Leakage Remediation

This repository contains leakage-controlled re-evaluation code and result tables for the H2C/TDA biomarker project.

H2C is a 37-gene panel identified from a Topological Data Analysis (TDA) workflow on TCGA-BRCA RNA-seq data. The analyses here test whether the original classification and survival findings remain valid after removing common sources of validation leakage.

Repository Structure

classification/
  code/
    leakage_free_classification.py
    leakage_free_classification_with_combat.py
  results/
    leakage_free/
    leakage_free_combat/

survival/
  code/
    leakage_free_survival.py
  data/
    brca_survival_cohort.tsv
  results/
    leakage_free/

inputs/
  raw_preprocessing_inputs.tsv

scripts/
  download_raw_preprocessing_inputs.sh

manuscript_revision_results/
  README.md
  generate_manuscript_revision_artifacts.py
  figures/
    fig5_classification_performance.{png,pdf}
    fig7_km_oos_survival.{png,pdf}
  tables/
    table3_classification_fold_internal_cv.*
    survival_*.*

Classification Re-Evaluation

The strict classification audit is in:

classification/results/leakage_free_combat/

It evaluates tumor/normal classification after:

• fitting ComBat within each training fold;
• applying held-out fold transformation using only training-fold ComBat estimates;
• recomputing TDA and Euclidean feature rankings inside each training fold;
• evaluating each held-out fold once.

Main result:

Feature set	Classifier	AUC mean	F1 mean	Accuracy mean
Euclidean_top200_train	LogisticRegression	0.998991	0.992740	0.986895
Euclidean_top200_train	RandomForest	0.998848	0.993233	0.987714
TDA_top200_train	RandomForest	0.998138	0.990543	0.982664
H2C_train_rule_top37	RandomForest	0.981446	0.973847	0.951393
H2C_train_rule_top37	LogisticRegression	0.702863	0.947700	0.900704

Interpretation: H2C remains strong with RandomForest under strict leakage control, but it is classifier-sensitive and is not uniquely superior to Euclidean Top200.

Survival Re-Evaluation

The survival audit is in:

survival/results/leakage_free/

It compares H2C, Euclidean, and Random gene sets under the same scoring and modeling rules. The primary rows are the out-of-sample multivariate results.

Main result:

Gene set	Score method	HR per SD	Cox p	Log-rank p	C-index
H2C_37	pca	0.973121	0.684033	0.921137	0.727071
H2C_37	cox_beta	1.226483	0.002658	0.000000456	0.766799
Euclidean_top37	pca	1.062423	0.394671	0.012743	0.725779
Euclidean_top37	cox_beta	1.127145	0.089836	0.001228	0.727601
Random_37	pca	0.964446	0.603271	0.222353	0.735156
Random_37	cox_beta	1.048946	0.548799	0.001177	0.733135

Interpretation: H2C retains adjusted prognostic signal under supervised out-of-sample Cox-beta scoring. PCA-based survival signal is not supported.

Manuscript Revision Artifacts

Camera-ready table and figure artifacts are provided in: manuscript_revision_results/.

This folder contains manuscript-ready replacements for Fig. 5 and Fig. 7, table-ready CSV/Markdown summaries, and a small generation script. It contains only derived artifacts; raw expression matrices, patient-level score tables, model checkpoints, and controlled-access data are not included.

Reproduction Notes

The committed result tables are self-contained. Re-running the scripts requires the original project repositories and preprocessing artifacts to be available as sibling directories:

Data-preprocessing/
FindVar/
FindVar-Survival-Analysis/
Leakage-Remediation/

Raw TCGA preprocessing files and large model/data artifacts are not committed to this repository.

Required External Artifacts

Most processed project artifacts are tracked in the existing project repositories:

Artifact	Expected source
TAE model checkpoint	Data-preprocessing/TAE/models/tae_dim32_cosine.pth
32-dimensional latent matrix	Data-preprocessing/TAE/results/woutSMOTE/latent_32d_cosine.csv
Cleaned TCGA TPM matrix	Data-preprocessing/data_preprocessing/cleaned_tcga_tpm_for_TAE.csv
BRCA gene statistics	Data-preprocessing/data_analysis/output_brca_patients/00_BRCA_All_23368_Genes_Statistics.csv
TDA gene importance ranking	FindVar/phase3_gene_traceback/results/gene_importance_full.csv
Original classification summary	FindVar/phase4_biological_interpretation/results/classification_results.csv
Survival helper module	FindVar-Survival-Analysis/phase7_survival_analysis/survival_analysis.py

The following raw preprocessing inputs are not tracked in the existing GitHub repositories and must be supplied separately to fully rerun the strict fold-internal ComBat classification script:

Data-preprocessing/GSE62944_06_01_15_TCGA_24_548_Clinical_Variables_9264_Samples.txt.gz
Data-preprocessing/GSE62944_RAW/GSM1536837_06_01_15_TCGA_24.tumor_Rsubread_TPM.txt.gz
Data-preprocessing/GSE62944_RAW/GSM1697009_06_01_15_TCGA_24.normal_Rsubread_TPM.txt.gz

These are public NCBI GEO/FTP supplementary input files referenced by the original preprocessing workflow. They were not developed in this project.

The exact URLs, file sizes, and SHA-256 checksums are listed in:

inputs/raw_preprocessing_inputs.tsv

To download and verify them into a workspace containing Data-preprocessing/, run:

bash scripts/download_raw_preprocessing_inputs.sh /path/to/workspace

Data Availability

The survival re-evaluation uses TCGA-BRCA survival and expression data from UCSC Xena. The strict classification ComBat run requires the original TCGA preprocessing input files used by the H2C/TDA project; those raw files are intentionally excluded from GitHub.