Félicien LE LOUEDEC (1,2), Mélanie WHITE-KONING (1), Fabienne THOMAS (1,2), Florent PUISSET (1,3), Étienne CHATELUT (1,2)
(1) Cancer Research Center of Toulouse (CRCT), Inserm U1037, Team 14 “Dose Individualization of Anticancer Drugs”, Université Paul Sabatier, Toulouse, FRANCE. (2) Department of Pharmacology, Institut Universitaire du Cancer de Toulouse – Oncopole, Toulouse, FRANCE. (3) Department of Pharmacy, Institut Universitaire du Cancer de Toulouse – Oncopole, Toulouse, FRANCE.
Objectives: Therapeutic drug monitoring consists in measuring the concentration of a drug in plasma, and to increase or decrease the dose depending on the value of a PK endpoint correlated to toxicity and/or efficacy. For oral anticancer drugs, such as pazopanib, the validated PK endpoint is the 24-hour trough concentration, hence the biological interpretation of the measured concentration should be straightforward.
However, in real-life settings, this interpretation may be hampered by an incorrect sampling time (i.e. other than 24h after intake) or an irregular administration scheme. In addition, dose adaptation may be complicated by non-linear pharmacokinetics. A bayesian approach, with an estimation of individual PK parameters and a re-estimation of the drug concentration, may be useful.
An approach consisting in the rough analysis of concentration data with NONMEM or Monolix is not sustainable, as it requires skills in PK modelling and interpretation of the data. Existing dose-optimisation software such as DoseMe, BestDose, or ADAPT are not free of charge, have a limited drug database and restricted settings and thus confirm the need for a new tool.
Methods: A unique tool in R was developed for data entry, parameter estimation and simulation tasks. It consists in a maximum a posteriori bayesian estimator based on the R package mrgsolve for ordinary differential equation (ODE) solving. The whole program has been made into a shiny app as a user-friendly interface.
A dataset of 51 patients treated by pazopanib and extensively sampled at day 15 (n = 32) or day 28 (n = 19) was used to validate the performance of the tool. Three population PK models were used to predict 24h post-dose concentrations (T24) based on various earlier sampling times (6h, 8h and 12h post-dose). Model 1 [1] can be considered as an external model. Model 2 and 3 are structurally close to published PK models [1,2] but population parameter values were obtained from a dataset that included the 51 patients. The estimated T24 was compared to the actual T24 by the calculation of RMSE and MAPE. The analyzes were conducted both in NONMEM and R for comparison purpose.
Results: The bayesian estimator is a slightly modified version of the example provided on the mrgsolve blog [3] with, in particular, the possibility of taking into account an additive error model for residual variability. Briefly, the estimation of individual PK parameters follows 3 steps. First, mrgsolve estimates a concentration at the chosen sampling time based on the dosing regimen (provided by the user) and a vector of PK parameters. The latter includes typical values, which are fixed, and a random component, referred to as etas, which are estimated. Secondly, an objective function value (OFV) is computed using the First-Order Conditional-Estimation with Interaction. The formulae take into account the estimated and observed concentrations with the residual variability, and the individual and population PK parameters with the inter-individual variability [4]. Lastly, a quasi-newtonian algorithm (newuoa from the minqa package) estimates the most likely eta values which minimize the OFV. Using these individual bayesian-estimated parameters it is possible, on the one hand, to estimate a concentration at a convenient sampling time (e.g. 24h post-dose), and on the other hand, to perform simulations of particular dose regimens.
At 12h post-dose, MAPE / RMSE found with Model 1, 2 and 3 are 3.7% / 19.2%, 3.8% / 16.6% and 5.4% / 17.6 % respectively; at 8h post-dose, 4.8% / 23.2%, 4.8% / 21.2% and 6% / 19.1 %; and at 6h post-dose, 6.9% / 23.7 %, 6.9% / 22.1% and 8% / 20.4 %. NONMEM and R returned the same estimates both in terms of eta values and concentrations.
Conclusions: A simple and easy-to-use tool for individual dosing of oral anticancer drugs was developed. The quality of predictions differed according to models and sampling times. The simple model (3) was the worst in terms of predictions. Model 2 returned the best estimations but was developed based on these concentration data. Thus, we implemented model 1 (Yu et al [1]) for MAP-Bayesian analyzes, and recommend drawing samples at least 12h after dose intake for clinical application. As it is fully-developed in R, this tool does not require any additional licensed software such as NONMEM or Monolix and can be used on any computer, while offering identical performances.
References:
[1] Yu H, van Erp N, Bins S, Mathijssen RHJ, Schellens JHM, Beijnen JH, et al. Development of a Pharmacokinetic Model to Describe the Complex Pharmacokinetics of Pazopanib in Cancer Patients. Clinical Pharmacokinetics 2016;56:293–303. https://doi.org/10.1007/s40262-016-0443-y.
[2] Imbs D-C, Diéras V, Bachelot T, Campone M, Isambert N, Joly F, et al. Pharmacokinetic interaction between pazopanib and cisplatin regimen. Cancer Chemotherapy and Pharmacology 2016;77:385–92. https://doi.org/10.1007/s00280-015-2953-y.
[3] Generate MAP Bayes Parameter Estimates n.d. https://mrgsolve.github.io/blog/map_bayes.html (accessed February 16, 2020).
[4] Kang D, Bae K-S, Houk BE, Savic RM, Karlsson MO. Standard Error of Empirical Bayes Estimate in NONMEM® VI. Korean J Physiol Pharmacol 2012;16:97. https://doi.org/10.4196/kjpp.2012.16.2.97.
Reference: PAGE () Abstr 9379 [www.page-meeting.org/?abstract=9379]
Poster: Clinical Applications