Answers are grounded in this paper and cite their sources.
Background/Objectives: Transcriptomic representations should be tested against matched expression references on held-out biological units. Methods: PRI-Audit evaluates low-rank pseudobulk, spatial-graph, and spot-topic implementations using grouped validation, paired AUROC differences, uncertainty estimates, and applicable source diagnostics. Results: Four of five principal or boundary comparisons had reference AUROC 1.000 and were non-evaluable for the specified gain. In mixed-disease GSE206391 (12 patients), PCA gained 0.002 (conditional 95% interval, −0.024 to +0.038); a 16-HVG model-refit sensitivity had median gain 0.068 (2.5th to 97.5th percentile range, −0.014 to +0.160). One post-review artificial gene rearrangement passed the numerical and shuffle gates, but its random-rearrangement sham had the same shuffle p value (0.010), and its conditional interval crossed zero. A label-conditioned synthetic-channel benchmark passed all three conditional criteria in 100/100 replicates at mixing 0.50 (mean ΔAUROC +0.143). A separate expression-matrix simulation with smooth artificial fields showed no positive incremental AUROC. None of these artificial constructions was biologically calibrated. Conclusions: No candidate was shown to satisfy the complete evidence rule. The evidence is inconclusive about potentially meaningful gains; the simulations and artificial signals do not estimate real biological power. Upstream fold-local processing could not be verified.
No takes yet. Share an insight, caveat, or question.
Deng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: