Doublets in Single-Cell RNA-Seq: Why Two Cells in One Droplet Fabricate Cell Types That Were Never There

Zetobit · Bioinformatics Insight Series

Doublets in Single-Cell RNA-Seq: Why Two Cells in One Droplet Fabricate Cell Types That Were Never There

Single-cell RNA-seq promises one transcriptome per cell. But sometimes two cells land in the same droplet and get one barcode — and their blended profile doesn't look like an error. It looks like a discovery: a novel cell type, a rare intermediate, a transition no one had seen before.

By Kanna Nandakumar, PhD Zetobit LLC

The entire premise of single-cell RNA-seq is the word single. Each droplet is supposed to capture one cell and tag its transcripts with one barcode, so that each barcode reconstructs one cell's expression profile. The technology's power — resolving cell types, states, and trajectories that bulk RNA-seq averages away — rests on that one-to-one correspondence between barcode and cell.

Doublets break it. When two cells are co-encapsulated in a single droplet, their combined mRNA is tagged with the same barcode, and the analysis sees one "cell" with a transcriptome that is the sum of two. This is not a rare edge case to be waved away — it is a routine, rate-governed consequence of how droplet capture works, and the resulting artifacts are uniquely dangerous because they don't look like noise. They look like biology.

The rate is set when you load the chip

Doublet formation is governed by Poisson statistics: the more cells you load, the higher the chance any given droplet catches two.¹ For common droplet platforms the rule of thumb is roughly 0.8% doublets per 1,000 cells recovered — so about 4% at 5,000 cells and roughly 8% at 10,000.² That scaling creates a direct tension: recovering more cells per run, which everyone wants for statistical power and cost, also multiplies the doublet burden. "Super-loading" the chip is efficient precisely because computational doublet removal is expected to clean up afterward.³

The determining factor is the number of cells put into the machine, not the number finally called — so if recovery is lower than expected, the true doublet rate among recovered cells can be higher than the headline figure implies.² At a target of 10,000 cells, on the order of one in twelve barcodes may be two cells wearing one label. In a dataset of tens of thousands of cells, that is thousands of fabricated profiles.

Why doublets are worse than random noise

A doublet's transcriptome is a blend of its two constituents, and where that blend lands in expression space depends entirely on what got mixed. This produces two fundamentally different failure modes.

A homotypic doublet joins two cells of the same type. Its profile looks like a slightly larger version of a normal cell of that type — so it barely diverges from real singlets and is nearly impossible to detect from expression alone.⁴ These mostly inflate counts and library size without creating a visibly false population; they are the doublets that survive filtering.

A heterotypic doublet joins two different types — say a T cell and a B cell — and its blended profile expresses markers of both. In the clustering that defines cell types, it lands between the two parent clusters, where no real cell sits. Enough such doublets form a phantom cluster that reads as a novel intermediate cell type or a transitional state, complete with plausible-looking co-expression of two lineages' markers.⁵ This is the doublet that generates false discoveries.

Doublets combining two distinct cell types express both lineages' markers and cluster between the real populations, where no genuine cell resides. Enough of them form a phantom "intermediate" cluster that is easily mistaken for a novel or transitional cell type.

The failure mode is a false discovery, not a dropout. A phantom intermediate cluster co-expressing two lineages' markers is exactly what a real transitional cell state would look like. Without doublet-aware analysis, there is no way from the expression matrix alone to tell "newly discovered cell type" from "two cells that shared a barcode" — and the more novel and exciting the finding, the more important that distinction becomes.

How detection works — and where it's blind

Since most experiments have no ground truth for which barcodes are doublets, the dominant computational strategy is to simulate them: create synthetic doublets by averaging the profiles of random pairs of observed cells, then flag real cells that sit suspiciously close to these artificial doublets in expression space. This is the engine behind widely used tools like DoubletFinder, Scrublet, and scDblFinder.⁶

The approach works, but it inherits the same blind spot as the biology. Because it detects cells that resemble blends of different profiles, it identifies heterotypic doublets well — DoubletFinder reports over 90% sensitivity for them — but is largely insensitive to homotypic doublets, which don't diverge enough from singlets to flag.⁴ And it is not solved: a systematic benchmark found the best method reached a mean precision-recall AUC of only about 0.537 across real datasets, meaning even state-of-the-art detection leaves substantial error behind.⁷ Doublet removal does measurably improve downstream results — cleaner clusters, better differential-expression and trajectory inference — but the improvement varies by method and is far from complete.⁷

Two further catches matter in practice. Most detectors require an assumed doublet rate as input, which cannot be directly measured in a standard experiment and, if set wrong, mis-thresholds the calls.⁶ And a common, silent error is running detection on a merged multi-sample dataset without splitting by capture: the tool then treats it as one enormous load and infers an absurdly high doublet rate, corrupting the result.²

What actually works

Robust doublet handling combines experimental design with careful computation:

• Constrain the rate at the bench. Loading fewer cells lowers the doublet burden directly; the cost/power trade-off of super-loading should be a deliberate choice, not a default.¹
• Use experimental multiplexing where possible. Cell hashing (oligo-tagged antibodies) or genotype-based demultiplexing of pooled donors flags droplets carrying two samples directly — catching cross-sample doublets that expression-based tools miss.⁸
• Run simulation-based detection with the right settings. Apply DoubletFinder / Scrublet / scDblFinder per capture, not on merged data, and supply a defensible expected rate tied to cells loaded.⁶
• Leverage multi-omic signal when available. CITE-seq surface-protein and VDJ (paired-receptor) data can reveal impossible combinations — two distinct T-cell receptors in one "cell" — exposing doublets that transcriptome-only methods, especially for homotypic cases, cannot.⁵
• Treat novel intermediate clusters with suspicion. A population co-expressing two lineages' markers and sitting between their clusters is a doublet hypothesis until proven otherwise — validate with independent markers, orthogonal assays, or multiplexing before claiming a new cell type.⁵

As with the other blind spots in this series, the trustworthy analysis states its own guards: the expected doublet rate and its basis, whether detection ran per capture, which tool and threshold were used, and whether any surprising intermediate population was checked against the doublet hypothesis. A cluster reported without that scrutiny looks like a finding but may be an artifact of loading.

The takeaway

Doublets are the single-cell field's most characteristic artifact because they violate the assumption the whole method is built on — one barcode, one cell — and they do it in the most seductive way, by manufacturing populations that look like discoveries. The rate is set the moment you load the chip; the damage is worst exactly when two different cells merge into a plausible-looking intermediate; and detection, though genuinely useful, is incomplete and blind to same-type doublets by construction. The discipline is to constrain doublets experimentally, detect them carefully, and hold every novel intermediate population to the question the data can't answer on its own: is this a new kind of cell, or just two old ones sharing a label?

References

1. McGinnis CS, Murrow LM, Gartner ZJ. DoubletFinder: doublet detection using artificial nearest neighbors (Poisson doublet-formation rate from cells loaded; super-loading). Cell Syst. 2019. cell.com
2. scDblFinder documentation — ~1% doublets per 1,000 cells captured (≈0.8% per 1,000 loaded adjusting for recovery); determining factor is cells loaded; split multi-sample data by capture. Bioconductor. bioconductor.org
3. DoubletFinder and sample multiplexing as complementary; computational removal enables super-loading. Cell Syst. 2019. sciencedirect.com
4. DoubletFinder — >90% sensitivity for heterotypic doublets; insensitive to homotypic doublets (do not diverge from singlets). Cell Syst. 2019. sciencedirect.com
5. Double-jeopardy: scRNA-seq doublet/multiplet detection using multi-omic profiling (CITE-seq/VDJ reveal hybrid droplets; existing tools mainly catch heterotypic doublets). Cell Rep Methods. 2021. PMC8262260
6. scDblFinder: simulation-based detection (artificial doublets from averaged pairs; expected-rate parameter; per-capture use). F1000Research / Bioconductor. bioconductor.org
7. Xi NM, Li JJ. Benchmarking computational doublet-detection methods (best mean AUPRC ≈0.537; doublet removal improves clusters/DE/trajectories, variably). Cell Syst. 2021. sciencedirect.com
8. Cell hashing (Stoeckius et al.) and genotype-based demultiplexing (Kang et al.) for experimental doublet detection. Cell Rep Methods refs. 2021. PMC8262260

Zetobit LLC · Bioinformatics Insight Series · Lexington, KY · zetobit.com