Core Concepts
Large language models can effectively synthesize the common biological functions represented by gene sets, providing insights beyond traditional functional enrichment analysis.
Abstract
The study evaluates the ability of five large language models (LLMs) - GPT-4, GPT-3.5, Gemini Pro, Mixtral Instruct, and Llama2 70b - to analyze gene sets and propose concise names describing their common biological functions.
Evaluation Task 1: The authors benchmarked the LLMs against gene sets derived from the curated Gene Ontology (GO) database. They found that GPT-4 was able to propose names highly similar to the GO-assigned names in 73% of cases, often capturing a more general concept. The other LLMs showed varying degrees of performance, with Llama2 70b performing the worst.
Evaluation Task 2: The authors then explored the LLMs' ability to analyze gene sets derived from 'omics data, such as transcriptomics, proteomics, and CRISPR screens. They found that in 32% of cases, GPT-4 was able to identify novel functions not reported by classical functional enrichment analysis. Independent review indicated that these novel insights were largely verifiable and not hallucinations.
The study highlights the potential of LLMs as valuable assistants in functional genomics, able to rapidly synthesize common gene functions based on their broad biomedical knowledge. While LLM outputs require careful validation, the authors conclude that these models can provide researchers with a new and powerful tool for gene set interpretation.
Stats
"GPT-4 confidently recovered the curated name or a more general concept in 73% of cases when benchmarked against canonical Gene Ontology gene sets."
"In 32% of cases, GPT-4 identified novel functions for gene sets derived from 'omics data that were not reported by classical functional enrichment analysis."
Quotes
"The ability to rapidly synthesize common gene functions positions LLMs as valuable 'omics assistants."
"Notably, we found that even a very lenient overlap requirement (JI ≥10%) left 87% of gene sets lacking annotation by GO terms."
"Of these non-enriched gene sets, 37% had been confidently processed by GPT-4, yielding a novel functional name synthesized from outside of the GO corpus."