CoChat
Sign Up
Cited by 476Single-cell and spatial transcriptomicsGene expression and cancer classificationGenomics and Phylogenetic Studies

The DOI matches its Crossref record; 73 of the 73 references it cites were found in a registry or an index.

Feature selection and dimension reduction for single-cell RNA-Seq based on a multinomial model

Open access
Townes, F. William, Hicks, Stephanie C., Aryee, Martin J., Irizarry, Rafael A.·2019·Genome Biology
DOI 10.1186/s13059-019-1861-6PMC PMC6927135

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Single-cell RNA-Seq (scRNA-Seq) profiles gene expression of individual cells. Recent scRNA-Seq datasets have incorporated unique molecular identifiers (UMIs). Using negative controls, we show UMI counts follow multinomial sampling with no zero inflation. Current normalization procedures such as log of counts per million and feature selection by highly variable genes produce false variability in dimension reduction. We propose simple multinomial methods, including generalized principal component analysis (GLM-PCA) for non-normal distributions, and feature selection using deviance. These methods outperform the current practice in a downstream clustering assessment using ground truth datasets.

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 · 10.1186/s13059-019-1861-6

Cites

73 of 73 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Multinomial model adequately characterizes sampling distributions of technical and biological replicates negative control data. a Fraction of zeros is plotted against the total number of UMI in each droplet for the technical replicates. b As a but for cells in the biological replicates (monocytes). c After down-sampling replicates to 10,000 UMIs per droplet to remove variability due to the differences in sequencing depth, the fraction of zeros is computed for each gene and plotted against the log of expression across all samples for the technical replicates data. The solid curve is theoretical probability of observing a zero as a function of the expected counts derived from the multinomial model (blue) and its Poisson approximation (green). d As c but for the biological replicates (monocytes) dataset and after down-sampling to 575 UMIs per cell. Here, we also add the theoretical probability derived from a negative binomial model (red)
Fig. 1 Multinomial model adequately characterizes sampling distributions of technical and biological replicates negative control data. a Fraction of zeros is plotted against the total number of UMI in each droplet for the technical replicates. b As a but for cells in the biological replicates (monocytes). c After down-sampling replicates to 10,000 UMIs per droplet to remove variability due to the differences in sequencing depth, the fraction of zeros is computed for each gene and plotted against the log of expression across all samples for the technical replicates data. The solid curve is theoretical probability of observing a zero as a function of the expected counts derived from the multinomial model (blue) and its Poisson approximation (green). d As c but for the biological replicates (monocytes) dataset and after down-sampling to 575 UMIs per cell. Here, we also add the theoretical probability derived from a negative binomial model (red)CC BY · PubMed Central
Fig. 2 Example of how current approaches to normalization and transformation artificially distort differences between zero and nonzero counts. a UMI count distribution for gene ENSG00000114391 in the monocytes biological replicates negative control dataset. b Counts per million (CPM) distribution for the exact same count data. c Distribution of log2(1+CPM) values for the exact same count data
Fig. 2 Example of how current approaches to normalization and transformation artificially distort differences between zero and nonzero counts. a UMI count distribution for gene ENSG00000114391 in the monocytes biological replicates negative control dataset. b Counts per million (CPM) distribution for the exact same count data. c Distribution of log2(1+CPM) values for the exact same count dataCC BY · PubMed Central
Fig. 3 Current approaches to normalization and transformation induce variability in the fraction of zeros across cells to become the largest source of variability which in turn biases clustering algorithms to produce false-positive results based on distorted latent factors. a First principal component (PC) from the technical replicates dataset plotted against fraction of zeros for each cell. A red to blue color scale represents total UMIs per cell. b As a but for the monocytes biological replicates data. c Using the technical replicates, we applied t-distributed stochastic neighbor embedding (tSNE) with perplexity 30 to the top 50 PCs computed from log-CPM. The first 2 tSNE dimensions are shown with a blue to red color scale representing the fraction of zeros. d As c but for the biological replicates data. Here, we do not expect to find differences, yet we see distorted latent factors being driven by the total UMIs. PCA was applied to 5000 random genes
Fig. 3 Current approaches to normalization and transformation induce variability in the fraction of zeros across cells to become the largest source of variability which in turn biases clustering algorithms to produce false-positive results based on distorted latent factors. a First principal component (PC) from the technical replicates dataset plotted against fraction of zeros for each cell. A red to blue color scale represents total UMIs per cell. b As a but for the monocytes biological replicates data. c Using the technical replicates, we applied t-distributed stochastic neighbor embedding (tSNE) with perplexity 30 to the top 50 PCs computed from log-CPM. The first 2 tSNE dimensions are shown with a blue to red color scale representing the fraction of zeros. d As c but for the biological replicates data. Here, we do not expect to find differences, yet we see distorted latent factors being driven by the total UMIs. PCA was applied to 5000 random genesCC BY · PubMed Central
Fig. 4 GLM-PCA dimension reduction is not affected by unwanted fraction of zeros variability and avoids false-positive results. a First GLM-PCA dimension (analogous to the first principal component) plotted against the fraction of zeros for the technical replicates with colors representing the total UMIs. b As a but using monocytes biological replicates. c Using the technical replicates, we applied t-distributed stochastic neighbor embedding (tSNE) with perplexity 30 to the top 50 GLM-PCA dimensions. The first 2 tSNE dimensions are shown with a blue to red color scale representing the fraction of zeros. d As c but for the biological replicates data. GLM-PCA using the Poisson approximation to the multinomial was applied to the same 5000 random genes as in Fig. 3
Fig. 4 GLM-PCA dimension reduction is not affected by unwanted fraction of zeros variability and avoids false-positive results. a First GLM-PCA dimension (analogous to the first principal component) plotted against the fraction of zeros for the technical replicates with colors representing the total UMIs. b As a but using monocytes biological replicates. c Using the technical replicates, we applied t-distributed stochastic neighbor embedding (tSNE) with perplexity 30 to the top 50 GLM-PCA dimensions. The first 2 tSNE dimensions are shown with a blue to red color scale representing the fraction of zeros. d As c but for the biological replicates data. GLM-PCA using the Poisson approximation to the multinomial was applied to the same 5000 random genes as in Fig. 3CC BY · PubMed Central
Fig. 5 Dimension reduction with GLM-PCA and feature selection using deviance improves Seurat clustering performance. Each column represents a different ground truth dataset from [15]. a Comparison of dimension reduction methods based on the top 1500 informative genes identified by approximate multinomial deviance. The Poisson approximation to the multinomial was used for GLM-PCA. Dev. resid. PCA, PCA on approximate multinomial deviance residuals. b Comparison of feature selection methods. The top 1500 genes identified by deviance and highly variable genes were passed to 2 different dimension reduction methods: GLM-PCA and PCA on log-transformed CPM. Only the results with the number of clusters within 25% of the true number are presented
Fig. 5 Dimension reduction with GLM-PCA and feature selection using deviance improves Seurat clustering performance. Each column represents a different ground truth dataset from [15]. a Comparison of dimension reduction methods based on the top 1500 informative genes identified by approximate multinomial deviance. The Poisson approximation to the multinomial was used for GLM-PCA. Dev. resid. PCA, PCA on approximate multinomial deviance residuals. b Comparison of feature selection methods. The top 1500 genes identified by deviance and highly variable genes were passed to 2 different dimension reduction methods: GLM-PCA and PCA on log-transformed CPM. Only the results with the number of clusters within 25% of the true number are presentedCC 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 476

Related

See all

Metadata

Authors
Townes, F. W., Hicks, S. C., Aryee, M. J., & Irizarry, R. A.
Affiliations
  • Harvard University (US)
  • Johns Hopkins University (US)
ORCID
Venue
Genome Biology
Publisher
Springer Science and Business Media LLC
Published
2019-12
Volume
20
Issue
1
Pages
295-295
ISSN
  • 1465-6906 (ISSN-L)
  • 1474-760X
DOI
10.1186/s13059-019-1861-6
PMC
PMC6927135
OpenAlex
W2998065416
Cited by
471
References
68
Citations by year
2026: 49 · 2025: 56 · 2024: 78 · 2023: 76 · 2022: 70 · 2021: 80 · 2020: 48 · 2019: 14
FWCI
17.71
Type
journal article
Language
en
Update notices
correction (2020-07-22)
Open access
Licence
Locations
  • Genome biology (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
  • National Institutes of Health
  • National Human Genome Research Institute (R00HG009007, R01HG005220, P41HG004059)
  • National Cancer Institute (T32CA009337)
  • National Institute of General Medical Sciences (R01GM083084)
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, feature selection, dimensionality reduction, multinomial model, unique molecular identifiers, highly variable genes, generalized principal component analysis, cell clustering
MeSH terms
Models, Statistical, Sequence Analysis, RNA, 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