Félicien Le Louedec1,2, Laura Morvan1, Fabienne Thomas1, Florent Puisset1, Étienne Chatelut1
1Institut Claudius Regaud, Institut Universitaire du Cancer de Toulouse Oncopole, Centre de Recherche en Cancérologie de Toulouse, INSERM U1037, Université Paul Sabatier, 2Pharmetheus AB (current affiliation)
Introduction Conventional therapeutic drug monitoring (TDM) of oral kinase inhibitors consists in comparing a concentration, measured exactly at the time of trough and steady-state, to a target value linked to efficacy and/or adverse events [1]. With model-informed precision dosing (MIPD), a patient dose could be individualized with the following algorithm: (1) Start of treatment at the approved dose. (2) TDM at any time, maximum a posteriori (MAP) Bayesian estimation of pharmacokinetics (PK) parameters, prediction of a trough concentration at steady-state (Cmin,ss) and comparison to an efficacy target value. (3a) If predicted Cmin,ss > target, the same dose is kept. (3b) If predicted Cmin,ss < target, simulation a posteriori of a finite number of available dosing regimen and selection of a dose with a simulated Cmin,ss > target [2]. The objective of this work was to report a comprehensive fit-for-purpose validation of the MIPD of sunitinib and pazopanib to predict Cmin,ss from any single concentration observed during an interdose interval. It included the external validation of existing PK models and the possible advantage of flattened priors [3] or model-averaging [4]. Methods Single concentration data from 41 renal cancer patients treated with sunitinib or pazopanib, included in the SUP-R trial (NCT02555748), were analyzed. For each drug, four prediction scenarios were studied: concentrations measured 2 (C1T2) or 6 hours (C1T6) after an intake were analyzed in order to predict individual Cmin,ss either at the current cycle (C1T0) or at the next cycle (C2T0, 6 weeks later). The motivation was to evaluate the performance of predictions at the current cycle, and if this performance would still be valid at the next cycle with a possible dose modification informed by the estimated PK parameters. For each scenario, 5 or 6 models from the literature describing the PK of sunitinib (+ active metabolite ND-sunitinib) or pazopanib, respectively, were tested. In addition to single-model analyses, a model-averaging procedure was applied. Finally, each parameter estimation was conducted with four levels of flattened priors: 1 (original model), 0.3, 0.1, and 0.03. The quality of predictions was measured with bias (mean prediction error, MPE) and precision (root mean square error, RMSE). All the analyses were conducted using the R package mapbayr [5]. Results For sunitinib, the precision of prediction of C1T0 from C1T2 varied between the models, with RMSE ranging from 17.2% (Houk et al [6]) to more than 70%. The precision of prediction of C1T0 from C1T6 varied to a similar extent between the models: RMSE ranging from 15.6% (Yu et al [7]) to more than 70%. The precision of prediction of C2T0 was slightly worse than to predict C1T0 (about +2% to +7% of RMSE, depending on the model and the sampling time). Model averaging did not outperform any single model, but had a consistent precision whatever the sampling time (C1T2 or C1T6) and the Cmin,ss to predict (C1T0 or C2T0), with RMSE between 20.1% and 23.6%. Flattened priors did not markedly improve or worsen the predictions, except for models that initially performed very badly (e.g., RMSE decreased from >70% to 17.7% in a scenario with Khosravan et al [8]). Bias varied between models and scenarios but was generally small, e.g. MPE between -4.4% to +5.1% with model-averaging depending on the scenario. For pazopanib, performances were similar or slightly worse than for sunitinib. The precisions of prediction of C1T0 and C2T0 from C1T2 were at the best 22.8% (Mourey et al [9]) and 25.9% (Imbs 2016 et al [10]), respectively, but all the models had an RMSE > 35% if the estimation was done from a C1T6 concentration. Model-averaging performed consistently well between the scenarios: precisions of prediction of C1T0 and C2T0 from C1T2 of 27.2% and 24.8%, respectively, and bias < 5%. Flattened priors had no influence on MPE and RMSE. Conclusions For both drugs, model averaging performed with consistency across the sampling and forecasting scenarios. It should be preferred to the choice of a single original model. The random intra-individual variability of PK worsened the quality of prediction at the next cycle. Thanks to the dedicated embedded features of mapbayr (data assembling, MAP estimation, model averaging, a posteriori simulations etc), this framework was successfully implemented in our laboratory within a shiny application where TDM data are interpreted to individualize patient dose.
[1] Verheijen R et al, Clin Pharmacol Ther, 2017. [2] Minichmayr et al, Adv Drug Deliv Rev, 2024. [3] Hughes et al, CPT Pharmacometrics Syst Pharmacol, 2021 [4] Uster et al, Clin Pharmacol Ther, 2021 [5] Le Louedec et al, CPT Pharmacometrics Syst Pharmacol, 2021 [6] Houk et al, Clin Cancer Res, 2009 [7] Yu et al, Br J Clin Pharmacol, 2015 [8] Khosravan et al, J Clin Pharmacol, 2010 [9] Mourey et al, J Geriatric Oncol, 2021 [10] Imbs et al, Cancer Chemother Pharmacol, 2016
Reference: PAGE 33 (2025) Abstr 11526 [www.page-meeting.org/?abstract=11526]
Poster: Clinical Applications