Anna Mc Laughlin 1, Martin Bergstrand 1, Justyna Kanska 2, Ana Ruiz-Garcia 3, Simone Filosto 2, Rhine Shen 2
1 Pharmetheus (Uppsala, Sweden), 2 Kite, A Gilead Company (Santa Monica, USA), 3 Gilead Sciences (Foster City, USA)
Objectives: Chimeric antigen receptor (CAR)-T cell therapies display pharmacokinetics (PK) characterized by distribution, expansion, and elimination. These PK are often characterized via piecewise-linear models [1-6], in which parameters are switched on and off to reproduce the observed profiles. While these models describe the PK adequately, the lack of a mechanistic basis limits the suitability for covariate exploration and extrapolations/simulations. The aim of this work was to establish a fit-for-purpose modeling framework to support model-informed drug development (MIDD) for CAR-T cell therapies [7]. Specifically, we (1) developed a parsimonious, mechanism-based population PK (popPK) model with a latent tumor size (TS) component (Model 1); (2) validated its structural validity against an integrated popPK-TS dynamic model (Model 2); and (3) demonstrated its utility by identifying significant covariates explaining variability in a clinical dataset.
Methods: Data from 473 patients with relapsed or refractory large B cell, follicular, or mantle cell lymphoma receiving axicabtagene autoleucel (studies ZUMA-1, ZUMA-5, ZUMA-7) or brexucabtagene ciloleucel (study ZUMA-2) was used for model development. Model 1 was developed using peripheral blood (PB) CAR-T cell PK and baseline sum of the product of diameters (SPD) data from all patients. Model 2 was an extension of Model 1, in which the latent tumor component was replaced with a TS dynamic component, and √SPD was used as a dependent variable, consistent with previous TS modeling for B cell lymphoma [8]. Parameters were estimated using PB CAR-T cell PK and longitudinal TS data from 95 evaluable patients enrolled in ZUMA-1. The structural validity for PK of Model 1 was assessed by comparing the parameter estimates and the predictive performance for the PK component shared by both models. To explore Model 1’s ability to identify covariates explaining interindividual variability (IIV) in PK parameters, stepwise covariate modeling with adaptive scope reduction [9] was performed.
Results: The PK component shared by Models 1 & 2 was a two-compartment (CMT) model with linear elimination from the central CMT. Patients’ individual CAR-T cell doses were modeled as i.v. infusions into the central CMT. In Model 1, the individual baseline SPD was used to initialize the latent tumor CMT. In Model 2, the TS at baseline was a combination of sensitive and resistant tumor [10-12]. Both models postulated TS-driven expansion of CAR-T cells and CAR-T-cell-dependent reduction in TS, where TS was the amount in the latent tumor CMT in Model 1 and the amount in the sensitive tumor CMT in Model 2. In Model 2, the CAR-T cell driven decline in observed TS lagged behind the TS-driven CAR-T cell expansion. Integrating concepts from pre-clinical TS modeling [13], this delay was accounted for by including a transit CMT between the sensitive tumor killed by CAR-T cells and eliminated tumor. Additionally, mixture modeling with three populations was used to capture different TS dynamic patterns. For population 1 (25.2% of patients), the fraction of sensitive tumor at baseline (FR) was set to 1, resulting in profiles of sustained undetectable tumor after CAR-T cell infusion. FR was estimated for populations 2 & 3. Population 2 (53.9% of patients) captured profiles of TS decline up to a detectable plateau, and population 3 (20.9% of patients) characterized profiles of TS increase with/without initial decline. While FR was similar in populations 2 & 3, the estimated growth rate of resistant tumor was higher in population 3 vs. population 2. Both models characterized the observed PK data well and the estimated PK parameters were highly congruent between the approaches. The covariate analysis for Model 1 identified product/disease type, the % effector cells in the infusion product, prior bendamustine therapy, baseline body weight, and baseline albumin as statistically significant. However, only the effect of product/disease type resulted in deviations of the predicted Cmax and AUC0-28d by >25% vs. the reference.
Conclusion: The ability of the mechanism-based, parsimonious popPK model to characterize the CAR-T cell PK as well as a fully integrated popPK-TS dynamic model makes it a useful tool for MIDD of CAR-T cells and provides a framework for credentialing the impact of patient and product parameters on efficacy and safety. TS response metrics derived from the popPK-TS dynamic model will be explored as potential early predictors for treatment outcome.
References:
[1] Stein AM et al. CPT:PSP (2019) 8, 285-95.
[2] Wu LS et al. Clin Transl Sci (2022) 15, 3000-11.
[3] Ogasawara K et al. Clin Pharmacokinet (2023) 60, 1621-33.
[4] Chartier M et al. Clin Pharmacokinet (2024) 53, 1283-99.
[5] Hu H et al. Clin Cancer Res (2025) 31, 1163-71.
[6] Wu F et al. Clin Pharmacokinet (2025) 64, 119-32.
[7] Mc Laughlin AM et al. CPT:PSP (2023) 12, 1577-90.
[8] Chan P et al. PAGE 27 (2018) Abstr 8691 [www.page-meeting.org/?abstract=8691]
[9] Svensson RJ and Jonsson NE CPT:PSP (2022) 11, 1210-1222.
[10] Chatterjee M et al. CPT:PSP (2017) 6, 29-39.
[11] Netterberg I et al. Clin Cancer Res (2020) 26, 4892-900.
[12] Bender BC et al. PAGE 26 (2017) Abstr 7344 [www.page-meeting.org/?abstract=7344]
[13] Simeoni M et al. Cancer Res (2004) 64, 1094-101.
Reference: PAGE 34 (2026) Abstr 12225 [www.page-meeting.org/?abstract=12225]
Poster: Drug/Disease Modelling - Oncology