AI-CURA, an automated LLM workflow for high-accuracy genetic variant classification

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Large language models (LLMs) show promise in medical applications but are underexplored for rare disease diagnosis in clinical genetics.
• Advancements in LLM reasoning and transparency can enhance clinical workflows.

Purpose of the Study:

• To develop and evaluate AI-CURA, a framework for automated genetic variant classification using LLMs.
• To assess LLM performance in interpreting literature-based evidence for variant classification according to established guidelines.

Main Methods:

• AI-CURA integrates automated non-literature evidence assessment with LLM-supported literature evidence review.
• Two LLMs, DeepSeek-R1 and o3-mini-high, were tested for summarizing literature evidence.
• Prompt engineering and ACMG-rule-specific knowledge bases were utilized for DeepSeek-R1 optimization.

Main Results:

• DeepSeek-R1 outperformed o3-mini-high, achieving high sensitivity and 100% specificity in interpreting literature-based ACMG rules.
• AI-CURA demonstrated high concordance with human experts in classifying 150 variants.
• The framework successfully reanalyzed 150 ClinVar variants with conflicting interpretations.

Conclusions:

• AI-CURA offers a robust LLM-based framework for automated genetic variant classification in rare disease diagnosis.
• The study highlights the potential of LLMs, particularly DeepSeek-R1, to improve efficiency and accuracy in clinical genetics workflows.
• AI-CURA facilitates automated variant reanalysis, addressing challenges with conflicting interpretations.