An AI foundation model developed at MIT can predict which healthy individuals will develop Alzheimer's biomarkers and which will respond best to lifestyle interventions, researchers reported at the 14th International Conference on Learning Representations in Rio de Janeiro in April. The model, called FINGERS-7B, integrates genomic, protein, metabolic, and microbiome data to find multi-omic signatures that could lead to new biomarkers and drug targets for Alzheimer's disease.
The work highlights a growing trend in AI for Science & Research: using foundation models to find patterns across multi-omic data.
The model, trained on one of the largest lifestyle prevention trials that combines exercise, diet, and cognitive training, uses 7 billion parameters to spot patterns across different biological data types. Adrián Noriega de la Colina, a researcher at MIT and co-lead developer, said the approach "allows us to read through all of those languages simultaneously and find patterns that are invisible to any of us individually."
How FINGERS-7B reads biological languages
FINGERS-7B operates like a chatbot for biological information, recognizing relationships without needing to understand what individual genes, proteins, or microbes do. "They can figure out relationships and correlations," said Arvid Gollwitzer, scientist at Broad Institute and model co-developer, "without ever being taught translation directly." The model had learned biomarker patterns from previous research and applied them to new data from the FINGER prevention study.
Predicting cognitive decline and drug targets
By analyzing multi-omic signatures - combinations of multiple biological signals rather than a single protein or microbe - the model identified gut microbiome signatures that predicted cognitive decline within three years with 89 percent accuracy. It also flagged which participants would benefit most from the lifestyle intervention and highlighted four potential drug targets from the microbiome data.
Neurologist Timothy Chang, director of the UCLA California Alzheimer's Disease Center, who was not involved in the study, called the approach "interesting," noting that it transfers research methods from other domains, like microbiome science, to Alzheimer's.
AI's broader role in Alzheimer's research
Other research groups are also applying AI to Alzheimer's. Chang's team has used AI to scan electronic health records for undiagnosed cases and is now working to predict how cognitive scores change with age using real-world data. These models, though harder to develop because patient visits are irregular, may better reflect everyday clinical settings.
The model is open source, allowing other scientists to use and build on it. The team describes FINGERS-7B as a hypothesis-generating model: its predictions about biomarkers and risks must be validated in further studies before any clinical use.
Why this matters for science and research professionals
FINGERS-7B shows how foundation models can turn large, heterogeneous biological datasets into testable hypotheses. For researchers, the open-source code offers a starting point to explore multi-omic interactions in Alzheimer's and other diseases. The ability to predict individual responses to interventions could shift prevention strategies toward personalized approaches, though the findings still require rigorous validation.
Your membership also unlocks:
