Physiologically Based Pharmacokinetic Modeling to Describe the CYP2D6 Activity Score-Dependent Metabolism of Paroxetine, Atomoxetine and Risperidone

The cytochrome P450 2D6 (CYP2D6) genotype is the single most important determinant of CYP2D6 activity as well as interindividual and interpopulation variability in CYP2D6 activity. Here, the CYP2D6 activity score provides an established tool to categorize the large number of CYP2D6 alleles by activity and facilitates the process of genotype-to-phenotype translation. Compared to the broad traditional phenotype categories, the CYP2D6 activity score additionally serves as a superior scale of CYP2D6 activity due to its finer graduation. Physiologically based pharmacokinetic (PBPK) models have been successfully used to describe and predict the activity score-dependent metabolism of CYP2D6 substrates. This study aimed to describe CYP2D6 drug–gene interactions (DGIs) of important CYP2D6 substrates paroxetine, atomoxetine and risperidone by developing a substrate-independent approach to model their activity score-dependent metabolism. The models were developed in PK-Sim®, using a total of 57 plasma concentration–time profiles, and showed good performance, especially in DGI scenarios where 10/12, 5/5 and 7/7 of DGI AUClast ratios and 9/12, 5/5 and 7/7 of DGI Cmax ratios were within the prediction success limits. Finally, the models were used to predict their compound’s exposure for different CYP2D6 activity scores during steady state. Here, predicted DGI AUCss ratios were 3.4, 13.6 and 2.0 (poor metabolizers; activity score = 0) and 0.2, 0.5 and 0.95 (ultrarapid metabolizers; activity score = 3) for paroxetine, atomoxetine and risperidone active moiety (risperidone + 9-hydroxyrisperidone), respectively.


Introduction
Differences in CYP2D6 activity have been described as early as the 1970s [1,2] and have since been a major focus of clinical research, as CYP2D6 is involved in the metabolism of approximately 20% of clinically relevant drugs [3]. Polymorphic expression of the CYP2D6 gene has been identified as the single most important determinant of CYP2D6 activity leading to a substantial interindividual and interpopulation variability observed in the pharmacokinetics of CYP2D6 substrates [4]. For instance, homozygous carriers of loss-offunction alleles (genetic poor metabolizers) show no detectable CYP2D6 activity [3] and are consequently unable to biotransform drugs via CYP2D6 [4]. In contrast, individuals carrying multiple copies of a normal function allele (genetic ultrarapid metabolizers) generally display increased CYP2D6 activity compared to homozygous carriers of wildtype alleles (genetic normal metabolizers) [5] and show accelerated biotransformation of CYP2D6

Clinical Study Data
An extensive literature search was conducted to gather individual and aggregated plasma concentration-time profiles after intravenous and oral administrations in single

Clinical Study Data
An extensive literature search was conducted to gather individual and aggregated plasma concentration-time profiles after intravenous and oral administrations in single and multiple-dose regimes of paroxetine, atomoxetine, and risperidone. Additionally, population or individual demographics (sex, age, weight, and height) alongside CYP2D6 activity (phenotype, genotype, or activity score) were extracted from the respective studies. The collected plasma concentration-time profiles were split into a training dataset for model

PBPK Base Model Building
The paroxetine and atomoxetine PBPK base models were developed using a sequential approach. First, appropriate quantitative structure-activity relationship (QSAR) methods to estimate partition coefficients and cellular permeabilities were selected by the smallest residual error for fitting simulations of intravenous administrations (paroxetine) or all studies of the training dataset (atomoxetine) to their observed data. Second, simulations of administrations of oral solutions were optimized against the respective clinical data to estimate intestinal permeability. Third, parameters for CYP2D6-independent metabolism were informed by fitting simulations of single and multiple-dose oral administrations in poor metabolizers of CYP2D6 to their respective observed data. Finally, parameters for CYP2D6-mediated metabolism were optimized for studies of the training dataset where the volunteers were extensive metabolizers. Here, the term "extensive metabolizers" was used to group populations that were either phenotyped via traditional phenotyping methods or populations, which were not phenotyped.
For the risperidone base model, a published PBPK model by Kneller et al. [21] was used.

PBPK Base Model Evaluation
The performance of the presented models was evaluated using graphical and statistical methods. First, predicted plasma concentration-time profiles were compared graphically with measured data from the respective clinical studies by plotting model population predictions (arithmetic mean ± SD) together with observed data points. For this purpose, virtual populations of 1000 individuals were created based on the population characteristics reported in the respective publication. System-dependent parameters, such as age, body weight, height, organ weights, blood flow rates, and tissue composition, were varied by the implemented algorithm in PK-Sim ® . Second, the plasma concentration values of all studies using the predicted arithmetic mean of the population were plotted against the corresponding observed values in goodness-of-fit plots. In addition, model performance was evaluated by a comparison of predicted to observed area under the concentrationtime curve (AUC) and maximum plasma concentration (C max ) values. All AUC values (predicted as well as observed) were calculated from the time of the first concentration measurement to the time of the last concentration measurement (AUC last ).
Finally, as quantitative measures of the model performance, the mean relative deviation (MRD) of all predicted plasma concentrations (Equation (1)) and the geometric mean fold error (GMFE) of all predicted AUC last and C max values (Equation (2)) were calculated.
where c i = predicted plasma concentration that corresponds to the i-th observed concentration, c i = i-th observed plasma concentration, and k = number of observed values.
where P i = predicted AUC last or C max value of study, p i = corresponding observed AUC last or C max value of study, i, and m = total number of studies.

Local Sensitivity Analysis
Local sensitivity of the AUC 0-24 h of paroxetine, atomoxetine, risperidone or 9-hydroxyrisperidone to single parameter changes was analyzed for simulations of single orally administered standard doses of paroxetine, atomoxetine, and risperidone, respectively. Parameters were included in the analysis if they have been optimized (intestinal permeabilities and k cat values), if they are associated with optimized parameters (K M values) or if they might have a strong impact due to calculation and QSAR methods used (lipophilicities, pK a values and fractions unbound (f u )). A detailed description of the model sensitivity analysis is provided in Section S1.4 of Supplementary Materials S1. Overviews of all varied parameters for the respective compounds are provided in Sections S2.2.3, S3.2.3 and S4.2.3 of Supplementary Materials S1.

DGI Model Building
CYP2D6-dependent clearance processes were modeled using Michaelis-Menten kinetics according to Equation (3): where v = reaction velocity at substrate concentration S, V max = maximum reaction velocity, K M = Michaelis-Menten constant, k cat = catalytic rate constant, and E = enzyme concentration. The CYP2D6 DGI models for paroxetine, atomoxetine, and risperidone were developed based on two previously published models for the CYP2D6 substrates metoprolol [17] and dextromethorphan [15]. Relative k cat values, defined as the ratio of k cat values for populations with a variant activity score and the k cat for populations with an activity score of 2 (corresponding to 100% of CYP2D6 activity), were calculated according to Equation (4): where k cat, rel, AS = i = k cat for the investigated activity score relative to AS = 2 and k cat, AS = i = k cat for activity score i. Activity scores were assigned according to the current consensus [5]. CYP2D6 k cat, rel values used to describe the activity score-dependent metabolism of metoprolol [17] and dextromethorphan [15] in their respective DGI PBPK models, were analyzed using OLS regression (polynomial of degree 2, no intercept). For the paroxetine, atomoxetine and risperidone models, CYP2D6 k cat values were optimized for studies, which reported plasma concentration-time profiles of populations with two wildtype alleles (AS = 2) and were set to 0 for poor metabolizers of CYP2D6 (AS = 0) as they were assumed to show no CYP2D6 activity [3]. Subsequently, k cat values for all other modeled activity scores were calculated using the polynomial equation obtained from the OLS regression of metoprolol and dextromethorphan k cat values. Here, CYP2D6 K M values as well as CYP2D6 reference concentrations were kept constant over the whole range of modeled activity scores.

DGI Model Evaluation
To evaluate the performance of the presented DGI models, as well as the implemented scale of CYP2D6 activity score-dependent metabolism derived from the published metoprolol and dextromethorphan PBPK DGI models, predicted plasma concentration-time profiles were plotted alongside their respective observed data. Plasma concentration-time profiles for populations with variant activity scores were compared to profiles of a population with normal activity (AS = 2) in studies reporting activity scores or genotypes, whereas plasma concentration-time profiles for variant phenotypes were compared to those of the extensive metabolizer phenotype, where only CYP2D6 phenotypes were reported. Furthermore, predicted DGI PK ratios (AUC last and C max ratios) (Equation (5)) were evaluated for study  (2)) for the predicted PK ratios.
DGI PK ratio = PK DGI PK reference (5) where PK DGI = AUC last or C max of either a variant activity score or a variant phenotype; PK reference = AUC last or C max of either AS = 2 or the extensive metabolizer phenotype, respectively. Additionally, steady-state exposures (AUC ss ) of model compounds were predicted for different CYP2D6 activity scores. Here, simulations were performed for individuals with different activity scores after multiple oral doses of 40 mg paroxetine, 40 mg atomoxetine, or 2 mg risperidone.

PBPK Base Model Building
A total of 57 plasma concentration-time profiles obtained from 29 published clinical trials were used for the development and the evaluation of the paroxetine, atomoxetine, and risperidone PBPK models and are summarized in Table 1. Clinical study tables providing comprehensive information such as individual and population demographics (sex, age, weight, and height) and CYP2D6 activity (phenotype, genotype, or activity score) as well as the assignment of the study to the respective dataset are presented in Sections S2.1.2, S3.1.2 and S4.1.2 of Supplementary Materials S1 for paroxetine, atomoxetine, and risperidone, respectively.
For the paroxetine PBPK model, a total of 33 plasma concentration-time profiles where paroxetine was administered as an intravenous infusion (four profiles) or orally in single (16 profiles) or multiple (13 profiles) doses were used to develop the paroxetine PBPK model. Here, administered doses ranged from 10 to 70 mg of paroxetine. The paroxetine PBPK model incorporates CYP2D6-and CYP3A4-dependent metabolism of paroxetine [22] as well as irreversible inhibition of CYP2D6 and CYP3A4 [23]. Additionally, an unspecific hepatic clearance process and renal elimination via passive glomerular filtration were included. A schematic overview of implemented paroxetine metabolic pathways is provided in Figure 2a. Drug-dependent model parameters for paroxetine are presented in Section S2.1.1 of Supplementary Materials S1. The atomoxetine PBPK model was developed using 12 plasma concentration-time profiles after oral administrations of atomoxetine in single (nine profiles) and multipledose administrations (three profiles) with doses of administered atomoxetine ranging between 20 and 50 mg. The atomoxetine PBPK model includes metabolism via CYP2D6 and CYP2C19 [24] as well as a passive glomerular filtration process. Figure 2b depicts the pathways implemented in the atomoxetine model. Atomoxetine drug-dependent model parameters are presented in Section S3.1.1 of Supplementary Materials S1.
An overview of risperidone model pathways as published by Kneller et al. [21] is provided in Figure 2c. Risperidone and 9-hydroxyrisperidone drug-dependent parameters are listed in Section S4.1.1 of Supplementary Materials S1.  , and an unspecific clearance process [22]. Moreover, paroxetine is a mechanism-based inhibitor of CYP2D6 and CYP3A4 resulting in an irreversible auto-inhibition of paroxetine metabolism [23,55]. Paroxetine metabolites were not included as model compounds. (b) Atomoxetine is metabolized via CYP2D6 and CYP2C19 [24]. Atomoxetine metabolites were not included as model compounds. (c) Risperidone is metabolized to its active metabolite 9-hydroxyrisperidone via CYP2D6 and CYP3A4 [56]. Moreover, other metabolites are formed via CYP2D6-and CYP3A4-mediated metabolism [56]. 9-Hydroxyrisperidone is metabolized via an unspecific hepatic clearance process. Other risperidone metabolites were not included in the model. CLhep: unspecific hepatic clearance, and CYP: cytochrome P450.

PBPK Base Model Evaluation
The three presented models could accurately predict the plasma concentrations for their respective model compounds. A representative selection of plots displaying predicted compared to observed plasma concentration-time profiles for paroxetine (a-f), atomoxetine (g-i) and risperidone and its metabolite 9-hydroxrisperidone (j-l) is shown in Figure 3. . Implemented metabolic pathways for the modeled compounds. (a) Paroxetine is metabolized via CYP2D6, CYP3A4, and an unspecific clearance process [22]. Moreover, paroxetine is a mechanism-based inhibitor of CYP2D6 and CYP3A4 resulting in an irreversible auto-inhibition of paroxetine metabolism [23,55]. Paroxetine metabolites were not included as model compounds.
(b) Atomoxetine is metabolized via CYP2D6 and CYP2C19 [24]. Atomoxetine metabolites were not included as model compounds. (c) Risperidone is metabolized to its active metabolite 9hydroxyrisperidone via CYP2D6 and CYP3A4 [56]. Moreover, other metabolites are formed via CYP2D6-and CYP3A4-mediated metabolism [56]. 9-Hydroxyrisperidone is metabolized via an unspecific hepatic clearance process. Other risperidone metabolites were not included in the model. CL hep : unspecific hepatic clearance, and CYP: cytochrome P450.
Additionally, plots displaying predicted compared to observed plasma concentrationtime profiles of the three compounds alongside their respective GMFEs for AUC last

Local Sensitivity Analysis
Local sensitivity analyses were performed using simulations after oral administrations of the respective standard doses for paroxetine (20 mg), atomoxetine (40 mg) and risperidone (2 mg). Parameters with associated sensitivity values >0.5 (100% parameter value perturbation resulting in a >50% change of predicted AUC) were considered sensitive.

DGI Model Building
An OLS regression of CYP2D6 k cat, rel values was performed for the published metoprolol and dextromethorphan models to derive a substrate-independent scale of activity score-dependent metabolism for the newly developed models of paroxetine, atomoxetine, and risperidone. The results of the OLS regression are displayed in Figure 4.

Local Sensitivity Analysis
Local sensitivity analyses were performed using simulations after oral administrations of the respective standard doses for paroxetine (20 mg), atomoxetine (40 mg) and risperidone (2 mg). Parameters with associated sensitivity values >0.5 (100% parameter value perturbation resulting in a >50% change of predicted AUC) were considered sensitive. Here, lipophilicity (literature value) and fu (literature value) were sensitive parameters for the paroxetine model. Lipophilicity

DGI Model Building
An OLS regression of CYP2D6 kcat, rel values was performed for the published metoprolol and dextromethorphan models to derive a substrate-independent scale of activity score-dependent metabolism for the newly developed models of paroxetine, atomoxetine, and risperidone. The results of the OLS regression are displayed in Figure 4.  [15,17]. The solid line represents the regression curve (degree = 2, intercept = 0), and symbols represent kcat, rel values for the different activity scores. AS: activity score, kcat, rel: kcat relative to AS = 2, and R 2 : coefficient of determination.  [15,17]. The solid line represents the regression curve (degree = 2, intercept = 0), and symbols represent k cat, rel values for the different activity scores. AS: activity score, k cat, rel : k cat relative to AS = 2, and R 2 : coefficient of determination.
Input values and results of the regression analysis as well as CYP2D6 model k cat values for the modeled activity scores (calculated using the Equation given in Figure 4) are shown in Table 2. Here, paroxetine, atomoxetine, and risperidone CYP2D6 model k cat values were calculated by multiplying interpolated k cat, rel values with the optimized baseline k cat value (activity score 2).

DGI Model Evaluation
The newly developed DGI models were evaluated using clinical studies, which stratified their subjects by CYP2D6 activity score or phenotype. These studies either provided the activity score for the investigated population, the CYP2D6 phenotype, or CYP2D6 genotypes of all study participants. Simulations were performed using the corresponding k cat values with respect to activity score ( Table 2)  Overall, predicted DGI AUC last and C max ratios were in good agreement with observed DGI ratios, highlighting the good performance of the DGI models predicting the activity score-dependent metabolism of paroxetine, atomoxetine, and risperidone, with 22/24 DGI AUC last and 22/24 C max ratios within the prediction success limits proposed by Guest et al. [58] as depicted in Figure 6. The predicted DGI AUC last ratios showed mean GMFEs of 1.37, 1.25 and 1.11 whereas the overall GMFEs of predicted DGI C max ratios were 1.33, 1.28 and 1.16 for paroxetine, atomoxetine, and risperidone, respectively. Predicted to observed DGI AUC last and C max ratios for all studies are presented in Sections S2.3.3, S3.3.3, and S4.3.3 of Supplementary Materials S1.
served DGI ratios, highlighting the good performance of the DGI models predicting the activity score-dependent metabolism of paroxetine, atomoxetine, and risperidone, with 22/24 DGI AUClast and 22/24 Cmax ratios within the prediction success limits proposed by Guest et al. [58] as depicted in Figure 6. The predicted DGI AUClast ratios showed mean GMFEs of 1.37, 1.25 and 1.11 whereas the overall GMFEs of predicted DGI Cmax ratios were 1.33, 1.28 and 1.16 for paroxetine, atomoxetine, and risperidone, respectively. Predicted to observed DGI AUClast and Cmax ratios for all studies are presented in Sections S2.3.3, S3.3.3, and S4.3.3 of Supplementary Materials S1. Simulations of steady-state plasma concentration-time profiles and AUCss values in individuals with different activity scores after multiple oral doses of 20 mg paroxetine, 40 mg atomoxetine or 2 mg risperidone and a comparison of the corresponding AUCss values are given in Figure 7. Predicted DGI AUCss ratios were 3.4, 13.6 and 2.0 for poor metabolizers (activity score 0) compared to normal metabolizers (activity score 2) for paroxetine, atomoxetine and risperidone active moiety (risperidone + 9-hydroxyrisperidone), respectively. Conversely, predicted DGI AUCss ratios were 0.2, 0.5 and 0.95 for ultrarapid metabolizers (activity score 3).

Discussion
In this study, whole-body PBPK models of paroxetine, atomoxetine, and risperidone, including its active metabolite 9-hydroxyrisperidone, are presented. A total of 57 studies were used for model building and evaluation. CYP2D6 DGIs were modeled by implementing CYP2D6 activity score-dependent metabolism of the respective compounds for

Discussion
In this study, whole-body PBPK models of paroxetine, atomoxetine, and risperidone, including its active metabolite 9-hydroxyrisperidone, are presented. A total of 57 studies were used for model building and evaluation. CYP2D6 DGIs were modeled by implementing CYP2D6 activity score-dependent metabolism of the respective compounds for activity scores ranging from 0 to 3. Parameters for the CYP2D6 activity score-dependent metabolism for the presented models were derived from previously published models [15,17] via interpolation and represent a substrate-independent approach of modeling CYP2D6 DGIs. All three models showed good performance as highlighted in the model evaluation sections.
Previously published PBPK models of paroxetine, atomoxetine and risperidone typically implemented CYP2D6 DGIs by adjusting model parameters such as CL int , K M or k cat values based on traditional CYP2D6 phenotypes (extensive and poor metabolizer) [59,60] or specific CYP2D6 genotypes such as CYP2D6*1/*1 and CYP2D6*10/*10 [43], whereas the presented model can accurately describe both traditional phenotypes as well as CYP2D6 activity scores, allowing the models to predict compound plasma concentrations for all relevant genotypes to provide a finer graduation of CYP2D6 activity [17].
The CYP2D6 activity score-dependent metabolism was modeled by adjusting CYP2D6 k cat values based on the activity score of the respective individual or population. Hence, the k cat value serves as a surrogate parameter, reflecting changes in both in vivo reference concentration [61] and in vitro V max [62] that typically occur due to polymorphisms in the CYP2D6 gene. Here, K M and CYP2D6 reference concentrations were fixed over the whole range of modeled activity scores and phenotypes [18]. Model CYP2D6 k cat values for extensive metabolizers were consistently lower compared to normal metabolizers (activity scores 1.25-2, see Table 2 and Sections S2.1.1, S3.1.1 and S4.1.1 of Supplementary Materials S1 for paroxetine, atomoxetine and risperidone, respectively). Specifically, the CYP2D6 k cat values for extensive metabolizers were 35%, 30% and 36% lower compared to activity score 1.25 for paroxetine, atomoxetine and risperidone, respectively. This is caused by the limited intersection between the aforementioned activities, mainly due to the often arbitrary definition of the extensive metabolizer phenotype [16].
While the presented approach of modeling CYP2D6 DGIs based on the activity score categories was a necessary simplification, it also represents one of the limitations of the presented study and, by extension, the activity score system itself. As suggested by van der Lee et al., CYP2D6 DGIs may be more accurately described using a continuous scale approach, reflecting the effect of CYP2D6 allelic variants on the pharmacokinetics of CYP2D6 substrates in vivo and in vitro compared to the activity score system [9]. Additionally, certain CYP2D6 alleles have been described to display substrate-specific effects in vitro and in vivo. For instance, the CYP2D6*17 allele, classified as a reduced function (activity score 0.5) allele, shows increased activity in risperidone metabolism when compared to the wildtype *1 allele [63]. These effects are not considered in the classification of alleles using the activity score system [61]. Regardless, current clinical guidelines by the Dutch Pharmacogenetics Working Group (DPWG), the Clinical Pharmacogenetics Implementation Consortium (CPIC) and other major institutions in this field, are based on the activity score system [5]. Moreover, a allele-specific modeling approach would drastically increase model complexity and would require an extensive amount of in vitro and in vitro model input data [18]. Hence, the approach presented in this study was deemed a more practical choice in the context of PBPK modeling. The presented paroxetine model includes metabolism via CYP2D6 and CYP3A4 and an unspecific hepatic clearance pathway as a surrogate pathway for metabolism via other CYP enzymes that were reported to metabolize paroxetine in vitro [22]. Here, additional experimental in vitro data such as K M and V max values were available for paroxetine metabolism via CYP1A2, CYP2C19 and CYP3A5. However, these enzymes were described to have a smaller effect on paroxetine kinetics compared to CYP2D6 and CYP3A4 [22]. Furthermore, CYP3A4 was implemented to describe the effect of auto-inhibition via CYP3A4 on the pharmacokinetics of paroxetine [23,55], especially in poor metabolizers of CYP2D6. No metabolite of paroxetine was implemented in the model due to a lack of reported metabolite plasma concentrations in the published literature, presumably due to the chemical and metabolic instability of major metabolite paroxetine-catechol [22,35]. Furthermore, paroxetine has been suggested as a substrate of P-glycoprotein (P-gp) in the published literature [39,64]. However, while a moderate affinity of paroxetine to P-gp was observed in in vitro experiments [65], genetic polymorphisms in the ABCB1 gene were described to have no significant on paroxetine plasma concentrations in vivo [66]. Hence, the authors did not implement active transport via P-gp in the model. Regardless, the model was able to describe paroxetine plasma concentrations for all doses (20-70 mg) in published clinical studies.
CYP2D6 has been described to be mainly responsible for atomoxetine metabolism, as atomoxetine AUC was found to be increased by 400% in poor metabolizers of CYP2D6 compared to extensive metabolizers [45]. In the presented PBPK model, atomoxetine metabolism was described by implementing CYP2D6 and CYP2C19. Although 4-hydroxyatomoxetine has been reported to be primarily formed via CYP2D6, only a total of four plasma concentration-time profiles of 4-hydroxyatomoxetine were reported in the published literature [41,67]. Thus, the metabolite was not explicitly modeled. However, as more clinical studies reporting 4-hydroxyatomoxetine plasma concentrations become available, the presented PBPK model of atomoxetine can be extended to cover the formation of 4-hydroxyatomoxetine.
While CYP1A2, CYP2B6, CYP2C9 and CYP3A4 were also described to contribute to the metabolism of atomoxetine in vitro, their relative contribution to atomoxetine depletion was found to be far smaller compared to CYP2C19 [24]. Hence, CYP2C19 serves as a surrogate pathway for multiple CYP enzymes in the presented model. While CYP2C19 is also polymorphically expressed, and CYP2C19 DGIs in CYP2C19 have been described in the literature [68], they were considered negligible, as CYP2D6 accounts for approximately 90% of atomoxetine oral clearance in normal metabolizers of CYP2D6 [45] and the CYP2C19 k cat value was below the sensitivity threshold for atomoxetine model parameters (see Section S3.2.3 of Supplementary Materials S1).
To describe the activity score-dependent metabolism of risperidone, an established parent-metabolite model was used [21] and adapted. The model includes metabolism via CYP2D6 and CYP3A4 for risperidone as well as active transport via P-gp for both risperidone and its active metabolite 9-hydroxyrisperidone [69].
Simulations of steady-state plasma concentrations for the modeled compounds revealed that, although k cat, rel values implemented in the respective DGI models for the different activity scores were the same, AUC ss values behaved differently between the three compounds. For instance, AUC ss DGI ratios for CYP2D6 poor metabolizers (activity score 0) were 3.4, 13.6 and 2.0, whereas AUC ss DGI ratios for ultrarapid metabolizers (activity score 3) were 0.2, 0.5 and 0.95 for paroxetine, atomoxetine and risperidone, respectively. Here, different model-specific factors might influence predicted AUC ss DGI ratios. For risperidone, the total active moiety was considered (risperidone + 9-hydroxyrisperidone). Here, a decrease in the risperidone AUC with increasing activity scores typically infers an increase in 9-hydroxyrisperidone AUC, partially compensating the effect of CYP2D6 DGIs on the AUC of the total active moiety, and the overall pharmacodynamic effect of risperidone, as 9-hydroxyrisperidone has similar activity compared to risperidone [70]. Conversely, the paroxetine model includes auto-inhibition via mechanism-based inhibition of CYP2D6 and CYP3A4, reducing the differences between DGI AUC ss values for the different modeled activity scores.

Conclusions
This study presents whole-body PBPK models of paroxetine, atomoxetine, and risperidone. The models implement CYP2D6 activity score-dependent metabolism informed from previously published PBPK models of CYP2D6 substrates and have been successfully used to predict the plasma concentrations of their model compounds both in non-DGI and DGI scenarios with various CYP2D6 activity scores. The final PBPK model files will be made publicly available in the Clinical Pharmacy Saarland University GitHub model repository (http://models.clinicalpharmacy.me/). Due to the mechanistic implementation of human physiology and important pharmacokinetic pathways, the models allow for knowledgebased scaling to special populations and can serve as the basis for future investigations of CYP2D6 drug-drug-gene interaction (DDGI) scenarios.