Artem Dolgun (1), Victor Sokolov (1), Henning Schmidt (2), Nidal Al-Hunity (3), Gabriel Helmlinger (3), Kirill Peskov (1)
(1) M&S Decisions, Moscow; (2) IntiQuan, Switzerland; (3) Quantitative Clinical Pharmacology, Early Clinical Development, IMED Biotech Unit, AstraZeneca, Boston, USA
Objectives: Quantitative systems pharmacology (QSP) is becoming an established modeling methodology used in drug development [1]. QSP modeling has found successful applications at different stages of pharmaceutical R&D [2]. Key factors limiting further expansion of QSP expansion are: (1) methodological issues, e.g., as related to complexities of parameter uncertainty estimations and model identifiability issues originating from typically sparse experimental data and often originating from multiple sources; (2) a lack of unified modeling tools, which would enable an analysis through a consistent modeling workflow, from data inputting to model development and simulations to reporting. We thus developed an integrated workflow intended to enable QSP type analyses, with industrialization features to allow for the processing of various data types (at both study and subject levels) and to provide programming support with broad analytical functionality [3]. In this work, we feature a solution for the industrialized development of mechanistic and semi-mechanistic systems pharmacology models, for both individual- and study-level data, in an R environment.
Methods: The workflow was implemented in R software (Version 4.3) and is further based on the IQR systems pharmacology and pharmacometrics toolbox (version 0.6.1) developed by IntiQuan [4], as well as the AZR tool (version 0.0.0.9) developed by AstraZeneca [5]. Graphical user interface for the workflow was developed in R shiny (version 1.0.5). IQR’s parameter estimation method [6] was used for the estimation of parameters using study-level data. This estimation method is based on a maximum-likelihood estimator, using a trust-region optimizer. The gradient and the Hessian are determined with high numerical precision through the use of symbolically derived parameter sensitivity equations. Confidence intervals for the estimated parameters are determined via the Fisher Information Matrix. In addition, IQR enables the calculation of point-wise finite sample confidence intervals through likelihood profiling, using the algorithm published in [7]. Estimation of parameters based on individual-level data was performed through NONMEM or Monolix fitting algorithms.
Results: The workflow for the development of mechanistic QSP models was established in a highly flexible syntax, using available IQR and AZR tools. The workflow enables the following steps: compilation of a standardized dataset; data exploration; parameter estimation procedures; model diagnostics; sensitivity/identifiability analyses; assessment of uncertainty around predicted mean; and generation of a report. A standardized ‘.csv’ dataset is used as input to the abovementioned packages, with a unified structure for the different data types; this allows for calling of population software, through the R environment, for individual-level data and manual handling without programming software, if study-level data were to be used for parameter estimation. Once the dataset is compiled, it is ready for automated processing, to the last phase of the workflow.
Goodness-of-fit metrics and a rich toolkit for model testing, e.g., Observed vs. Predicted and Residual plots as well as longitudinal profiles are incorporated into the main framework routine.
Sensitivity analysis is implemented in the workflow using a host of curves and tornado plots for selected timepoints. Identifiability of estimated parameters is assessed by measuring the gradient and the Hessian for the objective function, and objective function profiling.
Monte-Carlo simulations were applied to generate multiple trial data by simulating experimental data from mean and SE and, therefore, to obtain estimation of parameter distributions taken into observed uncertainty in the data.
Automated generation of a QSP modeling report is the last step of the established workflow. The workflow is compliant with RSTAN, which supports the development of fully Bayesian QSP models based on the Hamiltonian MC methods.
Conclusion: In this work, we propose an industrialized workflow for the development of mechanistic and semi-mechanistic systems pharmacology models, applicable at both study- and individual-level data, using R-based packages, and further integrated with commonly used software in Pharmaceuticals, such as NONMEM and Monolix. Such a functional, integrative workflow enables all key steps, from data inputting to report generation, necessary to perform and apply QSP modeling.
References:
[1] Helmlinger G, et al. Eur J Pharm Sci. 2017 Nov 15;109S:S39-S46. doi: 10.1016/j.ejps.2017.05.028.
[2] Peterson MC, et al. CPT Pharmacometrics Syst Pharmacol. 2015 Mar;4(3):e00020. doi: 10.1002/psp4.20.
[3] Gadkar K, et al. CPT Pharmacometrics Syst Pharmacol. 2016 May;5(5):235-49. doi: 10.1002/psp4.12071.
[4] Sunnåker M, et al. ACoP7, 2016.
[5] Fox R, et al. ACoP8, 2017.
[6] http://www.intiquan.com/iqr-tools/
[7] Kaschek et al. doi: https://doi.org/10.1101/085001
Reference: PAGE 27 (2018) Abstr 8700 [www.page-meeting.org/?abstract=8700]
Poster: Methodology - Model Evaluation