thebench.bio #1 Ecosystem for the Bio-Tech Industry
All RNA-seq questions

Question 17 of 20 Advanced Interpretation

How would you test whether a gene expression difference is driven by changing cell composition?

Short answer

Check whether the changed genes are markers of particular cell types, estimate cell-type proportions by deconvolution against a matched reference, and see whether the effect survives adjusting for them. Confirm with an independent measure such as histology or flow cytometry when it matters.

Why

Bulk RNA-seq measures a weighted sum of every cell in the sample. If a treatment recruits immune cells or kills tumor cells, bulk expression changes even when no single cell changes its expression. The first clue is the gene list itself: a block of T cell or neutrophil markers rising together points to cell numbers rather than regulation.

Deconvolution estimates the proportions. In a benchmark of deconvolution pipelines, Avila Cobos and colleagues found that methods worked best on data kept in linear scale rather than log scale, that the normalization choice had a dramatic effect on some methods, and that no method gave accurate proportions when the reference lacked a cell type present in the mixture. The estimated proportions can then go into the design as covariates, to ask whether the treatment effect remains once composition is accounted for.

This is an interpretation tool, not a correction to apply blindly. If the treatment's real effect is to change composition, adjusting for composition removes the very effect of interest, so the conclusion is about mechanism, not about whether the result is real.

How to work it through

  1. Look at the top changed genes and ask whether they are known markers of one cell type.
  2. Choose a reference profile from the same tissue and species, ideally single-cell data, that includes every expected cell type.
  3. Run deconvolution on counts in linear scale and check that estimated proportions are plausible against histology or known biology.
  4. Compare proportions between groups.
  5. Refit the differential expression model with the proportions as covariates and see which genes keep their effect.
  6. Confirm key composition changes with an orthogonal measurement.

An example

In a hypothetical drug study on liver biopsies, 300 genes rise with treatment, and many are macrophage and neutrophil markers. Deconvolution estimates immune cells at 8 percent of the signal in controls and 15 percent in treated samples. With the immune proportion in the design, 240 of the 300 genes lose significance, while a set of 45 hepatocyte metabolic genes remain.

The report says the drug increases immune infiltration and separately changes hepatocyte metabolism, and a pathologist's count of immune cells on stained sections confirms the infiltration.

Where the answer stops holding

Deconvolution estimates are noisy, and they fail when the reference misses a cell type or comes from a different tissue state. Covariate adjustment is unreliable when composition is nearly collinear with treatment. Cell states that change expression within one cell type cannot be separated from composition by bulk data alone.

A strong answer covers

  • Bulk expression is a weighted sum of cell types
  • Marker genes in the hit list are the first clue
  • Deconvolution works best in linear scale with a complete, matched reference
  • Adjusting for composition answers a mechanism question and can remove a real effect

Common mistakes

  • Reporting increased marker gene expression as upregulation within cells
  • Using a reference that lacks a cell type present in the tissue
  • Adjusting for composition by default and discarding a treatment effect that works through composition