From Symptoms to Brain Cells: Two AI Tools Map Individual Gene Networks in Alzheimer’s Disease

Differences in Alzheimer’s disease often lie beneath the same diagnosis: some people experience marked cognitive decline, some also have depression or insomnia, and others retain better cognitive abilities despite pathological changes in the brain. To understand these differences, researchers need to connect each person’s symptoms to the details of brain cells and gene activity. PASCode and iBrainMap, developed by a team at the University of Wisconsin–Madison, are attempting to fill this gap.

In an announcement on September 29, the university described the two studies, published in Nature Medicine and Nature Communications, respectively, and based on large-scale data from the PsychAD consortium. This dataset includes 1,494 postmortem brain tissue donors and more than 6.3 million cell nuclei, primarily from the dorsolateral prefrontal cortex, covering multiple brain diseases and control cases. These large numbers therefore describe a shared data resource and should not be taken directly as the sample size for each Alzheimer’s disease analysis.

PASCode addresses the question of which cells are associated with a particular disease manifestation. According to its public software documentation, it takes single-cell sequencing data and group labels, such as disease and control, integrates multiple cell abundance analyses and ranking results, and then uses a graph attention network to score individual cells. A study abstract listed by PsychAD states that the team identified approximately 1.5 million phenotype-associated cells from 584 donors with Alzheimer’s-related manifestations and compared 27 brain cell subclasses. These scores indicate the degree of association; they do not establish that a particular cell caused a symptom.

This analysis allows researchers to further distinguish whether different states exist within the same cell type. For example, the PASCode study linked some microglial populations to Alzheimer’s pathology and identified reactive astrocytes that may be associated with cognitive resilience. Cognitive resilience refers to the phenomenon in which cognitive function is relatively preserved despite a pathological burden in the brain. These findings help narrow the scope of subsequent experiments to test which cell states may be protective.

iBrainMap shifts the focus to the individual: it uses each donor’s data to infer interactions among cell types and gene regulatory networks, then assesses the importance of cells, genes, and connections to disease manifestations. It also links genetic variants to regulatory changes at the individual level, exploring why molecular pathways in different people may lead to different outcomes. The analyses described by the university covered symptoms including weight loss, insomnia, and depression, and identified potential disease subgroups with distinct molecular characteristics.

Regarding validation, the PsychAD abstract states that iBrainMap’s cross-cohort analysis covered a total of more than 1,900 individual brain samples; PASCode also validated its findings in external Alzheimer’s disease and major depressive disorder datasets. The team also made code, pretrained models, and interactive resources publicly available. PASCode provides tutorial notebooks, while iBrainMap provides an analysis workflow and a Docker environment, enabling other researchers to reproduce and examine the results.

However, a gap remains between validation across datasets and clinical usability. These tools primarily analyze postmortem brain tissue, and this does not yet establish that they can reliably diagnose living patients, predict disease progression, or guide medication use. The regulatory relationships inferred by the models also require confirmation through functional experiments. A more concrete use at present is to prioritize target exploration, helping researchers select biological questions from millions of cellular data records that warrant individual validation.