Ana Yisel Caballero Alfonso 1, Hanna Leithner 2, Jorin Diemer 1, Venetia Karamitsou 1, Behnam Amiri 2, Susana Proenca 1, Cleo Demeester 1, Jure Fabjan 2, Christian Maass 2, Alexander Kulesza 1, Stephan Schaller 1
1 ESQlabs GmbH (Saterland, Germany), 2 MPSlabs, ESQlabs GmbH (Saterland, Germany)
Objectives:
Microphysiological systems (MPS), including liver spheroids and organoid-based platforms, are increasingly applied to assess hepatotoxic risk during drug development [1]. While these models provide greater physiological relevance, translating in vitro potency metrics (e.g., IC₅₀) into in vivo drug-induced liver injury (DILI) remains challenging due to complex exposure dynamics and inter-individual variability. To address this, we present a comprehensive Predictive Safety Ecosystem comprising two parallel, complementary streams. Stream 1 focuses on a hybrid mechanistic–machine learning (ML) framework to refine MPS-derived hepatotoxicity metrics. Stream 2 expands this into a full in vitro-to-in vivo extrapolation (IVIVE) workflow using a Physiologically Based Pharmacokinetic-Quantitative Systems Toxicology (PBPK-QST) platform built within the Open Systems Pharmacology (OSP) Suite [2]. This second stream incorporates quantitative Adverse Outcome Pathways (qAOPs) and virtual populations for humans (EU Horizon ONTOX project) [3] and dogs (NC3Rs CRACK IT Virtual Second Species) [4] to predict population-level DILI incidence.
Methods:
Stream 1 (Hybrid Mechanistic-ML): A reduced-order intracellular exposure model was parameterized to represent MPS-specific geometry and experimental conditions. PBPK simulations were used to estimate human in vivo intracellular liver exposure. Compound-specific correction factors reflecting peak-driven (Cmax) and cumulative (AUC) exposure were derived and applied to in vitro concentration–toxicity metrics across three liver-on-chip datasets (89–144 compounds). Principal component analysis (PCA) and ML classification were applied to the corrected feature space.
Stream 2 (PBPK-QST & Virtual Populations): The QST platform translates intracellular exposure into perturbations of liver cellular or metabolic processes that lead to steatosis (lipid accumulation dynamics), cholestasis (bile acid transport inhibition), or necrosis (hepatocyte death), thereby quantifying existing AOPs. These perturbations are then read out as phenotype-specific DILI markers [5]. To address inter-individual variability and to account for the often idiosyncratic nature of DILI, a large virtual population approach was used. For human risk assessment (ONTOX), DILI-pattern-specific parameter distributions were cross-informed by compound-specific surrogates (in vitro / MPS). For preclinical translation, a virtual dog population was developed (CRACK IT challenge) to simulate canine toxicokinetics and hepatotoxicity (serving as a “Virtual Second Species”), and to infer the prevalent damage pattern from study-level biomarker data.
Results:
Stream 1: PCA revealed that mechanistically corrected exposure features reorganized the feature space into distinct, partially orthogonal axes corresponding to peak- and duration-driven exposure, highlighting that DILI risk is not governed by a single exposure dimension. While ML models incorporating these features showed modest aggregate predictive gains over baselines, feature importance analyses consistently ranked mechanistically corrected exposure metrics among the most informative predictors, revealing multiple exposure modes underlying toxicity.
Stream 2: The PBPK-QST platform successfully translated in vitro and MPS readouts into dynamic in vivo biomarker trajectories. The qAOP models accurately predicted the dose-dependent onset of steatosis and cholestasis for reference compounds. The virtual human population (ONTOX) successfully captured the low-frequency incidence of idiosyncratic DILI events by mimicking both exposure variability and pre-existing conditions that favor further liver injury. Furthermore, the Virtual Second Species dog model accurately reproduced canine toxicokinetic profiles and liver enzyme elevations for reference hepatotoxins, which, by its dynamic and dose-dependent nature, allows for extrapolations and thus serves as an alternative to reduce animal use (fewer dosing groups or even replacing chronic studies) in preclinical safety testing.
Conclusion:
The ESQlabs/MPSlabs Predictive Safety Ecosystem demonstrates that combining data-driven exposure corrections with mechanistic QST modelling provides a robust, scalable framework for DILI prediction. While Stream 1 enhances the biological interpretability of MPS assays by revealing nuanced exposure-toxicity relationships, Stream 2 bridges the translational gap. By leveraging qAOPs and virtual populations, the platform successfully predicts clinical and preclinical DILI incidence, supporting mechanistic hypothesis testing, nuanced risk prioritization, and advancing New Approach Methodologies (NAMs) in regulatory safety assessment.
References:
References to Add at the Bottom
[1] Ewart, L., Apostolou, A., Syed, S. A., et al. (2022). Performance assessment and economic analysis of a human Liver-Chip for predictive toxicology. Communications Medicine, 2(1), 154.
[2] Lippert, J., Burghaus, R., Edginton, A., et al. (2019). Open Systems Pharmacology Suite—An Open Source Collaborative Trendsetter for Comprehensive Pharmacometric Modeling and Simulation. CPT: Pharmacometrics & Systems Pharmacology, 8(12), 873-890.
[3] Vinken, M., Benfenati, E., Busquet, F., et al. (2021). ONTOX: ontology-driven and artificial intelligence-based repeated dose toxicity testing of chemicals for next generation risk assessment. Toxicology, 459, 152862.
[4] NC3Rs. (2022). CRACK IT Challenge 42: Virtual Second Species. National Centre for the Replacement, Refinement and Reduction of Animals in Research. Available from: https://nc3rs.org.uk/crackit/virtual-second-species
[5] Spinu, N., Cronin, M. T., Enoch, S. J., et al. (2020). Quantitative adverse outcome pathways: a review. Archives of Toxicology, 94(5), 1497-1510.
Reference: PAGE 34 (2026) Abstr 11889 [www.page-meeting.org/?abstract=11889]
Poster: Drug/Disease Modelling - Other Topics