Cells are not replicates: permutation calibration of the unit-of-analysis error in a pooled sleep single-cell design (GSE137665)

Background. In single-cell RNA sequencing the number of cells is not the number of independent observations for a between-animal intervention. When animals are pooled before capture, the animal label is destroyed and the between-animal variance component becomes unestimable. Methods. We re-analysed GSE137665 (mouse brainstem, cortex and hypothalamus across normal sleep, 12-h sleep deprivation and recovery sleep). The deposited matrix holds 29,571 cells × 18,634 genes in nine libraries, each a suspension pooled from three animals. We fitted the same region-adjusted model for the sleep-deprived vs normal-sleep contrast at two units of analysis - individual cells (df = 29,566) and library means (n = 9, df = 4) - and calibrated both by permuting condition labels within brain region (216 assignments). We also decomposed the variance into within- and between-library components, and audited the study's RNAscope validation layer and its own deposited differential-expression output. Results. Under the permutation null the cell-level test returned a mean of 6,324 genes at p < 0.05 (median 6,489), an empirical type-I error rate of 37.7% at a nominal α = 0.05; the library-level test returned 3.0%. The observed cell-level result (4,688 genes) lies below its own null mean (permutation p = 0.759). The Kruskal-Wallis test used by the source study gives a median per-gene p-value of 1.9 × 10⁻⁴ , against 0.5 for a calibrated test. At the library level no gene passes FDR < 0.1 or Bonferroni; the smallest effect detectable at 80% power (0.278) exceeds the 99th percentile of observed effects (0.113). Point estimates agree between the two units (r = 0.993), so the unit of analysis changes the inference, not the estimate. The RNAscope layer shows the same error: 10,419 − 21,227 counted cells per panel from three brains per group, with a critical between-brain ICC of 0.023–0.146. Conclusions. No valid population-level test can be constructed from these data, and because the defect lies in the design it cannot be repaired retrospectively. Sleep scRNA-seq reports should state n = animals, avoid pooling before capture, and calibrate cell-level tests by permutation whenever animals are few.