Reported liver toxicity of food chemicals in rats extrapolated to humans using virtual human-to-rat hepatic concentration ratios generated by pharmacokinetic modeling with machine learning-derived parameters.

Pharmacokinetic data are not generally available for evaluating the toxicological potential of food chemicals. A simplified physiologically based pharmacokinetic (PBPK) model has been established to evaluate internal exposures to chemicals in rats or humans with no reference to in vitro or in vivo experimental data. In this study, reported liver toxicity levels in rats were extrapolated to humans using virtual hepatic concentration-time curves (AUC) as the interspecies factor. Virtual liver exposures to 27 lipophilic food chemicals (octanol-water partition coefficient logP >1) with reported rat hepatic lowest-observed-effect levels (LOELs) of ≤1000 mg/kg/day were generated using PBPK models with input parameters obtained entirely in silico via machine learning algorithms. The resulting virtual rat and human liver AUCs were correlated (n = 27, r = 0.52, p < 0.01). However, AUCs for the phenolic compounds emodin, isoeugenol, and tert-butylhydroquinone, which have reported rat LOEL values of ≤300 mg/kg/day, were located outside the relatively wide 95% confidence interval, indicating more extensive hepatic elimination in rats than in humans. In vitro depletion of tert-butylhydroquinone in rat liver fractions via sulfation was confirmed to be faster than that in humans. For emodin, isoeugenol, and tert-butylhydroquinone, human-to-rat AUC ratios ranged from 10- to 13-fold; consequently, their extrapolated human hepatic LOEL values were estimated as ≤30 mg/kg/day, i.e., one order of magnitude smaller than the rat LOELs. Despite the small number of lipophilic food chemicals considered here, the PBPK modeling approach using in silico-generated input parameters for rats and humans has the potential to facilitate toxicological studies.