Paper. Wang, W., Cen, Y., Lu, Z., Xu, Y., Sun, T., Xiao, Y., Liu, W., Li, J.J., & Wang, C. (2024). scCDC: a computational method for gene-specific contamination detection and correction in single-cell and single-nucleus RNA-seq data. Genome Biology, 25, 136. https://doi.org/10.1186/s13059-024-03284-w (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.
scCDC: a computational method for gene-specific contamination detection and correction in single-cell and single-nucleus RNA-seq data¶
What this covers¶
A computational method, scCDC, for finding and correcting ambient-RNA contamination gene by gene in droplet-based single-cell and single-nucleus RNA-seq data, for the field of scRNA-seq quality control and preprocessing.
The question¶
In droplet-based scRNA-seq and snRNA-seq, RNA leaked from lysed or ruptured cells floats free in the suspension ("ambient RNA") and gets co-encapsulated with unrelated cells, inflating their apparent expression of genes that are not really theirs. Existing correction tools (DecontX, SoupX, CellBender, scAR) all estimate a global contamination profile and subtract it from every gene in every cell. The authors ran into the problem directly while generating snRNA-seq data from mouse mammary glands: markers expected to be restricted to one cell type — milk-protein genes such as Wap and Csn2 in differentiated alveolar cells, adipocyte genes such as Acaca and Ghr — were turning up at low levels in almost every cell type. Applying the four existing tools revealed two opposite failure modes: DecontX, CellBender, and SoupX's automated mode left the heavily contaminated markers barely corrected, while SoupX's manual mode and scAR stripped real signal from housekeeping and lowly expressed genes that were not contamination sources. No prior evaluation had asked whether correction strength should match how contaminated a given gene actually is, and three of the four tools also require empty-droplet barcodes, routinely discarded before data are shared publicly.
The approach¶
Examining empty droplets across several datasets, the authors found ambient RNA is not spread evenly across the transcriptome but dominated by a small set of "super-contaminating" genes (milk- protein genes in lactating mammary gland, insulin genes in pancreas). This motivated a gene-specific strategy: identify which particular genes cause contamination, and correct only those.
scCDC runs in two stages on data already clustered into cell types. Detection: for each gene in each cluster, it compares the observed entropy of the gene's count distribution to an expected entropy predicted from a curve fitted across genes assumed non-contaminating. Ambient RNA counts are abundant but low-variance, so contamination suppresses observed entropy below expectation; this gap is the "entropy divergence." A gene whose divergence is significant in more than a tunable fraction of clusters (default 50%, the "restriction factor") is a "global contamination-causing gene" (GCG). Correction: for each GCG, clusters where it is not really expressed are pooled as a baseline, and a Youden-index threshold sets how many counts to subtract where it is genuinely expressed. A "contamination ratio" — its share of total UMI counts in the non-expressing clusters — quantifies how contaminated it is. Working from the cell-containing droplets themselves, scCDC needs no empty-droplet data, unlike SoupX, CellBender and scAR.
What it found¶
On simulated PBMC data with three artificially contaminated genes, scCDC correctly flagged all three, with contamination-ratio estimates tracking the true contamination level. Across the authors' own mammary gland datasets and thirteen public scRNA-seq/snRNA-seq datasets, scCDC identified GCGs (72 and 106 in the lactating and virgin mammary gland datasets) that were mostly known cell-type markers, matched the top contaminants found directly in empty droplets, and were confirmed as highly expressed by bulk RNA-seq; it also recovered previously reported contaminants elsewhere, including Ins1/Ins2 in pancreas, LYZ in a PBMC dataset, and cross-species reads in the mixed human/mouse "barnyard" benchmark.
Comparing correction across eight datasets with empty-droplet data retained, DecontX, SoupX's automated mode, and CellBender under-corrected the most heavily contaminated GCGs, while SoupX's manual mode and scAR stripped counts from many of 66 tested housekeeping genes in more than 95% of cells. scCDC corrected the GCGs without touching those housekeeping genes, a pattern that held in real data, in a pancreas dataset with spike-in-based ground truth, and in synthetic data spanning low, medium and high contamination. DecontX's under-correction tended to appear once a GCG's contamination ratio exceeded roughly \(3.16\times10^{-4}\); the authors recommend running scCDC followed by DecontX together when a dataset mixes highly and lowly contaminating genes, since scCDC can fail to flag some lowly contaminating ones. Downstream, correcting with scCDC let a gene-co-expression network method recover lactation marker genes (Wap, Csn2, Csn3, Glycam1) as a coherent alveolar-cell module, and recovered pancreatic islet markers (Gcg, Sst) as central genes of their respective modules — associations invisible before correction.
Limits and context¶
scCDC requires cells to already be clustered into types, a dependency it shares with DecontX but not with SoupX, CellBender or scAR; the paper treats this as an open issue, noting only that known cell-type identification did not appear much degraded by contamination in the datasets tested. The method can miss lowly contaminating genes and leaves them uncorrected, which is why the authors recommend scCDC plus DecontX in combination rather than claiming scCDC alone suffices. They argue against relying on the widely used mixed-species "barnyard" dataset as a sole decontamination benchmark, since its contamination level was the lowest of the fifteen datasets examined. None of the five compared methods, including scCDC, fully zeroed out contaminating counts where a GCG should not be expressed; a small residual remained in every case. DecontX's convergence behaviour may need further tuning, the authors note, calling this speculation rather than a tested finding, and flag single-cell proteomics as an untested future application.
Citation¶
Wang, W., Cen, Y., Lu, Z., Xu, Y., Sun, T., Xiao, Y., Liu, W., Li, J.J., & Wang, C. (2024). scCDC: a computational method for gene-specific contamination detection and correction in single-cell and single-nucleus RNA-seq data. Genome Biology, 25, 136. https://doi.org/10.1186/s13059-024-03284-w
Open access (CC BY 4.0), available at the DOI above and via PubMed Central (PMC11112958).