A decision tree–based integrated testing strategy for tailor-made carcinogenicity evaluation of test substances using genotoxicity test results and chemical spaces. (14th December 2018)

• Record Type: Journal Article
• Authors: Fujita, Yurika; Honda, Hiroshi; Yamane, Masayuki; Morita, Takeshi; Matsuda, Tomonari; Morita, Osamu
• Is Part Of: Mutagenesis. Volume 34:Number 1(2019)
• Page Start: 101
• Page End: 109
• Publication Date: 2018-12-14
• Subjects: Mutagenesis -- Periodicals; Mutagenicity Tests -- Periodicals; Mutagens -- Periodicals; Mutagenesis; Periodicals; 576.542
• DOI: 10.1093/mutage/gey039
• Languages: English
• ISSNs: 0267-8357

Abstract

Abstract: Genotoxicity evaluation has been widely used to estimate the carcinogenicity of test substances during safety evaluation. However, the latest strategies using genotoxicity tests give more weight to sensitivity; therefore, their accuracy has been very low. For precise carcinogenicity evaluation, we attempted to establish an integrated testing strategy for the tailor-made carcinogenicity evaluation of test materials, considering the relationships among genotoxicity test results (Ames, in vitro mammalian genotoxicity and in vivo micronucleus), carcinogenicity test results and chemical properties (molecular weight, logKow and 179 organic functional groups). By analyzing the toxicological information and chemical properties of 230 chemicals, including 184 carcinogens in the Carcinogenicity Genotoxicity eXperience database, a decision tree for carcinogenicity evaluation was optimised statistically. A decision forest model was generated using a machine-learning method—random forest—which comprises thousands of decision trees. As a result, balanced accuracies in cross-validation of the optimised decision tree and decision forest model, considering chemical space (71.5% and 75.5%, respectively), were higher than balanced accuracy of an example regulatory decision tree (54.1%). Moreover, the statistical optimisation of tree-based models revealed significant organic functional groups that would cause false prediction in standard genotoxicity tests and non-genotoxic carcinogenicity (e.g., organic amide and thioamide, saturated heterocyclic fragment and aryl halide). In vitro genotoxicity tests were the most important parameters in all models, even when in silico parameters were integrated. Although external validation is required, the findings of the integrated testing strategies established herein will contribute to precise carcinogenicity evaluation and to determine new mechanistic hypotheses of carcinogenicity.