A novel prediction model to evaluate genotoxicity based on a gene signature in metabolically competent human HepaRGTM cells.

Genotoxicity testing is essential to ensure the safety of newly developed substances for human health. Traditionally, a stepwise standardized approach is applied, starting with a battery of in vitro tests. Despite its wide applicability and high sensitivity, the current in vitro genotoxicity battery is facing several limitations, among which the high number of misleading positive results which provoke unnecessary and costly follow-up in vivo animal studies [1-3]. To improve the predictive capacity of genotoxicity testing, mechanistic information at the molecular level is needed which can be obtained via new approach methodologies including gene expression biomarkers [4-7]. In this context, we previously developed GENOMARK to identify genotoxic substances in metabolically competent human HepaRG™ cells. GENOMARK consists of 84 biomarker genes, derived from whole genome transcriptomics data and selected to cover diverse modes of action, including bulky adduct formation, DNA alkylation, cross-linking, radical generation causing DNA strand breaks, inhibition of tubulin polymerization and base analogues [8, 9]. Briefly, cells are exposed for 72 hours to a concentration of the test chemical inducing a low level of cytotoxicity. Afterwards, gene expression levels are evaluated with RT-qPCR and results are automatically analyzed with a support vector machine (SVM)-based model to classify chemicals as genotoxic or non-genotoxic. In the present study, an improved prediction model was developed based on a different algorithm called random forest (RF). To this extent, the existing reference dataset of 24 chemicals was enlarged to 38 by selecting additional in vivo genotoxic and non-genotoxic chemicals for which new test data were generated. Next, two supervised machine learning algorithms in R software, i.e. SVM and random RF were applied on the extended dataset and their predictive capacity was compared. Both prediction models showed the same predictive accuracy (i.e. 92.3%), although the RF model displayed a higher sensitivity and a lower specificity compared to the SVM model. As the RF model is also less sensitive to outliers, this model was selected and further applied to predict the genotoxicity of 6 misleading positive chemicals. Four of these (2-Methyl-4-isothiazolin-3-one, 4-Amino-3-nitrophenol, Sodium benzoate and Dihydroxyacetone) were correctly classified as non-genotoxic by the GENOMARK biomarker. One chemical was classified as equivocal (hydroxybenzomorpholine) and one as genotoxic (1-napthol). These results demonstrate that GENOMARK could be useful as a follow-up of the in vitro genotoxicity test battery to de-risk misleading positives in a Weight of Evidence approach and may contribute to the paradigm shift in genetic toxicology to move towards a more human-relevant genotoxicity testing.