HADES: An Explainable AI Approach to Drug-Induced Liver Injury Prediction
Researchers propose that predicting drug-induced liver injury (DILI) should be viewed as a problem of generating explainable hypotheses instead of a simple binary classification task. To support this, they have created the DILER Benchmark, which enhances curated molecules with mechanistic hepatotoxicity hypotheses sourced from biomedical literature. Additionally, they introduce HADES, an agentic system that produces clear reasoning traces by integrating predictions at the molecular level, metabolite breakdown, structural insights, and evidence of toxicity pathways. When assessed using DILER, HADES surpasses current models in binary classification while also offering mechanistic perspectives. This research is outlined in a preprint available on arXiv (2605.02669v1).
Key facts
- Drug-induced liver injury (DILI) is a leading cause of late-stage clinical trial attrition.
- Existing computational predictors rely on binary classification, limiting generalization and mechanistic insight.
- DILI prediction is better posed as an explainable hypothesis-generation problem.
- The DILER Benchmark extends beyond binary labels with mechanistic hepatotoxicity hypotheses from literature.
- HADES is an agentic system for transparent and auditable reasoning traces.
- HADES combines molecular-level predictions, metabolite decomposition, structural understanding, and toxicity pathway evidence.
- HADES outperforms existing models in binary classification on the DILER Benchmark.
- The preprint is available on arXiv with ID 2605.02669v1.
Entities
Institutions
- arXiv