2.1 Participants and Design of Clinical Studies Analyzed
The population pharmacokinetic analysis was performed using combined data from three phase I studies conducted in healthy subjects and subjects with RA and two phase II studies conducted in subjects with RA (Table
1). Upadacitinib was administered as immediate-release capsules in all studies included in the analysis. The study protocols were approved by the Institutional Review Boards/Ethics Committees of the study sites, and all the participants gave written informed consent prior to participation in the studies. The studies were conducted according to the International Conference on Harmonization Guidelines for Good Clinical Practice.
Table 1
Overview of phase I and II studies included in the upadacitinib population pharmacokinetic analysis
Phase I | |
1 | Healthy subjects | 56 | Single dose, randomized, placebo-controlled 17 samples up to 72 h post-dose | 1, 3, 6, 12, 24, 36, 48 mg | No | |
2 | Healthy subjects | 44 | Multiple dose, randomized, placebo-controlled 11 samples up to 12 h post day 1 morning dose and 18 samples up to 72 h post day 14 dose | 3, 6, 12, 24 mg bid for 14 days | No | |
| Subjects with mild to moderate RA | 14 | Multiple dose, randomized, placebo-controlled 11 samples up to 12 h post days 1 and 26 morning doses and 17 samples up to 48 h post day 27 dose | 6, 12, 24 mg | Yes | |
3 | Healthy Japanese and Chinese subjects | 45 | Single dose, randomized, placebo-controlled 17 samples up to 72 h post dose | 3, 6, 24 mg | No | – |
Multiple dose, randomized, placebo-controlled 11 samples up to 12 h post day 1 morning dose and 18 samples up to 72 h post day 14 dose | 18 mg bid for 14 days |
Phase II | |
4 | Subjects with moderate to severe active RA (MTX-IR) | 300 | Randomized, placebo-controlled, dose-ranging Single predose trough sample at weeks 2, 4, 6, 8, and 12 Samples at 1, 2, 3 h after the morning dose on day 1 and week 8 in ~ 30% of subjects | 3, 6, 12, and 18 mg bid and 24 mg qd | Yes | |
5 | Subjects with moderate to severe active RA (anti-TNF IR) | 276 | Randomized, placebo-controlled, dose-ranging Single sample at weeks 2, 4, 6, 8, and 12 Samples at 1, 2, 3 h after the morning dose on day 1 and week 8 in ~ 30% of subjects | 3, 6, 12, and 18 mg bid | Yes | |
In the three phase I studies (studies 1, 2, and 3 in Table
1), healthy subjects were eligible to participate if they were 18–55 years of age, with a body mass index within 19–29 kg/m
2, and judged to be in good general health. In addition to healthy subjects, study 2 enrolled a cohort of subjects who were 18–75 years of age with a body mass index within 19–39 kg/m
2 at screening, had a diagnosis of RA for at least 6 months, and had been receiving a stable dose of methotrexate for at least 3 months. Study 3 enrolled subjects of Japanese or Chinese origin. In all three studies, individuals were excluded if they had any clinically significant condition (except stable RA in the patient cohort of study 2), abnormalities, infection, or any clinically significant findings at screening, as determined by the principal investigator. Inhibitors or inducers of drug-metabolizing enzymes were not allowed within 30 days of enrollment. Subjects were randomized to receive single doses of upadacitinib or placebo in study 1, multiple doses of upadacitinib or placebo for 14 days in healthy subjects and 26 days in subjects with RA in study 2, and single or multiple doses of upadacitinib or placebo for 14 days in study 3.
In the two phase II studies (studies 4 and 5), adult males and females 18 years of age or older who had active RA for at least 3 months were enrolled in the studies. Subjects were enrolled in study 4 if they had inadequate response to methotrexate, and were excluded if they had received any biologic RA therapy. Subjects were enrolled in study 5 if they had inadequate response to at least one prior anti-TNF therapy. In both studies 4 and 5, subjects were excluded if they had an estimated glomerular filtration rate (eGFR) of < 40 mL/min/1.73 m
2 or serum aspartate transaminase (AST) or alanine transaminase (ALT) > 1.5 × upper limit of normal at screening. Strong inhibitors or inducers of CYP3A were prohibited throughout both studies. Subjects in each study were randomized to receive different doses of upadacitinib or placebo for 12 weeks. Further details on the design and inclusion and exclusion criteria of studies 4 and 5 have been previously reported [
17,
18].
In studies 1–3, serial blood samples were collected from each subject over the time periods delineated in Table
1. In studies 4 and 5, a single predose blood sample was collected from each subject at each study visit. In approximately 30% of subjects in studies 4 and 5, blood samples were collected at 1, 2, and 3 h after dose at the day 1 and week 8 visits. Plasma concentrations of upadacitinib were determined using a validated liquid chromatography method with mass spectrometric detection at AbbVie (North Chicago, IL, USA), as previously described [
16].
Only 4% of upadacitinib concentrations for samples collected after initiation of dosing were found to be below the lower limit of quantitation (LLOQ) in the dataset. For each subject, concentrations below the LLOQ during the dosing period, and the first concentration below the LLOQ after the last dose, were imputed with 1/2 LLOQ [
20]; subsequent concentrations below the LLOQ after the last dose were censored in the analysis. Upadacitinib concentrations from the phase II studies in RA were plotted versus time since the last dose for each patient individually to identify potential outliers. Only 0.9% of all dataset concentrations were flagged as clear outliers that likely resulted from inaccurate dosing records and were excluded from the analysis.
2.2 Pharmacokinetic Model Development
A non-linear mixed-effects modeling approach was utilized to analyze upadacitinib concentration versus time data. NONMEM® 7.3 software (ICON Development Solutions, Hanover, MD, USA) was utilized.
The population pharmacokinetic model was built in a stepwise manner. The structural model was developed first, and then models for intersubject variability (ISV) were added. The ISV was modeled using exponential error models:
$$P_{ni} = \theta_{n} \cdot \exp (\eta_{ni} ),$$
where
Pni,
θn, and
ηni are the individual parameter estimate, population parameter estimate, and intersubject random effect, respectively, for subject
i and parameter
n. The
ηn values were assumed to be normally distributed with a mean of 0 and a variance of
ωn2.
The residual error was assumed to be a combination of additive and proportional error terms:
$$C_{ij} = \widehat{C}_{ij} \cdot \left( {1 + \varepsilon_{1ij} } \right) + \varepsilon_{2ij} ,$$
where
Cij is the measured plasma concentration for subject
i at time
j,
Ĉij is the corresponding model-predicted plasma concentration, and
ε1ij and
ε2ij are normally distributed independent residual random error terms with a mean of zero and variances of
σ2:
$$\varepsilon_{n} \sim N\left( {0,\;\sigma_{n}^{2} } \right).$$
The following covariates were evaluated: (1) for upadacitinib apparent clearance (CL/F): population (subjects with RA versus healthy subjects), baseline serum bilirubin concentration, baseline serum creatinine, baseline creatinine clearance, baseline eGFR, body weight, age, baseline serum AST concentration, baseline serum ALT concentration, baseline DAS28-CRP, baseline high-sensitivity C-reactive protein (hsCRP) serum concentration, upadacitinib dose, sex, race, baseline alcohol use, baseline tobacco use, concomitant use of pH-modifying medications (e.g. proton pump inhibitors or antacids), concomitant use of CYP3A inhibitors, concomitant use of CYP2D6 inhibitors, and CYP2D6 metabolic phenotype (determined based on CYP2D6 genotype); (2) for the apparent volume of distribution of the central compartment (Vc/F): population, sex, race, body weight, upadacitinib dose, and age; (3) for the first-order absorption rate (Ka): population, sex, race, baseline alcohol use, upadacitinib dose, age, and concomitant use of pH-modifying medications; and (4) for relative oral bioavailability: upadacitinib dose. For healthy subjects, missing baseline values of DAS28-CRP and hsCRP were imputed as 0.96 and 0.15 mg/L, respectively. For all other missing continuous covariate information, the population median value was assigned. Continuous covariates were evaluated using power models (centered on the median values in the datasets; referred to thereafter as the reference values), whereas categorical covariates were evaluated using indicator variables. The covariate that resulted in the largest drop in the objective function value (OFV) in a univariate covariate search was included in the model first, and then the process was repeated to identify the second covariate to be included in the model, and so forth.
The pharmacokinetic models were fit to the data using the first-order conditional estimation with interaction (FOCEI). The stochastic approximation expectation maximization (SAEM) algorithm with importance sampling employed within NONMEM was utilized at the last step to facilitate incorporation of interoccasion variability (IOV) on the lag time, as described in detail in the “
Results” section. The likelihood ratio test was used for hypothesis testing to discriminate among alternative nested models. Since the OFV provided by NONMEM is approximately Chi square distributed, the differences between OFV values were used to guide model building. When comparing nested models, one additional model parameter, corresponding to one degree of freedom in the higher-order model, was considered significant if it lowered the OFV by more than 6.63, corresponding to
p < 0.01.
The significance of covariates was evaluated through forward inclusion and backward elimination procedures using the likelihood ratio test with p value thresholds of p < 0.01 and p < 0.001, respectively. This was implemented on the last model estimated with FOCEI prior to inclusion of IOV with SAEM and importance sampling since the implementation of importance sampling is associated with Monte Carlo noise in the objective function that can, in certain cases, confound statistical comparisons. In addition to the likelihood ratio test, selection of the models was based on the following criteria: the observed and predicted plasma concentrations from the preferred model were more randomly distributed across the line of unity (a straight line with zero intercept and a slope of one) than alternative models, the conditional weighted residuals of the preferred model showed less systematic bias than alternative models, the preferred model showed physiologically reasonable and statistically significant estimates (95% confidence intervals [CIs] did not include zero) of the parameters and their standard errors, and visual predictive checks (VPCs) showed appropriate characterization of observed trends and variability.
VPCs and prediction-corrected VPCs [
21] were examined based on 400 simulated replicates of the dataset generated using NONMEM and Perl Speaks NONMEM (PSN 4,
http://psn.sourceforge.net/index.php, Uppsala University, Uppsala, Sweden). In order to estimate CIs of the model parameters, 200 bootstrap replicates were constructed by randomly sampling
N subjects from the original dataset with replacement, where
N was the number of subjects in the original dataset. Model parameters were estimated for each bootstrap replicate and the resulting values were used to estimate medians and CIs. Bootstrap statistics were based on successfully completed runs.
Simulations with 200 replicates using the demographics of the RA patient population from the two phase II studies were performed to explore the impact of covariates on upadacitinib area under the concentration–time curve (AUC) and maximum concentration (Cmax) at steady state. Parameter uncertainty was incorporated using the covariance matrix obtained from the bootstrap analysis. While exploring the impact of one of the identified covariates on exposure, other significant covariates were fixed to the reference values.