henhaobio
Sign in

Nucleic Acids ResearchBioinformaticsPeer-reviewed journal article

edgeR v4: powerful differential analysis of sequencing data with expanded functionality and improved support for small counts and larger datasets

Yunshun CheniD, Lizhong CheniD, Aaron T. L. LuniD, Pedro L. BaldoniiD, Gordon K. SmythiD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (24 Nov 2024) · CC BY 4.0

edgeR v4 adds a continuous negative binomial for fractional counts, bias-corrected quasi-likelihood for small counts, C-based fitting, and new transcript, exon, methylation and fold-change-threshold analyses.

Classical quasi-likelihood theory approximates deviances with a chisquare law that underestimates quasi-dispersions when average counts fall below 1, and fractional transcript counts otherwise need rounding. Version 4 corrects this bias, scales to larger designs and extends a widely used negative binomial GLM framework to transcript, splicing, methylation and pathway analyses.

Key results

  1. SuggestedReplacing the saddlepoint-based chisquare approximation with moment-matched scaled chisquare deviances on fractional df gives nearly unbiased quasi-dispersion estimates, even when fitted means are near zero or NB-dispersions are large.

    Simulated Poisson counts (100 samples, 10,000 genes, true quasi-dispersion 1): classical mean-deviance estimates are biased negatively or positively depending on expected count, while v4 adjusted deviances are unbiased at all count sizes. Relies on limma 3.62.0 fitFDistUnequalDF1, called by glmQLFit in edgeR v4.4.0. · Correcting quasi-likelihood for bias; Figure legends

  2. ClaimedA continuous generalization of the NB distribution lets edgeR model non-integer counts without rounding. This makes divided-count transcript-level DE possible: overdispersion from read-to-transcript ambiguity is estimated from resamples and divided out.

    Gamma functions replace binomial coefficients, so probabilities match NB values at integer counts and vary smoothly between them. catchSalmon, catchKallisto and catchRSEM estimate per-transcript ambiguity quasi-dispersions from bootstrap or Gibbs samples. This article gives no benchmark numbers; the performance gains are cited to prior work. · Divided counts; Supporting fractional counts; Transcript-level differential expression

  3. SuggestedThe v4 QL pipeline can skip estimateDisp by using a constant NB-dispersion taken from highly expressed genes. Low-level fitting is now pure C, and a CompressedMatrix class saves memory by storing only the repeat structure of offsets, weights and dispersions.

    A oneway multi-group analysis of 1000 samples and 10000 genes runs in about 30 seconds on a laptop, covering dispersion estimation, GLM fitting and DE testing. A more complex design takes under two minutes. No hardware details or timing comparison with v3 or other tools are reported. · Quasi-likelihood; Larger datasets; C implementation; CompressedMatrix class saves memory

  4. ClaimedglmTreat brings limma's TREAT to NB GLMs for testing against a fold-change threshold. It uses the expected rather than the maximal type I error rate as the p-value, which gains power while keeping FDR control.

    glmTreat works under both the LRT and QL F-test pipelines. The authors contrast it with ad hoc filtering by fold-change cut-off and then p-value, which favours low-expression genes and can lose FDR control. This article presents no simulation or power figures. · Testing differences relative to a fold-change…

  5. ShownedgeR serves as an analysis engine across Bioconductor, is extensively documented, and now accepts SummarizedExperiment containers directly.

    212 Bioconductor packages list edgeR as a dependency or suggestion (November 2024). v4.4.0 has 253 documented exported functions or methods and a 138-page User's Guide with 10 worked case studies (gene, exon, transcript, time course, BS-seq, CRISPR-Cas9 and shRNA-seq screens). · Object-oriented programming; Documentation; Discussion

Methods, data and reproducibility

edgeR models counts with log-linear quasi-negative-binomial GLMs. It moderates dispersions with weighted-likelihood and limma-style empirical Bayes and tests with quasi-F or likelihood ratio tests. v4 adds a continuous NB for fractional counts, moment-matched deviance bias correction, divided counts and pure C fitting.

  • TCGA breast tumour RNA-seq genewise counts, classified into five intrinsic subtypes by PAM50 signatures; used for the MDS plot example · human · n = 882 samples
  • Mammary gland RNA-seq (Fu et al., analysed by Chen et al.) used for BCV, quasi-dispersion and MD plots; lactating vs pregnant basal cells · mouse · GSE60450
  • RNA-seq experiment (Fu et al.) with two Foxp1 exons silenced in a knockout group; used for the differential exon usage plot · GSE118617
  • Simulated Poisson counts with true quasi-dispersion 1, used to compare classical and bias-adjusted deviance estimators · n = 100 samples, 10,000 genes
  • Oneway multi-group dataset used to illustrate runtime of the v4 QL pipeline · n = 1000 samples, 10000 genes
Methods & tools
edgeR v4 (glmQLFit, glmQLFTest, glmTreat, diffSpliceDGE, filterByExpr, normLibSizes), limma (empirical Bayes, fitFDistUnequalDF1, TREAT), limma gene set tests (goana, kegga, camera, fry, roast, mroast, romer), TMM and TMMwsp normalization, Salmon, kallisto, RSEM, Bismark, Subread featureCounts / HTSeq-count, Seurat (via Seurat2PB pseudo-bulk aggregation), 10x Genomics Cell Ranger (via read10X), Bioconductor SummarizedExperiment
Software & environment
R/Bioconductor package; edgeR v4 released October 2023. v4.4.0 is converted to pure C with no Rcpp dependency. The new QL pipeline relies on limma 3.62.0 (fitFDistUnequalDF1).
Assay conditions & compute
About 30 seconds on a laptop for a oneway analysis of 1000 samples by 10000 genes; under two minutes for more complex designs. No CPU or memory specifications are given.
Reproducibility gaps
No scripts are provided for the Figure 1 simulation or the runtime illustration, and the laptop hardware is unspecified.
Limitations (5)
  • The classical chisquare deviance approximation underestimates quasi-dispersions for genes with average count below 1, which is common in single-cell data with large NB-dispersions. The v4 correction yields residual df that are fractional and can fall below 1. · stated by the authors
  • Gene-level pipelines are suboptimal on raw transcript counts: read-to-transcript ambiguity disrupts the mean-variance trend and weakens empirical Bayes dispersion estimation. The divided-count remedy needs bootstrap or Gibbs resamples from the quantifier. · stated by the authors
  • Very low-count genes must be filtered out. They cannot reach significance even when truly differential, and NB-dispersion estimation is unreliable for them. · stated by the authors
  • As a software review, the article shows almost no head-to-head benchmarks. Claims of better performance for whole-genome BS-seq, pseudo-bulk single-cell DE and transcript DE rest on cited studies, not data shown here. · noted in this summary
  • Scalability is supported by one laptop timing, with no hardware details and no comparison against edgeR v3 or other tools, so the speed gains of the C rewrite and the memory savings of CompressedMatrix are not quantified. · noted in this summary

Read this if you run bulk, pseudo-bulk single-cell, transcript-level or methylation differential analyses in edgeR and need to know what changed in the v4 quasi-likelihood pipeline.

Citations
860
FWCI
246.57
Field percentile
top 1%

Abstract

edgeR is an R/Bioconductor software package for differential analyses of sequencing data in the form of read counts for genes or genomic features. Over the past 15 years, edgeR has been a popular choice for statistical analysis of data from sequencing technologies such as RNA-seq or ChIP-seq. edgeR pioneered the use of the negative binomial distribution to model read count data with replicates and the use of generalized linear models to analyse complex experimental designs. edgeR implements empirical Bayes moderation methods to allow reliable inference when the number of replicates is small. This article announces edgeR version 4, which includes new developments across a range of application areas. Infrastructure improvements include support for fractional counts, implementation of model fitting in C, and a new statistical treatment of the quasi-likelihood pipeline that improves accuracy for small counts. The revised package has new functionality for differential methylation analysis, differential transcript expression, differential transcript and exon usage, testing relative to a fold-change threshold and pathway analysis. This article reviews the statistical framework and computational implementation of edgeR, briefly summarizing all the existing features and functionalities but with special attention to new features and those that have not been described previously.

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
THISDimont '14Pachter '26Li '22Love '14Varet '15Baldoni '24
71% semantic matchBioinformatics
Emmanuel Dimont, Jiantao Shi, Rory Kirchner+1 more·bioRxiv·Aug 25, 2014