Summary

A study used single-nucleus data from human prefrontal cortex to map genetically regulated gene-expression signals across 12 brain disorders. Its cell-type and ancestry-aware models identified associations that bulk-tissue analyses missed.

A study published in Nature on 23 September 2026 used single-nucleus gene-expression data to map genetic associations with 12 neuropsychiatric and neurodegenerative disorders. Its models resolved signals to particular brain cell populations and included data from European, African and admixed American ancestry groups, revealing associations that analyses of bulk brain tissue often miss.

Contents

How the study mapped genetic signals

The researchers built their models using the PsychAD Consortium’s single-nucleus RNA-sequencing atlas of the dorsolateral prefrontal cortex. The atlas contains more than 6 million nuclei from 1,494 donors. For model training, the team used data from 1,359 genotyped donors: 920 of European ancestry, 321 of African ancestry and 118 of admixed American ancestry.

The models estimate genetically regulated expression, or the part of a gene’s expression that can be predicted from genetic variants. The researchers trained 94 single-nucleus transcriptomic imputation models across 32 cellular populations, then combined them with genome-wide association study data for 12 disorders. This approach—single-nucleus transcriptome-wide association study, or snTWAS—tests whether genetically predicted expression of a gene is associated with genetic risk for a disorder. It does not measure every kind of gene-activity change in the brains of people with that disorder.

The team also applied ancestry-matched models in the Million Veteran Program, a cohort of approximately 600,000 participants, to examine associations across disorders and ancestry groups.

Cell-type resolution revealed more associations

In European-ancestry data, models at the cellular-class and subclass levels could reliably predict expression for 11,021 and 11,052 protein-coding genes, respectively. The comparable pooled-nucleus model covered 9,147 genes, while a bulk-tissue model covered 8,959. Finer-resolution models also identified more significant gene–trait associations: the study reported 3,003 unique associations at subclass level, compared with 2,470 at class level and 1,254 in the pooled-nucleus analysis.

The additional signals were enriched for genes already linked to neurological or psychiatric symptoms. Bulk and pooled-nucleus analyses showed this enrichment for five and four of the 12 disorders; class-level models showed it for 11, and subclass models for all 12. The study classified 22.5% of its significant associations as novel because they were not identified by either bulk-tissue TWAS or a separate gene-based analysis. Here, “novel” describes the comparison with those analyses, rather than proving that a gene has never been linked to disease in any context.

The cell-level results also separated signals that bulk measurements can blend together. For Alzheimer’s disease, for example, the study found that the predicted-expression association for BIN1 varied substantially between cell types. Its pathway analyses linked Alzheimer’s-related signals in immune cells and microglia to processes including tau-protein binding and regulation of neuron death. These are statistical and pathway-level findings that help prioritize candidate genes and cellular contexts for further study.

What the findings offer—and their limits

Across ancestry groups, the researchers reported broadly similar patterns among highly ranked associations, alongside signals found in particular groups. Multi-ancestry fine-mapping also recovered African-ancestry associations missed by European-only fine-mapping. This supports using ancestry-matched models rather than assuming that models trained in one population will capture genetic regulation equally well in another.

The study’s main resource is a more detailed map of where genetically regulated expression may connect to brain-disorder risk. Such maps can guide follow-up work on candidate genes and pathways, but the analysis is not a treatment study and does not by itself establish that changing a prioritized gene would alter disease. Its reference data came from the dorsolateral prefrontal cortex, so the results describe that region rather than the full cellular landscape of the brain. East and South Asian ancestry groups were not included in model training because the available sample sizes were insufficient for the study’s analyses.

Sources