CoChat
Sign Up
Cited by 279Single-cell and spatial transcriptomicsGene expression and cancer classificationGenomics and Phylogenetic Studies

The DOI matches its Crossref record; 58 of the 59 references it cites were found in a registry or an index.

A systematic evaluation of single cell RNA-seq analysis pipelines

Open access
Vieth, Beate, Parekh, Swati, Ziegenhain, Christoph, Enard, Wolfgang, Hellmann, Ines·2019·Nature Communications
DOI 10.1038/s41467-019-12266-7PMC PMC6789098

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

The recent rapid spread of single cell RNA sequencing (scRNA-seq) methods has created a large variety of experimental and computational pipelines for which best practices have not yet been established. Here, we use simulations based on five scRNA-seq library protocols in combination with nine realistic differential expression (DE) setups to systematically evaluate three mapping, four imputation, seven normalisation and four differential expression testing approaches resulting in ~3000 pipelines, allowing us to also assess interactions among pipeline steps. We find that choices of normalisation and library preparation protocols have the biggest impact on scRNA-seq analyses. Specifically, we find that library preparation determines the ability to detect symmetric expression differences, while normalisation dominates pipeline performance in asymmetric DE-setups. Finally, we illustrate the importance of informed choices by showing that a good scRNA-seq pipeline can have the same impact on detecting a biological signal as quadrupling the sample size.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-019-12266-7

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Study Overview. a The data sets yielding raw count matrices: We use scRNA-seq data sets2,16 representing 5 popular library preparation protocols. For each data set, we obtain multiple gene count matrices that result from various combinations of alignment methods and annotation schemes (see also Supplementary Figs. 1 and 2, and Supplementary Tables 1 and 2). b The simulation setup: Using powsimR10 distribution estimates from real count matrices, we simulate the expression of 10,000 genes for two groups with 384 vs 384, 96 vs. 96 and 50 vs. 200 cells, where 5, 20 or 60% of genes are DE between groups. The magnitude of expression change for each gene is drawn from a narrow gamma distribution (X ~ Γ(α = 1, β = 2)) and the directions can either be symmetric, asymmetric or completely asymmetric. To introduce slight variation in expression capture, we draw a different size factor for each cell from a narrow normal distribution. c The analysis pipeline: The simulated data sets are then analysed using combinations of four count matrix preprocessing, seven normalisation and four DE approaches. The evaluation of these pipelines focuses on the outcome of the confusion matrix and its derivatives (TPR, FDR, pAUC, MCC), deviance in library size estimates (RMSE) and computational run time
Fig. 1 Study Overview. a The data sets yielding raw count matrices: We use scRNA-seq data sets2,16 representing 5 popular library preparation protocols. For each data set, we obtain multiple gene count matrices that result from various combinations of alignment methods and annotation schemes (see also Supplementary Figs. 1 and 2, and Supplementary Tables 1 and 2). b The simulation setup: Using powsimR10 distribution estimates from real count matrices, we simulate the expression of 10,000 genes for two groups with 384 vs 384, 96 vs. 96 and 50 vs. 200 cells, where 5, 20 or 60% of genes are DE between groups. The magnitude of expression change for each gene is drawn from a narrow gamma distribution (X ~ Γ(α = 1, β = 2)) and the directions can either be symmetric, asymmetric or completely asymmetric. To introduce slight variation in expression capture, we draw a different size factor for each cell from a narrow normal distribution. c The analysis pipeline: The simulated data sets are then analysed using combinations of four count matrix preprocessing, seven normalisation and four DE approaches. The evaluation of these pipelines focuses on the outcome of the confusion matrix and its derivatives (TPR, FDR, pAUC, MCC), deviance in library size estimates (RMSE) and computational run timeCC BY · PubMed Central
Fig. 2 Expression Quantification. a Read alignment and assignment rates per library preparation protocol stratified over aligner and annotation. The lighter shade represents the percentage of the total reads that could be aligned and the darker shade the percentage that also was uniquely assigned (see also Supplementary Fig. 3). For comparability, cells were downsampled to 1 million reads/cell, with the exception of 10× Genomics data that were only sequenced to on average 60,000 reads/cell. Hence, these data are farther from saturation and have a higher UMI/read ratio. b Number of genes per UMI with >1 reads for BWA and STAR alignment using the SCRB-seq data set and GENCODE annotation. Colours denote number bins of UMIs. c Number of genes detected per Library Preparation Protocol stratified over Aligner and Annotation (i.e. at least 10% nonzero expression values) (see also Supplementary Fig. 4). d Estimated mean expression, dispersion and gene dropout rates for SCRB-seq and Smart-seq2 data using STAR, BWA or kallisto alignments with GENCODE annotation (see also Supplementary Fig. 7). e Mean-dispersion fitting line applying a cubic smoothing spline with 95% variability bands for SCRB-seq and Smart-seq2 data using STAR, BWA or kallisto alignments with GENCODE annotation (see also Supplementary Fig. 8). f The effect of quantification choices on the power (TPR) to detect differential expression stratified over library preparation and aligner. The expression of 10,000 detected genes over 768 cells (384 cells per group) were simulated given the observed mean-variance relation per protocol. Five percent of the simulated genes are differentially expressed following a symmetric narrow gamma distribution. Unfiltered counts were normalised using scran. Differential expression was tested using limma-trend (see also Supplementary Fig. 9)
Fig. 2 Expression Quantification. a Read alignment and assignment rates per library preparation protocol stratified over aligner and annotation. The lighter shade represents the percentage of the total reads that could be aligned and the darker shade the percentage that also was uniquely assigned (see also Supplementary Fig. 3). For comparability, cells were downsampled to 1 million reads/cell, with the exception of 10× Genomics data that were only sequenced to on average 60,000 reads/cell. Hence, these data are farther from saturation and have a higher UMI/read ratio. b Number of genes per UMI with >1 reads for BWA and STAR alignment using the SCRB-seq data set and GENCODE annotation. Colours denote number bins of UMIs. c Number of genes detected per Library Preparation Protocol stratified over Aligner and Annotation (i.e. at least 10% nonzero expression values) (see also Supplementary Fig. 4). d Estimated mean expression, dispersion and gene dropout rates for SCRB-seq and Smart-seq2 data using STAR, BWA or kallisto alignments with GENCODE annotation (see also Supplementary Fig. 7). e Mean-dispersion fitting line applying a cubic smoothing spline with 95% variability bands for SCRB-seq and Smart-seq2 data using STAR, BWA or kallisto alignments with GENCODE annotation (see also Supplementary Fig. 8). f The effect of quantification choices on the power (TPR) to detect differential expression stratified over library preparation and aligner. The expression of 10,000 detected genes over 768 cells (384 cells per group) were simulated given the observed mean-variance relation per protocol. Five percent of the simulated genes are differentially expressed following a symmetric narrow gamma distribution. Unfiltered counts were normalised using scran. Differential expression was tested using limma-trend (see also Supplementary Fig. 9)CC BY · PubMed Central
Fig. 3 Normalisation choices determines DE-analysis performance, not count preprocessing. The data in panels a–c are based on Smart-seq2 data, all panels are based on two groups of 384 cells, STAR alignment with GENCODE annotation was used for quantification. a The root mean squared error (RMSE) of estimated library size factors per normalisation method is plotted for 20% asymmetric DE-genes (see also Supplementary Fig. 11) (Box and whisker plot with centre line = median, bounds of box = 25th and 75th percentile, whiskers = 1.5 * interquartile range from the lower and upper bounds of the box). b The discriminatory ability determined by the partial area under the curve (mean pAUC ± s.d.) based on DE testing with limma-trend for normalisation without spike-ins per DE-pattern. The grey ribbon indicates the mean pAUC ± s.d. given simulated size factors (see also Supplementary Figs. 13–15). c Using spike-ins for normalisation for 60% completely asymmetric DE-genes. d Effect of preprocessing the count matrix for 20% asymmetric DE-genes without spike-ins. Counts were either left asis (‘none’), filtered or imputed prior to normalisation. The derived scaling factors were then used for normalisation and DE testing was performed on raw counts using limma-trend (see also Supplementary Figs. 16–18). This procedure was applied to the full count matrix (circle) and to the count matrix downsampled to 10% of its original sequencing depth (triangular). Missing data points are due to failing imputation runs with the sparser data
Fig. 3 Normalisation choices determines DE-analysis performance, not count preprocessing. The data in panels a–c are based on Smart-seq2 data, all panels are based on two groups of 384 cells, STAR alignment with GENCODE annotation was used for quantification. a The root mean squared error (RMSE) of estimated library size factors per normalisation method is plotted for 20% asymmetric DE-genes (see also Supplementary Fig. 11) (Box and whisker plot with centre line = median, bounds of box = 25th and 75th percentile, whiskers = 1.5 * interquartile range from the lower and upper bounds of the box). b The discriminatory ability determined by the partial area under the curve (mean pAUC ± s.d.) based on DE testing with limma-trend for normalisation without spike-ins per DE-pattern. The grey ribbon indicates the mean pAUC ± s.d. given simulated size factors (see also Supplementary Figs. 13–15). c Using spike-ins for normalisation for 60% completely asymmetric DE-genes. d Effect of preprocessing the count matrix for 20% asymmetric DE-genes without spike-ins. Counts were either left asis (‘none’), filtered or imputed prior to normalisation. The derived scaling factors were then used for normalisation and DE testing was performed on raw counts using limma-trend (see also Supplementary Figs. 16–18). This procedure was applied to the full count matrix (circle) and to the count matrix downsampled to 10% of its original sequencing depth (triangular). Missing data points are due to failing imputation runs with the sparser dataCC BY · PubMed Central
Fig. 4 Evaluation of DE tools. The expression of 10,000 genes over 768 cells (384 cells per group) were simulated given the observed mean-variance relation per protocol. Twenty percent of the simulated genes are differentially expressed following an asymmetric narrow gamma distribution. Unfiltered counts were normalised using simulated library size factors or applying normalisation methods. Differential expression was tested using T-Test, limma-trend, MAST or edgeR-zingeR. The discriminatory ability of DE methods is determined by the partial area under the curve (mean pAUC ± s.d.) for the TPR-FDR curve (see also Supplementary Figs. 19–25)
Fig. 4 Evaluation of DE tools. The expression of 10,000 genes over 768 cells (384 cells per group) were simulated given the observed mean-variance relation per protocol. Twenty percent of the simulated genes are differentially expressed following an asymmetric narrow gamma distribution. Unfiltered counts were normalised using simulated library size factors or applying normalisation methods. Differential expression was tested using T-Test, limma-trend, MAST or edgeR-zingeR. The discriminatory ability of DE methods is determined by the partial area under the curve (mean pAUC ± s.d.) for the TPR-FDR curve (see also Supplementary Figs. 19–25)CC BY · PubMed Central
Fig. 5 Evaluation of analysis pipeline. a, b The expression of 10000 genes over 768 cells were simulated and 5, 20 or 60% of the genes were differentially expressed following a symmetric or asymmetric narrow gamma distribution. This simulation setup was applied to protocols, alignments, annotations, preprocessing of counts, normalisation and DE tools. For each analysis set, the Matthew Correlation Coefficient (mean MCC ± s.d.) was averaged over 20 simulations and rescaled to [0, 1] interval. The MCC was used as a response variable in beta regression models with log-log link function. a The contribution of each covariate in the full model (~Protocol + Aligner + Annotation + Preprocessing + Normalisation + DE-Tool). b Performance according to sample size, 1 good and 1 naive pipeline (see also Supplementary Fig. 26). c–e The expression of ~1000 human PBMcs profiled with 10× Genomics were processed using the good and naive pipeline. Cell types were identified with SingleR classification using the Blueprint Epigenomics Reference. Cell types are represented in a UMAP, for good c and naive d pipeline, respectively. True marker genes, i.e., given by the reference, per pairwise comparison of cell types for the good and naive pipeline are given in e where genes needed to have a adjusted p-value < 0.1, absolute log2 fold change threshold (>0.1) and expressed in at least 10% of the cells to be considered. f Pipeline recommendations for UMI and Smart-seq2 data
Fig. 5 Evaluation of analysis pipeline. a, b The expression of 10000 genes over 768 cells were simulated and 5, 20 or 60% of the genes were differentially expressed following a symmetric or asymmetric narrow gamma distribution. This simulation setup was applied to protocols, alignments, annotations, preprocessing of counts, normalisation and DE tools. For each analysis set, the Matthew Correlation Coefficient (mean MCC ± s.d.) was averaged over 20 simulations and rescaled to [0, 1] interval. The MCC was used as a response variable in beta regression models with log-log link function. a The contribution of each covariate in the full model (~Protocol + Aligner + Annotation + Preprocessing + Normalisation + DE-Tool). b Performance according to sample size, 1 good and 1 naive pipeline (see also Supplementary Fig. 26). c–e The expression of ~1000 human PBMcs profiled with 10× Genomics were processed using the good and naive pipeline. Cell types were identified with SingleR classification using the Blueprint Epigenomics Reference. Cell types are represented in a UMAP, for good c and naive d pipeline, respectively. True marker genes, i.e., given by the reference, per pairwise comparison of cell types for the good and naive pipeline are given in e where genes needed to have a adjusted p-value < 0.1, absolute log2 fold change threshold (>0.1) and expressed in at least 10% of the cells to be considered. f Pipeline recommendations for UMI and Smart-seq2 dataCC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Cited by

Sorted by citations

See all 279

Related

See all

Metadata

Authors
Vieth, B., Parekh, S., Ziegenhain, C., Enard, W., & Hellmann, I.
Affiliations
  • Ludwig-Maximilians-Universität München (DE)
  • Max Planck Institute for Biology of Ageing (DE)
  • Karolinska Institutet (SE)
Corresponding author
Hellmann, Ines
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2019-10-11
Volume
10
Issue
1
Pages
4667-4667
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-019-12266-7
PMC
PMC6789098
OpenAlex
W2980147288
Cited by
265
References
59
Citations by year
2026: 10 · 2025: 13 · 2024: 24 · 2023: 41 · 2022: 38 · 2021: 82 · 2020: 55 · 2019: 1
FWCI
15.38
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • Open access LMU (Ludwid Maxmilian's Universitat Munchen) (submittedVersion, cc-by)
Funders
Deutsche Forschungsgemeinschaft (SFB 1243)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene expression and cancer classification
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
Keywords
single-cell RNA sequencing, data analysis pipeline, differential gene expression, library preparation protocols, normalization, imputation, read mapping, simulated data, sample size, best practices
MeSH terms
RNA-Seq, Animals, Electronic Data Processing, Chromosome Mapping, Computer Simulation, Mice, Single-Cell Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review