Varvara Ustinova (1), Kirill Zhudenkov (1), Leonid Stolbov (1), Gabriel Helmlinger (2), Kirill Peskov (1, 3), Victor Sokolov (1)
(1) M&S Decisions, Moscow, Russia; (2) Clinical Pharmacology & Toxicology, Obsidian Therapeutics, Cambridge, MA, USA; (3) Computational Oncology group, I.M. Sechenov First Moscow State Medical University, Moscow, Russia
Introduction: Quantitative systems pharmacology (QSP) is a modeling methodology used in biomedical sciences and pharmaceutical R&D, to generate and test hypotheses when the relevant questions require a deeper understanding of biological, pharmacological, and delivery mechanisms [1]. QSP models are often built based on heterogeneous aggregated study-level data implemented into a single quantitative framework. Various methods can be applied to address this challenge, including the search for study-level covariates, implementation of regressors, or introduction of random effects on specific parameters. In this research communication, the latter option was tested and applied towards a previously developed QSP model which benchmarked multiple anti-PCSK9 monoclonal antibodies (mAbs) and small interfering RNA (siRNA), with between-study variability accounted for by regressors and random effects.
Objectives: The objectives of this work were to develop and assess a fully integrated QSP modeling workflow in a Monolix environment, which would account for between-study variability via incorporation of a hierarchical model structure.
Methods: We considered a PCSK9 QSP model that describes the physiological interplay among key dyslipidemia biomarkers, including PCSK9, LDL cholesterol (LDLc), VLDL cholesterol, HDL cholesterol, apoliprotein B, lipoprotein A and triglycerides, and their dynamic modulation upon treatment with anti-PCSK9 mAbs or siRNA. The original base model included 17 ODEs, 21 estimated parameters and 2 regressors, either taken or imputed from published sources. In order to assess between-study variability, two regressors (baseline PCSK9 and LDLc concentrations) were reconsidered as parameters with random effects. The model was parametrized using aggregated study-level data upon mAb (alirocumab and evolocumab – 958 data points from 14 studies) and siRNA (inclisiran, ALN-PCS – 548 data points from 3 studies) treatments, with subsequent external validation on independent mAb (48 data points from 1 study) and siRNA (54 data points from 1 study) datasets.
Results: The QSP model was successfully implemented and assessed in a Monolix environment. Estimated parameter values were within the ranges of 95% CI’s of previously analyzed values [2]. Model quality was evaluated using multiple criteria: (i) change in the objective function value (including -2 log-likelihood and Akaike information criterion), (ii) inspection of goodness-of-fit plots (observations vs. predictions, residuals and random effects plots, etc.), and (iii) precision of parameter estimates (based on estimated relative standard error values). The best model was determined using a convergence assessment tool (multiple parameter estimation with different, randomly generated, initial parameter values). Given the model complexity vs. the amount of available data, several PD parameters were fixed in order to overcome parameter non-identifiability, once baseline LDLc and PCSK9 levels were implemented as parameters with random effects instead of regressors. Random effects for PCSK9 and LDLc baseline levels were used to assess between-study variability, in addition to model-driven uncertainty in parameter estimates when analyzing the relationship between plasma PCSK9 and LDLc as well as lipid/lipoprotein biomarker responses following plasma LDLc lowering under different anti-PCSK9 therapies. Subjects with heterozygous familial hypercholesterolemia were found to have significantly higher baseline levels of both PCSK9 and LDLc, which were reflected in the corresponding covariates. Variability in baseline levels affected primarily model predictions for mAb treatment, due to the differences in mechanisms of action between mAbs and siRNA modalities.
Conclusions: This research highlights challenges in applying a mixed-effects methodology within the QSP model development framework using aggregated study-level data. We successfully evaluated and adapted a QSP model of lipoprotein homeostasis and anti-PCSK9 therapies with random effects in a Monolix environment. The presented workflow may be effectively applied towards a wide range of model-informed drug development projects.
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] Sokolov V, et al. J Lipid Res. 2019 Sep;60(9):1610-1621. doi: 10.1194/jlr.M092486
Reference: PAGE () Abstr 9320 [www.page-meeting.org/?abstract=9320]
Poster: Methodology - Model Evaluation