The problem. RNA-seq counts are discrete, over-dispersed, and — with two or three replicates per group — desperately under-powered per gene. Naive per-gene tests either miss real signal or drown in false positives. DESeq2 is the standard answer, and it runs inside both my bulk RNA-seq pipelines (via the pydeseq2 port), so I wanted to understand what it does to my numbers rather than trust the volcano plot.

The idea. Three moves. Median-of-ratios normalization corrects for library size robustly, so a handful of highly differential genes don’t skew the size factors. Empirical-Bayes dispersion shrinkage fits a smooth mean–dispersion trend across all genes and pulls each gene’s noisy dispersion estimate toward it — borrowing strength across the transcriptome to survive tiny replicate counts. Log-fold-change shrinkage does the same for effect sizes, pulling low-count genes’ wild LFCs toward zero so the ranking isn’t dominated by noise. A Wald test and Benjamini–Hochberg FDR finish it.

Why it matters. Reading the source changes how I present my own results. That “shrinkage” is why a gene with a huge raw fold change can rank below one with a modest, well-measured change — and being able to explain that is the difference between running a tool and understanding it. It’s also a clean lesson in empirical Bayes: share information across features to stabilize per-feature estimates, an idea that recurs everywhere in genomics.

Verdict. Deservedly canonical, and honest about its model assumptions (negative binomial, a fitted dispersion trend that can misbehave for very low counts or unusual designs). Next time I’ll annotate one of my pipeline’s volcano plots with the pre- and post-shrinkage LFCs as the worked example.