Derivation of a Human In Vivo Benchmark Dose for Perfluorooctanoic Acid From ToxCast In Vitro Concentration-Response Data Using a Computational Workflow for Probabilistic Quantitative In Vitro to In Vivo Extrapolation

Abstract

A computational workflow which integrates physiologically based kinetic (PBK) modeling, global sensitivity analysis (GSA), approximate Bayesian computation (ABC), and Markov Chain Monte Carlo (MCMC) simulation was developed to facilitate quantitative in vitro to in vivo extrapolation (QIVIVE). The workflow accounts for parameter and model uncertainty within a computationally efficient framework. The workflow was tested using a human PBK model for perfluorooctanoic acid (PFOA) and high throughput screening (HTS) in vitro concentration-response data, determined in a human liver cell line, from the ToxCast/Tox21 database. In vivo benchmark doses (BMDs) for PFOA intake (ng/kg BW/day) and drinking water exposure concentrations (µg/L) were calculated from the in vivo dose responses and compared to intake values derived by the European Food Safety Authority (EFSA). The intake benchmark dose lower confidence limit (BMDL₅) of 0.82 was similar to 0.86 ng/kg BW/day for altered serum cholesterol levels derived by EFSA, whereas the intake BMDL₅ of 6.88 was six-fold higher than the value of 1.14 ng/kg BW/day for altered antibody titer also derived by the EFSA. Application of a chemical-specific adjustment factor (CSAF) of 1.4, allowing for inter-individual variability in kinetics, based on biological half-life, gave an intake BMDL₅ of 0.59 for serum cholesterol and 4.91 (ng/kg BW/day), for decreased antibody titer, which were 0.69 and 4.31 the EFSA-derived values, respectively. The corresponding BMDL₅ for drinking water concentrations, for estrogen receptor binding activation associated with breast cancer, pregnane X receptor binding associated with altered serum cholesterol levels, thyroid hormone receptor α binding leading to thyroid disease, and decreased antibody titer (pro-inflammation from cytokines) were 0.883, 0.139, 0.086, and 0.295 ng/ml, respectively, with application of no uncertainty factors. These concentrations are 5.7-, 36-, 58.5-, and 16.9-fold lower than the median measured drinking water level for the general US population which is approximately, 5 ng/ml.

Keywords: bayesian; in silico; in vitro; physiologically based kinetic; reverse dosimetry.

FIGURE 1

PBK model for PFOA was evaluated by reproducing Figures 2–4 from Worley et al. (2017). The solid, colored lines represent simulations of ExposedDW (drinking water PFOA concentrations) which were set to the highest concentration reported for each water authority; these were 0.04 μg/L (red), 1.0 μg/L (green), and 4.9 μg/L (blue) for North Alabama, Lubeck Public Service District, and Little Hocking Water Association, respectively. The simulations were for 30 years with a further 10-year postexposure period. The corresponding serum PFOA concentrations are shown as solid colored symbols. The vertical gray shaded band between 100,000 and 120,000 h (11.4–13.7 years) highlights the point where steady state was judged to have been achieved; the model predictions from this period were used for QIVIVE calculations.

FIGURE 2

Lowry plots of the eFAST quantitative measure of the most sensitive parameters identified by Morris screening following oral (drinking water) exposure. The total effect of a parameter S_Ti comprised the main effect S_i (black bar) and any interactions with other parameters (gray bar) given as a proportion of variance. The ribbon, representing variance due to parameter interactions, is bounded by the cumulative sum of the main effects (lower bound of ribbon) and the minimum of the cumulative sum of the total effects (upper bound of ribbon). (A) For CL, liver cell concentrations (upper panel) and (B) for CA, serum concentrations (lower panel).

FIGURE 3

Comparisons of concentration response profiles simulated in the rejection phase were run for each dose concentration. A typical example is shown for a target concentration of 1,035.175 μg/L. (A) 500 concentration response profiles following 120,000 h exposure (upper panel) and (B) retained samples within a relative error of 7.5% (lower panel).

FIGURE 4

Typical predicted in vivo dose–response curves for Intake (upper panels) and ExposedDW (lower panels) for each of the in vitro datasets. These were for estrogen receptor binding activation leading to breast cancer (A) and (E), pregnane X receptor binding leading to hepatic steatosis (B) and (F), thyroid hormone receptor a binding leading to thyroid disease (C) and (G), and immunotoxicity (pro-inflammation from cytokines) (D) and (H). The curves for the modes only are shown. Benchmark dose values were calculated from such curves for the mode, lower and upper bounds (2.5 and 97.5%) of the credible intervals (Table 5).