FormalPara Key Points

Pitavastatin and atorvastatin PBPK models were successfully developed by considering the contribution of both metabolizing enzymes and transporters to their overall pharmacokinetics. The statin PBPK models reasonably predicted the observed exposure change due to OATP1B1 polymorphism or clinical DDIs with Cytochrome P450 (CYP) inhibitors or OATP1B1 inhibitors.

Further analysis demonstrated that OATP1B1 inhibition by rifampin or cyclosporine in the existing inhibitor models needs to be approximately tenfold stronger to recapitulate the observed DDI with these two inhibitors.

K i values of inhibitor drugs on OATP1B1 need to be carefully evaluated with known substrates for a comprehensive DDI risk assessment to avoid possible under-prediction of clinical DDI.

1 Introduction

Statins are a class of drugs prescribed widely among 20 million Americans to treat the hypercholesterolemia [1]. Currently marketed statins in the US include simvastatin, lovastatin, atorvastatin, fluvastatin, pravastatin, rosuvastatin, and pitavastatin. Although statins are generally well-tolerated, patients concomitantly receiving multiple medications are at an increased risk of adverse events including myalgia, myopathy and rhabdomyolysis, due to increased exposure of statins causing by drug interactions in the presence of co-medications [2]. For example, in a cross-sectional study of 2742 dyslipidemia patients, a total of 198 statin-related drug-drug interactions (DDIs) were identified (7.2%) among those aged 54–75 years who were receiving a mean of 3.8 and 5.8 concomitant medicines [3]. Thus, it is crucial to consider DDI risks when co-administering other drugs in patients receiving statin therapy. Mechanisms that cause statin-related DDIs should be understood for risk management.

Although statins have similar mechanism of action, their disposition pathways differ [4]. Lipophilic statins, including simvastatin, lovastatin, atorvastatin, and fluvastatin, are extensively metabolized by CYP3A4 or CYP2C9. Hydrophilic statins, including pitavastatin, pravastatin and rosuvastatin, are mainly eliminated unchanged with minimal metabolism by CYP enzymes [5]. Drug transporters in the liver, intestine, and kidneys can also influence the disposition of statins. All statin drugs are substrates of the membrane transporter organic anion transporting polypeptide 1B1 (OATP1B1) (encoded by SLCO1B1) [6–9]. Pharmacogenetic studies in statin taking patients with different SLCO1B1 polymorphisms have revealed that systemic exposure of most statins were significantly increased in patients with genotype SLCO1B1 c.521 CC, a polymorphism with reduced OATP1B1 function [10]. The rank order of mean systemic exposure change (in fold) of various statins in subjects with variant alleles as compared to wild type are as follows: simvastatin (4.8) > pitavastatin (3.3) > atorvastatin (3.2) > pravastatin (2.0) > rosuvastatin (1.6) > fluvastatin (1.1), indicating the importance of OATP1B1 in the disposition of some statins [11, 12]. Because different transporters and CYP enzymes contribute to the disposition of different statins, the exposure change of each statin by the same perpetrator (inhibitor) drug can be quite different. As such, quantitative prediction of statin DDI is complicated.

Elsby et al. developed a static model to predict the AUC change of statins in the presence of transporter or enzyme inhibitors [13]. By adopting the extended clearance concept [14], we developed mechanistic static models for all 7 U.S.-marketed statins to predict DDIs in the presence of enzyme and/or transporter inhibitors (unpublished) [15]. Our analysis showed that 60% (33/56 DDI cases) of predicted AUC ratios of statin AUC in the presence and absence of inhibitors (AUCRs) were within 80–125% of the observed AUCRs. While static models are relatively straightforward to apply, they are not able to consider the dynamic concentrations of victim  (substrate) drugs and inhibitors, especially concentrations in different tissues with varying expression levels of transporters or enzymes. These limitations hinder quantitative understanding and prediction of DDIs involving multiple transporters or complex enzyme-transporter interplay [16].

Physiologically based pharmacokinetic (PBPK) models integrate human physiologically-dependent system parameters with drug-dependent parameters to predict drug pharmacokinetics in the presence of intrinsic or extrinsic factors [17]. The use of PBPK to assess DDIs mediated by drug transporters is emerging, especially in studying the role of OATPs in the disposition of statin drugs [18–20]. For example, the development of PBPK models for rosuvastatin, pravastatin, and simvastatin have been reported previously [19, 21, 22].

To establish a framework to evaluate statin-associated risk, we developed PBPK models for pitavastatin and atorvastatin using in silico, in vitro and human pharmacokinetic data. These models considered quantitative contribution from both enzymes and transporters to the disposition of these drugs. We then used these models to simulate statin pharmacokinetic in subjects with varying SLCO1B1 polymorphism or in subjects co-administered with various CYP enzymes and/or transporter inhibitors. The results (1) quantified key factors that determined the disposition of these two statins and confirmed the significant role of OATP1B1, and (2) demonstrated the utility of PBPK in further understanding of knowledge gaps of predicting transporter-based DDIs.

2 Methods

PBPK models were developed using a population-based PBPK software, SimCYP® (SimCYP Ltd, a Certara company, Sheffield, UK, Version 14.1). Unless otherwise stated, simulations with virtual population used 5 trials with 10 healthy subjects for each trial (50 subjects in total) in software’s built-in healthy volunteer virtual population. For both statins, software’s Advanced Dissolution, Absorption, and Metabolism (ADAM) model and permeability-limited liver distribution model were used to parameterize drug absorption and liver disposition, respectively. Data from in silico, in vitro absorption, distribution, metabolism and excretion (ADME), and observed pharmacokinetic data in healthy subjects with wild type OATP1B1 (encoded by SLCO1B1) were used to develop the models (see Sects. 2.1 and 2.2). To simulate the effect of SLCO1B1 genetic polymorphism, sub-virtual populations were created according to OATP1B1 abundance relative to wild type transporter, named as extensive transporter (ET, see 2.3). To simulate the effect of inhibitors on statin pharmacokinetic profile, inhibitor models from software’s drug library or those published in the literature were directly used (See 2.4).

2.1 Pitavastatin Model Development

Pitavastatin drug-dependent parameters are listed in Table 1 and the section focuses on transporter parameterization. In addition to in silico and in vitro ADME data, the observed mean pharmacokinetic data of pitavastatin in healthy subjects receiving 2 mg pitavastatin via intravenous infusion or 2 mg oral dose (training dataset) [23] were used to optimize several parameters. Hepatic metabolism of pitavastatin is assumed to be rate-limited by hepatic uptake transporters. Direct application of intrinsic clearance (CLint,T) values for OATP1B1 and OATP1B3 determined in vitro (58.4 and 5.1 µL/min/million cells, respectively) [24] under-estimated systemic clearance of the drug. As such, empirical scaling factors for OATP1B1 and OATP1B3 were optimized to be 18 and 18, respectively, according to intravenous and oral pharmacokinetic data mentioned above [23].

Table 1 Drug-dependent parameters for pitavastatin and atorvastatin PBPK models
Full size table

The final pitavastatin PBPK model was verified by simulating the single oral dose of 4 mg pitavastatin [23], the exposure changes of pitavastatin in different genotypes of SLCO1B1, or in the presence of OATP1B1 inhibitors as described below. The simulated results were compared to the clinical observations to determine the model performance.

2.2 Atorvastatin Model Development

Atorvastatin drug-dependent parameters are listed in Table 1 and the section focuses on transporter parameterization. Because there was no intravenous plasma concentration-time profile data available for atorvastatin, oral pharmacokinetic profiles in healthy subjects [25, 26] were used during model development to optimize intrinsic clearance of hepatic uptake transporter. Atorvastatin is extensively metabolized by CYP3A4 and actively uptaken into hepatocytes through OATP1B1 [27]. Because the uptake process may be the rate-limiting step for the overall clearance of atorvastatin, intrinsic clearance of CYP3A (CLint,CYP3A) may be unidentifiable. We tried two approaches to determine CLint,T and CLint , CYP3A. First, we obtained CLint , CYP3A by retrograde analysis with reported value of intravenous atorvastatin clearance of 37.5 L/h [28]. The derived CLint,CYP3A was 8 µl/min/pmol of recombinant CYP according to the design of the software. (In SimCYP, the unit of CLint,CYP3A is µl/min/pmol if obtained from the retrograde method.) Then, we optimized CLint , T for OATP1B1. The optimization began with reported in vivo CLint,T of 910 ml/min/kg [29, 30], or 360 µl/min/million cells based on software’s extrapolation algorithm. The CLint,T can be further used to calculate the maximum transporter-mediated uptake (J max) of atorvastatin (277.2 pmol/min/million cells) based on a Michaelis-Menten Constant (K m) of 0.77 µM (assuming the lowest atorvastatin K m value from the reported range of 0.77–15 µM, the University of Washington Metabolism and Transport Drug Interaction Database, UWDiDB). With the input of these parameters, we simulated the pharmacokinetic of atorvastatin after 20 mg single oral administration (training dataset). The initial simulation of atorvastatin over-estimated the AUC, thus an empirical scaling factor of 4 for OATP1B1, obtained by sensitivity analysis using the same training set pharmacokinetic dataset (e.g. plasma concentration-time profile, AUC and C max) after 20 mg oral dosing, was decided [25, 26] (Supplementary Fig. 1). In the second approach, we directly used CLint,CYP3A based on the in vitro assays with human liver microsomes (184 µL/min/mg protein for ortho-hydro metabolite pathway, and 219 µL/min/mg protein for para-hydroxy metabolite pathway, respectively), and then optimized CLint,T as described in the first approach. (In SimCYP, the unit of CLint, CYP3A is µl/min/mg protein if the values are obtained from in vitro microsome assay. [31]) This modeling approach resulted in similar simulation of atorvastatin pharmacokinetic after 20 mg oral dosing as Approach 1, suggesting that CLint,CYP3A is unidentifiable and hepatic uptake seems to be the rate-limiting step of the overall clearance of atorvastatin.

Atorvastatin is a substrate of Breast Cancer Resistance Protein (BCRP) encoded by ABCG2, and Keskitalo et al [32] found that subjects with the ABCG2 c.421AA genotype showed a 72% and 46% increase in atorvastatin AUC and C max, respectively, as compared to subjects with the c.421CC genotype, suggesting the important role of BCRP in the absorption/disposition of atorvastatin. However, due to the limited data, BCRP and OATP transport CLT could not be simultaneously fitted with certainty, particularly using oral data alone. Because the t 1/2 of the atorvastatin was not changed in the ABCG2 genotype, we assume that BCRP mainly plays a role in the absorption of atorvastatin. The intrinsic clearance of atorvastatin via intestinal BCRP was optimized manually to match the simulated C max and time to C max (T max) to those reported in the literature [33]. The contribution of biliary clearance of atorvastatin from BCRP was directly adopted from the in vitro hepatocyte uptake studies (1.4 µl/min/mg protein, assuming 1 million cell = 1 mg protein) [34], which appears to contribute minimally to overall hepatic clearance.

The final atorvastatin PBPK model was verified by simulating the single oral dose of 40 mg [35], the exposure changes of atorvastatin in different genotypes of SLCO1B1, or in the presence of CYP3A and OATP1B1 inhibitors as described below. The simulated results were compared to the clinical observations to determine the model performance.

2.3 Simulation of Exposure Change of Statins in Populations with Different Genotypes of SLCO1B1

In SimCYP, the CLint.T of transporter is scaled to whole organ clearance by multiplying a series of scaling factors including milligram microsomal protein per gram of liver (MPPGL), transporter abundance (abundance of various transporter allelic variants), and total liver weight. Different SLCO1B1 genotypes include ET (extensive transporter), IT (intermediate transporter), PT (poor transporter), and UT (ultra-rapid transporter). For our studies described in this manuscript, all population simulations were conducted by using SimCYP-defined virtual healthy population, unless otherwise specified (e.g., simulations in populations with SLCO1B1 c.521 TT or c.521 CC). For a virtual healthy population, the percentage of each transporter phenotype in a population would be determined by the frequency values of transporter phenotype under transporter phenotype table in SimCYP virtual population, which are 0.64, 0.02, 0.27, and 0.07 for ET, PT, IT, and UT, respectively. The default relative abundance values for each SLCO1B1 genotype are 1, 0.68, 0.37, and 1.47 for ET, IT, PT, and UT, respectively. This relative abundance determines OATP1B1 intrinsic clearance in non-ET subjects as a fraction of that in ET in simulations of the pharmacokinetic of statins in these subjects.

We assumed that population with SLCO1B1 c.521 TT was consisted of subjects with 100% of ET, and population with SLCO1B1 genotype c.521 CC was consisted of subjects with 100% of PT. Two populations were created separately for SLCO1B1 genotype c.521 TT population and SLCO1B1 genotype c.521 CC population by adjusting the population frequency of ET or PT to 1 while keeping all other values as 0, respectively. The plasma concentration-time profile of pitavastatin or atorvastatin in SLCO1B1 genotype c.521 TT or in SLCO1B1 genotype c.521 CC was simulated in each respective population (5 trials with 10 subjects in each trial).

2.4 Inhibitor Drug Models

Models for itraconazole and its metabolite hydroxyl-itraconazole were directly used from SimCYP software compound library. Erythromycin PBPK model parameters were adopted from the model published by Yeo et al. [36]. Gemfibrozil PBPK inhibitor model published by Varma et al. [37] (with the consideration of its metabolite gemfibrozil 1-O-β-glucuronide) was directly applied as the inhibitor model. Cyclosporine model from Jamei, et al [21] and rifampin model from Varma et al [38] were directly applied (also see Sect. 2.6). The inhibitory kinetic parameters on CYP enzymes and transporters for these inhibitors are shown in Table 2.

Table 2 Inhibitor PBPK models
Full size table

2.5 Measurement of Prediction Performance

Both atorvastatin and pitavastatin have significant inter-study variability in pharmacokinetics [39]. As such, we used AUC ratios (with/without co-medication or between different genotypes) as metrics to evaluate model performance. Effect of SLCO1B1 polymorphism was measured by statin AUC ratio (AUCR), the ratio between mean statin AUC in PT subjects to that in ET subjects. The magnitude of drug interaction was represented by AUCR (mean AUC of statin in the presence of an inhibitor drug compared to the mean AUC in the absence of an inhibitor drug). The prediction performance of PBPK models was determined by an R value (R is the ratio of mean simulated AUCR to mean observed AUCR, R = AUCRs/AUCRo). There is no consensus with regard to acceptance criteria of a PBPK model [40]. A criterion that predicted pharmacokinetic parameters within twofold of the observed data has been used frequently to assess the performance of a PBPK model [40]. Although this twofold criterion is arbitrary, for the purpose of our analysis and known inter-study variability, an R value within a twofold range (0.5–2.0) was regarded satisfactorily. Furthermore, when observed plasma concentration-time profiles were available, the models were considered acceptable if the clinical observations were within 90% confidence interval (CI) of the simulated mean plasma concentration curve, an approach used in other PBPK studies [41, 42]. The similar analysis on change in C max as a result of DDI or different genotypes of SLCO1B1 was conducted and the results were shown in Supplementary Table 1.

2.6 Evaluation of the Impact of Inhibitor K i for OATP1B1 on DDI Prediction

For DDI prediction with rifampin and cyclosporine, a sensitivity analysis on their K i for OATP1B1 was conducted by changing the K i value to up to 100 fold-lower used in the published models. SimCYP’s built-in sensitivity analysis function was used to conduct the sensitivity analysis.

3 Results

3.1 Performance of Pitavastatin and Atorvastatin Models for Predicting Pharmacokinetic Profiles via Intravenous and Oral Routes of Administration

Figure 1 shows that the pitavastatin substrate model described pitavastatin pharmacokinetic profile after single intravenous infusion of 2 mg (Fig. 1a), which was used during model development to optimize several parameters (Table 1) [23]. The model was also able to independently describe pitavastatin pharmacokinetic profiles after single oral doses of 2 mg and 4 mg in healthy subjects (Fig. 1b, c) [23].

Fig. 1
figure 1

The simulation of plasma concentration-time profile of pitavastatin. Simulation of pitavastatin pharmacokinetic profile using the PBPK model (mean profile) following a intravenous infusion of 2 mg pitavastatin (model development), b single oral administration of 2 mg and c single oral dose of 4 mg pitavastatin (model verification). Simulations were conducted with virtual healthy subjects as described in Methods. Data points are observed mean values [23]

Full size image

Figure 2 shows that the atorvastatin substrate model well described atorvastatin pharmacokinetic profile following 20 mg [25] or 40 mg of single oral administration in healthy subjects [35], in which 40 mg data set was not used during the model development.

Fig. 2
figure 2

The simulation of plasma concentration-time profile of atorvastatin. Simulation of atorvastatin pharmacokinetic profile using the PBPK model (mean profile) following a single oral dose of 20 mg atorvastatin (model development) (a), and a single oral dose of 40 mg atorvastatin (model verification) (b). Data points are observed mean values (20 mg [25, 26], 40 mg [35]). Simulations were conducted with virtual healthy subjects as described in Methods. Data from two different studies (with 20 mg) [25, 26] were presented by different symbols

Full size image

3.2 Simulation of Exposure Change in Populations with Different Genotypes of SLCO1B1

Simulation of the effect of SLCO1B1 polymorphism on systematic exposure of pitavastatin or atorvastatin was conducted under the assumptions of relative transporter abundance with SLCO1B1 genotypes as described in Methods. Both pitavastatin and atorvastatin models predicted an exposure change in the presence of SLCO1B1 polymorphism to be within twofold of the observed values (Table 3). The pitavastatin model slightly under-estimated the change in C max as a result of different SLCO1B1 genotypes (R value is 0.39) (Supplementary Table 1).

Table 3 The effect of SLCO1B1 polymorphism or inhibitors on AUC changes of atorvastatin or pitavastatin
Full size table

3.3 DDI Prediction

The predicted mean AUCRs and C max R for DDIs with gemfibrozil, itraconazole, and erythromycin were within two-fold of observed values (Table 3) (Supplementary Table 1). Figure 3 and Table 3 compared the simulated and observed pharmacokinetic profiles of pitavastatin and atorvastatin, respectively, in the presence of gemfibrozil.

Fig. 3
figure 3

a Simulation of pitavastatin pharmacokinetic profile when the drug is coadministered with gemfibrozil. Mean plasma concentration-time profile of oral administration of 4 mg pitavastatin in the presence of 600 mg gemfibrozil. Squares represent mean observed plasma concentration-time profile of pitavastatin dosed in the presence of gemfibrozil. Dotted line represents 5th and 95th percentile of simulated mean data, respectively. Observed data were from Drugs@FDA [23]. b Simulation of atorvastatin pharmacokinetic profile when the drug is coadministered with gemfibrozil. Mean plasma concentration-time profile of oral administration of 40 mg atorvastatin in the presence of 600 mg gemfibrozil. Squares represent mean observed plasma concentration-time profile of atorvastatin dosed in the presence of gemfibrozil. Dotted line represents 5th and 95th percentile of simulated mean data, respectively. Observed data were from Ref. [81]

Full size image

The simulated pitavastatin or atorvastatin plasma concentration-time profiles in the presence of cyclosporine (200 mg co-administrated with pitavastatin [23], 5.2 mg/kg co-administrated with atorvastatin[43] or in the presence of single intravenous infusion or oral dose of rifampin (600 mg) significantly under-estimated the observed magnitude of clinical DDI (Table 3). The R values of AUCR (See Sect. 2.6) for pitavastatin co-administered with rifampin and cyclosporine were 0.32 and 0.41, respectively; and R values of AUC for atorvastatin co-administered with rifampin and cyclosporine were 0.31 and 0.25, respectively (Table 3). Similar under-estimation in the change of C max was found for atorvastatin DDI with rifampin and cyclosporine, or pitavastatin DDI with cyclosporine (Supplementary Table 1).

3.4 Evaluation of Inhibitor K i for OATP1B1 on the Performance of DDI Prediction

The significant under-predictions with rifampin and cyclosporine may be caused by several factors. To explore one potential factor, we conducted sensitivity analyses by assuming that OATP1B1 inhibition potency in both inhibitors’ PBPK models may be lower than the actual inhibition potency in vivo. The K i values for OATP1B1 used in models were adopted from publically reported values of 0.014 and 0.9 µM for cyclosporine [44] and rifampin [19], respectively. However, a wide range of K i values for OATP1B1 have been reported for cyclosporine and rifampin (0.014–1.0 and 0.41–3.1 µM, respectively based on UWDiDB). The analyses showed that decreasing initial K i values by tenfold for both rifampin and cyclosporine significantly improved the DDI prediction (Table 3; Fig. 4). For AUCR, the R values of atorvastatin DDI using these modified K i values were 0.86 and 1.14 for cyclosporine and rifampin, respectively; the R values of pitavastatin DDI using these modified K i values became 1.02 and 0.94 for cyclosporine and rifampin, respectively (Table 3). For C max R, the modification in K i improved the predictions, with the R values of atorvastatin DDI became 0.68 and 1.33, for rifampin and cyclosporine, respectively; while the R values of pitavastatin DDI using these modified K i values became 0.48 for cyclosporine. Figures 5 and 6 show the simulated concentration-time profiles of atorvastatin and pitavastatin in the presence of rifampin with the modified K i value, respectively, which were close to the clinical observations.

Fig. 4
figure 4

The impact of different K i values for OATP1B1 on AUCR of atorvastatin DDI with rifampin (a) and cyclosporine (b). Automatic sensitivity analysis (ASA) of the K i values of rifampin (a) and cyclosporine (b) for OATP1B1 inhibition was conducted with SimCYP stimulator V14.1 using pitavastatin or atorvastatin PBPK models with rifampin or cyclosporine as the inhibitor drugs, respectively. The predicted AUCR was plotted with the corresponding K i used in ASA. The K i used from the published model was marked with “initial K i” in the figure. K i inhibitory constant, OATP organic anion transporting polypeptide, AUCR ratio of area under the concentration-time curve, PBPK physiologically based pharmacokinetic, ASA automatic sensitivity analysis, DDI drug-drug interaction, Obs observed

Full size image
Fig. 5
figure 5

Simulation of atorvastatin-rifampin drug-drug interaction with OATP1B1 K i of 0.09 µM. Mean (solid line) plasma concentration-time profile of oral administration of 40 mg atorvastatin in the presence of single intravenous infusion of 600 mg rifampin. Triangles and squares represent mean observed plasma concentration-time profile of atorvastatin dosed in the presence or absence of rifampin, respectively. The K i of rifampin on OATP1B1 was lowered by tenfold to 0.09 µM from the initial 0.9 µM. Observed data were from Ref. [79]. OATP organic anion transporting polypeptide, K i inhibitory constant

Full size image
Fig. 6
figure 6

Simulation of pitavastatin DDI with rifampin using a K i value of 0.09 µM for OATP1B1 inhibition. Mean (solid line) plasma concentration-time profile of oral administration of 4 mg pitavastatin in the presence of single oral dose of 600 mg rifampin. Squares represent mean observed plasma concentration-time profile of pitavastatin dosed in the presence of rifampin. The K i of rifampin on OATP1B1 was lowered by tenfold to 0.09 µM from the initial K i of 0.9 µM. The dashed lines are 95th and 5th percentile of simulated mean data, respectively. Observed data was taken from Ref. [23]. OATP organic anion transporting polypeptide, K i inhibitory constant, AUC area under the concentration-time curve

Full size image

4 Discussion

4.1 Performance of Statin PBPK Models

In this study, we developed substrate PBPK models for pitavastatin and atorvastatin, with the consideration of the contributions from both enzymes (CYP2C9/2C8/UGT for pitavastatin and CYP3A4 for atorvastatin) and transporters (OATP1B1/3 for pitavastatin and OATP1B1/BCRP for atorvastatin). Due to limitations in in vitro-in vivo extrapolation (IVIVE) of transporter-mediated clearance, scaling factors (e.g. intersystem extrapolation factor (ISEF) or empirical scaling factor) are frequently applied during PBPK model development to capture the observed in vivo clearance [45]. As shown by Jones et al, a wide range of scaling factors were needed for each of 7 substrate drugs investigated (including some statins), suggesting significant gaps in translating in vitro transporter kinetic data using PBPK models. This gap cannot be filled with a common scaling factor [46]. For the pitavastatin model, we used the CLint.T for OATP1B1 and OATP1B3 measured in hepatocytes, which were 58.4 and 5.1 µL/min/million cells, respectively [24]. The difference in CLint.T suggests that the fractions of transporter for OATP1B1 and OATP1B3 are different, and OATP1B1 is the major transporter for hepatic uptake of pitavastatin. The initial model with these in vitro parameters cannot describe the clinically observed 2 mg oral dose or intravenous pharmacokinetic data, and the same empirical factor of 18 for OATP1B1 and OATP1B3 has to be applied to capture clinical data. For atorvastatin, we found that we need to apply a scaling factor of 4 to CLint.T of OATP1B1 (presented in Jmax and K m) obtained from hepatocyte study in order to recapitulate clinical observation.

Clinical studies showed that the exposure of pitavastatin increased by 3.1-fold in subjects with SLCO1B1 c.521 CC (PT) genotype as compared to those with the wild type c.521 TT (ET) [47]. Similarly, an increase of 2.4-fold in the exposure of atorvastatin was found in subjects with the SLCO1B1 c.521 CC genotype as compared to those with the wild type [48]. The reduced function of OATP1B1 in subjects with genotype SLCO1B1 c.521 CC could be considered as a genetic specific inhibition of OATP1B1, which would be considered as a more specific impairment of OATP1B1 function. Predictions of the AUCR based on different SLCO1B1 genotypes indicated that the current PBPK models appear to describe the contribution of OATP1B1 to the disposition of these two statins. The predicted AUCR in DDI studies with gemfibrozil, itraconazole, and erythromycin were within twofold range of clinically observed AUCR. The good predictions of DDIs with multiple inhibitors that inhibit enzymes and/or transporter OATP1B1 further lend confidence of the final substrate PBPK models.

Besides being a substrate of OATP1B1, atorvastatin is also metabolized by CYP3A, and thereby, it is interesting to know which pathway (OATP1B1 or CYP3A) is the rate-limiting step of atorvastatin elimination. A sensitivity analysis was conducted on the K i of OATP1B1 and CYP3A4, respectively (Fig. 7). If decreasing the K i of cyclosporine for OATP1B1 by 100-fold, the change in predicted AUCR could be as large as tenfold (a steeper slope), while same fold-change (100-fold) in its K i for CYP3A4 only resulted in a 1.6-fold change in the predicted AUCR (a flatter slope). Thus, the data suggested that in the presence of a drug inhibiting both CYP3A4 and OATP1B1, systemic exposure of atorvastatin was more affected by OATP1B1 inhibition than by CYP3A4 inhibition. This sensitivity analysis supported that OATP1B1 was the rate-limiting step for the hepatic clearance of atorvastatin. Our observation is consistent with the findings from Maeda et al [49]. In their study, the exposure of atorvastatin increased 12-fold in the presence of the OATP1B1 inhibitor rifampin, but did not change in the presence of the strong CYP3A4 inhibitor itraconazole.

Fig. 7
figure 7

The impact of inhibition of OATP1B1 (a) and CYP3A4 (b) on the predicted AUCR of atorvastatin DDI with cyclosporine. With atorvastatin PBPK model as the substrate model and cyclosporine as the inhibitor model, sensitivity analysis on OATP1B1 K i (a) or CYP3A4 K i (b) was conducted, respectively. The different OATP1B1 K i or CYP3A4 K i values were plotted with corresponding predicted AUCR. K i inhibitory constant, OATP organic anion transporting polypeptide, AUCR ratio of area under the concentration-time curve, PBPK physiologically based pharmacokinetic, CYP cytochrome P450

Full size image

Despite the encouraging results of independent verification of PBPK models of pitavastatin and atorvastatin with regard to SLCO1B1 polymorphism and DDI with some inhibitors (itraconazole, erythromycin, and gemfibrozil), significant under-predictions were observed when published cyclosporine and rifampin PBPK models were directly used to simulate their effects on the exposure of pitavastatin and atorvastatin (Table 3). The simulated inhibitor pharmacokinetic profile of rifampin or cyclosporine is close to clinical observation, which excludes the possibility that the poor prediction of DDI is as a result that inhibitor model had not been adequately developed (Supplementary Fig. 2). Similar under-prediction of DDI with cyclosporine was also found in other published PBPK models. The published pravastatin PBPK model under-estimated the clinical cyclosporine-pravastatin DDIs, when the reported geometric mean in vitro K i was used in cyclosporine model [19]. The rosuvastatin PBPK model developed by Jamei et al [21] also seemed to under-estimate the clinically observed rosuvastatin-cyclosporine DDIs. In a recent PBPK analysis, Snoeys and colleagues also updated rifampin PBPK model with K i for OATPs that are approximately tenfold lower than those reported by others in order to predict the risk of clinical DDI [50]. Recently, Zhang et al. developed an atorvastatin PBPK model to predict PK of atorvastatin and its two metabolites and the effect of CYP3A modulators on atorvastatin pharmacokinetic. The author demonstrated reasonable prediction of the effect of single dose rifampin on atorvastatin pharmacokinetic, primarily via OATP inhibition; however, the author stated that OATP parameters were “optimized with the data from combination with a single dose of rifampin” and no details on the optimization and K i of inhibitor were given. Furthermore, although the predicted AUCR as a result of rifampin DDI was similar to observation in Zhang et al.’s model, the change in C max is significantly under-estimated (5.7 predicted vs. 12.7 observed). In our study, we demonstrated that OATP K i values for cyclosporine and rifampin need to be lowered by tenfold in order to predict their effects on the pharmacokinetics of pitavastatin and atorvastatin (Table 3). With a lower K i, the predicted AUCR and change in C max are close to clinical observation.

In addition to OATP inhibition potency, other reasons may contribute to under prediction of cyclosporine-statin or rifampin-statin DDI. First, the current DDI simulations were conducted in healthy subjects, while the exposure of cyclosporine may differ between heart transplant patients and virtual healthy subjects [51]. Second, for both statin substrates, we directly used hepatic passive diffusion clearance (CLpd, Table 1) based on in vitro findings, whereas we optimized empirical scaling factor values for active uptake processes only. Both passive and active uptake clearances are important in determining drug exposure in the liver, and one cannot completely rule out the possibility of a lower or a higher in vivo CLpd. Models with lower CLpd may require higher CLint,T by OATP to describe observed pharmacokinetic profiles during model development. Consequently, simulated genetic effect and DDI may be greater. If so, one can imagine the amount of additional simulations needed to investigate relative contribution of hepatic uptake by CLpd and transporter CLint,T. Third, concurring disposition pathways that may be affected by cyclosporine and rifampin have not been appropriately incorporated in the model for pitavastatin and atorvastatin. For example, we did additional sensitivity analysis to investigate the effect of lowering OATP1B1 K i of cyclosporine on predicted DDI with pravastatin and rosuvastatin. As described above, the pravastatin model under-estimated the clinically observed pravastatin-cyclosporine DDI [19]. Although lowering the OATP1B1 K i of cyclosporine slightly increased the predicted AUCR, it failed to fully recapitulate the tenfold clinically observed AUCR (Supplementary Fig. 3). Inhibition of other pathways might contribute to clinically observed AUCR. Cyclosporine is also an inhibitor of MRP2, and MRP2 contributes to the renal clearance of pravastatin, which is not considered in the pravastatin model. Similarly, the tenfold decrease in OATP1B1 K i from 0.014 µM to 0.0014 µM only slightly increase the simulated AUCR of rosuvastatin from 1.3 fold to around 2.3 fold, and can’t fully recapitulate the clinically observed 7.1 fold change in AUC [21] (Supplementary Fig. 4).

Atorvastatin and pitavastatin have inter-conversion of acid form and lactone forms. However, the administrated statin is in acid form, and in clinical studies, acids are major moiety for hepatic uptake by OATPs [27]. Because the main purpose of our modeling is to build a framework to evaluate DDI risk of pitavastatin and atorvastatin with enzyme or transporter inhibitors, we assumed that observed statin plasma pharmacokinetics are representative for the acid form. However, the possible impact of acid-lactone form inter-conversion on clinical DDIs could not be excluded.

In summary, additional analyses are needed to confirm potency of OATP inhibition by cyclosporine and rifampin. In the meantime, one needs to fully characterize disposition mechanisms for substrate drugs and adequately incorporate these mechanisms into PBPK models.

4.2 Using in Vivo DDI Information to Inform DDI Mechanisms

Given the complexity of drug transport process into cells, obtaining accurate inhibitory kinetic parameters with in vitro assays is always challenging, and K i determined in vitro is often different from assay to assay, or from laboratory to laboratory. In addition, several studies have shown substrate-dependent inhibition potency for some transporters [52, 53]. A recent study showed substrate-dependent inhibition potency for OATP1B1 while using different probe substrates, for example, estradiol-17β-glucuronide, estrone-3 sulfate, and sulfobromophthalein [54]. Estradiol-17β-glucuronide was the most sensitive in vitro OATP1B1 probe substrate among the three substrates tested [54]. It was also shown that for some inhibitors (e.g. cyclosporine, saquinavir, ritonavir), a more complicated inhibitory mechanism (i.e., time-dependent inhibition) on OATP1B1 existed [44, 55]. Pre-incubation of cyclosporine could further potentiate its inhibitory effect on OATP1B1 while testing with several OATP1B1 substrates [44]. Therefore, it may not be appropriate to use a universal “K i” for various victim (substrate) drugs.

PBPK modeling offers a great tool to evaluate the underlying inhibition potency using the observed DDI data, if uncertainties exist. For example, Kato, et al applied PBPK models to estimate in vivo K i of 11 CYP inhibitors, and found discrepancy between in vitro and in vivo K i values [56]. Recently, in the cimetidine and metformin PBPK models developed by Burt, et al [57], the clinically observed AUC ratio of DDI could only be recapitulate with the PBPK model while further decreasing cimetidine K i for OCT2 up to 1000-fold or K i for both OCT1 and OCT2 up to 500-fold. While an in vitro K i may not represent in vivo inhibition potency, an apparent “in vivo” K i could be estimated by fitting the pharmacokinetic prediction to the clinical exposure change of the victim drugs. This approach was applied in the study of Hsu, et al to derive an in vivo inhibition constant on renal transporters by probenecid [58]. The optimized probenecid K i values on the same renal transporters were different for different victim drugs. We used the similar approach as Hsu, et al to evaluate the influence of K i on DDI prediction with rifampin and cyclosporine as inhibitors, where “erroneously high” K i values might be one of the factors for under-prediction. Lowering K i of rifampin, by tenfold to 0.09 µM, significantly improved the rifampin-atorvastatin DDI prediction. In vitro data suggested that the K i value of rifampin for OATP1B1 is similar using either pitavastatin or atorvastatin as the substrate (1.07 µM vs. 3.08 µM, respectively) (UWDiDB). Therefore, we used the same modified OATP1B1 K i and applied it to predict pitavastatin-rifampin DDI. Applying the lower K i of 0.09 µM to the rifampin PBPK model improved the prediction for pitavastatin-rifampin DDI as well.

The K i values for cyclosporine varied 6.3-fold when different OATP1B1 substrates were used in the in vitro transporter inhibition assays [54], suggesting a possible substrate-dependent inhibition of cyclosporine on OATP1B1. The reported K i of cyclosporine on OATP1B1 used initially in our models was a value optimized with tacrolimus [44]. However, a further decrease in K i was needed to recapitulate the observed clinical DDI data with our models, suggesting possible substrate-dependent inhibition of cyclosporine on OATP1B1. It is of note that the in vitro K i of cyclosporine on inhibiting OATP1B1 while using pitavastatin or atorvastatin as the substrates were similar (0.23 µM with pitavastatin vs. 0.16 µM with atorvastatin) [54]. Therefore, in our study, we applied the same optimized cyclosporine K i based on atorvastatin DDI studies to the pitavastatin model that improved prediction.

5 Conclusions

Our PBPK models described the pharmacokinetics of pitavastatin and atorvastatin, and reasonably described systemic exposure change due to OATP1B1 (SCLO1B1) polymorphism or clinical DDIs with several OATP1B1 and CYP inhibitors (Table 3). However, DDIs with rifampin and cyclosporine were under-predicted using OATP1B1 in vitro K i values reported from literatures. Sensitivity analysis revealed that decreasing K i for OATP1B1 improved predictions, indicating possible discrepancy between in vitro and “in vivo” K i values for these two inhibitors. K i values of inhibitor drugs on OATP1B1 need to be carefully evaluated with known substrates for a comprehensive DDI risk assessment. The under-predictions suggest that modeling drug transporter kinetics using PBPK requires further research to improve overall predictive performance for drug transporters.