Predicting hepatotoxicity using ToxCast in vitro bioactivity and chemical structure

Area of Science:

• Toxicology and Cheminformatics
• Environmental Health Sciences
• Computational Biology

Background:

• The U.S. Tox21 and EPA ToxCast programs utilize high-throughput screening to assess environmental chemical bioactivity.
• Predictive models of chemical toxicity are crucial for public health and environmental safety.
• Understanding the link between in vitro bioactivity and in vivo toxicological outcomes is essential.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting in vivo hepatotoxicity of environmental chemicals.
• To compare the predictive performance of models based on chemical structure, bioactivity, and hybrid descriptors.
• To identify key bioactivity and structural features associated with rodent hepatotoxicity.

Main Methods:

• A dataset of 677 chemicals with known rat liver histopathology outcomes was used.
• Chemicals were described by 711 ToxCast bioactivity descriptors and 4,376 chemical structure descriptors.
• Six machine learning algorithms (LDA, NB, SVM, CART, KNN, ENSMB) were employed to build classifiers.
• 10-fold cross-validation and feature selection were used for performance evaluation.

Main Results:

• Hybrid classifiers, integrating bioactivity and structure, achieved the highest balanced accuracy for predicting hypertrophy (0.84), injury (0.80), and proliferative lesions (0.80).
• Chemical structure classifiers showed higher sensitivity, while bioactivity classifiers demonstrated higher specificity.
• CART, ENSMB, and SVM algorithms performed best, with nuclear receptor activation and mitochondrial functions frequently identified as predictive features.

Conclusions:

• Combining in vitro bioactivity and chemical structure data significantly enhances the prediction of rodent hepatotoxicity.
• High-throughput screening data from ToxCast and ToxRefDB are valuable resources for linking chemical properties to adverse outcomes.
• The study highlights the utility of machine learning and hybrid approaches for characterizing environmental chemical hazards.