Population pharmacokinetics–pharmacodynamics of sunitinib in pediatric patients with solid tumors

Data were pooled from two Phase I/II clinical trials in pediatric patients with solid tumors (studies ADVL0612 and ACNS1021), including predominantly high-grade glioma, ependymoma, brain stem glioma, or sarcoma. Details of these trials have been published previously [8‐10]. The original trials were approved by the institutional review boards at all participating centers and The National Cancer Institute Pediatric Central institutional review board also approved study ACNS1021. All participants in the original trials or their parent/legal guardian signed a document of informed consent and assent was obtained as appropriate. A detailed description of the eligibility and exclusion criteria can be found in the Online Resource Methods. Patients were between 18 months and 22 years of age and received sunitinib at a starting dose of 15 or 20 mg/m² on a schedule of 4 weeks on treatment followed by 2 weeks off treatment (schedule 4/2) [8‐10]. In total, data for analysis were available from 59 patients (ADVL0612, n = 35; ACNS1021, n = 24).

Sample collections (1.5 mL) for PK analysis were completed at pre-specified visits. For study ADVL0612, the series of samples for PK analysis was conducted at 0, 1, 2, 4, 6, 8–10, and 24–48 h post-dose on day 1 of cycle one and pre-dose (trough) assessments were made on days 7, 14, 21, and 28 of cycle one. For study ACNS1021, the series of samples for PK analysis were taken at 2, 4, 6–8, and 24 ± 1 h post-dose on day 1 of cycle one. In addition, pre-dose (trough) assessments were made on days 7, 14, 21, and 28 of cycle one, and days 1 and 28 of cycle two. The majority of the safety assessment data that were to be used for PK–PD modeling were captured at each study visit, but some safety assessments were performed less frequently per the specific requirements of each study protocol. Plasma samples were analyzed for the determination of sunitinib and SU012662 concentrations using a sensitive, specific, and validated liquid chromatography with tandem mass spectrometry assay (BASi, West Lafayette, IN), as previously published [8‐10]. For study ADVL0612, calibration standard responses met acceptance criteria over the range of 1–200 ng/mL for sunitinib and 1–100 ng/mL for SU012662, using a quadratic weighted (l/concentration²) regression. The lower limit of quantification (LLOQ) for both sunitinib and SU012662 was 1 ng/mL. The between-day assay accuracy, expressed as percent relative error, for quality control (QC) concentrations, ranged from − 3.3–1.3% for sunitinib and − 1–3.8% for SU012662 for the low, medium, high, and diluted QC samples. Assay precision, expressed as the between-day percent coefficient of variation of the mean estimated concentrations of QC samples, was ≤ 5.1% for the low (3.00 ng/mL), medium (100 ng/mL), and high (150 ng/mL) concentrations of sunitinib, and was ≤ 9.5% for the low (3.00 ng/mL), medium (50.0 ng/mL), high (75.0 ng/mL), and diluted (15.0 ng/mL) concentrations of SU012662. For study ACNS1021, calibration standard responses met acceptance criteria over the range of 0.100–60.0 ng/mL for sunitinib, and 0.100–20.0 ng/mL for SU012662. The LLOQ for both sunitinib and SU012662 was 0.100 ng/mL. The between-day assay accuracy for QC concentrations ranged from 3.7 to 7.3% for sunitinib and − 3.3%–5.3% for SU012662 for low, medium, high, and diluted QC samples. Assay precision was ≤ 7.5% for low (0.300 ng/mL), medium (30.0 ng/mL), high (45.0 ng/mL), and diluted (9.00 ng/mL after dilution) concentrations of sunitinib, and was ≤ 13.7% for the low (0.300 ng/mL), medium (9.00 ng/mL), high (15.0 ng/mL), and diluted (3.00 ng/mL after dilution) concentrations of SU012662.

The PK and PK–PD modeling approaches were described in detail in the analysis plan before initiating the analyses. Regulatory guidance and quality control were taken into account. A systematic multistep approach to model development consisted of base model development, random effects model development, full model development, final model development, assessment of model adequacy (goodness-of-fit), and assessment of model predictive performance (validation). Analysis was conducted using nonlinear mixed-effects modeling methodology, as implemented in NONMEM (v7.1.2, University of California at San Francisco, CA), and using the first-order conditional estimation method with interaction to estimate all parameters. During model development, the goodness-of-fit of different models to the data were evaluated using change in the objective function, visual inspection of different scatterplots, precision of the parameter estimates, and decreases in inter-individual and residual variability. The base model consisted of a two-compartment model with first-order absorption and lag time (t_lag) to fit sunitinib and SU012662 concentrations. The type of base models used for the PK–PD modeling portion were transit compartments in series with feedback loop models or indirect response models, as used previously [16]. Sequential PK–PD models for the safety and tolerability endpoints were built using the final PK model–predicted sunitinib concentrations. SU012662 data were not included in this process because a previous study showed that inclusion of the predicted SU012662 concentrations did not improve the model objective function value (OFV) [16]. During the modeling portion, safety endpoints were used when there was a sufficient number of patients with PK and PD data (i.e., at least two patients with both PK and PD data, either both baseline and post-baseline for continuous safety endpoints, or post-baseline for categorical safety endpoints). The safety and tolerability endpoints assessed were related to the most common safety events observed with sunitinib use [5, 6], which were absolute neutrophil count (ANC), alanine transaminase (ALT), aspartate transaminase (AST), diastolic blood pressure (BP), hemoglobin, lymphocyte count, platelet count, white blood cell (WBC) count, hand-foot syndrome, fatigue, nausea, intracranial hemorrhage, and vomiting. For categorical safety endpoints, PK–PD modeling was not used. Instead, the relationships between the average daily plasma exposures (i.e., the accumulated dose divided by the time after the first dose, divided by the individual sunitinib clearance × 1000) up to the time of earliest worst grade Common Terminology Criteria for Adverse Events (CTCAE) and the incidence rate were explored by looking at different CTCAE grade incidence rates at the lower and upper halves for the mean plasma exposures, as well as using ordered logistic regression approaches, where there was a sufficient number of events (i.e., more than one event). Interindividual variability in the PK or PK–PD parameters was modeled using multiplicative exponential random effects of the form θ_i = θ × e^ηi, where θ is the typical or central value of the parameter, and the empirical Bayes prediction of the interindividual random effect (η) was a random variable normally distributed with a mean 0 and variance ω².

Based on prior experience, a group of potential covariates was examined with respect to the PK and PK–PD parameters (Table 1). Covariates were tested for significance in a stepwise manner using a stepwise covariate model (SCM) building procedure and statistical criteria of α = 0.01 for the forward inclusion step. The full model was then subjected to a backward elimination step with a statistical criterion of α = 0.001. In the SCM approach, linear and power functions were tested for any continuous covariates on PK parameters, with the SCM model reporting the best one selected (i.e., the one with the lower OFV). Baseline body weight and body surface area (BSA) were examined to ensure that the body size measure with the largest effect was included in the final model, and other measures related to body size that were highly correlated with body weight or BSA, such as lean body weight, body mass index, and height, were not included in the covariate analysis. BSA was calculated based on the method of DuBois and DuBois [i.e., BSA = 0.20247 × height (m)^0.725 × weight (kg)^0.425]. No adaptation was required because no patients had undergone prior amputation. In addition, for both sunitinib and SU012662, baseline bodyweight was tested in a separate SCM run where BSA was replaced with bodyweight and the final model with the lower OFV was selected.

For the validation of the base and final models, visual predictive check (VPC) techniques comprising 1000 simulations were carried out, and the median and upper and lower bounds of the 95% prediction interval (PI) for PK or PD profiles were compared against the observed data median and confidence intervals (CIs). The number of observations not within the 95% PI was to remain within 5% of the total number of observations, and the mean prediction profile was expected to follow the observed mean profile. In additon, the 90% PIs for the median and the lower and upper bounds of the 95% PIs were identified to ensure that they included the observed median and 95% CI bounds. Bootstrapping techniques (1000 boostrap datasets) were applied to generate the nominal 95% CIs around the point estimates and to confirm the 95% CIs generated by the base or final models based on asymptotic standard errors generated from the NONMEM covariance step.

The sunitinib and SU012662 base models comprised a two-compartment model with first-order absorption and elimination rates, consistent with previous work. Based on previous observations [13], a conversion of 21% of sunitinib to SU012662 was assumed to bring the magnitude of the parameters for SU012662 to a more physiologically relevant level. In both the sunitinib and SU012662 base models, the diagnostic plots were satisfactory. The effects of extreme outliers on the population PK parameter estimates and on the diagnostic plots were tested in both models, and no outliers met the criteria for exclusion from the datasets [i.e., |conditional weighted residual (CWRES)|> 6]. Bootstrap results were consistent with the population parameter estimates, indicating that the base models were stable and the parameter estimates represented the final datasets adequately.

Using SCM, the effects of different covariates on apparent clearance (CL/F) and apparent central volume of distribution (Vc/F) in both models were examined. For sunitinib, the estimated typical values for CL/F and Vc/F were 24.1 L/h and 1070 L, respectively, and for SU012662 the values were estimated to be 10.9 L/h and 1030 L, respectively. For sunitinib, the effect of BSA on CL/F was statistically significant (P < 0.001) using a linear function: CL/F = 24.1 L/h · [1 + 0.557 (BSA–1.47)], and for SU012662, the effect of BSA on CL/F was statistically significant (P < 0.001) using a power function: CL/F = 10.9 L/h · (BSA/1.47)^0.843. For sunitinib, the effect of BSA on Vc/F was also statistically significant (P < 0.001) using a power function: Vc/F = 1070 L · (BSA/1.47)^1.47, and for SU012662, the effect of BSA was also statistically significant on Vc/F (P < 0.001) using a power function: Vc/F = 1030 L · (BSA/1.47)^1.72. Therefore, for sunitinib and SU012662, higher BSA was associated with greater CL/F and Vc/F. Baseline bodyweight was also tested in a separate SCM run where BSA was replaced with bodyweight. However, the final model with BSA was the one with the lower OFV and hence was selected. The effects on PK parameters of the other covariates tested (Table 1) were not found to be statistically significant for sunitinib or SU012662.

Goodness-of-fit diagnostic plots for the observed versus individual-predicted or population-predicted values, as well as CWRES versus time or population-predicted values, in the final PK models for sunitinib and SU012662 plasma concentrations demonstrated the adequacy of the models (Online Resource Fig. S2A and S2B, S3A and S3B, S4A and S4B, and S5A and S5B). Moreover, the prediction- and variance-corrected VPC plots of sunitinib and SU012662 plasma concentration demonstrated the similarity between the predicted and the observed data during the first 12 h post-dose (Online Resource Fig. S6) and up to 3000 h post-dose (Fig. 1a, b). A summary of sunitinib and SU012662 PK parameters from the final models and following bootstrapping are shown in Table 3. Bootstrap results were consistent with the population parameter estimates, indicating that the final models were stable and that the population parameter estimates represented the dataset adequately. For sunitinib, inclusion of baseline BSA as a covariate into the final model reduced interindividual variability on CL/F and Vc/F by 25.9 and 78.0%, respectively, and for SU012662, interindividual variability on CL/F and Vc/F was reduced by 28.7 and 53.9%, respectively.

PK–PD models were based solely on sunitinib data, as described previously [16]. For ANC, hemoglobin, lymphocyte count, platelet count, and WBC count, a sequential transit compartments in series with feedback loop PK–PD model was used as the base model. For ANC, lymphocyte count, platelet count, and WBC count, the model included a maximum sunitinib effect (E_max) on the proliferation rate constant (k_prol) in the stem cell compartment, whereas for hemoglobin, the model included an effect first-order rate constant (k_PD) effect on the k_prol constant in the stem cell compartment. For ALT, AST, and diastolic BP, a sequential indirect response PK–PD model was used as the base model. For ALT and AST, the models included a k_PD effect on the output/elimination rate constant (k_out), and for diastolic BP, the model included a k_PD effect on the input rate constant (k_in). Each of these models has been used previously to describe the time course of each endpoint following sunitinib dosing [16]. Following SCM analysis for each endpoint, no covariates (Table 1) were found to statistically significantly (P > 0.001) affect the concentration at half maximum effect (EC₅₀) or k_PD; therefore, the final model used was the base model.

Goodness-of-fit plots for the observed values versus individual-predicted or population-predicted values, as well as CRWES values versus time or predicted values, in the final PK–PD models for the continuous safety endpoints (ANC, ALT, AST, diastolic BP, hemoglobin, lymphocyte count, platelet count, and WBC count) demonstrated the adequacy of the models (Online Resource Fig. S2C–J, S3C–J, S4C–J, and S5C–J). Moreover, the prediction- and variance-corrected VPC plots of each safety endpoint demonstrated the similarity between the predicted and the observed data (Fig. 1c–j), further supporting that the PK–PD models described the time course of each of the safety and tolerability endpoints adequately. Summaries of the PK–PD parameters from each of the final models for the continuous safety endpoints are shown in Table 4. For each of these endpoints, the mean from the model bootstrap runs was consistent with those from the model run, indicating that the population parameter estimates from the models adequately represented the datasets.

Logistic regression analysis was used to determine the relationship between the incidences of the worst AE grade for each safety and tolerability endpoint and the calculated average plasma sunitinib concentration. The incidences of observed worst grade ≥ 1 AEs were neutropenia (n = 23 patients), increased ALT (n = 15), increased AST (n = 13), hypertension (n = 10), decreased hemoglobin (n = 1), lymphopenia (n = 16), thrombocytopenia (n = 13), and leukopenia (n = 20). When patients were stratified by average plasma sunitinib concentration (< median vs. ≥ median, respectively), incidences were neutropenia [17.2% (5/29 patients) vs. 60.0% (n = 18/30)], increased ALT [6.9% (2/29) vs. 43.3% (n = 13/30)], increased AST [3.4% (1/29) vs. 40.0% (n = 12/30)], hypertension [0% vs. 33.3% (n = 10/30)], lymphopenia [13.8% (4/29) vs. 40.0% (n = 12/30)], thrombocytopenia [3.4% (1/29) vs. 40.0% (n = 12/30)], and leukopenia [24.1% (7/29) vs. 56.7% (n = 17/30)]. There was a higher probability of neutropenia, increased ALT, increased AST, hypertension, lymphopenia, thrombocytopenia, and leukopenia with higher average plasma sunitinib concentrations (Fig. 2a–g). Because there was only one incident of decreased hemoglobin, logistic regression analysis could not be performed for this endpoint.

For the categorical safety and tolerability endpoints of fatigue, nausea, vomiting, intracranial hemorrhage, and hand-foot syndrome, PK–PD modeling was not used, but the association between plasma exposure and incidence was explored when there was a sufficient number of AEs. The incidences of observed worst grade ≥ 1 AEs were fatigue (n = 24 patients), nausea (n = 7), vomiting (n = 7), intracranial hemorrhage (n = 5), and hand-foot syndrome (n = 1). When patients were stratified by average plasma sunitinib concentration (< median vs ≥ median, respectively), the incidences were fatigue [24.1% (n = 7/29) vs. 56.7% (n = 17/30)], nausea [0% vs. 23.3% (n = 7/30)], vomiting [3.4% (n = 1/29) vs. 20.0% (n = 6/30)], intracranial hemorrhage [3.4% (1/29) vs. 13.3% (n = 4/30)], and hand-foot syndrome [0% vs. 3.3% (n = 1/30)]. Using logistic regression analysis, there was a higher probability of fatigue, nausea, and vomiting with higher average plasma sunitinib concentrations (Fig. 2h–j). Because there were only five incidents of intracranial hemorrhage and one incidence of hand-foot syndrome, the logistic regression analysis could not be performed for these endpoints.