Researchers from the University of California San Diego School of Medicine have shown that large language models (LLMs) like GPT-4 could assist in automating functional genomics research, which aims to understand the roles of genes and their interactions. One of the most commonly used methods in this field is gene set enrichment, which involves comparing experimentally identified gene sets to existing genomics databases to uncover their functions. However, many novel biological insights often lie outside the limitations of current databases. By leveraging artificial intelligence (AI) to evaluate gene sets, scientists can potentially save countless hours of meticulous work and move closer to automating one of the key techniques for deciphering how genes collaborate to impact biological processes.
In their analysis of five different LLMs, the researchers determined that GPT-4 performed the best, achieving a 73% accuracy rate in recognizing common functions of curated gene sets from a well-known genomics database. Notably, when tasked with analyzing random gene sets, GPT-4 declined to assign labels in 87% of cases, showcasing its ability to assess gene sets while minimizing inaccurate information. Additionally, GPT-4 provided comprehensive explanations to justify its naming decisions.
Although more research is necessary to fully harness the capabilities of LLMs for automating functional genomics, this study underscores the importance of ongoing investment in LLM development and their uses in genomics and precision medicine. To facilitate this, the researchers established a web portal aimed at helping other scientists integrate LLMs into their functional genomics projects. More broadly, these findings illustrate the potential of AI to transform scientific research by amalgamating complex information to produce new, testable hypotheses much more rapidly.
The study, published in Nature Methods, was directed by Trey Ideker, Ph.D., a professor at UC San Diego School of Medicine and UC San Diego Jacobs School of Engineering, along with Dexter Pratt, Ph.D., a software architect in Ideker’s team, and Clara Hu, a doctoral candidate in biomedical sciences within Ideker’s group. The research received partial funding from the National Institutes of Health.
