An In Silico Test Battery for Rapid Evaluation of Genotoxic and Carcinogenic Potential of Chemicals
INTRODUCTION
According to the FDA Guidance for Industry, impurities identified below the ICH qualification thresholds may be evaluated for genotoxicity and carcinogenicity based on structural activity relationship (SAR) assessments using computational software. This work is an extension of our previous study focusing on computational assessment of genotoxic impurities in drug products. Our new approach relies on a battery of probabilistic QSAR models supplemented by a knowledge-based expert system that identifies structural fragments, potentially responsible for hazardous activity. The analysis was based on experimental data obtained from FDA, and involved 21 endpoints corresponding to different mechanisms of toxic action: mutagenicity, clastogenicity, carcinogenicity, etc. Probabilistic models for most endpoints were derived using GALAS (Global, Adjusted Locally According to Similarity) modeling methodology developed in our group. The updated list of alerting groups contained 70 distinct substructures. The expert system was highly sensitive, recognizing >90% of potent carcinogens, as classified by the FDA. Sensitivity of probabilistic models ranged from about 60% to more than 90%, while maintaining high (>80%) specificity of predictions for the majority of considered assays. These results show that the described computational platform ensures sufficient prediction accuracy for rapid genotoxicity/ carcinogenicity profiling of various chemicals.
According to the FDA Guidance for Industry, impurities identified below the ICH qualification thresholds may be evaluated for genotoxicity and carcinogenicity based on structural activity relationship (SAR) assessments using computational software. This work is an extension of our previous study focusing on computational assessment of genotoxic impurities in drug products. Our new approach relies on a battery of probabilistic QSAR models supplemented by a knowledge-based expert system that identifies structural fragments, potentially responsible for hazardous activity. The analysis was based on experimental data obtained from FDA, and involved 21 endpoints corresponding to different mechanisms of toxic action: mutagenicity, clastogenicity, carcinogenicity, etc. Probabilistic models for most endpoints were derived using GALAS (Global, Adjusted Locally According to Similarity) modeling methodology developed in our group. The updated list of alerting groups contained 70 distinct substructures. The expert system was highly sensitive, recognizing >90% of potent carcinogens, as classified by the FDA. Sensitivity of probabilistic models ranged from about 60% to more than 90%, while maintaining high (>80%) specificity of predictions for the majority of considered assays. These results show that the described computational platform ensures sufficient prediction accuracy for rapid genotoxicity/ carcinogenicity profiling of various chemicals.
