Finding rare cell populations without inventing them
How to look for a 0.5% population in single-cell data, and the four artefacts that produce a convincing rare cluster that is not real.
Do this on your own data: a population lost at the QC step never reaches the clustering you are trying to fix. The free QC report shows exactly how many cells each threshold removes on your data, per sample: omics.hstgenomics.com/qc-check. A worked example including the small clusters: 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.
Rare populations are where the interesting biology usually is, and also where most irreproducible single-cell findings come from. The difficulty is that the artefacts look exactly like discoveries: a small, distinct cluster with its own marker genes.
First, can you even see it?
Detection is a sampling problem. To recover about 50 cells of a population present at 1%, you need roughly 5,000 cells per sample; at 0.1% you need 50,000. Below about 20–30 cells a cluster will not survive clustering at any sensible resolution, and its "markers" are computed from too few observations to trust. If your target is rare and your budget is fixed, enrichment by sorting beats sequencing more cells — the arithmetic is here.
The four things that fake a rare population
- Doublets. A T cell plus a monocyte forms a consistent hybrid profile, so doublets cluster together and look like a novel intermediate state. Check for co-expression of mutually exclusive lineage markers and for UMI counts near the sum of two parents. Detection done properly.
- Ambient RNA. In a tissue dominated by one cell type, every droplet carries background from it. Clusters defined mainly by that background appear in every sample and go away after correction.
- Stress and dissociation artefacts. Cells that were dying during dissociation share a signature — heat-shock and immediate-early genes, elevated mitochondrial content. That signature clusters, and it is a processing artefact, not a cell type.
- One-sample clusters. If a "rare population" comes almost entirely from a single sample, it is more likely a batch or individual artefact than biology. Always plot cluster composition by sample before believing anything.
The checks that separate real from artefact
- Present in multiple subjects. A real rare type appears, at similar proportion, in most individuals.
- Coherent identity. Its markers should make biological sense together, not be a mixture of two known programmes.
- Stable across resolutions. Real populations persist when you cluster more finely and merge sensibly when you cluster coarsely. Artefacts appear at one setting and vanish at others.
- Recoverable independently. Score the cells for the candidate signature in a second dataset — public data is fine — and see whether the same population emerges.
- Not explained by QC. Compare its QC metrics with the rest. Half the "novel" clusters we are asked to re-examine are cells with distinctive mitochondrial or complexity profiles.
How to report one honestly
Give the proportion per subject, not pooled. Say how many cells it contains. State which checks it passed and which it did not. A rare population reported with its limitations is a finding; the same population reported as a clean discovery is a correction waiting to happen.