eToxPred custom models and predictions

An interactive web-based database for exploring chemical compounds across multiple toxicological and pharmacological categories. Browse 3000+ unique chemicals with SMILES notation, toxicity predictions, and custom classification scores.

⚠️ Disclaimer

This database may contain classification errors. The data was aggregated from multiple sources and some compounds may be incorrectly categorized. If you find a misclassified compound, please open an issue with the compound name and the correct classification.

This is an experimental reference tool, not an authoritative source. Predictions are based on molecular structure pattern matching and have significant limitations. Do not base medical or safety decisions on this data alone.

🧪 Features

• 3000+ Chemical Compounds across 4 major categories (4,177 training entries, deduplicated by SMILES)
• SMILES Notation for each compound with molecular structure data
• Toxicity Predictions using EtoxPred model (tox-score & SA-score)
• Custom Classification Scores for carcinogenicity, psychoactivity, and endocrine disruption
• Interactive Search & Filtering with real-time results
• Dark-Themed UI optimized for readability
• Sortable Columns for easy data exploration
• Static SQLite Database - runs entirely client-side (via sql.js)

📂 Project Structure

chemical-database-repo/
├── site/
│   ├── index.html          # Main web interface
│   └── chemicals.db        # SQLite database with all chemical data
├── training_data/
│   ├── carcinogens.smi            # 736 known carcinogens
│   ├── endocrine_disruptors.smi   # 732 endocrine disruptors
│   ├── nootropics.smi             # 887 nootropic compounds
│   └── psychoactive_drugs.smi     # 1616 psychoactive substances
├── etoxpred/
│   ├── custom_scores.csv   # Classification predictions
│   └── results_*.csv       # EtoxPred toxicity scores per category
└── README.md

🚀 Quick Start

Simply open site/index.html in any modern web browser. No server required - the entire database runs client-side using sql.js.

# Option 1: Direct file
open site/index.html

# Option 2: Local server (optional)
cd site && python3 -m http.server 8000
# Visit http://localhost:8000

📊 Database Schema

The database contains a single chemicals table:

Column	Type	Description
id	INTEGER	Primary key
name	TEXT	Chemical name (e.g., "Caffeine")
smiles	TEXT	SMILES notation (e.g., "CN1C=NC2=C1C(=O)N(C(=O)N2C)C")
category	TEXT	Comma-separated categories (nootropic, carcinogen, etc.)
tox_score	REAL	EtoxPred toxicity prediction (0-1, higher = more toxic)
sa_score	REAL	Synthetic accessibility score (0-1, normalized, higher = harder to synthesize)
carc_score	REAL	Carcinogenicity prediction (0-1)
psych_score	REAL	Psychoactivity prediction (0-1)
endo_score	REAL	Endocrine disruption prediction (0-1)

📖 How It Works

Web Interface (site/index.html)

A single-page application that runs entirely in the browser:

• sql.js: WebAssembly SQLite engine loads the database
• No backend required: Everything runs client-side
• Real-time search: Filters as you type
• Sortable columns: Click headers to sort
• Category filtering: Quick filtering by compound class
• Score visualization: Color-coded toxicity indicators

Features:

• Dark theme optimized for chemical data
• Responsive design (mobile-friendly)
• Inline help guide with scoring explanations
• Displays SMILES, categories, and all prediction scores

🧬 Training Data

Categories

Category	Count	Description
Nootropics	887	Cognitive enhancers, smart drugs, memory supplements
Psychoactive Drugs	1616	CNS-active compounds (stimulants, depressants, hallucinogens)
Carcinogens	736	Known or suspected cancer-causing agents
Endocrine Disruptors	732	Compounds that interfere with hormone systems

SMILES Format

All training data uses simplified molecular-input line-entry system (SMILES):

C1=CC=C(C=C1)O                          # Phenol
CN1C=NC2=C1C(=O)N(C(=O)N2C)C            # Caffeine
CC(C)CC1=CC=C(C=C1)C(C)C(=O)O           # Ibuprofen

• Tab-separated values: SMILES<tab>Name
• Canonical SMILES from PubChem ensures consistency
• Each compound appears once per category file
• Compounds may exist in multiple category files

🔬 Toxicity Scoring

EtoxPred Scores

The database includes predictions from the EtoxPred model:

• Tox-score (0-1): Overall toxicity prediction

  • 0.0-0.3: Low toxicity (green)
  • 0.3-0.7: Moderate toxicity (yellow)
  • 0.7-1.0: High toxicity (red)

• SA-score (1-10): Synthetic accessibility

  • 1-3: Easy to synthesize
  • 4-7: Moderate difficulty
  • 8-10: Very difficult to synthesize

Custom Classification Scores

Machine learning models trained on the labeled training data:

• carc_score: Carcinogenicity prediction (0-1)
• psych_score: Psychoactivity prediction (0-1)
• endo_score: Endocrine disruption prediction (0-1)

Training Methodology

Each classifier is trained as a binary one-vs-rest model:

1. Features: Morgan fingerprints (ECFP4, radius=2, 1024 bits) via RDKit
2. Positives: Compounds from the target category's .smi file
3. Negatives: ~4,900 FDA-approved drugs only (not compounds from other categories, to avoid label leakage)
4. Balancing: Negatives downsampled to 2:1 ratio vs positives
5. Split: 80/20 train/test with scaffold split (Murcko decomposition) to prevent structural leakage
6. Model: ExtraTreesClassifier (400 trees, min_samples_split=10, min_samples_leaf=3)
7. Output: Probability score (0-1) from predict_proba

Nootropics are not used in training (too heterogeneous) — they're only scored by the trained models.

Test Set Performance (Scaffold Split)

Model	ROC-AUC	PR-AUC	Accuracy
Carcinogen	0.92	0.91	76%
Psychoactive	0.86	0.72	84%
Endocrine Disruptor	0.82	0.69	82%

Note: Recall on positives is moderate (42-53%) due to conservative predictions — the models favor precision over recall.

Known Limitations

• Prodrugs invisible: Structure-based fingerprints can't predict metabolic activation (e.g., 1,4-butanediol → GHB scores only 0.36 despite being psychoactive via metabolism)
• Small training sets: ~700-1100 positives per category limits generalization
• FDA negatives bias: Negatives are biased toward drug-like molecules

🛠️ Dependencies

Runtime (Web Interface)

• None! Pure HTML/CSS/JavaScript
• sql.js loaded from CDN

Retraining Models

The training script is in scripts/. Uses uv for dependency management:

cd scripts
uv sync
uv run python train_classifiers.py

Or with pip:

pip install rdkit scikit-learn pandas numpy joblib
python scripts/train_classifiers.py

📝 Data Sources

• Chemical Names: Curated from scientific literature, databases, and regulatory lists
• SMILES Notation: PubChem REST API
• Toxicity Predictions: EtoxPred (pre-trained model)
• Classifications: Labeled training data from multiple sources

🤝 Contributing

Contributions welcome! Ways to help:

1. Add more chemicals: Submit .smi files or chemical name lists
2. Improve predictions: Train better classification models
3. Enhance UI: Add visualizations, filters, or export features
4. Documentation: Add compound descriptions or sources

📄 License

• Code: GNU General Public License v3.0 - see LICENSE
• Data: Chemical structures from PubChem (public domain)
• Models: EtoxPred and custom classifiers are provided as-is for research use

⚠️ Disclaimer

This database is for research and educational purposes only. Toxicity predictions are computational estimates and should not be used for:

• Medical decision-making
• Regulatory compliance
• Consumer product safety assessments
• Legal or forensic purposes

Always consult authoritative sources (EPA, FDA, ECHA) and perform proper experimental validation for any toxicological or pharmacological claims.

🔗 Links

• SMILES Notation
• PubChem
• EtoxPred
• sql.js

---

Built with 🧪 chemistry and ⚡ code