Objective:

To evaluate the effectiveness of the aiDIVA hybrid artificial intelligence system in ranking disease-causing variants for rare disease cases.

Approach:

• System Development: aiDIVA combines evidence-based scoring, machine learning, phenotype matching, inheritance information, and large language models to generate a ranked list for clinical assessment.
• Training Data: The model was trained using over 155,000 variants from ClinVar, including approximately 90,000 pathogenic and 65,000 benign variants.
• Evaluation Methodology: The system was benchmarked against 3,041 rare disease cases previously solved by genetics specialists, with separate models for dominant and recessive inheritance.

Key Findings:

• aiDIVA placed the causal variant within the top three in 97% of cases and at rank one in 88% among cases with variants in ClinVar or the Human Gene Mutation Database.
• In an independent group of 1,014 cases, aiDIVA placed 93% of causal variants within the top three and 96% within the top 10.
• aiDIVA identified 45 previously unsolved cases as newly solved after expert review.

Interpretation:

aiDIVA is a decision-support tool that aids in variant triage and periodic reanalysis, but final classification requires expert review.

Limitations:

• Language-model responses were inconsistent in a small proportion of tests and could cite nonexistent or irrelevant literature.
• Sending phenotype data to cloud-based models raises privacy considerations.
• aiDIVA does not currently support structural variants and has lower performance for in-frame insertions and deletions compared to single-nucleotide variants.

Conclusion:

aiDIVA demonstrates high accuracy in ranking disease-causing variants, supporting its potential use in clinical settings.

Sources:

• npj Genomic Medicine