A pipeline for generating large virtual dog populations to predict preclinical drug-induced liver injury

Objectives: Physiologically Based Pharmacokinetic (PBPK) and Quantitative Systems Pharmacology (QSP) models are increasingly used in drug development ([ 1], [ 2]), including for Drug-Induced Liver Injury (DILI)-a leading cause of drug attrition. Traditional DILI assessment methods are limited, especially for rare adverse events. Existing computational approaches ([ 3],[ 4],[ 5]) are often closed-source, lack explainability, or fail to capture idiosyncratic DILI [ 6]. We present an endeavor to build an open-source DILI model and scalable workflow using R and the Open Systems Pharmacology (OSP) Suite [ 7] to generate large virtual canine populations (>10,000) reproducing three major DILI phenotypes with realistic incidence. A part of the CRACK IT Virtual Second Species Challenge [ 8], our PBPK-QSP approach integrates a mechanistic representation of drug kinetics and toxicity mechanisms. Cloud computing and machine learning are utilized to enable the rapid generation of diverse virtual patients. Methods: Sobol Global Sensitivity Analysis was used to identify the most influential QSP model parameters affecting key liver biomarkers (ALT, AST, ALP, necrosis, and liver fat percentage). Latin Hypercube Sampling was then applied to generate a diverse set of candidate canine VPs by ensuring uniform coverage of biologically plausible parameter values. Each candidate VP was tested for biological realism by simulating baseline biomarker levels at steady state and rejecting any VP outside predefined reference ranges. Plausible VPs were then simulated under documented drug regimens known to induce specific DILI phenotypes: Acetaminophen (IV, 5g) for acute liver failure, Bosentan (IV, 400mg) for cholestasis, and a chronic Amiodarone treatment (oral, 100mg q12h for 7 days, then 50mg q24h for 1 year) for steatosis, with the goal of reproducing real-world incidence rates. Given the rarity of steatosis, Bayesian Optimization was employed as a machine learning method to efficiently generate high-risk VPs. The workflow was parallelized and deployed on cloud infrastructure to enable fast execution of the PBPK-QSP simulations, rapid virtual patient generation, and scalable analysis. Results: The virtual dog population showed substantial variability in liver biomarkers at baseline and in treatment response, reflecting real-world diversity in DILI susceptibility. Baseline values (IU/L) included ALT: median 27, IQR 15–52 (ref. 12–118); AST: median 20, IQR 17–33 (ref. 15–66); ALP: median 51, IQR 18–93 (ref. 5–131). Simulated drug exposure reproduced observed inter-individual differences in DILI outcomes, with acetaminophen-induced necrosis severity, bosentan-induced cholestasis, and amiodarone-induced steatosis aligning with published data. Conclusions: Our DILI virtual population pipeline is part of an open-source initiative to be integrated into the CRACK IT Virtual Dog Suite using R and the OSP Suite. It integrates PBPK and QSP modeling, machine-learning enriched parameter sampling and cloud resources, enabling efficient and robust DILI risk assessment in dog that accounts for inter-patient variability, idiosyncratic reactions and dose-effect responses while maintaining full transparency and reproducibility.