Answers are grounded in this paper and cite their sources.
Ranking cell types by a gene-set score is routine, and the order is often read as biology, though the procedure assumes the set carries information beyond a cell's RNA content. We built an audit framework–housekeeping substitution, an expression-matched permutation null, residualisation against technical covariates and a positive-control arm and applied it to lactylation scoring in 16 public single-cell datasets (113,419 cells) and to three non-metabolic signatures (hypoxia, stemness, senescence), two failing the same controls. A 40-gene housekeeping panel matched or beat the 100-114-gene set in 8 of 15, expression-matched random sets in 10 of 15, and none survived Benjamini-Hochberg correction (minimum q = 0.106). Within-cell coupling to detected-gene count was positive in all 15 (median +0.574). The cross-cell-type collinearity statistic proved untestable: +0.991 in simulation with no signal. Its residualised variant is demoted to a span-gated diagnostic, readable only where detection rate spans 0.15, and non-monotonic. Under two established implementations coupling fell by a median of 60%, never vanishing, and the top-ranked type agreed in only 7 of 13 datasets. Coupling also operates within a cell type: the score is higher in the deeper half in 114 of 116 comparisons, and depth-matching leaves the score-to-depth effect ratio flat at 0.55. Hypoxia and stemness signatures of 28 genes fail the same control (median SSR 0.436 and 0.440), so this is not specific to metabolism; a senescence panel with genuine cell-type restriction passes (1.298). The lactylation set carries 1.67 times the gene-level variance of expression-matched sets; controlling for depth and variance together lowers median SSR from 0.825 to 0.363. Variance matching leaves the coupling intact (97.2% positive) whereas an analytic Pearson residual removes it (56.0%), locating the fault in the summed-score construction, not the normalisation. A lineage-marker positive control passed, so the limitation concerns the estimator, not lactylation biology.
No takes yet. Share an insight, caveat, or question.
Jiadong Fu (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: