C-BRAIN Uses Federated AI to Speed Alzheimer’s Research

For Alzheimer’s researchers, the challenge is not simply generating more data. Findings can remain scattered across institutions, and reviewing evidence across a growing body of work takes time that research teams could spend testing ideas. C-BRAIN, a consortium of 18 institutions, is building AI tools and data infrastructure intended to help scientists move from research questions to analyses more quickly.

The effort, led by Dr. Randall Bateman, launched in November 2025. Its members span academic, philanthropic, pharmaceutical and research organizations. The consortium is working with NVIDIA on systems that can use research data held by participating institutions, including unpublished findings, negative results and governed proprietary data.

The scale of the underlying need is large: the source reports 57 million people worldwide living with dementia and more than 10 million new dementia cases diagnosed each year. It also reports that annual National Institutes of Health funding for Alzheimer’s research reached $3.9 billion after increasing sevenfold over 15 years. Despite that investment, no treatment has been found to stop, prevent or reverse the disease.

Bringing hidden research data into analysis

C-BRAIN calls data that has not been available to AI tools or researchers “dark data.” This can include results that were never published, as well as information that institutions hold under access and governance rules. The consortium’s premise is that leaving such material out of analysis limits what researchers can learn from the evidence already collected.

Its planned infrastructure combines centralized and federated computing. Some data is placed directly in secure systems for AI tools to use. In the federated approach, models learn from data held at individual institutions, while the data itself stays within those institutions’ systems. This design aims to make analysis across multiple data sources possible while respecting local privacy and governance requirements.

That setup brings practical engineering demands: systems must work across institutional boundaries, handle varied datasets and support model training without moving all data into one repository. C-BRAIN currently operates 64 NVIDIA data center GPUs on premises, alongside cloud nodes on Microsoft Azure. The consortium says it expects to add NVIDIA GPU capacity, including Blackwell systems, as its work expands.

The collaboration also includes NVIDIA’s BioNeMo open developer platform and the MONAI framework, which C-BRAIN is incorporating into its research pipelines. The technical work is relevant to spatial omics, where models must capture relationships among biological measurements tied to location in tissue. C-BRAIN chief technology officer Dr. Ade Ojewole said that substantial optimization will be needed as the team pushes current machine learning architectures to model those data.

Two tools put AI into research workflows

C-BRAIN has built two tools aimed at different stages of scientific work. OpenScientist, developed with Lawrence Berkeley National Laboratory in an effort led by research scientist Justin Reese, takes a scientist’s question and dataset, then semi-autonomously proposes hypotheses, runs analyses and produces results designed to be checked and audited.

InsightEngine, built with 387 Labs, focuses on evaluating hypotheses against scientific evidence. It can analyze data supplied by a user, search a biomedical knowledge graph and compare claims with the literature. The goal is to help researchers spot unsupported claims, including possible AI-generated errors, before committing time and resources to experiments.

Published case studies report that OpenScientist completed analyses in minutes that domain scientists estimated would take weeks to months. Each run cost less than $10. These results describe particular case studies, not a general guarantee that every scientific analysis will see the same time or cost savings. The work still depends on the quality of the input data, the research question and whether scientists can verify and interpret the output.

C-BRAIN says its tools currently reach more than 200 research groups. At that scale, the consortium estimates that the capacity released could exceed one million research hours annually. That figure is a projection of potential capacity, not a reported count of hours already saved. Its practical value will depend on whether teams can fit the tools into their workflows and use the time gained for productive follow-up research.

From pattern finding to testable mechanisms

The consortium’s longer-term goal is a multimodal biological foundation model trained on the breadth of its data. The planned inputs include neuroimaging, multiple types of omics data, fluid biomarkers and longitudinal clinical measures. Combining these sources could help researchers examine how different biological signals relate to brain aging and Alzheimer’s disease over time.

C-BRAIN aims by its third year to identify new causal mechanisms of brain aging and Alzheimer’s. That is a more demanding target than finding correlations across datasets: a useful result must point to a biological process that researchers can test. The consortium has described this as moving from patterns in data toward testable scientific principles.

For now, the clearest reported evidence is operational: OpenScientist has produced analyses quickly and at low stated cost in published case studies, while the broader infrastructure is still being assembled. The next measure of progress will be whether these tools generate findings that other scientists can validate and build on. C-BRAIN’s approach could increase the reach of existing research data, but its contribution to new treatments will depend on what those analyses reveal and how well experiments confirm them.