Kevin Mcnally1, Hiroshi Momiji1, Anthonia Onasanwo1, Pauline Bogdanovich1, Frederic Bois1, Masoud Jamei1, Céline Brochot1
1Certara UK
Introduction: Sensitivity Analysis (SA) describes a class of methods for studying the relationships between the inputs and outputs of a model [1]. It is generally expected that SA forms part of PBPK-based regulatory submission to address data gaps, uncertainty, and variability in key model parameters [2]. Variance-based methods are regarded as the gold standard for global sensitivity analysis (GSA) [3]; however this approach requires many thousands of evaluations of the model under study, which is not feasible for complex PBPK models that take several minutes for a single evaluation. A meta-modelling approach [4], whereby sensitivity analysis is conducted on a statistical approximation to the PBPK model, has the potential to reduce the computational burden of GSA by several orders of magnitude with little loss of precision, thus making variance-based GSA feasible for computationally expensive PBPK models [5]. Objectives: The objectives of this work were: 1) to develop a prototype workflow in R for meta-model-based GSA of the Simcyp Simulator models; 2) to study the correspondence between (variance-based) sensitivity indices calculated using the meta-model approximation and the PBPK model. Workflow development used a relatively fast PBPK model where variance-based GSA was computationally feasible. Methods: A Simcyp V24 workspace for the oligonucleotide Vutisiran following a single sub-cutaneous dose was used in workflow development. Five uncertain parameters of the model were studied in GSA: the fraction unbound in plasma, the equilibrium dissociation rate constant, the endocytosis uptake rate in adipose, the redistribution in adipose, and the nonspecific clearance by tissue nucleases. The spacefillr package was used to generate the first 256 points of a five-dimensional Sobol sequence. The Simcyp-R package was used to run the PBPK models at each of these design points and extract PK summary measures. Bespoke R scripts were written for building and validating Gaussian Process (GP) regression models using design points and output data on the summary measures Tmax, Cmax and AUC of the Vutisiran plasma concentration: the first 64, 128 and all 256 points were successively used in the GP models to investigate the impact of sample size on the method precision. The sobolSalt method from the sensitivity package was used for computation of sensitivity indices, with predications from GP models used as an approximation to the PBPK model outputs. GSA results were benchmarked against the equivalent routine run in Simcyp using 16000 runs of the PBPK model. Results: GP regression models were successfully fitted to output data on Tmax, Cmax and AUC. Visual checks of observations against predictions supplemented by calculations of Nash–Sutcliffe efficiency suggested 128 points were required to train the GP model for an acceptable quality of fit. The sensitivity indices calculated using PBPK model outputs and GP models differed by less than 1% in all cases when at least 128 design points were used to train the GP models. A computational efficiency gain of more than two orders of magnitude was achieved compared to a conventional GSA. Conclusions: The meta-modelling approach uses information in both model inputs and model outputs and consequently allows massive efficiency gains to be achieved. Results suggest that so long as GP models has been appropriately validated, the GSA results achieved using this approximation are reliable. The relative efficiency gain for larger models will be much greater: the few hundred runs of the model required to build meta-models would make a variance-based GSA feasible for even the most computationally expensive models supported by Simcyp
[1] A. Saltelli et al. (2008). Global Sensitivity Analysis. The Primer. John Wiley & Sons (2008). [2] EMA (2019). Guideline on the reporting of physiologically based pharmacokinetic (PBPK) modelling and simulation. EMA/CHMP/458101/2016. [3] A. Saltelli et al. (2019). Why so many published sensitivity analyses are false: A systematic review of sensitivity analysis practices. Environmental Modelling & Software (2019) 114:29-39. [4] Oakley, J. and O’Hagan, A. (2004). Probabilistic Sensitivity Analysis of Complex Models: A Bayesian Approach. Journal of the Royal Statistical Society: Series B, 66(3). [5] Lumen, A., McNally, K., George, N., Fisher, J.W. and Loizou, G.D. (2015). Quantitative Global Sensitivity Analysis of a Biologically Based Dose-response Model for the Thyroid Endocrine System. Frontiers in pharmacology: Predictive Toxicity, 6.
Reference: PAGE 33 (2025) Abstr 11569 [www.page-meeting.org/?abstract=11569]
Poster: Drug/Disease Modelling - Absorption & PBPK