Evaluation of Existing QSAR Models and Structural Alerts and Development of New Ensemble Models for Genotoxicity...

Area of Science:

• Toxicology
• Computational Chemistry
• Risk Assessment

Background:

• Regulatory agencies require efficient methods for prioritizing chemicals based on genotoxicity.
• Existing genotoxicity datasets are large but require harmonization for effective analysis.
• Accurate prediction of genotoxic potential is crucial for risk-based chemical management.

Purpose of the Study:

• To develop and validate an in silico scheme for predicting genotoxicity.
• To create a harmonized genotoxicity dataset for model training and evaluation.
• To assess the performance of quantitative structure-activity relationship (QSAR) models and structural alerts for genotoxicity prediction.

Main Methods:

• Compiled and harmonized a large genotoxicity dataset (54,805 records) from public sources.
• Categorized substances as genotoxic, non-genotoxic, or inconclusive based on assay outcomes.
• Employed QSAR models (TEST, VEGA) and OECD Toolbox structural alerts for in silico predictions.
• Developed a Naïve Bayes consensus model integrating multiple prediction methods.

Main Results:

• The harmonized dataset included 8442 chemicals, with 2728 identified as genotoxic.
• Individual QSAR tools and structural alerts achieved balanced accuracies between 57% and 73%.
• The best consensus model demonstrated a balanced accuracy of 81.2%, with 87.24% sensitivity and 75.20% specificity.

Conclusions:

• The developed in silico scheme shows significant promise for prioritizing chemicals for genotoxicity assessment.
• This computational approach can serve as an effective first step in ranking large numbers of substances.
• The consensus model offers improved accuracy over individual prediction methods for genotoxicity evaluation.