Study design and participants
The clinical trial was carried out in six African centres: three in Tanzania, two in Burkina Faso and one in Kenya. Written informed consent from a parent/guardian was required to enrol children younger than 5 years in the trial, who were infected by P. falciparum, as confirmed by microscopy (density between 2000 and 200,000 asexual parasites/µL), and with fever equal to or higher than 37.5 °C. Exclusion criteria were children with body weight less than 5 kg, signs of severe/complicated malaria, febrile conditions caused by diseases other than malaria, a known hypersensitivity to the study drugs, a mixed plasmodium infection, a history of anti-malarial treatment in the 2 weeks preceding the trial or 4 weeks in case of mefloquine and piperaquine, prior participation in a therapeutic trial within 3 months or inability to tolerate oral medication. Patients were followed up to day 63 after start of treatment or to the first recurrence of infection. The study protocol was reviewed and approved by national and independent ethics committees of all participating centres.
Of the 945 patients enrolled in the trial, 473 were randomized to the ASMQ arm (one of them was never dosed) and 472 were randomized to the AMLF arm. The pharmacokinetic analysis described here was performed on the 472 patients who received ASMQ.
Administered doses for these patients were one or two dispersible tablets containing 25 mg AS and 55 mg MQ once a day for three consecutive days to children aged from 6 to 11 months and from 12 to 59 months, respectively. Clinical and parasitological examinations were scheduled at baseline, i.e. before drug administration, at day 0 (D0), D1, D2, D3, D7, D14, D21, D28, D35, D42, D49, D56 and D63 and on any other day if the patient spontaneously returned and parasitological reassessment was required (as per protocol). A margin of ± 2 days to the assigned day of visit was allowed from D7 onward. In case of recurrence of parasitaemia on D7, D14, D21, D28, D35, D42, D49, and D56 the date was recorded and the type of recurrence was determined by PCR (appearance of new infection, malaria recrudescence, missing PCR information or undetermined type).
According to the study protocol, the first fifty children from Kenya enrolled in the ASMQ arm underwent intensive blood sampling: at baseline, on D0 after drug administration (until 6 h after first dosing), D2 (until 6 h after the third dose), D3 (72 h after first dose), D7 and on one other occasion on day 28, 35, 42, 49, 56 or 63. Two blood samples, at baseline and on D7, were collected for all the other participants. Additionally, for all patients with recurrence of parasitaemia, a blood sample was taken on the day of failure.
Analytical methods
The mass spectrometry assay for AS, DHA and MQ used for the analysis of study samples is an adaptation of a previously published validated multiplex method [
13]. The assay has been further improved by the use of stable isotopically labelled internal standards for MQ (mefloquine-d9) and DHA (DHA-13Cd4) to circumvent the potential matrix effect that may affect the accuracy of mass detection.
The mobile phase was delivered at a flow rate of 0.3 mL/min on a 2.1 mm × 75 mm XSelect HSS 3.5 μm column (Waters, Milford, MA, USA), using solvent A (2 mM ammonium acetate + 0.1% FA) and solvent B (MeCN + 0.1% FA) distributed according to the following stepwise gradient program: 98% A: 0 min; 98% A → 15% A: from 0.0 min → 13.0 min followed by a re-equilibration step to the initial solvent proportions. The retention time of mefloquine/mefloquine-d9, DHA/DHA-13Cd4 and artesunate is 7.4 min, 8.2 min and 9.2 min, respectively. The chromatographic system was coupled to a triple stage quadrupole (TSQ) Quantum Ion mass spectrometer (MS) from Thermo Fischer Scientific (Waltham, MA, USA) equipped with an Ion Max electrospray ionization (ESI) interface. The limits of quantification (LOQ) of the method are 2.5 ng/mL for MQ and 2 ng/mL for AS and DHA.
Plasma samples were isolated by centrifugation and stored at − 20 °C until batch analysis. Briefly, 100 μL of plasma sample were mixed with 50 µL internal standard (DHA-13Cd4 at 130 ng/mL; mefloquine-d9 at 43 ng/mL) and extracted with 600 µL of acetonitrile. The supernatant (700 µL) was evaporated under nitrogen at room temperature and was reconstituted in 150 µL of MeOH/ammonium acetate 2 mM (1:1) adjusted with formic acid at 0.1%, vortex-mixed and centrifuged again. The samples were maintained at +5 °C in autosampler racks throughout the analytical series. The injection volume was 20 μL.
The method is precise (with mean inter-day CV % < 10%), and accurate (inter-day deviation from nominal values < 5%). Since its initiation, the laboratory has participated in the Pharmacology Proficiency Testing Programme for anti-malarial drugs (
http://www.wwarn.org/toolkit/qaqc) organized by the World Wide Antimalarial Resistance Network WWARN (
http://www.wwarn.org/).
Pharmacokinetics analysis
Non-linear mixed effects modelling program (NONMEM
®, version 7.3) [
14] with the Perl-Speaks NONMEM
® (PsN) toolkit (version 3.7.6) [
15] was used to estimate average population pharmacokinetic parameters and their associated between-subject variability (BSV) and to identify factors that influence them. MQ and AS/DHA pharmacokinetic models were developed on the data collected from 50 Kenyan patient subjects with extensive sampling. Molar units were used for AS/DHA pharmacokinetic analyses. Because of the very fast rate of AS and DHA elimination and the selection of the trial sampling times, an external model validation could only be performed for MQ on the clinical trial data not used for model-building. Graphical exploration and statistical analyses were performed by means of the R package (version 2.15.1, R Development Core Team,
http://www.r-project.org/).
Structural and statistical model
A stepwise modelling approach was undertaken to identify models that best described the MQ and AS/DHA pharmacokinetics. Multi-compartment dispositions with first-order absorption and elimination processes were compared for MQ. Due to the restricted amount of AS and DHA data, drug and metabolite pharmacokinetics were modelled simultaneously and directly described by means of a one compartment model with linear absorption and elimination. Moreover, since AS is rapidly and almost completely hydrolysed in DHA, its elimination was assumed to occur exclusively via irreversible conversion to DHA [
16,
17]. An adequate AS absorption rate constant (K
a) estimation could not be made because of the small number of samples collected right after dose intake (one sample at maximum for each enrolled child on the first and third treatment day). K
a was thus fixed to 3.2 h
−1, the mean of previously published estimates retrieved from papers using a first-order process to depict AS absorption [
17,
18].
Parameterization was performed in terms of clearances (CL for drugs and CLM for metabolite), inter-compartmental clearance (Q), central (VC for drugs and VM for metabolite) and peripheral (VP) volumes of distribution and Ka. The metabolic conversion rate from AS to DHA was estimated by CL/Vc as previously discussed. AS and MQ relative bioavailability (F1, fixed to 100% and with estimated BSV) were also tested for AS/DHA and MQ to account for dose variation with respect to the nominal value due to the administration of water dispersible tablets. Since the ASMQ combination is administered orally, the pharmacokinetic parameter estimates represent apparent values.
Exponential errors were assumed to capture BSV in all the pharmacokinetic parameters. Proportional, additive and combined proportional-additive error models were compared to describe drugs and metabolite intra-patient (residual) variability. Finally, the correlation between AS and DHA concentration measurements was tested using the L2 function in NONMEM®.
Covariate analysis
Available covariates were: body weight (BW), height/length, age, sex, creatinine, total bilirubin (BIL), aspartate (AST) and alanine (ALT) aminotransferases, haemoglobin (Hb), haematocrit (Ht), total parasitaemia and co-medications categorized as CYP3A4 inducers. Visual inspection of the correlation between post hoc individual estimates of the pharmacokinetic parameters and the available patients’ characteristics was initially conducted to identify potential physiologically plausible relationships. Creatinine clearance was not evaluated since MQ elimination occurs mainly through non-renal processes and AS is completely converted into DHA, which is eliminated via glucuronidation [
16]. A stepwise forward insertion/backward deletion approach was then undertaken. Potential covariates influencing the kinetic parameters were first incorporated one at a time and tested for significance (univariate analysis). Sequential multivariate combinations of the identified factors were investigated to discard redundancies and to build an intermediate model with all the most important covariates (multivariate analysis). Finally, backward deletion consisted of removing covariates one at a time from the intermediate model, starting from the most insignificant until no further deterioration of the model was observed.
The influence of body weight on all MQ and DHA pharmacokinetic parameters (PAR) was tested using allometric scaling:
$$ PAR = \theta *\left( {\frac{BW}{MBW}} \right)^{PWR} $$
(1)
with θ PAR population estimate, MBW the median population body weight and PWR the function power fixed to 0.75 for clearances and 1 for volumes of distribution [
19]. A linear relationship between the typical value of a parameter and all the other covariates (continuous centered on the population median; dichotomous coded as 0 and 1) was used. Additionally, AST, ALT and BIL were implemented in the model as dichotomous variables, by introducing a boundary condition, i.e. below or exceeding 1.5 times the upper limit of normal (ULN). Children’s age was used to investigate the impact of organ maturation on MQ and DHA clearances, using the following equations, in addition to the simple linear one:
$$ CL = \theta *\frac{1}{{1 + \left( {\frac{AGE}{{TM_{50} }}} \right)^{ - Hill} }} $$
(2)
$$ CL = \theta *\left( {MAT_{mag} + \left( {1 - MAT_{mag} } \right)*\left( {1 - e^{{ - AGE*K_{mat} }} } \right)} \right) $$
(3)
where Hill is the sigmoid power, TM
50 the AGE at 50% of maturation, MAT
mag, the maturation magnitude for age, and K
mat the age maturation rate constant [
20,
21]. The population median covariate value was assigned to patients with missing information.
The acute phase of malaria is associated with altered gastrointestinal motility and an increased likelihood of vomiting. In the three-daily dose ASMQ regimen, the second dose is administered when the patient is in an improved state of health, thanks to the first dose of AS, that kills most of the parasites [
22]. The potential impact of parasitaemia on AS and MQ F1 was studied using a linear model of log-transformed (base 10) parasite counts measured at baseline of each ASMQ administration day. Missing parasitaemia information was imputed at the median value of the specific study day. Treatment day (0 vs. 1 and 2), considered as a surrogate marker of the rapid improvement in health due to the first AS dose, was also evaluated on AS and MQ F1. Since parasite counts and treatment day are correlated, differences in individual day 0 F1 due to parasitaemia at enrolment were explored, i.e. baseline parasite counts recorded at the first treatment day, by combining these two covariates. Furthermore, it was hypothesized that a patient’s clinical condition affects MQ K
a and this was tested by integrating the effect of the treatment day (0 vs. 1 and 2) on K
a.
Terminal half-lives (t1/2), maximum concentration (Cmax), and time to achieve Cmax (tmax) for all the three drugs, MQ area under the curve to infinite (AUC0–inf) and AS and DHA AUC0–24 after the first and the third ASMQ intake were computed using final pharmacokinetic parameter estimates and classic pharmacokinetic equations or NONMEM integration, as appropriate.
Parameter estimation, model selection and exclusion criteria
MQ and AS/DHA concentrations were fitted using the first-order conditional (FOCE) method with interaction. AS and DHA non-zero concentrations measured more than a week after last drug intake were thought unreliable and thus omitted from the analysis. Other missing variables (unreported concentration measurements, dose intake or sampling times, inconsistent date/time of dose intake and sampling) were also omitted. Data below the quantification limit (BQL) of the assays were handled by setting the first of a series of BQL samples at LOQ/2 and as missing all the others (M6 method) [
23].
Diagnostic goodness-of-fit plots, along with differences in the NONMEM® objective function value (ΔOFV), were employed to discriminate between nested models. Since a ΔOFV between any two hierarchical models approximates a χ2 distribution, a change of more than 3.84 (p < 0.05) and 6.63 (p < 0.01) points was considered statistically significant for one additional parameter in model-building or forward insertion and backward-deletion covariate steps, respectively. Akaike’s information criterion (AIC) was used for non-hierarchical models. Shrinkage was also evaluated. Sensitivity analyses removing outlying data with absolute conditional weighted residuals (CWRES) greater than 4 or potentially unreliable covariate values and concentration measurements were finally performed to avoid any potential bias in parameter estimation and covariate exploration.
Model validation and assessment
The stability of the final MQ and AS/DHA models was assessed by means of the bootstrap method implemented in PsN-Toolkit [
15]. Median parameter values with their 95% confidence interval (CI
95%) were derived from 2000 replicates of the initial datasets and compared with the original estimates. Prediction-corrected visual predictive checks (pcVPC) were also performed using the PsN-Toolkit and the R package Xpose4 by simulations based on the final pharmacokinetic models with variability using 1000 children [
15,
24]. Moreover, the final MQ pharmacokinetic model was validated using concentrations collected from participants not used in initial model development. The accuracy and precision of the model were estimated by means of prediction error (MPE) and root mean square error (RMSE), using log-transformed concentrations, for the entire dataset and also for each study site [
25].
Comparison between mefloquine exposures in children and adult volunteers and patients
Median and 90% prediction interval (PI
90 %) of children and adult concentration–time profiles were obtained through simulations (n = 1000) using the final pharmacokinetic model described above and published MQ pharmacokinetic models including BSV and intra-individual variability, respectively. A literature search allowed the identification of two pharmacokinetic models developed in adults receiving the same fixed dose formulation of ASMQ as the one administered to the children enrolled in this clinical trial [
26,
27]. The investigated populations consisted of Indian adult patients and Thai adult patients and volunteers, administered with 400 mg of MQ once per day over three consecutive days. MQ disposition was described by a two compartment model with linear elimination in both analyses. A first-order and a single transit compartment models in Julien et al. [
26] and Reuter et al. [
27], respectively, characterized the absorption phase. The two models were implemented in NONMEM
®, fixing simulated individuals’ body weight to the corresponding median population value. Administered MQ doses were 110 mg and 400 mg over three consecutive days for children and adults, respectively. MQ drug exposure was quantified by computing median and PI
95% AUC over the whole study period (AUC
0–day63) by NONMEM integration for all the simulated population/model.
Mefloquine pharmacokinetic–pharmacodynamic analysis
This exploratory analysis was carried out on MQ data collected from all children participating in the trial with complete dosing history information that did not drop out in the early days of the study. Model predicted MQ cumulative AUC (AUC0–dayx) on study days 7, 28, 42, and 63 were calculated by concentration integration in NONMEM®. The relationship between recrudescence of infection (response variable, coded as 0/1) and model predicted AUC0–dayx (independent variable) on study days 7, 28, 42, and 63 was inspected by means of logistic regression using STATA (StataCorp. 2013. Stata Statistical Software: Release 13. College Station, TX: StataCorp LP). The independent variable was log-transformed (using base 2) and cantered on its median value. The level of significance was set at 0.05.