Federal Medical Scientists Turn to LLMs as Research Co-Pilots Across FDA, NCI and Argonne
核心洞察
Federal agencies including the FDA, NCI and Argonne National Laboratory (搜索) are deploying large language models (LLMs) to accelerate drug development and biomedical research workflows.
The FDA's Elsa app is used by 80% of agency employees to review thousands of direct-to-consumer ad submissions, while NCI's Ask NanCI enables conversational interaction with scientific content.
Secure, centralized data governance underpins these tools, with agencies building internal application layers rather than feeding sensitive data to commercial LLMs.
Federal medical research agencies are increasingly deploying large language models (LLMs) as "research co-pilots," using the technology to address bottlenecks in language and knowledge synthesis across drug development and biomedical research. The Food and Drug Administration (搜索) (FDA), the National Cancer Institute (NCI), and the U.S. Department of Energy-sponsored Argonne National Laboratory (搜索) have each built customized LLM applications tailored to their specific scientific workflows, according to agency officials.
Medical research-associated LLM use has expanded in recent years due in part to the models' increasingly nuanced architecture and the availability of healthcare and biomedical data, such as electronic health records, says Wes Anderson, quantitative medicine scientist at the Critical Path Institute (搜索), a nonprofit that works with researchers and other entities to accelerate drug development. "The current generation of LLMs — which can make connections between different words in a sentence and be trained to respond to prompts and follow instructions — are being used to address bottlenecks that involve language and knowledge synthesis," Anderson says.
Faster Fact Finding Through Retrieval-Augmented Generation
Organizations may use prebuilt models from providers or adapt and host models in their own environments for specific tasks. "You're able to build tools on top of them and do things like create retrieval-augmented generation pipelines that you can give direct access to the information, run some code to answer some of the questions you may be interested in or generate code," Anderson says.
At the NCI, the agency introduced NanCI, followed by Ask NanCI, a feature that lets researchers interact with scientific content in a conversational way. "You can ask any question — for example, 'Tell me the major outcome of this paper,' or a specific question about a figure in a paper," Zahir says. "Ask NanCI will use the information in it and large language modeling to generate a response."
Google Cloud (搜索) provides the underlying cloud infrastructure and access to LLMs that power the Ask NanCI function, Zahir says, and also assists with updates to NanCI, which is available as an iOS app and web application accessible on any device, including Android smartphones. "NanCI has the potential to become a more integrated part of the research ecosystem — hopefully helping not just with discovery but with collaboration, identifying trends and perhaps even helping inform research directions," she says. "It represents a shift toward an AI-native scientific infrastructure, where AI tools can actively support how science is done, rather than just storing information."
Secure Data Governance as a Foundation
In 2024, Argonne National Laboratory (搜索) launched Argo, a custom interface that provides secure access to large language models such as Google's Gemini and OpenAI's GPT-5.4, says Matthew Dearing, the lab's AI for operations technical lead. The Lemont, Ill., research center's goal is to provide the core technology that researchers need to leverage generative AI and LLMs for data analysis, content generation and other objectives.
Communication between LLMs, users and any software they build occurs on the lab's network. "In the early days, the main large language model companies made it very clear that if you gave them your prompts and data, they were going to save and reuse that in training the next versions of models," Dearing says. "We either have the large language models running internally on systems that live inside Argonne that we can control, or we communicate with copies of large language models run through a secured network."
Security was also a central consideration for the FDA when building its LLM-based Elsa app. Data that nonclinical laboratory and other study sponsors have submitted to the agency does not meet the security requirements to be put into a government-issued version of a commercial LLM, says Chief AI Officer Jeremy Walsh. "We had to say, 'How do we get access to all of the model capabilities that exist and tailor the functionality specifically to the FDA's use cases?'" he says. "We built out an application layer that sits between the users and models to give multimodal support for working with large documents — building out workspaces for them, doing custom prompt libraries."
As Elsa — whose infrastructure involves Amazon Web Services (搜索) and Google Cloud (搜索) components — was being developed, the agency worked to centralize its internal divisions' separate data silos, which included consolidating 40 application and submission systems into one platform. "Right now, if you want to analyze an application, you'd have to get the files off a file-share, upload them to Elsa and then ask your questions," Walsh says. "With the next release, you can just do it right from Elsa — tell it to analyze this application this way, and it will grab all of the data, process it and analyze it. We're able to do things we couldn't do last year because the data was all over the place."
Expediting Research While Managing Risk
Currently, 80% of FDA employees use Elsa — which they can access with an internet connection and their agency-issued Dell laptop — to complete work such as reviewing thousands of direct-to-consumer ad submissions to identify ones that might qualify for a compliance violation warning letter. "Was the AI making a decision?" Walsh says. "No. Was it helping filter and categorize risk profiles associated with direct-to-consumer ads? Absolutely. Then, those were all reviewed by the experts at the FDA, and some they threw out."
LLMs can potentially produce problematic hallucinations and other inaccurate responses. Some agencies, including the FDA, have issued guidance on how to approach the different risk levels that the models' use can present, Anderson says. As the technology advances, understanding where failures can exist — and how to ensure models are continuously evaluated to confirm their research output is accurate — can help foster credibility.
"There's a lot of testing around how more autonomy can lead to increased productivity in different areas — evidence synthesis, biomedical text extraction, finding patients for clinical trials," Anderson says. "But as these processes become more automated, we need more validation. We aren't trying to use these models to replace scientists. We're trying to augment some of the more tedious parts of our workflows."