A multiparametric in vitro strategy for small molecule drug-induced liver injury risk mitigation in early drug development.

Drug-induced liver injury (DILI) remains a major cause of preclinical and clinical attrition, reflecting persistent gaps in the predictive translation of preclinical safety data to human outcomes. To address these gaps, we established a refined multiparametric framework for early DILI risk prediction using a balanced reference set of 170 drugs with established clinical outcomes. Thirty physicochemical, mechanistic endpoints, and exposure-related features were systematically evaluated to identify high-specificity predictors of intrinsic DILI risk. Cytotoxicity in primary human hepatocytes and human liver microtissues, mitochondrial toxicity (i.e. mitotoxicity), and bile salt export pump inhibition were identified as top predictive features. Incorporation of clinical exposure data to derive margins of safety substantially improved assay performance, emphasizing the importance of exposure normalization. A multiparametric, flag-based hazard framework integrating ≥2 hazard flags achieved 98% specificity and a positive likelihood ratio of 18, providing a pragmatic approach for early compound triage and confidence in early decision-making. Random Forest models further enhanced predictive performance, identifying exposure, predicted lipophilicity (e.g. cLogD), mitotoxicity, cytotoxicity, and hepatic clearance as major contributors to DILI risk, with exposure-informed models achieving a balanced accuracy of 0.72. Application of the framework to contemporary Genentech portfolio molecules demonstrated translational utility in refining risk categorization, particularly for compounds with borderline or 0-1 hazard flags. Collectively, this work establishes a mechanistically informed and machine learning-enabled platform that combines interpretability, specificity, and practicality to guide safer molecule design in drug discovery and reduce hepatotoxicity-related attrition in clinical development.