Integration of Machine Learning-Based Pathogenicity Prediction and Phenotype Matching Improves Variant Prioritization in Rare Clinical Testing.

Genome and exome sequencing have become central to diagnosing rare hereditary diseases, but each test returns thousands of variants that a clinical scientist must review by hand to find the one responsible for the patient's condition. This manual interpretation is the main bottleneck in clinical genomics. To reduce it, we developed DiagAI, a machine-learning system that ranks the variants found in a patient and returns a short list of the most likely causal candidates. DiagAI combines three sources of evidence: a pathogenicity score from the Universal Pathogenicity Predictor (UP2), a model we trained to estimate how damaging a variant is on the five-tier scale of the American College of Medical Genetics and Genomics (ACMG); a phenotype-matching score from PhenoGenius, which weighs how well a gene's known clinical features match the patient's symptoms (encoded as Human Phenotype Ontology, or HPO, terms); and expert rules covering inheritance pattern and sequencing quality. We evaluated DiagAI on 966 exomes from adults investigated for kidney disease of unknown cause, of which 196 had a confirmed genetic diagnosis. We first tested UP2 on its own by ranking 62 confirmed disease-causing missense variants that were absent from its training data: UP2 placed the causal variant within the top 100 candidates in 87% of cases, compared with 61% for the widely used tool REVEL. Across the 196 diagnosed exomes, the full DiagAI shortlist contained the causal variant in 94.9% of cases when the patient's symptoms were provided and in 90.8% when they were not, with a typical shortlist of about 10 variants. When symptoms were provided, the single top-ranked variant was the correct diagnosis in 74% of cases, versus 42% without symptoms, exceeding the performance of the established tools Exomiser and AI-MARRVEL on the same cohort. DiagAI produces compact, accurate shortlists that can reduce the manual interpretation workload as diagnostic sequencing volumes continue to grow.