Benchmarking large language models for replication of guideline-based PGx recommendations

PGx testing can markedly improve treatment safety and efficacy, yet adoption in health systems remains slow. Our latest peer-reviewed study-now published in The Pharmacogenomics Journal (Nature Portfolio)-benchmarks how well large language models (LLMs) reproduce guideline-based pharmacogenomic (PGx) recommendations and outlines a practical path from static, one-time advice to dynamic guidance updated in real time.
Nature

Key findings

General-purpose models often produced incomplete or unsafe outputs when asked for guideline-conformant PGx recommendations.

Domain adaptation (including targeted fine-tuning and structured prompting) substantially improves accuracy and reduces latency, making PGx-specific LLMs more suitable for clinical decision support.

Why this matters
LLMs purpose-built for PGx can help close the gap between genetic results and bedside actionability-supporting safer, faster, and more consistent medication decisions across diverse patient populations.

Looking ahead
We have already released a closed model, Deneb, capable of generating recommendations in real time. Comparative testing against the latest versions (o3, o3-pro, grok-4, claude-4, etc.) will be shared soon.