PopED: An extended, parallelized, nonlinear mixed effects models optimal design tool
Introduction
The idea of designing scientific experiments in an optimal way dates back to the beginning of the 20th century [1]. R.A. Fisher first demonstrated that the information contained in experimental data is limited and dependent on the experimental setup [1]. A formal mathematical inequality was established by Cramer and Rao (independently from each other) stating that the inverse of the Fisher information matrix (FIM) is a lower bound for the covariance of any unbiased estimator [2], [3]. Therefore, the Cramer–Rao bound provides the essential basis for the optimized planning of an experiment, suggesting that the experimental design variables should be chosen such that the FIM is maximal. For linear and nonlinear regression models without mixed effects parameters the FIM can be analytically calculated as demonstrated in the work by Atkinson and Donev [4]. In contrast to that, the calculation of the FIM for nonlinear mixed effect models (NLMEM) has to rely on approximations since analytic solutions do not exist. A closed form solution for the first-order approximated FIM for NLMEM was initially derived by Mentré et al. [5]. Later on this approach was extended and implemented in a computer program by Retout et al. [6]. Today, several software packages for the optimization of population experimental designs (designs for nonlinear mixed effects models) are available and have been compared by Mentré et al. [7]. The programs differ in their implementations of the FIM calculation, the programming language and in the set of available features. In this paper we discuss new features of the program PopED (Population Experimental Design) initially developed by Foracchia et al. [8].
This paper consists of three parts. First, the theoretical basis of population optimal design is presented in Section 2. In the second part the implementation of PopED is described in detail. This includes (i) a listing of the different FIM approximations available (Section 3.1 and in more detail in the appendix), (ii) a description of the numerical methods used for the calculation of the FIM and the optimization of the designs (Sections 3.2 Design criteria, 3.3 Design optimization, 3.4 Calculation of derivatives, 3.5 Solution of differential equations) and (iii) a presentation of the software architecture used for the graphical user interface (GUI) and the calculation engine (Sections 4.1 Graphical user interface, 4.2 Calculation engine, 4.3 Parallelization of PopED). Section 4 also includes a description of how computations can be parallelized in PopED. The third part of the article presents four examples selected to demonstrate the strengths of the software (Sections 5.1 Example 1: Group size optimization, 5.1.1 Results, 5.1.2 Conclusions, 5.2 Example 2: Constraint/cost optimization, 5.2.1 Results, 5.2.2 Conclusions, 5.3 Example 3: Optimization with different FIM approximations, 5.3.1 Results, 5.3.2 Conclusions, 5.4 Example 4: Parallel execution). Even though all examples consider pharmacometric applications of optimal design it should be pointed out that the application of the software is not restricted to this area as demonstrated for example in the work by Sjögren et al. [9].
Section snippets
Optimal design
The fundamental role of the Cramer–Rao inequality for the theory of optimal design has been mentioned above. Formally, the relationship between the inverse of the FIM and the variance-covariance matrix (COV) of any unbiased estimator, as established by the Cramer–Rao inequality, can be formulated aswhere y is the observed data, q are the design specific variables and Θ the model parameters. It is clear from this equation that a maximal FIM (or minimal inverse of the FIM) is
Fisher information matrix
In PopED the user can choose among an extensive collection of different FIM approximation methods. In this section, the different approximations are listed and discussed briefly. A more mathematical description is given in the appendix.
First-order approximation: An approximation of the FIM for NLMEMs was first proposed by Mentré et al. [5] and further developed and improved by Retout et al. [6], [10]. The first-order (FO) approximation available in PopED is based on the approach described by
Graphical user interface
The graphical user interface (GUI) (Fig. 1) is written in C# .NET Framework 2.0 and communicates with the calculation engine, written in MATLAB, via xml files (Fig. 2). The GUI enables usage of PopED without extensive knowledge of MATLAB or programming in general. However, the GUI only presents design- and optimization settings with a more user-friendly interface and hence all settings enabled via the GUI can also be entered directly in the PopED MATLAB configuration file (function_input.m). A
Example 1: Group size optimization
This example illustrates the application of the group size optimization, the differences between the GS and FA algorithm and the efficiency translation feature. The pharmacometric model for this example was a simple dose–response model, described by the following equation:where Dg is the vector of doses given to group g, ɛ is the residual variability random variable , θ1,2,3 are the fixed and η1,2,3 the associated random effect variables .
Discussion
This paper shows several new features in optimal design for NLMEM all implemented in the optimal design tool PopED. The tool is open source and freely accessible online (http://poped.sf.net). Most of the features in PopED can be used by the free software FreeMat [30] but some additional functionality are limited to MATLAB [24] users (see Table 2 for details).
Several other tools in NLMEM optimal design are available, written in MATLAB or R and with graphical user interfaces [7]. However, the
Conflict of interest
The authors declare no conflict of interest.
Acknowledgements
The parallel computations were performed on resources provide by SNIC through Uppsala Multidisciplinary Center for Advanced Computational Science (UPPMAX) under Project p2010057.
References (47)
- et al.
Development and implementation of the population Fisher information matrix for the evaluation of population pharmacokinetic designs
Computer Methods and Programs in Biomedicine
(2001) - et al.
POPED, a software for optimal experiment design in population kinetics
Computer Methods and Programs in Biomedicine
(2004) - et al.
Implementation of OSPOP, an algorithm for the estimation of optimal sampling times in pharmacokinetics by the ED, EID and API criteria
Computer Methods and Programs in Biomedicine
(1996) - et al.
Robust experiment design via stochastic approximation
Mathematical Biosciences
(1985) - et al.
The use of a modified Fedorov exchange algorithm to optimise sampling times for population pharmacokinetic experiments
Computer Methods and Programs in Biomedicine
(2005) - et al.
A program for individual and population optimal design for univariate and multivariate response pharmacokinetic–pharmacodynamic models
Computer Methods and Programs in Biomedicine
(2007) R. A. Fisher and the design of experiments, 1922–1926
The American Statistician
(1980)Methods of estimation
Mathematical Methods of Statistics
(1946)Information and accuracy attainable in the estimation of statistical parameters
Bulletin of the Calcutta Mathematical Society
(1945)- et al.
Optimum Experimental Designs
(1992)
Optimal design in random-effects regression models
Biometrika
Optimal experimental design for assessment of enzyme kinetics in a drug discovery screening environment
Drug Metabolism and Disposition
Further developments of the Fisher information matrix in nonlinear mixed effects models with evaluation in population pharmacokinetics
Journal of Biopharmaceutical Statistics
Some considerations on the Fisher information in nonlinear mixed effects models
mODa 9 – Advances in Model-Oriented Design and Analysis
Bayesian optimal designs for pharmacokinetic models: sensitivity to uncertainty
Journal of Biopharmaceutical Statistics
Robust population pharmacokinetic experiment design
Journal of Pharmacokinetics and Pharmacodynamics
A comparison of three methods for selecting values of input variables in the analysis of output from a computer code
Technometrics
Simultaneous optimal experimental design on dose and sample times
Journal of Pharmacokinetics and Pharmacodynamics
Model-Oriented Design of Experiments
The convergence of a class of double-rank minimization algorithms 1. General considerations
IMA Journal of Applied Mathematics
Cited by (63)
Optimizing study design in LPS challenge studies for quantifying drug induced inhibition of TNFα response: Did we miss the prime time?
2022, European Journal of Pharmaceutical SciencesUse of prior knowledge and extrapolation in paediatric drug development: A case study with deferasirox
2019, European Journal of Pharmaceutical SciencesTowards a consistent understanding of the metal hydride reaction kinetics: Measurement, modeling and data processing
2018, Journal of Alloys and CompoundsLeast squares polynomial chaos expansion: A review of sampling strategies
2018, Computer Methods in Applied Mechanics and EngineeringCitation Excerpt :In addition to the greedy algorithms discussed here, other techniques such as simulated annealing (SA) [80,81], genetic algorithms (GA) [82,83], semi-definite programming (SDP) [84,85] and pattern search [59,86] have been employed in the literature to construct alphabetic optimal designs. Moreover, software packages such as JMP [87], POBE [88], AlgDesign [89], PopED [90,91] and Gosset [92] have been developed to construct ODE using the algorithms discussed here. We note that the majority of the software packages for the construction of ODE are developed for scenarios where low-order (mostly quadratic) polynomials are employed, hence, limiting the applicability of such packages for high-order PCE’s.
Computational aspects of experimental designs in multiple-group mixed models
2024, Statistical Papers
- 1
Authors contributed equally to this work.
- 2
Present address: School of Pharmacy, University of Queensland, Brisbane, Australia.