Reducing Animal Use in Toxicological Research through Open-Source Individualized (PB)PK/QST Modeling: A Case Study with Phenobarbital in Dogs

Objectives: New Approach Methodologies (NAMs)—including in vitro systems, computational modeling, and organ-on-chip technologies—are increasingly being developed, validated, and integrated into regulatory science frameworks worldwide. As part of ongoing efforts to advance NAMs in toxicological research, we present an open-source modeling framework using R and the Open Systems Pharmacology (OSP) Suite [ 1] that integrates physiologically based pharmacokinetics (PBPK), classic pharmacokinetics (PK), and quantitative systems pharmacology/toxicology (QSP/T). Our framework is part of a broader platform developed for the CRACK IT Virtual Second Species Challenge [ 2]. The model focuses on phenobarbital in dogs, which has rare but potentially severe hepato- and hematotoxic effects. It leverages individual-level data from a small study of eight dogs [ 3], reflecting a common challenge in drug development and toxicological assessment, where limited sample sizes constrain traditional approaches. Our workflow utilizes individualized parameter fitting, which can enhance toxicological insights through optimal animal study design and reduce the need for animal use. Methods: A PBPK and a PK model were calibrated to single-dose phenobarbital data and validated against repeat dosing scenarios. A QSP/T model describing the dynamic life cycle of three common liver biomarkers—ALT, AST, and ALP—and three bone marrow biomarkers—neutrophils, thrombocytes, and red blood cells—was fit to data from [ 3] and [ 4], respectively. An individualized fitting approach was used to capture inter-animal variability. A genetic algorithm-based optimization strategy, incorporating lognormal variability in key pharmacokinetic and pharmacodynamic parameters identified via Sobol global sensitivity analysis, was applied to generate individualized parameter sets reproducing observed drug concentrations and biomarker trajectories. Dose escalation and prolonged administration were then simulated to link individual variability with susceptibility to toxicity. Results: The model successfully captured the diversity in pharmacokinetic profiles and biomarker responses and revealed a spectrum of sensitivity profiles consistent with the experimental data. For instance, for one dog with observed baseline ALP of 17 IU/L the model predictions achieved an RMSE of 0.65 IU/L and an R 2 of 0.99 while for another dog with observed baseline ALP of 51 IU/L the model demonstrated an RMSE of 1.57 IU/L and an R 2 of 0.99, indicating high accuracy in capturing both baseline and treatment-related biomarker dynamics. In the extrapolation study, high-dose and chronic liver and bone marrow toxicity could be explored pointing to low susceptibility of the dogs in the study [ 3]. Conclusions: Our work highlights the utility of mechanistic PBPK-PK-QSP/T modeling combined with individualized parameter fitting to recapitulate inter-animal variability in drug disposition and biomarker dynamics. The open-source framework contributes to ongoing NAMs efforts by providing a reproducible, data-driven approach to evaluate compound toxicity and reduce animal use.