GDC Cohort Copilot: an AI copilot for curating cohorts from the genomic data commons

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The Genomic Data Commons (GDC) offers harmonized cancer genomics data via a unified platform.
• Current GDC cohort creation tools can be challenging for users, especially new ones, due to numerous descriptors.
• Natural language offers a more intuitive way for users to describe desired cohorts.

Purpose of the Study:

• To introduce GDC Cohort Copilot, an open-source tool for curating GDC cohorts using natural language.
• To develop and evaluate large language models (LLMs) for generating GDC cohort filters.
• To provide an interactive interface for refining automatically generated cohorts.

Main Methods:

• Developed GDC Cohort Copilot, an open-source tool utilizing LLMs to translate natural language cohort descriptions into GDC filters.
• Evaluated multiple LLMs, including a locally-served GDC Cohort LLM and GPT-4o.
• Implemented the tool as a containerized Gradio app for accessibility.

Main Results:

• GDC Cohort Copilot successfully generates GDC cohort filters from natural language descriptions.
• The locally-served, open-source GDC Cohort LLM outperformed GPT-4o in generating accurate GDC cohorts.
• The tool provides an interactive interface for further cohort refinement.

Conclusions:

• GDC Cohort Copilot enhances the usability of the GDC for cancer genomics data exploration.
• Open-source LLMs can be effectively tailored for specialized bioinformatics tasks.
• The tool facilitates easier cohort curation and analysis within the GDC ecosystem.