PRI-Audit: A Protocol for Evaluating Incremental Value of Transcriptomic Representations with Matched Expression Baselines and Grouped Cross-Validation
Year: 2026
Type: article
Abstract: 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.
Source: Genes
Authors: Kaiwen Deng, Liwei Liu
Institution: Dalian Jiaotong University
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7235753909
