11 Apr 2025

As advanced biomedical technologies have allowed scientists to gather growing and increasingly complex datasets, even the most brilliant human minds grappling with the windfall can’t possibly keep up. Yawning gaps have spread out between data acquisition, scientific learning, and translation into tangible benefits for people suffering from diseases like Alzheimer’s. Closing these gaps is the goal of the Artificial Intelligence Biomedical Research Scientist Initiative, which is forming a consortium of Alzheimer’s and engineering researchers in academia, pharma, and the AI tech industry. The group aims to dramatically accelerate scientific discoveries and their impact on human health.

Randall Bateman of Washington University in St. Louis leads the initiative. He pitched the idea to some 60 scientists in a meeting alongside the AD/PD conference, held April 1 to 5 in Vienna. The consortium will build a platform of AI tools that draw on published data in PubMed, elsewhere, and on unpublished data contributed by scientific labs. This AI assistant is meant to suit specific needs of different labs, and help scientists in every aspect of their work, from building hypotheses to interpreting experimental findings in the broader context of all studies within the field. While the scientists hope the platform will ultimately deploy across the biomedical spectrum, it will cut its teeth on the study of AD, a 30-year-long pathogenesis process being investigated via numerous disciplines that generate heaps of increasingly multimodal data.

The consortium will focus on using AI for basic research, helping scientists discover mechanisms of the disease, and figuring out how to counteract them. It will not tackle clinical applications, such as diagnostic tools, clinical trial data analysis, or treatment monitoring. Smaller projects in this area presented results at AD/PD (see below).

Still in its development phase, the consortium has attracted stakeholders and collaborators from nearly 60 organizations so far (image below). They include academic scientists who focus on AD or AI, biotech/pharma companies with AD drugs in their pipeline, heavy hitters in the AI tech industry, including Microsoft and Google, funders, and government agencies. The consortium is slated to begin its work by July. Getting the project off the ground will cost around $10 million. Bateman estimated that fueling the computational needs of the platform will require a total of $100 million.

Click to Enlarge

Emerging Partnership. An AI Biomedical Research Scientist Consortium is forming with leaders from academia, pharma/biotech, AI companies, and other organizations. [Courtesy of Randall Bateman, Washington University.]

How will the group accomplish its lofty goal? In Vienna, Bateman laid out a three-year plan. In the first year, biomedical and AI scientists are to join forces to beef up a robust AI knowledge base. They will train large language models (LLMs), knowledge graphs, and graph-based retrieval augmented generation (RAG) on the scientific literature from PubMed and other public databases. The idea is that scientists will tap this knowledge base to conduct advanced biomedical queries that will go beyond a typical search—rendering and interpreting related findings, picking out scientific trends, and coming up with hypotheses.

Crucially, the consortium hopes that unpublished datasets, currently sitting in files and spreadsheets of labs around the world, will be integrated into this knowledge base. This could spare scientists from repeating other groups’ mistakes, or repeating accurate negative findings to a research question at hand. The idea is that such data will help the AI assistant distinguish—and weigh more heavily—findings that have drawn convergent evidence over the years from findings reported once but never confirmed and built-upon by others.

Another charge for the project’s first year is that it develop a “reviewer three” system. This is a virtual critical reviewer scientists can tap to get feedback on their manuscripts and grant proposals, and to help them decide which remaining experiments might strengthen their conclusion and boost their chance of publication.

Fit for Purpose. The goal is that scientists will tailor AI systems to address their unique research areas, such as imaging, neuropathology, or omics. [Courtesy of Randall Bateman, Washington University.]

In year two, the AI platform is to begin its transformation into an assistant that supports scientists across the research cycle, from searching existing literature, forming hypotheses, designing experiments to test those hypotheses, analyzing data, and interpreting it in the context of existing literature within and across scientific fields.

If all goes well, the AI assistant in the third year will get a promotion to “autonomous AI biomedical scientist.” By then, principal investigators ought to be able to direct the artificial “scientist” to design and run experiments. Bateman envisions a scene in which a scientist comes into the lab in the morning and asks the AI scientist to explain and interpret the outcome of the experiments it ran overnight and offer suggestions for what to do next. He projects that these AI scientists could accelerate AD research by orders of magnitude.

Is this achievable in three years? When Bateman tossed the question to the crowd of Alzheimer’s scientists in the room, most thought no. Among people who work in AI, the response is the opposite. “They think our timeline is too long,” Bateman said, adding “They’re on an exponential bend of the curve, and they expect things in months.” When pressed by Henne Holstege of VUMC Amsterdam, Bateman said he gives the question a “fair 50 percent chance, maybe higher.” Five years ago, he would have said, “not in my lifetime.” AI development in general, and its integration into aspects of AD research, has accelerated greatly since Alzforum last covered the topic in 2023 (Jul 2023 news).

At this consortium meeting, AI scientists showcased a few AI prototypes, and attendees toyed around with them. Justin Reese of Lawrence Berkeley National Lab offered up Alzassistant.org, a chat GPT-based model trained on Alzheimer’s literature from PubMed and Wikipedia. It answered some questions correctly, while missing the mark on more nuanced queries. These prototypes are in the development phase; to improve, they need access to full text, not just abstracts, and scientist input.

Cassie Mitchell of Georgia Institute of Technology, Atlanta, presented a strategy to build an AI scientist as an integrated set of knowledge bases, which are themselves structured layers of different types of information. Individual knowledge base layers can be increasingly specific, from cross-domain information relevant to several biomedical fields, to neuroscience, and then to Alzheimer’s specifically. By integrating such layers and using LLMs to draw relationships between them, scientists can get answers to their questions that are based on information beyond the realm of their siloed field, but retain enough specificity to be useful, Mitchell said (image below).

AI Onion? A cross-domain knowledge layer integrates multiple sources of information to answer research questions. [Courtesy of Cassie Mitchell, Georgia Tech.]

Next up was Sam Rodriques, a theoretical physicist/bioengineer turned AI scientist who co-founded FutureHouse. Found in 2023 with the goal of creating an AI scientist within a decade, this San Francisco-based nonprofit tries to help humans make sense of biology as research data expands exponentially. Noting that the complexity of biology means scientists miss impactful discoveries hidden within existing data, Rodriques said FutureHouse is developing better ways of drawing connections between seemingly disparate findings across fields.

For example, Future House has built PaperQA2. In response to a query, this “superhuman scientific literature search” extracts information from thousands of papers and synthesizes them into an exhaustively referenced article that answers a question or summarizes a topic. Rodriques claims that PaperQA2 produces summaries that outperform humans in accuracy. Bateman believes this technology will be transformative going forward.

At this point, the consortium is not wed to any one technology. Rather, the idea is to incorporate a suite of them that users will be able to fine-tune for their individual purposes.

This effort is but one example of how AI is making itself felt in AD research. In a correspondence paper published April 1 in Nature Medicine, Sandrine Andrieu of the geroscience institute IHU HealthAge, Toulouse, France, along with Bateman and 11 other leaders in academia, industry, and healthcare, articulate a broader vision for how AI can advance their field (Andrieu et al., 2025). It includes digital biomarkers for early detection, optimized recruitment, retention, and monitoring in clinical trials.

Some examples were already on display at AD/PD. For example, Michael Weiner of the University of California, San Francisco, reported promising results with the voice clinical dementia rating scale (vCDR). This is an AI-driven prototype of the electronic version of the CDR (eCDR) that UCSF has been developing in collaboration with Washington University and the U.K.-based AI company Novoic.

The original, in-person CDR has for many years been the gold standard for staging MCI and dementia, but efforts to digitize it have been difficult because it involves responsive conversation with a highly trained rater, Weiner said. Early in the assessment, an interviewer asks the patient’s care partner to describe a recent event in their lives, and then later asks the patient about the event to probe their memory. Weiner said that the AI-driven vCDR conducts these interviews with a striking degree of “empathy,” accuracy, and consistency. It deftly steered interviewees back on track when their thoughts wandered off topic. The eMCI had never been able to pull this off. As a result, it elicited much less actionable information than the vCDR does. The vMCI can converse in 30 languages, is available 24/7, delivers automated results, and avoids the problem of variance between raters, Weiner said.

In their correspondence paper, Andrieu et al. emphasized that AI-driven initiatives must be anchored in the needs and experiences of patients and their families. “The convergence of AI and Alzheimer’s disease research marks a pivotal moment and a chance to redefine our approach to one of the most pressing health challenges of our time,” they wrote. “Let us seize this opportunity to foster innovation that is both scientifically rigorous and deeply compassionate.”—Jessica Shugart

Comments

No Available Comments

Make a Comment

To make a comment you must login or register.

References

News Citations

1. Artificial Intelligence Is Everywhere You Look in Dementia Research 12 Jul 2023

Paper Citations

1. Andrieu S, Bateman RJ, Bereczki E, Bose N, Brookes AJ, Doraiswamy PM, Kivipelto M, Kohlhaas S, Kourtis L, Middleton LT, Moore GJ, Petrella JR, Postlethwaite N, Sainz-Fuertes R. Harnessing artificial intelligence to transform Alzheimer's disease research. Nat Med. 2025 Apr 1; Epub 2025 Apr 1 PubMed.

External Citations

1. PaperQA2

Further Reading

No Available Further Reading