Scholarlink

<p>Pharmacogenomic guidelines are essential for tailoring drug therapy to individual genetic profiles, but current curation workflows are manual, resource‑intensive, time‑bound, and limited in coverage. We introduce an agentic AI system for automated, scalable generation of CPIC-style recommendations using large language models (LLMs) guided by structured evidence. Our modular pipeline retrieves and processes full-text biomedical literature and FDA drug labels, extracts clinically relevant entities with high accuracy (91.9% across 22 articles), aggregates findings across studies, and generates phenotype-specific dosing recommendations for gene–drug pairs. In expert evaluations of 24 random recommendations, our system significantly outperformed leading LLM baselines (GPT-5, Claude, Grok) in clinical clarity and guideline concordance. These results demonstrate the feasibility of using evidence-grounded, agentic AI for end-to-end pharmacogenomic evidence synthesis, offering a path toward broader population coverage, faster updates, and more consistent and explainable decision support.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An agentic AI system for automated pharmacogenomic recommendation generation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月15日

查看全文
  • Mike Zack,
  • Anton Savinkov,
  • Danil Stupichev,
  • Alex Moore,
  • David Sokolov,
  • Igor Trifonov,
  • Anastasia Yankovskiy,
  • Kirill Reshetnikov,
  • Nurkyz Ydyrysova,
  • Allan Gobbs

摘要

Pharmacogenomic guidelines are essential for tailoring drug therapy to individual genetic profiles, but current curation workflows are manual, resource‑intensive, time‑bound, and limited in coverage. We introduce an agentic AI system for automated, scalable generation of CPIC-style recommendations using large language models (LLMs) guided by structured evidence. Our modular pipeline retrieves and processes full-text biomedical literature and FDA drug labels, extracts clinically relevant entities with high accuracy (91.9% across 22 articles), aggregates findings across studies, and generates phenotype-specific dosing recommendations for gene–drug pairs. In expert evaluations of 24 random recommendations, our system significantly outperformed leading LLM baselines (GPT-5, Claude, Grok) in clinical clarity and guideline concordance. These results demonstrate the feasibility of using evidence-grounded, agentic AI for end-to-end pharmacogenomic evidence synthesis, offering a path toward broader population coverage, faster updates, and more consistent and explainable decision support.