Gemini 2.5 Pro Tested in Multi-Agent Drug Safety Causality Framework

A team led by researchers including Maurizio Sessa posted a preprint on October 23, 2026 describing MARCA, Multi-Agent Retrieval-Augmented Causality Assessment, a system built to help pharmacovigilance teams keep pace with rising volumes of individual case safety reports. Rather than a single model producing a single answer, MARCA splits the Naranjo causality scale into four agent roles: one retrieves relevant scientific literature, one checks the drug’s own label, one classifies objective evidence, and one performs the clinical reasoning step that combines the other three into a final judgment. Gemini 2.5 Pro and DeepSeek-V3.2 both ran inside the pipeline, with the agents’ rationales feeding into a human-assisted scoring process rather than replacing the reviewer outright.

The authors tested the framework against a reference standard built by consensus among human experts, covering 215 individual case safety reports and 1,562 drug-event combinations pulled from the FDA Adverse Event Reporting System for the first two quarters of 2025. Agreement varied sharply by question: question-level agreement reached 96.8 percent on two of the ten Naranjo items but fell to 85.4 percent on the question asking whether alternative causes could explain the adverse event, the one item the authors flag as the hardest for the system to match human judgment on. Final causality classification, which rolls all ten questions into one category, landed at 79.4 percent agreement with a Cohen’s kappa of 0.587.

A smaller retrospective comparison found that labelling scores with retrieval-augmented generation rather than without it raised agreement sharply on two of the ten questions, though the authors describe that comparison as exploratory rather than evidence of prospective effectiveness.

The numbers

• 215 individual case safety reports and 1,562 drug-event combinations made up the consensus reference standard.
• Final causality classification agreement between MARCA and human experts was 79.4 percent.
• Gemini 2.5 Pro, one of the two models driving the agent pipeline, is listed on the Google API price table at $1.25 per million input tokens for prompts of 200,000 tokens or less.

The paper does not report what running the four-agent pipeline across 1,562 drug-event combinations cost in API spend, which leaves a gap for any safety team trying to budget a deployment against the published per-token rate.