Reconstructed by a model. cargnelli-2026-benchmarking-decontamination.pdf from papers/cargnelli-2026-benchmarking-decontamination — papers · cargnelli-2026-benchmarking-decontamination, licensed CC BY 4.0. Converted 2026-10-02 from .pdf. The original is a PDF with no usable text layer. A model read the pages and wrote this markdown: the prose is a paraphrase in places and every equation is unverified. Treat it as a pointer into the original, never as a citable source.

Figures

Figure 1: Setup for benchmarking ambient RNA removal methods.

Schematic diagram showing the benchmarking workflow. Three species-mixing, genotype-mixing, simulated, and negative control datasets were processed using 7 state-of-the-art ambient RNA removal tools and evaluated in terms of their ability to remove ambient RNA, conserve endogenous gene expression and effect on cell type labelling, clustering, and markers. Based on these, we provide recommendations for different methods depending on which types of data are available.

The diagram shows three columns: "Ground truth strategy" (Species mixing, with varying Complexity; Genotype mixing; Synthetic data; Control, Smart-seq2), feeding into "Computational decontamination tools" (CellBender, scAR, FastCAR, DecontX, SoupX, scCDC, CellClear), feeding into "Evaluation" (Remove ambient / retain endogenous; Neighborhood similarity, Marker genes, Cluster quality; Label transfer, Species labels).

Figure 2: Characteristics of ambient RNA.

A) Histogram or ridge plots of ambient load (fraction of total UMIs not derived from the cell) for each cell for the indicated datasets and samples. B) Scatterplot showing the ambient load on genes and their expression level in empty droplets. C) Scatterplot showing ambient load on genes and their endogenous expression frequency. D) Barplot showing the percentage of variance in ambient load across cells explained by cellular parameters, including the fraction of UMIs mapped to exons (Exons), mitochondrial genes (Mito), to protein-coding genes (PC), to ribosomal genes (Ribo), as well as the total number of UMIs (UMIs) and features detected (Feat).

The panel row A shows, from left to right, a histogram and three ridge plots of "Percent of cells" / "Samples" against "Ambient load (%)" for four datasets. Row B shows four scatterplots of log(ambient expr.) against log(background expr.). Row C shows four scatterplots of log(ambient expr.) against endogenous expr. freq. Row D shows four barplots of "Explained variance (%)" against cellular parameters (Exons, Feat, Mito, PC, Ribo, UMIs).

Figure 3: Method performance.

A) Dotplot of ambient removal scores (shown by color) and endogenous retain scores (shown by dot size) for all seven methods. For SoupX and DecontX, reduced mode (RED) runs were also included that only use filtered matrices across all seven datasets. B) Scatterplot showing the fraction of endogenous RNA retained and ambient RNA removed for all datasets. The color of the dot indicates the method. C) Scatterplot showing the fraction of endogenous RNA retained and the logged sum of UMIs for each gene after scAR correction in the low complexity species-mixing dataset.

Panel A is a dotplot with methods (CellBender, CellClear, FastCAR, DecontX, SoupX, DecontX\(^{RED}\), SoupX\(^{RED}\), scAR, scCDC) as rows and dataset categories (Low/Med/High species mixing, Strain mixing, Syn/SS2 synthetic, Negative control) as columns, with a color scale for "Ambient removal score" (Best to Worst) and a size scale for "Endogenous retain score" (Best to Worst). Panel B plots "Fraction endogenous RNA retained" (y-axis) against "Fraction ambient RNA removed" (x-axis), with points colored by method (SoupX\(^{RED}\), SoupX, FastCAR, DecontX\(^{RED}\), scCDC, CellClear, CellBender, scAR, DecontX). Panel C plots "Fraction endogenous RNA retained" (y-axis, from -1.00 to 1.00) against "log UMIs" (x-axis, 0 to 15).

Figure 4: Influence of ambient RNA on biological interpretation.

A) UMAP embeddings of a synthetic dataset showing the ground truth, as well as the observed data after ambient contamination (uncorrected) and after correction of the observed data for the indicated methods. Each dot was colored in cell type colors if correctly assigned using unsupervised clustering, or in red if misassigned. B) Barplot showing average module scores for marker genes defined in the ground truth dataset from A. C-F) Dotplots showing the average of the indicated metrics relative to the uncorrected data for contaminated or control datasets across four different tasks: cellular neighborhood similarity (C), marker gene selectivity (D), batch integration (E) and label transfer (F).

Panel A shows UMAP embeddings (labelled UMAP1 vs UMAP2) for: Ground truth, Uncorrected (ARI = 0.92), CellBender (ARI = 0.94), CellClear (ARI = 0.44), scAR (ARI = 0.87), DecontX (ARI = 0.91); and in a second row: DecontX\(^{RED}\) (ARI = 0.92), SoupX (ARI = 0.81), SoupX\(^{RED}\) (ARI = 0.81), FastCAR (ARI = 0.69), scCDC (ARI = 0.90). Cells are colored by labels HCC1500, HS578T, MCF12A, or Misassignment. Panel B is a barplot of "Average module score" (y-axis, 0 to 0.6) for Uncorrected, scAR, CellBender, SoupX, SoupX\(^{RED}\), DecontX, DecontX\(^{RED}\), CellClear, scCDC, FastCAR, with a dashed reference line near 0.5. Panels C-F are dotplots of methods (scAR, CellBender, SoupX, SoupX\(^{RED}\), DecontX, DecontX\(^{RED}\), CellClear, scCDC, FastCAR) against "Contaminated" and "Controls" columns, for metrics Purity/LISI (C), Entropy/Moran's I/Gini (D), LISI/bASW/PCR/ARI (E), and Purity/RMSE/cASW/ARI (F), each colored by "Relative to uncorrected" score.

Figure 5: Summary of benchmark.

Qualitative and quantitative characteristics of each tested methods summarized across datasets and tasks.

Method	CellBender	CellClear	scAR	DecontX	DecontX\(^{RED}\)	SoupX	SoupX\(^{RED}\)	FastCAR	scCDC
Runtime (h:mm:ss)	1:11:15	1:58:51	29:06	3:14	5:11	1:44	1:21	0:30	8:04
Uses GPU	Yes	No	Yes	No	No	No	No	No	No
Peak memory (Gigabytes)	22.54	50.56	27.19	10.75	7.45	15.02	12.99	5.39	29.53
Documentation	Excellent	Very poor	Excellent	Good	Good	Good	Good	Fine	Good
Language	Python	Python	Python	R	R	R	R	R	R
Input format	Raw	Raw / Filtered	Filtered*	Raw / Filtered	Filtered	Raw / Filtered	Filtered*	Raw / Filtered	Filtered*
Ambient removal	0.604 ± 0.328	0.100 ± 0.130	0.679 ± 0.240	0.391 ± 0.242	0.370 ± 0.278	0.297 ± 0.188	0.361 ± 0.224	0.198 ± 0.442	0.201 ± 0.446
Endogenous retainment	0.877 ± 0.084	0.662 ± 0.301	0.444 ± 0.200	0.890 ± 0.085	0.950 ± 0.049	0.939 ± 0.062	0.944 ± 0.058	0.927 ± 0.144	0.983 ± 0.046
Biological interpretation	0.524	0.527	0.473	0.741	0.505	0.396	0.440	0.167	0.428
Overcorrection score	0.712	0.439	0.137	0.804	0.836	0.818	0.842	0.420	0.720
Overall	1	9	7	2	4	5	3	8	6

---

← Methods and materials · Up: contents