henhaobio
Sign in

Cell SystemsBioinformaticsPeer-reviewed journal article

Scrublet: computational identification of cell doublets in single-cell transcriptomic data

Samuel L. WolockiD, Romain LopeziD, Allon Moshe KleiniD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Single-cell RNA-sequencing has become a widely used, powerful approach for studying cell populations. However, these methods often generate multiplet artifacts, where two or more cells receive the same barcode, resulting in a hybrid transcriptome. In most experiments, multiplets account for several percent of transcriptomes and can confound downstream data analysis. Here, we present Scrublet (Single-Cell Remover of Doublets), a framework for predicting the impact of multiplets in a given analysis and identifying problematic multiplets. Scrublet avoids the need for expert knowledge or cell clustering by simulating multiplets from the data and building a nearest neighbor classifier. To demonstrate the utility of this approach, we test Scrublet on several datasets that include independent knowledge of cell multiplets.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (9 Jul 2018) · CC BY 4.0

Scrublet detects scRNA-seq doublets by simulating doublets from random cell pairs and scoring cells by simulated-doublet neighbor fraction, reaching AUC 0.99 on human-mouse mixtures.

Multiplets make up several percent of droplet and well-based scRNA-seq captures and can create spurious clusters or bridges; experimental detection is not applicable to every design, and prior heuristics (high total counts, manual marker curation) assume similar RNA content or need expert annotation. Scrublet offers a clustering-free, data-driven classifier that also estimates what fraction of doublets are detectable in a given embedding.

Key results

  1. ShownOn a 10x mixture of human HEK293 T and mouse NIH3T3 cells with species labels hidden, Scrublet recovered cross-species doublets almost perfectly, and its detectable doublet fraction was close to the theoretical expectation.

    ROC AUC 0.99 (recall 98%, precision 96%) vs 0.88 for total transcript counts, which fell to 15% precision at 90% recall; ϕD = 54% vs 50% expected for cross-species doublets; 6,164 human, 5,915 mouse cells and 741 mixed multiplets; 15,642 shared-name genes. · Performance on dataset #1: human-mouse cell…; Pre-processing and doublet detector parameters

  2. ShownOn PBMCs pooled from eight genotyped donors with demuxlet genotype-based doublet labels, Scrublet outperformed marker co-expression and total-count heuristics, and ϕD tracked the actual recall across thresholds; a hybrid of Scrublet and local total counts scored highest.

    AUC 0.88 vs 0.77 for marker co-expression and 0.81 for total counts; Scrublet plus local total counts hybrid 0.93. At the chosen threshold, recall 49% vs ϕD 55%, precision 66%; 89% of doublets with ground-truth neighborhood score >0.4 recovered at 77% precision; 68% of 773 detected true doublets showed high marker co-expression. · Performance on dataset #2: peripheral blood…; Pre-processing and doublet detector parameters; Ground truth doublet score; Hybrid doublet score (Scrublet + total counts)

  3. ShownIn Splatter simulations, Scrublet found nearly all between-cluster (Type B) doublets when clusters were well separated. It failed on poorly separated groups and on within-cluster (Type A) doublets, and its estimated detectable fraction predicted the recall it actually achieved.

    Up to 99% of between-cluster doublets recovered at 99% precision; recall under 10% for poorly separated groups; branching continuum: up to 92% of between-branch doublets (>10% toward opposite endpoints) at 98% precision; 10 replicates of 5000 cells and 2000 genes, 5% doublets, k = 40, r = 5. · Performance on simulated data; Splatter simulations

  4. SuggestedWhen one PBMC sample was loaded at two cell concentrations, Scrublet called about twice as many doublets at the higher concentration. Its predicted doublet states grew in relative abundance while singlet states stayed flat or shrank.

    Predicted doublets 4.1% in PBMC-8 k (8,391 cells) vs 1.8% in PBMC-4 k (4,352 cells); mapping PBMC-8 k cells onto 30 spectral clusters of PBMC-4 k showed Scrublet-flagged clusters gaining relative abundance, while singlet clusters changed little or decreased; d = 0.03 and 0.06. · Performance on dataset #3: peripheral blood…; Pre-processing and doublet detector parameters; Mapping PBMC-8 k to PBMC-4 k

  5. SuggestedIn Kit+ mouse bone marrow progenitors forming a differentiation continuum, Scrublet flagged doublet bridges between committed progenitor lineages, though no ground truth was available. It missed erythroblast-macrophage aggregates because the macrophage parent never appears as a singlet.

    4,273 cells analyzed; 146 cells with score >0.28 removed, a second pass removed 34 more and a third found none; flagged cells co-expressed lineage markers (e.g., Car1 with Mpo or Ngp), while only some had above-average counts; 37 C1qa+ Hba-a1+ cells were missed. · Prediction of doublets in a cell state continuum; Pre-processing and doublet detector parameters; Marker gene co-expression score

Methods, data and reproducibility

Random pairs of observed cells are combined into simulated doublets (count-weighted averaging of PCA coordinates). Observed and simulated cells share a PCA embedding and kNN graph, and each cell is scored by its fraction of simulated-doublet neighbors. A threshold between the bimodal simulated-score peaks gives doublet calls and the detectable doublet fraction ϕD.

  • Splatter-simulated scRNA-seq data with discrete clusters or a branching continuum, 5% doublets · n = 10 replicates of 5000 cells x 2000 genes per condition
  • 10x Genomics mixture of human HEK293 T and mouse NIH3T3 cells, restricted to genes with identical human and mouse names · human and mouse (cell lines) · n = 6,164 human, 5,915 mouse cells; 741 mixed multiplets · 10x Genomics dataset hgmm_12k (v2.1.0)
  • PBMCs pooled from eight genotyped human donors, with demuxlet singlet/doublet calls · human · n = 14,619 cells, 35,635 genes · GSM2560248
  • One PBMC sample split and loaded at two cell concentrations (10x Genomics PBMC-4 k and PBMC-8 k) · n = 4,352 (PBMC-4 k) and 8,391 (PBMC-8 k) cells · 10x Genomics datasets pbmc4k, pbmc8k (v2.1.0)
  • Kit+ hematopoietic progenitors from mouse bone marrow (library batches 2, 3 and 4) · mouse · n = 4,273 cells after filtering · GSM2388072
Methods & tools
Scrublet (Python), Splatter R package v1.0.3, principal component analysis (PCA), k-nearest-neighbor graph classifier, V-score (baseline-corrected Fano factor) gene selection, demuxlet (genotype-based doublet labels), force-directed graph layout, spectral clustering, SciPy scipy.linalg.expm diffusion smoothing, SPRING, total transcript count thresholding (baseline), marker gene co-expression score (baseline)
Software & environment
Python implementation of Scrublet (version not stated); Splatter R package v1.0.3 for simulations; SciPy scipy.linalg.expm for graph smoothing.
Assay conditions & compute
Only a qualitative runtime is given: classifying datasets of tens of thousands of cells takes a few minutes. Hardware and memory are not reported.
Reproducibility gaps
Doublet thresholds for the empirical datasets were set by eye from simulated-score histograms (simulations used a 95th-percentile cutoff). The 10x dataset links in the text are split by whitespace (e.g., hgmm_12 k, pbmc4 k). The Scrublet version is not stated.
Limitations (6)
  • Doublets are detectable only if both parent states also appear as singlets. In Kit-sorted mouse marrow, 37 putative erythroblast-macrophage doublets (C1qa+ Hba-a1+) were missed because Kit-negative macrophages had no singlet representation. · stated by the authors
  • The method assumes multiplets are rare and simulates only doublets (justified as >97% of multiplets at a <5% multiplet rate with full dissociation). Cell clumps with a stereotyped composition could degrade performance. · stated by the authors
  • Performance depends on manifold structure: within-cluster (Type A) doublets are essentially invisible. ϕD is only 50% for two equal clusters and may fall lower on complex continua or near branch points. · stated by the authors
  • The Scrublet plus total-counts hybrid needed extra weight fitting (tuned to maximize AUC on the same PBMC data), and the authors note its benefit may vary across datasets. · stated by the authors
  • Thresholds were placed by eye on empirical datasets, and preprocessing settings (number of PCs, gene filters, k, r, d) were chosen per dataset. Threshold sensitivity is shown only for the demuxlet PBMC data; robustness to the other choices is not tested. · noted in this summary
  • Only two datasets have orthogonal doublet labels, and both miss same-species or same-donor doublets. The concentration and hematopoiesis tests rely on indirect evidence, so true precision on realistic tissues remains uncertain. · noted in this summary

Read this if you process droplet or well-based scRNA-seq data and need a clustering-free way to flag doublets, or want to estimate how many doublets your embedding can actually reveal.

Citations
2,755
FWCI
59.53
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMcGinnis '19Bais '19Ttoouli '25DePasquale '20DePasquale '19Bernstein '19
75% semantic matchBioinformatics
Christopher S. McGinnis, Lyndsay M. Murrow, Zev J. Gartner·Cell Systems·Apr 1, 2019· 4,515 citations