Advances in artificial intelligence (AI) are beginning to reshape functional genomics, a field focused on understanding gene functions and interactions. A recent study from the University of California San Diego School of Medicine demonstrates how large language models (LLMs), including GPT-4, could automate key processes, potentially streamlining the analysis of complex genetic datasets.
Challenges in Functional Genomics
Functional genomics relies heavily on gene set enrichment analysis, a method for identifying the biological significance of gene sets by comparing them to reference databases. While these databases are extensive, they are often limited in scope, leaving researchers to manually interpret more novel or complex gene sets. Automating this process with AI could significantly reduce labor-intensive tasks, freeing scientists to focus on higher-level research questions.
Evaluating Large Language Models
To assess the utility of LLMs in functional genomics, the research team tested five models, finding GPT-4 to be the most effective. The model achieved a 73% accuracy rate when tasked with identifying common functions for curated gene sets from a standard genomics database. According to the researchers, GPT-4 also exhibited a cautious approach when confronted with random gene sets, declining to provide a name in 87% of cases—an indicator of its ability to minimize erroneous or “hallucinatory” results.
In addition to accuracy, GPT-4 demonstrated a capacity to generate detailed explanations supporting its analyses. This ability to articulate reasoning offers researchers valuable insights into the AI's decision-making process, fostering greater confidence in its utility.
Applications and Future Directions
While the study represents a significant step forward, the researchers emphasize the need for further refinement and testing before LLMs can be fully integrated into functional genomics workflows. To facilitate broader adoption, they developed a web portal designed to help other scientists incorporate AI tools into their research.
The findings highlight the potential of AI in genomics. Beyond functional genomics, LLMs could contribute to hypothesis generation by synthesizing complex information at unprecedented speeds, opening doors to more efficient experimental design.
Balancing Optimism with Caution
Despite the promising results, the researchers caution that AI models like GPT-4 are not a replacement for human expertise. Instead, these tools are envisioned as collaborators—augmenting the capabilities of researchers by automating routine tasks and enabling a deeper focus on innovative questions. As AI continues to develop, its integration into functional genomics may redefine the pace and scope of discovery in the field. However, ensuring accuracy, minimizing biases, and fostering transparency will remain essential to its success.
Citation:
Hu, M., Alkhairy, S., Lee, I. et al. Evaluation of large language models for discovery of gene set function. Nat Methods (2024). https://doi.org/10.1038/s41592-024-02525-x