Incorporating LLM-Derived Information into Hypothesis Testing for Genomics Applications.

Area of Science:

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

Background:

• Statistical hypothesis testing is crucial in genomics research.
• Large language models (LLMs) offer novel ways to extract biological information.
• Integrating LLM insights into genomics analysis remains an underexplored area.

Purpose of the Study:

• To develop strategies for incorporating LLM information into genomics hypothesis tests.
• To leverage gene embeddings from LLMs to improve statistical power in genomics studies.
• To introduce a novel frequentist and Bayesian (FAB) framework for hypothesis testing.

Main Methods:

• Generated gene embeddings using OpenAI's GPT-3.5 model from text inputs.
• Analyzed the principal subspace of gene embeddings to identify biological signals.
• Developed three hypothesis tests within a frequentist and Bayesian (FAB) framework, guided by LLM embeddings.

Main Results:

• Biological signals in genomics datasets were found to reside near the principal subspace of LLM-derived gene embeddings.
• The proposed FAB hypothesis tests demonstrated increased statistical power compared to classical methods.
• Successful application of the FAB tests in three distinct real-world genomics datasets.

Conclusions:

• LLM-derived gene embeddings provide valuable prior information for genomics hypothesis testing.
• The FAB framework effectively integrates LLM insights to enhance statistical power in genomics.
• This approach offers a promising direction for advancing statistical methods in genomic data analysis.