Evaluating Generative AI's Ability to Identify Cancer Subtypes in Publicly Available Structured Genetic Datasets

Area of Science:

• Genomics and Bioinformatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Genetic data integration into clinical care is crucial for disease diagnosis and treatment.
• Barriers to genetic data integration include electronic health record (EHR) structure, testing costs, and result interpretability.
• Artificial intelligence (AI) offers potential solutions, particularly large language models (LLMs), for analyzing genetic information.

Purpose of the Study:

• To reevaluate the capabilities of LLMs, specifically GPT models, in supervised prediction tasks using structured gene expression data.
• To assess the effectiveness of LLMs in analyzing real-world genetic data for cancer research.
• To compare LLM performance against traditional machine learning approaches for genetic data analysis.

Main Methods:

• Utilized GPT models for supervised prediction tasks on structured gene expression datasets.
• Employed traditional machine learning methods as a benchmark for comparison.
• Focused on predicting cancer subtypes as a key application.

Main Results:

• GPT models demonstrated effectiveness in supervised prediction tasks using gene expression data.
• AI models, including LLMs, show potential for analyzing real-world genetic data.
• The study provides evidence for AI's utility in generating real-world evidence from genomic data.

Conclusions:

• LLMs can be effectively applied to structured genetic data for predictive tasks in cancer research.
• AI-driven analysis of genetic data holds significant promise for clinical applications and real-world evidence generation.
• Further development of LLMs trained on comprehensive genetic datasets could enhance their utility in genomics.