Anna Mikhailova1,2, Kirill Peskov1,2,3, Gabriel Helmlinger4, Victor Sokolov1,3
1M&S Decisions LLC, 2Research Center of Model-Informed Drug Development, I.M. Sechenov First Moscow State Medical University, 3Marchuk Institute of Numerical Mathematics RAS, 4Quantitative Medicines
Objectives: Dapagliflozin is a sodium-glucose cotrasporter-2 inhibitor originally developed for the treatment of type 2 diabetes mellitus and subsequently approved in several other indications, including heart failure and chronic kidney disease [1]. Several clinical trials (CTs) have been conducted, and multiple pharmacokinetic (PK) models have been published for this drug. Based on this body of data and models, we sought to develop a generalized PK model integrating all available information and data. Such a model would allow to support post-approval activities (e.g., local registration or life cycle management); it may also serve as a part of a broader mechanistic model with multiple pharmacodynamic and clinical endpoints, that explicitly incorporates dapagliflozin mechanisms of action. In this work, we propose a seamless Bayesian workflow for the development of minimal physiologically-based pharmacokinetic (mPBPK) models; the workflow was tested in three Bayesian packages. Methods: The developed workflow consists of three parts: (1) systematic data search (SDS); (2) model calibration and diagnostics; (3) model validation and evaluation. The SDS was performed using the PubMed database, to identify and digitize available quantitative data on dapagliflozin PK and related existing PK models. Data from parallel-group CTs were used for model calibration. Parameters were estimated using three Bayesian modeling packages: Nimble (v1.1.0) with RxODE2 (v2.1.2), MCSim (v6.2.0), and Stan (v2.32.6) with Torsten (v0.89.0). Model convergence was assessed by numerical criteria (Gelman-Rubin statistics [2], effective sample size (ESS) [2], etc.) and graphical diagnostics. Software performance was compared by numerical metrics (Watanabe-Akaike Information Criterion (WAIC) [3], ESS, computational time, etc.) and graphical diagnostics. The mathematical model of dapagliflozin renal excretion was adapted from the Sokolov et al. publication [4]. The model was validated using dapagliflozin urine recovery (UR) and PK data from CTs with a cross-over design. Additionally, a local sensitivity analysis (SA) was conducted using the one-at-a-time method. Results: 10 dapagliflozin PK models were identified through SDS. All models consisted of 2 compartments, 6 out of 10 models incorporating delayed absorption. Consequently, a two-compartment model with transit compartments was selected for further analysis. Prior distributions of model parameters were chosen in accordance with historical information. For typical value parameters, prior distributions were set to lognormal, with the mean and variance equal to the weighted mean and variance of the corresponding parameters in the published models. For residual error variance, a uniform prior distribution within the (0, 100] range was set, as no historical information was available for this parameter. Model convergence was achieved in all programming packages. Posterior distribution medians, obtained by Torsten, were ka = 11.5 1/h, ktr = 10.7 1/h, Vd = 71.3 L, Cl = 17.1 L/h, Vp = 130.0 L, Q = 8.2 L/h. 95% credible intervals (CI), obtained by MCSim and Nimble, overlapped with the Torsten derived intervals. Posterior predictive distributions generated by different programs exhibited comparable descriptive performances: 95% CI of these distributions overlapped; WAIC metrics differed by less than 1%. Stan exhibited superior sampling efficacy compared to Nimble and MCSim – demonstrating a 5.1-fold increase in efficiency in the central posterior region and a 8.5-fold increase in the tails. However, this efficiency came at a cost, with Stan requiring an average 5.7 times longer runtime. Predicted UR of dapagliflozin ranged from 2.1% to 4.1% with a mean of 3.2%, while the observed UR varied from 0.8% to 4.0%, yielding a mean value of 2.0%. Although the predicted UR interval overlapped with that observed in the CTs, the mean value was arguably overestimated. SA indicated that the model was particularly responsive to Cl, GFR and the fraction unbound parameters; these are indeed critical to account for most of the between-trial variability. Conclusion: In this work, a Bayesian approach was employed to develop a mPBPK model for dapagliflozin, building upon and generalizing from previously published PK models and data related to the drug. The workflow established in this research is seamless and can be adapted for other types of mechanistic models. Supported by the Russian Science Foundation (Grant Number 23-71-10051)
[1] PRESCRIBING INFORMATION FARXIGA. FDA, 2014. Viewed: 20.02.2025 [Online]. Can be accessed on: https://www.accessdata.fda.gov/drugsatfda_docs/label/2024/202293s031lbl.pdf [2] A. Vehtari, A. Gelman, D. Simpson, B. Carpenter, P.-C. Bürkner, Rank-Normalization, Folding, and Localization: An Improved R ^ for Assessing Convergence of MCMC (with Discussion), Bayesian Anal., 16, 2. 2021, doi: 10.1214/20-BA1221. [3] S. Watanabe, Asymptotic Equivalence of Bayes Cross Validation and Widely Applicable Information Criterion in Singular Learning Theory, 2010, arXiv. doi: 10.48550/ARXIV.1004.2316. [4] V. Sokolov et al., A mechanistic modeling platform of SGLT2 inhibition: Implications for type 1 diabetes, CPT Pharmacometrics & Syst. Pharm., 12, 6, 831–841. 2023, doi: 10.1002/psp4.12956.
Reference: PAGE 33 (2025) Abstr 11574 [www.page-meeting.org/?abstract=11574]
Poster: Methodology - Estimation Methods