Development of a genotoxicity/carcinogenicity assessment method by DNA adductome analysis

Abstract

Safety evaluation is essential for the development of chemical substances. Since in vivo safety evaluation tests, such as carcinogenesis tests, require long-term observation using large numbers of experimental animals, it is necessary to develop alternative methods that can predict genotoxicity/carcinogenicity in the short term, taking into account the 3Rs (replacement, reduction, and refinement). We established a prediction model of the hepatotoxicity of chemicals using a DNA adductome, which is a comprehensive analysis of DNA adducts that may be used as an indicator of DNA damage in the liver. An adductome was generated with LC-high-resolution accurate mass spectrometer (HRAM) on liver of rats exposed to various chemicals for 24 h, based on two independent experimental protocols. The resulting adductome dataset obtained from each independent experiment (experiments 1 and 2) and integrated dataset were analyzed by linear discriminant analysis (LDA) and found to correctly classify the chemicals into the following four categories: non-genotoxic/non-hepatocarcinogens (−/−), genotoxic/non-hepatocarcinogens (+/−), non-genotoxic/hepatocarcinogens (−/+), and genotoxic/hepatocarcinogens (+/+), based on their genotoxicity/carcinogenicity properties. A prototype model for predicting the genotoxicity/carcinogenicity of the chemicals was established using machine learning methods (using random forest algorithm). When the prototype genotoxicity/carcinogenicity prediction model was used to make predictions for experiments 1 and 2 as well as the integrated dataset, the correct response rates were 89 % (genotoxicity), 94 % (carcinogenicity) and 87 % (genotoxicity/carcinogenicity) for experiment 1, 47 % (genotoxicity), 62 % (carcinogenicity) and 42 % (genotoxicity/carcinogenicity) for experiment 2, and 52 % (genotoxicity), 62 % (carcinogenicity), and 48 % (genotoxicity/carcinogenicity) for the integrated dataset. To improve the accuracy of the toxicity prediction model, the toxicity label was reconstructed as follows; Pattern 1: when +/+ and −/− chemicals were used from the toxicity labels +/+, +/−, −/+ and −/−; and Pattern 2: when +/+, +/−, and −/+ other than −/− were replaced with the label "Others". As a result, chemicals with only +/+ and −/− toxicity labels were used and the correct response rates were approximately 100 % for the measured data in experiment 1, 53 %–66 % for the data in experiment 2, and 59–73 % for the integrated data, all of which were 10 %–30 % higher compared with the data before the label change. In contrast, when the toxicity labels were replaced with −/− and “Others”, they reached nearly 100 % in the measured data from experiment 1, 65 %–75 % in the data from experiment 2, and 70 %–78 % in the integrated data, all of which were 10 %–50 % higher compared with the data before the label change.