Tom Haber (1), Tom Jacobs (2), Xavier Woot de Trixhe (2), Frank van Reeth (1)
Universiteit Hasselt, (1) Janssen Research and Development (2)
Introduction: DiffMEM is an open-source software package, freely available on Bitbucket[1], for rapid optimization of non-linear mixed effect models using ordinary and partial differential equations (ODE/PDE). During parameter estimation, these ODE/PDE’s are frequently integrated numerically since parameters change between individuals and across iterations. This can be computationally expensive, resulting in long runtimes for model fitting. Consequently, this can seriously limit building, testing and using such models. However, through heavy use of parallelism, extensive low-level optimizations and algorithmic innovation [3], DiffMEM is able to rapidly fit such complex models. Duty-cycles are shortened while building a model, adaptive/optimal trial design is enabled, and model validation becomes possible, since these typically require a lot of simulation-estimation steps. DiffMEM is under active development with interfaces to R and Python, and an exciting innovation is the concept of model fitting as a web-service which completely decouples the software/hardware configuration from the fitting process.
Objectives:
- Examine estimation properties of the SAEM routines implemented in DiffMEM and compare with several other methods and implementations based on the methodology from Plan et al. [2]
- Compare the performance between DiffMEM and NONMEM in several real-life use-cases
Methods: Plan et al. [2] considered a sigmoid Emax model with one hundred individuals with observations at four doses (rich design) and two doses (sparse design) to study nine approaches for ML estimation. One hundred simulated datasets were generated for each of eight scenarios. The SAEM estimation routine in DiffMEM is tested on the same datasets to compare its results to the other nine approaches.
DiffMEM was used for the parameter estimation on the ODE based PK/PD models from real-life use-cases by Dunne et al. [4] and de Winter et al. [5]. The estimation was performed on the original models without alterations. The former model consists of two-compartment PK model and turn-over PD model. Dunne et al. [4] used a sequential estimation method: PK parameters are estimated in a first step and are subsequently fixed while fitting PD parameters in the second step. In our comparison, we will only focus on this second step. The latter is a population PK model based on a turn-over model for HbA1c consisting of 1347 type-2 diabetic patients. In both settings, we compare performance of DiffMEM with parallel NONMEM.
Results: The SAEM routine in DiffMEM consistently produces comparable results to other SAEM approaches investigated by Plan et al. [2]. With the real-life use-cases, DiffMEM is able to produce similar results, but runtime is much lower: model fitting using parallel NONMEM with 24 cores takes respectively 3 hours and 8 hours whereas using DiffMEM the runtime is reduced to 10 and 20 minutes.
Conclusions: The study suggests that DiffMEM is a viable open-source alternative for parameter estimation of non-linear mixed effect models while being an order of magnitude faster in delivering results. The open-source nature allows for relatively easy extension to novel types of models.
References:
[1] https://bitbucket.org/tomhaber/diffmem
[2] Performance Comparison of Various Maximum Likelihood Nonlinear Mixed-Effects Estimation Methods for Dose–Response Models, Elodie L. Plan, Alan Maloney, France Mentré, Mats O. Karlsson and Julie Bertrand, 2012
[3] Fast derivatives of likelihood functionals for ODE based models using adjoint-state method, Valdemar Melicher, Tom Haber and Wim Vanroose, 2017
[4] The method of averaging applied to pharmacokinetic/pharmacodynamic indirect response models, Adrian Dunne, Willem de Winter, Chyi-Hung Hsu, Shiferaw Mariam, Martine Neyens, Jose Pinheiro and Xavier Woot de Trixhe, 2015
[5] A dynamic population PK/PD model to assess the effect of once daily versus twice daily dosing regimens on the relationship between canagliflozin plasma exposure and HbA1c response, Willem de Winter, Adrian Dunne, Chyi-Hung Hsu, Xavier Woot de Trixhe, Damayanthi Devineni, David Polidori and Jose Pinheiro, 2015
Reference: PAGE 27 (2018) Abstr 8625 [www.page-meeting.org/?abstract=8625]
Poster: Methodology - Estimation Methods