The problem. Like DESeq, edgeR faces the fact that RNA-seq counts are discrete and overdispersed, and that most experiments have too few replicates to estimate each gene’s variance on its own. Test naively and the low-count genes — the bulk of the table — throw false positives.

The idea. edgeR also uses the negative binomial, but its signature is empirical-Bayes moderation of the dispersion: it estimates a common trend across all genes, then shrinks each gene’s noisy per-gene estimate toward it. The exact test (and later the GLM/quasi-likelihood framework) then gives calibrated p-values even from two or three replicates.

Why it matters. This is the sibling of DESeq2, the engine of my aml_rnaseq_nf pipeline. Reading edgeR next to yesterday’s DESeq paper sharpens the point that the two dominant DE toolkits agree on the count model and differ mainly in how they borrow strength — trended dispersions here, fitted mean-variance shrinkage there. Knowing both is knowing the design space, not just one button.

Verdict. Foundational and still actively used — edgeR and DESeq2 remain the two default choices a decade on. Read it as the other half of a comparison, not a competitor to pick a winner from. Pairs directly with DESeq and the FDR paper.