The use of artificial intelligence and big data for the safety evaluation of US food-relevant chemicals

Environmental contaminants, naturally occurring toxicants, pesticide residues, and food additives are the four chemical-associated categories of six for food safety established by the Food and Drug Administration. The direct food additives, which are intentionally added to food, are the main focus of this case study, and the indirect food additives, such as pesticides, natural toxicants, and environmental residues will also be discussed. This study is attempting to investigate how artificial intelligence tools developed using big data could support the hazard evaluation of food additives. Automated read-across technology, that is, the read-across-based structure activity relationships (RASAR) tool, was utilized to generate predictions, which were compared with traditional animal testing methods to assess utility for providing estimates of chemical toxicity for food-relevant substances. This was conducted using Underwriters Laboratories (UL) Cheminformatics Tool Kit followed by descriptive statistics and performance-based validation with datasets retrieved from sources such as the European Chemicals Agency, the US Environmental Protection Agency, the Occupational Safety and Health Administration, the European Food Safety Authority, and other literature. In our analysis, the main findings indicate that more direct food additives than indirect food additives are in the training data and there were more non-toxicants than toxicants, which was expected for food-related substances. Most results were at “very strong” and “strong” reliability level. For 123 cases, where classifications could be retrieved from other sources for a preliminary validation, 83% of the RASAR results matched with the toxicological assessment results confirming that in silico tools can robustly generate predictions for informing on the potential of food-use chemical toxicity.