Preprint: ClinGeX, a 'skill‑centered' AI, speeds genomic interpretation but lacks peer review or outcomes data

A new arXiv preprint from researchers in China says a “skill-centered” artificial intelligence agent for clinical genomics, called ClinGeX, substantially outperformed comparator systems on a 223-case benchmark and helped specialists reach closer agreement with expert reference answers while cutting interpretation time in a multicenter reader study. But the work has not been peer-reviewed, the source code is not publicly available, and the study did not show better patient outcomes.

That matters because clinical genomic interpretation is one of the most labor-intensive parts of precision medicine. Specialists must weigh formal rules such as the ACMG/AMP guidelines for classifying genetic variants, along with evidence from public databases including ClinVar and gnomAD, scientific literature and computational tools. As genome sequencing becomes more common, that evidence-gathering and synthesis has become a bottleneck.

The paper, “Advancing Clinical Genomic Analysis with a Skill-Centred Agent,” was posted on arXiv on Oct. 29 by researchers affiliated with Shanghai Jiao Tong University, Shanghai AI Laboratory, Xinhua Hospital affiliated to Shanghai Jiao Tong University School of Medicine, The Chinese University of Hong Kong and Shanghai Chest Hospital. The authors describe ClinGeX as a large-language-model agent built not around free-form chatbot responses, but around 105 curated executable skills: 27 guide skills, 40 search skills and 38 analysis skills. Their argument is that selecting and orchestrating tested skills for each case makes the system more traceable and less prone to hallucinated evidence.

In automated testing on 223 temporally held-out benchmark cases, the preprint reports that ClinGeX reached 71.7% macro-averaged task-level concordance with expert reference classifications, compared with 45.7% for the strongest comparator. The comparator systems in that benchmarking included OpenAI DeepResearch and Biomni.

The main human study was a randomized, counterbalanced, multicenter human-AI collaboration study involving 45 clinical-genomics readers and nine adjudicators across 135 retrospectively collected, clinically derived cases from three medical centers. The researchers evaluated three scenarios: criterion-level variant analysis, genomic interpretation in rare disease and molecular oncology decision-making. In criterion-level variant analysis, ClinGeX-assisted readers achieved 51.1% case-level concordance with the expert panel reference, versus 22.2% for clinicians working without AI and 15.6% for readers assisted by DeepSeek. In rare-disease interpretation, concordance was 57.8% with ClinGeX assistance, compared with 28.9% for clinicians alone and 31.1% with DeepSeek assistance. In molecular oncology decision-making, the figures were 64.4% for ClinGeX-assisted readers, 31.1% for clinicians alone and 20.0% for DeepSeek-assisted readers.

The paper also reports faster work, based on self-reported interpretation times. Median time for criterion-level variant analysis fell from 60 minutes without AI to 15 minutes with ClinGeX, the authors said. For rare-disease interpretation, median time fell from 30 minutes to 15 minutes.

The authors also described a one-month, non-interventional real-world assessment meant as an exploratory deployment study rather than a test of clinical benefit. In that exercise, 11 experienced professionals submitted feedback on 96 self-selected analysis episodes. They reported fully or partially adopting ClinGeX’s conclusions in 94.8% of tasks, and said the tool saved more than 15 minutes in 77.3% of difficult tasks. The paper says the Ethics Committee of Xinhua Hospital approved research use of the clinical data and that participants gave informed consent.

Still, the study leaves major open questions. The authors said they tested only three scenarios, so broader generalizability is unproven, and some secondary comparisons were not statistically powered. The exploratory deployment tracked workflow measures such as adoption and time savings, not diagnostic yield, treatment changes or patient outcomes. The work also remains hard to independently reproduce: “Although the source code for ClinGeX is not publicly available at present,” the preprint says, while adding that data are available “upon reasonable request.” For now, the study is best read as an early, unreviewed signal that a skill-based AI approach may help genomic specialists work faster and more consistently — not as established clinical evidence.