Field Guide home

Week 8 ยท 7 minutes

Differential Expression: What It Tells You and What It Does Not

A practical interpretation guide for RNA-seq results

Beginner friendly Tested commands Reproducible assets

Differential Expression: What It Tells You and What It Does Not

Takeaway: Differential expression can tell you which genes changed between conditions, but it cannot by itself prove mechanism, causality, or clinical relevance.

What The Test Actually Asks

Differential expression asks whether observed gene counts differ between groups more than expected under a statistical model.

It depends on:

  • count data
  • sample metadata
  • normalization
  • dispersion estimation
  • study design
  • multiple-testing correction

The result is model-based evidence. It is not automatic biology.

The Columns People Misread

Column Meaning Common mistake
log fold change size and direction of change treating tiny effects as important
p-value evidence under the model ignoring multiple testing
adjusted p-value corrected evidence across many tests treating it as biological truth
base mean average expression level ignoring low-expression instability

The most useful genes are often not just statistically significant. They are statistically supported, biologically plausible, and relevant to the question.

A Better Interpretation Rubric

For each result, ask:

  1. Is the gene expressed enough to trust the estimate?
  2. Is the fold change large enough to matter?
  3. Is the adjusted p-value convincing?
  4. Is the direction biologically plausible?
  5. Is the signal consistent across samples?
  6. Is the result robust to reasonable preprocessing choices?
  7. Does independent evidence support it?

Volcano Plots Are Not Conclusions

Volcano plots are useful summaries, but they can encourage shallow interpretation. A labeled point on a volcano plot is a hypothesis, not an answer.

Always inspect:

  • normalized counts
  • sample-level plots
  • batch structure
  • pathway context
  • known markers or controls
  • outliers

A Tiny Example

Suppose GENE1 has:

gene log2 fold change adjusted p-value base mean
GENE1 0.18 0.00001 900
GENE2 2.4 0.04 6
GENE3 1.2 0.003 180

GENE1 is statistically strong but may have a small effect. GENE2 has a large effect but low expression, so the estimate may be unstable. GENE3 may be the more interesting first follow-up if the biology fits.

That is the point: interpretation needs more than one column.

Design Formula Matters

A design formula tells the model what comparison to test and what covariates to account for.

Examples:

~ condition
~ batch + condition
~ donor + condition

If batch is confounded with condition, no formula can magically create a clean experiment. If donor pairing matters and is ignored, the result can be misleading.

Common Mistakes

  • Ranking genes by p-value only.
  • Ignoring effect size.
  • Testing too many contrasts without a plan.
  • Using the wrong design formula.
  • Not accounting for paired samples.
  • Treating adjusted p-value cutoffs as universal laws.
  • Overinterpreting pathway enrichment.
  • Making clinical claims from exploratory results.

Save This: DE Result Triage

Check Good sign
expression not driven by near-zero counts
effect size large enough to matter biologically
adjusted p-value survives multiple-testing correction
sample consistency not driven by one outlier
design covariates match the experiment
biology direction fits known or testable biology
validation independent data or experiment can support it

What To Watch Next

Experts still debate shrinkage, filtering, complex designs, pseudobulk single-cell differential expression, and when pathway-level interpretation should be primary or secondary. A careful analyst treats differential expression as a prioritized hypothesis list, not as final proof.

Credits and References

Join the discussion on GitHub.

Ask a question, suggest an example, or share how you used this guide. Comments are powered by GitHub Discussions through giscus.