Multi-Omics Breast Cancer Subtype Classification with TCGA-BRCA

This project compares single-omics and multi-omics machine learning models for breast cancer subtype classification using TCGA-BRCA data from Kaggle.

The main question is simple:

Does integrating multiple omics layers improve subtype classification compared with using each omics block alone?

Results

Late-fusion multi-omics modeling performed best on the holdout test set.

Method	Feature set	Model	Accuracy	Macro F1	Weighted F1
Multi-omics	Late fusion	Averaged probabilities	0.892	0.815	0.885
Single-omics	Protein	XGBoost	0.863	0.800	0.853
Multi-omics	Early fusion	Logistic regression	0.804	0.677	0.806
Single-omics	Copy number	Random forest	0.784	0.581	0.774
Single-omics	RNA expression	Logistic regression	0.765	0.559	0.762
Single-omics	Mutation	Random forest	0.667	0.341	0.632

The best multi-omics model improved macro F1 by +0.015 over the best single-omics model.

Dataset

Dataset: BRCA Multi-Omics TCGA on Kaggle

The dataset contains 705 breast tumor samples and four omics blocks:

Omics block	Prefix	Features
Copy number	cn_	860
Mutation	mu_	249
RNA expression	rs_	604
Protein / phosphoprotein	pp_	223

The Kaggle files used here do not include a PAM50 subtype label. The supervised target in this project is a derived clinical_subtype based on ER, PR, and HER2 receptor status. Ambiguous or missing receptor-status rows are excluded from supervised modeling.

Modeling cohort after filtering:

Class	Samples
HR+/HER2-	335
Triple-negative	90
HR+/HER2+	57
HER2-enriched-like	25

Methods

The workflow uses a fixed random seed and a stratified train/validation/test split:

Split	Samples
Train	303
Validation	102
Test	102

Preprocessing:

• Copy number, RNA expression, and protein features: median imputation and standard scaling
• Mutation features: most-frequent imputation and preserved as binary 0/1
• Feature selection: SelectKBest with ANOVA F-statistics inside model pipelines
• Main model-selection metric: macro F1

Single-omics baselines:

• Logistic regression
• Random forest
• Linear SVM
• XGBoost when available

Multi-omics models:

• Early fusion: concatenate all omics features before modeling
• Late fusion: train one model per omics block and average predicted class probabilities

Explainability

The best model is a late-fusion ensemble, so the main explanation uses grouped permutation importance by omics block. The score is the mean macro-F1 drop after shuffling one omics block on the holdout test set.

Omics block	Mean macro-F1 drop
Protein	0.300
RNA expression	0.265
Copy number	0.125
Mutation	0.061

Top feature-level signals included mu_TP53, mu_PIK3CA, mu_GATA3, pp_ER.alpha, pp_HER2, and pp_HER2.pY1248.

Repository Structure

tcga-brca-multiomics-classification/
├── app/
├── data/
│   ├── raw/
│   ├── processed/
│   └── external/
├── models/
│   ├── checkpoints/
│   └── metrics/
├── notebooks/
├── reports/
│   ├── figures/
│   └── tables/
├── src/
│   ├── data/
│   ├── features/
│   ├── models/
│   └── visualization/
├── README.md
├── requirements.txt
└── .gitignore

Reproduce The Project

Create and activate a virtual environment:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt

Download the Kaggle files:

python -m src.data.download_data

Run the pipeline:

python -m src.visualization.run_eda
python -m src.features.run_preprocessing
python -m src.models.train_baselines
python -m src.models.train_fusion
python -m src.models.explain
python -m src.visualization.final_figures

Open the notebooks:

python -m jupyter lab

Key Outputs

Final figures:

• reports/figures/final_summary_panel.png
• reports/figures/final_model_comparison_macro_f1.png
• reports/figures/final_best_model_confusion_matrix.png
• reports/figures/final_omics_importance.png
• reports/figures/final_subtype_distribution.png

Final tables:

• reports/tables/final_model_comparison.csv
• reports/tables/final_key_results.csv

Model artifacts:

• models/checkpoints/single_omics_*.joblib
• models/checkpoints/multiomics_*.joblib

Limitations

• The Kaggle dataset used here does not include PAM50 subtype labels, so this project predicts a receptor-status clinical subtype proxy.
• The HER2-enriched-like class is small, with 25 total samples before splitting.
• Results are based on one stratified holdout split. A larger study should use repeated cross-validation and external validation.
• Feature importance values are useful for model inspection, but they should not be interpreted as causal biological effects.

Next Steps

• Add an external PAM50 label source if sample identifiers can be matched safely.
• Add repeated cross-validation for more stable model comparison.
• Add a Streamlit demo for interactive prediction and model inspection.
• Compare late fusion with calibrated weighted fusion.