Multi‑Omic Cancer Type Classification (TCGA PanCancer)

A reproducible pipeline and notebook for classifying tumour types from multi‑omics data (RNA‑seq, DNA methylation, miRNA‑seq) using the TCGA Pan‑Cancer Atlas dataset.

Highlights

• Goal: Predict CANCER_TYPE across 18 cancers from ~549k features per patient.
• Modality fusion: RNA‑seq, DNA methylation (beta values), and miRNA‑seq.
• Dimensionality reduction: PCA → 300 components after StandardScaler.
• Best model: L1‑regularised Logistic Regression (saga) on PCA features.
• Test performance: 97.45% accuracy, 0.97 weighted F1 (stratified 80/20 split).
• Key insight: DNA methylation signals were consistently the strongest contributors to discriminating cancer types.
• Challenge: Extreme class imbalance—rare classes (support < 10) remain hard to model.

The accompanying notebook (notebooks/Pancan_project.ipynb) contains exploratory analysis, model training, and comparisons (LogReg, RandomForest, XGBoost). A report is available in reports/.

Repository Structure

pancan-multiomic-classifier/
├─ src/
│  ├─ train.py            # Train/evaluate pipeline from CSV/Parquet paths
│  └─ utils.py            # I/O helpers and metrics
├─ notebooks/
│  └─ Pancan_project.ipynb
├─ reports/
│  └─ Pancan_Project_Report.docx
├─ data/
│  ├─ raw/                # Place downloaded TCGA files here
│  └─ processed/          # Intermediate files (PCA, splits)
├─ models/                # Trained models and artefacts
├─ figures/               # Exported plots
├─ requirements.txt
├─ environment.yml
├─ CITATION.cff
├─ LICENSE
├─ .gitignore
└─ README.md

Data

This project uses multi‑omics data from the TCGA Pan‑Cancer Atlas (commonly distributed on Kaggle/TCGA mirrors). You will need three aligned tables per sample:

• RNA‑seq (e.g., TPM)
• DNA methylation (beta values)
• miRNA‑seq (RPM)

Note: Large datasets are not included. See src/utils.py for expected input formats and column conventions. Minimal sample schema is documented inline.

Reproduce

1) Create environment

python -m venv .venv
# Windows: .venv\Scripts\activate
# macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt

Or with conda/mamba:

mamba env create -f environment.yml
mamba activate pancan-ml

2) Prepare data

Place modality files under data/raw/ and ensure sample IDs align across modalities.

3) Train baseline

python src/train.py   --rna data/raw/rna.csv   --meth data/raw/methylation.csv   --mirna data/raw/mirna.csv   --label_column CANCER_TYPE   --output_dir models   --n_components 300

This will:

• scale → PCA (n=300)
• fit LogisticRegression (saga, penalty='l1')
• print accuracy/F1, save model.joblib, label_encoder.joblib, and metrics.json
• write a confusion matrix to figures/confusion_matrix.png

4) Evaluate on your split

If you provide a --test_size (default 0.2) and --random_state, the script will generate a stratified split and report accuracy and weighted F1.

Results Summary

• LogReg (opt): 97.45% accuracy, 0.97 weighted F1. Rare classes improved vs trees/boosting.
• RandomForest / XGBoost: ~95% accuracy; struggled with rare classes (sometimes F1=0.0).
• PCA impact: Reducing ~549k features to 300 components delivered >100× speedups with minimal accuracy loss.
• Biology: DNA methylation carried strong discriminative signal between cancers.

See reports/Pancan_Project_Report.docx for the full narrative and notebooks/ for the analysis steps.

Next Steps (Roadmap)

• Variance‑targeted PCA (n_components=0.95) to probe rare‑class signals.
• Imbalance methods: SMOTE/ADASYN or class‑balanced ensembles.
• Modality‑specific models: Methylation‑only baselines; early/late fusion strategies.
• Interpretability: Map high‑loading features → gene sets/pathways.

License

Released under the MIT License (see LICENSE).

Citation

If you use this work, please cite via CITATION.cff or:

Meacci, D. (2025). Multi‑Omic Cancer Type Classification (TCGA PanCancer). GitHub repository.

(Sync test)