Quantitative structure-activity relationship models for genotoxicity prediction based on combination evaluation...

Area of Science:

• Toxicology
• Computational Chemistry
• Bioinformatics

Background:

• Mutagenicity poses significant human health risks.
• Conventional toxicity prediction methods are limited in scope and predictive power.
• Existing models often lack integrated experimental data, hindering accuracy.

Purpose of the Study:

• To develop robust mutagenicity prediction models by integrating diverse experimental data.
• To enhance the predictive ability and application scope of computational toxicology tools.
• To establish an effective early warning system for compound mutagenicity.

Main Methods:

• Data from GENE-TOX, CPDB, and CCIS were partitioned using the weight-of-evidence method.
• Inclusion of in vivo, in vitro, prokaryotic, and eukaryotic experimental data following ICH guidelines.
• Development of nine sub-models and three fusion models using Random Forest (RF), Support Vector Machine (SVM), and Back Propagation (BP) algorithms.

Main Results:

• Fusion models demonstrated excellent predictive performance across all algorithms.
• The RF, SVM, and BP fusion models achieved prediction accuracies of 83.4%, 80.5%, and 79.0%, respectively.
• Area Under the Curve (AUC) values reached 0.853 (RF), 0.897 (SVM), and 0.865 (BP).

Conclusions:

• Integrated QSAR models effectively predict compound mutagenicity.
• The developed fusion models offer a reliable early warning system for potential mutagens.
• This approach enhances the assessment of chemical safety and reduces reliance on single experimental types.