AI agent combines executable skills for clinical genomic analysis

What the study found

In an arXiv preprint, Qiaoyu Zheng and colleagues describe ClinGeX, a large-language-model agent that retrieves and orchestrates 105 curated executable skills for genomic analysis. On 223 cases held out by time, the authors report 71.7% macro-averaged task-level concordance with expert reference classifications, compared with 45.7% for the strongest comparator.

The team also reports a counterbalanced study in which 45 clinical genomic professionals assessed 135 clinically derived cases under three conditions: routine practice without AI, DeepSeek assistance and ClinGeX assistance. With ClinGeX, raw case-level concordance was 51.1% for criterion-level variant analysis, 57.8% for rare-disease genomic interpretation and 64.4% for molecular oncology decisions. The abstract gives comparator figures of 15.6%, 31.1% and 20.0%, respectively, and says reported interpretation time fell in all scenarios.

In a separate exploratory, non-interventional assessment lasting one month, 11 experienced professionals gave feedback on 96 self-selected analysis episodes. They reported frequent adoption of conclusions and time savings.

Why it matters

Genomic interpretation can be labor-intensive even as sequencing becomes more accessible. The authors suggest that combining executable domain skills could help specialists interpret genomic evidence; concordance with expert references provides one way to assess that potential.

Limits and unknowns

The abstract does not specify which of the two collaboration-study comparator conditions corresponds to each reported comparator figure, or provide detailed timing results, uncertainty estimates or descriptions of the benchmark cases and skills. The feedback episodes were self-selected, and reported adoption and time savings do not establish effects on patient outcomes. The work is a preprint, and the abstract does not establish clinical effectiveness or safety.