Volume 8, Issue 5 p. 296-307
Article
Open Access

Physiologically-Based Pharmacokinetic Models for CYP1A2 Drug–Drug Interaction Prediction: A Modeling Network of Fluvoxamine, Theophylline, Caffeine, Rifampicin, and Midazolam

Hannah Britz

Hannah Britz

Clinical Pharmacy, Saarland University, Saarbrücken, Germany

Search for more papers by this author
Nina Hanke

Nina Hanke

Clinical Pharmacy, Saarland University, Saarbrücken, Germany

Search for more papers by this author
Anke-Katrin Volz

Anke-Katrin Volz

Clinical Pharmacy, Saarland University, Saarbrücken, Germany

Search for more papers by this author
Olav Spigset

Olav Spigset

Department of Clinical and Molecular Medicine, Norwegian University of Science and Technology, Trondheim, Norway

Department of Clinical Pharmacology, St. Olav University Hospital, Trondheim, Norway

Search for more papers by this author
Matthias Schwab

Matthias Schwab

Dr. Margarete Fischer-Bosch-Institute of Clinical Pharmacology, Stuttgart, Germany

Department of Clinical Pharmacology, University Hospital Tübingen, Tübingen, Germany

Department of Pharmacy and Biochemistry, University Tübingen, Tübingen, Germany

Search for more papers by this author
Thomas Eissing

Thomas Eissing

Clinical Pharmacometrics, Bayer AG, Leverkusen, Germany

Search for more papers by this author
Thomas Wendl

Thomas Wendl

Clinical Pharmacometrics, Bayer AG, Leverkusen, Germany

Search for more papers by this author
Sebastian Frechen

Sebastian Frechen

Clinical Pharmacometrics, Bayer AG, Leverkusen, Germany

Search for more papers by this author
Thorsten Lehr

Corresponding Author

Thorsten Lehr

Clinical Pharmacy, Saarland University, Saarbrücken, Germany

Correspondence: Thorsten Lehr ([email protected])Search for more papers by this author
First published: 14 February 2019
Citations: 27

Abstract

This study provides whole-body physiologically-based pharmacokinetic models of the strong index cytochrome P450 (CYP)1A2 inhibitor and moderate CYP3A4 inhibitor fluvoxamine and of the sensitive CYP1A2 substrate theophylline. Both models were built and thoroughly evaluated for their application in drug–drug interaction (DDI) prediction in a network of perpetrator and victim drugs, combining them with previously developed models of caffeine (sensitive index CYP1A2 substrate), rifampicin (moderate CYP1A2 inducer), and midazolam (sensitive index CYP3A4 substrate). Simulation of all reported clinical DDI studies for combinations of these five drugs shows that the presented models reliably predict the observed drug concentrations, resulting in seven of eight of the predicted DDI area under the plasma curve (AUC) ratios (AUC during DDI/AUC control) and seven of seven of the predicted DDI peak plasma concentration (Cmax) ratios (Cmax during DDI/Cmax control) within twofold of the observed values. Therefore, the models are considered qualified for DDI prediction. All models are comprehensively documented and publicly available, as tools to support the drug development and clinical research community.

Study Highlights

  • WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

☑  Physiologically-based pharmacokinetic (PBPK) models are a valuable tool to investigate and predict the drug–drug interaction (DDI) potential of investigational drugs. A publicly available library of thoroughly and transparently evaluated models of relevant perpetrator and victim drugs used in clinical studies is needed to accelerate the drug development process.

  • WHAT QUESTION DID THIS STUDY ADDRESS?

☑  The aim of this study was to provide whole-body PBPK models of the most important cytochrome (CYP)1A2 perpetrator and victim drugs and to evaluate them for their application in PBPK DDI modeling.

  • WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

☑  This study provides publicly available and transparently built and evaluated PBPK models of fluvoxamine and theophylline. Both models integrate the current knowledge on relevant pharmacokinetic (PK) mechanisms, including the impact of different genotypes and smoking on the PK of fluvoxamine.

  • HOW MIGHT THIS CHANGE DRUG DISCOVERY, DEVELOPMENT, AND/OR THERAPEUTICS?

☑  The developed PBPK models are ready to use for their application in DDI modeling and might help to support the drug development process.

Cytochrome P450 (CYP)1A2 is an important enzyme for the metabolism of several endogenous substances (e.g., melatonin), and it is involved in the elimination of 15% of all therapeutic drugs.1 CYP1A2 is exclusively expressed in the liver, where it accounts for about 13% of total CYP content in liver microsomes.2 The expression of CYP1A2 can be markedly induced by smoking, whereas rifampicin, a strong CYP3A4 inducer, shows only a moderate potential to induce CYP1A2.1, 3 Well-known substrates of CYP1A2 include caffeine and theophylline, which are mainly metabolized via CYP1A2 (fractions metabolized of 0.954 and 0.7,5, 6 respectively) and can, therefore, be used as sensitive CYP1A2 substrates to evaluate the activity of CYP1A2 in vivo.7 The most important inhibitor of CYP1A2 is fluvoxamine.

The US Food and Drug Administration (FDA) specifies caffeine as a sensitive clinical index substrate and fluvoxamine as a strong clinical index inhibitor for CYP1A2. Furthermore, they recommend considering a clinical study in smokers for investigational drugs that are CYP1A2 substrates.8 Theophylline is classified as a sensitive clinical substrate and rifampicin as moderate clinical inducer of CYP1A2.9

Physiologically-based pharmacokinetic (PBPK) modeling is a valuable method, recognized by the FDA and the European Medicines Agency, to explore and quantitatively predict the pharmacokinetics (PK) of drugs, to evaluate drug–drug interactions (DDIs), and to support clinical study design, dose selection, and labeling.8, 10-12 The FDA furthermore supports the prediction of DDI studies with weak and moderate index inhibitors and inducers as an alternative to prospective clinical studies, if the sponsors can demonstrate adequate model performance using clinical data from DDI studies with strong index perpetrators.8

The aim of this study was to develop a PBPK DDI network for CYP1A2 and thereby to extend the library of publicly available PBPK models for DDI prediction.13, 14 For this purpose, whole-body PBPK models of fluvoxamine and theophylline have been developed and existing models of caffeine,15 rifampicin,13 and midazolam13 have been expanded and coupled for mutual validation of the DDI performance of these five models. The evaluation of the single models and of the network was accomplished by prediction of multiple clinical DDI studies, demonstrating their performance with different victim or perpetrator drugs. Figure 1 shows the successfully developed CYP1A2 PBPK DDI network, with caffeine and theophylline as sensitive substrates, fluvoxamine as a strong inhibitor, and rifampicin and smoking as moderate inducers (owing to the lack of strong CYP1A2 inducers). The evaluation of the final fluvoxamine PBPK model, including the fluvoxamine fraction metabolized via CYP2D6, was supported by a post hoc population pharmacokinetic (PopPK) analysis to confirm the PBPK results concerning the impact of CYP2D6 poor metabolism and smoking on the metabolism of fluvoxamine. The supplementary document (Supplement S1) to this paper was devised as comprehensive documentation and reference guide and provides detailed information on the single models and modeled DDI studies, including all model parameters, plots, and quantitative assessments of model performance.

Details are in the caption following the image
Cytochrome P450 (CYP) 1A2 drug–drug interaction (DDI) network. Schematic illustration of the developed CYP1A2 DDI network with fluvoxamine and rifampicin as CYP1A2 perpetrator drugs and theophylline and caffeine as CYP1A2 victim drugs. Midazolam was used as CYP3A4 victim drug for fluvoxamine. Dark green lines indicate induction by rifampicin or smoking, and the red and orange lines indicate inhibition by fluvoxamine.

Methods

Software

PBPK modeling was performed with PK-Sim and MoBi modeling software version 7.3.0 (part of the Open Systems Pharmacology Suite,16 www.open-systems-pharmacology.org). Parameter optimization was accomplished using the Monte Carlo algorithm implemented in PK-Sim. Sensitivity analysis was performed within PK-Sim. PopPK analysis was performed with NONMEM version 7.3 (ICON Development Solutions, Ellicott City, MD). Digitization of published plasma concentration-time profiles was accomplished using GetData Graph Digitizer version 2.26.0.20 (S. Fedorov). PK parameter analysis was performed with MATLAB version R2013b (The MathWorks, Natick, MA). Graphics were compiled with R version 3.5.1 (The R Foundation for Statistical Computing, Vienna, Austria) and RStudio version 1.1.453 (RStudio, Boston, MA). SAS version 9.4 (SAS Institute, Cary, NC) was used for statistical analysis and graphics of the PopPK analysis.

PBPK model building

Fluvoxamine and theophylline PBPK model building was started with an extensive literature search to collect physicochemical parameters, information on absorption, distribution, metabolism, and excretion processes and clinical studies of i.v. and p.o. administration of fluvoxamine and theophylline in single-dose and multiple-dose regimens.

The PBPK models were built based on healthy individuals, using the reported mean values for age, weight, height, and genetic background for each study protocol. If no information on these parameters could be found, a healthy male European individual, 30 years of age, with a body weight of 73 kg and a height of 176 cm was used.

To model the specific metabolic clearance, relevant CYP enzymes were implemented in accordance with literature, using the PK-Sim expression database reverse transcription-polymerase chain reaction profiles17 to define their relative expression in the different organs of the body. For more details see Table S6 in Supplement S1. Glomerular filtration and enterohepatic cycling were enabled, as they are active under physiological conditions.

To build the data sets for PBPK modeling, the reported observed plasma concentration-time profiles were digitized and divided into “training data set” and “test data set.” Model parameters that could not be informed from experimental reports were optimized by simultaneously fitting the model to all measured plasma concentration-time profiles assigned to the training data set. To limit the parameters to be optimized during model building, the minimal number of processes necessary was implemented into the model. Model evaluation was carried out based on the clinical data of the test data set. Descriptive (training data set) and predictive (test data set) performance of the model for all published clinical studies is transparently presented in Supplement S1.

PBPK model evaluation

Model performance was evaluated with different methods. The predicted population plasma concentration-time profiles were compared with the plasma concentration-time profiles observed in the clinical studies. Furthermore, predicted plasma concentration values of all studies were compared with the observed plasma concentrations in goodness-of-fit plots. In addition, the performance was evaluated by comparison of predicted to observed area under the plasma curve (AUC) and peak plasma concentration (Cmax) values. As quantitative measures of the descriptive and predictive performance of the models, the mean relative deviation (MRD) according to Edginton et al.18 and the geometric mean fold error (GMFE) were calculated. MRD was calculated for all observed plasma concentrations according to Eq. 1.

urn:x-wiley:21638306:media:psp412397:psp412397-math-0001(1)

with log10 cobs = logarithm of the observed plasma concentration, log10 cpred = logarithm of the predicted plasma concentration, and N = number of observed values. An MRD value ≤ 2 characterizes an adequate prediction.

The GMFE was calculated for all observed AUC and Cmax values according to Eq. 2.

urn:x-wiley:21638306:media:psp412397:psp412397-math-0002(2)

with pred PK parameter = predicted AUC or Cmax value, obs PK parameter = observed AUC or Cmax value, and n = number of studies. A GMFE value below two characterizes an adequate prediction.

PopPK model building and evaluation

Fluvoxamine PBPK model evaluation was supported by a post hoc PopPK analysis to quantify the effect of CYP2D6 poor metabolism and the impact of smoking on fluvoxamine clearance and to compare the results to the effect sizes predicted by the PBPK model.

PopPK analysis, model evaluation, and simulation were performed using nonlinear mixed-effects modeling techniques implemented in NONMEM. A full description of the PopPK methodology is available in Supplement S1.

DDI network building

In addition to the evaluation methods described above, a CYP1A2 DDI network was built to evaluate the DDI performance of the developed models (Figure 1). Fluvoxamine was used as a CYP1A2 and CYP3A4 inhibitor theophylline and caffeine as CYP1A2 victim drugs, rifampicin as CYP1A2 and CYP2E1 inducer, and midazolam as a CYP3A4 victim drug. Mathematical implementation of the drug interaction processes in general is specified in Supplement S1. All induction and inhibition processes were modeled using interaction parameter values from in vitro experimental reports without further adjustment or fitting.

DDI network evaluation

All predicted DDI simulations were evaluated by comparison of predicted vs. observed victim drug plasma concentration-time profiles alone and during co-administration, DDI AUC ratios (Eq. 3), and DDI Cmax ratios (Eq. 4).

urn:x-wiley:21638306:media:psp412397:psp412397-math-0003(3)
urn:x-wiley:21638306:media:psp412397:psp412397-math-0004(4)

As a quantitative measure of the prediction accuracy for each DDI interaction, GMFEs of the predicted DDI AUC ratios and DDI Cmax ratios were calculated according to Eq. 2.

Sensitivity analysis

Sensitivity of the final PBPK models to single parameters (local sensitivity analysis) was calculated, measured as relative changes of the AUC of one dosing interval in steady-state conditions for simulations of the highest recommended doses for fluvoxamine (300 mg once daily) and theophylline (500 mg once daily), respectively.

Parameters were included into the analysis if they have been optimized (Table S1b or S2b in Supplement S1), if they might have a strong influence due to calculation methods used in the model (fraction unbound) or if they had significant impact in former models (solubility, blood/plasma ratio, and glomerular filtration rate fraction).

Sensitivity to a parameter is calculated as the ratio of the relative change of the simulated AUC to the relative variation of the parameter around the value used in the final model according to Eq. 5.

urn:x-wiley:21638306:media:psp412397:psp412397-math-0005(5)

with S = sensitivity of the AUC to the examined model parameter, ΔAUC = change of the AUC, AUC = simulated AUC with the original parameter value, Δp = change of the examined model parameter value, and p = original model parameter value. A sensitivity value of + 1.0 signifies that a 10% increase of the examined parameter causes a 10% increase of the simulated AUC.

Virtual population characteristics

To predict the variability of the simulated plasma concentration-time profiles, virtual populations of 100 individuals were generated, containing European, Asian, or Japanese individuals. The percentage of male and female individuals and the age and weight ranges were set corresponding with the reported demographics. If not specified, virtual populations containing 50 male and 50 female individuals 20–50 years of age were used, without specific body weight or height restriction as implemented in the software. For details on study populations see Tables S1a, S2a, S7a, S8a, S9a, and S10a in Supplement S1. In the generated virtual populations, corresponding organ volumes, tissue compositions, blood flow rates, etc. were varied by an implemented algorithm within the limits of the International Commission on Radiological Protection,19, 20 Tanaka and Kawamura,21 or Japanese22 databases. In addition, the reference concentrations of the implemented CYP enzymes were set to be distributed with the default variabilities for their expression available in PK-Sim. Table S6 in Supplement S1 summarized the implemented enzymes with their reference concentrations and variabilities.

With these populations, simulations were generated and compared with observed data. As the observed data were reported in terms of arithmetic means and SDs, simulated 68% population prediction intervals were plotted that correspond to the range span of ± 1 SD around the mean assuming normal distribution.

Results

PBPK model building and evaluation

The final PBPK models of fluvoxamine and theophylline precisely describe and predict the plasma concentration-time profiles following i.v. and p.o. administration for a large range of administered doses.

Plots of population predicted compared with observed plasma concentration-time profiles of all studies obtained from literature are shown in linear as well as in semilogarithmic plots in Figure 2a (selected fluvoxamine studies), Figure 3a (selected theophylline studies), and Figures S1a, S1b, S2a, and S2b of Supplement S1 (all studies). Goodness-of-fit plots are presented in Figure 2b (fluvoxamine), Figure 3b (theophylline), and Figures S1c and S2c of Supplement S1. MRD values of all studies are listed in Tables S1c and S2c of Supplement S1.

Details are in the caption following the image
Fluvoxamine plasma concentrations. (a) Population predictions of selected fluvoxamine plasma concentration-time profiles compared with observed data in linear (left panel) and semilogarithmic plots (right panel). The upper panel shows i.v. application, the lower panel p.o. administration of fluvoxamine. Observed data are shown as dots ± SD.34, 35 Population simulation arithmetic means are shown as lines; the shaded areas illustrate the 68% population prediction intervals. (b) Predicted compared with observed fluvoxamine plasma concentration values of all clinical studies. Line of identity and 0.5-fold to 2.0-fold acceptance limits are shown as black lines. The 0.8-fold to 1.25-fold limits are shown as grey lines. Details on dosing regimens and study populations are listed in Table S1a of Supplement S1. Predicted and observed pharmacokinetic parameters are summarized in Table S1d of Supplement S1.
Details are in the caption following the image
Theophylline plasma concentrations. (a) Population predictions of selected theophylline plasma concentration-time profiles compared with observed data in linear (left panel) and semilogarithmic plots (right panel). The upper panel shows i.v. application, the lower panel p.o. administration of theophylline. Observed data are shown as dots ± SD.36, 37 Population simulation arithmetic means are shown as lines, and the shaded areas illustrate the 68% population prediction intervals. (b) Predicted compared with observed theophylline plasma concentration values of all clinical studies. Line of identity and 0.5-fold to 2.0-fold acceptance limits are shown as black lines. The 0.8-fold to 1.25-fold limits are shown as grey lines. Details on dosing regimens and study populations are listed in Table S2a of Supplement S1. Predicted and observed pharmacokinetic parameters are summarized in Table S2d of Supplement S1.

Predicted compared with observed AUC and Cmax values of all studies with calculated GMFEs are listed in Tables S1d and S2d of Supplement S1. Plots showing the correlation of predicted to observed AUC and Cmax values of all studies are presented in Figures S1f and S2d of Supplement S1.

For fluvoxamine PBPK model development, 26 different clinical studies with PK blood sampling were used, with 9 studies assigned to the training data set (Table S1a in Supplement S1). The fluvoxamine PBPK model applies metabolism by CYP1A2, CYP2D6, and glomerular filtration.

To distinguish between fluvoxamine metabolism in CYP2D6 extensive metabolizers (EMs) and poor metabolizers (PMs), the CYP2D6 catalytic rate constant (kcat) of PMs was set to zero. This assumption was made because CYP2D6 PMs were characterized by absent CYP2D6 enzymatic activity,23 which results in a predicted 1.5-fold increase of the fluvoxamine AUC in CYP2D6 PMs compared with CYP2D6 EMs (observed: 1.3-fold increase24). Population predictions of fluvoxamine plasma concentration-time profiles compared with observed data for CYP2D6 EMs and PMs are shown in Figure S1d of Supplement S1.

Furthermore, the final model is able to describe the influence of smoking on the PK of fluvoxamine. Smoking is the strongest known inducer of CYP1A2 and results in higher metabolism of CYP1A2 substrates.1 As no detailed information on the frequency, duration, and amount of smoking was available from literature, the induction of CYP1A2 was implemented as a static 1.38-fold increase in enzyme activity. This factor was optimized based on the study of Spigset et al.,25 resulting in a 39% reduction of the fluvoxamine AUC in smokers (observed: 31% reduction). Population predictions of fluvoxamine plasma concentration-time profiles compared with observed data for nonsmokers and smokers are shown in Figure S1e of Supplement S1. Drug-dependent parameters of the final fluvoxamine model are listed in Table S1b of Supplement S1. System-dependent parameters are given in Table S6 of Supplement S1.

Sensitivity analysis of a simulation of 300 mg fluvoxamine p.o. once daily with a sensitivity threshold of 0.5 reveals that the fluvoxamine model is sensitive to the values of lipophilicity (optimized), CYP2D6 catalytic rate constant (optimized), CYP2D6 Michaelis-Menten constant (literature value), and fraction unbound (literature value; see Figure S1g of Supplement S1).

For theophylline PBPK model development, 40 different clinical studies with PK blood sampling and additional fraction excreted unchanged to urine measurements and CYP1A2 fraction metabolized information were used, with 13 clinical studies assigned to the training data set (Table S2a in Supplement S1). The theophylline PBPK model applies metabolism by CYP1A2, CYP2E1, and glomerular filtration with reabsorption in the renal tubulus.

In the model, CYP1A2 metabolizes theophylline with high affinity but low capacity, whereas CYP2E1 metabolizes theophylline with low affinity and high capacity, as described in the literature,26 resulting in a good prediction of the observed concentration dependency of theophylline metabolism. About 95% of an administered theophylline dose are excreted with the urine but only 14–17% as unchanged drug.27, 28 Due to the lack of valid in vitro data on renal tubular reabsorption transporters for theophylline, the glomerular filtration rate fraction was optimized to a value of 0.22 to describe the fraction of theophylline excreted unchanged to urine. Drug-dependent parameters of the final theophylline model are listed in Table S2b of Supplement S1. System-dependent parameters are given in Table S6 of Supplement S1.

Sensitivity analysis of a simulation of 500 mg theophylline p.o. once daily with a sensitivity threshold of 0.5 reveals that the theophylline model is sensitive to the values of fraction unbound (literature value), CYP1A2 catalytic rate constant (optimized), and CYP1A2 Michaelis-Menten constant (literature value; see Figure S2e of Supplement S1).

DDI network modeling

For the CYP1A2 DDI network modeling, eight different clinical DDI studies were available, consisting of two studies of fluvoxamine with theophylline, three studies of rifampicin with theophylline, two studies of fluvoxamine with caffeine, and one study of fluvoxamine with midazolam. The victim drug plasma concentration-time profiles of these studies, before and during perpetrator treatment, were predicted and compared with observed data. Tables S7a, S8a, S9a, and S10a of Supplement S1 list the administration protocols and study population details of the clinical DDI studies. The parameters to model the CYP1A2, CYP2E1, and CYP3A4 induction and inhibition processes are described in Supplement S1. Population predictions of plasma concentration-time profiles of the different victim drugs before and during coadministration are presented in linear as well as semilogarithmic plots in Figure 4 (selected studies) and Figures S7a, S8a, S9a, and S10a of Supplement S1 (all studies). All victim drug plasma concentration-time profiles before and during coadministration with fluvoxamine or rifampicin are well-predicted over the full range of reported administration protocols.

Details are in the caption following the image
Plasma concentration-time profiles of the drug–drug interaction (DDI) network. Population predictions of selected plasma concentration-time profiles compared with observed data for the fluvoxamine-theophylline, rifampicin-theophylline, fluvoxamine-caffeine, and fluvoxamine-midazolam DDIs in linear (left panel) and semilogarithmic plots (right panel). Observed data are shown as dots ± SD.38-41 Population simulation arithmetic means are shown as lines, and the shaded areas illustrate the 68% population prediction intervals. Details on dosing regimens and study populations are listed in Tables S7a, S8a, S9a, and S10a of Supplement S1. Predicted and observed pharmacokinetic parameters are summarized in Tables S7b, S8b, S9b, and S10b of Supplement S1.

Figure 5 shows the correlation of predicted to observed DDI AUC ratios and DDI Cmax ratios of the modeled DDI studies as a visualization of the performance of the entire network. Table 1 lists the corresponding DDI AUC ratio and DDI Cmax ratio values shown in Figure 5, with calculated GMFE values for each perpetrator–victim drug combination, demonstrating the good performance of the developed models when applied for DDI prediction.

Details are in the caption following the image
Correlation of predicted to observed drug–drug interaction (DDI) area under the curve (AUC) ratios, and DDI peak plasma concentration (Cmax) ratios. The left panel illustrates the predicted compared with observed DDI AUC ratios, the right panel illustrates the predicted compared with observed DDI Cmax ratios of the fluvoxamine-theophylline, rifampicin-theophylline, fluvoxamine-caffeine, and fluvoxamine-midazolam DDIs. Fluvoxamine interaction studies are shown as dots and rifampicin interaction studies are shown as triangles. The colors represent the different victim drugs. The line of identity and the 0.5-fold to 2.0-fold acceptance limits are shown as straight black lines. The curved grey lines are the prediction acceptance limits proposed by Guest et al.42Study references, dosing regimens, and values of predicted and observed DDI AUC ratios and DDI Cmax ratios are listed in Table 1.
Table 1. DDI AUC ratios, DDI Cmax ratios, and GMFE values of DDI studies
Perpetrator drug Victim drug Observed DDI AUC ratio Predicted DDI AUC ratio Pred/ObsDDI AUC ratio Observed DDI Cmax ratio PredictedDDI Cmax ratio Pred/ObsDD Cmax ratio Reference
Fluvoxamine Theophylline
50 mg p.o., q.d./b.i.d. 3.21 mg/kg p.o., s.d. 2.40 2.16 0.90 1.09 1.08 0.99 Orlando 200643
50/100 mg p.o., q.d. 257 mg p.o., s.d. 2.70 3.10 1.15 1.16 1.13 0.97 Rasmussen 38
GMFE (range) 1.13 (1.11–1.15) 1.02 (1.01–1.03)
Pred/Obs within twofold 2/2 2/2
Rifampicin Theophylline
600 mg p.o., q.d. 3.95 mg/kg i.v. (30 minutes) 0.83 0.87 1.05 0.98 Powell-Jackson 198544
600 mg p.o., q.d. 5.19 mg/kg i.v. (20 minutes) 0.81 0.71 0.89 1.19 0.98 0.82 Boyce 198439
600 mg p.o., q.d. 355.5 mg p.o., s.d. 0.87 0.75 0.87 0.90 0.89 0.99 Powell-Jackson 198544
GMFE (range) 1.12 (1.05–1.16) 1.11 (1.01–1.21)
Pred/Obs within twofold 3/3 2/2
Fluvoxamine Caffeine
50/100 mg p.o., q.d. 200 mg p.o., s.d. 5.40 4.99 0.92 1.06 1.07 1.01 Jeppesen 199640
100 mg p.o., b.i.d 250 mg p.o., s.d. 14.90 7.03 0.47 1.44 1.06 0.74 Culm-Merdek 200545
GMFE (range) 1.51 (1.08–2.12) 1.17 (1.01–1.36)
Pred/Obs within twofold 1/2 2/2
Fluvoxamine Midazolam
50 mg p.o., b.i.d. 10 mg p.o., s.d. 1.38 1.51 1.09 1.40 1.34 0.95 Lam 200341
GMFE 1.09 1.04
Pred/Obs within twofold 1/1 1/1
  • AUC, area under the plasma concentration-time curve; Cmax, peak plasma concentration; DDI, drug-drug interaction; GMFE, geometric mean fold error; Pred/Obs, predicted/observed; -, no data available.

PopPK modeling of fluvoxamine

The PK of fluvoxamine were best described by a one-compartment model with zero-order absorption with a lag time and linear elimination from the central compartment. As shown in Table S11 of Supplement S1, parameter estimates were precise. Goodness-of-fit plots (Figure S11a in Supplement S1) and visual predictive checks (Figure S11b in Supplement S1) demonstrate the good descriptive performance of the model.

The impact of CYP2D6 phenotype on total clearance of fluvoxamine was best described as a categorical covariate. Volunteers who are CYP2D6 PMs show a 22% lower total clearance of fluvoxamine compared with EMs. Furthermore, fluvoxamine clearance was found to be 28% higher in smokers compared with nonsmokers.

Discussion

The developed PBPK models of fluvoxamine and theophylline reliably describe and predict plasma concentration-time profiles over the full range of published doses and administration protocols. Their good descriptive and predictive performance has been demonstrated by comparison of predicted to observed plasma concentration-time profiles, AUC and Cmax values, calculation of MRDs and GMFEs, as well as with the prediction of different DDIs. Although the populations used for model predictions were carefully generated according to the reported study demographics, CYP1A2 and CYP2D6 show high interindividual variability, and information on smoking status and CYP2D6 phenotype were lacking in most of the study reports. This could explain why a small percentage of the fluvoxamine and theophylline studies cannot be accurately predicted using the same kcat values for all studies.

There are two previously published PBPK models of fluvoxamine: a minimal PBPK model (three compartments)29 and a model built on the basis of few clinical studies (four studies).30 For theophylline, one PBPK model has been previously reported, developed to predict the disposition of theophylline during pregnancy.31 All three models have not been challenged by prediction of DDIs. The whole-body PBPK models presented in this study have been built using a multitude of clinical studies, are transparently documented, and they have been evaluated in a DDI network.

To describe the metabolism of fluvoxamine, CYP1A2 and CYP2D6 were implemented into the PBPK model. Model building was started with the working hypothesis that CYP2D6 accounts for up to 60% of fluvoxamine metabolism.32 However, our PBPK analysis suggested a higher fraction of fluvoxamine metabolized by CYP1A2 than by CYP2D6. This result was supported by the finding that fluvoxamine total apparent clearance (CL/F) in CYP2D6 PMs (no CYP2D6 activity) was only 25% lower than in CYP2D6 EMs.32 (Taking into account that CYP2D6 is also expressed in the intestine, CYP2D6 PMs might show a higher bioavailability of fluvoxamine, reducing CL/F, and thereby further reducing the impact of CYP2D6 poor metabolism on fluvoxamine clearance.)

To confirm this relatively small impact of CYP2D6 poor metabolism on fluvoxamine PK, a PopPK analysis of fluvoxamine was conducted. The reduction of fluvoxamine CL/F in CYP2D6 PMs compared with EMs was quantified at 22%. This is the first reported compartmental analysis of fluvoxamine, which is in very good agreement with the noncompartmental result for reduction of CL/F in CYP2D6 PMs of 25%.32

Simulation of fluvoxamine fraction metabolized using the final PBPK model and a single dose of 50 mg predicts fractions metabolized of 20% by CYP2D6 and of 71% by CYP1A2, which is very close to the PopPK analysis result. Neither fraction metabolized information nor the CYP2D6 PM fluvoxamine plasma profiles were used during the final PBPK model parameter optimization. Fitting the catalytic rate constants of CYP2D6 and CYP1A2 and, therefore, the contribution of both enzymes to fluvoxamine metabolism to get a good description of the nonlinear PK of fluvoxamine for the different doses administered already resulted in a model that accurately describes the fractions metabolized.

The inducing effect of smoking on the metabolism of fluvoxamine is also well-described by the PBPK model, with AUC ratios smoking/nonsmoking of 0.61 predicted and 0.69 observed. The fluvoxamine PopPK analysis gives a 28% higher CL/F of fluvoxamine in smokers compared with nonsmokers. The small overprediction of the fluvoxamine Cmax in smokers could be attributed to gastrointestinal effects of smoking that reduce the absorption of fluvoxamine but were not accounted for in the model. However, due to a lack of more detailed information on the frequency, duration, and amount of smoking, the induction of CYP1A2 could only be implemented as a static increase of the enzyme activity. To model this CYP induction in a mechanistic and dynamic way, for example, to predict the return of CYP1A2 activity to baseline when smoking is stopped before a surgical intervention, as well as to validate the estimated factor on CYP1A2 enzyme activity for the smoking population, more data are needed.

The developed theophylline model can be used for prediction of plasma concentration-time profiles following i.v. administration or p.o. administration of syrup, solution, or immediate-release formulations. As the reported plasma concentration profiles of the different sustained release dosage forms strongly vary with the mechanism used for prolongation of drug release, sustained release or enteric coated theophylline formulations were not considered in the current investigation. If needed, the model can be easily extended by implementation of sustained release drug dissolution profiles.33

The DDIs presented in this study have been modeled using reported experimental values to inform all necessary interaction parameters. This approach is followed as an additional means of model evaluation, predicting all available reported clinical DDI studies, and comparing the observed data to model predictions. The caffeine,15 rifampicin,13 and midazolam13 PBPK models applied have been evaluated and described elsewhere. The existing rifampicin model has been extended to predict the induction of CYP1A2 and CYP2E1 by rifampicin. The DDI performance of the enhanced rifampicin model has been successfully evaluated with the data of three different clinical rifampicin-theophylline DDI studies.

The presented CYP1A2 DDI network demonstrates the good performance of all models for DDI prediction over the full range of reported DDI study protocols. This has been shown by victim drug concentration-time profiles, DDI AUC ratios, DDI Cmax ratios, and corresponding GMFE values. All DDIs of fluvoxamine with the sensitive CYP1A2 victim drugs theophylline and caffeine are well predicted. The moderate inhibition of CYP3A4 by fluvoxamine was successfully implemented and evaluated by prediction of the fluvoxamine-midazolam DDI. Due to the present lack of models for CYP2C19 victim drugs, the strong inhibition of CYP2C19 by fluvoxamine could not be tested. However, fluvoxamine CYP2C19 interaction parameters are reported and can be easily implemented into the presented fluvoxamine PBPK model.

In summary, a PBPK CYP1A2 DDI network has been successfully developed. Whole-body PBPK models of fluvoxamine and theophylline have been carefully built and evaluated by DDI prediction using different kinds of perpetrator (induction, competitive inhibition, and mixed inhibition) and victim drugs (CYP1A2 and CYP3A4). Furthermore, a previously developed model of rifampicin has been expanded with parameters for CYP1A2 and CYP2E1 interaction and tested. The resulting PBPK network of fluvoxamine, theophylline, caffeine, rifampicin, and midazolam adequately predicts the observed data of all clinical DDI studies reported for combinations of these drugs and, therefore, all models are considered ready to use for DDI prediction. The newly developed models of fluvoxamine and theophylline are transparently documented and the model files, also including DDI model files, are provided as Supplementary Material to this paper (Data S1-S6) as well as in the Open Systems Pharmacology repository (www.open-systems-pharmacology.org), to extend the library of publicly available PBPK models for DDI prediction. They can be applied to help understand and characterize the DDI potential of investigational drugs, to inform the design of clinical trials, or to generate dose recommendations for comedication.

Funding

This project has received funding from the European Union's Horizon 2020 Research and Innovation Programme under grant agreement No. 668353. M.S. was supported by the Robert Bosch Stiftung, Stuttgart, Germany.

Conflict of Interest

T.E., T.W., and S.F. are employees of Bayer AG. A.K.V. performed this work at the Saarland University. Since January 2018, she is an employee of the Federal Institute for Drugs and Medical Devices (BfArM) and declares no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. No potential conflicts of interest were disclosed by the other authors.

Author Contributions

H.B., N.H., A.K.V., and T.L. wrote the manuscript. N.H., O.S., M.S., T.E., T.W., S.F., and T.L. designed the research. H.B., N.H., A.K.V., S.F., and T.L. performed the research.

Disclaimer

The views expressed in this article are the views of A.K.V. and do not necessarily reflect the views and opinions of the BfArM.