Facilitating structural population PK model building for pharmacometricians and beyond
Brigitta Bodak (1), Karsten Kuritz (1), Anne Kuemmel (1)
(1) IntiQuan GmbH, Basel, Switzerland
Introduction: Population pharmacokinetic (PK) analyses not only require a pharmacometric understanding but also a decent knowledge of the software used for the analysis, e.g., NONMEM or Monolix. In particular, for NONMEM, a significant level of post processing of estimation output is required to derive diagnostics and, thereby, to evaluate the models properly. Pirana and psn are examples for software tools that remove some of the coding effort for the modeler [1]. Aside diagnostics, more and more approaches are evolving to also automate the decision process during model development, starting from stepwise covariate modeling (SCM [2]) for covariate model development to new approaches spanning the full model development (AMD [3], SAMBA [4]). However, all tools are usually developed for either NONMEM or Monolix such that no tool is applicable across estimation softwares. Also, automated decision processes are based on single optimization metrics like the Bayesian Information Criterion (BIC), while the model selection still should be considering more aspects like estimation robustness and clinical relevance.
Objectives: Establish a software tool to facilitate population PK base model development that
(i) is independent of the used estimation software,
(ii) saves time for the modeler on coding, and
(iii) provides sufficient information (diagnostics, parameter estimation tables) to enable decision making.
Methods:
IQRtools [5] is an R package enabling modeling across estimation softwares. It provides functionality to support programming of the data into a standardized data format which is compatible with both NONMEM and Monolix. Furthermore, it generates and exports models to the selected estimation software, runs them, and performs post-processing of the results, including generation of diagnostic plots (e.g., goodness of fit plots) and parameter estimation tables.
We implemented the function test_PKmodels as an extension to IQRtools in order to generate and run models typically needed in the base model development. Based on a review of population PK modeling use cases [6], a structural model space was defined covering more than 85% of selected final models. This includes all combinations of (i) 1 to 3 compartmental models, (ii) zero or first order absorption, (iii) with or without lag time, (iv) linear, saturable elimination, or a combination thereof, (v) additive, proportional error, or a combination of these. Using the function, the modeler can select the models in this space to be automatically generated and run in the selected estimation software, whilst it automatically post-processes the results.
Results: The use of test_PKmodels was demonstrated on an artificial dataset that was simulated with a 2-compartmental model with first order absorption without lag time. The simulated data was programmed in the general dataset format of IQRtools. The dataset was graphically explored to define the model space to be considered to describe the data. The provided functionality allowed us to set up the base PK model development for this dataset in a short concise script with minimal coding effort. That is, models with 2- and 3- compartments, linear and saturable elimination, zero and first order absorption with and without lag time, and proportional as well as additive+proportional residual errors were tested. The review of the automatically generated model comparison tables and diagnostic plots revealed that for the simulated data, the original structural model was the most appropriate model to describe the data.
Conclusions: Supporting the modeler with tools to reduce coding effort enables them to focus on the model selection process. Using test_PKmodels, estimations can be performed using the preferred modeling software. A large set of models can be coded by a limited number of lines of R code, typically around 10 to 20 lines. We have shown that by the provided diagnostics and model comparison tables, all relevant information to make a decision is gathered in one place.
We believe that a function like test_PKmodels could even be valuable beyond population PK model development. E.g., the function would allow data programmers to confirm that data sets are created appropriately by checking whether standard models can deliver reasonable individual fits. Furthermore, if NCA is impossible, e.g., due to sparse sampling, the best base model could be used to derive key PK parameters from the individual fits.
References:
[1] Keizer R.J., van Benten M., Beijnen J.H., Schellens J.H., Huitema A.D. Piraña and PCluster: a modeling environment and cluster infrastructure for NONMEM. Comput Methods Programs Biomed. (2011) 101, 72-9.
[2] Jonsson E. N., Karlsson M. O. Automated covariate model building within NONMEM. Pharmaceutical Research (1998) 15, 1463–1468.
[3] Chen X., Hamdan A., Wang S., Yang T., Nordgren R., Belin S., Huang Z., Carter S.J., Buatois S., Abrantes J.A., Hooker A.C., Karlsson M.O., Development of a tool for fully automatic model development (AMD), PAGE 30 (2022) Abstr 10091.
[4] Prague M., Lavielle M., SAMBA: A novel method for fast automatic model building in nonlinear mixed-effects models. CPT Pharmacometrics Syst Pharmacol. (2022) 11, 161–172.
[5] https://https://iqrtools.intiquan.com
[6] Schmidt H., Radivojevic A. Enhancing population pharmacokinetic modeling efficiency and quality using an integrated workflow. J Pharmacokinet Pharmacodyn (2014) 41, 319–334.