Queimo/tox21_tabicl_clamp_classifier

Tox21 TabICLv2-CLAMP Classifier

Submission API for the Tox21 leaderboard. This Gradio Server / ZeroGPU Space uses frozen 768-dimensional CLAMP molecular embeddings and one fitted TabICLv2 classifier for each of the twelve Tox21 endpoints.

The model is fitted on the public ml-jku/tox21 development data. Features are standardized with training statistics and clipped to [-6, 6]. TabICLv2 uses eight estimators and the tabicl-classifier-v2-20260212.ckpt checkpoint.

API

• —GET /metadata reports the leaderboard capabilities.
• —GET /healthz reports service health.
• —POST /predict accepts {"smiles": ["CCO", "c1ccccc1"]}.

The prediction route requests a ZeroGPU allocation for up to 600 seconds and accepts batches of up to 1,000 SMILES. Invalid SMILES receive a probability of 0.5 for every endpoint.

Reproduce training

pip install -r requirements.txt
python train.py

train.py downloads only the public Tox21 development set, recreates the features and preprocessing state, fits all endpoint contexts, and writes them under checkpoints/. This repository is for research use and is not a clinical toxicity assessment tool.