AI maps cell-level signatures of Alzheimer’s disease phenotypes in millions of brain nuclei

An artificial intelligence framework has sifted through more than 6 million single-cell transcriptomes from human prefrontal cortex to pinpoint the specific cell subpopulations, genes, and pathways that may underlie different Alzheimer’s disease (AD) phenotypes—from cognitive resilience to depression. The work, published today in Nature Medicine, was led by Daifeng Wang at the University of Wisconsin–Madison and Panos Roussos at the Icahn School of Medicine at Mount Sinai, with first author Chenfeng He.

The researchers developed a tool called PASCode (Phenotype Associated Single Cell encoder), which combines multiple statistical methods and a graph neural network to score how strongly each individual cell is associated with a particular clinical phenotype. Applying PASCode to the PsychAD consortium data—covering over 6 million nuclei from 1,494 donors, including 584 with detailed AD-related phenotypes—they identified roughly 1.5 million phenotype-associated cells. These cells were then used to home in on meaningful biological differences.

Key findings include the prioritization of microglia as the most AD-relevant cell class, with specific subpopulations linked to known AD pathology pathways like tau binding and amyloid-β clearance. The analysis also revealed astrocyte subtypes with opposing neuroprotective and neurotoxic gene expression programs that may explain cognitive resilience—why some people with severe AD pathology maintain normal cognition. In cognitively impaired AD donors, the framework flagged excitatory/inhibitory imbalance and mitochondrial dysfunction in neurons as likely contributors.

Notably, the team also examined depression in AD, a common comorbidity that is often understudied. They found that depression-associated cells, particularly in astrocytes and oligodendrocytes, showed distinct inflammatory signatures, with IL18 emerging as a key marker. Shared gene expression patterns between AD and depression suggest potential common mechanisms, including endoplasmic reticulum stress in a specific astrocyte subtype (Astro_WIF1).

The PASCode framework and its pre-trained models are available as an open-source tool with a web application, and the results were validated in independent datasets such as SEA-AD, ROSMAP, and an MDD cohort. The authors caution that the current version assesses one phenotype at a time and treats disease progression in a linear fashion, so the tool may not yet capture the full complexity of AD’s heterogeneous continuum.

Still, by pinpointing cell subpopulations and pathways that are invisible at whole-tissue level, PASCode offers a new way to identify biomarkers and therapeutic targets for precision medicine in AD—and potentially other complex diseases.