ClinGeX builds genomic workflows from expert-curated resources
Advancing Clinical Genomic Analysis with a Skill-Centred Agent · arXiv:2610.40556
In plain words
Genomic analysis can require specialists to combine information from databases, research papers, guidelines and patient records. ClinGeX is a language-model agent that uses a curated library of executable skills to build a workflow for each question.
On a test of 426 evidence-based questions, its rubric score was higher than the unenhanced DeepSeek model and five single-strategy additions. This result applies to that test; it does not show that patients had better outcomes.
The research
Clinical genomic analysis depends on evidence spread across databases, literature, guidelines and patient information. Because each case can require different steps and expert judgment, can an agent organize selected resources into a workflow that supports interpretation?
ClinGeX selects and combines executable skills from curated genomic resources so a language-model agent can build a workflow for each analysis question.
Can a system help specialists combine fragmented genomic evidence while adapting its steps to each case? ClinGeX addresses this with 105 skills based on selected guidelines, databases, tools, calculations and prediction models; it selects relevant skills and prerequisites to build a task-specific workflow.
In a separate controlled study, professionals’ conclusions met expert-panel criteria more often with ClinGeX assistance than with routine or DeepSeek assistance in each of three scenarios: variant analysis, rare-disease interpretation and molecular-oncology decisions. These observed comparisons measure agreement with expert references, not patient outcomes.
The collaboration study covered three scenarios, and agreement with expert references is a proxy for interpretive quality, not a patient outcome. The exploratory routine-use assessment had no control condition and did not test effects on diagnosis, management or health.
What the evidence shows
- A changing evidence problem
Clinical genomic analysis asks specialists to combine variant, disease, inheritance and patient information with evidence from databases, literature and guidelines. The needed steps differ by case, so a fixed sequence of tools may not fit every question.
ClinGeX is a language-model agent designed to assemble a task-specific workflow from a curated library of executable skills.
- Skills shape each workflow
The library contains 105 skills based on selected guidelines, databases, tools, calculations and prediction models. For a request, ClinGeX selects relevant skills, expands the selection with related skills and prerequisites, then gives an agent detailed instructions for the needed steps.
The agent executes that workflow and synthesizes the evidence. The design organizes selected resources; it does not establish that every clinical-genomics resource or task is covered.
- A test of the full design
A controlled comparison tested the complete ClinGeX design and five single-strategy additions against an unenhanced DeepSeek-V4 model on the same 426 evidence-grounded questions. Two independent language-model judges scored responses against question-specific rubrics, and researchers averaged the judges’ scores and then the questions.
ClinGeX scored 83.2%, compared with 72.2% for the unenhanced model and 74.0–78.4% for the single-strategy additions. The result supports the full design on this test, not a universal advantage across clinical work.
- Evidence with professional help
In a separate study, professionals assessed cases under routine, DeepSeek-assisted and ClinGeX-assisted conditions. Expert panels judged whether each conclusion met the scenario’s complete criteria.
ClinGeX-assisted assessments had higher observed agreement in all three tested scenarios. This extends the evidence from system-only tests to professional work on study cases, while measuring agreement with expert references rather than patient outcomes.
- What the evidence supports
Together, the tests suggest that the skill-based agent can support genomic interpretation in the evaluated tasks. The human-AI study covered three scenarios, and the exploratory routine-use feedback had no control group. Neither establishes that ClinGeX improves diagnosis, treatment decisions or health outcomes.
Deep
- Original paper authors
- Qiaoyu Zheng, Weike Zhao, Yanjie Fan, Ye Cao, Mengqi Jiang
- Submitted to arXiv
- License
- arXiv non-exclusive