← Back to Home

Beyond Average Expression: Study Identifies 27 Overlooked Gene–Cell Clues in Alzheimer’s Disease

Analyzing brain cell nuclei from 627 donors, TriSCOPE combines machine learning with genetic evidence to identify disease signals missed by conventional methods. This candidate list can guide mechanistic research, but clear limits remain for individual diagnosis.

By SURL BioNews

To find molecular clues to Alzheimer’s disease, researchers often compare gene expression between patients and controls to see which genes become more active and which become quieter. However, the absence of a marked change in a gene’s average expression does not mean it is unrelated to disease. A study published November 9 in Nature Communications cross-compared multiple analyses to identify 27 genetically supported gene–cell type pairs that conventional differential expression analysis had missed.

The study used a framework called TriSCOPE across six single-nucleus RNA sequencing datasets, totaling approximately 4.2 million nuclei from 627 donors, to analyze six major brain cell types and 41 cell subpopulations. All data came from the dorsolateral prefrontal cortex. The large number of cells allowed researchers to examine different cell states in detail, but coverage remained concentrated in a single brain region.

TriSCOPE brings together three analytical approaches: machine learning models that consider multiple genes simultaneously, gene expression comparisons that account for differences among donors, and “pseudotime” analysis that infers cell states from expression patterns. The aim is to identify signals that do not stand out when tested individually but still provide disease-related information when considered alongside other genes and factors such as age, sex, and APOE genotype.

The research team then compared genetic signals for Alzheimer’s disease risk with those regulating gene expression to examine whether the two might share the same genetic variant. This provided additional support for 27 pairs, including JAZF1 and SYK in microglia. These genes were informative in the models but did not meet the significance threshold for differential expression. This makes the candidate mechanisms worth further investigation, but a shared genetic signal alone still cannot establish that a gene directly causes disease.

For validation, key candidate signals, including ARL17B, LINGO1, and APOE, were replicated in independent cohorts, rather than appearing only in the initial data. This does not mean that all 27 pairs have individually received the same level of confirmation. The study also found that expression patterns associated with cognitive impairment did not fully match those associated with pathological changes in the brain, reminding researchers that explaining cognitive function and explaining the accumulation of pathology may require different molecular clues.

The team also released analysis code covering predictive modeling, differential expression, pseudotime inference, and genetic colocalization workflows. The repository documents five random data splits for the predictive models and seven random seeds for trajectory analysis. It also includes synthetic microglial expression data and expected outputs, allowing other researchers to test the prediction workflow before obtaining access to controlled human data. Such demonstrations help assess the methods but cannot replace biological validation in real cohorts.

The practical value of this work is to narrow the range of candidates for follow-up experiments. Because the analysis relies on cross-sectional data from postmortem brain tissue, even an ordering of states shown by pseudotime cannot be treated as an actual record of disease progression in the same person. It remains unclear whether these signals appear before disease onset or arise as subsequent responses to the disease, and they cannot be used to predict an individual’s future disease course. Advancing the candidate list into therapeutic targets will require functional experiments and studies better able to establish temporal order.

References

  1. Nature Communications
  2. AdiVM / GitHub
  3. Genome News