Elba Raimúndez (1), Rüdiger Siek (1), Andreas Kovar (1), and Ahmed A. Suleiman (1)
(1) Sanofi, Germany
Introduction: Automated model development workflows to identify the optimal population pharmacokinetic (PK) model – as implemented in Pharmpy open-source software package [1,2] for example – have recently gained popularity. Finding the optimal model structure is also a recurrent challenge in systems biology where often the considered models are highly complex. FAMoS (Flexible Algorithm for Model Selection) is an open-source tool for performing automated model selection aimed at analysing complex dynamical systems [3,4]. FAMoS can search in large model spaces in a flexible, efficient, and dynamical way by performing the classical forward and backward search algorithms. Additionally, it incorporates a swap search algorithm, a novel non-local method, that can prevent the model search from ending up in local minima. Considering the scope and capabilities of FAMoS, it can potentially boost workflows implemented in Pharmpy. In this work, we propose and evaluate the combination of FAMoS with Pharmpy on a collection of models.
Objectives: Evaluate the performance of the hybrid FAMoS-Pharmpy approach with synthetic datasets and compare the results with using Pharmpy alone.
Methods: Synthetic simulated datasets with richly and sparse sampled profiles were generated using RxODE [5]. Richly sampled profiles were simulated for 4 different dose levels (10, 30, 60 and 120 mg) of 20 subjects as single dose, and multiple dose (with 5 doses daily administered). Sparse sampled profiles were simulated for a multiple single dose level (120 mg with 10 doses daily) of 200 subjects consisting of 4 random time point samples during the first and last dose interval per subject. To do this, models with a 2-compartment disposition were employed in combination with (1) linear clearance or combined with target-mediated drug disposition (TMDD), and (2) with first-order absorption or intravenous administration. Inter-individual variability (IIV) was included in clearance and set to 30%, and the proportional residual error (RUV) set to 10%. The same PK parameter values were used for all models. Parameter estimation was performed using NONMEM [6] via the Pharmpy package. In the hybrid approach Pharmpy was employed for the compartmental-structure search step and Famos for the IIV and RUV search step.
Results: Pharmpy could successfully find the true 2-compartment disposition model for all models except one. For the first-order absorption synthetic datasets, only an instantaneous absorption was identified. For the TMDD datasets, all final models contained a Michaelis-Menten elimination component (alone or combined with first-order clearance) indicating that the TMDD component was correctly captured. Overall, for the first search step, we saw a very good agreement between the inferred models by Pharmpy with the true model structures. For the subsequent IIV and RUV search steps, we found that all models using both approaches converged to the true IIV and RUV models. Although for IIV, when using Pharmpy alone we saw that for a few models the optimal IIV structure contained one additional IIV parameter which was estimated to be very close to zero (1e-06). Moreover, the final model structures found by both approaches converged to the same objective function values and retrieved the same individual predictions. In terms of performance, we observed an overall consistent reduction in CPU time by using our hybrid FAMoS-Pharmpy approach compared to Pharmpy alone. The largest difference was found in the IIV search, where our hybrid approach was in average 50.97 (6.72 – 104.63) times faster than using Pharmpy alone. An improvement was also reported for the RUV search being in average 6.26 (3.99 – 7.72) times faster. In the latter, we attribute the lower performance improvement to the reduced model search space for the RUV models, where generally only 3 combinations are tested: additive, proportional, and combined error models.
Conclusions: The integration of optimization methods used in the systems biology field has the potential to boost the pharmacometrics workflows. The findings of this evaluation suggest that FAMoS provides an efficient model selection tool which can enhance the current automated model search functionality of the Pharmpy workflow by integrating both tools. While the current results suggest considerably more efficient workflows with similar modeling conclusions, further testing and extensions, such as covariate model development, must be explored.
References
[1] Nordgren, R et al. (2023) Pharmpy: a versatile open-source library for pharmacometrics. PAGE2023 Abstract 10508 [https://www.page-meeting.org/default.asp?abstract=10508]
[2] https://github.com/pharmpy/pharmpy
[3] Gabel, M et al. (2019) FAMoS: A Flexible and dynamic Algorithm for Model Selection to analyse complex systems dynamics. PLoS computational biology, 15(8), e1007230.
[4] https://github.com/GabelHub/FAMoS_Py
[5] Wang, W et al. (2016) A Tutorial on RxODE: Simulating Differential Equation Pharmacometric Models in R. CPT: pharmacometrics & systems pharmacology, 5(1), 3–10.
[6] Bauer, RJ (2019) NONMEM Tutorial Part I: Description of Commands and Options, With Simple Examples of Population Analysis. CPT: pharmacometrics & systems pharmacology, 8(8), 525–537.
Reference: PAGE 32 (2024) Abstr 10857 [www.page-meeting.org/?abstract=10857]
Poster: Methodology - New Modelling Approaches