Development, Early Verification and Application of a North American Pediatric PBPK Model, Comparison With the North European Pediatric Population

ABSTRACT

Physiologically based pharmacokinetic (PBPK) models are increasingly used in pediatric drug development and matching such models to the target population is an important consideration. A US pediatric PBPK model was developed incorporating four sub‐groups describing physiology and enzymatic/transporter levels for White, African American, Hispanic/Latino and Asian subjects. Virtual populations could then be created reflecting the demographics of the clinical study design. Ten different drugs (14 studies) were used to verify the pediatric PBPK model performance. Predicted PK parameters showed reasonable agreement with observed data, 50% of the predicted: observed values were within 0.8–1.25‐fold, 74% within 0.67–1.5‐fold, and 97% within 0.5–2‐fold. There was minimal difference for most drugs in performance between the mixed US and North European pediatric populations. This was partly due to reported studies having low numbers of subjects and under‐representation of some ethnicities. Power calculations showed that to detect a pharmacokinetic difference for the CYP3A4/5 substrates midazolam, sirolimus and tacrolimus between African American and White American children with a power of 80%, 25, 18 and 12 subjects respectively, in each group would be needed. Developing adult and pediatric PBPK models to closely match study populations is important due to their increased regulatory use and with the move to their application in personalized medicine. Such models may also help to bridge some of the ethnic differences not captured in pediatric clinical studies due to the low numbers of subjects.

Study Highlights

• What is the current knowledge on the topic?

  • ○
    A number of different adult and pediatric PBPK models have been built to represent specific populations including North European, Japanese, and Chinese. Recently, an adult US PBPK model incorporating White, African American, Hispanic, and Asian sub‐groups has been described and verified.
• What question did this study address?

  • ○
    Whether a pediatric US PBPK model incorporating the sub‐populations could be established, verified, and applied to clinical situations.
• What does this study add to our knowledge?

  • ○
    For certain key drugs where underlying genetic and biochemical changes between ethnic sub‐groups result in altered drug PK, it is important that these changes are incorporated into pediatric PBPK models. Overall predictive performance of the model was good and indicates that some clinical PK studies may be under‐powered to detect differences between population sub‐groups.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    Application of the model to study design suggests that for certain drugs (midazolam, sirolimus, and tacrolimus) over‐recruitment from certain ethnic groups may be necessary to generate sufficient study power to detect a difference in PK.

1. Introduction

The importance of understanding the contribution of ethnic differences in drug pharmacokinetics (PK) and how these may impact on dose and drug–drug interaction (DDI) liability is well known [1]. The USA is an increasingly diverse population; latest census data indicates that 57.8% are white, 18.7% Hispanic or Latino, 12.1% African American, and 11.4% other ethnicities [2]. There are known ethnicity related differences in PK within the US population, a notable example is tacrolimus where 93% of African American compared to 17% of White Americans express the CYP3A5*1 gene resulting in a need for higher doses in the former subjects [3]. Given the increasing conduct of international clinical trials, the US Food and Drug administration (FDA) has an expectation that any drug trials include data that are representative of the US patient population [4]. They have issued draft guidance for industry recommending that sponsors should include a race and ethnicity diversity plan as part of new investigational drug application [5]. Regulators, including the FDA, the European Medicines Agency (EMA), and the Japanese Pharmaceutical and Medical Devices Agency (PMDA), have published details on bridging drug efficacy across different populations based on the International Conference on Harmonization (ICH) guideline E5: ‘Ethnic Factors in the Acceptability of Foreign Clinical Data’ [6].

Applying model informed drug development (MIDD) tools such as physiologically based pharmacokinetic (PBPK) models has been advocated for bridging between ethnicities. Several adult PBPK models have previously been reported for North European [7], Japanese [8], Chinese [9] and a mixed US populations [10]. These models are ideally suited for this role as they consider the complex interplay between physiological parameters and drug‐related characteristics and can be used to predict the PK of drugs in subjects with differing genetic backgrounds and ethnicities.

Consideration of ethnic differences is also important in pediatric clinical drug trials but increased ethical constraints and difficulties in recruitment make this more difficult to achieve especially in the younger age groups [11]. Extending PBPK models for different ethnicities into the pediatric population is possible by incorporating physiological developmental changes, including the ontogeny of specific enzymes and transporters. Pediatric PBPK models have been developed for the North European [12], Japanese [13], and Chinese [14] populations.

To address drug pharmacokinetic differences due to ethnicity the main aim of this study is to create four North American pediatric sub‐population PBPK models based on NHANES database classification: White, non‐Hispanic (White), African American, Hispanic‐Latino including Mexican (Hispanic), and Asian. Creating the separate populations allows for flexibility in virtual clinical trial design and close matching of the virtual study populations with those used in the clinical trials. This manuscript describes the development of these populations, their verification against observed data and three case studies demonstrating their clinical and research application.

2. Methods

A summary of the workflow for this study is presented in Figure 1. Both the existing adult US and North European pediatric models were used as a starting point for US pediatric PBPK model development. All the observed data used for model building, comparison, and application refers to American children unless stated otherwise.

FIGURE 1.

Summary of US pediatric population development, verification and application. The light blue shading represents existing models and data and the dark blue, new population models and data. AA, African American; AAG, alpha‐1 acid glycoprotein; AsA, Asian American; DDI, drug–drug interaction; GFR, glomerular filtration rate; His, Hispanic/Latino; HSA, human serum albumin; HT, height; W, White American; WT, weight.

2.1. US Pediatric Population Model Development

Physiological and biochemical data were obtained from the literature (Pubmed, EBSCO) and from the NHANES database [15] with additional height and weight data obtained from the WHO growth charts [16] for the 0 to 1 year age range assuming no ethnic difference. For each physiological parameter, a search strategy similar to that used in compiling the adult US population [10] was applied with additional terms pediatric, neonate, infant, child, and adolescent. The existing equations for the North European Pediatric population (Sim‐Pediatric) were evaluated against the observed US pediatric physiological data for the four sub‐populations. Where the fit could be improved, the observed data were used to develop new age‐related equations using the Solver function in Microsoft Excel 365 (Microsoft, Redmond, WA, USA). Values were optimized based on reducing the sum of squares and visual predictive checks. Where no specific data were available or where no corresponding changes had been made to the adult US PBPK populations for example, tissue volumes and blood flows, the Sim‐Pediatric equations were assumed. Further information is provided under Table S4.

The phenotype and abundance data for CYP and UGT enzymes and for transporters in each of the four adult US populations has recently been presented [10] and this data was used in the US pediatric populations. Based on a recently published model qualification for the Chinese pediatric population in children 9–14.5 years [14], the hepatic expression values for CYP2B6 in the American Asian pediatric population have been revised to 24 ρmol/mg microsomal protein. The expression of intestinal enzymes CYP3A4 and CYP3A5 in all four US pediatric populations was left the same as for the North European population at 65.4 and 17.8 nmol/average adult small intestine; the same ontogeny function as previously described [12] was also applied.

The equations describing the ontogeny of enzymes and transporters were left the same as the North European pediatric population (Sim‐Pediatric) [12] except for CYP2C19, where the Upreti ontogeny [17] was applied based on recent experience in simulating pantoprazole pharmacokinetics in US pediatrics [18].

2.2. Drug Study Selection

A literature search was performed in PubMed to identify clinical pharmacokinetic studies performed in the US pediatric population, search terms included ‘Drug’, pediatric, neonate, infant, child, adolescent, US population, African American, Hispanic, Latino White, Asian. Age‐related pharmacokinetic parameter data was extracted from each publication including AUC, C_max, and CL. Where available, concentration‐time data was extracted using GetData Graph Digitizer v2.22. Preferred studies gave details of race and ethnicity of the participants allowing this to be replicated in the simulations; but where this information was missing, population average 2021 US census values were assumed [2]. The drugs and studies identified are shown in Table S1.

2.3. PBPK Model Performance Verification

All model building, verification, and application steps were completed using V24 of the Simcyp population‐based simulator. Most of the drugs studied were existing compound files within the Simcyp simulator. The input values for the remaining drugs and sources of data are shown in Table S2. For morphine and pantoprazole, the compound files were exactly as previously published; for sirolimus and tacrolimus, some modifications were made as listed in Table S2, and consequently, some further verification for these revised files was undertaken. The main metabolic route for all drugs is shown in Table S3.

For each drug and study the simulation trial design was matched as closely as possible to the clinical study in terms of route of administration, ethnic mix of the study (White, African American, Hispanic, Asian), phenotype mix in population, age range, number of subjects in study (n), number of trials (such that n*trials was > 100 subjects), fraction of females, dose, duration of study and study units (Table S1). The multiple population trials design feature was used to ensure the correct ethnic mix in the study for the US sub‐groups; where this was not stated, the average % proportions 58% White, 20% Hispanic, 14% African American and 8% Asian were applied.

Simulations performed in the mixed US population were also run in the North European pediatric population (Sim‐Pediatric) with the same trial design but ignoring ethnic mix to allow comparison of results between the two approaches.

2.4. PBPK Model Application

Three application examples were included in this analysis. The first application example was to use the developed US pediatric PBPK model to investigate the power of some of the included studies to detect a likely difference between the different US population groups for example, for drugs partly metabolized by CYP3A5 including tacrolimus, sirolimus and midazolam. Simulations were run with an enriched trial design based on 100 subjects in each sub‐group. Separate power calculations were performed between the White and African American populations for these drugs to determine the minimum number of subjects to detect a difference in AUC with 80% power.

The second was to replicate a population pharmacokinetic study by Sadhasivan et al. [19] showing a 1.23‐fold increase in morphine clearance in African American compared to White children. The initial simulation included separate populations of 113 White and 29 African American but subsequently separate populations of 100 in each were run; other details are the same as in Table S1. A sensitivity analysis was performed on OCT1 expression in the African American population to capture the clearance difference seen in the clinical study.

The final application example was to replicate the clinical data presented by Minnick et al. [20] comparing the maintenance doses of tacrolimus between the African American and White subjects in pediatric heart transplant populations aged 7–18 years. The trial design is shown in Table S1, the initial dose was set to a 0.07 mg/kg 12 hourly dose, and the dose was compared between both populations based on the recommended therapeutic range for C_min (9–12 ng/mL). The dose in the White population was then adjusted so that C_min fell within the therapeutic range.

2.5. Data Analysis

Predictive performance was evaluated by visual inspection of the simulated concentration‐time profiles and 5th and 95th percentiles with the observed data and by comparing the PK parameters, AUC, C_max, and CL with observed clinical results for each drug and study. PK parameter predictions were classified by fold error: 0.5 to 2‐fold; 0.67 to 1.5‐fold; and 0.8 to 1.25‐fold. Overall predictive accuracy was assessed for both the US mixed and NEC pediatric populations based on average fold error (AFE) Equation (1) and average absolute fold error (AAFE) Equation (2) as measures of bias and precision, respectively.

$$\mathrm{AFE}={10}^{1/\mathrm{N}\sum \log \left(\frac{\text{Predicted}\mathrm{PK}}{\text{Observed}\mathrm{PK}}\right)}$$ (1)

$$\text{AAFE}={10}^{1/\mathrm{N}\sum \left|\log \left(\frac{\text{Predicted}\mathrm{PK}}{\text{Observed}\mathrm{PK}}\right)\right|}$$ (2)

Further details on how power calculations were performed within the PBPK model are shown in the Supporting Information.

3. Results

3.1. Development and Verification of the US Pediatric Systems Data

The derived age‐related equations for the four US pediatric sub‐populations are shown in Table S4; these included new equations for HT with age, WT with HT and age, hematocrit (HCT) with age below 1 year, and alpha‐1 acid glycoprotein (AAG) for the Asian population. All other equations for these parameters, including human serum albumin (HSA), were left the same as the Sim‐Pediatric (North European) population. The performance of the HT and WT equations with age for males and females in the four US populations is shown in Figures S1A and S1B, and likewise for haematocrit, HSA, and AAG in Figures S2A and S2B. Note there are gaps in some of the NHANES observed data for HSA and AAG. Overall, the equations with applied variability appear to adequately describe the observed data.

The glomerular filtration rate in the simulated adult US populations is calculated from creatinine values generated within the PBPK model; the same values are applied in the pediatric populations for subjects ≥ 15 years. The predicted and observed creatinine values are shown for the four populations in males and females in Figure S3. The GFR in the pediatric subjects < 15 years is calculated using the previously published formula based on body surface area [12], for the African American male population, an arbitrary correction factor of 0.8 was applied to the pediatric formula to ensure a smooth transition with adult data. The interface at 15 years of age between the pediatric and adult calculated GFR values generated in the four populations and males and females is shown in Figure S4. For the drugs selected for this study, a summary of the enzyme and transporter phenotype frequency and hepatic expression are shown in Table S5; this is based on the adult data [10].

3.2. Model Verification

A total of 10 drugs (14 studies) were included in the model verification. Adult verification data is presented in Figure S5 A/B for tacrolimus and C/D for sirolimus; overall, the observed data is captured within the 5th and 95th percentiles of the predictions.

Based on prediction of concentration‐time data and PK parameter data, overall performance of the mixed US pediatric population model was good. The concentration‐time profiles for studies presenting observed data are shown in Figure 2, most of the observed values are within the 5th and 95th percentiles for the predictions, the oral midazolam concentration data is slightly under‐predicted. The observed and predicted PK parameters are shown in Table 1, 50% of the P:O PK parameter values were within 0.8 to 1.25‐fold, 74% were within 0.67 to 1.5‐fold, and 97% within 0.5 to 2‐fold. The only value not within 2‐fold was the C_max for ruxolitinib, which was overpredicted, but this could be due to concurrent disease (Cancer patients).

FIGURE 2.

Predicted plasma concentration‐time profiles in the US pediatric population for oral atomoxetine, Age 7–14 years (A), oral montelukast, Age 6–8 years (B), IV pantoprazole 0.8 mg/kg, Age 2–14 years (C), IV pantoprazole 1.6 mg/kg, Age 2–14 years (D) IV midazolam, Age 0.5–12 years (E), IV sildenafil, Age 0.8–5.3 years (F) oral midazolam, Age 0.5–16 years (G), IV morphine, Age 4–14.4 years (H), oral sirolimus (blood), Age 3–18 years (I), oral ruxolitinib, Age 2.4–21.4 years (J). The open circles are observed data, the solid black line is the mean prediction and gray lines are 5th and 95th percentiles for the prediction.

TABLE 1.

Summary of the predicted results from the mixed US Pediatric PBPK population compared to observed values.

Drug	Study	Route	SD/MD	Observed data			Predicted data			P:O data
				AUC	Cmax	CL	AUC	Cmax	CL	AUC	Cmax	CL
Atomoxetine	Witcher et al. [21]	O	SD MD	645 2250	144 584	0.46 0.48	750 2276	146 582	0.53 0.48	1.16 1.23	1.02 0.997	1.16 1.01
Efavirenz	Starr et al. [22]	O	MD1 MD2	218 244			320 349			1.47 1.43
Efavirenz	Saitoh et al. [23]	O	MD3 MD4 MD5 MD6			5.81 7.00 5.70 3.01			3.9 4.9 3.6 2.1			0.68 0.69 0.63 0.70
Midazolam	Reed et al. [24]	IV	SD	281		0.62	283	292	0.6	1.01		0.97
Midazolam	Reed et al. [24]	O	SD	137	55.6	2.40	217	95.6	1.9	1.58	1.72	0.79
Montelukast	Knorr et al. [25]	O	MD	2928	495		2604	484		0.89	0.98
Morphine	Sadhasivan et al. [19]	IV	SD			1.41			1.3			0.94
Morphine	Emoto et al. [26]	IV	SD			1.83			1.7			0.92
Pantoprazole	Kearns et al. [27]	IV	SD7 SD8	7160 14,320	5700 10,320	0.20 0.20	9525 19,049	6011 12,022	0.12 0.12	1.33 1.33	1.05 1.16	0.60 0.60
Pantoprazole	Kearns et al. [27]	O	SD	7741	2970	0.26	7884	2140	0.15	1.02	0.72	0.59
Ruxolitinib	Loh et al. [28]	O	MD	6.72	1.26	11.5	8.8	2.8	13.4	1.31	2.18	1.17
Sildenafil	Hill et al. [29]	IV	SD	533		0.64	526		0.67	0.99		1.05
Sirolimus	Emoto et al. [30]	O	MD			18			18.5			1.03
Tacrolimus	Zheng et al. [31]	O	MD9 MD10		33.1 65.4			18.8 111			0.57 1.70

Note: 1 = 11.7 mg/kg; 2 = 14.2 mg/kg; 3 = mixed genotype; 4 = CYP2B6 GG; 5 = CYP2B6 GT; 6 = CYP2B6 TT; 7 = 0.8 mg/kg; 8 = 1.6 mg/kg; 9 = EM, 10 = PM. SD refers to single dose and MD to multiple dose studies.

The performance of the NE pediatric PBPK population in predicting the US observed data is shown in Table S6 and was similar to the US pediatric population predictions. 51% of the P:O PK parameter values were within 0.8 to 1.25‐fold, 74% were within 0.67 to 1.5‐fold, and 97% within 0.5 to 2‐fold. The two populations are compared directly in Table 2. The AFE for the P:O AUC and CL values was 1.01 and 1.05 for the US and NE populations, respectively, and the AAFE was 1.24 in both populations.

TABLE 2.

Comparison of the predicted to observed AUC and Clearance ratios between the mixed US population and the north European pediatric populations.

Drug	Study	Route	SD/MD	P:O US pediatric		P:O NE pediatric
				AUC	CL	AUC	CL
Atomoxetine	Witcher et al. [21]	O	SD MD	1.16 1.23	1.16 1.01	1.03 1.10	1.36 1.22
Efavirenz	Starr et al. [22]	O	MD1 MD2	1.47 1.43		1.37 1.30
Efavirenz	Saitoh et al. [23]	O	MD3		0.68		0.67
Midazolam	Reed et al. [24]	IV	SD	1.01	0.97	1.03	0.95
Midazolam	Reed et al. [24]	O	SD	1.58	0.79	1.44	0.89
Montelukast	Knorr et al. [25]	O	MD	0.89		0.99
Morphine	Sadhasivan et al. [19]	IV	SD		0.94		0.97
Morphine	Emoto et al. [26]	IV	SD		0.92		0.88
Pantoprazole	Kearns et al. [27]	IV	SD4 SD5	1.33 1.33	0.60 0.60	1.64 1.64	0.60 0.60
Pantoprazole	Kearns et al. [27]	O	SD	1.02	0.59	1.31	0.77
Ruxolitinib	Loh et al. [28]	O	MD	1.31	1.17	1.35	1.08
Sildenafil	Hill et al. [29]	IV	SD	0.99	1.05	1.05	0.97
Sirolimus	Emoto et al. [30]	O	MD		1.03		0.97

Note: 1 = 11.7 mg/kg; 2 = 14.2 mg/kg; 3 = mixed genotype; 4 = 0.8 mg/kg; 5 = 1.6 mg/kg. SD refers to single dose and MD to multiple dose studies. The data was compared in terms of average fold error (AFE) for precision and absolute average fold error (AAFE) for bias. Based on combined AUC and CL data the AFE was 1.013 and 1.049 in the US and NEC pediatric populations respectively whilst AAFE was 1.236 and 1.241.

3.3. Model Application

The first application example Investigates the difference in CYP3A5 phenotype frequency (gut and hepatic) and expression within the US population using enriched populations. Results are shown in Figure 3 for midazolam, sirolimus and tacrolimus. For sirolimus and tacrolimus the drugs were administered orally, the geometric mean sirolimus AUC_last and C_max were 127.5 ng/mL.h and 18.6 ng/mL in the White American compared to 89 ng/mL.h and 11.9 ng/mL in the African American, the fraction escaping gut metabolism (Fg) was 0.39 and 0.36 in the White and African American populations, respectively. The geometric mean tacrolimus AUC_last and C_max were 235.8 ng/mL.h and 64.3 ng/mL in the White American compared to 147.6 ng/mL.h and 40.6 ng/mL in the African American, the Fg was 0.19 and 0.16 in the White and African American populations, respectively. All three drug files include metabolism by CYP3A5 and as shown in Table S5 the number of extensive metabolizer (EM) in the African American population is higher (72.5%) compared to White subjects (14.3%) with Hispanic and Asian being 43 and 48% respectively. Despite the slightly lower hepatic expression for African American (78.6 ρmol/mg) compared to White (108 ρmol/mg) when the enzyme is expressed there is a higher fraction of the drug metabolized by this enzyme in African American compared to White populations resulting in higher clearance (pie charts Figure 3), Hispanic and Asian populations appear to have intermediate CL. For the African American population, the higher frequence of CYP3A5 EM in the gut has also contributed to the lower AUC and C_max values compared to the White population. Using the data analysis power calculation tool in Simcyp the number of African American and White subjects needed in each population for the study to detect a difference at 80% power for midazolam, sirolimus and tacrolimus were 25, 18 and 12, respectively (Figure 3 bottom graphs). This largely reflects the CYP3A5 fm for these substrates in each population related to enzyme expression and phenotype.

FIGURE 3.

Comparison of the concentration‐time profiles between the different North American sub‐groups for drugs partially metabolized by CYP3A5, iv Midazolam in plasma, age 0.5–12 years (A) oral sirolimus in blood, age 3–18 years (B) and oral tacrolimus in blood, age 6.3–13.1 years (C). The median fraction metabolized by CYP3A5 and CYP3A4 and other minor pathways between the White and African American populations are shown for each corresponding drug in the pie charts A1, B1 and C1. The results from the PBPK power calculations (100 subjects in each White, African American, Hispanic and Asian pediatric groups) for each drug are shown in A2, B2 and C2, the black line is the power calculation and dashed lines show the number in each group needed for 80% power.

The second application example is in Figure 4 showing the simulated concentration‐time profiles for morphine between the White and African American (100 subjects in each group) as well as a modified African American pediatric population. The faster clearance in the African American compared to White populations is due to higher numbers of extensive OCT1 transporter phenotype subjects at 82% compared to 53%. Despite this difference, there was only a simulated 1.03 increase in CL (L/kg/h) in the African American compared to the White subjects, whereas the clinical study reported a 1.23‐fold change [19]. Sensitivity analysis suggested that a 2‐fold increase in relative OCT1 abundance was needed in African American subjects to recover this change, and when this value was implemented in the simulations it resulted in a 1.2‐fold higher clearance compared to the White population.

FIGURE 4.

Application examples comparison of morphine iv plasma concentration time profile between White and African American pediatric subjects (Age 6–15.9 years) using the default populations (dashed black and gray lines) and a modified African American population with doubled OCT1 activity (solid black line).

The final application example is shown in Figure 5 where the reported steady state dose of tacrolimus post heart transplant is investigated in African American and White pediatric subjects and in relation to recommended trough blood concentrations between 9 and 12 ng/mL. The clinical study [20] reports final doses of 0.07 and 0.045 mg/kg every 12 h in African American and White, respectively; the simulations support these dose levels, and administering 0.07 mg/kg/12 hourly to the White pediatric population resulted in simulated mean blood trough concentrations of 18.9 ng/mL, which are above the therapeutic range.

FIGURE 5.

Simulated tacrolimus blood concentration time profiles in the white American, dose = 0.07 mg/kg (A) = 0.045 mg/kg (B) and African American, dose = 0.07 mg/kg (C) pediatric populations (Age 3–18 years) based on the mean reported long term doses following heart transplant. Doses from Minnick et al. [20]. Orange and blue lines are upper (12 ng/mL) and lower (9 ng/mL) therapeutic range.

4. Discussion

The increased use of both adult and pediatric PBPK models for regulatory submissions [32, 33] and the move towards application in personalized medicine [34, 35] reinforce the need to match the simulated individuals with that of the population. The regulatory qualification of these models must be in relation to the context of use with the expectation that if a pediatric study was conducted in a specific population, then the underlying model would reflect this population [36]. Ethnicity and age‐associated differences in physiology, genetics, and biochemistry can result in altered drug pharmacokinetics [1, 14]. Because all these differences can be captured within PBPK models, these models represent an approach to extrapolate between adult and pediatric ethnic groups or for supplementing observed clinical trial data with simulations where recruitment within a particular sub‐group within a population is difficult [37, 38]. The US population is increasingly diverse, and with clinical drug trials becoming more global, the US Food and Drug administration (FDA) has an expectancy that any such drug trials include data that are representative of the US patient population; this includes for pediatric medicines [39].

This study presents the information required to build a mixed US pediatric PBPK population that includes the main ethnic groups; however, data available to verify system parameters within each ethnic group are generally limited. The changes to CYP, UGT, and transporter phenotype and hepatic expression are key factors that result in altered PK between the different US population groups; these are based on detailed literature searches performed whilst developing the adult US populations [10]. In this study, the Asian value for hepatic CYP2B6 has been updated based on prediction of efavirenz PK data in the older pediatric Chinese population [14] and is now set to the same values for the adult North European population at 21.6 ρmol/mg microsomal protein.

Overall model performance for the mixed US populations, across the age range 0.5 to 18 years, was good with 97% of predicted PK parameters within 2‐fold of the observed values. For certain drugs such as montelukast no PK difference is expected between the different pediatric US population groups as CYP2C8 hepatic expression and phenotype are the same. For other drugs for example, efavirenz and pantoprazole some PK differences may be expected due to the different phenotypic mix between ethnic populations and in the case of pantoprazole higher CYP2C19 expression in the American Asian population than other sub‐groups. For drugs such as midazolam, sirolimus and tacrolimus differences in PK are expected to exist between the ethnic US sub‐groups due to the involvement of CYP3A5 and relatively large phenotypic differences in the frequency of CYP3A5 expression between them [40]. Detecting any differences in PK in the different ethnic US populations is dependent on study design in terms of subject recruitment. For the sirolimus study [30] 23 of the 27 subjects were white and just 4 were African American, enriched power calculation simulations (100 subjects in White and African American groups) in application example 1, suggested that a minimum of 18 subjects would be needed in each group to detect a difference at 80% power. In general (with some exceptions) the US pediatric studies identified for model qualification in this study do not always reflect the ethnic mix seen in the general US population and some are under‐powered to detect any differences. In agreement with a recent study [38] there is general under‐representation of certain ethnic groups with lack of subjects in the African American, Asian and Hispanic groups in the reported clinical studies. In some cases, the ethnic mix is not reported in the study, and the US population mean distribution was assumed in the simulations.

For most drugs investigated in this study, pediatric US and NEC PBPK model (based on > 82% white) predictions are very similar on a population level both in terms of AFE and AAFE. As already stated, this is expected in certain cases where little difference exists in enzyme phenotype and expression between the populations, for example, CYP2C8. In these cases, regulators can generally accept results from either PBPK model with no need to extrapolate clinical trial results between the two populations [4]. In other cases, differences in drug PK between the US and NE populations may be expected depending on the mix of subjects included in the US population; in these cases, further data or use of MIDD models may be needed to fill in the gaps.

A morphine application example is presented. A study by Fukuda et al. 2013 [41] showed that OCT1 genotype plays a significant role in iv morphine PK and that relatively high frequencies of defective OCT1 variants among Caucasians may explain their lower morphine clearance by around 20%, CL was 1.26 L/min in OCT1 wild type *1*1 compared to 1.02 in OCT1 homozygotes for example, *2*2 or *5*5.

This has been further explored by Emoto et al. [26] who used a PBPK model to show the effect of OCT1 genotype on morphine PK in children, 3 to 4‐fold lower transporter activity (J_max) was assigned to *2/*2 and *2/*3 genotyped subjects compared to heterozygotes *1/*2, *1/*3 and wild type *1/*1 subjects in order to replicate the measured PK data.

Morphine PK differences due to race and genetic variation are further explored in a study by Sadhasivan et al. [19]. When equal number of White US and African American subjects (100 each) are simulated the differences in morphine PK can be seen due to the implemented OCT1 phenotype changes in the PBPK models [41, 42, 43] but the overall change in morphine clearance within the clinical study is not recovered (1.23‐fold increase in African American). To recover this difference the hepatic OCT1 expression had to be doubled. No literature evidence was found suggesting that OCT1 or UGT2B7 expression (within the same genotype) may be increased in African American compared to White subjects and further research is needed here. This example shows an advantage of the PBPK approach in that enriched population simulations can be performed to explore potential ‘what if’ differences that may not be apparent in clinical trials with relatively low numbers of subjects.

In the third application example the dose of tacrolimus in a White and African American US pediatric population to achieve a C_trough between 9 and 12 ng/mL was investigated. In agreement with the clinical study [20] the dose needed in the two populations was 0.045 and 0.07 mg/kg 12 hourly; this reflects the different CYP3A5 phenotypes between the two populations. In this case, according to our simulations the clinical study was adequately powered to detect a difference between the two populations with 18 White and 12 African American subjects.

There are several study assumptions and limitations related to data availability both for verification of certain system parameters and the overall models. The WHO growth charts were used to define height and weight below 1 year of age assuming no difference related to ethnicity; more research is needed to confirm or refute this assumption, but studies on growth related skeletal parameters suggest no ethnic differences in this age range [44]. Measured GFR values across the age bands in each of the US sub‐populations using pure markers such as inulin are not available; this is particularly important for the male African American population where a correction factor of 0.8 was applied to the existing pediatric BSA based GFR equation (valid from term birth to 15 years) [12] to ensure a smooth transition to adult values. This may reflect the use of the CKD‐EPI creatinine fit without race for the adult African American populations [10] and further research is needed. Model verification is needed on a broader range of drugs both renally cleared, metabolized by other enzymes or where transporters are involved, and across the whole pediatric age range particularly including neonates and infants.

The ontogeny of all the enzymes and transporter is assumed to be the same as for the North European population because insufficient data was available to derive ontogeny related to US ethnicity. A significant number of the published ontogeny studies are from US‐based researchers [45, 46] and so more likely reflect the mixed US population. Although the details of clinical samples are often given including ethnicity, low numbers mean they are usually pooled in determining age‐related changes. Further research is needed on the influence of ethnicity on enzyme and transporter ontogeny. Moving forward, consideration should be given to study design and powering US pediatric studies adequately to properly investigate known or suspected PK differences between the ethnic groups as well as age.

As verification of our model versus clinical data was scarce on neonatal data, our conclusions should be carefully weighed with respect to (pre‐term) neonates and infants until 6 months of age. A major information gap remains on the interplay of age and genotype, especially in the youngest infants when neither enzyme and transporter activity nor GFR are fully matured. Maturation may obscure genotype influences, resulting in the disappearance of a genotype–phenotype relationship [47]. Also, alternative excretion pathways may need to be considered, for example, caffeine, which is almost completely metabolized in adults and older children but renally cleared in neonates [48].

In conclusion, a US pediatric population has been developed and some verification undertaken with 10 drugs (14 studies) eliminated by various pathways. There is a potential application for US pediatric PBPK models to aid in bridging between the different population groups where sufficient recruitment in a particular sub‐group may not always be possible [38]. In addition, the models may also have application in study design, by running enriched population groups to define the power of a study needed to see pharmacokinetic differences. There is clearly a need to expand the number of drugs assessed in the model and ultimately to undertake formal qualification of the population covering the major elimination routes across the whole pediatric age range; this will ensure confidence in the models. These efforts are ongoing and, as recently expressed [49] should be seen as a community effort involving academia, drug industry, model provider, and regulators.

Author Contributions

All authors wrote the manuscript. Trevor N. Johnson, Fiskani Kondowe, Iain Gardner, Joyce van der Heijden, and Karen Rowland Yeo designed the research, Fiskani Kondowe, Trevor N. Johnson, and Joyce van der Heijden performed the research. Fiskani Kondowe, Joyce van der Heijden, Karen Rowland Yeo, and Trevor N. Johnson analyzed the data.

Funding

Joyce van der Heijden was funded by the Bill & Melinda Gates Foundation (INV‐023795). Fiskani Kondowe was funded by the EU H2020 Marie Sklodowska‐Curie Innovative Training Networks (ITN) grant Dohartnet (H2020‐MSCAITN‐2018‐812660).

Conflicts of Interest

Trevor N. Johnson, Iain Gardner, and Karen Rowland Yeo are all employees of Certara Predictive Technologies. All other authors declare no conflicts of interest.

Supporting information

Data S1: Supporting information—Power calculations.

Table S1: Summary of the US clinical studies and trial designs for the verification examples included in this study.

Table S2: Input parameters for the non‐default Simcyp compound files.

Table S3: Summary of the elimination pathways for each drug file based on simulations in 100 healthy North European adults (20–50 years).

Table S4: Summary of the algorithms used to describe the Simcyp pediatric US populations.

Table S5: Phenotype and expression data for the four US sub‐populations focussing on the enzymes and transporters involved in drug disposition of the compounds included in the study.

Table S6: Summary of the predicted results from the North European Pediatric PBPK population (Sim‐Pediatric) compared to observed values.

Figure S1A: Performance verification of the age height (top line) and age weight (bottom line) for the male and female African American and White American populations. The light gray symbols are the predicted values, and the black symbols are observed values from the 2022/23 NHANES database.

Figure S1B: Performance verification of the Age height (top line) and Age weight (bottom line) for the male and female Hispanic‐Latino and Asian American populations. The light gray symbols are the predicted values, and the black symbols are observed values from the 2022/23 NHANES database.

Figure S2A: Performance verification of the haematocrit with age (top line), albumin with age (middle line) and alpha 1‐acid glycoprotein (AAG) with age (bottom line) for the male and female African American and White American populations. The light gray symbols are the predicted values, and the black symbols are observed values from the 2022/23 NHANES database.

Figure S2B: Performance verification of the haematocrit with age (top line), albumin with age (middle line) and alpha 1‐acid glycoprotein (AAG) with age (bottom line) for the male and female Hispanic‐Latino and Asian American populations. The light gray symbols are the predicted values, and the black symbols are observed values from the 2022/23 NHANES database.

Figure S3: Performance verification of the creatinine concentration with age (15–25 years) in the male and female African American and White American populations (top line) and male and female Hispanic‐Latino and Asian American populations (Bottom line). The light gray symbols are the predicted values, and the black symbols are observed values from the 2022/23 NHANES database.

Figure S4: Performance of the Glomerular filtration rate (GFR) with age (0–25 years) in the male and female African American and White American populations (top line) and male and female Hispanic‐Latino and Asian American populations (Bottom line). The gray symbols are the predicted values showing the interface of the pediatric GFR equation with the creatinine calculated values from 15 years of age.

Figure S5: Performance verification for the blood concentration time profiles of in adult subjects. (A) Mancinelli et al. tacrolimus 5 mg oral dose, (B) Moller et al. tacrolimus 0.05 mg/kg oral dose, (C) Brattstrom et al. Sirolimus 1 mg/m² oral dose, (D) Leung et al. sirolimus 42 mg oral dose. Black lines are the mean simulations and the gray lines are 5th and 95th percentiles, white filled circles are the observed data.