Alternative dosing regimens for atezolizumab: an example of model-informed drug development in the postmarketing setting

Data from atezolizumab monotherapy studies, including the phase 1 study PCD4989g (NSCLC and UC cohorts only; ClinicalTrials.gov ID, NCT01375842) [3, 10‐12] and phase 3 studies OAK (NSCLC, GO28915; ClinicalTrials.gov ID, NCT02008227) [13] and IMvigor211 (UC, GO29294; ClinicalTrials.gov ID, NCT02302807) [14] (Table S1), were used in the ER analyses based on clinical data cut-off dates of December 2, 2014; July 7, 2016; and March 13, 2017, respectively. OAK and IMvigor211, randomized studies of atezolizumab monotherapy vs chemotherapy, were used in tumor growth inhibition (TGI) modeling analyses. PK simulations were performed based on the popPK model [6] developed with Phase I data (n = 472) from PCD4989g and JO28944, the phase 1 study conducted in Japanese patients [15] and validated with data from over nine studies, including phase 3 studies OAK, IMvigor211, and IMpassion130 (WO29522; ClinicalTrials.gov ID, NCT02425891) [16]. Safety data for select patient subgroups from PCD4989g (using a data cut-off date of March 31, 2016) and OAK were summarized. The protocols for studies from which data were analyzed were approved by institutional review boards and/or independent ethics committees at each site. All patients provided written informed consent.

ER analyses were performed to inform any relationships between PK metrics and ORR, OS, grade 3–5 adverse events (AEs), and AE of special interest (AESI) endpoints evaluated in previous clinical studies based on cycle 1 data to minimize potential bias due to both confounding with baseline prognostic factors [17, 18] and time-dependent variation in clearance that has been observed for atezolizumab and other checkpoint inhibitors [4, 8, 19‐22]. These analyses were conducted using pooled data from atezolizumab-treated patients with NSCLC or UC (from PCD4989g, OAK, and IMvigor211) for whom exposure data were available, except as noted below for overall survival (OS). Exploratory ER analyses were performed using cycle 1 maximum serum concentration (C_max), C_min, and area under the concentration–time curve (AUC; time 0–21 days), as recommended [21] to minimize the effect of response-dependent time-varying clearance observed previously for anti-PD-1 and anti-PD-L1 agents [20]. AUC (time 0–21 days), C_max, and C_min were derived at cycle 1 based on individual PK parameters estimated using cycle 1 data only and the previously developed popPK model [6]. The efficacy endpoints evaluated were investigator-assessed confirmed Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1) objective response rate (ORR; secondary endpoint in all studies) and OS (primary endpoint in OAK and IMvigor211). ORR analyses used data from atezolizumab-treated patients with NSCLC or UC in PCD4989g, OAK (first 850 randomized patients), and IMvigor211, whereas OS analyses used data from OAK (first 850 randomized patients) and IMvigor211 only. The safety endpoints evaluated included AEs of grades 3–5 per National Cancer Institute Common Terminology Criteria for Adverse Events version 4 and Medical Dictionary for Regulatory Activities version 20.1 (primary endpoint in PCD4989g, also evaluated in OAK and IMvigor211) and AESIs (evaluated in all studies). AESIs, conditions suggestive of autoimmune disorder, were defined previously [11].

ORR and AEs were evaluated as binary endpoints (yes/no) and studied vs exposure as a continuous variable using logistic regression. The Wald test P value was reported for each logistic regression, along with proportions/frequencies and their 95% CIs computed for quartiles of exposure. For OS data, to mitigate confounding factors between patients’ baseline information and atezolizumab clearance and exposure, TGI-OS modeling [23, 24] was performed. To be evaluable in this analysis (TGI evaluable), patients needed to have ≥ 1 posttreatment sum of longest diameters (SLD) assessment. The impact of individual baseline prognostic factors and TGI metrics (estimated tumor shrinkage and tumor growth rates in a biexponential longitudinal model of the SLD of the target lesions per RECIST 1.1) on OS were explored using Kaplan–Meier and Cox regression analyses, and a parametric multivariate regression TGI-OS model was built. The final TGI-OS model was validated by simulation in its ability to describe OS distributions and hazard ratios (HRs) compared with a control in different subgroups (notably by exposure quartiles). For the HR simulations, TGI metric estimates and baseline covariates for control patients were taken from previous analyses [24, 25]. Exposure metrics were tested on the final multivariate model after adjustment for confounding with prognostic factors. A “tumor type” factor could be incorporated in the model if appropriate.

To simulate PK parameters of varying regimens of atezolizumab [840 mg q2w, 1200 mg q3w, 1680 mg every 4 weeks (q4w), and 20 mg/kg q3w], Monte Carlo simulations were performed using the popPK model of atezolizumab, including covariate effects, previously developed using PCD4989g data [6] to obtain virtual individual PK profiles at cycle 1 and steady state. In the popPK model used for the PK simulations, bodyweight, albumin, tumor burden, treatment-emergent antidrug antibody (ADA) status, and gender were found to have a statistically significant impact on atezolizumab PK [6]. A single replicate of 500 patients was simulated for each regimen. A seed number was provided in the control stream to ensure reproducibility of the simulations. Random effects were sampled from the previously estimated distribution, and the residual error was not taken into account for individual predictions. Virtual patients per dosing regimen were assumed to have a 1:1 male:female ratio (males weighing 85 kg and females weighing 64 kg, median body weight in the phase 1 database used to develop the popPK model). Other covariates affecting atezolizumab PK parameters were set to the median or most frequent category for the categorical covariates: albumin level of 40 g/L, baseline tumor size of 63 mm, and negative for ADAs. Four dosing regimens were simulated: 1200 mg q3w, 20 mg/kg q3w (i.e., 1700 mg for males and 1280 mg for females), 840 mg q2w, and 1680 mg q4w. To assess the impact of body weight on exposure after the fixed-dose regimen, 500 virtual patients per quartile of body weight with median albumin level, baseline tumor size, and negative for ADAs were assigned a dose of 840 mg q2w or 1680 mg q4w. The distribution of body weight in the phase 1 population of patients was divided by quartiles as follows: 36.5–63.7, 63.7–77.0, 77.0–90.9, and 90.9–168.0 kg. The 500 individual body weights were sampled in each quartile assuming a truncated normal distribution. To maintain the correlation between sex and body weight, the proportion of females was set to 80% in the first quartile, 50% in the second quartile, 25% in the third quartile, and 10% in the last quartile, as observed in the phase 1 database used to develop the popPK model.

Atezolizumab exposure metrics (cycle 1: AUC [calculated using the trapezoidal method; time 0–21 days], C_max, and C_min; steady state: AUC [dose/clearance], C_max, and C_min) were derived from the simulated individual PK profiles and summarized across individuals for each dosing regimen. To compare several dosing regimens involving different dosing intervals (every 2, 3, or 4 weeks), steady-state weekly AUC data were also derived.

A prediction-corrected visual predictive check (pcVPC) was performed based on the prior phase 1 popPK model (external evaluation). The phase 1 popPK model was used to derive the individual PK parameter estimates based on atezolizumab observed concentration–time profiles in IMpassion130. PK data for atezolizumab-treated patients in IMpassion130 were simulated (1000 replicates) using actual dosing and patient covariates (body weight, sex, ADA status, albumin level, and tumor burden) and the phase 1 popPK model. Observed atezolizumab peak (C_max) and trough (C_min) concentrations in IMpassion130 were compared with corresponding predictive distributions.

AE frequencies were summarized for subgroups of patients: (1) from PCD4989g who received atezolizumab 20 mg/kg q3w based on C_max values in relation to predicted C_max for the 1680 mg q4w regimen, (2) from PCD4989g based on the atezolizumab dose received, and (3) from PCD4989g and OAK based on body weight quartiles (lowest quartile vs quartiles 2–4). In the PCD4989g study, adverse events were collected until 90 days following the last administration of study treatment or until initiation of the subsequent anticancer therapy, whichever occurred first (Figure S6, footnote 1). In these analyses, whether or not AESIs required the use of corticosteroids was also specified.

Data set preparation, exploration, visualization, and analysis, including descriptive statistics, were performed using R version 3.4.3 and Comprehensive R Archive Network packages. Non-linear mixed-effect modeling using the first-order conditional estimation algorithm with interaction [Non-Linear Mixed-Effect Modeling tool (NONMEM) version 7.3; ICON Development Solutions, Ellicott City, MD, USA] [26] was used for Bayesian estimation of individual PK parameters. Logistic regression used the generalized linear model function in R with family “binomial” (variance = binomial; link = logit). Monte Carlo PK simulations were implemented using NONMEM version 7.3, and simulation data sets to assess were created using R.

Online resource Table S1 summarizes the patient populations and dosing regimens for the analyses in this study. A total of 164 of 1042 evaluable patients with exposure data (15.7%) achieved a confirmed investigator-assessed ORR. No difference in ORR was observed between patients with NSCLC (15.6%, n = 501) or UC (15.9%, n = 541), so tumor type was not included in the logistic regression models. Neither AUC nor Cmin at cycle 1 was significantly related to the probability of response (Fig. 1).

For OS, a multivariate model was developed to account for baseline prognostic factors and TGI metrics. Median OS in OAK patients with NSCLC [n = 388 TGI evaluable of 425 intention-to-treat (ITT) patients (91%)] was 467 days (95% CI 402–508 days) and in IMvigor211 patients with UC [n = 382 TGI evaluable of 467 ITT patients (82%)] was 344 days (95% CI 290–383 days). Based on this difference, tumor type was incorporated in the model. Of 770 TGI-evaluable patients, 764 had exposure data. Individual estimates of Log(tumor growth rate [KG]) and baseline factors (e.g., Eastern Cooperative Oncology Group performance status > 0; tumor size; albumin, lactate dehydrogenase, and alkaline phosphatase levels; PD-L1 status; and tumor type) were independent predictors of OS (Table S2). Of note, after accounting for baseline covariates in the final model, cycle 1 atezolizumab exposure (AUC, C_min, or C_max) was not significant (P > 0.01). The model performed well in simulating OS distribution and HRs by exposure quartiles for each tumor type even if exposure was not in the model. Comparisons of predicted and observed OS data are included in online resource Fig. S1 and Fig. S2. The flat ER relationship of atezolizumab was also illustrated in a simulation of the HRs by AUC quartiles after adjusting for baseline covariates (fixed to median values) and resampling KG as estimated (Fig. 2).

Pooled atezolizumab exposure–safety analyses were performed on all atezolizumab-treated patients with locally advanced or metastatic NSCLC or UC with exposure data (n = 1228). AEs of grade ≥ 3 and AESIs occurred in 209 (17.0%) and 298 (24.3%) of 1228 patients, respectively. AE frequencies were similar in patients with NSCLC compared with UC (14.9% vs 19.6% for grade ≥ 3 AEs; 24.6% vs 23.9% for AESIs); therefore, tumor type was not included in the logistic regression models. Neither safety analysis (incidence of grade ≥ 3 AEs or AESIs) showed any statistically significant ER relationship with the cycle 1 exposure metrics, AUC and C_max (Fig. 3).

Simulated atezolizumab exposure profiles for four dosing regimens—1200 mg q3w, 840 mg q2w, 1680 mg q4w, and 20 mg/kg q3w—are presented in Fig. 4. A summary of the exposure metrics associated with each dosing regimen is shown in Table 1.

For the 840 mg q2w dosing regimen, the predicted C_min was 13% lower at cycle 1 and 16% higher at steady state than that for the 1200 mg q3w dosing regimen and was also in excess of the C_min target concentration of 6 μg/mL [7] by > tenfold. The 840 mg q2w predicted C_max was lower than the 1200 mg q3w C_max at cycle 1 and steady state. For the 1680 mg q4w dosing regimen, the predicted C_min was 14% higher at cycle 1 and 6% lower at steady state than that for the 1200 mg q3w dosing regimen and was also in excess of the C_min target concentration of 6 μg/mL by > tenfold. The 1680 mg q4w-predicted C_max was 12% higher at cycle 1 and 0.8% higher at steady state relative to the predicted geometric mean C_max for the 20 mg/kg dosing regimen, and was consistent with observed exposures for the 20 mg/kg q3w dosing regimen in PCD4989g [6, 27]. The predicted weekly AUC for the regimens of 840 mg q2w and 1680 mg q4w at steady state were higher than those simulated for 1200 mg q3w by 3.5% and 4.8%, respectively.

When considering fixed-dose regimens, since clearance and volume are impacted by body weight in the atezolizumab popPK model [6], patients with lower body weight would be expected to exhibit higher atezolizumab exposure relative to heavier patients. To further evaluate the q2w and q4w regimens, C_min or C_max was simulated by quartiles of body weight for dose levels of 840 mg q2w and 1680 mg q4w (Table S3). For the 1680 mg q4w regimen, the predicted C_max values for the lowest body weight quartile (< 63.7 kg, with a majority of females) were 692 and 950 μg/mL for cycle 1 and steady state, respectively, which is within the range of the observed C_max values for 1200 mg q3w and 20 mg/kg q3w [6, 27]. For the 840 mg q2w regimen, the predicted C_min values for the highest body weight quartile (> 90.9 kg, with a majority of males) were 58 and 158 μg/mL for cycle 1 and steady state, respectively, which is within the range of the observed C_min values for 1200 mg q3w and above the C_min target concentration of 6 μg/mL.

As an external evaluation of phase 1 popPK model and to confirm the 840 mg q2w PK simulations, the PK of the atezolizumab plus nab-paclitaxel q2w arm from the IMpassion130 study were simulated based on baseline patient covariates (pcVPC). Four-hundred forty-three (of 445) atezolizumab-treated patients had evaluable serum samples for PK analysis, for a total of 2232 samples. Results are presented in Fig. 5. Both dose 1 and steady-state exposure metrics were similar to those predicted for the 840 mg q2w dosing regimen based on the phase 1 popPK model. A trend toward underprediction of the median and fifth percentile of atezolizumab exposure data (troughs) after longer-term administration (doses 2, 4, 6, 14, and 30 +) was observed for the popPK model, consistent with the time-dependent clearance of atezolizumab [4].

Table S4 provides a safety summary for 20 mg/kg q3w atezolizumab-treated patients in PCD4989g, with observed C_max during cycle 1 relative to the mean predicted C_max of the 1680 mg q4w regimen. In general, AE frequencies were similar between these groups. Similar results were obtained in groups based on the PCD4989g patients’ modeled C_max (i.e., individual predictions estimated by the popPK model) relative to the mean predicted C_max of the 1680 mg q4w regimen.

Table S5 provides a summary of atezolizumab exposure by dose group. In a dose range from 10 mg/kg q3w to 20 mg/kg q3w and 1200 mg q3w, the median treatment duration ranged from 2.07 to 9.48 months, and the median number of doses ranged from 4 to 14.5. Table S6 provides a safety summary for PCD4989g patients by dose group. The overall safety profile was consistent among the 15 mg/kg q3w, 20 mg/kg q3w, and 1200 mg q3w dose groups. Patients in the 10 mg/kg q3w dose group demonstrated increased frequency of serious AEs and treatment-related AEs relative to the other dose groups. This may be due to the longer safety follow-up and the lower number of patients in this dose group relative to the other dose groups.

Table S7 provides a safety summary for PCD4989g and OAK patients by body weight. Median body weight in the 20 mg/kg treatment group in PCD4989g was 78.2 kg (Q1–Q3, 63.7–93.0 kg), and the overall safety profile was generally similar between patients in the lowest (n = 37) and upper 3 (n = 109) body weight quartiles. A higher incidence of grade 3–5 AEs (48.7% vs 37.3%) in the lowest body weight quartile subgroup was observed, which was due to grade 3 AEs (38.8% vs 27.8%). Evaluation of grade 3 AEs did not identify any individual AE preferred term with a  ≥ 2% difference between subgroups. Serious AEs with a  ≥ 5% difference between subgroups included fatigue and asthenia (both common to malignancy) as well as pneumonia and cardiac tamponade (known complications of thoracic cancers), with all such events occurring infrequently. In the lowest body weight subgroup, only asthenia and respiratory complications led to study treatment withdrawal; no action with respect to study treatment was taken for the other events. To assess the impact of body weight in a larger cohort of patients, AE data from OAK (1200 mg q3w dosing) were also analyzed. Median body weight was 71.0 kg (Q1–Q3, 59.5–82.2 kg). No differences between the lowest (n = 152) and upper 3 (n = 442) body weight quartiles were observed.