Using the super-learner to predict the chemical acute toxicity on rats

Area of Science:

• Computational toxicology
• cheminformatics
• machine learning

Background:

• Traditional median lethal dose (LD50) testing faces ethical and cost challenges.
• Existing quantitative structure-activity relationship models often lack precision and interpretability.

Purpose of the Study:

• To develop a robust predictive framework for acute toxicity using ensemble learning.
• To improve the accuracy and interpretability of LD50 predictions.

Main Methods:

• Developed a Super-learner ensemble model using 16 meta-models, 4 molecular descriptors, and machine learning algorithms.
• Incorporated data filtering and applicability domain methods to enhance model reliability.
• Utilized a dataset of 9843 compounds for LD50 prediction.

Main Results:

• The Super-learner model achieved R² values of 0.61 (cross-validation) and 0.64 (test set), outperforming individual models.
• An R² of 0.76 was achieved within the applicability domain, demonstrating high prediction accuracy.
• The model showed improved performance and a broader applicability domain compared to previous studies.

Conclusions:

• The Super-learner framework provides an effective, cost-efficient, and accurate method for predicting chemical toxicity.
• An online tool was launched to facilitate rapid LD50 predictions and chemical safety assessments.
• This work offers valuable technical support for chemical risk assessment processes.