TRPV1 Classification Models: Systematic Evaluation

Supporting information for the manuscript:

Title: TRPV1 classification models: systematic evaluation across algorithms and molecular representations
Authors: Mohamed Diwan M. AbdulHameed and Anders Wallqvist
Manuscript Status: Under review

---

Overview

This repository provides a comprehensive benchmark of machine learning models for TRPV1 modulation classification using IC50 and EC50 endpoints.

Dec 2025 update: The codebase has been fully updated to eliminate duplication, use repository-relative paths, and provide a unified analysis pipeline.

---

Key Features

• 7 ML algorithms: Logistic Regression, Random Forest, XGBoost, LightGBM, SVM, KNN, Gaussian Naive Bayes
• 4 molecular representations: Morgan fingerprints, RDKit fingerprints, MACCS keys, Mordred descriptors
• Validation: repeated stratified cross-validation with scaffold-based external test sets
• Statistical analysis: repeated-measures ANOVA and Tukey HSD for model comparison
• Interpretability: SHAP-based model interpretation
• Endpoint support: IC50, EC50, or both via command-line arguments

---

Quick Start

Installation

git clone https://github.com/BHSAI/TRPV1.git
cd TRPV1

conda env create -f environment.yml
conda activate trpv1_ml_benchmark

python check_dependencies.py

See INSTALLATION.md for detailed setup instructions and troubleshooting.

---

Run Complete Pipeline

# Preprocessing for both endpoints
python code/preprocessing/run_preprocessing.py --endpoints IC50 EC50

# Analysis for both endpoints
python code/analysis/run_analysis.py --endpoints IC50 EC50

Outputs are written to endpoint-specific subfolders under:

• results/
• figures/

---

Run Individual Steps (examples)

# Preprocessing
python code/preprocessing/01_standardize_smiles.py --endpoint IC50
python code/preprocessing/04_scaffold_split.py --endpoint IC50

# Analysis
python code/analysis/01_cross_validation_fingerprints.py --endpoint IC50
python code/analysis/07_generate_master_table.py --endpoint IC50
python code/analysis/06_visualize_dashboard.py --endpoint IC50

For full script documentation, see:

• code/preprocessing/README_UNIFIED_PIPELINE.md
• code/analysis/README.md

---

Repository Structure (High Level)

TRPV1/
├── code/            # Preprocessing, analysis, utilities
├── data/            # Raw and processed datasets
├── results/         # Tables and evaluation outputs
├── figures/         # Plots for manuscript and SI
├── models/          # Final models

---

Project Scope

This benchmark evaluates TRPV1 classification performance across algorithms, molecular representations, and validation regimes.

Machine Learning Algorithms (7)

• Linear: Logistic Regression
• Tree-based: Random Forest, XGBoost, LightGBM
• Instance-based: K-Nearest Neighbors
• Probabilistic: Gaussian Naive Bayes
• Kernel-based: Support Vector Machine (RBF)

Molecular Representations (4)

• Morgan fingerprints: radius = 2, 2048 bits
• RDKit fingerprints: 2048 bits
• MACCS keys: 166-bit keys
• Mordred descriptors: 1600+ descriptors

Validation Strategy

• Internal: 5×5 repeated stratified cross-validation
• External: scaffold-based test split (20%)
• Statistics: repeated-measures ANOVA + Tukey HSD
• Applicability domain: SDC-based analysis

Model Interpretation

• SHAP analysis: feature importance and dependence plots
• Fingerprint interpretation: bit-level / substructure analysis

---

Performance Metrics

All models are evaluated using:

• ROC-AUC: Area under ROC curve
• PR-AUC: Area under precision-recall curve
• MCC: Matthews correlation coefficient
• G-Mean: Geometric mean of sensitivity and specificity
• F1 Score: Harmonic mean of precision and recall
• Accuracy: Overall classification accuracy
• Sensitivity (Recall): True positive rate
• Specificity: True negative rate
• Precision: Positive predictive value

---

Citation

If you use this code or data, please cite:

@article{abdulhameed2025trpv1,
  title={TRPV1 classification models: systematic evaluation across algorithms and molecular representations},
  author={AbdulHameed, Mohamed Diwan M. and Wallqvist, Anders},
  journal={[Journal Name]},
  year={2025},
  note={Under review}
}

---

Contributing

This repository supports a manuscript under review. For questions or bugs, please open a GitHub issue and include relevant details (command run, endpoint, and error messages).

---

Acknowledgments

• RDKit community for molecular processing tools
• Scikit-learn contributors for ML infrastructure
• The TRPV1 research community for bioactivity data

---

Last Updated: December 2025
Manuscript Status: Under review