Back to Blog
single-cell RNA-seq bioinformatics scientific consulting

When Not to Use Single-Cell: Cost, Replication, and the Questions Bulk Answers Better

Single-cell RNA-seq is often the wrong, expensive default. A practitioner's guide to when bulk answers your question better — and why 'dropout' isn't the reason to reach for it.

When Not to Use Single-Cell: Cost, Replication, and the Questions Bulk Answers Better

A grant gets funded, a new question lands, and the analysis plan opens with “we’ll do single-cell.” It’s a reasonable default, and it got that way for reasonable reasons: reviewers expect it, the field moved there fast, and it’s genuinely transformative for the questions it was built for. Nobody chooses it carelessly — it’s just that the choice increasingly gets made before the question is fully specified.

Sometimes it’s exactly right. Often, though, it’s an expensive route to an answer bulk RNA-seq would have reached more cleanly, with more statistical power, for a fraction of the cost. The most common failure we see isn’t picking single-cell — it’s picking it before anyone has asked where the statistical power is going to come from. That question has a clear answer, and it’s worth asking early, because it’s much cheaper to answer before the tissue is dissociated.

This is the honest counterpart to our single-cell vs. spatial decision framework: before you compare single-cell to spatial, it’s worth asking whether you need single-cell resolution at all.

The cost is not the per-sample sticker price

When people budget single-cell, they budget the library prep and sequencing. That’s the smallest part of the real cost.

The hidden multiplier is replication. Here is the point that trips up the most experiments: if your question is about differences between conditions — treated vs. control, responder vs. non-responder, disease vs. healthy — your statistical power comes from biological replicates (donors, animals, patients), not from the number of cells. Sequencing 20,000 cells from two mice does not give you the power of 20,000 independent observations. It gives you roughly the power of two, with a very detailed picture of those two. Cell number isn’t irrelevant — capture too few cells of a given type and that type’s per-donor estimate gets noisy, and rare populations have a floor below which they can’t be tested reliably — but past that floor, adding cells shows steeply diminishing returns while adding replicates keeps paying off. To make a population-level claim, you need enough independent subjects, and every subject is a full single-cell library. Single-cell doesn’t reduce your sample-size requirement — it multiplies the per-sample cost of meeting it.

The analysis is the other hidden cost. A bulk RNA-seq differential expression project is a well-trodden path: alignment, counts, DESeq2 with a sensible design, quality control, done. A single-cell project adds cell filtering, doublet detection, ambient-RNA correction, normalization, integration across samples, clustering, annotation, and only then the biology — each step a decision that reshapes the result. That’s more analyst time, more compute, and more places for the answer to quietly depend on a parameter nobody wrote down. It’s real value when you need it and pure overhead when you don’t.

”Dropout” is not the reason to reach for single-cell — and not the reason to fear it

A lot of single-cell folklore is organized around “dropout” — the idea that the technology randomly fails to detect expressed genes, leaving technical zeros scattered through the matrix. It shaped a generation of tools built to impute or model away those zeros, and it gets cited as both a weakness of single-cell and a reason you need special methods.

The framing is largely outdated for modern data. For UMI-based droplet protocols — the 10x-style workflows most labs actually run — the evidence is that the data are not zero-inflated. The apparent excess of zeros largely disappears once you resolve cell-type heterogeneity: the zeros look excessive mainly when a mixed population is modeled as one. Account for that, and for the fact that you’re sampling only a small number of transcripts from each cell, and the observed zeros are about what a plain count model predicts — a Poisson within a homogeneous population, or a negative binomial where there’s genuine extra dispersion. Most zeros aren’t a malfunction to be corrected; they’re a lowly expressed gene not being captured in a shallow sample, or a gene that cell genuinely isn’t expressing (Kim et al., Genome Biology 2020; Jiang et al., “Statistics or biology: the zero-inflation controversy,” Genome Biology 2022).

Two practical consequences follow. First, “single-cell has too much dropout” is a weak reason to avoid it — the sparsity is mostly honest low-depth sampling, not noise. Second, and more useful: you generally don’t need zero-inflation-specific models or aggressive imputation on UMI data — imputation carries a real risk of introducing structure that wasn’t in the data, and the benefit is protocol- and question-dependent. The sparsity is a feature of shallow per-cell sampling, and the right response is usually to respect it, not to paper over it.

The false-discovery trap that makes bulk look better than it is

Here is where the choice of technology and the choice of analysis get tangled — and where a substantial share of published single-cell differential expression rests on a design that inflates false positives.

The intuitive way to test “is gene X different between disease and healthy?” in single-cell data is to treat every cell as an independent sample and run a test across all of them. Thousands of cells per group, tiny p-values, long gene lists. It feels powerful. It is a false-discovery machine. Cells from the same donor are not independent replicates; treating them as such conflates within-donor variation with between-condition variation. A benchmark across eighteen datasets with matched bulk ground truth found that every top-performing method aggregated cells within a biological replicate first, while methods comparing individual cells performed poorly — with false positives concentrated in highly expressed genes and in genes that simply varied a lot between replicates (Squair et al., “Confronting false discoveries in single-cell differential expression,” Nature Communications 2021).

The community-recommended fix is pseudobulk: sum each cell type’s counts within each donor to form one profile per sample, then run ordinary bulk DE tools (DESeq2, edgeR) across those donor-level profiles. It’s more conservative, it respects your real replication structure, and — tellingly — it converges on the same statistical footing as a bulk experiment. Pseudobulk is the most robust default, not the only defensible answer: a mixed-effects model with a donor random effect respects the same replication structure and is the main alternative, and can help where cell counts per donor are very uneven — though benchmarks are mixed on whether it beats pseudobulk, with at least one head-to-head finding pseudobulk the more robust choice (Murphy & Skene, Nature Communications 2022). If your endpoint is a cell-type-resolved differential expression list across conditions, single-cell buys you the per-cell-type stratification, not a shortcut around needing real biological replicates. When the cell types are already known and separable, that stratification is sometimes cheaper to get another way.

The questions bulk answers as well or better

Bulk RNA-seq is not the primitive precursor to single-cell. For a large class of questions it’s the more powerful, more defensible tool:

  • Condition-level expression differences in a definable tissue or population. If you want to know what changes between treated and control in a tumor, a sorted cell population, or a whole tissue, bulk gives you every gene at high depth with straightforward, well-validated statistics. You can afford many more biological replicates for the same money — and replicates, not cells, are what power the comparison. See our note on sizing omics experiments before you spend the grant.
  • Homogeneous or pre-sorted populations. If you’ve already isolated your population by FACS or MACS, you’ve done the “single-cell” separation at the bench. Bulk-sequencing sorted populations often answers the question with more depth per gene and none of the doublet or ambient-RNA bookkeeping.
  • Large cohorts, eQTL, and biomarker screens. When N is in the dozens or hundreds — clinical cohorts, population studies — bulk is the only economically sane choice, and depth per sample matters more than single-cell resolution.
  • Known cell types, one dominant signal. If the biology lives in a tissue’s overall expression program and you don’t need to resolve who’s doing what, single-cell resolution is answering a question you didn’t ask.

When single-cell earns its cost

None of this is an argument against single-cell — it’s an argument for spending it deliberately. Single-cell is the right tool, and worth every dollar, when:

  • You’re discovering, not confirming. Unknown cell states, unexpected subpopulations, heterogeneity you can’t pre-sort for.
  • Rare cells carry the biology. A population under a percent or two that a bulk average would erase entirely.
  • Composition is the answer. Shifts in cell-type proportions between conditions — something bulk can only infer through deconvolution.
  • Trajectories and dynamics. Differentiation, pseudotime, cell-state transitions that only exist at single-cell resolution.

For these, there is no bulk substitute, and the analysis overhead is the price of a question you couldn’t otherwise ask. The standard toolchains — Seurat 5, Scanpy, scDblFinder for doublets, ambient-RNA correction where warranted — are mature and worth the investment. The point is to know which question you’re in.

The decision test we actually use

When a client comes to us with “we’re planning single-cell,” we don’t start with the platform. We ask three questions:

  1. What’s the unit of the claim — a cell type/state, or a condition? If the endpoint is a condition-level difference in a known population, bulk with proper replication is often the stronger design.
  2. Do you need to discover populations, or do you already know them? Discovery and heterogeneity justify single-cell; a known, separable population often doesn’t.
  3. Where does your statistical power come from? If the honest answer is “biological replicates,” then the number of cells is largely a resolution choice, not a power choice — enough to clear the per-cell-type floor, and after that budget replicates first, resolution second.

Answer those honestly and the technology usually chooses itself. The failure mode isn’t picking single-cell; it’s picking it by default, discovering the power problem after the money’s spent, and asking us to rescue a comparison that never had the replicates to support it.

The best time to have this conversation is before the tissue is dissociated. If you’re weighing single-cell against bulk for an upcoming study — or trying to get a defensible answer out of data you’ve already generated — our analysis and study-design team is glad to think it through with you.

References

Back to Blog