A team centered at MIT has released FINGERS-7B, described as the first AI foundation model purpose-built to make Alzheimer's disease preventable. Announced on April 26, 2026 and presented the following day at the ICLR conference in Rio de Janeiro, the model integrates lifestyle, clinical, genomic, and proteomic data from tens of thousands of at-risk individuals to surface biomarkers of preclinical Alzheimer's, years before memory symptoms appear.
Key Highlights
- First open-source AI foundation model for Alzheimer's prevention, with weights, training code, and evaluation pipelines released publicly
- Reports four times more accurate preclinical diagnosis and a 130 percent improvement in responder stratification versus prior methods
- Deployed in the AD Workbench, a secure cloud environment run by the Alzheimer's Disease Data Initiative and used by researchers worldwide
- Built on the WW-FINGERS network of 30,000 participants across 40 countries
- Trained and shipped in roughly ten months from a 100,000 dollar seed grant by MIT's Aging Brain Initiative
Details
FINGERS-7B is the modeling core of a broader platform the team calls FINGERPRINT, which pairs the foundation model with AI agents that run automated multi-omic analyses. Rather than treating each data domain as a separate puzzle, the system learns jointly from lifestyle, clinical records, biomarkers, genomic data, and proteomic signals to find the cross-domain patterns that single sources of data miss.
Adrian Noriega, an MIT-Novo Nordisk AI Fellow and FINGERPRINT co-lead, framed the approach in biological terms. "Each of us carries a biological fingerprint, basically a unique combination of signals that reveal disease risk and, if properly understood, could enable prevention and treatment of Alzheimer's disease," he said. Noriega co-leads the project with Arvid Gollwitzer, a Broad Institute research scholar who led the design and training of the model.
Given an individual's data, the model is designed to estimate Alzheimer's risk, forecast the likely course of cognitive decline, and predict the effect of candidate interventions, from dietary change to drug therapies.
Impact
The biggest near-term implication is for prevention trials. Earlier risk prediction could help researchers identify candidates for intervention studies before symptoms set in, when lifestyle changes or therapies may have a better chance of altering the course of disease. The reported 130 percent improvement in responder stratification points to a second use: sorting participants into more precise subgroups for clinical research, making trials more targeted and less blunt.
Because the model is open source and hosted inside the AD Workbench, outside groups can validate it on their own cohorts without moving sensitive patient data or rebuilding infrastructure. "Someone was going to build the foundation model stack for Alzheimer's prevention," Gollwitzer said. "It should be open, and it should be now."
Background
The work builds on Professor Miia Kivipelto's original FINGER study, which focused on cognitively unimpaired but at-risk older adults and inspired the global WW-FINGERS network now spanning 40 countries and 30,000 participants. MIT's Aging Brain Initiative, directed by Picower Professor Li-Huei Tsai, seeded the project with a 100,000 dollar grant to Noriega and mechanical engineering professor Giovanni Traverso. Within about ten months, the team trained FINGERS-7B, deployed it in the AD Workbench, and opened the weights.
Institutional partners include the Broad Institute, Yale University, Imperial College London, and Brigham and Women's Hospital. Industry partners include Alamar Biosciences and Novo Nordisk. Earlier this year, the Davos Alzheimer's Collaborative and the FINGERS Brain Health Institute announced a partnership to use FINGERPRINT to support globally inclusive prevention research.
What's Next
External validation on independent cohorts will be the next test of whether the reported accuracy gains hold up outside the WW-FINGERS data. Because the model and code are open, that validation can happen in parallel across multiple research groups rather than waiting on a single sponsor. For the broader AI community, FINGERS-7B is also a notable data point in the rise of multi-omic foundation models, a class of systems that learn across biological data layers rather than within one.
Source: Picower Institute at MIT