Dear edgeR/limma team,

I read through the user guides and papers of limma and egdeR that were published over the years, especially the latest edgeR v4 and voomByGroup papers. While I basically understand the conceptual differences and assumptions of each tool, is there a general and intuitive recommendation in which situations (for example RNA-seq and scRNA-seq) one would choose edgeR v4 over flavour of limma/voom/voomByGroup and vice versa?

Thank you!

The short answer is that we generally prefer edgeR for applications with lots of small counts and limma for complex designs with random effects and outlier samples. But there is a large middle ground for which both packages are equally as good. For limma, we recommend the voomLmFit() pipeline, which depends under the hood on edgeR as well.

We generally prefer edgeR for Hi-C, ChIP-seq, ATAC-seq, CUT&Tag, methylation, transcript-specific analyses, and haplotype-specific analyses. We generally use limma for a wide range of other RNA-seq experiments.

With edgeR v4, it is possible to keep a lot more low count genes in an RNA-seq analysis than we usually do with limma, and that is something we haven't fully explored yet. It surely will have implications for single-cell and spatial RNA-seq.

We continue to develop both packages, and the relative performances will depend on the latest improvements. We recently compared edgeR to limma for DTU analyses (Baldoni et al 2025). edgeR (v4 QL) and limma (voomLmFit) were both better than any of the competitors, with edgeR slightly better than limma. However, limma would have beaten any version of edgeR prior to April 2025, and edgeR and limma gave very similar results on the real data case study.

Reference

Baldoni PL#, Chen L#, Li M, Chen Y, Smyth GK (2025). Dividing out quantification uncertainty enables assessment of differential transcript usage with limma and edgeR. Nucleic Acids Research 53(22), gkaf1305. https://doi.org/10.1093/nar/gkaf1305

I would use edgeR's quasi-likelihood model for any analysis with a smaller number of observations (like 3 vs 3 or similar), but if you have more samples, or if you have nesting that will require a linear mixed model, then limma-voom is the way to go. These days I rarely use the quasi-likelihood model.