NIH Advances AI Use with Agentic Tools to Boost Health Research

The National Institutes of Health (NIH) is developing its first AI Strategic Plan, which will shape future strategies of biomedical AI capabilities at the agency, as leaders explore how agentic AI can improve health research.

Susan Gregurick, NIH’s associate director for Data Science and director of the Office of Data Science Strategy, told GovCIO Media & Research in a statement that NIH has significantly invested in AI to support both research and development and administrative efforts over the past decade. The new plan will focus on high impact AI cases, utilizing AI for agency operations, best practices for AI validation in health care delivery and collaboration with the Food and Drug Administration to evaluate clinical AI tools.

“This strategic plan will foster synergy across programs, enhance transparency and expedite the research and development and translation of AI discoveries to benefit patients,” said Gregurick.

She added that, as NIH considers the growing number of AI applications, the agency is committed to ensuring future AI policies and guidelines to ensure patient privacy.

“As these tools become more sophisticated, they may increasingly run the risk of exposing the underlying data on which they were trained, potentially posing privacy risks to research participants,” Gregurick explained.

Leveraging Agentic AI to Improve Research

One example of NIH’s expanded AI use is the development of a new agentic AI tool at the National Library of Medicine (NLM), GeneAgent, which creates more accurate and informative descriptions of biological processes and their functions in gene set analysis compared to preexisting systems. The agency is relying on frameworks to ensure effective use and reliable outputs as AI becomes more integrated into operations.

NLM Senior Investigator Zhiyong Lu told GovCIO Media & Research that the tool performs functional analysis of novel gene sets from existing databases. He added that GeneAgent can serve as a “powerful tool for the discovery of therapeutic targets” because of its high accuracy rate.

“By comparing the original predictions with known knowledge retrieved from existing human-curated databases, GeneAgent automatically compiles a verification report that either supports or refutes each prediction,” Lu said in a statement to GovCIO Media & Research.

After examining the verification results, GeneAgent will summarize a prediction for researchers and provide an explanatory analysis to add context and justification for the predicted function. Zhiyong highlighted a new framework NIH used in the development of GeneAgent to ensure output reliability.

“In this study, we leveraged an agentic AI framework that incorporates an automated self-verification mechanism to mitigate the hallucination issues commonly observed in general-purpose GenAI tools,” he said.

NIH is also a partner agency in the National AI Research Resource (NAIRR) pilot, which is developing a national infrastructure to connect U.S. researchers with resources for AI research. The pilot is featured in the White House’s AI Action Plan. The plan calls on NAIRR to increase AI research opportunities, a part of President Trump’s larger goal to make the U.S. the global leader in AI.

NIH’s Commitment to Securing Health Data

As AI becomes commonplace, Gregurick said researchers have developed several privacy-preserving techniques to mitigate data security risks.

“NIH is considering the degree of potential risks posed by data leakage from generative AI and may develop future AI policies and guidelines to address these risks,” said Greguirk. “Currently, NIH has provided guidance that researchers who are using genomic controlled access data to train generative AI models may not share the model, including model parameters, except with collaborators.”