Population Pharmacokinetic Analysis of Ixazomib, an Oral Proteasome Inhibitor, Including Data from the Phase III TOURMALINE-MM1 Study to Inform Labelling

Ixazomib plasma concentration–time data were collected from adult patients with advanced hematologic or non-hematologic malignancies (i.e., MM, lymphoma, light-chain amyloidosis or various solid tumours) who had participated in one of ten clinical trials (Table 1; [5, 7, 8, 13‐19]). Full eligibility criteria for study participants are described in each of the original publications. Patients received oral (BSA-based or fixed) or IV (BSA-based) doses of ixazomib. Ixazomib was administered using once-weekly (days 1, 8, and 15 of a 28-day cycle) or twice-weekly (days 1, 4, 8, and 11 of a 21-day cycle) dosing regimens. All procedures performed in studies involving human participants were in accordance with the International Conference on Harmonisation Good Clinical Practice guidelines and appropriate regulatory requirements.

Plasma concentrations of ixazomib were measured pre-dose and at various timepoints post-dose during ixazomib treatment (Supplementary Table 1). Concentrations of ixazomib were measured using a validated liquid chromatography/tandem mass spectrometry assay. A reverse-phase gradient method, running at a flow rate of 0.3 mL/min on a Fortis Phenyl, 2.1 × 50 mm, 5-μm column (Fortis Technologies Ltd, Neston, UK), provided sample stacking and separation for the analytes. Ixazomib and the internal standard (¹³C₉-ixazomib) were ionised in the positive ion spray mode and detected through multiple reaction monitoring of mass transition pairs at 343.1 → 109.0 m/z and 352.1 → 115.0 m/z, respectively. Assay linearity was achieved over a concentration range of 0.5–500 ng/mL for ixazomib, and precision ranged from 1.7 to 6.1% coefficient of variation, with a bias of −4.0 to 2.3% [12].

Patients with at least three PK samples with concentrations above the lower limit of quantification (LLOQ) were included in the analysis. Ixazomib concentrations below the limit of quantification during active treatment were imputed as LLOQ/2 (0.25 ng/mL) or discarded according to the modified M6 method, in which only the first sample of a sequence of samples below the limit of quantification was set to LLOQ/2 and subsequent samples in the sequence below the limit of quantification were discarded [20].

Population PK data analysis was performed using NONMEM software (Version 7.2, ICON Development Solutions, Hanover, MD, USA [21]) for non-linear mixed-effects models, running under Perl-speaks-NONMEM (PsN [22] 4.2.0) on a grid of CentOS 7.0 linux servers, and the Intel Fortran compiler, Version 12. Analysis of results was performed using R (Version 3.0.0 or later [23]). The NONMEM models were developed to describe the log-transformed ixazomib concentrations, and estimates were based on the first-order conditional estimation method. The model was developed in a step-wise manner, guided by a pre-specified analysis plan. First, the structural model was optimised by fitting a fixed-effects model to the observed data. Subsequently, the statistical model describing the within- and between-patient variability was developed, resulting in the base model. Finally, statistically significant and clinically relevant covariates were added to the base model to produce the final model. The population PK model was developed using both IV and oral data. Structural model selection proceeded from a one-compartment model to two- and three-compartment models, and was also guided by the results of the first population PK analysis that clearly demonstrated multi-phasic disposition kinetics of ixazomib, which were best described by a three-compartment model [13].

Inter-individual variability was implemented by exponential random-effect models, corresponding to log-normally distributed individual parameters:where θ _i and θ _TV,i are the individual and typical values of the parameter, η _i is an individual-level random variable sampled from a normal distribution with a mean of zero and variance of ω ², and represents the ith individual’s deviation from the typical value.

The residual variability was described by an additive error model on the log-transformed concentrations:where Y _ik is the kth observed ixazomib concentration in the ith individual, C _ik is the corresponding individual model-predicted concentration, and ε _ik is the residual error component sampled from N(0, σ ²). The variance of the residual error was either constant or time dependent [24] with variance defined as a parametric function of time:where sd(ε) is the standard deviation of the unexplained random variability, parameterized as a function of the standard deviation at the beginning of the treatment period (SD₁), the asymptotic standard deviation (SD₀) and the rate constant of change in SD (K _SD).

Based on typical and individual parameter estimates, the characteristic half-lives associated with a three-compartment model (t _1/2,α, t _1/2,β and t _1/2,γ), as well as the steady-state volume of distribution were derived [25].

Covariates were tested for statistical significance and clinical relevance, based on the pre-defined list of relevant parameter-covariate relationships (Supplementary Table 2). All covariates were baseline covariates, except concomitant administration of CYP-modulatory drugs, which was included as a time-dependent covariate. Covariates were tested on CL and volume parameters that included a random effect. Model selection was done on the basis of a log-likelihood ratio test by iteratively adding the most significant of the evaluated covariate-parameter relationships until no more statistically significant covariates (at p = 0.01) that also led to a reduction of at least 10% in the corresponding random-effect variance were found. Subsequently, the least significant covariate was eliminated in a backward deletion procedure until no more non-significant covariates (at p = 0.001) were left in the model.

Continuous covariates were included in the population PK model as power functions, while categorical covariates were implemented as factors:where θ _TV,i is the typical parameter for patient i, defined as a function of the population typical value, θ _TV,POP and the individual contributions from continuous (X _Cont) and binary (X _Cat, with values 0 and 1) covariates. θ ₁ and θ ₂ represent the respective covariate coefficients.

The objective function value (OFV) was used to initially evaluate whether alternative models represented an improvement in overall fit compared with the candidate model. If the alternative model better fit the data, more elaborate diagnostic tools were used to further assess the quality of the model. Alternative models were compared based on the overall model fit (i.e., NONMEM OFV), precision of parameter estimates (% relative standard error), condition number, random-effect and residual error model variances, and based on the following diagnostic tools: goodness-of-fit plots, prediction-corrected visual predictive check (VPC) and bootstrap. In the VPC, quartiles of prediction-corrected observed data were superimposed on the 1000 corresponding simulated datasets to demonstrate the ability of the model to adequately describe the mean trend and variability in the observed data. Model robustness and parameter estimates were assessed by bootstrap based on 1000 resampled datasets (with replacement).

Pharmacokinetic data were available from 755 adult patients, including 632 patients with MM, 80 patients with advanced solid tumours, 28 patients with lymphoma and 15 patients with light-chain amyloidosis (Table 1; [5, 7, 8, 13‐19]); 347 patients were treated in the phase III TOURMALINE-MM1 study, in which patients with relapsed/refractory MM received once-weekly oral ixazomib in combination with lenalidomide and dexamethasone. Patients in studies C16005, C16008, C16013 and a cohort in TB-MC010034 also received ixazomib in combination with lenalidomide and dexamethasone (n = 168); all other studies used single-agent ixazomib (n = 240). Most patients received oral ixazomib (n = 647), but patients in two phase I studies received IV ixazomib (n = 108); 560 patients received weekly dosing and 195 patients received twice-weekly dosing.

A total of 10,199 PK samples were available; 172 (1.7%) were flagged for exclusion from the analysis because of suspected date and/or sampling time recording errors. Of the 354 (3.5%) samples below the limit of quantification, 124 (1.2%) samples were flagged for exclusion, while the remaining 230 samples (2.3%) were simply imputed to LLOQ/2. Given the small number of samples below the limit of quantification, a more complex imputation method (e.g., M3, M4) was not examined as this would have increased the complexity of the model without providing substantial improvements in the model fit. The demographics and baseline characteristics of patients included in the analysis dataset are summarised in Table 2.

Geometric mean and individual dose-normalised PK data obtained from the phase III TOURMALINE-MM1 study are shown in Fig. 1a. Dose-normalised ixazomib concentration–time profiles for phase I and I/II studies with 21-day treatment cycles (twice-weekly ixazomib dosing) and 28-day treatment cycles (weekly ixazomib dosing) are shown in Fig. 1b and c, respectively. In all three plots, the observed ixazomib concentration was normalised by dividing each individual sample concentration by the preceding ixazomib dose. Plots of raw concentrations (not dose normalised) are shown in Supplementary Figure 1.

The observed ixazomib plasma concentration–time data after IV or oral administration were adequately described by a three-compartment model with linear distribution and elimination kinetics, including first-order linear absorption with a lag time describing the oral dose PK profile (Fig. 2). The developed model included log-normally distributed patient-level random effects on systemic CL, absolute bioavailability of an oral dose (F) and the volume of the second peripheral compartment (V ₄). Individual CL and F were correlated, with coefficient of correlation estimated at 82%. The value of F estimated for patients receiving IV ixazomib was based on the estimated correlation between F and CL. Residual unexplained variability of the log-transformed concentrations was described by an additive error model with time-varying variance described by an exponential decline in the standard deviation (SD) from a starting level (SD₁) towards a steady-state plateau (SD₀). This residual error model has been previously described by Karlsson et al. [24] and was implemented in the analysis, as a constant additive residual error model showed bias in the conditional weighted residuals and VPCs with respect to time. The SD in the time-varying error model decreases over the first 2–4 h post-dose towards a plateau, reflecting the lower variability in the observed data at later time points.

A covariate model was built based on pre-specified covariate-parameter relationships (Supplementary Table 2). Covariates were only tested on CL and volume parameters that included a random effect. The improvement in OFV, statistical significance, and change in the corresponding inter-individual variances associated with the CL and V ₄ parameters following univariate covariate additions are shown in Supplementary Table 3. The addition of univariate covariates corresponded to the first step in the forward addition procedure. Among the parameter-covariate relationships tested, only BSA and body weight met the predefined selection criteria on V ₄ by leading to a statistically significant improvement of the model fit, and reducing the random-effect variance by at least 10%. In fact, no other covariates reduced the random-effect variance meaningfully, even when a more conservative criterion of relative reduction of inter-individual variability by ≥5% was used. Because BSA and body weight are strongly correlated body size-related covariates, only one was included in the model. BSA had a small advantage to body weight, both in terms of reducing the OFV and the random-effect variance. Consequently, BSA was included on V ₄ as the only covariate in the final model.

The PK parameter values estimated for the final population PK model are presented in Table 3. The final model provided parameter estimates with good precision (relative standard error ≤22%). Furthermore, shrinkage of individual random effects was reasonable (≤37%) with shrinkage of individual CL estimates of 21% in the final model. The geometric mean (95% confidence interval) of the individual estimated terminal disposition phase half-life (t _1/2,γ) was 9.53 (9.32–9.75) days and the steady-state volume of distribution was 543 (534–551) L. These estimates were similar to the population estimates for t _1/2,γ (9.35 days) and steady-state volume of distribution (527 L). The absolute oral bioavailability (F) was estimated to be 58%.

To evaluate the appropriateness of the statistical components of the final model, a prediction-corrected VPC was performed [26]. Figure 3 shows a VPC based on data from the first 4 weeks of all studies (Fig. 3a) and from the phase III TOURMALINE-MM1 study (Fig. 3b). Apart from a few time bins, the observed median and 2.5th–97.5th percentiles of data fall within the corresponding 95% confidence interval of the model simulations, indicating that the model adequately describes the central tendency and total variability of the data.

Figure 3c shows the goodness-of-fit plots of the final model based on data from all studies and data from the phase III TOURMALINE-MM1 study. The plots show that both population and individual predictions vs. the observed data scatter randomly around the line of unity, indicating an unbiased fit. As expected, individual predictions are noticeably closer to the line of unity than corresponding population predictions, illustrating the impact of the random-effects model for explaining variability in the observed data. Figure 3d shows the patient-specific random effects (ETA) on V ₄ estimated with the base and final models, plotted against individual BSA, and confirms that including BSA as a covariate on V ₄ in the final model removes the trend in the random effect on V ₄ vs. BSA that was observed with the base model.

Based on the developed final model, the individual predicted exposure following a single oral 4-mg ixazomib dose was calculated using the individual estimated CL and absolute bioavailability as AUC_∞ = 4 mg × F_i/CL_i. To evaluate trends in individual predicted exposure, AUC_∞,_i was described using linear regression models including individual covariates as predictors. This corresponds to a graphical covariate screening where individual parameters are plotted vs. covariate values. If no trend is observed, the parameter can be described by the population typical value and random variability. Figure 4 shows the correlation between key continuous covariates and individual predicted exposure and the prediction of the linear regression at the 5th and 95th percentile of continuous covariates relative to the median value. Across the analysis population, the 5th and 95th percentiles of individual predicted exposures were −50 and +98% relative to the median AUC_∞. For each continuous covariate, the magnitudes of percent difference in AUC_∞ at the 5th or 95th percentiles relative to the AUC_∞ at the median of individual covariate values were <20%, and are therefore well below the variability in AUC_∞ in the study population. This was consistent with these covariates not being identified as statistically or clinically meaningful covariates on CL. Similarly, as shown in Fig. 5, for categorical covariates, the magnitude of AUC_∞ in different categories relative to the most common category were <20% in all cases with the exception of Asian race, where the mean AUC_∞ was 35% higher than in White race group. However, there was substantial overlap in exposure between the White and Asian race groups. Taken together, these analyses confirm the absence of clinically meaningful effects of the evaluated covariates on the systemic exposure (AUC_∞) of ixazomib in cancer patients.

The final population PK model was simulated to show the typical concentration–time profile following weekly treatment with oral ixazomib 4 mg in a 28-day treatment regimen (dosing on days 1, 8 and 15). All model parameters were set to the respective typical values with V ₄ normalised to the median of individual BSA values reported in the analysis dataset (1.88 m²). Figure 6 shows the simulated concentration–time profile for a typical patient during the first two treatment cycles.