Leakage-Safe Cross-Cohort Transcriptomic Survival Analysis in PDAC

This repository contains a corrected reanalysis of the project originally reported as Batch-Harmonized Machine Learning Framework for Cross-Cohort RNA Biomarker Discovery in Pancreatic Adenocarcinoma.

Version 2 correction notice (2026-09-12): The original analysis does not establish a prognostic classifier, external validation, or five validated biomarkers. GSE71729 is an Agilent microarray containing mixed primary, metastatic, normal, and cell-line specimens; the original outcome also reduced survival to alive/dead status and ignored follow-up time and censoring. The v1 scripts and artifacts are preserved for provenance but must not be used as evidence of predictive performance.

Read the complete forensic audit and interpretation before using the results.

Corrected scientific question

Can a transcriptomic survival model developed in one pancreatic ductal adenocarcinoma cohort retain discrimination in a completely held-out cohort measured on another platform, without letting the held-out samples influence normalization, feature selection, tuning, or model fitting?

The current answer is not convincingly. The corrected analysis finds modest internal TCGA discrimination but no demonstrated cross-platform transport.

Headline results

Analysis	Patients / events	C-index (95% patient-bootstrap CI)	Interpretation
TCGA repeated nested 5-fold CV	176 / 92	0.630 (0.567-0.694)	Modest internal discrimination
Train TCGA, hold out GSE85916	79 / 57	0.527 (0.435-0.607)	No demonstrated transport
Train GSE85916, hold out TCGA	176 / 92	0.576 (0.508-0.647)	Weak reverse-direction transport
Original fixed five-gene Cox, train TCGA, test GSE85916	79 / 57	0.569 (0.476-0.657)	Confidence interval includes chance

The original five genes—LAMC2, DKK1, ITGB6, GPRC5A, and MAL2—remain exploratory hypotheses. All have adverse univariate associations in TCGA, but this is the original discovery cohort. In the two-cohort random-effects analysis, none survives BH correction across the five and heterogeneity is high. LAMC2 is only nominally significant: HR 1.58 per cohort SD, 95% CI 1.04-2.42, unadjusted p=0.034, BH FDR=0.170.

Cohorts and eligibility

Dataset	Platform	Prognostic use in v2	Patients / events
TCGA-PAAD PanCancer Atlas	RNA-seq; cBioPortal RSEM snapshot	Development and internal validation	176 / 92
GSE85916	GPL13667 Affymetrix Human Genome U219 array	Retrospective whole-cohort holdout	79 / 57
GSE71729	GPL20769 Agilent microarray	Provenance/sample-composition audit only	Not eligible

GSE71729 contains 145 primary tumors, 61 metastases, 17 cell lines, 46 normal pancreas samples, and 88 distant-site normal samples. It is excluded from prognostic modelling because the repository has no auditable patient-level survival-time/event join for its primary tumors.

The fast TCGA reanalysis uses the cBioPortal PanCancer Atlas RSEM snapshot. A manuscript-grade final analysis should pin and reprocess a GDC STAR-count release and preserve raw-file provenance.

Leakage controls in v2

• One primary tumor per patient.
• Positive overall-survival time and explicit death/censoring indicator.
• Surv(time, event) rather than alive/dead classification.
• Repeated nested patient-level cross-validation for model development.
• MAD filtering, imputation, centering/scaling, outcome-guided screening, hyperparameter tuning, and penalized Cox fitting learned only from training partitions.
• Entire cohorts held out from model development.
• No joint ComBat, cohort-wise test scaling, or fitting to the held-out cohort distribution.
• Sample-wise percentile-rank transformation for the cross-platform analysis.
• Cohort-specific Cox estimates and random-effects meta-analysis for candidate biomarkers.
• Machine-readable sample flows, predictions, selected features, input URLs, byte sizes, MD5 checksums, package metadata, and confidence intervals.

Reproduce the fast v2 analysis

Run from the repository root with R. Base R plus the recommended survival package are required.

Rscript scripts/v2/01_audit_gse71729.R
Rscript scripts/v2/02_prepare_tcga.R
Rscript scripts/v2/03_nested_cox_tcga.R
Rscript scripts/v2/06_prepare_gse85916.R
Rscript scripts/v2/04_loco_cox.R
Rscript scripts/v2/05_meta_analyze_candidates.R
Rscript scripts/v2/07_validate_fixed_panel.R

Downloads are cached under data/v2/cache/ and are excluded from version control. See scripts/v2/README.md for the cohort schema, outputs, and extension instructions.

Main v2 outputs

• results/v2/gse71729_sample_audit.csv
• results/v2/tcga_sample_flow.csv
• results/v2/gse85916_sample_flow.csv
• results/v2/tcga_nested_cv_performance.csv
• results/v2/tcga_oof_predictions.csv
• results/v2/tcga_selected_gene_stability.csv
• results/v2/loco_performance.csv
• results/v2/loco_predictions.csv
• results/v2/fixed_five_external_performance.csv
• results/v2/candidate_cohort_cox.csv
• results/v2/candidate_random_effects_meta.csv
• results/v2/tcga_input_manifest.csv
• results/v2/gse85916_input_manifest.csv
• results/v2/run_manifest.txt

What the original analysis got wrong

The detailed evidence is in V2_AUDIT.md. In brief:

• GSE71729 was incorrectly described as platform-matched RNA-seq.
• All 357 heterogeneous GSE71729 samples were combined without specimen filtering.
• Unlabelled predictions were described as external validation.
• Alive/dead status replaced time-to-event analysis.
• ComBat was fitted jointly to training and purported external samples.
• Full-data preprocessing and top-variance selection preceded Random Forest OOB scoring.
• “Training Accuracy” was actually OOB accuracy.
• “Class Balance” values were class-specific recalls, and the raw-model recalls were reported incorrectly.
• The Shiny app's hard-coded 92.6% accuracy is unsupported; the serialized corrected Random Forest has approximately 64.0% OOB accuracy.

Interpretation and next phase

Version 2 is a retrospective correction, not a clinical model. GSE85916 was selected during the repair and should be treated as exploratory external testing rather than a prospectively locked confirmation cohort.

The next phase is a multi-cohort transportability study that evaluates locked PDAC transcriptomic signatures across eligible survival cohorts and compares clinical-only, expression-only, and clinical-plus-expression models. It should report discrimination, calibration, time-dependent performance, biological signal preservation, and uncertainty under whole-cohort holdout validation.

Data sources

• GSE71729
• GSE85916
• GDC TCGA-PAAD
• cBioPortal PanCancer Atlas snapshot
• Original bioRxiv preprint

Repository history

The original scripts, figures, serialized models, and results remain in their existing directories as an auditable record. They are v1 artifacts and are superseded by scripts/v2/, results/v2/, and this correction notice. The original repository state is preserved at tag v1-original-preprint.

License

MIT License. See LICENSE.