AI-enhanced clinical trial-in-a-dish platform for improved DILI risk classification and mechanistic insights into hepatotoxicity

Drug-induced liver injury (DILI) remains a leading cause of clinical trial attrition and post-marketing drug withdrawals. Its prediction is hindered by the limited physiological relevance and interindividual variability captured in conventional preclinical models. To overcome this, we developed a human serum-derived educated spheroid system incorporating human blood sera from donors to generate liver spheroids that recapitulate human hepatic diversity. This platform enables clinical trial-in-a-dish studies and supports acute and chronic treatment regimens. Using a panel of drugs spanning the full DILI risk spectrum, we evaluated hepatotoxic potential through a proprietary AI-driven algorithm that integrates severity and incidence metrics at therapeutic concentrations. Our platform reliably distinguished low-risk from high-risk DILI compounds and recapitulated both dose-dependent and idiosyncratic toxicity profiles. Notably, ximelagatran-induced DILI was only detected under chronic exposure conditions, mirroring clinical outcomes. Transcriptomic profiling revealed innate immune activation in DILI-positive individuals. STRING analysis further implicated HLA-DRB1 and HLA-DQA1 interactions via VIM upregulation in macrophages and dendritic cells, suggesting a mechanistic link to immune-mediated iDILI. In exploratory prospective studies, our system predicted ribociclib-induced grade 3 DILI in one ER+/HER2− breast cancer patient and absence of DILI in two patients, consistent with clinical outcomes. These findings highlight the value of integrating our model with our AI-based mapping strategy to enable mechanistic classification of DILI, deconvolution of immune-related toxicity, and prediction of patient-specific risk. Our platform represents a step toward personalized hepatotoxicity assessment and improved translational toxicology strategies.