BRCA-PathwayML

Interpretable pathway-based machine learning for breast cancer prognosis: a four-cohort benchmark against PAM50-ROR.

Authors: Suhaan Thayyil, Eshaan Nidee

Overview

This repository contains the code, analysis pipelines, and results for a four-cohort study comparing a pathway-based survival machine learning model (Gradient Boosted Survival Analysis) against official genefu PAM50-ROR for breast cancer prognosis.

The pre-registered primary endpoint (TNBC meta-analyzed delta C-index >= +0.03, p < 0.05) was not met. The observed result was delta = +0.0144, 95% CI [-0.0472, +0.0760], p = 0.6466. All secondary analyses are exploratory.

Key Results

• Harmonized database: 4,532 samples across four cohorts (TCGA-BRCA, GSE96058/SCAN-B, METABRIC, GSE20685).
• Analysis sets: 4,003 patients in patient-characteristics summaries; 4,002 metric-evaluable patients in the official PAM50-ROR discrimination analysis.
• Headline model: Gradient Boosted Survival (internal TCGA CV C-index = 0.642, delta vs Cox = +0.042).
• TNBC primary endpoint: NOT MET (meta delta = +0.0144, p = 0.6466).
• Overall discrimination: comparable to PAM50-ROR across external cohorts.

Submitted Manuscript Crosswalk

The submitted BMC Artificial Intelligence manuscript is mapped to repository files in docs/SUBMITTED_MANUSCRIPT_REPOSITORY_CROSSWALK.md.

Reviewer-facing traceability helpers:

• scripts/check_submission_consistency.py verifies the central submitted values against repository CSV/JSON outputs.
• results/Table_S2_internal_cv_manuscript_crosswalk.csv maps internal cross-validation values between the manuscript summary and exact reproducible output.
• results/Table_1_denominator_crosswalk.csv documents patient-characteristics versus metric-evaluable denominators.

Repository Structure

src/              Python modules (models, survival analysis, data loading, features)
src/ml/           DeepSurv implementation
R/                Official genefu PAM50-ROR scoring (pam50.R)
scripts/          Analysis pipeline scripts
notebooks/        Jupyter notebooks (01-14: original analyses)
results/          All results tables and QC reports
figures/          Generated figures
paper/            Reviewer-facing LaTeX manuscript source and DOI-backed bibliography
data/             Raw, processed, and clinical data
models/           Trained model artifacts (gitignored)
tests/            Pytest test suite
docs/             Pre-registration endpoint, story lock, cohort notes

Manuscript LaTeX

The repository includes a reviewer-facing LaTeX manuscript source at paper/main.tex, a DOI-backed BibTeX bibliography at paper/main.bib, a compiled PDF at paper/main.pdf, and a Word export at paper/main.docx. The LaTeX source mirrors the submitted competitive-not-superior framing and is intended for traceability, not as a claim that the manuscript source is the journal's copyedited version.

Build commands:

cd paper && tectonic --keep-logs main.tex
cd .. && python3 scripts/build_manuscript_docx.py

Data Sources

• TCGA-BRCA: GDC Data Portal (https://portal.gdc.cancer.gov)
• GSE96058/SCAN-B: NCBI GEO (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE96058)
• METABRIC: cBioPortal (https://www.cbioportal.org/study/summary?id=brca_metabric)
• GSE20685: NCBI GEO (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE20685)

Raw data are not redistributed in this repository.

Pre-registration

The primary endpoint was locked before any external validation was run. See docs/PRIMARY_ENDPOINT.md (committed at 7783f4f0bfaaa6bdc611c78d33ccda621c6b243d, before any results commits).

Reproduction

1. Install Python dependencies: pip install -r requirements.txt
2. Install R with the genefu package for official PAM50-ROR scoring.
3. Pipeline scripts are in scripts/. Key steps:
   • scripts/harmonize_cohorts.py -- build unified cohort database
   • scripts/score_pathways.py -- compute pathway features
   • scripts/train_ml_zoo.py -- train survival models
   • scripts/external_validation.py -- run external validation
   • scripts/compute_clinical_baselines.py -- compute clinical genomic baselines
   • Phase 3 exploratory scripts: within_subtype_analysis.py, calibration_analysis.py, dca_analysis.py, stability_analysis.py
   • Sensitivity analyses: external_only_tnbc_sensitivity.py, feature_ablation.py, pathway_scoring_sensitivity.py

License

MIT License. See LICENSE.