Integration of Machine Learning-Based Pathogenicity Prediction and Phenotype Matching Improves Variant Prioritization...

Area of Science:

• Genomics
• Bioinformatics
• Medical Genetics

Background:

• Rare hereditary diseases are diagnosed using genome and exome sequencing, but manual variant interpretation is a major bottleneck.
• Thousands of genetic variants are identified per patient, requiring extensive manual review to pinpoint the causal one.

Purpose of the Study:

• To develop DiagAI, a machine-learning system designed to reduce the manual interpretation bottleneck in clinical genomics.
• To create a system that ranks genetic variants and provides a short list of likely causal candidates for rare hereditary diseases.

Main Methods:

• DiagAI integrates pathogenicity scores (Universal Pathogenicity Predictor - UP2), phenotype matching (PhenoGenius using Human Phenotype Ontology - HPO terms), and expert rules.
• The system was evaluated on 966 exomes from adults with undiagnosed kidney disease, including 196 with confirmed genetic diagnoses.
• Performance was compared against existing tools like REVEL, Exomiser, and AI-MARRVEL.

Main Results:

• DiagAI's UP2 component outperformed REVEL in ranking disease-causing missense variants.
• The full DiagAI system identified the causal variant in over 90% of diagnosed cases, with a typical shortlist of ~10 variants.
• When patient symptoms were included, DiagAI's top-ranked variant was correct in 74% of cases, surpassing other tools.

Conclusions:

• DiagAI effectively reduces the manual interpretation workload in clinical genomics by generating accurate, compact variant shortlists.
• The system demonstrates significant potential to improve diagnostic efficiency as sequencing volumes increase.
• Incorporating patient phenotype data further enhances DiagAI's diagnostic accuracy.