🧬 Mice Protein Expression Analysis

Unsupervised Learning on Brain Protein Data

---

📌 Project Overview

This project analyzes the Mice Protein Expression Dataset, which contains measurements of 77 protein expression levels from the brains of 72 mice.

The goal is to explore how:

• 🧬 Genotype (Control vs Trisomy)
• 🧠 Behavioral Training
• 💊 Treatment (Memantine vs Saline)

affect protein expression patterns using unsupervised machine learning techniques.

This project focuses on:

• Data Cleaning & Preprocessing
• Dimensionality Reduction (PCA, SVD, t-SNE, UMAP)
• Clustering (K-Means, Agglomerative, DBSCAN)
• Cluster Evaluation using Silhouette Score

---

📂 Dataset Information

• Total Mice: 72
• Total Protein Features: 77
• Categorical Variables:
  • Genotype → Control / Trisomy
  • Behavior → CS / SC
  • Treatment → Memantine / Saline

Each mouse belongs to one of 8 experimental groups: [Genotype]-[Behavior]-[Treatment] Example: c-CS-s

Dataset file location: dataset/Data_Cortex_Nuclear.csv

---

⚙️ Technologies Used

• Python
• Pandas
• NumPy
• Matplotlib
• Seaborn
• Scikit-learn
• UMAP

---

🔄 Project Workflow

---

1️⃣ Data Preprocessing

✔ Missing Value Handling using KNNImputer (k=5)
✔ One-Hot Encoding of categorical variables
✔ Feature Scaling using RobustScaler

---

2️⃣ Dimensionality Reduction

We applied multiple dimensionality reduction techniques:

• PCA (Principal Component Analysis)
• Truncated SVD
• t-SNE
• UMAP

UMAP preserved structure more effectively for clustering.

---

3️⃣ Clustering Algorithms Applied

Clustering was performed on:

• PCA-reduced data
• SVD-reduced data
• UMAP-reduced data

Algorithms used:

• 🔹 K-Means
• 🔹 Agglomerative Clustering
• 🔹 DBSCAN

---

📊 Clustering Performance Comparison

All models were evaluated using Silhouette Score.

Higher score → Better cluster separation.

Model	Silhouette Score
KMeans (PCA)	~0.38
KMeans (SVD)	~0.38
Agglomerative (PCA)	~0.29
Agglomerative (SVD)	~0.31
DBSCAN (PCA)	~0.57
DBSCAN (SVD)	~0.58
KMeans (UMAP)	0.6218
DBSCAN (UMAP)	~0.61

---

🏆 Best Clustering Model

Best Clustering Model: KMeans_UMAP

Best Silhouette Score: 0.6218

✅ Conclusion:

UMAP combined with K-Means produced the best cluster separation among all tested models.

This indicates that:

• UMAP preserved meaningful structure
• K-Means effectively separated the clusters in reduced space

---

📈 Clustering Performance Visualization

The following bar chart compares Silhouette Scores across models:

📁 Project Structure

Mice-Protein-Expression/
│
├── dataset/
│   └── Data_Cortex_Nuclear.csv
│
├── images/
│   └── clustering_performance.png
│
├── mice_protein_analysis.ipynb
├── README.md
└── requirements.txt

▶️ How to Run

1️⃣ Clone Repository

git clone https://github.com/muhammedshihab1001/mice-protein-expression.git
cd mice-protein-expression

2️⃣ Run Notebook

jupyter notebook

3️⃣ Open

mice_protein_analysis.ipynb

📈 Key Insights

• Proper scaling significantly improves clustering quality.

• PCA and SVD provide moderate separation.

• DBSCAN performs better than K-Means on linear reductions.

• UMAP provides the best structure preservation.

• KMeans + UMAP achieved the highest Silhouette Score (0.6218).

🚀 Future Improvements

• Try Spectral Clustering

• Hyperparameter tuning for DBSCAN

• Compare clusters with true experimental labels

• Apply supervised classification models

• Perform feature importance analysis

---

👨‍💻 Author

Muhammed Shihab P

Machine Learning Enthusiast