Pyrosage - Environmental Property Predictors

This repository contains datasets, models, and reproducible notebooks for predicting key physicochemical, environmental, and toxicity-related properties of chemical compounds. These predictors are intended to support virtual screening and safer chemical design, especially for applications like flame retardancy and environmental safety.

✅ Covered Properties

The following properties are included and modeled using open datasets:

• LogKow / LogP
• Bioaccumulation Factor (BCF)
• Soil/Water Partition Coefficient (Koc)
• Biodegradability
• Water Solubility
• Henry's Law Constant (KH)
• Aqueous Hydroxyl Rate (kAOH)
• Hydrolytic Stability
• pKa (acidic and basic)
• KOA (Octanol-Air Partition Coefficient)
• TBP (related to biodegradation)
• Fish Bioaccumulation / Toxicity
• Molecular Weight (computed)

📂 Repository Structure

/data              # CSV datasets used for training
/models            # Saved model files (e.g., .pkl or .onnx)
/notebooks         # Jupyter notebooks to train or evaluate models
README.md          # You're here

📘 Usage

1. Clone the repository:

   git clone https://github.com/ntua-unit-of-control-and-informatics/pyrosage

2. Install dependencies:

   pip install -r requirements.txt

3. Explore and run any notebook in /notebooks to reproduce or fine-tune the models.

⚖️ License

This project uses publicly available data (e.g. from VEGA/OPERA repositories) and models trained for research purposes. Please cite sources if you reuse the datasets.

Acknowledgments

Some of the datasets and models in this repository are reused or adapted from the work of Wang et al. (2023), titled "Applicability Domains Based on Molecular Graph Contrastive Learning Enable Graph Attention Network Models to Accurately Predict 15 Environmental End Points". This study introduced GAT-based QSAR models with applicability domain refinements for environmental property prediction.

📄 Reference: https://doi.org/10.1021/acs.est.3c03860