The GeneAgent system generates a verification report that outlines its accomplishments and shortcomings after verifying the accuracy of its initial predictions against data from reputable, expert-curated databases.

Researchers can gain a better grasp of how various diseases and conditions impact gene groups both individually and collectively by using the AI agent to assess high-throughput molecular data and find pertinent biological pathways or functional modules.

LLMs trained on vast volumes of text data from the internet create AI-generated content. These data are used by LLMs to identify trends and forecast which words might appear next in a phrase.

However, as LLMs are not made to check for accuracy, content produced by AI may be faked, deceptive, or false—a phenomena known as AI hallucinations. Circular reasoning, or fact-checking their generated results against their own data, is another tendency of LLMs that makes them appear more confident in the output, even when the information is incorrect.

When employing LLM tools for gene set analysis, which is the process of creating collective functional descriptions of grouped genes and their possible connections, it’s critical to avoid AI hallucinations. Hallucinations in the generated information were not specifically addressed in earlier studies that trained LLMs to summarize biological processes in a given gene set or provide answers to genomic inquiries.

By independently comparing its own claims to proven knowledge gathered in external, expert-curated databases, GeneAgent lessens this problem. Using 1,106 gene sets from pre-existing datasets with established functions and process names, the research team initially evaluated GeneAgent.

GeneAgent first produced a preliminary list of functional claims for every gene set. After that, it independently cross-checked these assertions against the curated databases using its self-verification agent module, producing a verification report that indicated whether each claim was denied, partially supported, or supported.

Two human specialists were then called in by the researchers to manually analyze ten randomly chosen gene sets totaling 132 claims in order to assess whether GeneAgent’s self-verification reports were accurate, somewhat accurate, or inaccurate. This was done in order to determine the accuracy of the self-verification process.

Experts found that 92% of GeneAgent’s self-verification reports were accurate, demonstrating the program’s strong self-verification capabilities, particularly in comparison to GPT-4. Their thorough analysis validated the model’s ability to reduce hallucinations and produce more trustworthy analytical narratives.