2022 - Ljubljana - Slovenia

PAGE 2022: Methodology - Estimation Methods
Nivea M. F. Voelkner

Implementation of Estimation Step Refinements to Support Runtime Reduction for Model Building in NONMEM® 7

Nivea M.F. Voelkner (1), Sonya C. Chapman (2), Aurelie Lombard (2), Jeroen Elassaiss-Schaap (1)

(1) PD-value B.V., Yalelaan 1, 3584 CL Utrecht, Netherlands, (2) Eli Lilly and Company, Arlington Square, Bracknell, UK

Introduction: Overall, as pharmacometrics advances as a powerful computational tool, model complexity describing biological systems has also increased. Computer runtime of a specific model is dependent on multiple factors such as model complexity, dataset size, convergence algorithm and the chosen software. Some effective approaches to decrease the runtime duration such as selecting the right estimation method and reducing the dataset may enhance utility and model efficiency. As pharmacometrics becomes an increasingly important factor in drug development and approvals, improving model runtimes can have a significant impact on timely decision making and regulatory submission timelines.

Objectives:

  • The aim of this study is to reduce runtime while maintaining model fit and predictive performance in NONMEM® 7.

Methods: A mechanistic pharmacokinetic (PK) model for abemaciclib previously published by Chigutsa et al. [1] was modified by dataset and model reduction. Estimation methods used in the reduced model were stochastic approximation expectation maximization (SAEM) and Monte Carlo importance sampling expectation maximization (IMP). The runtime of this reduced model was 84h (3.5 days) [data on file, Eli Lilly and Company]. To optimize this reduced model, the subroutine was adjusted for ADVAN7 or ADVAN5 instead ADVAN13. The subroutine update allowed the differential equations ($DES in NONMEM®) to be removed and instead substituted with rate constants for NONMEM® to self-build the equations. Also, the dataset was adjusted removing an intravenous (IV) group record of 11 subjects since those were continuously indicating ID output errors. MCETA was used as an estimation option to assess a different set of initial estimates (rather than the reduced model parameters as inputs) for the individual mode a posteriori (MAP) estimation. Additionally, Owen and Faure-Tezuka scrambling techniques were used to de-correlate the components of the quasi-random parameter samples, without affecting the basic low discrepancy property that improves the accuracy. The final model was selected based upon precision estimates, diagnostic plots, and visual predictive checks (VPCs). The ability to replicate the predicted abemaciclib exposure estimates at steady state was evaluated.

Results: Compared to the reduced model, which had a runtime of 84 h (~ 3.5 days) using SAEM+IMP methods and the ADVAN13 subroutine, the run time dropped to 17h26min using the ADVAN5 subroutine. The ADVAN7 subroutine was also evaluated, but NONMEM® displayed output errors during the iteration process. MCETA (1, 2 and 10) estimation options were then explored using the ADVAN5 subroutine model. MCETA = 2 model was more efficient and had a runtime decreased to 16h59min while MCETA = 1 or 10 had a longer runtime. This model was further improved with the inclusion of the “Sobol” method (RANMETHOD = [1|S|2]) (1 = Owen type scrambling method; 2= Feure-Tezuka method) as an estimation option decreasing the runtime to 11h44min for the final model. Exposures at steady state for the reduced model versus the final model were comparable to the MONARCH 2 study [1] predictions.

Conclusion: The mechanistic PK model for abemaciclib was optimized using estimation options, making it more easily applied to new clinical trial data in a timely manner. The optimized final model with shorter runtime (11h 44min; 0.5 days) maintained the model fit with satisfactory OFV compared with the reduced model (84h; 3.5 days). Therefore, SAEM+IMP estimation methods in combination with ADVAN5 subroutine, MCETA = 2 and scrambling method used as estimation options can be a useful alternative to ADVAN13 for complex PK models with multiple compartments.



References:
[1] Chigutsa, E. et al. CPT:PSP 2020; 9 (9): 523-533
[2] Elassaiss-Schaap, J. & Duisters, K. CPT:PSP 2020; 9 (5) 245-257
[3] Hamrén, B. et al. PAGE 2006; poster
[4] Karlsson, M.O. et al. PAGE 2008; Abstract 1361


Reference: PAGE 30 (2022) Abstr 9961 [www.page-meeting.org/?abstract=9961]
Poster: Methodology - Estimation Methods
Top