Aromatase Inhibitor — Machine Learning Regression Analysis (MSR & R² Evaluation)

This repository contains a supervised machine learning regression analysis of Aromatase inhibitors using multiple models.
The notebook used for this analysis is included as aromataseB.ipynb.

It compares various regression algorithms based on MSR (Mean Squared Residuals ≈ MSE) and R² (Coefficient of Determination) to identify which performs best for predicting aromatase-related activity.

📘 Project Summary

• Objective: Predict and analyze Aromatase inhibitor outcomes using regression-based ML models.
• Data Source: Processed molecular data (loaded within the notebook).
• Evaluation Metrics:
  • MSR / MSE — measures average squared prediction error.
  • R² Score — measures how well the model explains the variance of the data.

⚙️ Models Evaluated

The notebook trained and evaluated the following models:

• Support Vector Regressor (SVR)
• ExtraTreeRegressor (multiple variants)
• ExtraTreesRegressor
• GradientBoostingRegressor
• DecisionTreeRegressor
• XGBRegressor
• KNeighborsRegressor
• MLPRegressor
• LinearSVR
• SGDRegressor

📊 Results Summary

The top-performing models (by Test R²) from the notebook are shown below:

Model	Train MSE (≈ MSR)	Train R²	Test MSE (≈ MSR)	Test R²
SVR	0.782579	0.551344	1.061522	0.413114
ExtraTreeRegressor3	0.782579	0.551344	1.061522	0.413114
ExtraTreesRegressor2	0.782579	0.551344	1.061522	0.413114
ExtraTreeRegressor1	0.782579	0.551344	1.061522	0.413114
GradientBoostingRegressor	0.841112	0.517786	1.221255	0.412216

Best Model (Test R²): SVR (0.413)
Lowest MSR (Test): ExtraTreeRegressor variants (~1.06)

📈 Visualizations

Two plots summarize the performance of all regressors:

• figures/test_r2_by_model.png → Test R² comparison
• figures/test_mse_by_model.png → Test MSE (≈ MSR) comparison

These visuals show how well each model generalizes to unseen data.

🧠 Insights

• Models with higher Test R² and lower Test MSE (MSR) perform best.
• Tree-based models like ExtraTreeRegressor and GradientBoostingRegressor show solid performance but may overfit.
• Simpler models (e.g., LinearSVR) perform less effectively on this dataset.
• The gap between Train and Test results indicates the complexity of the data.

🧪 Reproduction

If you want to reproduce or modify this analysis:

1. Open the provided notebook:

   notebooks/aromataseB.ipynb
   

aromatase-ml-regression/ ├── README.md ├── requirements.txt ├── .gitignore ├── notebooks/ │ └── aromataseB.ipynb ├── results/ │ ├── model_metrics.csv │ └── model_metrics_table.md ├── figures/ │ ├── test_r2_by_model.png │ └── test_mse_by_model.png └── src/ └── parse_notebook_metrics.py

🧩 Technologies Used

Python

Pandas

NumPy

Scikit-Learn

XGBoost

Matplotlib

🧬 Purpose of This Study

Aromatase is an enzyme critical in estrogen biosynthesis. Understanding inhibitors through data-driven modeling can aid in developing better therapeutic candidates for hormone-dependent cancers. This project explores how machine learning can be applied to analyze and predict inhibitory activities.

External Test Results (Aromatase pIC50 — PaDEL PubChem 881-bit)

Model: Random Forest
External test size: ~614 compounds

Predicted vs Observed (External Test)

Calibration / Reliability (10 bins)

Headline Metrics (External Test)

Model	Test R²	Test MSE
Random Forest	0.505	0.901

Pipeline: RDKit standardization → PaDEL fingerprints (PubChem 881-bit) → stratified split with strict external test → kNN-distance applicability domain → calibration (reliability) → error-by-scaffold inspection.

Reproducibility: fixed seeds, saved metrics JSON, versioned splits.
Repo link: github.com/mike3119/aromatase-ml-regression

👨🏽‍🔬 Author

Akosu Michael Hemen Machine Learning Researcher | Computational Chemist