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

AI-CURA — near-fully automated LLM workflow (DeepSeek-R1) combining bioinformatic-tool assessment of non-literature criteria with LLM-based assessment of literature evidence, following ACMG/AMP rules and ClinGen recommendations

AI-CURA is an LLM framework that nearly fully automates genetic variant classification per ACMG/AMP and ClinGen recommendations, separating assessment of non-literature criteria (standard bioinformatic tools) from assessment of literature-based evidence (handled by the LLM). Among the tested models, DeepSeek-R1 outperformed o3-mini-high and achieved high sensitivity and 100% specificity on ACMG rules requiring understanding of literature-based evidence. On 150 variants curated by ClinGen experts, the model showed high concordance with human curators for the final diagnosis. AI-CURA can also be used for reanalysis, demonstrated on 150 ClinVar variants with conflicting interpretations.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Automating variant interpretation via LLM with expert-level concordance addresses one of the major bottlenecks in diagnostic genomics, especially in an exome/genome perspective where the volume of variants to classify is exploding. The explicit separation of literature-based and bioinformatic evidence, and the 100% specificity on literature rules, are reassuring for supervised use. Prospective routine validation on unselected cases is still needed before clinical deployment.