Tomás Sou1, Brigitte Lacroix2, Himanshu Naik2, Akash Khandelwal2, Neva Coello1
1Novartis, 2UCB Biopharma
Introduction: Covariate model building is a crucial step in pharmacometric analyses. The selected covariate model may be used to support individualised dosing and the design of clinical trials in drug development [1]. Following judicious covariate scoping, search algorithms may be used to facilitate covariate selection from a curated list of candidates. Automated procedures such as Stepwise Covariate Modelling (SCM), Conditional Sampling use for Stepwise Approach (COSSAC) [2] and Stochastic Approximation for Model Building Algorithm (SAMBA) [3] are often time-consuming requiring many iterations. Full model methods such as full fixed effect model (FFEM) and full random effect model (FREM) [4,5] are subject to challenges such as model instability and the need of data reconfiguration. The horseshoe prior is a special shrinkage prior symmetric around zero with fat tails and an infinitely large spike at zero [6]. In the Bayesian framework, this makes it ideal for sparse models that have many regression coefficients, although only a subset of them is non-zero. This work presents the implementation of regularised horseshoe prior as an alternative approach to support fast and efficient covariate selection in pharmacometric modelling. The procedure requires minimal preparation and provides uncertainty information with posterior distributions to guide covariate selection. Objectives: This work aims to demonstrate the implementation of regularised horseshoe prior as a fast and efficient approach to guide covariate selection in pharmacometric modelling. Methods: In this work, a population pharmacokinetic (PopPK) model developed for minzasolmin, an investigational treatment for Parkinson’s disease (PD), was used to evaluate covariate search algorithms. Initially, a base model was developed to describe the concentration-time profiles of the drug collected in healthy volunteers and PD patients from five clinical studies. The modelling dataset included 7727 plasma concentrations from 462 subjects. Baseline covariates available from the studies included age (AGE), body weight, sex (SEX), ethnicity, disease status (POPTYPE), alanine transaminase, aspartate transferase, renal function and the use of concomitant medications including CYP3A4 inhibitors (CYP3A4INH), CYP3A4 inducers and acid-reducing agents. Following base model development and covariate scoping, candidate covariates were evaluated for covariate model building. For covariate selection using horseshoe prior, empirical Bayes estimates and covariate values were extracted from the base model and standardised before the procedure. The regularised horseshoe prior (HSP) covariate selection procedure was implemented in R (version 4.3.1) supported by the brms package [7]. Covariates were selected based on the posterior distributions of their coefficients and were selected if the 95% credible intervals did not include zero. The selected covariates were then included in the final covariate model. For comparison, automated procedures available in the Monolix Suite 2023R1 including SAMBA, COSSAC and SCM were also performed. Results: The base model selected was a two-compartment disposition model with first-order absorption and a lag time before absorption. Following covariate scoping, candidate covariates were evaluated on key model parameters including apparent clearance (CL), apparent central volume of distribution (V1) and absorption rate constant (ka). The HSP procedure was able to evaluate all candidate covariates on a given parameter and provide posterior distributions of their coefficients efficiently in a single run. The HSP results suggested AGE, CYP3A4INH, SEX and POPTYPE as significant covariates on CL and SEX as a significant covariate on V1. The HSP selected covariates were comparable to the ones suggested by SAMBA, COSSAC and SCM. However, the HSP procedure was significantly faster requiring substantially shorter runtime (65.86 sec) compared to SAMBA (65x faster), COSSAC (295x faster) and SCM (1447x faster). Conclusion: The results show that the regularised horseshoe prior procedure was able support covariate selection comparable to other automated search algorithms with a much shorter runtime. The procedure can serve as an alternative approach to guide fast and efficient covariate model building in pharmacometric analyses.
[1] Sanghavi K et al. CPT Pharmacometrics Syst Pharmacol. 2024 May;13(5):710-728. [2] Aryal et al. CPT Pharmacometrics Syst Pharmacol. 2021 Apr;10(4):318-329. [3] Prague et al. CPT Pharmacometrics Syst Pharmacol. 2022 Feb;11(2):161-172. [4] Yngman et al. CPT Pharmacometrics Syst Pharmacol. 2022 Feb;11(2):149-160. [5] Jonsson et al. CPT Pharmacometrics Syst Pharmacol. 2024 Aug;13(8):1297-1308. [6] Carvalho et al. PMLR 2009;5:73-80. [7] https://CRAN.R-project.org/package=brms
Reference: PAGE 33 (2025) Abstr 11683 [www.page-meeting.org/?abstract=11683]
Poster: Methodology - Covariate/Variability Models