How many samples do you need for single-cell differential expression?

This is the question worth asking before the money is spent, and it is usually asked after. The short answer is that the number of cells barely matters for a between-group comparison, and the number of biological samples almost entirely determines what you can claim.

The short version: for a group comparison, plan on a minimum of three biological replicates per group and prefer five or more; adding cells to a two-versus-two design does not fix it. Cells buy you resolution within a sample — rarer populations, finer states. Donors buy you the ability to generalise.

Why cell count is the wrong knob

Treating every cell as an independent observation is the single most common statistical error in the field. Cells from the same donor are not independent: they share genotype, treatment history, batch, dissociation stress and capture chemistry. Pooling them produces enormous nominal sample sizes and correspondingly tiny p-values, and the result does not replicate.

The effect is measurable, and we measured it on a public dataset with eight donors profiled with and without interferon-β stimulation. Comparing cell-type proportions between conditions:

• Pooling all cells and testing: 4 cell types significant.
• The same data, tested at the donor level with each donor's own before and after: 0 significant.

All four were artefacts of pooling. In one of them the proportion moved from 24.5% to 22.5% — a two-point shift that looked convincing until it was tested against how much donors differ from each other.

What the donor count buys you

With a paired design (each donor contributes both conditions), the between-donor variability is removed from the comparison, and a handful of donors goes surprisingly far. With unpaired groups, you are estimating that variability from the data, and small numbers do it badly.

• n = 2 per group. Not analysable as a group comparison. You can describe what you saw; you cannot claim it generalises, and most reviewers will say so. Any analyst who runs a differential test on this without saying that is selling you a number, not an answer.
• n = 3 per group. The practical minimum. Fine for a strong, consistent effect such as a stimulation response. Underpowered for anything subtle, and one outlier donor can dominate.
• n = 5–6 per group. Where most designs should sit. Robust to one unusual sample and enough to say something about consistency across individuals.
• n ≥ 10 per group. Needed for patient cohorts, where the biological variability between people is the thing you are trying to see through rather than a nuisance.

How many cells, then?

Enough that every cell type you intend to test is present in reasonable numbers in every sample. A useful rule: aggregate each donor's cells within a cell type into one pseudobulk profile, and require at least around 20 cells per donor per cell type before that donor contributes to a test. In the eight-donor dataset above, the monocyte pseudobulk samples ranged from 126 to 779 cells — all usable. If a cell type would leave several donors below the floor, either capture more cells per sample or accept that the type is out of scope for the comparison.

Sequencing more deeply per cell does not help a group comparison at all. It helps detection of low-expressed genes within a cell, which is a different question.

The design that survives review

1. Pair where you can. Same donor before and after beats two independent groups, for the same total cost.
2. Spread conditions across batches. If every control was run on Monday and every treated sample on Friday, no amount of statistics separates the treatment from the day.
3. Decide the unit of replication before you start, and write it in the methods: "n refers to donors, not cells".
4. Aggregate to pseudobulk for the test. On the same public data, pseudobulk and a naive per-cell test disagreed on a substantial fraction of genes; the per-cell list is the optimistic one.

Worked examples with the code and every number are published on our pseudobulk sample and differential abundance sample. If you are planning a study and want a design checked before you spend the sequencing budget, send the details — that conversation is free and it is much cheaper than the alternative.