When it comes to data, “bigger is better” is no empty phrase. At the Alzheimer’s Association International Conference held last month in San Diego, California, scientists presented their first analysis of two of the largest single-nucleus RNA sequencing efforts in Alzheimer’s research to date. They included 1.2 million cells from 84 people in the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD) and 1.6 million cells from 478 participants in the Religious Orders Study and Memory and Aging Project (ROSMAP). These massive datasets taught scientists which neuronal subpopulations died early in AD, showed a subsequent rise in glia, and tied gliosis to memory decline in a subset of people.

  • Scientists have created two snRNA-Seq datasets from millions of cells.
  • Transcriptome analysis, plus quantitative neuropathology, showed sequential cellular changes during AD progression.
  • Gliosis marked people with worse neurodegeneration, faster cognitive decline.

“It’s exciting that different groups see similar abundance and gene-expression changes in single-nucleus data from AD donors, especially with distinct patient cohorts and sampling strategies,” Kyle Travaglini, Allen Institute for Brain Science, Seattle, wrote to Alzforum. “It gives us more confidence that our observations are robust and could lead to a better understanding of the disease process.” Evan Macosko of the Broad Institute of Harvard and MIT in Cambridge, Massachusetts, agreed. “The consistency makes me very optimistic about the future of single-cell analysis as a tool for studying disease mechanisms,” he wrote to Alzforum (full comments below).

Researchers at the Allen Institute, University of Washington, and Kaiser Permanente Washington Health Research Institute recently released the first fruits of their collaborative SEA-AD efforts. Published on the SEA-AD website, this database includes genotyping, single-cell transcriptomics, single-cell epigenomics, and spatial transcriptomics on cells from the middle temporal gyrus. The MTG is a portion of the temporal lobe that is important for learning and memory. “We wanted to start with an area that degenerates in mid-stage AD,” Eitan Kaplan of the Allen Institute told Alzforum.

The brain tissue came from people 65 years and older in two well-characterized cohorts, i.e., that of the UWash AD Research Center and the Adult Changes in Thought study led by UWash and Kaiser Permanente. The scientists sorted volunteers by their AD neuropathological change (ADNC) score, ending up with nine controls sans pathology and 12, 21, and 42 with low, intermediate, and high pathology, respectively. ADNC is a composite of Braak neurofibrillary tangle stage, Thal amyloid plaque phase, and CERAD plaque/neuropsychological testing score.

Travaglini presented the MTG snRNA-Seq data. The 1.2 million nuclei fell into 139 transcriptionally distinct clusters of 24 different brain cell types. Travaglini measured the proportion of each cell cluster in each participant, then assessed how cell-type proportions changed as AD worsened as per ADNC score. He used scCODA, a Bayesian algorithm that accounts for differences in all cell types simultaneously and “knows” that all cellular fractions must add up to one (Büttner et al., 2021). “If the proportion of a cell population increases, then all others must decrease to compensate,” Travaglini explained.

So which cells changed as AD pathology got worse? The proportion of layer 2/3 excitatory neurons, as well as that of two groups of inhibitory interneurons, oligodendrocytes, and oligodendrocyte precursor cells, all declined. The proportion of microglia and astrocytes grew. This suggests that the former cell types may be particularly vulnerable to plaques and tangles, while the latter may respond to the pathology.

To learn about how pathological and transcriptional changes might relate to one another, Victoria Rachleff, also at the Allen Institute, first performed a quantitative neuropathology analysis on cortical slices from all 84 SEA-AD participants. She labeled the tissue slices with antibodies against the neuronal marker NeuN, the microglial protein Iba1, the astrocytic glial fibrillary acidic protein, as well as pathology markers 6E10 for Aβ, AT8 for p-tau, and antibodies against α-synuclein or TDP-43. Using a machine-learning algorithm, Rachleff divided images of the slices into cortical layers, and then calculated how much of each marker was in each layer.

The researchers computed z-scores for each neuropathology variable, comparing the degree of pathology in each person to the average of all participants. Grouping similar z-scores together revealed how and when cell types increased or decreased in number as AD progressed. This process ordered participants from controls to those with extensive pathology according to their ADNC score. With increasing AD pathology came fewer neurons, smaller nuclei, more Aβ and tau accumulation, and more microglia surrounding amyloid plaques (see image below).

Picture of Progression. Grouping similar quantitative neuropathology data (rows) per participant (columns) paints a spectrum of changes, generally going from controls (dark blue, top left) to low (light blue), intermediate (dark orange), then high (light orange) AD pathology. [Courtesy of Kyle Travaglini, Allen Institute.]

A particular subset of participants, who had high ADNC scores, caught the researchers’ attention. Compared to the others, these people's brain samples had yielded consistently fewer nuclei for isolation and snRNA sequencing. They also had less DNA available for transcription as detected by ATAC sequencing, which measures genome-wide chromatin accessibility. Travaglini does not think these shortcomings were technical issues, because these measurements tracked with the participants' neuropathology data. For example, a person who had low-quality RNA and lots of transcriptional repression also had fewer layer 3 cortical neurons and had declined faster on the memory portion of the Cognitive Abilities Screening test (see image below).

Who Faded Fast? The worse a brain sample's quality of postmortem RNA, the fewer cortical neurons it contained (left). Participants with bad RNA also had faster memory decline (right, red) than controls (gray), or people with similar AD pathology but more intact RNA upon isolation (blue). [Courtesy of Kyle Travaglini, Allen Institute.]

Next, the scientists excluded these severely affected participants and once again deployed scCODA to recalculate how the cell proportions changed with pathology for the remaining groups. The same neuronal loss showed up, but the glial alterations no longer did. “This suggests that there is an underlying process driving neuron loss in all AD participants, and a secondary, inflammatory process driving the glial changes and, perhaps, worse cognitive decline in the severely affected donors,” Travaglini said at the conference.

Travaglini is now analyzing other brain regions affected at different stages of AD, including the medial entorhinal, dorsolateral prefrontal, and V1 visual cortices.

Gilad Green, Hebrew University of Jerusalem, Israel, presented another snRNA-Seq atlas at AAIC. His encompassed 1.6 million cells from prefrontal cortex tissue of 478 ROSMAP participants spanning the clinical spectrum from healthy to AD. Their average age was 89 years.

Green used an algorithm to sort the cells into 96 populations based on their gene expression, then correlated each with the presence of amyloid plaques, tau tangles, and cognitive decline. In his analysis, all three outcomes were associated with absence of a subset of somatostatin-expressing inhibitory neurons and presence of disease-associated microglia expressing markers of the so-called DAM state, e.g., APOE, TREM2, and GPNMB. In contrast, a population of astrocytes Green called disease-associated (DAA) were linked to tangles and cognition but not to plaques. This suggests that neuron loss and a microglial response of the DAM type occurs before reactive astroglia of this “DAA” type appear. In other words, the data suggest these inhibitory neurons might have died before astrocytes got worked up.

Increasingly, scientists are characterizing diseased glia by their gene expression, rather than using the acronyms of earlier studies (Jul 2022 news).

For a more complete sense of cellular changes during AD pathogenesis, Green measured the proportion of each cell population within each participant, summed them into a composite value, and then figured out how these whole-brain cell populations were related to one another, and to AD pathology. For this he used a pseudotime algorithm. To find out what that is, and what Green discovered, see Part 19 of this AAIC series.

Besides generating new, ever-larger RNA-Seq datasets, scientists are also finding success in compiling existing datasets into one big pool for more powerful analyses (see Part 18 of this series).—Chelsea Weidman Burke


  1. It’s exciting that different groups see similar abundance and gene-expression changes in single-nucleus data from AD donors, especially with distinct patient cohorts and sampling strategies. It gives you more confidence that what we’re observing is robust and could lead to a greater understanding of the disease process. It’ll be important going forward to bring these datasets together to understand their similarities and differences more closely, and to increase our statistical power to identify molecular changes brought by AD.

    Like Evan Macosko, we do see specific L1 inhibitory interneurons supertypes falling out (Lamp5- and Sncg-expressing types), but these aren’t the most strongly affected (Sst and Pvalb types from Layers 2 and 3). This may be because it’s a different brain region (prefrontal cortex versus middle temporal gyrus), resolution of cell type clusters, or differences in sampling strategy.

  2. I’m very excited by the SEA-AD database. In general, I think the Alzheimer’s research field will benefit enormously from open-access datasets that can be queried by all researchers. The lack of such openly available data has been a huge challenge in this field, and I think it’s now really changing. We plan to follow SEA-AD’s lead and release our full dataset for use—and provide an easy-to-use tool for users to integrate their own single cell datasets into our full analysis.

    The field of human genetics exploded once it became clear that data could be analyzed in a cumulative, progressive manner. That is to say, as more and more data for a particular trait was generated, the data were sufficiently robust and easy to integrate that meta-analyses were possible, so more and more could be discovered as more data arrived. We didn’t know if this would be true for single-cell analysis studies—transcription is far noisier, and the data type is so new. But what our integrative analysis shows is that many of these disease effects—certain cell vulnerabilities and gene-expression patterns associated with disease—are in fact quite consistent across datasets generated by different labs, from different cohorts. It makes me very optimistic about the future of single-cell analysis as a tool for studying disease mechanisms.

Make a Comment

To make a comment you must login or register.


News Citations

  1. Don’t Name It: Glial States Confound Easy Labels
  2. Pseudotime Simulates Disease Progression from Case-Control RNA-Seq
  3. Meta-Analysis of RNA-Seq Data is Robust, Finds Key Changes in AD

Paper Citations

  1. . scCODA is a Bayesian model for compositional single-cell data analysis. Nat Commun. 2021 Nov 25;12(1):6876. PubMed.

External Citations

  1. SEA-AD website
  2. Adult Changes in Thought
  3. ROSMAP 

Further Reading

No Available Further Reading