🧬 Breast Cancer Subtype Prediction via Multi-Omics Integration

📌 Project Overview

This project presents a machine learning pipeline for predicting Breast Cancer Subtypes (Luminal A, Luminal B, Basal, and HER2) using multi-omics data integration:

• Gene Expression (RNA-Seq)
• DNA Methylation

The main contribution of this project is a Multi-Stage Hybrid Feature Selection pipeline designed to address the curse of dimensionality in bioinformatics datasets.

From tens of thousands of biological features, the pipeline extracts a robust subset of 50 biomarkers, enabling accurate classification using a Random Forest classifier.

---

⚙️ Methodology & Pipeline

The workflow is implemented in 6 sequential Jupyter Notebooks, ensuring modularity and reproducibility:

1. Data Harmonization

   • Synchronize patients across clinical, RNA, and methylation datasets

2. Preprocessing

   • Train/Test split (to prevent data leakage)
   • Missing value imputation
   • Log2 transformation
   • MinMax normalization

3. Feature Selection – Phase 1 (Variance Threshold)

   • Remove quasi-constant features

4. Feature Selection – Phase 2 (ANOVA Filter)

   • Select top 500 statistically significant features using ANOVA (F-test)

5. Feature Selection – Phase 3 (ReliefF Wrapper)

   • Select top 150 context-aware features based on nearest neighbors

6. Feature Selection – Phase 4 (Hybrid + Modeling)

   • Apply mRMR (Minimum Redundancy Maximum Relevance)
   • Select final 50 features
   • Apply SMOTE for class balancing
   • Train and evaluate models: Random Forest, SVM, KNN

---

📂 Project Structure

Run notebooks in the following order:

Step	Notebook	Description
01	Reading.ipynb	Load datasets and synchronize patients
02	preprocessing.ipynb	Data split, imputation, normalization
03	Feature_Selection-1.ipynb	Variance Threshold
04	Feature Selection_2.ipynb	ANOVA feature selection
05	Feature Selection_3.ipynb	ReliefF selection
06	Feature Selection_4.ipynb	mRMR + SMOTE + modeling

---

📊 Results

Models were evaluated on a held-out test set.

Model	Accuracy	Features	Pipeline
🏆 Random Forest	84.55%	50	Variance → ANOVA → ReliefF → mRMR
SVM	82.73%	50	Same pipeline
KNN	68.18%	50	Same pipeline

ℹ️ The mRMR step significantly improved performance by removing redundant and highly correlated genes.

---

🛠️ Installation

Make sure you have Python installed, then run:

pip install numpy pandas matplotlib seaborn scikit-learn

If you used a specific mRMR library, also install:

pip install pymrmr

---

🚀 How to Run

1. Clone the repository

git clone https://github.com/YourUsername/Breast-Cancer-Subtype-Prediction.git
cd Breast-Cancer-Subtype-Prediction

2. Prepare data

Place the following datasets in the project root (or update paths inside Reading.ipynb):

• RNA-Seq data
• DNA Methylation data
• Clinical labels

3. Run notebooks

Execute notebooks in order from 01 → 06

📁 Intermediate processed files will be automatically saved in:

outputs/