⚙
AI-generated summary
by qwen3.7-flash, 2026-11-20
ⓘ
This endometriosis microarray analysis reveals that sample reuse, outcome-informed batch correction, and acquisition-date confounding compromise classifier validation, reducing performance from perfect to near-chance levels upon rigorous deduplication.
⚙
AI-generated deep summary
by qwen3.7-flash, 2026-11-20
· read from full text
ⓘ
This study reanalyzed publicly available transcriptomic data from the Gene Expression Omnibus to identify methodological flaws in existing endometriosis classification models. The authors found that significant portions of test samples were exact duplicates of training arrays and that batch correction algorithms inadvertently used outcome labels, leading to artificially inflated diagnostic accuracy metrics. Removing these duplicates and correcting for date-based confounding drastically reduced model performance, revealing that many previously reported high accuracies were artifacts of sample reuse and biased preprocessing rather than genuine biological signal. This paper is centrally about endometriosis — specifically critiquing the validity of gene expression classifiers used to diagnose the condition.
Abstract
Distinct Gene Expression Omnibus (GEO) accessions do not necessarily represent independent transcriptomic samples. We investigated exact array reuse, access to test diagnoses during batch correction, and cohort and acquisition-date confounding in endometriosis classifiers. We compared decompressed CEL files using SHA-256, reconstructed a matrisome elastic-net workflow, changed test labels supplied to ComBat, and conducted 3,582 constrained training-label refits. We also evaluated the effect of removing matching training copies in a five-gene Huang-inspired random forest and performed a 324-probe nested internal benchmark within GSE51981. Of 37 test CEL files, 26 (70.3%) were identical to training arrays; all had been scanned in 2004-2005. Changing a test diagnosis supplied to ComBat moved the fixed classifier score toward the supplied label in all 37 interventions. In a deduplicated, batch-constrained comparison, median nominal balanced accuracy was 0.881 when ComBat received true test outcomes and 0.549 without them; the outcome-informed arm performed better in 1,193 of 1,194 matched refit comparisons. In the separate Huang-inspired reconstruction, removing each target's matching training copy lowered pooled AUC from 1.000 to 0.323 (conditional 95% interval 0.105-0.571). Exploratory nested AUC was 0.733 for endometriosis versus other gynecological conditions, and date-held-out AUC in 2009 was 0.698. For endometriosis versus healthy controls, a date-only out-of-fold model achieved AUC 0.935, compared with 0.893 for the molecular model. Together, these findings show how exact array reuse, outcome-informed preprocessing, and acquisition-date structure can compromise the interpretation of nominally independent validation. They describe reconstructions and exploratory internal analyses, not corrected performance estimates for the published classifiers.
Full text
4,249 characters
· extracted from
oa-html
· click to expand
Abstract
Distinct Gene Expression Omnibus (GEO) accessions do not necessarily represent independent transcriptomic samples. We investigated exact array reuse, access to test diagnoses during batch correction, and cohort and acquisition-date confounding in endometriosis classifiers. We compared decompressed CEL files using SHA-256, reconstructed a matrisome elastic-net workflow, changed test labels supplied to ComBat, and conducted 3,582 constrained training-label refits. We also evaluated the effect of removing matching training copies in a five-gene Huang-inspired random forest and performed a 324-probe nested internal benchmark within GSE51981. Of 37 test CEL files, 26 (70.3%) were identical to training arrays; all had been scanned in 2004-2005. Changing a test diagnosis supplied to ComBat moved the fixed classifier score toward the supplied label in all 37 interventions. In a deduplicated, batch-constrained comparison, median nominal balanced accuracy was 0.881 when ComBat received true test outcomes and 0.549 without them; the outcome-informed arm performed better in 1,193 of 1,194 matched refit comparisons. In the separate Huang-inspired reconstruction, removing each target's matching training copy lowered pooled AUC from 1.000 to 0.323 (conditional 95% interval 0.105-0.571). Exploratory nested AUC was 0.733 for endometriosis versus other gynecological conditions, and date-held-out AUC in 2009 was 0.698. For endometriosis versus healthy controls, a date-only out-of-fold model achieved AUC 0.935, compared with 0.893 for the molecular model. Together, these findings show how exact array reuse, outcome-informed preprocessing, and acquisition-date structure can compromise the interpretation of nominally independent validation. They describe reconstructions and exploratory internal analyses, not corrected performance estimates for the published classifiers.
Competing Interest Statement
The authors have declared no competing interest.
Author Declarations
I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.
Yes
The details of the IRB/oversight body that provided approval or exemption for the research described are given below:
This study exclusively used publicly available, previously collected and de-identified human transcriptomic data from the NCBI Gene Expression Omnibus (GEO), including GSE51981, GSE6364, and other GEO datasets identified in the manuscript and Supplementary Material. The analyses involved secondary computational reanalysis of existing data. No new human participants were recruited, no new biological samples were collected, and no restricted-access patient-level data were obtained. Public repository: https://www.ncbi.nlm.nih.gov/geo/
I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.
Yes
I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).
Yes
I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.
Yes
Data Availability
All transcriptomic data analyzed in this study were obtained from publicly available repositories, including the NCBI Gene Expression Omnibus (GEO). Relevant dataset accession numbers and analysis details are provided in the manuscript and Supplementary Material. Derived results and supplementary analyses are included in the accompanying supplementary files. No new human samples or participant-level clinical data were collected for this study.
Text is read by the "Ask this paper" AI Q&A widget below.
Extraction quality varies by source — PMC NXML preserves structure
cleanly, OA-HTML may include some navigation residue, and OA-PDF can
have broken hyphenation. The publisher copy
(via DOI)
is the canonical version.