Interpreting a genetic variant requires crossing recommendations, bases and calculations. An agent who assembles executable skills significantly improves the accuracy and speed of 45 professionals.
Genetic sequencing has become affordable, but its interpretation remains a slow and meticulous expert work, which slows down the use of results in clinical decisions. A team of researchers presents ClinGeX, an agent based on a large language model that does not respond directly: it selects and chains, depending on the task, a part of 105 executable and verified "skills" (interpretation rules, database querying, calculations), while respecting their dependencies.
Measured gains with real professionals
Out of 223 cases subsequent to its design data, the agent meets the expert reference classification in 71.7% of the tasks, compared to 45.7% for the best competitor system. The authors then had 45 clinical genomics professionals work on 135 real-world cases, unaided, with DeepSeek or with ClinGeX. With the agent, the agreement with the reference reaches 51.1% for variant analysis, 57.8% for rare diseases and 64.4% for molecular oncology decisions, compared to 15.6%, 31.1% and 20% respectively under the other conditions, and the reported interpretation time decreases in all scenarios.
A month of exploratory testing in real conditions, without intervention on care, completed the study: eleven specialists evaluated 96 analyses and say that they often adopted the agent's conclusions while saving time. The authors see this as an argument for building medical assistants on executable business bricks rather than on model knowledge alone.
Read the full article on the original source
Go to the source →