4,000 'significant' genes, then 212 real ones

The problem

The lab compared microglia from three knockout and three wild-type mice using a cell-level Wilcoxon test. With thousands of nuclei per group, nearly every expressed gene reached significance, and the pathway analysis pointed everywhere at once. A reviewer asked for a 'more appropriate statistical framework that accounts for biological replicates'.

What we found

• The test treated each nucleus as an independent replicate, so the effective sample size was inflated from 6 animals to more than 9,000 nuclei.
• Between-animal variance was large. One wild-type mouse was an outlier with a strong interferon signature that drove a large share of the 'hits'.
• Ambient neuronal transcripts in microglia nuclei added further false positives.

How we rescued it

1. Corrected ambient RNA with CellBender and rebuilt the microglia annotation.
2. Aggregated counts per animal and cell type (pseudobulk) and tested with DESeq2, with a sensitivity analysis excluding the outlier animal.
3. Confirmed results with a mixed model with animal as a random effect (NEBULA).
4. Ran fgsea on shrunken fold changes, which converged on a coherent lipid-metabolism and phagocytosis program.

The result

The final analysis found 212 robust differentially expressed genes, consistent across methods and with or without the outlier animal. The central biological claim, a defect in lipid handling, became clearer and was supported by the lab's existing lipid-droplet staining.