thebench.bio #1 Ecosystem for the Bio-Tech Industry
All Single-Cell Analysis questions

Question 18 of 20 Advanced Practical

How would you design a single-cell experiment to detect a rare population without overclaiming its prevalence?

Short answer

Size cell numbers from the expected frequency so enough of the rare cells are captured, replicate across donors, and keep enrichment separate from prevalence estimates. Report prevalence per donor with uncertainty, measured in unenriched samples or by an orthogonal method.

Why

Detection and prevalence are different questions. To detect a population at 0.5 percent, you need enough cells that it forms a recognizable group. Capturing 5,000 cells gives about 25 expected rare cells, with a Poisson standard deviation of about 5, which may be too few to cluster cleanly.

Enrichment, such as sorting on a surface marker, makes detection much easier but destroys the original proportion, so prevalence must come from unenriched data or another method such as flow cytometry. Replication matters as much as depth. A population seen in one donor might be donor-specific, and cell-to-cell counts within one sample say nothing about donor-to-donor variability.

Abundance should then be tested across donors with methods that respect replicates, such as edgeR on cell counts or Milo neighborhoods, as the Heumos review describes. Analysis choices affect rare cells too. The same review notes analytical Pearson residuals suit identifying rare cell identities, and doublets of two common cell types can imitate a rare population, so doublet detection is essential.

How to work it through

  1. Write down the expected frequency and the minimum number of rare cells needed to call a cluster.
  2. Compute the total cells per sample needed to reach that number with margin.
  3. Plan several donors per condition so prevalence can be estimated across individuals.
  4. If you enrich by sorting, run an unenriched aliquot or parallel flow cytometry for prevalence.
  5. Remove doublets and low-quality cells before clustering.
  6. Report prevalence per donor with confidence intervals and confirm the population with an orthogonal marker.

An example

A team expects a population at about 0.3 percent of lung immune cells and wants at least 50 such cells per donor. That needs roughly 17,000 immune cells per donor, so they sort CD45-positive cells and load two lanes per sample, and also run flow cytometry on an unsorted aliquot. Across six donors, single-cell data detect 40 to 90 rare cells each.

Flow cytometry gives a prevalence of 0.2 to 0.5 percent, which they report alongside a per-donor confidence interval rather than a single pooled number.

Where the answer stops holding

Dissociation and capture can under-sample fragile cells, so even unenriched data may misstate prevalence. Poisson reasoning assumes random capture, which does not hold when cells clump. A population defined only by transcriptome may have no clean surface marker for flow confirmation.

A strong answer covers

  • Size cell numbers from expected frequency
  • Separate enrichment for detection from prevalence measurement
  • Replicate across donors and test across them
  • Remove doublets that can mimic rare populations
  • Report per-donor prevalence with uncertainty and orthogonal confirmation

Common mistakes

  • Reporting prevalence from a sorted, enriched sample
  • Pooling cells from all donors and quoting one percentage
  • Assuming a small cluster is real without checking doublet scores