panelbc — Focus-Anchored Breast-Cancer Prognostic Panel

A minimal, cheap, multiply-validated breast-cancer prognostic gene panel and the code that produced it. ICMR grant IIRPSG-2024-01-02447, Objective 3.

Starting from a mechanistically-defined protein complex (EEF1A2, IQGAP1, IQGAP2, FRG1), an unbiased genome-wide, leakage-free search yields a 7- or 9-gene panel that matches or beats the 50-gene PAM50 signature at a fraction of the genes, validated across four independent cohorts on three continents (TCGA, METABRIC, SCAN-B, GSE20685; ~6,800 patients).

Scientific note. The predictive signal is carried by the discovered genes (FLT3, CLIC6, SUSD3, ZIC2, P4HA2); the four complex genes are near-random as predictors on their own (C≈0.54) and are retained as the biological anchor and discovery scaffold, not as predictive features. See docs/.

Install

pip install -e .            # core scoring (numpy/pandas/scipy/scikit-learn/lifelines)
pip install -e ".[full]"    # + reproduction dependencies (xgboost, shap, sksurv, ...)
pip install -e ".[deep]"    # + torch for the deep-learning stage

Quick start — score your own cohort

import pandas as pd, panelbc
expr = pd.read_csv("my_expression.csv", index_col=0)   # samples x genes (any scale)
scores = panelbc.risk_score(expr, panel="9gene")       # higher = higher risk
groups = panelbc.risk_group(expr, panel="9gene")       # Low / Intermediate / High

or from the command line:

python scripts/predict.py --expr my_expression.csv --panel 9gene --out scores.csv

Reproduce the headline benchmark

With the cohort checkpoints in data/ (see docs/DATA.md):

python scripts/reproduce.py
# model           TCGA  METABRIC  SCAN-B    MEAN
# 7-gene         0.678     0.608   0.676   0.654
# 9-gene         0.710     0.617   0.688   0.671
# focus-only     0.569     0.532   0.516   0.539

Repository layout

panelbc/            installable package: locked panel defs (panels.json) + scoring
scripts/            predict.py (score a CSV), reproduce.py (headline benchmark)
analysis/           the 13 ordered stage scripts for the full study (01..13) + helpers
tests/              pytest suite (scoring correctness, direction, missing-gene handling)
docs/               DATA.md (how to obtain/rebuild cohort data), METHODS.md
requirements.txt    pip dependencies (core + extended tiers, with exact pins for reproduction)
pyproject.toml      installable package metadata (pip install -e .)

Install

pip install -r requirements.txt     # dependencies
pip install -e .                    # the panelbc package + panelbc-predict CLI
pytest tests/ -q                    # 5 tests, should all pass

The full analysis pipeline (analysis/)

Thirteen ordered stages documenting the complete study: dataset build (Aim I) -> transcriptional cross-talk (Aim II) -> genome-wide gene discovery -> leakage-free forward selection -> parsimony/cost frontier -> panel head-to-head -> deep learning (Phase II) -> 4th-cohort validation -> statistical rigor (calibration/DCA/clinical-independence) -> tumor-vs-normal (Aim III) -> (near-)exhaustive gene-set search -> KEGG pathway novelty -> immune-infiltration control. Stages 11-13 are the extended analyses:

• 11_exhaustive_search.py — numba-JIT ridge-Cox engine; exhaustive enumeration of all 3-5 gene panels + wide-beam 6-10, anchored and free. Shows the best free panel contains zero focus genes at every size and plateaus at ~0.70.
• 12_pathway_novelty.py — KEGG annotation of every panel gene; focus and discovered genes occupy disjoint pathway space (novel combination, not a known module). Requires NCBI + KEGG network access.
• 13_immune_control.py — confirms the panel is not an immune-infiltration surrogate (risk-immune correlation ~0; immune adds nothing to C-index). Turnkey: runs from the six checkpoints.

These are the faithful record of what was computed in a persistent analysis kernel; stages 01-10 share in-memory state and are meant to be read/run in order (see analysis/helpers_reference.py for the shared functions), while 11-13 are more self-contained. The installable panelbc package + scripts/ are the turnkey, tested surface.

Method summary

• Expression z-scored within each cohort before pooling — neutralises TPM/RSEM/microarray scale differences.
• Prognosis = penalised Cox; discrimination = Harrell's C-index.
• Leakage-free: survival-based gene selection done inside CV folds only.
• External validation = leave-one-cohort-out + a fully held-out 4th cohort.

Data availability

All cohorts are public (cBioPortal, GEO). GDC/UCSC-Xena were unreachable in the original environment, so TCGA RSEM was obtained from cBioPortal's PanCancer Atlas. See docs/DATA.md.

Citation

See CITATION.cff.

License

MIT — see LICENSE.