Juan Gonzalez Maffe, Sebastian Weber, Guillaume Baneyx, Lukas Andreas Widmer, Lada Markovtsova
Novartis Pharma AG
Introduction
The design of phase I trials in oncology follows mainly a dose-escalation (DE) design monitoring dose limiting toxicity (DLT) in the 1st treatment cycle to formally decide on the dose during the trial. Other data like pharmacokinetic (PK), pharmacodynamic (PD) and dose reductions/interruptions are considered less formally due to its sparsity/incompleteness at early stages. With toxicities occurring at later stages beyond cycle 1, the focus on cycle 1 is becoming a limitation in modern multi-cycle therapies. We introduce a Bayesian multi-cycle time to event (TTE) model which is formulated as an extension to existing approaches like the Bayesian Logistic Regression Model (BLRM) and relies on readily available actual dose history and possibly PK summaries. This work shows the application of a novel integrated population PD framework (“popPD”) to the analysis of safety events beyond cycle 1 by re-analyzing a past dose-escalation trial [1].
Objective
To support DE decisions on dose/regimen selection during the DE, we apply an integrated popPD framework linking a simplified PK model to a TTE model and having the following features:
- Uses the actual dose history to form a longitudinal exposure measure aligned with the known drug pharmacology
- Avoids overconfident predictions by accounting for uncertainty in the longitudinal exposure measure
- Uses partially observed patient data (eg, most recent cohort)
- Predicts outcomes for future patients at selected dose/regimens and inform DE decisions
Methods
For the popPD model, we first setup a simplified PK model which focuses on steady state kinetics. This includes reaching, maintaining, and leaving steady state. We hence use a linear one-compartment model with first order absorption, which is informed from available data during the trial, i.e. the actual dose history and non-compartmental analysis estimates of clearance and volume as well as observed Cmin and Cmax. The simplified PK model is linked to the time to first adverse event (AE) process by modeling the log of the hazard as a linear function of the log of the exposure resulting in a popPD model. This popPD model is called PK-TTE model in the rest of the abstract. We also consider a K-TTE model variant which only uses the dose history.
We use the data from trial NCT02375958 for application of this framework. The trial tested the experimental drug PCA062 given as infusion (in mg/kg) to patients with pCAD+ tumors [2, 3]. As the DE was completed, the DE data was emulated using the cut-off date for each DE meeting, with the PK data of the last two patients enrolled in a cohort set to missing. We focus here on all Gr3+ AEs. To align with the priors used for the BLRM that guided the DE as per protocol, we assume that the Gr3+ AEs rate at the end of cycle 1 is 25% for 3.0 mg/kg dose. All models are fitted in Stan [4] using the brms R package [5].
Results
The (P)K-TTE model for PCA062 was implemented from DE meeting 4 (n=15, 4 cohorts) onwards. The PK model has an adequate fit with good prediction for clearance/volume in patients with missing PK data. As the PK-TTE model includes the dose history and PK data, it is considered more reliable as it accounts for the actual patient exposure to drug rather than using the administered dose as a regressor. In comparison to the BLRM, (P)K-TTE models are preferrable since data from all treatment cycles are included. This results in a more efficient use of the data as some AEs (4 out of 11 events) occurred past cycle 1.
Conclusion
This work shows that a Bayesian (P)K-TTE model is a useful tool to incorporate exposure data in early DE decisions. These TTE models have the ability to incorporate event data from multiple cycles allowing a more efficient use of the trial data and to predict AE rates for the entire multi-cycle therapy rather than just the AE rate in cycle 1 as in the BLRM approach. Moreover, we show how the integrated popPD framework can leverage the common issue of incomplete PK data in DE trials. For patients with missing PK data, a model based imputation of the exposure is used in place of the actual data. The approach accounts for all relevant uncertainties to make predictions of toxicity rates. We consider the K-TTE model with the planned dose regimen as a replacement candidate for the BLRM while the (P)K-TTE models using actual dose history and available exposure data provide additional information for decision making on dose/regimen during DE.
[1] S Weber, J Gonzalez Maffe, G Baneyx, L Widmer, L Markovtsova. Novel integrated population PD modeling framework to inform decision making during Oncology phase I dose-escalation. To be submitted to PAGE 32 (2024) (OAK ID: 53342)
[2] Duca M, Lim DW, Subbiah V, Takahashi S, Sarantopoulos J, Varga A, D’Alessio JA, Abrams T, Sheng Q, Tan EY, Rosa MS, Gonzalez-Maffe J, Sand-Dejmek J, Fabre C, Martin M. A First-in-Human, Phase I, Multicenter, Open-Label, Dose-Escalation Study of PCA062: An Antibody-Drug Conjugate Targeting P-Cadherin, in Patients With Solid Tumors. Mol Cancer Ther. 2022 Apr 1;21(4):625-634. doi: 10.1158/1535-7163.MCT-21-0652. PMID: 35131875.
[3] J Kim, J Gonzalez Maffe, E Tan, J Sand Dejmek, C Fabre, C Meille. Combining BLRM and safety PKPD models to improve decision making in a phase I dose escalation study: case study of PCA062, an antibody drug conjugate targeting P-Cadherin. PAGE 28 (2019) Abstr 9108 [www.page-meeting.org/?abstract=9108]
[4] Stan Development Team. 2023. Stan Modeling Language Users Guide and Reference Manual, 2.32.2. https://mc-stan.org
[5] Bürkner P. C. (2017). brms: An R Package for Bayesian Multilevel Models using Stan. Journal of Statistical Software. 80(1), 1-28. doi.org/10.18637/jss.v080.i01
Reference: PAGE 32 (2024) Abstr 10822 [www.page-meeting.org/?abstract=10822]
Poster: Clinical Applications