Automatic translation of Bayesian pharmacometric models: the PharmML-to-WinBugs converter
Cristiana Larizza (1), Lorenzo Pasotti (2), Enrica Mezzalana (2), Elisa Borella (2), Gareth Smith (3), Paolo Magni (2)
(1) Laboratory for Biomedical Informatics
Objectives: PharmML is a markup language for pharmacometric models description, under development by the DDMoRe consortium, that will enable the tool-independent formulation, exchange and integration of models and tasks [1]. The Modelling Description Language (MDL), also under development, is a human-readable standard language aimed to facilitate model writing and enable, via automatic translation, the generation of PharmML-encoded models that can be converted into the desired target language [2]. This work describes the efforts undertaken for the development of a PharmML-to-WinBugs converter, which will support Bayesian model estimation tasks in fully integrated interoperable workflows.
Methods: Converter development included 4 main steps. 1) Features for Bayesian model support, currently under revision, were proposed for PharmML and MDL, including extensions for parametric, non-parametric and empirical prior distributions. 2) A number of increasingly complex models were defined in PharmML and WinBugs to describe all the relevant situations. 3) A PharmML-to-WinBugs translation tool was developed via Java, using libPharmML to read/validate PharmML files and some libraries to generate an intermediate model representation, suitable for final translation into WinBugs. 4) A conversion tool was developed via R to translate NONMEM-format data files into WinBugs input files, to create all the elements needed to use the model. PharmML 0.4, WinBugs 1.4.3, BlackBox 1.5, PKPD Model Library 1.2, and the WBDiff and WBDev interfaces were used.
Results: PharmML files are correctly translated into WinBugs and related Pascal files. The currently supported features are: single-subject and population models, algebraic and ordinary differential equations (ODEs), independent distributions at prior and inter-individual levels, multiple Observation Models with additive Gaussian error, time-varying continuous covariates, Function Calls, all Individual Parameter types, and transformation of Covariates, Individual Parameters and Observation Models. Three options are available to solve ODEs: 1) the inline ode block or 2) Pascal code via WBDiff, and 3) Pascal code via PKPD Library/WBDev, which can solve models with multiple dosing.
Conclusions: The converter supports a large number of situations of interest in Pharmacometrics. Its development will be completed to include the remaining PharmML features not currently implemented.
This work was supported by the DDMoRe project (www.ddmore.eu).
References:
[1] Swat MJ, Wimalaratne S, Kristensen NR. PharmML 1.0 - An exchange standard for models in Pharmacometrics. Proceedings of the PAGE meeting (2014) June 10-13, Alicante, Spain.
[2] Harnisch L, Matthews I, Chard J, Karlsson MO. Drug and Disease Mode Resources: a consortium to create standards and tools to enhance model-based drug development, CPT Pharmacometrics Syst Pharmacol (2013) 2(3): e34.
