NIH Develops GeneAgent AI to Improve Accuracy of Gene Set Analysis

The project was supported by the National Library of Medicine (NLM), a global leader in biomedical informatics research, and forms part of NIH’s broader mission to advance medical science through data-driven research.

Researchers at the National Institutes of Health (NIH) have developed GeneAgent, an artificial intelligence (AI) system powered by a large language model (LLM), designed to improve the accuracy of gene set analysis by minimizing false or misleading content, commonly known as AI hallucinations.

Unlike traditional LLMs that often rely on their own training data for output verification, GeneAgent takes a novel approach. It independently cross-checks its predictions, or “claims,” against expert-curated biological databases and generates a detailed verification report on each.

This allows for greater reliability in interpreting high-throughput molecular data and identifying critical biological pathways and gene interactions.

“AI-generated content can be false, misleading, or fabricated,” the researchers noted, describing the challenge of hallucinations in gene set analysis. Existing LLMs are prone to circular reasoning and tend to reinforce their own outputs, even when inaccurate.

GeneAgent addresses this limitation by applying a self-verification module. When tested on 1,106 known gene sets, the system first produced a list of biological function claims, then used external databases to independently assess and categorize each claim as supported, partially supported, or refuted.

To validate the AI’s self-verification capabilities, two human experts manually reviewed 10 randomly selected gene sets comprising 132 claims. The experts confirmed that GeneAgent made correct verification decisions in 92% of the cases, outperforming general-purpose LLMs like GPT-4 in accuracy and consistency.

In real-world testing, GeneAgent was also applied to seven novel gene sets derived from mouse melanoma cell lines. The system successfully identified potential new gene functions, suggesting promise for future drug target discovery in diseases such as cancer.

While the researchers acknowledge that LLMs still face limitations in reasoning and data scope, they emphasize GeneAgent’s significant progress in minimizing hallucinations through automated fact-checking.

The project was supported by the National Library of Medicine (NLM), a global leader in biomedical informatics research, and forms part of NIH’s broader mission to advance medical science through data-driven research.

Stay tuned for more such updates on Digital Health News