The impact of genetic polymorphisms on the pharmacokinetics of efavirenz in African children

Aims Using a model‐based approach, the efavirenz steady‐state pharmacokinetics in African children is characterized, quantifying demographic and genotypic effects on the drug's disposition. Simulations are also conducted allowing prediction of optimized doses of efavirenz in this population. Methods We modelled the steady‐state population pharmacokinetics of efavirenz in Ugandan and Zambian children using nonlinear mixed‐effects modelling. Individual mid‐dose efavirenz concentrations were derived and simulations explored genotype‐based dose optimization strategies. Results A two‐compartment model with absorption through transit compartments well described 2086 concentration‐time points in 169 children. The combined effect of single nucleotide polymorphisms (SNPs) 516G>T and 983T>C explained 44.5% and 14.7% of the variability in efavirenz clearance and bioavailability, respectively. The detected frequencies of composite CYP2B6 genotype were 0.33 for 516GG|983TT, 0.35 for 516GT|983TT, 0.06 for 516GG|983TC, 0.18 for 516TT|983TT, 0.07 516GT|983TC and 0.01 for 516GG|983CC. The corresponding estimated clearance rates were 6.94, 4.90, 3.93, 1.92, 1.36, and 0.74 l h−1 for a 15.4 kg child and median (95% CI) observed mid‐dose concentrations 1.55 (0.51–2.94), 2.20 (0.97–4.40), 2.03 (1.19–4.53), 7.55 (2.40–14.74), 7.79 (3.66–24.59) and 18.22 (11.84–22.76) mg l−1, respectively. Simulations showed that wild‐type individuals had exposures at the bottom of therapeutic range, while slower metabolizers were overexposed. Conclusions Dosage guidelines for African children should take into consideration the combined effect of SNPs CYP2B6 516G>T and 983T>C.


Introduction
Efavirenz is a non-nucleoside reverse transcriptase inhibitor (NNRTI) commonly used in first-line antiretroviral treatment (ART) for HIV-1 infected adults and children over 3 years old [1,2]. Due to its ease of dosing (the long half-life allows once daily administration), proven efficacy, ability to be used with anti-TB drugs, and availability of cheap generic formulations, it is especially widely used in Africa.
Several investigators have reported a high proportion of subtherapeutic efavirenz concentrations in children, highlighting the need for optimization of paediatric dosing guidelines [26][27][28]. The aim of this analysis was therefore to characterize the steady-state PK of efavirenz in the largest cohort of African children reported so far, quantifying demographic and genotypic effects on efavirenz disposition, and thus allowing prediction of optimized doses of efavirenz in this population.

Methods
PK and other data from two studies in African children from Uganda and Zambia were pooled together: CHAPAS-3 [29] (Children with HIV in Africa -Pharmacokinetics and Adherence/Acceptability of Simple antiretroviral regimens) and ARROW [28,30] (Anti-Retroviral Research for Watoto).

CHAPAS-3
Efavirenz was dosed once daily, either in the morning or at night, following modified WHO 2010 guidelines (Table 4), using a new paediatric double-scored 600 mg efavirenz tablet (provided by Cipla Pharmaceuticals, India) that can be split into two or three parts enabling administration of doses of 200, 300, 400 or 600 mg.
All included patients took part in sparse PK sampling on clinic visits at weeks 6, 36, and every 24 weeks thereafter until the end of the study. The self-reported time of the last dose was recorded. Additional intensive PK sampling was conducted in the first patients enrolled in each WHO weight band at week 6 and in all patients who acquired tuberculosis during the study (4-10 weeks after tuberculosis treatment initiation and again 4-10 weeks after tuberculosis treatment cessation). Children in this intensive PK substudy were advised to take efavirenz in the morning for 6 weeks prior to sampling and the drug intake on the PK day was observed. Plasma was separated and stored at À80°C until transportation on dry ice for drug concentration assay.
Plasma efavirenz concentrations from the intensive PK were assayed using ultra high-performance liquid chromatography (HPLC) at the Department of Clinical Pharmacy of Radboud University Nijmegen Medical Centre, Nijmegen, the Netherlands. The method produced linear results over the range of 0.0517 to 15.51 mg l À1 . The lower limit of quantification was 0.05 mg l À1 . The intraassay and interassay coefficients of variation (CV) were 1.01-5.31% and 0.1-1.63%, respectively. Relative error of the method ranged from 97.7% to 104.5% [31]. Plasma efavirenz concentrations from sparse PK were determined by liquid chromatography tandem mass spectrometry (LC-MS/MS) in the Division of Clinical Pharmacology, University of Cape Town, South Africa. The method was accurate over the range of 0.0195-20 mg l À1 . The lower limit of quantification was 0.0195 mg l À1 . The interassay CV and residual error (RE) were 3.59-5.78% and À2.57-0.92%, respectively, and the intraassay CV and RE were 1.50-9.67% and À6.00-4.28%, respectively [32].
Both laboratories participate in international quality assurance and proficiency testing schemes and are expected to have comparable standards, although no cross-validation was performed for the assays. Systematic differences between the labs and assays were tested as covariates in the PK model (see Population pharmacokinetic analysis/Covariates).

Arrow
Efavirenz was dosed once-daily, either in the morning or at night, according to modified WHO 2006 paediatric recommendations. The tested formulations included 50, 100 and 200 mg capsules; and half or whole 600 mg tablets (provided by the national ART programme in Uganda).
Children in a PK substudy [28] were sampled on two occasions: 36 and 40 weeks after starting ART. Eligible children were advised to take efavirenz in the morning for 4 weeks prior to sampling and drug intake on the PK day was observed. Samples were stored and assayed at the Department of Clinical Pharmacy of Radboud University Nijmegen Medical Centre, Nijmegen using the same methods as described above.
The distribution of the genotypes was evaluated for compliance with Hardy-Weinberg equilibrium using the exact test conducted using R-package "genetics" [35].

Population pharmacokinetic analysis
Model building. The steady-state efavirenz PK was analysed using nonlinear mixed-effects modelling with software NONMEM VII (version 7.3) [36] and the first-order conditional estimation method with interaction. PsN 4.4.0, Pirana and Xpose were used to facilitate modelling and for model diagnostics [37]. The model was developed and validated in accordance with standard methods described in the literature [38]. For the structural model, one-, two-, and three-compartment models with first-order absorption and elimination were tested, as well as time lag or transitcompartment absorption [39] and hepatic first-pass model [40]. Between-subject and -occasion variability (BSV, BOV) were tested on PK parameters assuming lognormal distribution. Residual unexplained variability (RUV) was tested using a combined proportional and additive error. Data below level of quantification (BLQ) were included in the analysis by imputing half of the lower limit of quantification (LLOQ) of the corresponding assay as suggested in Beal et al. (M6 method) [41]. Implausible samples and PK profiles were identified using extreme values of CWRESI and their exclusion was evaluated based on visual checks.
Model development and covariate selection was guided by the NONMEM objective function value (OFV), inspection of goodness-of-fit (GOF) plots and visual predictive checks (VPCs), biological plausibility and clinical relevance. OFV (proportional to À2 log-likelihood of the data) was assumed to be χ 2 -distributed and a drop of 3.84 or more between two hierarchical models after inclusion of one additional parameter (df = 1) was considered a significant improvement (P = 0.05). Stability and robustness of the final model, together with precision of its parameter estimates, was evaluated through a nonparametric bootstrap (n = 200).
Intensive and sparse data were included in the model development process in a stepwise manner as suggested in Svensson et al. [42], starting with intensive PK data from CHAPAS-3, followed by the intensive data from ARROW, and finally the sparse PK data from CHAPAS-3.
The model-derived empirical Bayesian estimates for the individual parameters were used to predict steady-state middose concentrations (measures 12 h after dose) for each sampling occasion and patient.
Covariates. Allometric scaling was added to the model at an early development stage as previously suggested [43]. The effect of maturation of metabolic pathways on PK parameters was tested using post-menstrual age (gestationadjusted age) as a predictor. Both a power function or a sigmoidal model with and without Hill coefficient were tested [43]. Besides weight and age, the other covariates tested were: tuberculosis co-treatment, study site, nucleoside reverse transcriptase inhibitor (NRTI) backbone, sex, weight-for-age Z-score (WAZ) and height-for-age Z-score (HAZ), drug formulation, the effect of splitting tablets used in CHAPAS-3 (inferred from total daily dose) and genotype information (SNPs listed above). The potential differences between assays and lab procedures for the quantification of drug concentrations were tested in the model as proportionality and correction factors on RUV.
Missing genotype values were imputed using mixture modelling with frequencies fixed to those observed in the rest of the cohort as previously suggested by Keizer et al. [44].
Simulations. The final model was used to simulate exposures after administration of efavirenz with the formulation given in CHAPAS-3 and using a dataset of subjects with a uniform distribution of weights ranging from 10 to 40 kg, in 0.1 kg steps (300 individuals simulated 100 times). Several dosing strategies were explored. To avoid generating implausibly extreme values, the maximum variability for each random effect was limited to three standard deviations. Data were analysed and plots generated using R [35].

Demographic results and samples
This analysis included data from 128 children from CHAPAS-3 and 41 children from ARROW. Relevant subject characteristics including the genotype frequencies for the tested SNPs are presented in Table 1. All tested genotypes were in Hardy-Weinberg equilibrium (HWE). For SNPs rs35599367, rs3826711 and CYP2B6*29, all patients were homozygous for the common allele and HWE was not calculated. The genotype information was missing for five children from ARROW and two from CHAPAS-3, who were assigned by the mixture model as follows: two as 516GG|983TT, four as 516GT|983TT, and one as 516TT|983TT.
From CHAPAS-3, 61 intensively sampled PK profiles (a total of 474 samples) and 510 sparse PK profiles (1002 samples, 1-2 per occasion) were available. The PK data were collected from 6 weeks after starting efavirenz up to a maximum of 132 weeks. There were up to 7 PK sampling visits per child. Of 14 children who acquired tuberculosis, nine had at least one intensively sampled PK profile on efavirenz with tuberculosis treatment. The only BLQ measurement from intensive PK substudy was a pre-dose measurement and all eight samples from that PK visit for that patient were excluded from the analysis as it was deemed likely not to be in steady state due to poor adherence. Within the sparse data, 15 samples were BLQ and were included by imputing half LLOQ, i.e. 0.00975 mg l À1 . From the ARROW study, 611 intensive PK samples from 82 PK visits (two visits per patient) were available. Data from one visit were discarded due to an implausible PK profile, possibly caused by mismatch of samples. No samples were BLQ.

Population pharmacokinetics
The data were best described using a two-compartment model with first-order elimination and transit compartment absorption [39]. Final parameter estimates, their precision (obtained through a bootstrap) and statistical significance for the inclusion of the covariate and random effects (based on drop in OFV) are presented in Table 2. The PK parameters were estimated relative to oral bioavailability whose typical value was fixed to one due to lack of intravenous data. Adequate fit of the model was confirmed by a GOF plot and VPC (see Supplementary Figures S1 and S2).
The effect of body size on all clearance and volume parameters was accounted for using allometric scaling, which significantly improved model fit (18-point drop in OFV) [43]. No effect of age on the maturation of clearance could be detected. After adjusting for body size, the main predictor of clearance was the effect of CYP2B6 genotype, categorized into six subgroups based on the combined effect of 516G>T and 983T>C SNP variants present in our population (Tables 2, 3). CYP2B6 genotype explained 44.5% and 14.7% of BSV in clearance and oral bioavailability respectively. Exclusion of individuals with missing genotype did not have a significant effect on final results.
The absorption rate constant (k a ) and the absorption mean transit time (MTT) were 1.6-fold larger and 1.4 times longer in ARROW compared with CHAPAS-3. Splitting of the new double-scored efavirenz tablets used in CHAPAS-3 was not found to affect efavirenz bioavailability. No other covariate (see Methods) was found to significantly improve the model fit. We did not detect any systematic differences between the assays and labs employed in the analysis.
The model fit was markedly improved by inclusion of a correction parameter to allow for larger residual unexplained variability for all samples obtained after self-recorded efavirenz intake. This includes all predose intensive PK samples and all the sparse PK samples from CHAPAS-3. The residual variability for those samples was twice as large as in the rest of the data.

Simulations
Simulations were performed to predict exposures in African children based on their weight and genotype, when dosed according to the regimen used in CHAPAS-3. Median mid-dose concentrations were comparable across weight bands but noticeable differences were observed between the CYP2B6 genotype subgroups ( Figure 1, Table 3).
A dose optimization strategy for African children was devised by categorizing subjects into four phenotypic subgroups based on their composite genotype vector 516G>T| 986T>C, similarly to Dooley et al. [32] (presented in Table 3). The proposed dose adjustment between metabolic subgroups is based on optimal ratios of 1:0.66:0.33:0.1 for EM: IM:SM:USM (extensive, intermediate, slow and ultra-slow metabolizers), respectively, and is outlined in Table 4.
The predicted exposures based on the suggested doseoptimization algorithm are presented in Figure 2 and Supplementary Table S1. The suggested dosing approach ensured adequate drug exposure in all simulated weight bands ( Figure 2, left panel), and reduced the differences due to metabolic status ( Figure 2, right panel).

Discussion
Efavirenz pharmacokinetics in Africans has previously been shown to be affected by the combined effect of SNPs 516G>T and 983T>C [8,12,14,17,32]. The current investigation confirms those findings and is the first analysis to quantify the effect of the CYP2B6 516G>T|983T>C SNP vector on efavirenz clearance in African children using nonlinear mixed-effects modelling. The use of modelling provides a tool to concomitantly account for multiple effects such as genotype and weight, and a platform to derive a dose adjustment strategy based on these effects.
Numerous studies in adults and children have reported a significant effect of CYP2B6 516G>T on efavirenz clearance. Our analysis shows that presence of one variant allele in 516G>T causes clearance to drop by 34%, while the reduction reaches 72% for homozygous mutants, which is in line with previously reported reductions of 20-47% and 58-80%, respectively [13,20,24,26,[45][46][47]. Our findings also show that the effect of SNP 516G>T is significantly modified by the 983T>C (i.e. in wild type 516G>T individuals presence of a single variant allele in 983T>C causes a 43% Efavirenz pharmacogenomics in African children drop in clearance and 89% if no functional allele is present), confirming associations found by Holzinger et al. [8] and in a number of African studies [12,14,32]. This polymorphism is virtually absent in individuals of European ancestry [8,9] and was not detected in a study of Cambodian patients [24]. The combined effect of the CYP2B6 516G>T|983T>C SNP vector on efavirenz clearance has previously been reported distinguishing four phenotypic subgroups [32]. In our study, we were able to further characterize this effect using six CYP2B6 516G>T|983T>C SNP variants. Similarly to previous reports we show that, despite its low prevalence, SNP 983T>C is not only a significant predictor of efavirenz clearance, but it is responsible for a larger drop in metabolic rate than 516G>T (i.e. 29% drop in clearance in 516G>T heterozygote vs. 43% drop in 983T>C heterozygote, when no other polymorphisms were present) [8,12]. No further significant genetic associations were detected. A genome-wide association study by Holzinger et al. identified rs4803419 as another significant polymorphism in CYP2B6 affecting efavirenz clearance [8]. The effect of SNP rs4803419 becomes significant only for homozygous mutants who are wild type for 516G>T and 983T>C and no such variants were present in our population. Although this finding was recently replicated in South African patients, the investigators concluded this effect was negligible in comparison to 516G>T and 983T>C [12].
Results from the simulations ( Figure 1) showed that, even though the dosage guidelines tested in CHAPAS-3 result in average mid-dose concentrations within target range of 1.0-4.0 mg l À1 [3], the effect of CYP2B6 genotype leads to large differences within each weight band. In particular, children with slower CYP2B6 genotypes (516GG|983CC, 516GT|983TC and 516TT|983TT) were overexposed, while the fastest metabolizers (516GG|983TT) achieved exposures at the bottom of the therapeutic range. Over 20% of children in the study with the 516GG|983TT genotype had efavirenz concentrations below the proposed minimum target concentration of 1.0 mg l À1 (Table 3). Moreover, our model indicates that disregarding the effect of the 983T>C SNP and basing dose optimization only on 516G>T could lead to exposures significantly higher than the therapeutic range [3] in~14% of African patients with 983TC or 983CC genotypes ( Table 3 in green). This suggests that genotype-based dose optimization in African children should take into account both 516G>T and 983T>C SNPs.
The only available guidelines on genotype-adjusted paediatric dosage were recently formulated for patients under 3 years of age by the Panel on Antiretroviral Guidelines for Adults and Adolescents and are currently being tested in the IMPAACT study P1070 [10,11]. The dose-adjusted strategy was developed based on results of an analysis by Salem et al.    Data are population median (5th-95th percentile) or number (percentage). Based on all patients from ARROW and CHAPAS-3 trials (missing genotype estimated by mixture model). *CL refers to typical population value estimated by the model for a patient with median weight of 15.4 kg (combined ARROW and CHAPAS-3 data). **CHAPAS-3 data only (due to differences in dosing between studies, see Methods). †Value for each PK visit estimated by the model, multiple measurements used to calculate geometric mean for every patient which were then used to calculate median and percentiles for each subgroup. MET, metabolic subgroup [32]: EM (extensive metabolizers), 516GG|983TT; IM (intermediate metabolizers), 516GG|983TC or 516GT|983TT, SM (slow metabolizers), 516TT|983TT or 516GT|983TC; USM (ultra-slow metabolizers), 516GG|983CC. The green shading indicates groups of patients who would be significantly overexposed if dose optimization were conducted based only on SNP 516GT.
exposures above the therapeutic range [3] in a large proportion of patients who are either heterozygous for 516G>T or wild type with 983TC or 983CC genotypes (Figure 3). Our recommendations were simplified to the previously described four metabolic subgroups determined by the composite genotype vector 516G>T|983T>C outlined in Table 3 [32]. The clearance between EM:IM:SM:USM drops as follows: 1:0.6:0.24:0.1, which is remarkably similar to the ratios detected by Dooley et al. in African adults (1:0.6:0.26:0.08, respectively) [32]. Our dosage algorithm presented in Table 4 (1:0.66:0.33:0.1) was adjusted to maximize the use of currently available solid formulations. A similar dose adjustment pattern (1:0.66:0.33) was previously successfully implemented based on phenotypic differences in an adult study by Mello et al. [50].
The results of the few dose reduction studies guided by CYP2B6 genotype conducted in developed countries highlighted improved treatment tolerability and costeffectiveness in adults [50][51][52][53]. These results were confirmed by recent cost-effectiveness analysis of this practice in American adults [54]. It could be speculated that in resourcelimited settings, cost and logistical challenges would make implementation of such practice difficult. Nonetheless, decreasing price and broader availability of genotyping technology and economic development open future scenarios in which genotype-based dosing approaches may be economically viable and beneficial even in developing countries. The IMPAACT P1070 study, which is currently being conducted in HIV-infected infants and children, should give more insight into the practical implications of genotyping in lowand middle-income countries.
Similarly to previous paediatric studies, the average clearance value (before inclusion of genotype effect) was higher than   findings in adults: 14.34 L h À1 vs. 7.5 to 11.7 L h À1 (both after scaling with allometry up to 70 kg) [5,21,24,45,46]. This is consistent with reports that clearance in children of age 1-4 years exceeds adult values [27,46,55]. In keeping with previously published paediatric models, the effect of size on clearance and volume was explained through allometric scaling [23,43,46]. Unlike Salem et al., we did not observe age-related maturation of clearance, but the previous analysis showed that 90% of maturation was reached by the age of 9 months [23] and the majority of patients in the current study were >3 years. We detected significant differences in absorption parameters between CHAPAS-3 and ARROW, possibly related to the use of different formulations (tablets in the former and mostly capsules in the latter). The formulations were assumed to be bioequivalent and indeed no formulation effect on bioavailability was detected.
Due to the availability of data from multiple sampling occasions within the same patient, it was possible to characterize both BSV and BOV in the PK parameters. Large BOV was identified for absorption parameters and bioavailability. Drug absorption is widely known to be a variable phenomenon, depending on occasion-specific factors, such as food intake, gastric emptying times and GI tract pH. In the current study, other factors may have contributed to inflating BOV (in particular bioavailability), including differences between actual and self-reported intake times for the sparse data, lack of accurate intake history before the last dose, lack of information on accompanying food consumption. The fact that   the information about intensively sampled occasions was more accurate was accounted for with the introduction of a scaling factor on residual unexplained variability (2-fold larger for sparse data). The current study had several limitations. As mentioned, the dosage timing and food co-administration was not recorded beyond the last intake. Previous studies showed that efavirenz PK was affected by adherence and food effects [2,21,25]. Polymorphisms in accessory pathways including CYP2A6 and CYP3A4 or UGT were not assessed and the effect of CYP2B6 785A>G was only evaluated in patients from the ARROW study. Despite several reports suggesting that efavirenz metabolism is affected by polymorphisms in those pathways, the genome-wide association study by Holzinger et al. showed that their effect was significantly less dramatic than for 516G>T and 983T>C [8].
Furthermore, no PK/PD relationship was explored in this analysis for either efficacy or toxicity, but currently accepted therapeutic ranges were used as cut-offs guiding dose optimization. These targets were generated in an adult cohort and have recently been brought into question, suggesting that lower efavirenz concentrations might be sufficient to provide viral suppression [17]; however, no alternative has been suggested to date.
Lastly, our study was underpowered to determine the effect of tuberculosis treatment on efavirenz concentrations, but recent findings suggest that the inducing effect of rifampicin on clearance is counterbalanced by a concentrationdependent inhibitory effect of isoniazid that could explain contradictory conclusions from previous studies [7,24,32].

Conclusions
Our study suggests that genotype-adjusted efavirenz dosage in African children should be based on the composite 516G>T|983T>C SNP vector, due to significant modification of clearance rates caused by SNP 983T>C genotype, whose prevalence in Africans is much higher than in other populations. Using nonlinear mixed-effects modelling, we quantified this effect and suggest that a dose optimization algorithm 1:0.66:0.33:0.1 (EM:IM:SM:USM, respectively) would provide more balanced drug exposures between individuals with a different metabolic status while maximizing the potential of using the new double-scored efavirenz tablets tested in the CHAPAS-3 study. The findings warrant further studies evaluating the genotype-based dosing approach and the feasibility of genotyping in resource-limited settings. Janssen, ViiV and Tandem Nano, as well as personal fees from Merck was issued a patent "Compositions of efavirenz". No other support from any organization for the submitted work; no financial relationships with any organizations that might have an interest in the submitted work in the previous 3 years and no other relationships or activities that could appear to have influenced the submitted work are to be declared for any of the authors. The Division of Clinical Pharmacology at the University of Cape Town would like to gratefully acknowledge Novartis Pharma for their support of the development of pharmacometrics skills in Africa.

Competing Interests
The content is solely the responsibility of the authors and does not necessarily represent the official views of any funders.

Supporting Information
Additional Supporting Information may be found in the online version of this article at the publisher's web-site: http://onlinelibrary.wiley.com/doi/10.1111/bcp.12934/suppinfo.

Table S1
Simulated values of mid-dose concentrations obtained after suggested dose optimization and proportions of patients of less than 1 mg l À1 , between 1 mg l À1 and 4 mg l À1 , and more than 4 mg l À1 Figure S1 Goodness of fit plots. Top left, observations vs. population predictions; top right, observations vs. individual predictions; bottom left, conditional weighted residuals vs. time after dose; bottom right, absolute values of individual weighted residuals vs. individual predictions Figure S2 Visual predictive check of the final model by 516GT|983TC SNP vector in semi-log scale. Hollow points, observations; red solid line, median of observed data; red line with breaks, 5th and 95th percentile of observed data; orange fill area, 95% CI of simulated median; blue fill area, 95% CI of simulated 5th and 95th percentile