DILI-context: a dose- and exposure-enriched knowledge base for translational liver safety assessment

Drug-induced liver injury (DILI) is a major cause of drug development failure and post-marketing withdrawal. Although assessing clinical hepatotoxic risk is critical for decision-making, most benchmark datasets represent DILI with categorical hazard labels rather than exposure-dependent risk measures. As a result, the current understanding of DILI risk focuses on hazard rather than context-dependent exposure risk. To address this limitation, we present DILI-Context, a dose- and exposure-enriched knowledge base that integrates hazard annotations with high-content regulatory and toxicological data sources. This framework incorporates therapeutic dose, treatment duration, and preclinical safety thresholds to contextualize hepatotoxic liability. Using this harmonized resource, we observe a consistent monotonic relationship between therapeutic dose and DILI severity, with higher daily dose requirements associated with greater hepatotoxic concern. Our results show that drugs with elevated DILI risk operate within significantly narrower therapeutic windows, demonstrated by systematic declines in both acute Therapeutic Index and chronic Margin of Safety across DILI categories. We also demonstrate that chronic no-effect thresholds (NOAEL-based metrics) more effectively discriminate DILI risk than lowest-effect endpoints (LOAEL). To further capture cumulative burden, we introduce Chronic Load Score that integrates dose magnitude and treatment duration. This metric significantly stratifies DILI concern levels and aligns with independent regulatory signals, including hepatic dose adjustment requirements in prescribing information. Together, these findings demonstrate that hepatotoxic liability is strongly linked to required systemic exposure and constrained safety margins. We further characterize the chemical-property landscape of included small molecules and provide a minimalist benchmark comparing Morgan fingerprint-only classifiers against classifiers enriched with exposure-derived covariates, demonstrating complementary predictive signals beyond structure alone under the cohort constraints of DILIrank 2.0. DILI-Context thus provides both a reproducible supplementary feature layer and an explicit resource for advancing exposure-aware DILI modeling.