Deep learning framework PASCode uncovers cellular basis of Alzheimer's disease heterogeneity

On September 24, 2026, researchers led by Daifeng Wang at the University of Wisconsin–Madison and Panos Roussos at Mount Sinai Hospital published a study in Nature Medicine that combines artificial intelligence with large-scale single-cell genomics to dissect the heterogeneity of Alzheimer's disease (AD) at single-cell resolution.

Using a PsychAD consortium dataset comprising single-nucleus RNA sequencing data from 1,494 donors and more than 6 million brain cells from the dorsolateral prefrontal cortex, the team developed PASCode (Phenotype Associated Single Cell Encoder), a deep learning framework that integrates multiple differential abundance algorithms to assign aggregate phenotype labels to cells and then trains a graph attention network to compute a phenotype association score (PAC score) for each cell. The method was validated on independent datasets including SEA-AD and ROSMAP.

PASCode identified approximately 1.5 million phenotype-associated cells across 27 brain cell types. Microglia and astrocytes ranked highest among AD-associated cells, with vascular leptomeningeal cells also prominent. SST-positive inhibitory neurons showed an approximately 10-fold relative reduction in AD patients. Analysis at the PAC level detected 1,148 differentially expressed genes, compared with only 220 at the traditional donor level, capturing AD-related microglial pathways such as tau binding and amyloid-beta binding.

The study also revealed cellular signatures of cognitive resilience. AD-resilient and AD-strict individuals differed fundamentally at the molecular level, with IT-type excitatory neurons and multiple inhibitory neuron subtypes showing the largest differences. An astrocyte subtype, Astro_PLSCR1, highly expressed the neuroprotective marker EMP1 in resilient individuals, while the strict group expressed the neurotoxic marker FKBP5. Additionally, astrocytes associated with both AD and depression (e.g., Astro_WIF1) showed activation of endoplasmic reticulum stress and inflammatory responses, suggesting shared cellular states and molecular mechanisms between the two conditions.

PASCode is now openly available to the research community as an open-source tool, a pretrained AI model, and an interactive web application.