How would you model a paired RNA-seq design where each donor gives a treated and a control sample?

Why

People differ from one another in baseline expression far more than many treatments change it. In an unpaired analysis that donor variation sits in the residual, widening the variance and hiding real effects. The DESeq2 documentation answers this directly: for paired samples, use a multi-factor design that includes the sample or donor as a term, and put the condition of interest at the end of the formula.

The fitted log fold change for condition is then an average of within-donor changes. The Bioconductor RNA-seq workflow uses exactly this structure for four airway cell lines, each treated and untreated, with the design ~ cell + dex. The same idea covers any blocking factor, such as a sequencing batch that holds both conditions.

Two limits are worth knowing. If a variable never varies within a donor, such as disease status in a patients against healthy controls study, a simple donor term absorbs it and the model cannot estimate it. DESeq2 reports this as a model matrix that is not full rank, and its documentation describes a nested design for individuals within groups.

Pairing also only helps if the pairs are real, so sample labels must be trustworthy.