Study design and data

The exposure–response analyses were conducted using data from the phase 3 CheckMate 9ER trial of nivolumab plus cabozantinib versus sunitinib in first-line advanced RCC. CheckMate 9ER was an open-label, randomized phase 3 trial of nivolumab (240 mg intravenously every 2 weeks) plus cabozantinib (40 mg orally QD) versus sunitinib (50 mg orally QD for 4 weeks on treatment, then 2 weeks off) in patients with previously untreated advanced or metastatic RCC [11, 13]. The primary endpoint was PFS per Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1 by blinded independent central review (BICR). Randomization was stratified by International Metastatic RCC Database Consortium (IMDC) risk score, tumor programmed death-ligand 1 (PD-L1) expression, and geographic region. Patients were ≥ 18 years of age with a clear-cell RCC histology and measurable disease per RECIST v1.1.

A total of 651 patients were randomized to receive nivolumab plus cabozantinib (n = 323) or sunitinib (n = 328) until disease progression or unacceptable toxicity, with a maximum duration of 2 years for nivolumab treatment. Dose holds to manage AEs were permitted for both cabozantinib and nivolumab. Dose reductions to manage AEs were only permitted for cabozantinib: from 40 to 20 mg QD, and then to 20 mg every other day (QOD).

Radiographic assessments were performed at screening, Week 12, then every 6 weeks until Week 60, then every 12 weeks until progression. Safety was assessed every 2 weeks. Blood samples for PK assessment of cabozantinib were collected in Cycles 1, 3, 4 and 7 (cycle = 2 weeks for nivolumab), at the same time as pre-dose PK samples for nivolumab, provided the time of the draw was approximately ≥ 8 h after the previous evening dose of cabozantinib to ensure sampling occurred after the Tmax of roughly 3–4 h [14]. Plasma concentration analyses for cabozantinib were performed by a validated liquid chromatography– tandem mass spectrometry assay (lower limit of quantitation was 0.5 ng/mL).

Population pharmacokinetic model

The PopPK analysis included plasma cabozantinib concentration–time data from CheckMate 9ER and 10 studies from a previously developed integrated PopPK model that characterized cabozantinib concentration data from healthy subjects and patients with various types of malignancies [15, 16]. Protocols for all studies were approved by institutional review boards of participating institutions and written informed consent was obtained from patients prior to enrollment.

The analytical data preparation has been previously described [15]. The previous PopPK model comprised a 2-compartment model with first-order elimination and dual-absorption (first-order + zero-order) processes [16]. That model was used as the base model for the current analysis. The impact of previously evaluated demographic covariates and cancer type on cabozantinib apparent clearance (CL/F) was re-evaluated in the updated dataset, which included data from CheckMate 9ER patients. An additional covariate effect related to combination therapy with nivolumab was also assessed. The final PK model was used to generate individual predicted cabozantinib exposures using dosing history from patients in the CheckMate 9ER trial for the exposure–response analyses.

Exposure–response analysis for time-to-event endpointsEfficacy and safety endpoints

Time-to-event analyses were performed to characterize the exposure–response relationship between cabozantinib exposure and PFS (defined as the time from randomization to radiographic progression per RECIST v1.1 by BICR, or death), cabozantinib dose modification, and each of the following AEs commonly associated with cabozantinib dose holds and reductions: palmar-plantar erythrodysesthesia (PPE; Grade ≥ 1), diarrhea (Grade ≥ 3), hypertension (Grade ≥ 3 [systolic BP > 160 mmHg or diastolic BP > 100 mmHg]), fatigue/asthenia (Grade ≥ 3), and alanine aminotransferase/aspartate aminotransferase (ALT/AST) elevation (Grade ≥ 3). An intended time-to-event analysis assessing the relationship of OS with cabozantinib exposure was not performed as there were too few events and limited follow-up at the time of the exposure–response analysis.

Dataset construction and preliminary analyses

Individual average cabozantinib concentrations from time zero to the event or censoring time (CAVG0T) were calculated from the estimated individual PK parameters of the final PopPK model and individual dosing history; CAVG0T was used for the exposure–response analysis dataset to link clinical endpoints with individual predicted cabozantinib exposure.

The average cabozantinib concentration, CAVG0T, in units of ng/mL was defined as (Eq. 1):

where t represents time in days, and AUC is the area under the concentration–time curve (concentration × time in hours). CAVG0T is a time-invariant exposure metric. For the cabozantinib dose-modification endpoint, individual predicted apparent clearance was used to represent cabozantinib exposure. Patients who had at least one documented cabozantinib dose without any measurable PK concentration were assigned population-level PK parameter estimates for prediction of exposure.

Kaplan–Meier (KM) plots were constructed for each efficacy and safety endpoint prior to model development. Patients were grouped into quartiles based on CAVG0T or apparent clearance for corresponding clinical endpoints. The preliminary analyses were performed to identify potential relationships between cabozantinib exposure and each clinical endpoint. Only time to the first event was considered for the analyses of safety endpoints.

Cox proportional hazards (CPH) model

CPH models were used to describe the relative hazard for each of the clinical efficacy and safety endpoints.

The general form of the CPH model is represented by the equation (Eq. 2):

$$h\,(t,X)\, = \,h_ (t) \cdot }\,(\beta \cdot X)$$

(2)

where h (t, X) denotes the hazard at time t, ho(t) is the background hazard function, β is a vector of the regression coefficients, and X is a matrix of covariates.

Model development

Separate CPH models were developed for each of the clinical endpoints using data from the CheckMate 9ER trial. The impact of cabozantinib exposure on relative hazard was evaluated during base model development. A drop in − 2 log likelihood (− 2LL) when including cabozantinib exposure in the CPH model was considered significant. Both linear (Eq. 3) and nonlinear (Eq. 4) functional forms were evaluated:

$$h\,(t,\,X_ )\, = \,h_ (t) \cdot \exp \,(\beta_ \cdot X_ )$$

(3)

$$h(t, ^}}}_)=_\left(t\right)\cdot \mathit\left(_ \cdot ^}}}_\right)\,where\ ^}}}_= \frac_}_+_}$$

(4)

where Xex is the cabozantinib exposure (e.g, CAVG0T), βex1 represents the slope in the log-linear model, βex2 represents the maximum drug effect in the Emax model, and EC50 represents cabozantinib exposure at which half of the maximal effect is achieved.

Covariate analysis

Covariate effects were assessed for the PFS model. Covariates were added to the base model simultaneously to form a full model followed by stepwise backward elimination to identify the most parsimonious model. Statistical significance of covariate-parameter relationships was assessed with the Wald test, at α = 0.1 for inclusion. At each step of the backward elimination procedure, the least significant covariate–parameter relationship was eliminated, and the procedure was repeated until all covariates–parameters met the inclusion criteria.

Covariates evaluated were based on clinical judgment and mechanistic plausibility and included: age, baseline PD-L1 + tumor expression (≥ 1% vs. < 1% or indeterminate), baseline IMDC score (0 vs. 1–2 vs. 3–6), baseline Karnofsky performance status (≥ 90 vs. < 90), prior adjuvant or neo-adjuvant therapy for localized or locally advanced RCC (Yes vs. No), prior nephrectomy (Yes vs. No), prior radiotherapy (Yes vs. No), baseline lactate dehydrogenase (LDH) level (≤ 1.5 × upper limit of normal [ULN] vs. > 1.5 × ULN), time from initial disease diagnosis to randomization (< 1 year vs. ≥ 1 year), sex, baseline body weight, tumor burden (sum of the diameter of target lesions at baseline, > median vs. ≤ median), baseline albumin (5th percentile vs. median, 95th percentile vs. median), baseline nivolumab clearance (5th percentile vs. median, 95th percentile vs. median), and liver metastasis (Yes vs. No).

Final model evaluation and simulations

The predictive performance of the PopPK model and time-to-event models was evaluated using visual predictive checks (VPC). Simulations were also performed using the final CPH model parameter estimates to predict the incidence of efficacy and safety endpoints for different dose levels.

Software

For PopPK modeling, analyses were performed using nonlinear mixed effects modeling methodology as implemented in the NONMEM software system, version 7.3 (ICON Development Solutions, Ellicott City, MD). For exposure–response analysis, time-to-event analyses were performed using the Cox proportional hazards regression (PHREG) procedure within SAS (version 9.4). Graphical analysis of the data or output from the models was performed using R software (version 3.6.1).