Fair re-estimation of published prostate cancer prognostic gene signatures reveals moderate but reproducible cross-cohort signal — reproducible archive

Reproducible archive for the manuscript "Fair re-estimation of published prostate cancer prognostic gene signatures reveals moderate but reproducible cross-cohort signal". Contains: (1) analysis scripts 01-14 (R/Python: Cox-ridge re-estimation, cross-cohort transfer, 500x bootstrap confidence intervals, machine-learning benchmarks, figure generation); (2) processed result tables and cohort matrices under data/processed/, plus the curated PCaDB signature inventory (RDS derivatives; third-party full texts are not redistributed); (3) English and Chinese manuscripts plus Supplement S1 (independent same-method reproduction, including the 30x5 per-signature C-index matrix and heatmap). Source data are public cohorts (TCGA-PRAD via GDC; GSE70768, GSE70769, GSE46602, GSE16560 via GEO) and are not redistributed here. v1.0.2 (2026-10-28): Table 5's per-signature external transfer C-index column recomputed from external_validation_cindex.tsv (median of four external cohorts) and rewritten in both manuscripts; added scripts/15_reporting_audit.R, data/processed/signature_reported_performance_curated.tsv, and the graphical abstract figure. v1.0.3 (2026-10-28): Table 4 numbers aligned to script output (ml_vs_signature_cindex.tsv); Li 2022 patient count corrected to 1,558; Table 1 citation order fixed per Vancouver; Figure 4 cited in text; performance range restated as per-signature four-cohort transfer median 0.480–0.714 (IQR 0.613–0.669, 25/30 ≥ 0.6); scripts referenced as 01–15; README data/raw path corrected to data/fulltext; added derived signature_reporting_audit.tsv. v1.0.4 (2026-10-28): Published Elastic-net external C-index values aligned to fully reproducible glmnet 4.1-8 re-run output (0.731 / 0.705 / 0.705; ≤0.002 drift from prior 4.1-8 estimates; Ridge coefficients near-saturated, byte-identical reproduction) and pinned in README. All three manuscripts, library documents, and README unified to these values; per-signature external-transfer range restated as 0.583–0.731. Archive rebuilt from the re-run pipeline (ml_vs_signature_cindex.tsv, rf_vs_signature_cindex.tsv, signature_reporting_audit.tsv). Author list corrected per author instruction; corresponding author left blank. v1.0.5 (2026-10-28): Manuscript and README Zenodo version labels corrected from v1.0.3 to v1.0.4 (concept DOI 10.5281/zenodo.23642008 unchanged, always resolves to latest); aligned with the EN/ZH/README library documents. No data, script, or result changes versus v1.0.4.