nlmixr: an open-source package for pharmacometric modelling in R
Rik Schoemaker (1), Matt Fidler (2), Yuan Xiong (3), Justin Wilkins (1), Mirjam Trame, Christian Laveille (4), Wenping Wang (2)
(1) Occams, The Netherlands, (2) Novartis Pharmaceuticals, USA, (3) Certara Strategic Consulting, USA, (4) Calvagone, France
Introduction: nlmixr is an open-source R package under development and freely available on GitHub[1]. It builds on RxODE[2], an R package for simulation of nonlinear mixed effect models using ordinary differential equations (ODEs), providing an efficient and versatile way to specify pharmacometric models and dosing scenarios, with rapid execution due to compilation in C. By combining the RxODE core with population-type estimation routines, a versatile pharmacometric ecosystem entirely contained within R becomes feasible. Currently, estimation routines comprise the nlme[3] package in R, a Stochastic Approximation Expectation Maximization (SAEM) algorithm [4], and a proof-of-concept FOCE-I implementation [5], as well as adaptive Gaussian quadrature for odd-type data. Both closed-form and ODE model definitions are included in nlmixr.
Methods: Richly sampled profiles were simulated for 4 different dose levels (10, 30, 60 and 120 mg) of 30 subjects each as single dose (over 72h), multiple dose (4 daily doses), single and multiple dose combined, and steady state dosing, for a range of test models: 1- and 2-compartment disposition, with and without 1st order absorption, with either linear or Michaelis-Menten (MM) clearance (MM without steady state dosing). This provided a total of 42 test cases. All inter-individual variabilities (IIVs) were set at 30%, residual error at 20% and overlapping PK parameters were the same for all models. A similar set of models was previously used to compare NONMEM and Monolix[6]. Additionally, a sparse data estimation situation was investigated where 500 datasets of 600 subjects each (150 per dose) were generated consisting of 4 random time point samples in 24 hours per subject, using a first-order absorption, 1-compartment disposition, linear elimination model. NONMEM®[5] with FOCE-I was used as a comparator to test the various nlmixr estimation routines.
Results: Theta parameter estimates were comparable across estimation methods. In comparison to NONMEM, nlmixr using nlme was always faster for ODEs (MM-models) and comparable for closed form models, but IIV estimates were regularly estimated close to 0% in nlmixr, whereas NONMEM provided estimates closer to the original simulation values. In contrast, both the SAEM and the FOCE-I estimation routines provided good IIV estimates at higher computational cost.
Conclusion: These findings provide further evidence that nlmixr may provide a viable open-source parameter estimation alternative for fitting nonlinear mixed effects pharmacometric models within the R environment.
References:
[1] https://github.com/nlmixrdevelopment/nlmixr
[2] Wang W et al. CPT:PSP (2016) 5, 3–10.
[3] Pinheiro J et al. (2016). nlme: Linear and Nonlinear Mixed Effects Models. R package version 3.1-126.
[4] Kuhn E and Lavielle M. Comput Stat Data An, 49:1020–1038, 2005.
[5] Beal SL et al. 1989-2011. NONMEM Users Guides. Icon Development Solutions, Ellicott City, Maryland, USA.
[6] Laveille C et al. PAGE 17 (2008) Abstr 1356 [www.page-meeting.org/?abstract=1356]
