Reproducibility package for "When Prognostic Cutpoints Fail to Travel: A Prespecified Multicohort Reproducibility Analysis of Glioblastoma Biomarkers"

This reproducibility package supports the manuscript “When Prognostic Cutpoints Fail to Travel: A Prespecified Multicohort Reproducibility Analysis of Glioblastoma Biomarkers.” The study evaluates 39 literature-informed molecular biomarkers across three independent glioblastoma transcriptomic cohorts: TCGA (n=153; 121 deaths), CGGA-693 (n=104; 92 deaths), and CGGA-325 (n=73; 64 deaths). The analysis was designed to distinguish four properties that are frequently conflated in prognostic biomarker research: continuous prognostic association, statistical significance after outcome-optimized dichotomization, internal numerical stability of an optimized cutpoint, and external transportability of a locked discovery threshold. The reproducibility package includes the analysis protocol and freeze record, R analysis code, complete derived statistical outputs for continuous Cox models, restricted cubic spline analyses, median dichotomization, maximally selected cutpoints with selection-adjusted inference, bootstrap cutpoint stability, percentile-locked and z-score-locked external threshold transport, random-effects meta-analysis, focused proportional-hazards sensitivity analyses, and the R software environment. Across 117 marker-cohort analyses, naive outcome-optimized dichotomization produced 52 nominally significant results. This decreased to 22 after selection-adjusted maximally selected rank-statistic inference and to 8 after Benjamini-Hochberg false-discovery-rate correction. None of the eight robust discovery cutpoints achieved concordant nominal external replication when transferred to independent cohorts without outcome-based re-optimization. Raw patient-level TCGA and CGGA source data are not redistributed in this record. Users should obtain the original source datasets directly from their respective repositories under the applicable terms of use. This deposit contains analysis code, documentation, and derived statistical outputs required to audit the results reported in the associated manuscript.