The problem. RNA-seq counts are heteroscedastic: a gene’s variance depends on its mean and on sequencing depth, which violates the equal-variance assumption of ordinary linear models. Count-based methods (edgeR, DESeq2) model this with negative-binomial GLMs. But the microarray world had years of refined linear-model tools in limma — could RNA-seq reuse them instead of reinventing?

The idea. voom estimates the mean–variance trend of the log-counts empirically, then assigns each observation a precision weight reflecting its reliability. Feed the log-counts and weights into limma’s empirical-Bayes linear-model pipeline, and the heteroscedasticity is accounted for. Suddenly the whole limma toolkit — flexible designs, contrasts, empirical-Bayes moderation — applies to RNA-seq, and simulations show it matches or beats count-based methods.

Why it matters. This is the third major differential-expression philosophy alongside edgeR and DESeq2, all of which I’ve now read — worth knowing because the choice affects results and limma-voom shines with complex experimental designs. It’s also a lesson in method reuse: rather than build anew, weight the data so an existing, battle-tested framework applies. That instinct shows up again in transfer learning on the ML side.

Verdict. A widely-used, flexible DE approach, especially strong for intricate designs; the mean–variance estimate needs enough genes and samples to behave. Read it as RNA-seq differential expression routed through the microarray era’s best linear-model tools.