Screening for genotoxicants in food: A data-driven approach using food composition data and machine learning based in silico models

Abstract

Foods represent complex mixtures of constituents and contaminants, some of which may pose risks to health through genotoxic effects. We investigated the current capabilities and limitations of a data-driven approach for the systematic identification of genotoxic substances in food. To this end, we used machine learning to develop quantitative structure–activity relationship (QSAR) models aimed at predicting outcomes for three in vitro genotoxicity assays: the bacterial reverse mutation assay (Ames test), the in vitro chromosomal aberration test (CAvit) and the in vitro micronucleus test (MNvit). These models were applied to screen for putative dietary genotoxicants using the FooDB compound dataset (n = 70,477) as a case study. Overall, 6.6 % of the FooDB compounds were predicted as positive by at least one in silico model, while 7.1 % were predicted as negative by all three models. Depending on the predicted endpoint, between 77 % and 82 % of the FooDB compounds fell outside the model’s applicability domain or gave an equivocal prediction. Interestingly, of the 4,683 FooDB compounds predicted to be positive in at least one in vitro assay, only 491 could be mapped to an experimental data point. As a strategy to progress from in silico screening to risk assessment, we propose a tiered approach that integrates in silico modelling, exposure assessment and experimental testing. While it has to be acknowledged that current food composition databases and in silico models still have limitations, this work illustrates that data-driven approaches hold great promise for enhancing the identification of genotoxicants in foods.