mixed in; and negative-control datasets with no contamination (including a non-droplet Smart-seq2 dataset), used to test for over-correction.
Each method-dataset pair was scored on an "ambient removal score" and an "endogenous retain score" (0–1 scale). The authors then traced the downstream consequences of decontamination through standard analysis tasks — unsupervised clustering, marker-gene scoring, batch integration (Harmony), and reference-based label transfer (Azimuth) — to see whether cleaning the data actually improved biological interpretation, not just the raw counts.
What it found¶
Ambient RNA is pervasive: across 97,357 real cells and nuclei from 49 samples, 81.56% showed detectable contamination, higher on average in single-nucleus (11.80%) than single-cell (7.79%) preparations. Contamination is dominated by a small number of ubiquitous, highly expressed mRNAs — the 10% most frequently expressed genes accounted for 82.30% of ambient signal on average — and its per-cell level was only weakly predicted by standard quality metrics, so stringent filtering reduces but cannot eliminate it.
No single method won across every dataset and metric. scAR and CellBender gave the strongest raw ambient-removal scores, but scAR also stripped out a large share (44.5% on average) of genuine endogenous signal, especially from lowly expressed genes, and could degrade batch integration and label transfer even on uncontaminated data — a strong tendency to over-correct. scCDC and FastCAR removed ambient RNA mainly from already highly expressed genes, limiting collateral damage but also yielding little improvement in interpretation. CellClear showed the opposite skew, removing signal mainly from lowly expressed genes, and stripped out nearly as much endogenous as ambient RNA. DecontX and SoupX sat in the middle, preserving endogenous RNA well and consistently improving downstream interpretation with comparatively low over-correction risk.
The authors recommend CellBender or DecontX (full mode, unfiltered matrix) when contamination is expected and the full matrix is available, and SoupX in reduced mode when only a filtered matrix can be obtained or there is no strong prior expectation of contamination, since it was least prone to over-correcting clean data. CellBender had the best overall weighted performance but the heaviest compute footprint (GPU, more memory). A general finding: applying any of these tools is not a safe default — several methods altered expression and biological scores even with no contamination present, so decontamination should be guided by prior evidence of ambient RNA, not applied automatically.
Limits and context¶
The authors note that each ground-truth strategy has a distinct limitation. Synthetic data, built from fitted per-gene distributions, lacks higher-order structure such as gene-gene covariance present in real data. Species-mixing gives a robust ground truth but places endogenous and ambient RNA in non-overlapping feature spaces (different species' genes), unlike real contamination where both compete for the same genes. Strain-mixing via SNPs is accurate but only measures contamination at genetically variable genes, a limited subset. The paper argues that further progress is constrained less by algorithms than by the lack of experimental designs where ambient and endogenous RNA share the same feature space while contamination can still be measured without bias — a gap not yet closed. Results and code are released as an open, extensible resource meant to be rerun as new methods and datasets appear, rather than as a final verdict.
Citation¶
Cargnelli, C. B., Nielsen, J. V. & Madsen, J. G. S. (2026). Benchmarking computational decontamination of ambient RNA. bioRxiv preprint, posted April 1, 2026. https://doi.org/10.64898/2026.01.13.699237 (preprint, CC BY 4.0 licence).