Structure-based QSAR Models to Predict Repeat Dose Toxicity Points of Departure.

Area of Science:

• Environmental toxicology
• Computational toxicology
• Chemical risk assessment

Background:

• Human health risk assessment is hindered by a lack of experimental in vivo toxicity data for most chemicals.
• Quantitative structure-activity relationship (QSAR) models offer a method for predicting chemical hazards using structural information.
• Accurate risk assessment necessitates quantitative point-of-departure (POD) values for low-dose extrapolation.

Purpose of the Study:

• To develop and validate QSAR models for predicting POD values (POD_QSAR) in repeat dose toxicity studies.
• To generate both point estimates and confidence intervals for POD_QSAR predictions.
• To assess the utility of POD_QSAR models in informing screening-level human health risk assessments.

Main Methods:

• Compiled an in vivo toxicity dataset for 3592 chemicals from the U.S. EPA's Toxicity Value database (ToxValDB).
• Developed two sets of QSAR models: one for point-estimate POD_QSAR and another for 95% confidence intervals.
• Utilized a random forest model with chemical structure and physicochemical descriptors; employed bootstrap resampling for confidence intervals.

Main Results:

• The best random forest QSAR model achieved an RMSE of 0.71 log₁₀-mg/kg/day and R² of 0.53 on an external test set.
• The second set of models generated 95% confidence intervals for POD_QSAR predictions, accounting for uncertainty.
• Enrichment analysis indicated that 80% of the most potent chemicals were identified within the top 20% of POD_QSAR predictions.

Conclusions:

• The developed POD_QSAR models provide reliable predictions for repeat dose toxicity, aiding in data-poor chemical assessments.
• These models can help prioritize chemicals for further testing and inform preliminary human health risk assessments.
• The approach demonstrates the potential of QSAR for enhancing the efficiency and scope of chemical risk evaluations.