An Ensemble of Trees, Self-Normalizing Networks and a Foundation Model for Tox21 Toxicity Prediction
Year: 2026
Type: preprint
Abstract: A seven-model ensemble (random forest, ExtraTrees, two LightGBM models, two self-normalizing neural networks and the CheMeleon foundation model) for the 12-task Tox21 toxicity benchmark, submitted to the JKU Linz Tox21 Reproducible Leaderboard. Trained only on the ml-jku/tox21 data. Cross-validation mean ROC-AUC 0.854 and official validation-set mean ROC-AUC 0.785. Code and training script: https://huggingface.co/spaces/ruebenkadosh/tox21-ensemble
Author: Kadosh, Rueben
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225392720