Quantitative prediction of repeat dose toxicity values using GenRA.

Area of Science:

• Toxicology
• Computational Chemistry
• cheminformatics

Background:

• Computational toxicology methods, including read-across, are increasingly used to predict chemical safety data.
• The generalized read-across (GenRA) tool previously integrated in vitro bioactivity and chemical descriptors to predict in vivo toxicity hazards.
• Accurate prediction of point of departure (POD) values is crucial for chemical risk assessment.

Purpose of the Study:

• To modify the GenRA tool for quantitative prediction of POD values from the US EPA's Toxicity Reference Database (ToxRefDB).
• To evaluate the performance of the modified GenRA tool in predicting various toxicity endpoints.
• To assess the impact of domain specificity on the accuracy of GenRA predictions.

Main Methods:

• The GenRA tool was adapted to predict oral Lowest Observed Adverse Effect Levels (LOAELs) for 1,014 chemicals.
• Chemicals were analyzed for systemic, developmental, reproductive, and cholinesterase inhibition effects.
• Predictions were evaluated globally and locally within clusters of structurally similar chemicals using Morgan fingerprints and Jaccard similarity.

Main Results:

• Global GenRA predictions yielded R-squared values ranging from 0.14 to 0.43 for different toxicity endpoints.
• Evaluating GenRA within local domains of structurally similar chemicals significantly improved prediction accuracy.
• Average R-squared values for local predictions reached 0.73 for systemic, 0.66 for developmental, 0.60 for reproductive, and 0.79 for cholinesterase inhibition LOAELs.

Conclusions:

• The modified GenRA tool demonstrates potential for predicting chemical PODs.
• Prediction accuracy is highly dependent on the chemical domain; localized predictions outperform global ones.
• Identifying appropriate local domains is critical for the effective application of GenRA in chemical safety evaluations.