Concentration–QTc analysis of quizartinib in patients with relapsed/refractory acute myeloid leukemia

The concentration–QTc (C-QTc) analysis included data from the phase 3 QuANTUM-R study, which was conducted at 152 sites in 19 countries [10]. The institutional review board or ethics committee at each site approved the study protocol, and all patients provided written informed consent per the Declaration of Helsinki and Good Clinical Practice.

In QuANTUM-R, the starting dose of quizartinib was 30 mg/day (26.5 mg free base), followed by an increase to 60 mg/day (53.0 mg free base) after 2 weeks if QTcF was ≤ 450 ms. Patients receiving a concurrent strong CYP3A inhibitor initiated quizartinib at 20 mg/day (17.7 mg free base), with an increase to 30 mg/day (26.5 mg free base) after 2 weeks if QTcF was ≤ 450 ms. The dose was reduced from 60 to 30 mg daily or from 30 to 20 mg daily if protocol-specified criteria for dose reduction were met (for example, QTcF > 480 ms, persistent grade ≥ 3 nonhematologic toxicity, or myelosuppression in patients achieving complete remission with incomplete platelet recovery or complete remission with incomplete hematologic recovery who had received ≥ 2 cycles of treatment).

Plasma concentrations of quizartinib and AC886 were measured by BASi (West Lafayette, IN, USA) using a validated liquid chromatography–tandem mass spectrometry method. The analytical range validated was from 2 to 2000 ng/mL for both quizartinib and AC886. For quizartinib, the ranges for within-run and between-run assay precision were 0.9 to 6.3% and 3.0 to 6.3%, respectively, for quality control samples at concentrations within the calibration curve; the accuracy ranged from − 11.2 to 7.7% and − 4.0 to 4.0%, respectively. The lower limit of quantification was 2 ng/mL. For AC886, the ranges for within-run and between-run assay precision were 0.3 to 7.4% and 2.2 to 6.7%, respectively, for quality control samples at concentrations within the calibration curve; the accuracy ranged from − 9.0 to 6.7% and − 3.0 to 2.1%, respectively. The lower limit of quantification was 2 ng/mL.

For each patient, centrally read triplicate QTc measurements (from three 12-lead ECGs, ≥ 5 min apart per time point) were obtained at screening; predose and 2, 4, and 6 h postdose on days 1 and 15 of cycle 1; predose and 2 to 4 h postdose on days 2 and 8 of cycle 1 and on day 1 of cycles 2 and 3; and at any time on day 1 of subsequent cycles. Pharmacokinetic (PK) blood samples were collected at each of the QTc measurement times after ECG was performed, until cycle 3 of treatment.

For the C–QTc analysis, QTc data and concentration records were matched by comparing the actual date/time of each replicate QTc measurement versus the date/time of the corresponding concentration measurement. The difference in time between the 2 measurements was computed and used to determine whether the replicate QTc measurements met the criteria to be matched with concentration. A difference of ≤ 30 min was allowed for all time points except for the nominal 24-h postdose QTc measurements, in which up to 90 min was allowed.

The demographic and clinical covariates evaluated for their effect on the C–QTc model parameters included age, baseline body weight, sex, race, baseline QTc, coadministration of QT-prolonging drugs, and selected electrolyte deficiencies (hypomagnesemia [serum magnesium concentration < 0.75 mmol/L], hypokalemia [serum potassium concentration < 3.5 mmol/L], and hypocalcemia [serum calcium concentration < 2.2 mmol/L]). For patients with serum albumin concentrations < 4.0 g/dL, the total serum calcium concentration was corrected according to the standard formula [12].

All exploratory data analyses were performed using SAS 9.4 (SAS Institute Inc) and KIWI Version 2 (Cognigen Corporation). The C–QTc modeling was performed using the computer program NONMEM Version 7.3.0 (ICON Development Solutions).

Exploratory data analyses and data visualization were used to determine the characteristics of the data, assess possible trends, confirm the appropriateness of the models tested, and verify model assumptions. These included inspection of QTc measurements versus time, concentration, and R–R interval (duration of ECG interval between consecutive waves) as well as the possible time delay between QTc and concentration (i.e., hysteresis).

The first-order conditional estimation method, with interaction as implemented in NONMEM, was used for the C–QTc analysis. QTcF data were modeled using quizartinib and AC886 concentrations as independent variables. The base model included the baseline term to estimate the typical value of baseline QTcF along with interindividual variability. To account for circadian variation, the time effect on baseline QTcF was also included in the model. Briefly, baseline QTcF data were split into 10 bins according to 24-h clock time (each bin had approximately the same number of samples), and the mean differences in QTcF for the 10 bins of time were estimated by the model. For determination of the structural model, various functions of quizartinib and AC886 concentrations were evaluated, with interindividual variability estimated for applicable parameters. Explored models included a linear model having separate slopes for quizartinib and AC886 concentrations and a maximum effect (E_max) model of quizartinib and AC886 concentrations, with and without a Hill coefficient, to account for sigmoidicity. Once the C–QTc model for QTcF was determined, ∆QTcF was obtained by subtracting the baseline terms.

Model selection was determined using several criteria, including decrease in NONMEM objective function value, inspection of goodness-of-fit plots, and decrease in estimates of interindividual variability (IIV) and/or residual variability (RV). Prediction-corrected visual predictive checks (pcVPCs) were performed for major steps during the model-building process. A detailed description of the covariate modeling and model building process is available in Online Resource 1.

The appropriateness of IIV and RV models was reevaluated during model refinement after covariate analysis. This procedure included reassessment of the assumed distributional form (normal or log-normal) of each IIV term as well as the functional form (additive or proportional) of the RV model. A number of models were evaluated, and the best model was selected based on numerical and graphical criteria.

The final C–QTc model was utilized to obtain the model-predicted mean ΔQTcF and 90% confidence interval (CI) at the maximum concentration at steady state (C_max,ss). Individual C_max,ss values for patients enrolled in QuANTUM-R who received quizartinib 60 mg were generated using the population PK model [13], then the geometric mean of quizartinib C_max,ss was obtained. When making predictions of ΔQTcF, the bootstrap method was used to account for multiple parameters in the model and the correlation between the parameters [14]. First, 1000 random samples of C–QTc model parameters were generated from the mean and covariance matrix of the final C–QTc model. Then, corresponding ΔQTcF values were computed at the C_max,ss; 90% CI was obtained by the 5th and 95th percentiles from the distribution.

The majority of patients were white (74.3%), with slightly more women (54.0%) than men. Patient age ranged from 19 to 81 years, with a median of 55 years, and body weight ranged from 39.5 to 147 kg, with a median of 70.0 kg. The median value for baseline QTcF in the analysis population was 414 ms (Table 1).

Figure 1 shows the observed mean ΔQTcF per visit from the QuANTUM-R study. All tested covariates were baseline values except for the use of QT-prolonging drugs and serum electrolyte imbalances, such as hypocalcemia, hypokalemia, and hypomagnesemia, which were tested as time-varying covariates.

QTcF was chosen as the preferred dependent variable for this analysis because it was used for dose increase/reduction criteria in the clinical study. Furthermore, the Fridericia correction method was recommended in a recent white paper on QTc analysis [14]. To confirm that the QT interval was properly corrected, scatterplots of triplicate QTcF values from pre- and post-treatment versus R–R interval, stratified by sex, were examined (Online Resource 2). The results suggested that QTcF was nearly independent of R–R interval and confirmed the use of QTcF as an adequate correction of QT interval in this population. Notably, QTcF did not increase when heart rate increased. Approximately 10% of patients had a > 25% increase in heart rate from baseline as well as a heart rate ≥ 100 beats per minute (i.e., R–R interval ≤ 600 ms) measured from centrally read ECGs, demonstrating an adequate number of patients to assess the effect of elevated heart rates on QTcF.

No consistent trend for hysteresis was observed in the drug concentration versus QTcF plot. Therefore, a direct relationship between drug concentration and QTcF was considered appropriate and was used as the base model structure.

To investigate possible variations in the QTc data, pretreatment QTcF data were plotted according to 24-h clock time using 10 bins with approximately the same number of samples in each bin. The plot showed considerable variation in pretreatment QTcF values versus time before administration of quizartinib, which was attributed to circadian rhythm. Therefore, the use of baseline correction was necessary. Because the majority of pretreatment QTc measurements in QuANTUM-R were obtained between 8:00 am and 5:00 pm, there were insufficient pretreatment data to support the inclusion of a full 24-h circadian rhythm function for the correction of baseline. Instead, a fixed time effect model for baseline correction was implemented [14, 15] by estimating the mean QTcF difference for each of the 10 bins (Online Resource 3).

To determine the structural model, a linear model with different slopes for quizartinib and AC886 and a sigmoid E_max model with separate E_max terms for quizartinib and AC886 were evaluated. The sigmoid E_max model had a lower Akaike information criterion (AIC) value than the linear model (AIC: 17,628.623 vs 17,784.369), supporting a better fit to the data than the linear model. This corroborated the trend exhibited in the exploratory data analysis plots, in which a possible plateauing of QTcF increase was observed at the higher concentration range. Therefore, the C–QTc data were modeled with E_max functions of quizartinib and AC886, including a baseline term and interindividual variabilities on E_max and baseline terms. Covariate analysis identified hypokalemia as a statistically significant predictor of baseline QTcF in the C–QTc model, with both quizartinib and AC886 concentrations as predictors of response. Further evaluation of IIV and RV functional forms during model refinement suggested that the model with log-normal distribution of IIV for the intercept, normal distribution of IIV for the E_max terms, and proportional RV best fit the data, as indicated by a lower objective function value and better precision of parameter estimates (Online Resource 4). The final model parameter estimates, along with corresponding precisions (relative standard error expressed as a percentage [% RSE]) for the final C–QTc model, are presented in Table 2. The model parameters for baseline QTcF, E_max for quizartinib, half maximal effective concentration (EC₅₀) for both quizartinib and AC886, and gamma for quizartinib were estimated with good precision (< 20% RSE), while the E_max and gamma for AC886 and the effect of hypokalemia on baseline QTcF were estimated with moderate precision (44.5%, 54.3%, and 31.5% RSE, respectively). Random effect parameters, IIV and RV, were estimated with good precision (≤ 30% RSE). The goodness-of-fit plots were reasonably unbiased and suggested that the model was able to characterize the observed C–QTc relationship well (Daiichi Sankyo, Inc. Data on file). The pcVPCs were performed by generating 1000 replicates of the analysis dataset. The pcVPC plot shown for quizartinib concentration, which plays a major role in QTc prolongation, indicates good concordance between observed data and model predictions (Fig. 2).

The equation for QTcF based on the final model is shown below:

where QTcF_it is the observed QTcF value in the ith patient at time t; p_t is the fixed time effect on QTcF; flag_hypoki,t is the indicator variable for hypokalemia in the ith patient at time t (1 = with hypokalemia, 0 = without hypokalemia); \(\eta_{i}\) is the between-patient random effect parameter assumed to be normally distributed with mean zero and variance Ω; \({\text{Cp}}(t)_{i,p}\) is the plasma concentration of quizartinib (ng/mL) at time t in the ith patient; \({\text{Cp}}(t)_{i,m}\) is the plasma concentration of AC886 (ng/mL) at time t in the ith patient; and \(\varepsilon_{it}\) is the residual within-patient variability assumed to be normally distributed with mean zero and variance σ². \({p}_{t}\) is modeled as:

where:

\(\theta\) is the fixed time effect shift on the intercept for each bin; and \({\rm I}[\mathrm{time}=t,t\ne 0]\) is the indicator variable, with the value of 1 if the condition in brackets is true and 0 if otherwise.

Figure 3 shows the model-predicted ΔQTcF and 90% CI overlaid with the observed ΔQTcF across the range of quizartinib (Fig. 3a) and AC886 (Fig. 3b) concentrations. The model-predicted ΔQTcF includes effects from both quizartinib and AC886. Additionally, Fig. 3 shows the model-predicted ΔQTcF effect of quizartinib alone (Fig. 3c) and AC886 alone (Fig. 3d), with corresponding 90% CI, demonstrating the relative contribution of quizartinib versus that of AC886 in QT prolongation.

The geometric mean C_max,ss of quizartinib was 401 ng/mL, and the corresponding geometric mean AC886 concentration at the time of the quizartinib C_max,ss was 204 ng/mL in patients in QuANTUM-R who received quizartinib 60 mg daily. The model-predicted mean ΔQTcF at these geometric mean C_max,ss values for quizartinib 60 mg daily was 21.1 ms (90% CI, 18.3–23.6 ms). Consequently, quizartinib belongs to the class of drugs that have increased QT-prolongation risk according to the US Food and Drug Administration E14 guidance [16]. Histograms of ∆QTcF at the geometric mean C_max,ss values of quizartinib and AC886 are provided in Online Resource 5. At the geometric mean C_max,ss of quizartinib, 51.5% and 77.6% of the observations were above 20 and 10 ms, respectively. At the geometric mean C_max,ss of AC886, 41.2% and 67.2% of the observations were above 20 and 10 ms, respectively.