African-compound-unsupervised-clustering-kmeans-pca

Unsupervised clustering and PCA analysis of African-origin compounds using KMeans (k=4). The project identifies molecular feature patterns, visualizes clusters in reduced PCA space, and extracts representative compounds for each group.

African Compound Clustering — PCA + KMeans (Unsupervised Learning)

This project performs unsupervised clustering on a curated dataset of African-origin chemical compounds to uncover structural patterns and feature-based groupings.

Using Principal Component Analysis (PCA) for dimensionality reduction and KMeans (k=4) for clustering, the notebook identifies the most representative compound from each cluster (closest to the centroid).

---

🧠 Project Summary

• Goal: Identify natural groupings in chemical compound data using unsupervised methods.
• Dataset: African database.xlsx — containing molecular descriptors such as MolWt, LogP, HBA, HBD, TPSA, and RotBonds.
• Techniques:
  • StandardScaler normalization
  • Principal Component Analysis (PCA)
  • KMeans clustering (k=4)
  • Representative compound selection (centroid + medoid options)

---

⚙️ Workflow

1. Data Preprocessing

   • Load Excel dataset (African database.xlsx)
   • Clean missing entries and scale features using StandardScaler

2. Dimensionality Reduction

   • PCA (2 principal components) for visualization
   • Variance ratio analyzed to confirm feature importance

3. Clustering

   • KMeans with k=4
   • Visualized clusters in PCA space
   • Cluster labels added to the dataframe

4. Cluster Representation

   • Representative compound (closest to centroid)
   • Optionally, medoid compound (minimum average intra-cluster distance)
   • Results saved as CSV

---

🧪 Results Example

Cluster	Representative SMILES	Distance to Centroid
0	C=C1CC/C=C(\C)CCC2=C(C)...	22.07
1	OC1=CC=CC=C1O	9.91
2	CC1(C)CC[C@]2(C(=O)O)...	20.26
3	OCC1O[C@@H](OC2=CC=...	13.24

---

📊 Visualizations

• PCA Scatter Plot — Compounds plotted by PC1 vs PC2, colored by cluster.
• Feature Distribution Charts — Boxplots per cluster.
• Cluster Centroid Distance Chart — Identifies representative compounds.

---

📂 Repository Structure

african-compound-clustering/ ├── README.md ├── requirements.txt ├── .gitignore ├── LICENSE ├── data/ │ └── African database.xlsx ├── notebooks/ │ └── unsupervisemodels.ipynb ├── results/ │ ├── cluster_representatives.csv │ └── cluster_medoids.csv └── figures/ ├── pca_clusters.png ├── feature_boxplots.png └── cluster_summary.png

---

---

🧩 Technologies Used

• Python
• Pandas
• NumPy
• Scikit-Learn
• Matplotlib
• Seaborn

---

👨🏽‍🔬 Author

Akosu Michael Hemen
Data Scientist | Machine Learning Researcher
GitHub: mike3119

---

📜 License

MIT License