PBPK Modeling of Impact of Nonalcoholic Fatty Liver Disease on Toxicokinetics of Perchloroethylene in Mice

Abstract

Background

Nonalcoholic fatty liver disease (NAFLD), a major cause of chronic liver disease in the Western countries with increasing prevalence worldwide, may substantially affect chemical toxicokinetics and thereby modulate chemical toxicity.

Objectives

This study aims to use physiologically-based pharmacokinetic (PBPK) modeling to characterize the impact of NAFLD on toxicokinetics of perchloroethylene (perc).

Methods

Quantitative measures of physiological and biochemical changes associated with the presence of NAFLD induced by high-fat or methionine/choline-deficient diets in C57B1/6J mice are incorporated into a previously developed PBPK model for perc and its oxidative and conjugative metabolites. Impacts on liver fat and volume, as well as blood:air and liver:air partition coefficients, are incorporated into the model. Hierarchical Bayesian population analysis using Markov chain Monte Carlo simulation is conducted to characterize uncertainty, as well as disease-induced variability in toxicokinetics.

Results

NAFLD has a major effect on toxicokinetics of perc, with greater oxidative and lower conjugative metabolism as compared to healthy mice. The NAFLD-updated PBPK model accurately predicts in vivo metabolism of perc through oxidative and conjugative pathways in all tissues across disease states and strains, but underestimated parent compound concentrations in blood and liver of NAFLD mice.

Conclusions

We demonstrate the application of PBPK modeling to predict the effects of pre-existing disease conditions as a variability factor in perc metabolism. These results suggest that non-genetic factors such as diet and pre-existing disease can be as influential as genetic factors in altering toxicokinetics of perc, and thus are likely contribute substantially to population variation in its adverse effects.

1. Introduction

A major challenge in estimating health risks associated with exposure to xenobiotics is population heterogeneity in the adverse health effects due to inter-individual variation in the biologically effective tissue dose (Bois, 2010). Inter-individual variance in toxicokinetics are due to genetic polymorphisms, age, sex, diet, and concomitant disease states. Understanding of the impact of these factors on toxicokinetics may refine the estimates of the variability in exposed populations.

Chronic disease conditions in a number of metabolism-active organs could have profound impacts on toxicokinetics; systemic and/or tissue concentrations of chemicals and metabolites may vary greatly, which in turn leads to variability in their adverse effects. Physiologically-based pharmacokinetic (PBPK) modeling is an invaluable tool to dissect the impact of disease conditions on toxicokinetics. The rapid advancement in PBPK modeling, as well as the increasing quantitative knowledge of disease-related pathophysiological changes, facilitate the development of chemical–disease models. Still, few PBPK models developed to date include the pathophysiological changes associated with chronic disease (Li et al., 2015a; Marsousi et al., 2017).

Nonalcoholic fatty liver disease (NAFLD) is a spectrum of the pathological states of in the liver, a condition that is of increasing impact on public health as its prevalence is on the rise (Younossi et al., 2018). NAFLD is a major type of chronic liver disease in the West and is steadily increasing worldwide, including the associated conditions of metabolic syndrome, hypertension, obesity and type 2 diabetes (Vanni et al., 2014). NAFLD severity ranges from nonalcoholic fatty liver (NAFL) to nonalcoholic steatohepatitis (NASH) (Popescu et al., 2016). NASH has the potential for progression to more grave liver diseases such as cirrhosis and hepatocellular carcinoma (Stephenson et al., 2018).

One example of liver disease-chemical metabolism connection are studies of the contribution of NAFLD to inter-strain variation in toxicokinetics of perchloroethylene (perc) (Cichocki et al., 2017a; Cichocki et al., 2017c; Cichocki et al., 2018). Perc is a ubiquitous environmental chemical found in ambient air, at hazardous waste sites, and in surface and ground water due to its extensive use, high volume production, and persistence in the environment (NRC, 2010; Cichocki et al., 2016). Exposure to perc is associated with both cancer and noncancer adverse effects in multiple organs in humans and animals (USEPA, 2011a; USEPA, 2011b; IARC, 2014). Investigation of the toxicokinetics of perc and its metabolites in multiple tissues is essential for understanding of perc-associated adverse effects, because delivered doses of perc and its specific metabolites modulate organ-specific toxicity (Lash and Parker, 2001; Guyton et al., 2014; Lash et al., 2014; Cichocki et al., 2016; Cichocki et al., 2017c; Luo et al., 2018; Luo et al., 2019). For example, neurotoxicity in humans is associated with the concentrations of perc in the central nervous system (Bushnell et al., 2005; USEPA, 2011a). Hepatotoxicity in mice is associated with hepatic concentrations of trichloroacetate (TCA), an oxidative metabolite of PERC (Bull et al., 1990; Corton, 2008). Kidney cancer in rats is associated with the reactive species of glutathione (GSH) metabolites of perc formed in the renal proximal tubules, including trichlorovinyl cysteine (TCVC)-sulfoxide (Elfarra and Krause, 2007) and N-acetyl trichlorovinyl cysteine (NAcTCVC)-sulfoxide (Lash and Parker, 2001; Cristofori et al., 2015).

The objective of this study was to model the impact of diet-induced NAFLD states on toxicokinetic variability of perc. We utilize new empirical data on the inter-individual variability due to the physiological and biochemical changes that were the result of NAFLD (Cichocki et al., 2017a), to extend the most recent population PBPK model for perc and its metabolites (Dalaijamts et al. 2018) beyond addressing genetic variability. This work is an important advance which allows for prediction of perc toxicokinetics in subjects with underlying pathological liver conditions, a very important emerging factor that may have profound impact on chemical metabolism and adverse health effects.

2. Methods

2.1. PBPK model updates

The PBPK model detailed herein is an update to the recent model for perc and its metabolites in mice reported in Dalaijamts et al. (2018). For detailed descriptions of the model structure and equations, see Dalaijamts et al. (2018) and the references therein. Updates consisted of changes in parameter initialization in order to include in vivo measured physiological and biochemical parameters (see Section 2.4.1. Prior assignment of model parameters). Full model code is provided in the Supplemental materials.

2.2. PBPK model parameters and baseline values

Most PBPK model parameters and their baseline values and sources, including physiological measurements (e.g., volumes and flows) and biochemical measurements from in vitro studies (e.g., partition coefficients (PCs), binding coefficients, and metabolism and clearance rates), are listed in Supplemental Table S–1, and were previously described by Chiu and Ginsberg (2011) and Dalaijamts et al. (2018). In this study, modifications were made in the baseline values to incorporate the quantitative measures of physiological and biochemical changes associated with the presence of varying stages of diet-induced NAFLD into the previously developed model. Specifically, default values of fat and liver volumes, and blood:air and liver:air PCs are replaced with in vivo measurements for C57BL/6J mice with three different diets that served as surrogates for various stages of diet-induced NAFLD. Animals fed low fat diet (LFD) were used for a control diet group (C57BL/6J-control), high fat diet (HFD) for inducing NAFL (C57BL/6J-NAFL), and methionine/folate/choline-deficient high-fat diet (MCD) for inducing NASH without fibrosis (C57BL/6J-NASH) (Cichocki et al., 2017a) (Table 1).

Table 1.

Measured physiological and biochemical parameters in 3 diets of male C57BL/6J mice obtained from an in vivo experiment ( mean ± S.D., n = 5/group) (Cichocki et al., 2017a).

Parameter	C57BL/6J-control	C57BL/6J-NAFL	C57BL/6J-NASH
Body weight (BW, kg)*	0.027 ± 0.002	0.033 ± 0.002#+	0.030 ± 0.002#
Fat/BW ratio*	0.075 ± 0.009	0.284 ± 0.026#+	0.173 ± 0.047#
Liver/BW ratio*	0.049 ± 0.005	0.047 ± 0.005#+	0.062 ± 0.008#
Blood:Air PC	18.1 ± 12.6	29.4 ± 6.3	35.7 ± 11.6
Liver:Air PC*	61.3 ± 9.9	149 ± 101	300 ± 158#

^*p < 0.05 by one-way ANOVA

^#p < 0.05 compared with control diet group

⁺p < 0.05 compared with NASH group, using Newman-Keuls post-hoc test.

2.3. In vivo toxicokinetic data

In addition to in vivo toxicokinetic data for male mice of three strains, including B6C3F1/J, NIH Swiss Webster (SW), and C57BL/6J from previously published model (Dalaijamts et al., 2018), new data from liver-diseased male mice of C57BL/6J strain from (Cichocki et al., 2017a; Cichocki et al., 2018) were added to the previous model in order to characterize inter-strain and inter-diet variability (Fig. 1). The in vivo toxicokinetic data for B6C3F1/J and SW mice, and LFD-fed (i.e., control) C57BL/6J mice were as detailed in the previous perc PBPK model (Chiu and Ginsberg, 2011; Dalaijamts et al., 2018). Data from male C57BL/6J -NAFL (HFD-fed) and - NASH (MCD-fed) mice (n=5/group) dosed with perc at 300 mg/kg using aqueous-based vehicle by oral gavage were added in the model for calibration and validation.

Fig. 1.

Experimental data used for the analyses were previously reported. B6C3F1/J hybrid, NIH Swiss Webster (albino) outbred mice, and C57BL/6J inbred mice were included in the model. Data for B6C3F1 and Swiss strains were initially collected by (Chiu and Ginsberg, 2011), while data for three diets of C57BL/6J mice was obtained from (Cichocki et al., 2017a).

We focused on male mice because the most experimental toxicokinetic data come from male mice. This is likely due to the fact that perc-induced tumor-development, including hepatocellular adenomas and carcinomas, are observed in male mice with higher incidence (Lash et al., 2014). Moreover, protein levels of CYP2E1 were slightly higher in liver of male mice as compared with female mice. This sex dependent difference was even more pronounced in kidneys where no detectable CYP2E1 protein was found in females. Protein levels of CYP2E1, as well as the oxidative metabolites of TCE and PCE, were higher in liver and kidney of male mice as compared with female mice (Luo et al., 2018). For these reasons, perc risk assessment based on male mice would presumably be protective of females.

Parameters analyzed included concentration-time course data on perc in serum, liver, kidney and fat, its oxidative metabolite, TCA, in plasma, liver and kidney, and GSH conjugation metabolites, TCVG, TCVC and NAcTCVC, in serum, liver and kidney. Serum and liver data are described in (Cichocki et al., 2017a) and kidney data are described in (Cichocki et al., 2018).

2.4. Bayesian Population PBPK Model Approach

A hierarchical Bayesian population approach was used to construct a statistical model for estimation of PBPK model parameters and their uncertainty and variability as previously described by (Bois, 2000a; Bois, 2000b; Hack et al., 2006; Chiu et al., 2009). The conceptual representation of the hierarchical Bayesian statistical model was updated from (Dalaijamts et al., 2018) and is shown in Fig. 2.

Fig. 2.

Updated scheme of Hierarchical Bayesian population model for PBPK model parameter uncertainty and variability in mice. Square nodes denote variables for which the values were observed, such as y_ij or φ_i; were fixed by the experimenters, such as E_ij and t_ij; or were fixed by us, such as the prior on M_θ and V_θ, represent a functional relationship [B=f(A)]. The population consists of strain/diet i, each of circle nodes represent uncertain or unobserved quantities, such as σ², θ_i, M_θ or V_θ; inverted trapeze denotes nonlinear deterministic model f_ij. Solid arrows denote a stochastic relationship represented by a conditional distribution [A→B means B ~ P(B|A)], dashed arrows which undergo one or more experiments j with exposure parameters E_ij with data y_ij collected at times t_ij (Gelman et al., 1996). Each “strain/diet” consists of animals (usually comprising multiple dose groups) of the same sex within a study. Mice within each group are presumed to be “identical,” with the same PBPK model parameters. The PBPK model depends on measured covariates ϕ_i (e.g., body weight and tissue volumes) and unobserved model parameters θ_i (e.g., V_Max) drawn from a population with mean M_θ and variance V_θ, each of which is uncertain and has a prior distribution assigned to it. The PBPK model produces outputs f_ij for comparison with the data y_ij. The difference between them (“residual error”) has a variance σ². In the mouse, all strains are assigned the same prior distribution Prσ of σ².

The mouse population model consists of strains/diets i, each of which undergoes one or more study/experiments j with various exposure E_ij routes and doses, each of which provides in vivo toxicokinetic (concentration-time course data) data y_ij collected at times t_ij (Fig. 1). Different animals of the same sex and strain in the same study (or series of studies conducted simultaneously) were treated as identical at the strain/diet level, with the same PBPK model parameters. The nonlinear (deterministic) PBPK model (shown structurally in Fig. 2 and described previously in (Dalaijamts et al., 2018)) predicts the toxicokinetics for comparison with the data y_ij, which are a function f_ij of exposure parameters E_ij, collected times t_ij, a set of physiological and chemical-specific parameters of unknown values θ_i, and a set of measured covariate parameters (φ_i). E_ij and t_ij are study-specific, and θ_i and φ_i are strain/diet-specific.

The parameters θ_i are drawn from a population with population means μ and population variances V, each of which reflects uncertainty about the true mean and variability among the population, and was assumed to be priori distributed log-normally. A priori knowledge of μ and V is available in the form of “standard” values for some parameters. Uncertainty in these means and variances was acknowledged under the form of a priori log-normal or log-uniformly distributions for the population means μ depending on available information and a standard inverse gamma distribution, with parameters α=1 and β=V₀ for the population variances V (see section below on a priori parameter values).

2.4.1. Prior assignment of model parameters for statistical analysis

The PBPK model parameters used for statistical analysis are similar to those of previously published model (Dalaijamts et al., 2018). In principle, a population PBPK model requires a large number of parameters to be estimated since each model parameter must be estimated for each experimental subject and for the population as a whole; this can lead to computational burden with greatly increased number of iterations to assess convergence. To reduce this burden, the parameters, which are most influential to model outputs, are selected for statistical estimation and the selection processes as described in a previous model (Dalaijamts et al., 2018). Prior uncertainty distributions of population mean and population variability for the selected parameters are summarized in Supplemental Material (Supplemental Table S–2). The assumptions of prior distributions reflecting the uncertainty in the population mean and variance of previously assigned parameters were previously described (Chiu et al., 2009; Chiu and Ginsberg, 2011; Dalaijamts et al., 2018). In summary, the informative prior distributions for the population mean of parameters (with independent data) were centered on either baseline values of model parameters or the posterior means from the previously published multi-strain mouse PBPK model of trichloroethylene (Chiu et al., 2014). For parameters without independent data either a truncated normal distribution centered on baseline values with high standard deviation (conjugate PCs) or log-uniform distributions with upper and lower bounds separated by at least 10⁴ (absorption and metabolism/ clearance of perc metabolites) priors were set as to minimize potential bias. The spread of the distributions for the population mean and variance is referred to as the coefficient of uncertainty and described in detail previously (Bois, 2000a; Hack, 2006; Chiu et al., 2009).

The empirical toxicokinetic data used in this study is subject to both measurement error and inter-individual variability within the group. Because neither experimental measurements nor the model predictions can fully reflect the true in vivo concentrations, error distributions are defined for likelihood of each kinetic endpoint modeled. The corresponding errors are assumed to be independent and log-normally distributed, with mean zero and variance δ² (on the log scale). The variance vector δ² called “residual error” estimate for the measurements were unknown and sampled with a log-uniform prior distribution with a lower bound of 0.01 to prevent the possibility of estimating an error of zero. It provides some quantitative measure of the degree to which there were deviations due to model misspecification, including any difficulties fitting multiple dose levels in the same study using the same model parameters. The detailed description of likelihood of the toxicokinetic data and their non-detects included in the analysis can be found in the previous model by (Dalaijamts et al., 2018).

2.4.2. Estimation of posterior parameter distributions

A hierarchical Bayesian population analysis using Markov Chain Monte Carlo (MCMC) simulations was conducted to estimate PBPK model parameters and characterize the uncertainty and inter-strain/diet variability in perc toxicokinetics. Bayesian approaches aim at identifying a probability distribution of the parameter vector, considering prior knowledge about the model parameters, and ‘update’ this prior knowledge by integration of new data based on Bayes’ theorem:

$$p({\theta }_i|Y)\propto p({\theta }_i)\ast p(Y|{\theta }_i)$$

The posterior distribution p(θ_i|Y) combines prior knowledge p(θ_i) about the parameter vector θ (θ∈ℝ^P×1) with the likelihood function p(Y|θ). The likelihood function represents the closeness to the data Y, where Y_ij = N(f_ij , σ²) with i=1,…,N_j. x_i,j represents the measurement of individual j (j=1,…,J) at time point t_ij. Application of Bayesian approach to our PBPK model were described in detail previously by (Chiu et al., 2009; Chiu et al., 2014), and updated scheme shown in Fig. 2.

MCMC simulation was utilized to obtain samples from the posterior distribution. Hierarchical Metropolis-Hastings algorithms within the Gibbs sampler (Geman and Geman, 1984; Gelfand and Adrian, 1990) was used in MCMC simulation. MCMC algorithm draws a sample along a Markov chain, which is constructed to have the posterior distribution as its long-run stationary distribution (Andrieu et al., 2003; Krauss et al., 2013; van Ravenzwaaij et al., 2018). Each iteration of the Markov chain generates a vector of parameters sampled from the parameter distributions and values of the posterior likelihood associated with that vector of parameters. The MCMC simulation, thereby, generates posterior (updated) parameter values at the population level, and parameter values for each strain/diet of mice in the current model. Strain/diet-specific posterior parameter distributions are estimated under the population distributions.

Analysis of PBPK model in conjunction with statistical model was performed using MCSim v5.6.5 software (Bois, 2009). To demonstrate the number of iterations of the Markov chain needed to get centered on the true population mean, we used four independent MCMC with different starting values each run out to 100,000 iterations (one of every 10^th iteration was recorded). The MCMC statistical model code is provided in Supplementary Materials.

2.5. Model evaluation

2.5.1. Evaluation of convergence

Convergence of the Markov chains to the posterior distribution was monitored using analysis of variance as described by (Gelman et al., 1996). The Gelman & Rubin shrink factors (potential scale reduction factor, R), a ratio of an upper bound and a lower bound of the variance in the target distribution, and Brooks-Gelman multivariate shrink factor are used to assess whether the independent MCMC chains have converged to a common distribution. As multiple, independent chains move closer together toward the same distribution, the ratio declines to unity. A convergence diagnostic “R” of 1.2 or less has been proposed as a criterion for acceptable convergence (Gelman et al., 2004). A visual inspection for dependency of posterior parameter distributions using cross correlations and convergence of the chains using traces and probability density functions for each posterior parameter distributions was performed as additional diagnostics. As is standard, the first half of the iterations for each chain were discarded as “burn-in” iterations, i.e., iterations for which the simulation had not yet converged, and the remainders used for inferences. In total, 100,000 iterations were performed for each of the four independent Markov chains, with the first 50,000 discarded.

2.5.2. Evaluation of posterior parameter distributions

Posterior distributions of the population parameters were checked as to whether they appear reasonable given the prior distributions, while posterior distributions of the strain/diet-specific parameters were estimated to see variability across strains. Inconsistency between the prior and posterior distributions of population mean parameters may indicate the insufficient broad priors (i.e., overconfidence in their specification), a misspecification of the model structure (e.g., leading to pathological parameter estimates), or a measurement error. Note that population variance of model parameters reflects both uncertainty and variability in the population.

2.6. Posterior model predictions

Posterior model predictions of internal toxicokinetics of perc and its metabolites were made following MCMC calibration using the population posterior distributions as well as strain/diet-specific posterior parameter values accounting for both inter-strain/-diet variability and parameter uncertainty. At a population level, parameter values for “random strains” were sampled from the population parameters (means and variances). Thus, the predictions were only conditioned on the population-level parameter distributions, representing a distribution across all the strains, and not on the specific predictions for that dataset. These strains then represent the predicted population distribution, incorporating variability in the population as well as uncertainty in the population means and variances and therefore compares predictions for the overall mouse population. In contrast to population level, because, strain/diet-specific prediction was sampled from posterior parameter distribution of the specific strain/diet, selectively, that prediction represents all datasets of the specific strain/diet.

Residual error estimated for each in vivo measurement provides some quantitative measure of the degree to which there were deviations due to intra-study variability, inter-individual variability, and measurement and model errors, including any difficulties fitting multiple dose levels in the same study using the same model parameters. Estimated residual error with geometric standard deviation of more than 3-fold was assumed to indicate that the model prediction does not fit the in vivo measurement adequately.

2.7. Posterior dose metric predictions

The model posteriori was used to compute surrogate dose metrics. Strain/diet-specific posterior dose metrics were estimated for perc metabolism and perc and its metabolite’s area under the concentration-time curve. For the dosimetry comparison, all strains modeled using a single dose of 300 mg/kg oral gavage of perc. Absorption parameters were set to aqueous vehicle for all strains except B6C3F1/J, for which an oil vehicle was assumed to correspond with that used in the majority studies with this strain. Amount metabolized were estimated at 36 hours after exposure.

3. Results

3.1. Model convergence analysis

All values of the convergence diagnostic potential scale reduction factors (R) were <1.05 for all population parameters, indicating convergence. Traces of four chains were well mixed, which were hardly distinguished by visual inspection for all parameters. Figures of cross-correlations, traces and probability density functions of posterior uncertainty distributions for each calibrated parameter population mean of the last-half of 100,000 iterations of four Markov chains are in Supplementary materials, Supplemental Figs. S1 and S2. From the last half of every MCMC chain, 1,250 iterations were randomly selected, yielding 5,000 sets of parameter values, which the inferences and predictions presented in the following were made.

3.2. Evaluation of posterior parameter distributions

Comparisons of the prior and posterior uncertainty distributions of population parameters are shown in Fig. 3a and 3b. All posteriors uncertainty distributions of population means are narrower than the corresponding informative and non-informative priors, indicating that substantial information on almost all parameters has been gained from the experimental data in the updated posteriors. The resulting posterior distributions are generally consistent with those in previous model (Dalaijamts et al., 2018) for almost all parameters. For oral exposure, because, the gastrointestinal absorption parameters are highly uncertain with unknown priori, they are tended to be among those that took the longest to converge and not well updated in posteriori, except a parameter of absorption of perc through duodenum (kADAq). The posterior population mean of kADAq was updated showing narrower distribution and was slightly lower than that of previous model. The most appreciable differences between posteriors in the current and the previous models were in population mean parameters of oxidative metabolism of perc, which were somewhat more strongly shifted the low end of the priors for affinity constants and high end of the priors for metabolic clearance in the current model. Overall, the Bayesian analysis of the updated PBPK model and data exhibited no major inconsistencies in prior and posterior parameter distributions. Therefore, there were no indications based on this evaluation of prior and posterior distributions either that prior distributions were overly restrictive or that model specification errors led to pathological parameter estimates.

Fig. 3.

Densities of prior and posterior parameter uncertainty distributions of (A) log-transformed population mean and (B) standard deviations for population variability, and (C) strain/diet-specific parameter posterior distributions. The distributions of parameter population mean in the current model were also compared with those predicted in the previous model (Dalaijamts et al., 2018).

Posterior distributions of parameters were also sampled for each mouse strain/diet in the framework of the population parameters, shown in Figure 3c. These strain/diet-specific values represent the sets of parameters for each strain/diet group, calibrated to the experimental data in the strain/diet. Posterior distributions for oxidative metabolism of perc are highly variable across strains and diets. In particular, the parameter for the saturable (high affinity) hepatic oxidative clearance of perc (lnClC) for B6C3F1/J mice was 10-fold lower than posterior population mean, while those for C57BL/6J-NAFL and C57BL/6J-NASH mice were 3-fold greater. By contrast, hepatic clearance through the linear (low affinity) oxidative metabolism pathway of perc was higher for B6C3F1/J (34-fold) and control diet C57BL/6J mice (2-fold) than posterior population mean, while it was lower for C57BL/6J-NASH (5-fold) and C57BL/6J-NAFL (1.5-fold) mice. The population variability across strains and liver disease conditions for these parameters is evident in the analysis of population GSD, in which posterior distributions for these parameters were shifted to 2- to 4-fold compared to the prior distributions. Moreover, strain/diet-specific kADAq showed variation across both strains and diets, where disease inducing diets were predicted to have 2-fold less than normal diet in the same strain. Regarding PKidTCA, strain/diet-specific posterior distributions showed variation across strains, but not diets, where all diets of C57 mice were predicted to have 1.4 – 2.6-fold greater PKidTCA compared to B6 and SW mice.

The summary statistics of the posterior distributions of population geometric mean, population geometric standard deviation of, and strain/diet-specific parameters are provided in Supplemental Table S–3.

3.3. Model fitting

A goodness of model fit was first evaluated by comparison between median model predictions of strain/diet-specific parameter posteriors and median values for each time-point of the in vivo data for the specific strains/diets of mice (Fig. 4a). A large majority of the predictions (81% for perc, 93% for TCA, and 75% for GSH conjugates – see Supplemental Fig. S–3) were with 2-fold of the data, indicating a good model fit. Posterior distributions of the geometric standard deviation (GSD) for the estimated residual error in each in vivo measurement of mouse population is illustrated in Fig. 4b, and summary statistics (median, 95% CIs) are presented in Supplemental Table S–4. Overall, the model predictions of toxicokinetics using strain/diet-specific parameter posteriors fit within 3-fold the in vivo measurements from (Cichocki et al., 2017a) (Fig. 4a). The estimated residual errors are less than 3-fold GSD (Fig. 4b) except for perc in liver and kidney and TCVC in liver, which have a larger residual error with median GSD up to 6.0 (Fig. 4b).

Fig. 4.

Global evaluation of goodness of model fit. Comparison of medians of strain/diet-specific model predictions (y-axis) with medians of observed toxicokinetic data (x-axis) (A) and posterior distributions of geometric standard deviation (GSD) of estimated residual error (95% CIs) (B). In plot (A), the diagonal lines indicate where data and predictions are equal and the shades are 2- and 3-fold difference. In plot (B), posterior distributions of estimated residual error in the current model was plotted against those predicted in the previous model (Dalaijamts et al., 2018). Vertical dashed line indicates 3-fold error.

Time-course comparisons for C57BL/6J mice fed the three different diets are shown in Figs. 5–7, with those for other strains in Supplemental Figs. S3–S5. In addition to strain/diet-specific posterior predictions, also shown are time-courses for a “random strain/diet” reflecting the range of inter-strain/diet variability predicted by the model. The predictions of the current model for B6C3F1/J mice, SW mice, and C57BL/6J mice on control diet are consistent with those of the previous version of the model (Dalaijamts et al., 2018).

Fig. 5.

PBPK model predictions of toxicokinetic profile (median and 95% CIs of chemical concentration) using both a population-generated “random strain” and strain/diet-specific parameter posterior distributions compared with in vivo data in male control C57BL/6J mice fed with a single dose of 300 mg/kg aqueous based oral gavage of perc (data from (Cichocki et al., 2017a)).

Fig. 7.

PBPK model predictions of toxicokinetic profile (median and 95% Cis of chemical concentration) using both a population-generated “random strain” and strain/diet-specific parameter posterior distributions compared with in vivo data in male C57BL/6J-NASH mice fed with a single dose of 300 mg/kg aqueous based oral gavage of perc (data from (Cichocki et al., 2017a)).

Regarding the toxicokinetic profiles of perc and its metabolites in two NAFLD models, the current model successfully predicts the concentration-time course of all GSH-conjugation metabolites, including TCVG, TCVC, and NAcTCVC, in both NAFL- and NASH-model mice with high accuracy using both population and diet-specific predictions (Figs. 6 and 7). Although the population mean and diet-specific model over-predicted the elimination of perc from blood and liver in both NAFL- and NASH-model mice, especially, from around 10 hours after exposure, which led the predictions to underestimate the central tendency of in vivo concentration-time course, overall, most data points are in the uncertainty range of population prediction. Diet-specific predictions slightly over-predict the elimination of perc from kidney for both NAFL- and NASH-model mice. Regarding the kinetics of TCA, diet-specific predictions were consistent with in vivo concentration-time course in plasma, liver and kidney. Notably, these predictions are at the high end of the population-wide predictions, indicating that TCA is eliminated more slowly in both NAFL- and NASH-model mice as compared to the “typical” mouse strain/diet.

Fig. 6.

PBPK model predictions of toxicokinetic profile (median and 95% CIs of chemical concentration) using both a population-generated “random strain” and strain/diet-specific parameter posterior distributions compared with in vivo data in male C57BL/6J-NAFL mice fed with a single dose of 300 mg/kg aqueous based oral gavage of perc (data from (Cichocki et al., 2017a)).

3.4. Posterior dose metric predictions

Fig. 8 shows PBPK model predictions across mouse strains/diets for the overall flux of perc metabolism given a single gavage dose of 300 mg/kg of perc, with summary statistics given in Supplemental Table S–5.

Fig. 8.

Density of uncertainty distributions of posterior strains-specific predictions of perc metabolism. All stain/diet variations used a single dose of 300 mg/kg oil (B6C3F1/J) and aqueous based oral gavage of perc. Amount metabolized were estimated at 36 hours after exposure.

With respect to oxidative metabolism, because posterior model predictions of metabolism were calibrated with the observed data on internal toxicokinetics of TCA for all strains and diets, distinct inter-strain and disease condition-associated variability with narrow uncertainty in oxidative metabolism was predicted across all strains/diets of mice in this model (Fig. 8, top row). The lowest clearance of perc was predicted to be through oxidative metabolism in B6C3F1/J mice, while it was predicted as highest in C57BL/6J mice fed NAFLD-inducing diets, with inter-strain variability of ~3-fold. Among the liver disease states, C57BL/6J mice fed normal diet were predicted to have less oxidative metabolism, and NASH group to have the most. The oxidative metabolism for SW mice was between them. Although a similar fraction of total TCA was produced (89% of total oxidative metabolism) across strains/diets, there was a slight difference in the production of TCA in liver, i.e., while almost all portions of TCA (99%) was expected to be produced in liver of B6C3F1/J mice, the kidney contributes 3 to 5% in the other strains following order of mice: C57BL/6J-NASH < C57BL/6J-NAFL < C57BL/6J-controls and SW mice.

With respect to GSH-conjugation, data were only available for different diets of C57BL/6J mice. Notable disease-associated variability appeared in GSH-conjugative metabolism (Fig. 8, bottom row), given the same exposure scenario described above. Total GSH-conjugation and TCVG production in kidney were 2-fold less in C57BL/6J mice fed NAFLD-induced diets than that in those fed normal diet. Due to the lack of the experimental data on toxicokinetics of GSH-conjugates in B6C3F1/J and SW mice, the posterior distributions of predicted GSH-conjugation were essentially based on the overall posterior population uncertainty and variability (i.e., “random strain”).

Area under the concentration-time curve (AUC) of perc and its metabolites in blood and various tissues were also estimated using strain/diet-specific posteriors, as shown in Supplemental Fig. S–6 and Supplemental Table S–6. AUCs for perc and TCA in all tissues were estimated at 36 hours and those of GSH-conjugates in all tissues at 24 h after exposure. From the AUCs, perc was highly distributed to fat, in which perc was 2 orders of magnitude greater than in blood, followed by that in liver and kidney, in which it was an order of magnitude greater. Although AUC of perc in blood was lowest in C57BL/6J-NASH group and highest in B6C3F1/J mice, the difference is negligible. In liver, AUC of perc in C57BL/6J-NASH mice is the highest and ~2-fold greater than that in SW mice, while that in C57BL/6J-NAFL mice are comparable with those in other strains. In kidney, AUC of perc in mice with NAFLD are lowest and 2.2- to 3-fold lower than that in other strains. Similar pattern appears in fat, i.e., AUC of perc in mice with NAFLD are ~3-fold lower than in other strains even though it is the site to which the greatest amount of perc was distributed.

AUCs of circulated free TCA in plasma and in liver are comparable at 36 h after exposure, although TCA in plasma is 2 orders of magnitude greater than perc in blood, while TCA in liver is an order of magnitude greater than perc in liver for all strains/diets. Regarding to variability across strains/diet, AUCs of TCA in blood is and liver the lowest for B6C3F1/J mice and the highest in C57BL/6J-NAFL mice with variability factor of >2-fold. In kidney, AUCs of TCA are lower compared to those in plasma and liver. Regarding to variability across strains/diet, in contrast to perc, AUC of TCA in kidney is the highest in diseased mice and the lowest in B6C3F1/J mice with variability factor of 5.5-fold, and those in SW and C57BL/6J-control mice are somewhere between them, indicating that TCA tends to be accumulated more in the kidney of diseased mice. In particular, 2 orders of magnitude greater TCA versus perc was accumulated in kidney of diseased mice 36 h after exposure.

AUCs of GSH-conjugates in C57BL/6J mice, in general, are 4 orders of magnitude lower than those of perc. AUCs of TCVG are highest in liver, which are 2 orders of magnitude greater than in blood, followed by those in kidney, indicating liver as a primary site for perc GSH-conjugation. Among diets, diseased mice have >2-fold lower AUCs in tissues compared to those in mice with normal-diet. AUCs of TCVC are highest in kidney, which are an order of magnitude greater than in liver, followed by those in blood, indicating kidney as a site for cysteine conjugation of TCVG. Disease-associated variability in AUC of TCVC is not apparent in blood and kidney, while C57BL/6J-NAFL mice had the highest and C57BL/6J-NASH mice had the lowest concentration of TCVC in the liver in 24 h after exposure, where AUC of TCVC was about half of that in C57BL/6J-NAFL mice. AUCs of NAcTCVC are highest in blood, which are 2- to 8-fold of those in kidney, followed by those in liver, indicating kidney as a site not only for the formation of a mercapturate, but also for active elimination of end products through either bioactivation and urinary excretion. Disease-associated variability in AUC of NAcTCVC is not apparent in blood, while C57/BL/6J-NASH mice had the lowest concentration of NAcTCVC in the liver and C57BL/6J-NAFL mice had the lowest concentration of NAcTCVC in the kidney.

4. Discussion

Recognizing the underlying causes for toxicokinetic and toxicodynamic variability is important in the risk assessment of chemicals because this information helps in quantitative characterization of individual differences in susceptibility. Physiological variance associated with genetic polymorphisms, sex (male vs. female), age (pediatric vs. young adults vs. elderly), health status (healthy vs. diseased), and even circadian rhythms may all underlie inter-person variability in response to xenobiotics (Zeise et al., 2013). Advanced population-based simulators for toxicokinetic parameters that include consideration for many of these factors are widely used in drug safety evaluations (Strougo et al., 2011; Jamei et al., 2013) and can be used as inputs in the PBPK approach (Haber et al., 2002; Hartmanshenn et al., 2016). Indeed, PBPK modeling is a promising method for estimating chemical exposure in specific health conditions and several models have been developed to predict the pharmacokinetics in patients with liver diseases through substantial extrapolation of pharmacokinetics from healthy individuals to patients (Edginton and Willmann, 2008; Johnson et al., 2010; Strougo et al., 2011; Li et al., 2015b; Jamwal et al., 2018). However, most of these focused on treatment-related drug disposition in patients under clinically severe disease conditions, i.e., liver cirrhosis, and there is a lack of PBPK model-based analyses characterizing the heterogeneity in toxicokinetics of environmental chemicals caused by underlying disease states.

In the past two decades, prevalence of chronic liver disease such as NAFLD has increased in the developed countries (Younossi et al., 2018). This has both raised concerns about how a chronic fatty liver condition may enhance the risk of chemical toxicity (Allard et al., 2019) and that the increase in NAFLD prevalence may be caused, in part, by chemical exposures (Heindel et al., 2017). It is without a doubt that NAFLD is a disease state that is known to contribute to the inter-individual variation in toxicokinetics and metabolic activation (biologically effective tissue dose) of chemicals (Cobbina and Akhlaghi, 2017). One recent example of such a link that pertains to a ubiquitous environmental contaminant are studies of perc (Cichocki et al., 2017).

Because of insufficient available human toxicokinetic data, modeling in mouse population serves as a surrogate to characterize and quantitate the extent of human variability (Rusyn et al., 2010; Chiu et al., 2014; Chiu and Rusyn, 2018). A murine model of the experimental NAFLD by Cichocki et al. provides evidence of the effects of underlying disease conditions on perc disposition and metabolism (Cichocki et al., 2017a; Cichocki et al., 2018). Indeed, physiological and biochemical changes due to NAFLD that were measured in vivo in C57BL/6J mice with NAFL and NASH (Cichocki et al., 2017a; Cichocki et al., 2018) provided empirical inputs into existing PBPK model. These data allowed us to describe an impact of NAFLD on the variation in toxicokinetics and metabolism of perc across strains and liver disease states. We found that after incorporating in vivo measurements of disease-induced alterations in body weights, tissue volumes (fat and liver), and blood:air and liver:blood PCs, we were able to adequately describe the variation in observed toxicokinetic profiles for perc and its metabolites across liver disease states. Specifically, we found that overall hepatic oxidative clearance of perc was much greater in NAFLD groups of C57BL/6J mice as compared to normal diet-fed mice both of the same strain as well as of different strains. Thus, the amount of perc metabolized to TCA was predicted to be highest in both NAFL and NASH mice, indicating that NAFLD is a susceptibility factor for the metabolic activation of perc by oxidation (Fig. 8). Whereas previously it was hypothesized that the accumulation of TCA observed in NAFLD mice was due primarily to reduced hepatic clearance (Cichocki et al., 2017a; Cichocki et al., 2018), use of the PBPK model developed in this study revealed that this could also be explained by increased production of TCA. Perc is metabolized by cytochrome P450 enzymes in both animals and humans, and believed to be primarily, by the 2E1 subclass of these enzymes, based on data for similar compounds (Lash and Parker, 2001). The enzymatic activity of CYP2E1 has been reported to be increased not only in mice fed a high-fat diet (Begriche et al., 2008; Abdelmegeed et al., 2012), but also in biopsy-verified human NASH patients in vivo using chlorzoxazone (Chalasani et al., 2003; Orellana et al., 2006; Fisher et al., 2009), with some other CYPs also exhibiting changes (CYP2A6 and CYP2C9 increased, CYP1A2, CYP2C19, and CYP2D6 decreased, CYP3A4 unchanged) (Fisher et al., 2009). These data are consistent with the hypothesis of increased production of TCA due to increased activity of some CYPs, particularly 2E1, in patients with NASH. Interestingly, the murine model of NAFLD also showed upregulation of the organic transporter genes (e.g., Slco1a1) after exposed to perc, which may, in turn, reabsorb TCA in this group (from unpublished data used to respond to reviewer’s comments in (Cichocki et al., 2018), see Supplemental Fig. S–7). Therefore, future work is warranted to explore variability in TCA clearance across strains and diets.

Data from mouse studies show that GSH conjugation only contributes approximately 0.3% of the total metabolism of perc (Luo et al., 2018); however, the role of these conjugative metabolites in perc-associated nephrotoxicity, through formation of reactive sulfoxides, is well-recognized (Cichocki et al., 2016). However, little is known about variation in GSH conjugation. Interestingly, the total amount of perc metabolized through GSH conjugation was 2-fold lower in C57BL/6J mice with NAFLD diets as compared to controls. These results corroborate the qualitative finding previously reported suggesting the potential protective factor of NAFLD for perc-associated kidney effects due to the reduced production and delivery of nephrotoxic glutathione conjugation-derived metabolites by (Cichocki et al., 2018). Our prediction of internal dosimetry shows that TCVG is most abundant in liver, whereas TCVC is most abundant in kidney across all conditions (Supplemental Fig. S–6), consistent with the primary formation of TCVG in the liver and its subsequent conversion to TCVC in the kidney (Cichocki et al., 2017a; Cichocki et al., 2018).

Parent compound toxicokinetics are more challenging to predict in the presence of NAFLD. Our model predictions of perc toxicokinetics underestimated the observed concentration-time profiles in blood and liver in NAFL and NASH mice, leading to the reduced delivered dose of perc in blood in both conditions, and in liver of NAFL mice. This may be associated with other physiological or toxicodynamic changes for which we have not accounted. In the experiment, the phenotypes of NAFLD disease included simple liver steatosis (NAFL) or steatohepatitis (NASH), with the apparent diffuse micro- and macro-vesicular steatosis in liver tissue and increased serum liver injury markers and liver triglycerides (Cichocki et al., 2017a). Other physiological abnormalities exist in the presence of NAFLD, including possible upregulated absorption of fatty acids in the gut, decrease in hepatic perfusion, portal vein and renal blood flows, increase in cardiac index and blood flows of other organs (Erdogmus et al., 2008; Farrell et al., 2008; Shigefuku et al., 2012; Anavi et al., 2017; Marsousi et al., 2017; Utsunomiya et al., 2017). Substantial changes in blood flow were demonstrated in human and animal models of fatty liver (Farrell et al., 2008; Shigefuku et al., 2012; Beers, 2014; Cocciolillo et al., 2014). In relation to the hepatic vascular network, decreased portal vein flow was found in patients with fatty liver (Erdogmus et al., 2008; Guiu et al., 2012). Accordingly, the cardiac output was increased in patients with cirrhosis (Marsousi et al., 2017). We did not incorporate these because they were not measured experimentally, and did not include them in the Bayesian analysis because of computational limitations in terms of length of time for convergence, but future PBPK modeling that accounts for those physiological abnormalities may help to fill the gap in predicting the perc disposition in blood and liver in individuals with NAFLD.

We conclude that the PBPK model developed in this study is a valuable addition to the characterization of toxicokinetics of perc in individuals with chronic liver disease. This model successfully incorporates the effects of pre-existing disease conditions on toxicokinetic variability of perc metabolites through combining the model for mice with normal livers with physiological and biochemical changes caused by the underlying NAFLD. Based on PBPK model predictions of the delivered dose in target organs, we posit that, given the same exposure scenario, individuals with NAFLD will have 1) elevated liver levels of TCA, a hepatotoxic metabolite of perc; and 2) lower kidney levels of the nephrotoxic GSH conjugation metabolites compared to healthy individuals. Although genetic variability in this study was limited to three strains of mice, the results nonetheless suggest that non-genetic factors such as diet and pre-existing disease conditions may be at least as influential as genetic factors in altering chemical toxicokinetics, and thus are likely to be a substantial contributor to population variation in susceptibility to chemical exposures. Therefore, more fully characterizing inter-individual toxicokinetic variation for use in health risk assessments of environmental toxicants requires incorporation not only of genetics, but also of disease states as an additional, health-related susceptibility factor. Overall, our population PBPK model will help health risk assessments of perc to more accurately predict human risk by integrating the effects of pre-existing NAFLD on inter-individual variation in biologically effective tissue dose, and serves as a proof-of principle for quantitatively addressing susceptibility due to pre-existing disease in risk assessment.

Highlights.

• We investigated the impact of NAFLD on toxicokinetic variability of perc in mice using a PBPK model.

• Disease-induced physiological and biochemical changes are incorporated into the model.

• Toxicokinetic variability across strains and disease-states of mice is evident.

• Diseased mice exhibit higher oxidative and lower conjugative metabolism of perc compared to healthy mice.

• Pre-existing diseases are likely contributors to susceptibility to chemical exposures.