Batch-Harmonized Machine Learning Framework for Cross-Cohort RNA Biomarker Discovery in Pancreatic Adenocarcinoma
Background: Transcriptomic prognostic signatures in pancreatic ductal adenocarcinoma (PDAC) are vulnerable to endpoint misspecification, heterogeneous specimen inclusion, platform effects and information leakage. We re-audited and replaced an earlier batch-harmonized alive/dead classifier. Methods: GSE71729 was audited for assay and composition but excluded from prognosis because its GEO Series Matrix does not provide an auditable patient-level survival join. Development used 176 TCGA-PAAD primary tumors with 92 deaths. A penalized Cox model was evaluated by repeated nested 5-fold cross-validation with filtering, scaling, outcome-guided screening and tuning learned inside each training partition. Cross-platform transport was tested without joint ComBat by holding out 79 survival-labelled primary tumors from GSE85916 (57 deaths), using a per-sample percentile-rank representation. Discrimination was summarized by Harrell's concordance index (C-index) with patient-bootstrap 95% confidence intervals (CIs). The five genes highlighted in version 1 were re-examined by cohort-specific Cox models and random-effects meta-analysis. Results: Internal TCGA C-index was 0.630 (95% CI 0.567-0.694). TCGA-to-GSE85916 C-index was 0.527 (0.435-0.607); reverse-direction transport was 0.576 (0.508-0.647). A fixed five-gene Cox model trained in TCGA gave external C-index 0.569 (0.476-0.657). None of the five random-effects associations survived Benjamini-Hochberg correction across the panel; heterogeneity was high (I2 74%-93%). Conclusions: The corrected data support exploratory transcriptomic signal within TCGA but not a validated transportable prognostic model or validated five-gene biomarker panel. Cohort mixing after batch correction is not evidence of biological preservation or external validity. The revised open workflow provides a reproducible basis for larger, prespecified multi-cohort studies.