Batch correction reshapes dual-transition gene lists in confounded GTEx-TCGA prostate RNA-seq

Abstract

Studying prostate field cancerization requires comparing truly healthy tissue against adjacent normal and tumor samples. Combining GTEx healthy prostate tissue with TCGA cancer data provides that three-group comparison; however, the design is fully confounded: every healthy sample comes from one database and every cancer sample comes from another. This study tested how much a dual-transition field-effect gene list changes when batch correction is applied to 820 public prostate RNA-seq samples (263 GTEx True Normal, 52 TCGA Adjacent Normal, 505 TCGA Tumor) from recount3. Before correction, samples separated by database source rather than tissue type. After ComBat-seq, although that separation was reduced, the three tissue groups did not form a clear ordered pattern. The dual-transition gene list fell from 643 genes without correction to 10 after correction, with only two genes shared between the two lists (RPL39P and RPL6P; Jaccard index 0.003). After correction, thousands of genes had small adjusted p-values in the field contrast, reflecting residual confounded signal rather than a clean biological result. Three of the ten corrected genes are pancreatic digestive enzymes (PRSS3, AMY2A, CELA3A), consistent with known contamination signals in public GTEx data. Sensitivity analyses across different fold-change thresholds, an alternative batch method, and bootstrap resampling all supported the same pattern. Within-TCGA tumor versus adjacent normal analysis provided a cleaner same-database comparison for tumor-associated expression changes. These results show that dual-transition field-effect gene lists from GTEx-TCGA prostate data are highly unstable under batch correction when batch and biology are fully confounded. Researchers using this design should treat uncorrected dual-transition lists with caution and report batch correction sensitivity as a standard check. Files: preprint PDF and supplementary figures/tables (zip). Analysis code and full result tables: https://github.com/kcncell/GTEX_TCGA_prostate_RNAseq. Public counts: recount3 (TCGA-PRAD, GTEx PROSTATE).