CoChat

Determining sequencing depth in a single-cell RNA-seq experiment

Open access
Zhang, Martin Jinye, Ntranos, Vasilis, Tse, David · 2020 · Nature Communications

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

Abstract

An underlying question for virtually all single-cell RNA sequencing experiments is how to allocate the limited sequencing budget: deep sequencing of a few cells or shallow sequencing of many cells? Here we present a mathematical framework which reveals that, for estimating many important gene properties, the optimal allocation is to sequence at a depth of around one read per cell per gene. Interestingly, the corresponding optimal estimator is not the widely-used plug-in estimator, but one developed via empirical Bayes.

Overview

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

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

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

Integrity signals

Registry

Crossref · 10.1038/s41467-020-14482-y

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Optimal sequencing budget allocation. a Description of the sequencing budget allocation problem. Consider estimating the underlying gene distribution (top) from the noisy read counts obtained via sequencing (bottom). With a fixed number of reads to be sequenced, deep sequencing of a few cells accurately estimates each individual cell but lacks coverage of the entire distribution (left), whereas a shallow sequencing of many cells covers the entire population but introduces a lot of noise (right). b Optimal tradeoff. The memory T-cell marker gene S100A4 has 41.7k reads in the pbmc_4k dataset. For estimating the underlying gamma distribution Xg~Gamma(rg,θg), the relative error is plotted as a function of the sequencing depth, where the optimal error is obtained at a depth of one read per cell (orange star) and is two times smaller than that at the current depth of pbmc_4k (red triangle). c Experimental design. To determine the sequencing depth for an experiment, first the relative gene expression level can be obtained via pilot experiments or previous studies (top left). Then the researcher can select a set of genes of interest (i.e., some marker genes highlighted as black dots), of which the smallest relative expression level p^(*) (MS4A1) defines the reliable detection limit. Finally, the optimal sequencing depth is determined as nreads^(*)=1∕p^(*) (top right). The errors under different tradeoffs are visualized as a function of the genes ordered from the most expressed to the least (bottom). The optimal sequencing budget allocation (orange) minimizes the worst-case error over all the genes of interest (left of the red dashed line), whereas both the deeper sequencing (green) and the shallower sequencing (blue) yield worse results.
Fig. 1 Optimal sequencing budget allocation. a Description of the sequencing budget allocation problem. Consider estimating the underlying gene distribution (top) from the noisy read counts obtained via sequencing (bottom). With a fixed number of reads to be sequenced, deep sequencing of a few cells accurately estimates each individual cell but lacks coverage of the entire distribution (left), whereas a shallow sequencing of many cells covers the entire population but introduces a lot of noise (right). b Optimal tradeoff. The memory T-cell marker gene S100A4 has 41.7k reads in the pbmc_4k dataset. For estimating the underlying gamma distribution Xg~Gamma(rg,θg), the relative error is plotted as a function of the sequencing depth, where the optimal error is obtained at a depth of one read per cell (orange star) and is two times smaller than that at the current depth of pbmc_4k (red triangle). c Experimental design. To determine the sequencing depth for an experiment, first the relative gene expression level can be obtained via pilot experiments or previous studies (top left). Then the researcher can select a set of genes of interest (i.e., some marker genes highlighted as black dots), of which the smallest relative expression level p^(*) (MS4A1) defines the reliable detection limit. Finally, the optimal sequencing depth is determined as nreads^(*)=1∕p^(*) (top right). The errors under different tradeoffs are visualized as a function of the genes ordered from the most expressed to the least (bottom). The optimal sequencing budget allocation (orange) minimizes the worst-case error over all the genes of interest (left of the red dashed line), whereas both the deeper sequencing (green) and the shallower sequencing (blue) yield worse results.CC BY · PubMed Central
Fig. 2 Empirical quantification of the optimal sequencing depth. a Simulations of error under different budget allocation. 3-std confidence intervals are provided. The top panel simulates the error for estimating the first principal direction using the plug-in estimator (blue) and the EB estimator (orange), respectively. Three budgets are considered, i.e., B1 = 0.6 k per gene, B2 = 3k per gene, B3 = 15k per gene. The depth (mean reads per cell per gene) ranges from 0.02 to 10. The result indicates that the optimal depth for the EB estimator is the same (~0.1) for all three budgets, validating the theory that the optimal depth is independent of the budget. The cases for the coefficient of variation and the Pearson correlation (bottom) also show similar qualitative behaviors. b Post hoc guidance for reliable estimation. We visualized the top 4k genes of some representative datasets (top), where a triangle residing in the green region means the Pearson correlation of corresponding genes can be reliably estimated (relative error < 10%). For example, we can reliably estimate the first 2k genes for the brain_1k dataset and all 4k genes for the brain_9k dataset. A more comprehensive result is summarized in the bottom table. For example, the first element (mean, 1k) shows that with 1k cells, a gene needs to have at least 0.1 reads per cell for reliably estimating the mean.
Fig. 2 Empirical quantification of the optimal sequencing depth. a Simulations of error under different budget allocation. 3-std confidence intervals are provided. The top panel simulates the error for estimating the first principal direction using the plug-in estimator (blue) and the EB estimator (orange), respectively. Three budgets are considered, i.e., B1 = 0.6 k per gene, B2 = 3k per gene, B3 = 15k per gene. The depth (mean reads per cell per gene) ranges from 0.02 to 10. The result indicates that the optimal depth for the EB estimator is the same (~0.1) for all three budgets, validating the theory that the optimal depth is independent of the budget. The cases for the coefficient of variation and the Pearson correlation (bottom) also show similar qualitative behaviors. b Post hoc guidance for reliable estimation. We visualized the top 4k genes of some representative datasets (top), where a triangle residing in the green region means the Pearson correlation of corresponding genes can be reliably estimated (relative error < 10%). For example, we can reliably estimate the first 2k genes for the brain_1k dataset and all 4k genes for the brain_9k dataset. A more comprehensive result is summarized in the bottom table. For example, the first element (mean, 1k) shows that with 1k cells, a gene needs to have at least 0.1 reads per cell for reliably estimating the mean.CC BY · PubMed Central
Fig. 3 EB estimates are consistent between deep and shallow datasets. a Top: for estimating the coefficient of variation (CV), the plug-in estimates become more inflated as the sequencing depth becomes shallower (from right to left along the x axis), whereas the EB estimates are consistent. 3-std confidence intervals are provided for this panel. Middle: both brain_1k and brain_1.3m are from the mouse brain, and hence each gene should have a similar CV value between the two datasets. This is indeed the case for the EB estimator (right), which is adaptive to different sequencing depths. However, as brain_1k is twice deeper than brain_1.3m, the plug-in estimates are biased that most points are above the 45-degree line (red). Bottom: distribution recovery for the gene GZMA from a dataset that is subsampled to be five times shallower (left). The EB estimator provides a reasonable estimation for both the zero proportion and the tail shape, resulting in a small total variation error (right). b Feature selection and PCA. The task is to first select features (genes) based on CV, and then perform PCA on the selected features. The results on the full data (pbmc_4k) and a subsampled (three times shallower) are compared. EB estimates are more consistent between the full data and the subsampled data for both the CV ranks (top) and the PCA plots (bottom).
Fig. 3 EB estimates are consistent between deep and shallow datasets. a Top: for estimating the coefficient of variation (CV), the plug-in estimates become more inflated as the sequencing depth becomes shallower (from right to left along the x axis), whereas the EB estimates are consistent. 3-std confidence intervals are provided for this panel. Middle: both brain_1k and brain_1.3m are from the mouse brain, and hence each gene should have a similar CV value between the two datasets. This is indeed the case for the EB estimator (right), which is adaptive to different sequencing depths. However, as brain_1k is twice deeper than brain_1.3m, the plug-in estimates are biased that most points are above the 45-degree line (red). Bottom: distribution recovery for the gene GZMA from a dataset that is subsampled to be five times shallower (left). The EB estimator provides a reasonable estimation for both the zero proportion and the tail shape, resulting in a small total variation error (right). b Feature selection and PCA. The task is to first select features (genes) based on CV, and then perform PCA on the selected features. The results on the full data (pbmc_4k) and a subsampled (three times shallower) are compared. EB estimates are more consistent between the full data and the subsampled data for both the CV ranks (top) and the PCA plots (bottom).CC BY · PubMed Central
Fig. 4 Gene module and gene network analysis. a Top: the EB-estimated Pearson correlation for some marker genes in pbmc_4k are visualized, ordered by different cell populations (top). The clear block-diagonal structure implies that the EB estimator is capable of capturing the gene functional groups. As a comparison, the plug-in estimator also recovers those modules but with a weaker contrast (bottom left panel, plug-in with 100%). Bottom: a subsample experiment further shows that the EB estimator can recover the module with 5% of the data. For the plug-in estimator, the first block (T cells) is blurred with 25% of the data, and the entire structure vanishes with 10% of the data. b Gene network based on the EB-estimated Pearson correlation using the pbmc_4k dataset. Most gene modules correspond to important cell types or functions, including T cells, B cells, NK-cells, myeloid-derived cells, megakaryocytes/platelets, ribosomal protein genes, and mitochondrially encoded protein-coding genes. c Left: the estimated Pearson correlations between all genes and LCK (1st panel) and CD3D (2nd panel), two known T-cell markers. There are three modes for the EB-estimated values, where the positive mode, the zero mode, and the negative mode correspond to genes in the same module, different modules, and irrelevant genes, respectively. The plug-in estimated values are nonetheless much closer to zero even for the truly correlated ones, indicating an artificial shrinkage of the estimated values. Right: two instances where the EB estimates are significantly different from the plug-in estimates. The axes represent read counts, and the color codes the number of cells. Both gene pairs are biologically validated (see Gene network analysis in Methods). See also Supplementary Figs. 11–12 for more examples.
Fig. 4 Gene module and gene network analysis. a Top: the EB-estimated Pearson correlation for some marker genes in pbmc_4k are visualized, ordered by different cell populations (top). The clear block-diagonal structure implies that the EB estimator is capable of capturing the gene functional groups. As a comparison, the plug-in estimator also recovers those modules but with a weaker contrast (bottom left panel, plug-in with 100%). Bottom: a subsample experiment further shows that the EB estimator can recover the module with 5% of the data. For the plug-in estimator, the first block (T cells) is blurred with 25% of the data, and the entire structure vanishes with 10% of the data. b Gene network based on the EB-estimated Pearson correlation using the pbmc_4k dataset. Most gene modules correspond to important cell types or functions, including T cells, B cells, NK-cells, myeloid-derived cells, megakaryocytes/platelets, ribosomal protein genes, and mitochondrially encoded protein-coding genes. c Left: the estimated Pearson correlations between all genes and LCK (1st panel) and CD3D (2nd panel), two known T-cell markers. There are three modes for the EB-estimated values, where the positive mode, the zero mode, and the negative mode correspond to genes in the same module, different modules, and irrelevant genes, respectively. The plug-in estimated values are nonetheless much closer to zero even for the truly correlated ones, indicating an artificial shrinkage of the estimated values. Right: two instances where the EB estimates are significantly different from the plug-in estimates. The axes represent read counts, and the color codes the number of cells. Both gene pairs are biologically validated (see Gene network analysis in Methods). See also Supplementary Figs. 11–12 for more examples.CC BY · PubMed Central
Fig. 5 Validation using smFISH data. a The estimated CV (top) and inactive probability (bottom, κ=2.5nreads) from the Drop-seq data are compared with the smFISH results. The EB estimates (right) are consistent with the smFISH results while there is a clear inflation for the plug-in estimates. b The sequencing budget tradeoff for estimating CV (top) and inactive probability (bottom, κ=1, i.e., estimating the zero proportion at 1 read per cell) for the gene MITF. The relative error is evaluated against the gold standard smFISH result. 3-std confidence intervals are provided.
Fig. 5 Validation using smFISH data. a The estimated CV (top) and inactive probability (bottom, κ=2.5nreads) from the Drop-seq data are compared with the smFISH results. The EB estimates (right) are consistent with the smFISH results while there is a clear inflation for the plug-in estimates. b The sequencing budget tradeoff for estimating CV (top) and inactive probability (bottom, κ=1, i.e., estimating the zero proportion at 1 read per cell) for the gene MITF. The relative error is evaluated against the gold standard smFISH result. 3-std confidence intervals are provided.CC BY · PubMed Central

Verification

Claims the paper makes

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

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

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

Show all 58 references

Research works printed but not linked: 2

  1. Mohammadi, S., Davila-Velderrain, J., Kellis, M. & Grama, A. DECODE-ing sparsity patterns in single-cell RNA-seq, Preprint at https://doi.org/10.1101/241646v2 (2018). · doi:10.1101/241646v2
  2. Ding, J. et al., Systematic comparative analysis of single cell rna-sequencing methods, Preprint at https://doi.org/10.1101/632216v2 (2019). · doi:10.1101/632216v2

Cited by

See all 112

Metadata

Authors
Zhang, M. J., Ntranos, V., & Tse, D.
Affiliations
  • Stanford University (US)
  • California Institute of Technology (US)
Corresponding author
Tse, David
Venue
Nature Communications
Published
2020-02-07
Volume
11
Issue
1
Pages
774-774
ISSN
2041-1723 (ISSN-L)
OpenAlex
W3004833389
Cited by
112
References
55
Citations by year
2026: 8 · 2025: 14 · 2024: 18 · 2023: 21 · 2022: 20 · 2021: 26 · 2020: 5
FWCI
5.83
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • CaltechAUTHORS (California Institute of Technology) (submittedVersion, cc-by)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
U.S. Department of Health & Human Services | NIH | National Human Genome Research Institute (R01HG008164)
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
sequencing depth, single-cell RNA sequencing, empirical Bayes, plug-in estimator, gene expression estimation, mathematical framework
MeSH terms
S100 Calcium-Binding Protein A4, Models, Theoretical, Reproducibility of Results, Gene Expression, In Situ Hybridization, Fluorescence, Sequence Analysis, RNA, Computational Biology, Gene Regulatory Networks, 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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Determining sequencing depth in a single-cell RNA-seq experiment. https://app.cochat.ai/papers/10.1038/s41467-020-14482-y
The paper
Zhang, M. J., Ntranos, V., & Tse, D. (2020). Determining sequencing depth in a single-cell RNA-seq experiment. Nature Communications. https://doi.org/10.1038/s41467-020-14482-y

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.