Prima Mente Unlocks Early-Stage Alzheimer’s Diagnostics with Epigenome Model

AI biology company, Prima Mente, is on a mission to tackle one of the greatest challenges in medicine, uncovering the molecular basis of neurodegenerative conditions. These aging-related diseases are characterized by subtle changes in gene regulation, cellular identity, immune signaling, and more that begin decades before symptoms emerge.  

While much progress in therapeutics has focused on designing new protein structures or analyzing fixed genome sequences, Prima Mente seeks to establish a new paradigm for precision diagnostics that uses AI to understand the dynamic epigenome, where aging biology happens in real-time.  

“Imaging and cognitive scores don’t necessarily tell you how to intervene or change the path of disease,” said Hannah Madan, PhD, co-founder at Prima Mente, in an interview with GEN Edge. “This is what epigenetics can start to unravel.” 

The company’s AI model, Pleiades, analyzes DNA methylation changes in cell-free DNA (cfDNA) derived from patient blood samples to diagnose neurodegenerative disease. The method presents an appealing alternative to invasive and costly brain biopsies, which are the current diagnostic standard. 

Strikingly, Pleiades demonstrated an 89% success rate for detecting early-stage Alzheimer’s and Parkinson’s in real-world clinical cohorts. That rate subsequently increased to 97% when the model was integrated with leading neurodegenerative protein biomarkers, such as pTau-217.  

Prima Mente is an NVIDIA Inception member and has partnered with NVIDIA Cloud Partner, Nebius, to leverage the DGX Cloud Lepton AI platform to train and develop Pleiades. The company, whose name translates to “first mind” and signals the approach of breaking down problems to first principles, was founded in 2022 and currently hosts over two dozen employees based across London, San Francisco, and Dubai.   

“The hypothesis we had from the beginning was that bigger models would help understand complex biology over time, which requires both compute and the ‘know-how’ for building these models,” said Madan, when describing the motivation for the Nvidia and Nebius collaboration.  

Pleiades is currently deployed in a live clinical study, called SANDBOX, across 15 sites in the U.K. The model will support the classification and stratification of patients with suspected neurodegenerative conditions to improve clinical outcomes and manage waiting lists in the U.K. National Health Service. Designing therapeutics in-house is also on Prima Mente’s horizon. 

In the fragment 

Pleiades is a 7 billion parameter epigenetic foundation model trained using 1.9 trillion tokens of epigenetic data representing human methylation and genomic sequences. Data was sourced from genomic and cfDNA across multiple cell types and sequenced with methods that preserve cytosine methylation state. 

By training on tissue-specific information, the model can accurately pinpoint cfDNA cell type of origin, identifying both brain components (neurons and microglia) and immune components (T-cells and B-cells), as predictors of Alzheimer’s disease. The team is building a virtual cell model, named Parthenon, to discover novel therapeutic targets for microglia-related cells. 

Notably, Pleiades demonstrated that cfDNA fragmentation patterns can act as a novel class of biomarkers for Alzheimer’s detection. These fragments reflect biological processes such as epigenetic state, nucleosome positioning, and mechanism of cell death.  

Results show that Pleiades can reconstruct cell-free DNA fragments with 83% per-nucleotide accuracy and a 91% match to original methylation patterns. Additionally, the model enriches signals from specific cell types in the blood by up to 20-fold to enable the detection of rare disease signatures from limited material. 

Looking ahead, Madan said Prima Mente aims to expand the model beyond epigenomics to proteomics, transcriptomics, and more clinical data. To support this goal, the company is currently conducting longitudinal studies that include both blood samples and brain tissue.  

Madan also emphasizes that the platform is generalizable and will be applied to more disease areas in 2026.