Clinical Study Design
The clinical trial (NCT01394016) was a multicenter, non-randomized, open-label, dose-escalation, phase I study of abemaciclib for the treatment of adult patients with advanced cancer. The study design consisted of a dose escalation phase (Part A) and six tumor-specific cohorts (Parts B [NSCLC], C [glioblastoma multiforme; GBM], D [breast cancer], E [melanoma], F [colorectal cancer], and G [hormone receptor-positive metastatic breast cancer in combination with fulvestrant]). Further details of the trial design, and patient inclusion and exclusion criteria, have been described previously [8].
During dose escalation, patients received abemaciclib in capsules on two different schedules: either at 50, 100, 150, or 225 mg every 24 h (q24h), or at 75, 100, 150, 200, or 275 mg every 12 h (q12h). In the tumor-specific cohorts, patients were treated on a q12h schedule at a dose no greater than the maximum tolerated dose of 200 mg q12h [8]. Doses were reduced when a patient experienced unacceptable toxicity; sequential dose reductions were permitted from 200 mg q12h to 150, 100, and 75 mg q12h. Individual dosing information was collected using patient diaries.
Patients were asked not to consume food 1 h before through to 1 h after taking a dose of abemaciclib. As abemaciclib is primarily metabolized by cytochrome P450 (CYP) 3A [12], patients were advised against both drinking grapefruit juice and taking inducers or strong inhibitors of CYP3A4 during the trial.
Patient Characteristics
Information such as date of birth, habits (e.g., alcohol consumption, smoking), historical diagnoses, and chronic conditions were self-reported by the patient. Clinical parameters such as weight and height were measured at visits to the investigative site. Creatinine clearance was estimated using the Cockcroft-Gault formula [13].
Pharmacokinetic/Pharmacodynamic (PK/PD) Sampling Schedule
The pharmacokinetic sampling schedule was designed to characterize both the single-dose and multiple-dose pharmacokinetics of abemaciclib. Patients received a single dose of abemaciclib on day-3, and then began continuous treatment starting on day 1 through day 28 of cycle 1. On day 28, patients received a single dose of abemaciclib; those enrolled on the q12h dosing regimen did not receive a second abemaciclib dose on day 28, in order to better characterize the steady-state half-life of abemaciclib. Blood samples for pharmacokinetic analysis were drawn on day-3 (pre-dose, 1, 2, 4, 6, 8, 10, 24, 48, and 72 h), day 15 (pre-dose, 1, 2, and 4 h), day 22 (pre-dose), and day 28 (pre-dose, 1, 2, 4, 6, 8, 10, and 24 h) of the first cycle. Plasma concentrations were assayed using a validated liquid chromatography with tandem mass spectrometry method [8].
The pharmacodynamic sampling schedule was designed to evaluate changes in p-Rb expression in epidermal keratinocytes at steady state as a result of abemaciclib-mediated cell cycle inhibition, while maintaining a relatively sparse schedule due to the procedural nature of the sampling technique. Skin biopsies for pharmacodynamic analysis were scheduled at baseline (prior to the first dose of abemaciclib; any time between day-14 through day-4) and on day 15 (pre-dose and 4 h post-dose) of cycle 1.
Base Pharmacokinetic Model Development
The base pharmacokinetic model was developed to describe the single and multiple oral dose pharmacokinetics of abemaciclib in patients with advanced cancer. A variety of model structures were explored, including linear versus non-linear absorption rate, delayed absorption (lag times, transit compartments), linear versus non-linear clearance, time-dependent clearance, and monophasic versus biphasic distribution. During model development, it was observed that the apparent clearance (CL/F) and apparent volume of distribution (V
d/F) parameters were strongly correlated, which was confirmed by incorporating a formal correlation assessment. As these parameters are both dependent on bioavailability, a relative bioavailability term (F
rel) was used to capture the inter-individual variability that could be attributed to this apparent correlation between CL/F and V
d/F. Further explorations with this model structure included dose and time dependence on F
rel.
For each new model structure, variability terms were investigated for all parameters; variability was assumed to be log-normally distributed. Inter-occasion variability and residual error models (additive, proportional, or combined) were evaluated with each assessment of inter-individual variability.
Final Population Pharmacokinetic (PopPK) Model Development
Once the base model was established, the impact of patient factors on the disposition of abemaciclib was assessed. Any inter-occasion variability and parameter correlations were removed from the base model for covariate evaluation to avoid parameter/relationship bias; these relationships were reconsidered for the final model once the covariate analysis was complete.
Patient factors tested comprised both categorical covariates (e.g., sex, alcohol consumption, and smoking status) and continuous covariates (e.g., age, body weight, serum albumin, aspartate transaminase, alanine transaminase, alkaline phosphatase, blood urea nitrogen, lactate dehydrogenase, serum creatinine, and creatinine clearance). Both serum creatinine and creatinine clearance were tested given the observed change in serum creatinine levels as a result of abemaciclib treatment [8]. Covariate relationships were considered for all combinations of pharmacokinetic parameters (those with inter-individual variability) and patient factors. For continuous covariates, relationships were first tested using a linear model (Eq. 1); if the linear model demonstrated significance (change [Δ] in objective function value [OFV] −3.84, p < 0.05) then a power model was tested (Eq. 2) and the covariate relationship with the lowest OFV was selected. For categorical covariates, relationships with the relevant pharmacokinetic parameters were evaluated using a categorical model (Eq. 3).
$$ P = \varTheta_{1} \cdot (1 + \varTheta_{2} \cdot ({\text{COV}} - {\text{MED}})) $$
(1)
$$ P = \varTheta_{1} \cdot \left( {\frac{\text{COV}}{\text{MED}}} \right)^{{\varTheta_{2} }} $$
(2)
$$ P = \varTheta_{1} \cdot (1 + \varTheta_{2} \cdot {\text{IND}}), $$
(3)
where P is the individual’s estimate of the parameter, Θ1 represents the typical value of the parameter, Θ2 represents the effect of the covariate, COV is the value of the covariate, and MED is the population median of the covariate. IND is an indicator variable with a value of 0 or 1 for a dichotomous categorical covariate, or with a value from 1 to n for various values of a categorical covariate (where n is the number of categories).
For any time-dependent pharmacokinetic parameters in the structural model (i.e., F
rel), the covariate relationship was tested separately on each of the initial and steady-state parameter terms. If either one was significant (ΔOFV −3.84, p < 0.05), the covariate was then tested on the global pharmacokinetic parameter. If the covariate relationship for the global pharmacokinetic parameter resulted in the same or greater change in OFV, the relationship was retained on the global parameter.
For covariates that vary over time (body weight, serum albumin, aspartate transaminase, alanine transaminase, alkaline phosphatase, blood urea nitrogen, and lactate dehydrogenase), the covariate relationship was assessed using the patient information collected at each visit. Where no new information was available, the last observation was carried forward.
Following covariate evaluation, a full model was developed by incorporating all individual covariate relationships that were identified in the covariate selection step. The significance of each of these potential covariates was evaluated using backward elimination, where at each iteration the least significant covariate not resulting in an increase of 10.828 or greater in OFV (p < 0.001) was removed. For each of the remaining covariates, if the increase in the inter-individual variability on its omission was less than 5% points, the covariate was removed.
Model Implementation and Selection
Pharmacokinetic parameter estimates, inter-individual variability estimates, and error terms were obtained by fitting a model to the concentration–time data by means of the non-linear mixed-effects modeling program, NONMEM® (version 7.2; Icon Development Solutions, Ellicott City, MD, USA). The first-order conditional estimation method with interaction was used for all analyses. Base model selection was based on significant decreases in OFV, goodness-of-fit plots, and visual predictive checks (VPCs). Final model selection was based on the forwards inclusion and backwards exclusion approach, and the criteria described in Sect. 2.5.
Pharmacokinetic/Pharmacodynamic (PK/PD) Model Simulations
A semi-mechanistic pre-clinical PK/PD model was previously developed to describe the relationship between abemaciclib pharmacokinetics and inhibition of p-Rb in mouse xenograft tumors [9]. By combining the human popPK model developed in the present work with the previously developed mouse PK/PD model, the p-Rb response to abemaciclib in human epidermal keratinocytes was predicted given the following key assumptions:
-
As abemaciclib protein binding is extensive (>95%) and in vitro values for mouse and human were within twofold, it was assumed that species differences in protein binding were negligible.
-
As the xenograft cell line was derived from a human tumor, it was assumed that there were no species-specific differences in potency for abemaciclib against CDK4 and 6 as measured by p-Rb inhibition.
-
It has previously been noted that the metabolites of abemaciclib are also inhibitors of CDK4 and 6 with similar potency values as the parent compound [14]. In the absence of in vivo data, it was assumed that the exposures of such active metabolites were comparable between species.
-
Finally, it was assumed that differences in biomarker matrix (i.e., tumor p-Rb vs. skin p-Rb) were negligible.