tox24-challenge-project

Overview

This project was developed for the Tox24 Challenge, aimed at predicting the activity of chemical compounds against the transthyretin (TTR) protein using SMILES representations. We employed a hybrid approach that integrates Graph Convolutional Neural Networks (GCNNs) to generate electrostatic potential (ESP) surfaces and 3D Convolutional Neural Networks (3D-CNNs) to predict compound activity from voxelized ESP data.

For detailed explanations on the project methodology, analysis, and results, please refer to the full project report Predicting_TTR_Activity.

Project Structure

tox24-challenge-project/
│
├── data/
│   ├── voxel_data/
│   │   ├── train/                      # Voxelized training data
│   │   └── test/                       # Voxelized test data
│   ├── tox24_challenge_train.csv       # Raw training data
│   └── tox24_challenge_test.csv        # Raw test data
│
├── outputs/
│   ├── checkpoints/
│   │   └── best_model.pth              # Trained model checkpoint
│   ├── intermediate/
│       ├── min_max_scaler.pkl          # Scaler for activity values
│       └── processed_train_dataset.csv
│
├── LICENSE                     
├── Predicting_TTR_Activity.pdf         # Project report
├── project.ipynb                       # Jupyter notebook with code
|── requirements.txt
└── README.md               

Methodology

1. Data Processing: SMILES strings are converted to 3D molecular structures using RDKit. The ESP-DNN model generates electrostatic potential surfaces, which are voxelized into 3D grids.
2. Model Architecture: A 3D Convolutional Neural Network (3D-CNN) is used to predict compound activity based on voxelized ESP data.
3. Training: The model is trained on the voxelized data with early stopping and evaluated using Root Mean Squared Error (RMSE). The best model is saved in outputs/checkpoints/best_model.pth.

Results

• Validation RMSE: 31.8
• Baseline RMSE: 21.8

While our model did not outperform the baseline, it provides a foundation for further exploration and improvements in this domain.

Installation and Usage

Prerequisites

Before starting, ensure that you have the following installed:

• Python 3.6 or higher
• JupyterLab (or Jupyter Notebook)
• Other necessary libraries are listed in requirements.txt.

1. Clone the repository:

   git clone https://github.com/helenhenryz/tox24-challenge-project.git
   cd tox24-challenge-project

2. Create a virtual environment (optional but recommended):

   python -m venv env
   source env/bin/activate  # On Windows use `env\Scripts\activate`

3. Install required packages:

   pip install -r requirements.txt

Running the Code

1. Data Preparation:

   • Ensure that the data files (tox24_challenge_train.csv, tox24_challenge_test.csv) are in the data/ directory.

   • The voxelized data should already be provided in data/voxel_data/train and data/voxel_data/test.

2. Run the Jupyter Notebook:

   • Open project.ipynb in Jupyter Notebook or JupyterLab.

   • Run the cells sequentially to execute the data processing, model training, and evaluation steps.

3. View the Results:

   • The notebook will display plots and output showing the model's performance.

   • The trained model is saved in outputs/checkpoints/best_model.pth.

Acknowledgements

This project uses the ESP-DNN model to generate electrostatic potential surfaces. The ESP-DNN model is licensed under the Apache License 2.0, and we acknowledge the work of the authors:

• Title: Practical High-Quality Electrostatic Potential Surfaces for Drug Discovery Using a Graph-Convolutional Deep Neural Network
• Authors: Prakash Chandra Rathi, R. Frederick Ludlow, Marcel L. Verdonk
• Journal: Journal of Medicinal Chemistry
• Publication Date: August 27, 2020
• DOI: 10.1021/acs.jmedchem.9b01129

We also acknowledge the use of the Tox24 Challenge data provided by Tetko (2024).

I am beyond grateful to my brother, Kevin Henry, whose constant support and technical expertise were the backbone of this project. His patience, thoughtful guidance, and willingness to dive into every challenge with me made this a true brother-sister collaboration. From coding hurdles to late-night problem-solving, he was always there, making this journey both rewarding and memorable. I couldn’t have done it without him!

Additionally, I would like to extend my sincere thanks to Chris Lumen Ben for generously providing access to his GPU, which was essential for conducting the computational experiments necessary for this study.

License

This project is licensed under the Apache License 2.0. See the LICENSE file for more details.

Contact

For any questions or issues, please contact Helen Henry at helenhenry2025@gmail.com.