AI-Assisted Variant Interpretation

AI-Assisted Variant Interpretation is a tool from Strand Life Sciences designed to accelerate germline variant interpretation workflows using large language models (LLMs). It targets clinical genomics teams seeking to reduce the time and manual effort involved in processing clinical notes, assigning HPO terms, and applying ACMG classification criteria.

The tool addresses two core stages of the variant interpretation process: converting unstructured clinical notes into structured phenotype data, and automating ACMG variant labelling for cases that require literature review.

Clinical Notes to HPO Conversion

• Clinical or doctor's notes arriving in image form are processed using Cloud Vision to extract text.
• Phenotypes are extracted from the text using a hybrid approach combining fine-tuned and zero-shot LLMs.
• Extracted phenotypes are converted to HPO terms by performing semantic search against a database of vectorised HPO terms.
• The publicly available HPO-to-gene database is then used to identify genes associated with the identified phenotypes.
• The workflow achieves a best recall of 94.7% for identifying genes associated with the underlying HPOs, compared to a best human performance of 96%.

Automated ACMG Labelling

• LLMs are used to accelerate the assignment of ACMG labels that require literature review, specifically for functional studies and segregational studies labels.
• The labelling workflow uses an LLM prompted with ACMG criteria and supplemented with gene- and variant-specific information retrieved from a literature search.
• The LLM-enabled ACMG labelling process delivers an average case time saving of approximately 5x.
• Overall, the workflow increases throughput from approximately 25 variants reviewed per person per day to approximately 40 variants reviewed per person per day, representing a roughly 1.5x improvement in analysis speed.