Doublet Detection and Ambient RNA Correction

Doublet Detection and Ambient RNA Correction

Single-cell RNA sequencing experiments rely on the assumption that each droplet or well contains exactly one cell. In practice, two failure modes are common. First, doublets occur when two or more cells are captured in the same droplet and sequenced together, producing a "hybrid" transcriptomic profile that does not correspond to any real cell type. Second, ambient RNA contamination arises from mRNA molecules released by lysed cells during dissociation; these free-floating transcripts are captured in droplets that contain a real cell, adding spurious expression signals unrelated to that cell's identity. Doublets are not merely a nuisance — they can distort downstream analyses. Because a doublet inherits the expression profiles of two parent cells, it may appear as a novel, intermediate cluster or inflate apparent transition states in trajectory analyses. Doublet rates scale with cell concentration during droplet loading: at typical 10x Genomics loading concentrations, roughly 0.8% of events per 1,000 cells loaded are doublets, so a 10,000-cell dataset may contain 5–10% doublets. Computational detection exploits two signals: (1) doublets often have anomalously high total UMI counts and numbers of detected genes, and (2) their expression profiles are linear combinations of two distinct cell-type signatures. Tools such as Scrublet , DoubletFinder , and scDblFinder generate artificial doublets by averaging pairs of observed cells, then identify real cells whose profiles sit near these artificial doublets in transcriptomic space. Ambient RNA contamination presents a different challenge: the contaminating transcripts are present at low levels across many cells, but they become conspicuous in cell types that should not express them. The classic example is hemoglobin genes ( HBB , HBA1 ) contaminating immune cell clusters, which can lead to the false annotation of a "erythroid-like" population. Decontamination tools such as DecontX (implemented in the celda package) and SoupX estimate, for each cell, the proportion of counts derived from the ambient pool and subtract a corrected expression matrix. Importantly, ambient correction should be applied before clustering and annotation, whereas doublet removal is typically performed after initial dimensionality reduction. In the R/Bioconductor ecosystem, scDblFinder integrates cleanly with SingleCellExperiment objects, while DecontX operates on the raw count matrix. The following example demonstrates a typical doublet-detection workflow in R: library(scRNAseq) library(Scater) library(scDblFinder) # Load a reference dataset and compute quality metrics sce For ambient RNA correction, DecontX models each cell as a mixture of its native expression profile and the ambient RNA background, estimating the contamination fraction per cell: library(celda) library(SingleCellExperiment) # DecontX requires raw (unnormalized) counts counts When interpreting these tools, students should keep several caveats in mind. Doublet detection scores are probabilistic, not definitive; in datasets with continuous differentiation trajectories, genuine transitional cells can be misclassified as doublets, so thresholds should be tuned and flagged doublets inspected rather than blindly discarded. Conversely, ambient correction can over-correct if the ambient pool is poorly estimated, particularly in heterogeneous samples where some "ambient" transcripts actually reflect real biology. Best practice is to report the estimated doublet rate and contamination levels, run downstream analyses both with and without correction, and verify that key biological conclusions are robust to these QC steps. Doublets arise when two cells share one droplet; their rate increases with cell loading concentration and can reach 5–10% in typical experiments. Doublets are identified computationally by combining high UMI/gene-count outliers with transcriptomic profiles resembling artificial doublets (e.g., via Scrublet , DoubletFinder , or scDblFinder ). Ambient RNA contamination stems from lysed-cell transcripts captured alongside real cells and is corrected with tools such as DecontX or SoupX . Ambient correction should precede clustering and cell-type annotation; doublet removal is usually applied after initial dimensionality reduction. Classic diagnostic signatures include hemoglobin genes appearing in immune clusters (ambient RNA) and unexpected intermediate clusters (doublets). Always report estimated doublet and contamination rates, and confirm that biological conclusions hold with and without correction. Practice exercise: Download a publicly available peripheral blood mononuclear cell (PBMC) dataset from GEO or the 10x Genomics website. Using scDblFinder , estimate the doublet rate and identify the top 10 predicted doublets. Then apply DecontX to the raw count matrix and generate a UMAP colored by (a) predicted doublet status and (b) estimated contamination fraction. Write a short report (300–500 words) describing whether hemoglobin or mitochondrial gene expression signals in your clusters are better explained by ambient RNA contamination, and discuss whether removing the predicted doublets changes the number of clusters you observe.