The reviewer asked you to re-analyse your single-cell data. Now what?

Almost every single-cell manuscript comes back with at least one analysis request. The panic is usually worse than the work — most requests fall into five buckets, and only one of them genuinely means starting over.

1. "Your QC thresholds are arbitrary"

This is the most common and the cheapest to answer. Reviewers rarely want different thresholds; they want evidence that your conclusions do not depend on them. The winning response is a small sensitivity table: re-run clustering at two or three alternative cut-offs (for example mitochondrial <5%, <10%, <15%) and show that cluster number and the marker genes of your key population are stable.

Report the numbers per sample, not pooled. If one sample loses 40% of its cells while others lose 5%, say so — volunteering that asymmetry buys more credibility than hiding it, and it usually explains a batch effect the reviewer is about to point out anyway.

2. "How do you know these are cell type X?"

Annotation challenges are answered with convergent evidence, not with more marker plots of the same three genes. Three independent lines usually settle it: canonical markers, an automated reference-based label transfer (SingleR, Azimuth, CellTypist) run blind to your manual labels, and a gene-set score. When all three agree, show them side by side in one figure. When they disagree, that disagreement is the finding — an ambiguous population deserves a sentence in the results, not a confident label.

3. "Please correct for batch"

The trap here is over-correction. Integration methods will happily merge biologically distinct samples into one smooth blob. Before you integrate, ask what the batch actually is: different chemistry versions and different days are real technical batches; treated versus untreated is not. Run the integration, then check that a population you know should be condition-specific has not been dissolved. If it has, you have removed your own result.

4. "Add trajectory / cell–cell communication / pathway analysis"

These are additive modules, not re-analyses. Each is a day or two of work on an already-processed object. The mistake is treating their output as a result rather than a hypothesis: a pseudotime ordering is a hypothesis about progression, a ligand–receptor score is a hypothesis about interaction. Phrase them that way in the text and reviewers stop arguing with them.

5. "Your differential expression is not statistically valid"

This one is real, and it is the request most often answered wrongly. Comparing thousands of cells from three mice against thousands from three other mice with a per-cell test treats cells as independent replicates — they are not. Your n is the number of animals or donors, not the number of cells. Aggregate counts per sample per cell type into a pseudobulk matrix and test with a bulk method (DESeq2, edgeR, limma-voom). The gene list will shrink, sometimes dramatically. That is not a loss; that is the false positives leaving.

How long each of these really takes

• QC sensitivity table — half a day, mostly re-running an existing pipeline.
• Annotation evidence figure — half a day if the object is processed.
• Batch integration plus sanity checks — one to two days.
• Trajectory / communication / enrichment module — one to two days each.
• Pseudobulk differential expression redo — one day, plus rewriting whatever conclusions change.

The one thing that saves you every time

Whether you do this yourself or hand it to someone else, insist that the analysis is reproducible: the script, the parameters, and the software versions stored with the results. Round two of review comes three months later, when nobody remembers which resolution parameter produced figure 3. If you cannot re-run it in an afternoon, you will end up redoing it from scratch.