Abstract / Summary
Drug-induced liver injury (DILI) remains a major challenge in drug development because existing in vitro models poorly predict clinical hepatotoxicity. We coupled patient-derived human liver organoids (HLOs) with high-content imaging and automated machine learning (ML) to detect hepatotoxicity-associated phenotypes by morphological profiling. HLOs from six DILI patients and one healthy donor (n = 7) retained donor-specific baseline phenotypes in hepatocyte maturity, metabolism, and morphology. ML models trained per line on MAD-robustized, well-level-aggregated median features achieved compound-level AUROCs of 0.88 (95% CI 0.86-0.90) to 0.91 (95% CI 0.89-0.93) and balanced accuracies of 0.81-0.84 in distinguishing 54 hepatotoxicants from 20 non-hepatotoxic controls. Averaging predictions across all seven lines at an optimized threshold gave 94.4% sensitivity (95% CI 84.6-98.8%) and 85.0% specificity (95% CI 62.1-96.8%), substantially outperforming HLO cytotoxicity assays and cancer cell line profiling. Lines were moderately-to-highly concordant (Pearson r = 0.66-0.85), and benchmark concentration modeling revealed reproducible inter-line differences in phenotypic sensitivity (median 4.0-fold, several compounds exceeding 20-fold). Exposure anchoring of benchmark concentrations to human-equivalent plasma concentrations, using in vitro mass-balance dosimetry and curated plasma free fractions, placed organoid points of departure within 100-fold of clinical Cmax for 60 of 66 compounds (23-fold residual offset, 95% CI 12-45), and placed two of the three misclassified negative controls an order of magnitude above clinical Cmax. Sulindac injury signatures were significantly enhanced by TNFα co-treatment, demonstrating capacity to detect pro-inflammatory-mediated DILI. This platform preserves inter-line variability while enabling reproducible, scalable multi-line consensus hazard classification; because compounds were tested over a common nominal concentration range, the readout defines hepatocellular hazard, which must be combined with exposure as an independent input to estimate clinical DILI risk.