Integrative population pharmacokinetic modeling of two BCMA-targeted and one CD19-targeted CAR-T therapies using full Bayesian inference with a student’s t-based M3 censoring approach for robust handling of outliers and BLQ data

Abstract

Background

Chimeric antigen receptor (CAR) T-cell therapies represent a novel class of adoptive immunotherapies characterized by dynamic in vivo expansion, contraction, and long-term persistence. These unique kinetic properties distinguish CAR-T therapies from traditional small-molecule and biologic agents and pose significant challenges for pharmacometric modeling, particularly due to high interindividual variability, frequent outliers, and a large proportion of concentrations below the lower limit of quantification (BLQ).

Objective

In this study, we developed and applied an integrated Bayesian population pharmacokinetic (PopPK) modeling framework to evaluate and compare the cellular kinetics of three CAR-T therapies: liso-cel (CD19-targeted), and ide-cel and orva-cel (both BCMA-targeted).

Method

Four modeling strategies were examined, combining two residual error structures (normal vs. Student's t) with two censoring methods (M1 vs. M3). Full Bayesian inference was implemented using Markov Chain Monte Carlo (MCMC) methods in NONMEM®.

Results and Conclusion

Our results demonstrate that M3 censoring is essential for preserving information from BLQ data and that the Student’s t-distribution offers superior robustness to outlier contamination compared to conventional normal error models. The combined use of M3 censoring and Student’s t residuals yielded the most accurate and stable parameter estimates across all phases of the CAR-T kinetic profiles. Additionally, comparative modeling under a unified Bayesian platform revealed both shared and target-specific kinetic features among the three therapies, with BCMA-targeted products demonstrating faster expansion kinetics than the CD19-targeted product. Our findings support routine use of the Student’s t–M3 configuration under a Bayesian framework for CAR-T PopPK modeling.