Bradleyyb37/Quantitative_toxicity_prediction

🧪 ActFound-based Toxicity Predictor

This Space provides EC10 toxicity predictions from SMILES strings using a fine-tuned ActFound meta-learning model.

🔧 How it works

• —Base model: ActFoundRegressor (meta-learning QSAR)
• —Input: SMILES → 2048-d Morgan fingerprint (via RDKit)
• —Inner loop: Meta-learning adaptation using a fixed support set (EC10_support.json)
• —Output: Predicted EC10 (e.g., -log10(EC) in BEAS-2b cells)

📂 Files

• —app.py – Gradio UI (single + batch prediction)
• —inference.py – Core inference logic (predict, predict_batch)
• —requirements.txt – Dependency versions
• —EC10_support.json – Support compounds and EC10 labels
• —actfound_finetuned_ec10.pt – Fine-tuned model weights (state_dict)
• —Actfound_demo/ – Model definitions (system_actfound.py, etc.)

⚠ Note: system_actfound.py must be edited so that all .cuda() calls are replaced with device-agnostic code (e.g., .to(device)).

✅ Usage (Web UI)

1. 1.Open the Space.
2. 2.Single Prediction tab:
3. 3.Enter a SMILES string.
4. 4.Click Predict.
5. 5.The EC10 value will be displayed on the right.
6. 6.Batch Prediction tab:
7. 7.Fill or paste a list of SMILES in the table.
8. 8.Click Predict Batch.
9. 9.A result table with predicted EC10 values will appear.

🔌 Programmatic API Example (Python)

If your Space is public (e.g., https://huggingface.co/spaces/yourname/actfound-ec10), you can call the Gradio API endpoint like:

import requests

SPACE_URL = "https://yourname-actfound-ec10.hf.space"

payload = {
    "data": ["CC1=CC=CC=C1"]  # single SMILES input for the main function
}

res = requests.post(f"{SPACE_URL}/api/predict", json=payload)
print(res.json())