Every number on this site traces back to a published script. The exact R scripts that produced the reported tables are frozen in the repo as they were run in September 2026, with a manifest mapping each table to its script, seed, and console log: code/frozen_tcga/ on GitHub.
Which script made which table
| Results section | Script | Seed scheme |
|---|---|---|
| As-written vs fixed (RNA only) | tcga-brca-real.R (V1), tcga-brca-v3.R (V3 adaptive + standard) | 20260908 / 20260909 |
| Five-arm benchmark (5 splits) | tcga-brca-5arm.R (run once per split), aggregate-5arm.R | 20260909 + run id |
| Stability study (S1 to S3) | stability_lib.R, stability_ef.R, stability_lf.R, aggregate_stability.R | consensus 7; runs 900 + r |
| Secondary endpoint (Basal) | tcga-brca-basal-v3.R | 20260909 |
| Simulation cross-check | fusion_sim_extracted.R | set inside the script |
The full table-by-table mapping, including which console log in code/frozen_tcga/logs/ captures each run's output, is MANIFEST.md. The per-run .rds objects and raw TCGA files are not committed; the summary CSVs and logs in logs/ capture each run's reported numbers.
How to rerun it
The complete setup is documented in REPRODUCE.md. The short version: R (≥ 4.3) with tidyverse, glmnet, grpreg, pROC, Matrix, patchwork, and SKAT; download the TCGA files (~410 MB) from cBioPortal with the committed dl_rna.py and dl_clinical.py scripts; then run the scripts in order, real, v3, five-arm, stability, basal, simulation. All seeds are fixed, so a rerun on the same data and package versions reproduces the runs bit-for-bit. The stability scripts are resumable: already-computed run/variant results are skipped on relaunch.
code/main.r, code/tcga_brca_multimodal_selection.R) are kept with explicit archival headers for provenance. They are not part of the rerun pipeline; use code/frozen_tcga/. One comprehensive script, tcga-brca-validation.R, is committed for provenance with its table mapping still to be confirmed rather than guessed.