How many cells, and how deep? Designing a single-cell experiment you can afford
Power thinking for single-cell studies: how many cells per sample, how many reads per cell, and why the number of subjects matters more than either.
Do this on your own data: if you already have a pilot run, its own numbers answer this better than any rule of thumb. Type five numbers from the Cell Ranger summary and get saturation and expected doublets computed in your browser: omics.hstgenomics.com/qc-self-check. A worked analysis at a typical depth: omics.hstgenomics.com/sample.
Want it done rather than explained? We do analysis only — send a count matrix, an h5ad or Seurat object, Cell Ranger output or a public accession and get publication-ready results with the code that produced them, at a fixed price agreed before any work starts: omics.hstgenomics.com/quote.
Three numbers decide what a single-cell experiment can answer: how many subjects, how many cells per subject, and how many reads per cell. Most planning goes into the second, most of the statistical power comes from the first, and the third is where money is wasted.
Subjects beat cells
If your question is "does this differ between conditions", your sample size is the number of animals, donors or patients — not the number of cells. Twenty thousand cells from two mice per group cannot answer it; two thousand cells from six mice per group can. Cells give you resolution within a sample; subjects give you the ability to generalise beyond it.
Practical floor: three subjects per group for an exploratory study, five or more when the effect is expected to be modest or the biology is variable. Human tissue is more variable than inbred mice, so the human floor is higher.
Cells per sample
The number you need depends on the rarest population you care about. To see roughly 50 cells of a population that makes up 1% of the tissue, you need about 5,000 cells per sample; for a 0.1% population you need 50,000. Work backwards from the rarest thing you must quantify, not from the platform's maximum.
Above about 10,000 cells per lane the doublet rate becomes the limiting factor — roughly 0.8% per 1,000 cells recovered on standard chemistry, about half that on high-throughput kits — so loading harder buys progressively less usable data. If you need many cells of a rare type, enrich by sorting rather than by loading more.
Reads per cell
The standard target for 3' gene expression is 20,000–50,000 reads per cell. Below about 10,000 you start losing lowly expressed genes; above 50,000 the curve is flat for most questions, because the library saturates.
The way to check is sequencing saturation in the Cell Ranger summary: under 50% means more reads would still add information, over 80% means more reads are mostly duplicates. Sequence one sample, look at saturation, then decide the depth for the rest. This single habit routinely saves more than it costs.
Nuclei versus cells
If the tissue does not dissociate well — brain, heart, muscle, frozen archival material — single-nucleus is the honest choice. Expect fewer genes per nucleus (often 1,000–3,000 versus 2,000–5,000 for cells), loss of cytoplasmic transcripts, and a different mitochondrial profile. What you gain is representation: dissociation-resistant cell types that would otherwise simply be missing from your data.
Where the budget usually goes wrong
- Two subjects per group, many cells each. Statistically unanswerable, however beautiful the UMAP.
- Sequencing every sample deeply before checking saturation on one.
- No pilot. A single lane on one sample tells you dissociation quality, ambient levels and cell recovery, and routinely changes the design.
- Budgeting nothing for analysis, then having no one to analyse the data you paid for — what to put in the grant.
A sane default for an exploratory study
Four to six subjects per group, 5,000–8,000 cells per subject, 25,000–30,000 reads per cell, one pilot sample sequenced and inspected before committing the rest. It is not the most cells anyone has ever published. It is a design that answers a question.