Food and Drug Administration reviews model for predicting drug related liver injury

1. The Food and Drug Administration is reviewing an artificial intelligence based tool that could help predict drug induced liver injury earlier in therapeutic development.
2. The review highlights how computational safety models are beginning to move upstream, before drugs reach broader human exposure.

The Food and Drug Administration (FDA) is reviewing an artificial intelligence based drug development tool designed to predict drug induced liver injury, a persistent source of clinical trial failure and post approval concern. The agency’s Center for Drug Evaluation and Research accepted a letter of intent for the tool, according to Reuters reporting published on June 3, 2026. The tool is intended to support earlier safety assessment by identifying compounds with a higher likelihood of hepatotoxicity before they advance deeper into development. Drug induced liver injury remains difficult to anticipate because preclinical toxicology models do not always capture human hepatic metabolism, immune mediated injury, mitochondrial dysfunction, or dose dependent risk. That challenge is familiar to clinicians who have seen promising therapies limited by liver enzyme elevations, black box warnings, or unexpected safety signals after wider use. A candidate can look effective in early testing and still develop serious hepatic concerns once exposure expands across more diverse patients. The appeal of this type of model is that it may add another layer of signal detection before a compound reaches larger human trials. If validated, earlier prediction could reduce failed studies, limit patient exposure to unsafe candidates, and help sponsors prioritize compounds with more favorable safety profiles. Still, this should not be interpreted as a replacement for toxicologists, pharmacologists, hepatologists, or clinical safety committees. Hepatic injury is biologically heterogeneous, and the performance of any model will need to hold up across mechanisms of injury, therapeutic classes, dosing patterns, comorbid disease, and polypharmacy. A model that performs well in a curated dataset may not generalize cleanly across real development programs. The most clinically useful version would be one that complements traditional toxicology rather than bypassing it. This review also reflects a broader regulatory shift toward evaluating artificial intelligence as part of the scientific infrastructure behind drug development. The signal for physicians is that artificial intelligence is no longer limited to bedside documentation, imaging, or triage. It is beginning to shape the earlier decisions that determine which therapies reach patients at all.