Physiologically Based Pharmacokinetic Models for Prediction of Complex CYP2C8 and OATP1B1 (SLCO1B1) Drug–Drug–Gene Interactions: A Modeling Network of Gemfibrozil, Repaglinide, Pioglitazone, Rifampicin, Clarithromycin and Itraconazole

PBPK models were developed using PK-Sim^® and MoBi^® modeling software (version 7.3.0, part of the OSP suite). Parameter optimizations (Monte-Carlo-algorithm) and sensitivity analyses were performed with PK-Sim^®. Clinical study data was digitized using GetData Graph Digitizer (version 2.26.0.20 ©, S. Fedorov). Pharmacokinetic parameter analyses were performed and plots were created with R (version 3.4.4, R Foundation for Statistical Computing, Vienna, Austria) and R Studio (version 1.1.423, RStudio, Inc., Boston, MA, USA).

PBPK model building was started with an extensive literature search to gain information on the physicochemical properties and absorption, distribution, metabolism, and excretion (ADME) processes of the drugs of interest as well as to obtain clinical studies (healthy individuals) of intravenous and oral administration in single and multiple doses. The plasma concentration–time profiles of the clinical studies were digitized and divided into an internal dataset used for model building and parameter optimization, and an external dataset used for model evaluation. Parameters that could not be informed from the literature were optimized, fitting the model simultaneously to the observed data of all studies assigned to the internal dataset.

The mean individuals used to simulate the different studies were modeled according to the respective study reports, with corresponding age, weight, height, sex, and ethnicity. If no information on these demographics was found, a 30-year-old male European was assumed, with the mean weight and height characteristics given in the PK-Sim^® database. The creation of virtual populations to compare predicted and observed population plasma concentration–time profiles is described in the ESM (Sect. 3.9).

Protein expression of enzymes and transporters was implemented according to the literature, using the PK-Sim^® database [10]. For more details, see ESM Table S3.9.1.

The PBPK models were evaluated by comparison of predicted population plasma concentration–time profiles to observed data. The observed data were mostly reported as arithmetic or geometric mean plasma concentration–time profiles with standard deviations. To compare the variability of predicted to observed profiles, 68% population prediction intervals were plotted, as this interval corresponds to the range of ± 1 standard deviation around the mean if normal distribution is assumed.

Furthermore, the predicted plasma concentration values were compared to their respective observed values in goodness-of-fit plots and the model performance was evaluated by comparison of predicted to observed area under the plasma concentration–time curve (AUC), maximum plasma concentration (C_max), apparent oral clearance (CL/F), and half-life values. For model evaluation, all AUC values were calculated from time zero to infinity (AUC_∞).

As quantitative measures of the model performance, mean relative deviations (MRDs) of the predicted plasma concentrations (see ESM Eq. 3.1) and geometric mean fold errors (GMFEs) of the AUC_∞, C_max, CL/F, and half-life values (see ESM Eq. 3.2) were calculated.

In addition to the evaluation methods described in Sect. 2.3, a CYP2C8 DDI network was built to evaluate the DDI performance of the developed models (Fig. 1). Gemfibrozil with its metabolite gemfibrozil 1-O-β-glucuronide was used as the CYP2C8 and OATP1B1 inhibitor, repaglinide as the CYP2C8 and OATP1B1 victim, and pioglitazone as the CYP2C8 victim drug. In addition, repaglinide is also a substrate of CYP3A4 and OATP1B3, adding DDI potential during coadministration of CYP3A4 and OATP1B3 perpetrators, such as itraconazole, clarithromycin, and rifampicin. Rifampicin was used as the inducer and competitive inhibitor of CYP2C8, CYP3A4, OATP1B1, and OATP1B3, interacting with repaglinide and pioglitazone. Interaction processes are modeled using the equations given in the ESM (Sect. 1).

The quality of the DDI modeling was evaluated by comparison of predicted to observed victim drug plasma concentration–time profiles when administered alone and during coadministration, DDI AUC ratios (Eq. 1) and DDI C_max ratios (Eq. 2). For DDI evaluation, all AUC values were calculated from time zero to the time of the last concentration measurement (AUC_last).

To assess the DDI modeling performance, the GMFEs of the predicted DDI AUC ratios and DDI C_max ratios were calculated according to ESM Eq. 3.2.

Sensitivity analyses were performed on the gemfibrozil, repaglinide, and pioglitazone models to investigate the impact of single model parameters on the predicted AUC at steady state given the highest recommended dose. Parameters were included into the analysis if they have been optimized, if they could have a strong influence on the pharmacokinetics due to their use in calculation of permeabilities or partition coefficients (e.g., fraction unbound), or if they had significant impact in formerly developed models (e.g., blood/plasma concentration ratio).

Sensitivity is calculated as the ratio of the relative change of the simulated AUC to the relative variation of the tested parameter around the parameter value used in the model, according to Eq. 3:where S is the sensitivity of the AUC to the tested model parameter, ΔAUC is the change of the AUC, AUC is the AUC calculated with the original model parameter value, p is the original model parameter value, and Δp is the change of the tested model parameter value.

Sensitivity analyses were performed using a relative perturbation of 100%. The threshold value for sensitivity was set at 0.5. A sensitivity value of + 0.5 indicates that a 100% increase of the model parameter value causes a 50% increase of the predicted AUC. In addition, the parameters were varied within a 0.03- to 30-fold range and the resulting fold changes of AUC were investigated in spider plots.

A detailed description of each model, including Tables listing the drug-dependent parameters, is given in ESM Sects. 3.3 (gemfibrozil and gemfibrozil 1-O-β-glucuronide model), 3.4 (repaglinide model), and 3.5 (pioglitazone model). System-dependent parameters were taken directly from the PK-Sim^® database or, if not available, they were gathered from literature, as summarized in ESM Table S3.9.1. The good model performance for both internal and external datasets is demonstrated by comparison of population predicted to observed plasma concentration–time profiles in Fig. 2 (representative studies for each compound) and in ESM Figs. S3.3.1, S3.3.2, S3.3.3, S3.3.4, S3.4.1, S3.4.2, S3.4.3, S3.4.4, S3.5.1, and S3.5.2 (all studies, semilogarithmic and linear plots). Goodness-of-fit plots are presented in ESM Figs. S3.3.5, S3.4.5, and S3.5.3 and MRD values for all studies are given in ESM Tables S3.3.3, S3.4.3, and S3.5.3. Correlation of predicted to observed AUC and C_max values is presented in ESM Figs. S3.3.6, S3.4.6, and S3.5.4, and the corresponding values are given in ESM Tables S3.3.4, S3.4.4, and S3.5.4, including calculated model GMFE values. Correlation of predicted to observed CL/F and half-life values is shown in ESM Figs. S3.3.7, S3.4.7, and S3.5.5 and the corresponding values are given in ESM Tables S3.3.5, S3.4.5, and S3.5.5. The results of sensitivity analyses are described in ESM Sects. 3.3, 3.4, and 3.5 and Figs. S3.3.8, S3.3.9, S3.3.10, S3.4.8, S3.4.9, S3.5.6, and S3.5.7.

For the development of the gemfibrozil parent–metabolite PBPK model, plasma concentration–time profiles of 23 studies (oral administration), including ten studies showing gemfibrozil 1-O-β-glucuronide concentrations (after oral administration of gemfibrozil) were used. Gemfibrozil 1-O-β-glucuronide was incorporated into the model because it is the major circulating metabolite of gemfibrozil and a mechanism-based inactivator of CYP2C8 [11], contributing significantly to the strong CYP2C8 inhibition by gemfibrozil. The gemfibrozil model applies an unspecified active uptake of gemfibrozil into hepatocytes [12], metabolism by uridine 5ʹ-diphospho-glucuronosyltransferase (UGT) 2B7 to form gemfibrozil 1-O-β-glucuronide and glomerular filtration. The metabolite model applies a hepatic uptake transport by OATP1B1 [12], an efflux transport into bile via multidrug resistance-associated protein (MRP) 2 and glomerular filtration.

For the development of the repaglinide model, plasma concentration–time profiles of 56 studies (intravenous and oral administration) as well as the fraction metabolized via CYP2C8 information were used. The model applies hepatic uptake via OATP1B1 and OATP1B3, metabolism by CYP2C8 and CYP3A4, and glomerular filtration. An unbound Michaelis-Menten contant (K_M) value determined for repaglinide uptake into untransfected primary human hepatocytes was used as the K_M value for the OATP1B1 and OATP1B3 transport processes [13]. For the metabolism via CYP2C8, a fraction metabolized of 89% was assumed, according to a clinical DDI study with gemfibrozil [14]. The remaining repaglinide is metabolized via CYP3A4 [15]. The model slightly overpredicts the low repaglinide plasma concentrations at later times after dosing (ESM Figs. S3.4.1, S3.4.2, and S3.4.5). This could be due to a misspecification of the CYP2C8 or OATP1B1/1B3 K_M values taken from the literature, or to an unknown mechanism that impacts the pharmacokinetics of repaglinide but is missing in the model. This limitation of the model affects the prediction of low repaglinide plasma concentrations, but C_max and AUC values are well-predicted (ESM Fig S3.4.6). For the investigation of DGIs, the model adequately describes the pharmacokinetics of repaglinide in carriers of the CYP2C8*3, SLCO1B1 521C and SLCO1B1 –11187A alleles. The implementation of polymorphic CYP2C8 and OATP1B1 is described in ESM Sects. 2.2, 2.3, and 3.4.

For the development of the pioglitazone model, plasma concentration–time profiles of 13 studies (oral administration), one study describing the fraction of pioglitazone excreted to urine, as well as the fraction metabolized via CYP2C8 information were used. Pioglitazone is reported to be predominantly metabolized by CYP2C8 [16], with no consistent information on the identity of other metabolic enzymes involved. Therefore, the model applies metabolism by CYP2C8, an unspecific hepatic clearance and glomerular filtration. For the metabolism via CYP2C8, a fraction metabolized of 70–75% was assumed [17]. For the investigation of DGIs, the model adequately describes the pharmacokinetics of pioglitazone in carriers of the CYP2C8*3 allele. The implementation of polymorphic CYP2C8 is described in ESM Sects. 2.2 and 3.5.

For the DDI network modeling, a total number of 34 DDI studies were available in literature and used to evaluate the modeled interactions (Fig. 1). Thereof, 23 studies describe the gemfibrozil–repaglinide DDI, four the gemfibrozil–pioglitazone DDI, and one study each the itraconazole–repaglinide DDI, the itraconazole–pioglitazone DDI, the gemfibrozil–itraconazole–repaglinide DDI, the gemfibrozil–itraconazole–pioglitazone DDI, the rifampicin–repaglinide DDI, the rifampicin–pioglitazone DDI, and the clarithromycin–repaglinide DDI. The previously developed models of itraconazole, rifampicin, and clarithromycin [9] were used without changes other than the addition of interaction parameters to model the CYP2C8 and OATP1B1/1B3 DDIs. A full description of the DDI modeling is given in ESM Sect. 4, including all interaction parameters, administration protocols, and study population demographics.

All DDI victim drug plasma concentration–time profiles are well-predicted using interaction parameters taken from the literature (listed in ESM Tables S3.3.2, S3.6.1, S3.7.1, and S3.8.1 summarizing the drug-dependent parameters of the perpetrator drugs), except for the hydroxy-itraconazole OATP1B1 and OATP1B3 inhibitory constant (K_i) values, which had to be optimized. To describe the DDIs with itraconazole, the solubility value of the previously developed itraconazole model was adjusted (see ESM Figs. S3.6.1 and S3.6.2). Itraconazole is a poorly soluble compound (8.0 mg/L in fasted state simulated intestinal fluid [18]), leading to variable absorption and, therefore, to large inter-individual differences in itraconazole plasma concentration–time profiles. In the gemfibrozil–itraconazole–pioglitazone interaction study [19], itraconazole concentration–time profiles before and during gemfibrozil coadministration are reported, in addition to the DDI victim drug concentrations. For itraconazole and hydroxy-itraconazole, a reduction of the AUCs is described if gemfibrozil is added to the itraconazole–pioglitazone DDI. Possible explanations for these finding proposed by Jaakkola et al. [19] and Niemi et al. [20] are the displacement of itraconazole from plasma proteins or a reduction of itraconazole bioavailability by gemfibrozil. Both hypotheses were tested with the models by either increasing the fraction unbound of itraconazole during the DDI or by reducing its solubility (to describe reduced bioavailability). The observed plasma concentration–time profiles of itraconazole and hydroxy-itraconazole could not be described properly by a change in fraction unbound, but the change of itraconazole solubility resulted in a good description of the observed data. To accurately model the observed itraconazole plasma concentrations of the gemfibrozil–itraconazole–pioglitazone interaction study [19], the solubility of itraconazole (capsule formulation, fasted state) was optimized to 14.5 mg/L at pH 6.5 in the absence of gemfibrozil and to 0.69 mg/L at pH 6.5 in the presence of gemfibrozil (shown in ESM Sect. 3.6), indicating a physicochemical DDI between gemfibrozil and itraconazole.

The second assumed physicochemical DDI affects pioglitazone during coadministration of both gemfibrozil and itraconazole (see Figs. 3c and ESM Fig. S4.7.1). Like itraconazole, pioglitazone is poorly soluble, with a reported solubility of 16.8 mg/L at pH 6.5 [21]. To accurately describe the pioglitazone plasma concentrations during gemfibrozil–itraconazole–pioglitazone coadministration, pioglitazone solubility was adjusted to 1.59 mg/L at pH 6.5 (shown in ESM Sect. 4.7).

The good DDI performance of the models is demonstrated in predicted compared to observed victim drug plasma concentration–time profiles before and during perpetrator drug coadministration in Figs. 3 and 4 (representative studies for each DDI) and in ESM Figs. S4.2.1, S4.2.2, S4.2.3, S4.2.4, S4.2.5, S4.2.6, S4.3.1, S4.3.2, S4.4.1, S4.5.1, S4.6.1, S4.7.1, S4.8.1, S4.9.1, and S4.10.1 (all studies, semilogarithmic and linear plots). For the gemfibrozil–repaglinide DDI, the time-dependency (Fig. 4 and ESM Figs. S4.2.3 and S4.2.4) and dose-dependency (ESM Figs. S4.2.5 and S4.2.6) of the interaction were also modeled and compared to clinical data.

As further evaluation of the performance of the entire DDI network, the correlation of predicted to observed DDI AUC ratios and DDI C_max ratios of all modeled DDI studies is shown in Fig. 5. The corresponding values are listed in ESM Tables S4.2.2, S4.3.2, S4.4.2, S4.5.2, S4.6.2, S4.7.2, S4.8.2, S4.9.2, and S4.10.2, including GMFE values calculated for each perpetrator–victim drug combination.

To show the utility of the models to individualize and improve drug therapy, dose adaptations for different DGI scenarios were calculated. Exemplarily, victim drug plasma concentration–time profiles in subjects with polymorphisms, and with polymorphisms during perpetrator drug coadministration, were simulated and compared to those of a CYP2C8 and SLCO1B1 wild-type individual without DDI (control). Then, potential dose adaptations were calculated, aiming to match control exposure (i.e., AUC in steady state). Figure 6 shows the different simulated scenarios, the extrapolated iso-exposure doses (% of control), simulations of plasma concentration–time profiles with the same dose for all individuals, and simulations of plasma concentration–time profiles with the extrapolated adjusted doses.

As expected, the worst-case scenario is the repaglinide administration to a patient carrying the CYP2C8 wild-type sequence and simultaneously homozygously the SLCO1B1 521C allele who is co-medicated with gemfibrozil and itraconazole, leading to a 49-fold increase in the repaglinide steady state AUC. According to the simulations, a repaglinide dose reduction by nearly 98% would be necessary to produce the same drug exposure in this individual as in the control person given the normal dose. A dose adjustment to simultaneously match both AUC and C_max values to the control profiles was not possible, as the drug half-life is changed due to the polymorphisms and the DDIs. The safety and efficacy of increased or decreased C_max values or the drug half-life following dose adaptations that were calculated to match the AUC in steady state cannot be predicted from the models.