Development of efficient CAR-T cell clinical study designs: Towards understanding of novel anticancer therapies through optimal experimental design
Fenja Klima1,2, Anna M. Mc Laughlin3, Robin Michelet1, Anika Winkler4, Wilhelm Huisinga2,5, Katy Haussmann6, Annette Künkele4, Andrew C. Hooker7, Charlotte Kloft1,2
1Department of Clinical Pharmacy and Biochemistry, Institute of Pharmacy, Freie Universitaet Berlin, 2Graduate Research Training Program PharMetrX, 3Pharmetheus AB, 4Charité – Universitätsmedizin Berlin, corporate Member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health, Department of Pediatric Oncology and Haematology, 5Institute of Mathematics, University of Potsdam, 6Charité – Universitätsmedizin Berlin, corporate Member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health, Stem Cell Facility, 7Department of Pharmacy, Uppsala University
Introduction: Chimeric antigen receptor-T cell (CAR-T) therapy is a novel cellular immunotherapy and has achieved impressive treatment outcomes for many haematological cancer patients, yet up to 60% of patients fail to enter long-term remission or eventually relapse [1]. While required for a deeper understanding of CAR-T kinetics, dynamics and early predictors of survival, the collection of informative clinical CAR-T data remains challenging: The choice of sample numbers and times is complex due to unique cellular kinetics, high interindividual variability, and resource-intensive bioanalysis. Furthermore, patient numbers are limited due to rare indications, high costs, and only recent introduction to clinical practice. To fill knowledge gaps and improve CAR-T therapy, the design of informative yet feasible clinical CAR-T data collection across different patient populations is thus essential. Through optimal experimental design (OED), this work aimed to inform robust designs for clinical data collection with regards to population size, sampling strategies and optimal CAR-T sampling windows. Methods: An OED framework was developed based on a mechanistic model of CAR-T kinetics and interaction with metabolic tumour volume (MTV) [2], using the R packages rxode2 and PopED [3]. To evaluate population size and sampling strategies, design variables were varied compared to a reference scenario commonly found in clinical CAR-T datasets [2,4-6] of 60 patients, CAR-T sampling at days 7, 14, 28 and measurement of MTV at days 0, 30, and 90: (1) Population size was varied between 10-150 patients in steps of 10; (2) the last CAR-T and/or MTV sampling time was removed or shifted between days 18-90 (CAR-T) or days 30-180 (MTV); (3) an additional early CAR-T sampling time during CAR-T expansion was explored between days 0-13. For each design scenario, 100 population replicates with simulated patient characteristics were evaluated. Patient characteristics were simulated from a truncated lognormal (initial tumour burden) or a binomial distribution (CAR-T reference or low expansion subpopulation) based on published phase 3 data of a similar patient population [4]. Subsequently, optimal CAR-T sampling windows were determined using design optimisations and applying insights on population size and sampling strategy: D-optimal design was used as design criteria and sampling windows were defined as 90% confidence intervals (CI) of optimal sampling times obtained across multiple populations or parameter vectors: To quantify the impact of different covariate distributions across populations, CAR-T sampling times were optimised for 100 population replicates with simulated patient characteristics. To quantify the impact of model parameter uncertainty, CAR-T sampling times were optimised using 100 parameter vectors obtained by Sampling Importance Resampling [7]. The addition of a small additive residual variability (RUV) to the original proportional RUV was investigated for assessments of CAR-T/MTV sample number or time to balance optimal sampling times and bioanalytical constraints, e.g. lower limit of quantification [8]. Model parameter bias and precision were assessed for key intermediate and final design scenarios using Stochastic Simulation and Estimation (SSE) with 500 samples. Results: Evaluating population size and sampling strategies revealed differences between parameters related to CAR-T kinetics, tumour-induced CAR-T expansion and CAR-T-induced tumour killing: (1) A population size of 60 patients was sufficient to inform all model parameters, i.e., expected relative standard errors (eRSE) =50%, while as few as 20 patients could inform parameters related to CAR-T kinetics only. (2) Especially for parameters related to tumour killing, eRSE increased upon removal of the last CAR-T and/or MTV sampling times. An SSE confirmed increased parameter imprecision and bias for these scenarios, highlighting the need for late sampling: The root mean square error increased up to 6-fold and bias up to 4-fold upon removal of the last CAR-T and MTV sampling times compared to the reference scenario. Shifting the last CAR-T sampling time backwards resulted in decreasing eRSE over time, favouring always the latest tested timepoint, even when clinically implausible or bioanalytically not feasible. This limitation was overcome by including a small additive RUV to identify plausible last CAR-T sampling times, indicated by a minimal eRSE between days 18-42 for all parameters. A sensitivity analysis across a 200-fold range of additive RUV values yielded robust results, i.e., very similar timepoints of minimal eRSE, supporting the use of this approach to account for bioanalytical constraints. Shifting the last MTV sampling time did not impact eRSE, suggesting a broad range of informative timepoints. (3) An additional early CAR-T sample did not lead to substantially reduced eRSE and was only beneficial at days 2-5, i.e., prior to the first reference sample. Based on these insights, the reference scenario was adapted to optimise three CAR-T sampling timepoints in a wide design space of days 1-14, 7-28 and 14-60, while MTV timepoints had shown little need for optimisation and were kept fixed. Optimal CAR-T timepoints [90% CI] across populations with different covariate distributions were identified at days 3 [3-4], 13 [12-15] and 27 [27-28], showing highly robust results independent of covariate distributions. Optimal sampling times obtained from different parameter vectors were days 4 [3-6], 14 [13-21] and 28 [23-32]. While the median optimal sampling times were very similar, the wider 90% CI highlighted the importance of accounting for uncertainty about true parameter values when informing clinical CAR-T sampling windows. An SSE yielded acceptable parameter precision and bias for a design of 20-60 patients, three CAR-T time windows and three MTV timepoints. The proposed design will be implemented for clinical CAR-T data collection in paediatric patients with haematological cancers at Charité – Universitätsmedizin Berlin. Conclusion: To inform a complex mechanistic CAR-T model with minimal data, a comprehensive OED framework was developed, and clinically feasible and flexible CAR-T clinical data designs were identified. Considering different covariate distributions and parameter uncertainty yielded robust designs with optimal sampling windows rather than timepoints, expected to cover heterogenic cancer patient populations and minimise bias in parameter estimates. OED enables efficient data collection, of special importance in settings with limited patient numbers and resources, and close collaboration with clinical partners ensures successful adoption to clinical trials and clinical practice. Ultimately, OED paves the way for timely and robust data analysis, thus contributing to leveraging the full potential of cancer immunotherapies and advance their development in clinical studies.