Can deconvolution of your existing bulk RNA-seq answer the question — or do you need single-cell?

If you already have bulk RNA-seq, deconvolution — estimating cell-type proportions computationally — costs a fraction of a new single-cell experiment. Sometimes that is the right answer and the money should not be spent. Sometimes it will give you a number that is confidently wrong. The difference is predictable, and worth knowing before you commit a budget.

The short version: deconvolution is reliable for large shifts in abundant cell types and unreliable for small differences, rare populations, and any question about cell state rather than cell proportion. In our benchmark against known ground truth, overall correlation was r = 0.77 and typical error was 7.5 percentage points — but one abundant cell type was overestimated by more than nine points.

What the benchmark measured

Most deconvolution comparisons are circular: the reference and the test data come from the same samples, so the method is graded on data it has effectively seen. We split eight donors into four whose cells built the reference and four whose cells built 30 synthetic mixtures with known composition. The method never saw the test donors. That is the honest version of the question — how well does it do on someone else's tissue?

• Overall correlation between estimated and true proportions: r = 0.77.
• Typical error: 7.5 percentage points.
• Best behaved: B cells (r = 0.93, error 6.2 points).
• Worst: CD14+ monocytes — true mean 15.2%, estimated mean 24.6%. A systematic overestimate of more than nine percentage points, in the most abundant myeloid population.

When deconvolution is enough

• Large, expected shifts in common cell types. A treatment that doubles a population from 10% to 20% will be visible even with several points of error.
• Screening a large cohort. If you have 200 bulk samples and want to know which ones look immunologically different, deconvolution is the only affordable option, and being approximately right across 200 samples beats being exactly right about three.
• Generating a hypothesis to test properly later. As a cheap first pass it is entirely defensible, as long as the writeup says that is what it is.

When it is not

• Differences smaller than the method's error. A shift from 15% to 18% is inside the noise. Reporting it as a finding is reporting the method, not the biology.
• Rare populations. Cell types below a few percent are estimated poorly and inconsistently; a doubling of a 1% population is not recoverable.
• Anything about cell state. Deconvolution returns proportions. If the question is whether a cell type changed — activated, exhausted, stressed — proportions cannot answer it, and no reference choice fixes that.
• Tissues without a good matched reference. Accuracy depends on how well the reference represents your tissue. PBMC references are mature; many solid tissues are not.

The practical decision

1. Write down the smallest difference that would change your conclusion. If it is under about ten percentage points in an abundant cell type, deconvolution alone will not settle it.
2. Ask whether the question is about how many or about what state. Only the first is a deconvolution question.
3. If deconvolution is used, insist that the report includes a benchmark on held-out data — not a correlation computed on the same samples that built the reference.
4. A common middle path: single-cell on a small number of representative samples to build a tissue-matched reference, then deconvolve the existing bulk cohort against it. That is usually far cheaper than running single-cell on everything, and it is a genuinely different answer from either extreme.

The benchmark above, with the code and per-cell-type numbers, is on the deconvolution sample page. If you have bulk data and are deciding whether to spend on single-cell, describe the question — knowing which of the two you need is worth more than either analysis done blind.