Skip to main content
Skip to main content
Back to Grand Rounds
Grand RoundsWeekly Evidence Brief

Genetics

Edition

30-Second Takeaway

  • High herbal remedy use can interact with pharmacogenomic enzymes in British South Asian patients.
  • Fine-tuned interpretable LLMs (RareDAI) can meaningfully improve test-selection accuracy versus base models.

Week ending May 23, 2026

Genetics, testing, and translation: community uptake, AI decision support, trial rigor, metabolic trait links, and ancestry in preclinical models

High herbal remedy use and CYP2C9 relevance in British South Asian adults considering PGx

BMC MEDICINEMay 16, 2026

In a survey of 553 British South Asian adults of Bangladeshi or Pakistani ancestry, 66% used herbal remedies and 72% reported medication inefficacy. Black seed, turmeric, and ginger were the most used herbs and each inhibits CYP2C9, a common pharmacogenomic enzyme. Fifty-eight percent were willing to provide DNA for PGx testing, but most demanded stronger data protections and had privacy concerns. Not using traditional medicines correlated with higher medication adherence (MARS-5), suggesting cultural practices affect adherence and PGx uptake.

RareDAI: interpretable LLM fine-tuning improves genetic test selection for rare disease

NPJ DIGITAL MEDICINEMay 20, 2026

RareDAI fine-tunes Llama 3.1 and Qwen 3 with chain-of-thought outputs to mimic clinician reasoning for test selection. Self-distillation fine-tuning improved performance by 10–20% across accuracy, precision, recall, and F1 on internal and external sets. The model produces interpretable reasoning before recommendations, aiding choices between gene panels and WES/WGS across systems. Clinical adoption requires local validation and oversight despite improved metrics.

Most RCTs labeled 'personalized' show low transparency and high bias

JOURNAL OF CLINICAL EPIDEMIOLOGYMay 18, 2026

This survey of 262 RCTs (2020–2022) found heterogeneous interventions labeled as personalized, individualized, or precision. Of 221 trials comparing personalized versus non-personalized controls, 70.6% reported favorable conclusions. Data sharing was rare (5% shared data; 0.4% shared code), and 68.6% of trials had high overall risk of bias. Labels were applied largely interchangeably and genomic features were uncommon among these trials.

References

Numbered in order of appearance. Click any reference to view details.

Additional Reads

Optional additional studies from this edition.

Edition context

Clinical signal

  • Ask about herbal/traditional medicine use and consider CYP2C9 interactions when interpreting PGx results.
  • Use AI test-selection tools only with local validation and clinician oversight.
  • Be cautious interpreting trials labeled 'personalized' given frequent high risk of bias.