CoChat
Sign Up
Cited by 204Single-cell and spatial transcriptomicsGene expression and cancer classificationInnovative Microfluidic and Catalytic Techniques Innovation

The DOI matches its Crossref record; 66 of the 66 references it cites were found in a registry or an index.

Normalizing and denoising protein expression data from droplet-based single cell profiling

Open access
Mulè, Matthew P., Martins, Andrew J., Tsang, John S.·2022·Nature Communications
DOI 10.1038/s41467-022-29356-8PMC PMC9018908

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Multimodal single-cell profiling methods that measure protein expression with oligo-conjugated antibodies hold promise for comprehensive dissection of cellular heterogeneity, yet the resulting protein counts have substantial technical noise that can mask biological variations. Here we integrate experiments and computational analyses to reveal two major noise sources and develop a method called "dsb" (denoised and scaled by background) to normalize and denoise droplet-based protein expression data. We discover that protein-specific noise originates from unbound antibodies encapsulated during droplet generation; this noise can thus be accurately estimated and corrected by utilizing protein levels in empty droplets. We also find that isotype control antibodies and the background protein population average in each cell exhibit significant correlations across single cells, we thus use their shared variance to correct for cell-to-cell technical noise in each cell. We validate these findings by analyzing the performance of dsb in eight independent datasets spanning multiple technologies, including CITE-seq, ASAP-seq, and TEA-seq. Compared to existing normalization methods, our approach improves downstream analyses by better unmasking biologically meaningful cell populations. Our method is available as an open-source R package that interfaces easily with existing single cell software platforms such as Seurat, Bioconductor, and Scanpy and can be accessed at "dsb [ https://cran.r-project.org/package=dsb ]".

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/s41467-022-29356-8

Cites

66 of 66 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Antibody-derived protein UMI count data noise source assessment. a 1 and 2: Experimental setup and potential noise sources in CITE-seq data. 3: protein-specific noise: if ambient antibody encapsulated in droplets constitutes a major source of protein-specific noise, values should be highly correlated with those in unstained control cells (top); if control cells contain information on noise not captured by empty drops, the correlation should be weak. 4: Cell-specific noise evaluated through the correlation between the background protein population mean and isotype controls across single cells. Created with BioRender.com. b Average protein log10(count + 1) of unstained control cells spiked into the stained cell pool prior to droplet generation (y-axis) versus that of droplets without a cell (x-axis). Pearson correlation coefficient and p value (two sided) are shown. c Density histograms of protein expression of lineage-defining proteins within major subsets in stained cells (black) and unstained controls (red) normalized together using dsb step I (ambient correction and rescaling based on levels in empty droplets). d A two-component Gaussian mixture model was fitted to the protein counts within each single cell; the distributions of the component means from all single cell fits (blue = ”negative” population; red = “positive” population) are shown, protein distributions from a randomly selected cell shown in the inset. e Comparison of Gaussian mixture models fit with between k = 1 and k = 6 subpopulations to dsb normalized protein values for n = 28,229 cells from batch 1 after dsb step I (ambient correction) but prior to step II, vs. the model fit Bayesian Information Criteria (BIC, using mclust R package definition of BIC where larger values correspond to a better fit) from the resulting 169,374 models. Boxplots show the median with hinges at the 25th and 75th percentile, whiskers extend plus or −1.5 times the inter quartile range. k = 2 component Gaussian mixtures have the best fit in more than 80% of cells (orange, right inset bar plot). f Pearson correlation coefficients among isotype controls and background component mean inferred by Gaussian mixture model (µ1 fitted per cell as in d); all corresponding p values (two sided) are <2e−16. g Scatter density plot between µ1, the mean of each cell’s negative subpopulation from the per-cell Gaussian mixture model (blue in c) versus the mean of the four isotype controls across single cells. Pearson correlation coefficient is shown (two-sided p value < 2e−16). h The distribution of the dsb technical component as calculated using a 2 component (x-axis) vs. 3 component (y-axis) mixture model to define the µ1 parameter, Pearson correlation coefficient, p value (two-sided) < 2e−16.
Fig. 1 Antibody-derived protein UMI count data noise source assessment. a 1 and 2: Experimental setup and potential noise sources in CITE-seq data. 3: protein-specific noise: if ambient antibody encapsulated in droplets constitutes a major source of protein-specific noise, values should be highly correlated with those in unstained control cells (top); if control cells contain information on noise not captured by empty drops, the correlation should be weak. 4: Cell-specific noise evaluated through the correlation between the background protein population mean and isotype controls across single cells. Created with BioRender.com. b Average protein log10(count + 1) of unstained control cells spiked into the stained cell pool prior to droplet generation (y-axis) versus that of droplets without a cell (x-axis). Pearson correlation coefficient and p value (two sided) are shown. c Density histograms of protein expression of lineage-defining proteins within major subsets in stained cells (black) and unstained controls (red) normalized together using dsb step I (ambient correction and rescaling based on levels in empty droplets). d A two-component Gaussian mixture model was fitted to the protein counts within each single cell; the distributions of the component means from all single cell fits (blue = ”negative” population; red = “positive” population) are shown, protein distributions from a randomly selected cell shown in the inset. e Comparison of Gaussian mixture models fit with between k = 1 and k = 6 subpopulations to dsb normalized protein values for n = 28,229 cells from batch 1 after dsb step I (ambient correction) but prior to step II, vs. the model fit Bayesian Information Criteria (BIC, using mclust R package definition of BIC where larger values correspond to a better fit) from the resulting 169,374 models. Boxplots show the median with hinges at the 25th and 75th percentile, whiskers extend plus or −1.5 times the inter quartile range. k = 2 component Gaussian mixtures have the best fit in more than 80% of cells (orange, right inset bar plot). f Pearson correlation coefficients among isotype controls and background component mean inferred by Gaussian mixture model (µ1 fitted per cell as in d); all corresponding p values (two sided) are <2e−16. g Scatter density plot between µ1, the mean of each cell’s negative subpopulation from the per-cell Gaussian mixture model (blue in c) versus the mean of the four isotype controls across single cells. Pearson correlation coefficient is shown (two-sided p value < 2e−16). h The distribution of the dsb technical component as calculated using a 2 component (x-axis) vs. 3 component (y-axis) mixture model to define the µ1 parameter, Pearson correlation coefficient, p value (two-sided) < 2e−16.CC BY · PubMed Central
Fig. 2 Assessment of dsb model assumptions and performance of dsb normalization on external datasets. Panels a–h Application of dsb to a publicly available dataset generated using 10X genomics “NextGem” chemistry measuring 29 proteins across ~5K cells. a The protein library size distribution of empty and cell-containing droplets used for dsb normalization. b UMAP of single cells based on dsb normalized protein values with colors representing clusters obtained from clustering cells on dsb normalized protein values. c Heatmap of the average of dsb normalized values per protein-based cluster shown in (b). d The distribution of CD14 and CD4 dsb normalized values. e As in Fig. 1e, Gaussian mixture model parameters fit to the dsb normalized values of each single cell after step I (ambient noise/background droplet based correction). The Bayesian Information Criterion (BIC) of the model vs. number of components in the model fit for each cell (n = 3774 cells). Boxplots show the median with hinges at the 25th and 75th percentile and whiskers extending plus or minus 1.5 times the inter quartile range. f As in Fig. 1f, Pearson correlation coefficient matrix of variables used to define each cell’s technical component; each isotype control and µ1, the Gaussian mixture model background mean across proteins for each cell. g As in Fig. 1g, Pearson correlation coefficient between the inferred cell-specific background mean µ1 from the Gaussian mixture model vs. the mean of isotype controls in each cell. h The relationship between each cell’s technical component and the cell’s protein library size (Pearson correlation coefficient shown as in Supplementary Fig 3a with 95% confidence interval in gray). i Summary statistics for the eight independent datasets assessed in this study; Cor 1 and 2 correspond to the Pearson correlation coefficient for assessing the relationships between variables shown in (h) and (g) across cells for each dataset.
Fig. 2 Assessment of dsb model assumptions and performance of dsb normalization on external datasets. Panels a–h Application of dsb to a publicly available dataset generated using 10X genomics “NextGem” chemistry measuring 29 proteins across ~5K cells. a The protein library size distribution of empty and cell-containing droplets used for dsb normalization. b UMAP of single cells based on dsb normalized protein values with colors representing clusters obtained from clustering cells on dsb normalized protein values. c Heatmap of the average of dsb normalized values per protein-based cluster shown in (b). d The distribution of CD14 and CD4 dsb normalized values. e As in Fig. 1e, Gaussian mixture model parameters fit to the dsb normalized values of each single cell after step I (ambient noise/background droplet based correction). The Bayesian Information Criterion (BIC) of the model vs. number of components in the model fit for each cell (n = 3774 cells). Boxplots show the median with hinges at the 25th and 75th percentile and whiskers extending plus or minus 1.5 times the inter quartile range. f As in Fig. 1f, Pearson correlation coefficient matrix of variables used to define each cell’s technical component; each isotype control and µ1, the Gaussian mixture model background mean across proteins for each cell. g As in Fig. 1g, Pearson correlation coefficient between the inferred cell-specific background mean µ1 from the Gaussian mixture model vs. the mean of isotype controls in each cell. h The relationship between each cell’s technical component and the cell’s protein library size (Pearson correlation coefficient shown as in Supplementary Fig 3a with 95% confidence interval in gray). i Summary statistics for the eight independent datasets assessed in this study; Cor 1 and 2 correspond to the Pearson correlation coefficient for assessing the relationships between variables shown in (h) and (g) across cells for each dataset.CC BY · PubMed Central
Fig. 3 Case study I: dsb improves interpretation of cell clusters derived from protein-based and joint mRNA–protein clustering. a UMAP plot of single cells labeled by cluster number (clustering was performed using dsb normalized protein values). b The distribution of protein expression of cluster 4 (highlighted with a gray box in (a)) using CLR (across cells) or dsb for normalization. c Median log + 1 protein levels (left) and CLR transformed across cells (as in (b), right) in cells from cluster 4 versus the level in empty droplets; proteins highlighted in red are comparable in expression to “positive” proteins after log transformation (left) and CLR transformation across cells (right) but are similar to background levels in empty droplets (identity line y = x shown in black). All proteins with median log10 expression >1 but <3.5 after dsb normalization are labeled with the protein name. d Similar to (c), but the y-axis shows the median dsb normalized values; proteins in red (those near the diagonal in (c)) are now residing below our uniformly applied dsb positivity threshold of 3.5, reflective their proximity with mean counts in empty droplets; proteins above the red line have median dsb normalized expression within the highlighted cluster 4 (see (a) and (b)) above 3.5, i.e., 3.5 standard deviations above ambient noise, ±adjustment for the cell intrinsic technical component. e The dsb normalized value vs. the median value in empty droplets of proteins within a subset of protein-defined clusters. A subset of proteins informative for cluster identification from B cell and dendritic cell subsets with a dsb value above 3.5 (red line) are annotated with the protein name within each panel and are labeled in red when below 3.5 within each subset. Proteins labeled for B cell subsets (C13: Unswitched B cells, C5 Transitional B cells, C12 Switched B cells) include B cell proteins CD20, CD19, IgD, and IgM, proteins labeled for the dendritic cell subsets (C16: pDC, C14: mDC) include innate cell markers CD1c, CD1d, CD34, CD14, CD16, and CD303. f UMAP plot of the same cells shown in (a) but the UMAP embeddings and clusters here were derived using Seurat’s weighted nearest neighbor (WNN) mRNA-protein multimodal algorithm applied to dsb normalized values. g Similar to (d) but derived using cells from WNN cluster 3; Pearson correlation coefficient and p value (two sided) are shown between median dsb normalized values and the 98th percentile expression value (log10) of the same protein in empty droplets. h Similar to (g) but for CLR normalized values. i. Differentially expressed genes (ROC test; see the “Methods” section) for cell in cluster 3 vs. other clusters.
Fig. 3 Case study I: dsb improves interpretation of cell clusters derived from protein-based and joint mRNA–protein clustering. a UMAP plot of single cells labeled by cluster number (clustering was performed using dsb normalized protein values). b The distribution of protein expression of cluster 4 (highlighted with a gray box in (a)) using CLR (across cells) or dsb for normalization. c Median log + 1 protein levels (left) and CLR transformed across cells (as in (b), right) in cells from cluster 4 versus the level in empty droplets; proteins highlighted in red are comparable in expression to “positive” proteins after log transformation (left) and CLR transformation across cells (right) but are similar to background levels in empty droplets (identity line y = x shown in black). All proteins with median log10 expression >1 but <3.5 after dsb normalization are labeled with the protein name. d Similar to (c), but the y-axis shows the median dsb normalized values; proteins in red (those near the diagonal in (c)) are now residing below our uniformly applied dsb positivity threshold of 3.5, reflective their proximity with mean counts in empty droplets; proteins above the red line have median dsb normalized expression within the highlighted cluster 4 (see (a) and (b)) above 3.5, i.e., 3.5 standard deviations above ambient noise, ±adjustment for the cell intrinsic technical component. e The dsb normalized value vs. the median value in empty droplets of proteins within a subset of protein-defined clusters. A subset of proteins informative for cluster identification from B cell and dendritic cell subsets with a dsb value above 3.5 (red line) are annotated with the protein name within each panel and are labeled in red when below 3.5 within each subset. Proteins labeled for B cell subsets (C13: Unswitched B cells, C5 Transitional B cells, C12 Switched B cells) include B cell proteins CD20, CD19, IgD, and IgM, proteins labeled for the dendritic cell subsets (C16: pDC, C14: mDC) include innate cell markers CD1c, CD1d, CD34, CD14, CD16, and CD303. f UMAP plot of the same cells shown in (a) but the UMAP embeddings and clusters here were derived using Seurat’s weighted nearest neighbor (WNN) mRNA-protein multimodal algorithm applied to dsb normalized values. g Similar to (d) but derived using cells from WNN cluster 3; Pearson correlation coefficient and p value (two sided) are shown between median dsb normalized values and the 98th percentile expression value (log10) of the same protein in empty droplets. h Similar to (g) but for CLR normalized values. i. Differentially expressed genes (ROC test; see the “Methods” section) for cell in cluster 3 vs. other clusters.CC BY · PubMed Central
Fig. 4 Case study II: application of dsb to tri-modal TEA-seq data unmasks a MAIT cell population obscured by noise in CLR normalization. Analysis of TEA-seq (transcriptome, epitopes, and accessibility) tri-modal single cell assay data. a dsb normalization of protein data from TEA-seq showing the distribution of CD4 and CD14 with the same 3.5 threshold used throughout the study. b UMAP plot of single cells and clusters derived by WNN joint mRNA–protein clustering with protein data normalized using dsb. c Bi-axial distribution of the alpha beta and va7.2 T cell receptor (TCR) proteins in cluster 14 cells normalized by dsb and d the same cells’ CLR normalized values. e Similar to Fig. 3g but here for cluster 14 from (b) using dsb or f CLR normalized values (y-axis); in both plots Pearson correlation coefficients and p values (two sided) are shown between normalized values (y axis) and values in empty droplets (x axis). g Differential expression analysis (ROC test) of genes in cluster 14 vs. other clusters. h Gene set enrichment of a MAIT cell signature constructed from FACS-sorted TCR-va7.2+/MAIT cells compared to other T cells (RNA-seq data from Park et al. 2019) with genes ranked by log2 fold change in cluster 14 cells vs. other cells as in (g).
Fig. 4 Case study II: application of dsb to tri-modal TEA-seq data unmasks a MAIT cell population obscured by noise in CLR normalization. Analysis of TEA-seq (transcriptome, epitopes, and accessibility) tri-modal single cell assay data. a dsb normalization of protein data from TEA-seq showing the distribution of CD4 and CD14 with the same 3.5 threshold used throughout the study. b UMAP plot of single cells and clusters derived by WNN joint mRNA–protein clustering with protein data normalized using dsb. c Bi-axial distribution of the alpha beta and va7.2 T cell receptor (TCR) proteins in cluster 14 cells normalized by dsb and d the same cells’ CLR normalized values. e Similar to Fig. 3g but here for cluster 14 from (b) using dsb or f CLR normalized values (y-axis); in both plots Pearson correlation coefficients and p values (two sided) are shown between normalized values (y axis) and values in empty droplets (x axis). g Differential expression analysis (ROC test) of genes in cluster 14 vs. other clusters. h Gene set enrichment of a MAIT cell signature constructed from FACS-sorted TCR-va7.2+/MAIT cells compared to other T cells (RNA-seq data from Park et al. 2019) with genes ranked by log2 fold change in cluster 14 cells vs. other cells as in (g).CC 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

Show all 65 references

Research works printed but not linked: 1

  1. Swanson, E. et al. Integrated single cell analysis of chromatin accessibility and cell surface markers. eLife. 10, e63632 (2021). · DOI · Paper page

Cited by

Sorted by citations

Related

See all

Metadata

Authors
Mulè, M. P., Martins, A. J., & Tsang, J. S.
Affiliations
National Institutes of Health (US)
Corresponding author
Tsang, John S.
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2022-04-19
Volume
13
Issue
1
Pages
2099-2099
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-022-29356-8
PMC
PMC9018908
OpenAlex
W4225731526
Cited by
197
References
65
Citations by year
2026: 39 · 2025: 43 · 2024: 41 · 2023: 36 · 2022: 27
FWCI
12.65
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, public-domain)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Apollo (University of Cambridge) (submittedVersion, cc-by)
  • Apollo
  • Apollo (cc-by)
Funders
  • Johns Hopkins University
  • National Institutes of Health
  • National Institute of Allergy and Infectious Diseases (ZIAAI001152)
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
  • Physical Sciences › Chemical Engineering › Fluid Flow and Transfer Processes › Innovative Microfluidic and Catalytic Techniques Innovation
Keywords
data denoising, CITE-seq, cellular heterogeneity
MeSH terms
Software, Gene Expression Profiling, 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