A new AI agent classifies genetic disease severity with expert-level accuracy, and it can show its work.
Researchers combined two AI techniques - a reasoning-and-acting loop and retrieval-augmented generation, which lets a model pull in outside documents instead of relying only on what it memorized - to build an autonomous system that grades severity for 10,211 conditions cataloged in the Human Phenotype Ontology. The agent applies severity guidelines from the American College of Medical Genetics and quality-of-life criteria from the American College of Obstetricians and Gynecologists, then searches PubMed for supporting evidence before making a call. Tested against expert-curated cases, it hit 93.55 percent accuracy, with 82.6 to 91.4 percent of its claims backed by direct evidence or valid inference rather than guesswork. Rolled up to the gene level across 8,738 pairs, the system flagged 3,283 recessive gene pairs as severe or profound, and its output matched an existing screening reference, Mackenzie's Mission, 95.2 percent of the time.
Commercial genetic screening panels currently vary widely in what they test for, partly because deciding how severe a condition is has been a slow, subjective, expert-by-expert process. An automated classifier that shows its reasoning and cites sources could let panel designers standardize which conditions get included, instead of leaning on inconsistent individual judgment calls.
The 95.2 percent match with Mackenzie's Mission is reassuring, but agreement with an existing list is the easiest bar a new tool can clear - the harder test is what happens when the agent disagrees with the experts, and who gets believed then.