SEA-AD Multiregion Gene Expression Viewer

The SEA-AD Multiregion Gene Expression Viewer is an interactive R Shiny application for exploring pseudobulk gene expression and the fraction of nuclei expressing each gene across brain regions, cell types, donors, and Alzheimer's disease progression. Identify candidate genes in a sortable statistics table, retrieve selected genes from public AWS storage, apply simultaneous metadata filters, create five types of plots, and share configured analyses through bookmarked URLs.

The viewer uses data from the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD), with gene-specific links into the Allen Brain Cell Atlas and GeneCards.

Launch the interactive viewer Go to GitHub

Getting started

Explore what the viewer can do through examples that revisit findings from our two Alzheimer’s disease studies, starting with the most recent. Each example opens a gene with its filters and plot settings already configured.

Examples from the multiregion paper (Travaglini et al., 2026)

Examine cell-type-selective gene expression and disease-associated changes across brain regions. The first examples revisit genes selected for RNAscope validation of hippocampal and entorhinal cortical cell types in Figure 1.

THSD4: CA1 neurons (Figure 1g)

Explore selective THSD4 expression in CA1 neurons within the hippocampus (HIP).

Show this example

GHSR: CA2 neurons (Figure 1g)

Explore selective GHSR expression in CA2 neurons within the hippocampus (HIP).

Show this example

RORB: EC IT_2 neurons (Figure 1h)

Examine RORB expression in EC IT_2 neurons within the medial entorhinal cortex (MEC).

Show this example

RELN: EC IT_3 neurons (Figure 1h)

Compare RELN expression in EC IT_3 neurons across the medial and lateral entorhinal cortex (MEC and LEC).

Show this example

The following examples revisit Figure 6, comparing gene expression in vulnerable and relatively protected SST cell types and examining changes with Alzheimer’s disease progression.

MME: changes in vulnerable SST types (Figure 6e)

Compare the decline in MME expression with Alzheimer’s disease in vulnerable SST cell types with its much lower expression in relatively unaffected SST types.

Show this example

HCN1: baseline differences in SST types (Figure 6g)

Compare HCN1 expression across SST cell types, with higher baseline expression in vulnerable types.

Show this example

SST: declining expression across regions (Figure 6i)

Explore decreasing SST expression with Alzheimer’s disease progression across SST cell types and brain regions.

Show this example

Examples from the MTG paper (Gabitto et al., 2024)

Extend findings from the middle temporal gyrus (MTG) to additional brain regions. These examples illustrate a pattern that is consistent across regions for APOE and a more region-specific pattern for OLIG2.

APOE: contrasting glial responses (Figure 6k)

Explore increasing APOE expression in microglia and decreasing expression in astrocytes with Alzheimer’s disease progression, extending the MTG findings across brain regions.

Show this example

OLIG2: region-specific changes in OPCs (Figure 7g)

Examine OLIG2 in oligodendrocyte precursor cells (OPCs), where it is most highly expressed. Its modest increase with Alzheimer’s disease in temporal cortex is not seen across the other brain regions.

Show this example

Other examples

Explore additional striking gene-expression patterns across cell types, brain regions, and disease progression.

PTPRG: a strong microglial progression pattern

Both studies report that Alzheimer’s disease-associated microglia emerge early and increase more sharply later in disease. Explore the sharp increase in PTPRG expression in Micro-PVM_3-SEAAD with the global continuous pseudoprogression score (CPS), comparing trajectories across brain regions.

Show this example

Data

Scientific source

The viewer uses the multiregion SEA-AD data resource associated with:

Travaglini, K. J., Gabitto, M. I., Ding, Y., et al. Multiregional single-cell profiling reveals shared and specialized cellular vulnerability in Alzheimer’s disease. bioRxiv (2026), preprint, not peer reviewed. View the publication DOI.

The study extends SEA-AD across ten neo- and allocortical regions and relates cell-type changes to regional and global Alzheimer’s disease pathological burden. See the SEA-AD website for data and resource access.

Pseudobulk expression

The source pseudobulk AnnData (.h5ad) objects contain summed counts by gene symbol. Each row is an aggregated observation with cell-type, donor, brain-region, and study annotations; Number of nuclei records the number of contributing nuclei. Gene partitions also contain the number of nuclei expressing each gene, used to calculate detection fractions.

Pseudobulk observations are aggregates, not individual nuclei. Multiple observations may come from the same donor.

Gene-level Parquet files

Expression is stored under counts_by_gene/gene=<gene_key>/. Each partition supplies sample_id, summed_counts, and a detection-count column named number_of_cells_expressing or number_of_nuclei_expressing. The current app requires detection counts and reads only the selected gene partition, joining it to metadata by sample_id.

The app displays counts per ten thousand total counts (CP10K), using the stored CPM scaling factor:

CP10K = summed_counts × (CPM_scaling_factor / 100)
fraction expressing = number_of_cells_expressing / Number of nuclei

Expression can be displayed as CP10K or ln(CP10K+1). This is library-size-normalized expression, not counts per 10,000 nuclei. For grouped dot plots, the fraction is calculated by summing expressing-nucleus counts and total nuclei before division.

Metadata and value sets

obs_metadata.parquet contains unique sample_id values and cell taxonomy, brain region, donor, study, clinical, neuropathology, and nucleus-count annotations.
gene_map.parquet maps gene_symbol to partition keys in gene_column.
CPM_scaling.csv supplies scaling factors in the same row and library order as observation metadata.
global_cps_multiregion.csv supplies donor-level CPS_Global, CPS_Global_ABeta, and CPS_Global_pTau, joined from donor_id to Donor ID. These remain continuous numeric values. Unmatched scores remain missing, not zero; observations without a finite selected score are excluded from trajectory plots.
value_sets.csv supplies categorical labels, order, and colors; unlisted values are retained with fallback colors.
Local table_column_definitions.csv supplies table-header descriptions, and abc_atlas_gene_urls.csv.gz supplies precomputed gene-specific ABC Atlas links. Both files are required at startup.
Optional www/tokens.csv, with token, type, and name columns, supplies the searchable glossary.

The app also derives a dementia grouping from Cognitive Status for exploratory comparisons.

AWS storage and access

The app anonymously reads metadata, CPS scores, value sets, and gene statistics at startup, then retrieves and caches requested gene partitions. Runtime AWS credentials are not required. The configured bucket is sea-ad-single-cell-profiling in us-west-2, under Multiregion_2026/previous_objects/parquet/Multiregion_2026. Browse the public Parquet resources.

Using the viewer

Left control panel

The left panel controls gene retrieval, stacked filtering, expression scaling, and plot configuration. Filter and plot controls appear after the first gene has loaded.

Gene

Type a gene symbol or select a row in the statistics table.
Select Get gene data.
The app retrieves the partition, joins metadata, calculates CP10K and fraction expressing, and caches the result.
Use Plot in ABC Atlas or GeneCards to explore the loaded gene in another resource.

The selector starts empty with APOE as an example when available. Filter and plotting controls appear after a gene is loaded.

Filter and scale

Filter metadata adds simultaneous filters. Categorical selections within a field use OR; filters across fields use AND. Numeric fields, including CPS scores, use range sliders.
Discard observations with zero counts removes zero-count observations for the selected gene. In two-gene correlation, both genes must have positive counts when this option is selected.
Under Additional plot options, Gene metric switches supported plots between expression and fraction expressing. Dot plots always use expression for color and fraction expressing for size.
Plot ln(CP10K+1) toggles transformed expression for trajectory, heatmap, dot, and distribution plots. Expression-mode two-gene correlations always use ln(CP10K+1).
Minimum Expr/Maximum Expr, or Minimum Frac/Maximum Frac, set display limits. Suggested limits update when data or relevant controls change; shared views restore their saved limits. There is no automatic-limits checkbox in the current interface.

Plot controls

Plot type selects one of five visualizations.
Plot-specific controls configure dimensions, facets, colors, trend lines, pagination, plot height, counts, and observation display. Hover over dropdown labels for concise control definitions.
Generate plot captures the current settings, creates the plot, and opens the plotting tab.
Share this view creates a URL bookmark containing the selected gene or genes, active filters, and plot configuration. Opening the link restores the state, retrieves the required gene data, and regenerates the plot.
Reset plot options restores defaults.
The optional alphabetical glossary below the plot controls supports exact lookup of abbreviations used in cell-type and anatomical labels.

Application tabs

Select gene of interest

This tab contains a searchable, sortable gene-statistics table with numeric range filters and header tooltips. Selecting a row fills the Gene symbol selector; click Get gene data to retrieve it. GeneCards is accessed through the sidebar after loading a gene, not through table hyperlinks.

The columns are gene_symbol, ln(mean_CP10K+1), frac_expr, dementia_log2fc, dementia_frac_diff, max_ROI, log2fc_ROI, max_frac_ROI, frac_diff_ROI, max_subclass, log2fc_subclass, max_frac_subclass, frac_diff_subclass, and gene_type. The app converts the source table’s log mean CPM column to log mean CP10K for display.

Exact-value multi-select filters below the table cover gene_symbol, max_ROI, max_subclass, and gene_type. Within-field selections use OR; across-field selections use AND.

Important: These precomputed summaries guide exploration and do not replace the source study’s analyses.

Plot gene of interest

The plotting tab supports five visualization types with matching plot-specific statistical summaries.

Trajectory across disease

Plot expression or fraction expressing against one of three continuous donor-level scores: CPS_Global (default), CPS_Global_ABeta, or CPS_Global_pTau. These are the available trajectory axes; developmental ages and other metadata are not progression-axis choices.

Facet by separates eligible cell-type or brain-region groups, or shows all data together.
Color by overlays an eligible metadata field or uses one color. Facet and color choices are updated after filtering, with a 50-value limit for these controls.
Trend line supports linear fit (default), LOESS, or no fit. Show 95% confidence (error of fit) toggles confidence bands.
Toggle individual observations and adjust facets per row, height per facet, facets per page, and page number. Tall plots scroll vertically.
Eligible facets receive a Spearman monotonic-trend test, with BH correction across eligible facets. Tests require at least five observations, three progression values, and variation in the selected metric.

Heatmap of mean expression

The heatmap displays mean expression or mean fraction expressing for every occupied combination of two dimensions. Heatmap and dot plots default to Subclass on the horizontal axis and Brain Region on the vertical axis.

Tile color represents the selected gene metric.
Show number of observations optionally labels each tile with its contributing observation count.
The selected minimum and maximum apply to the fill scale.
A sparse rook-neighbor Moran's I test with 999 permutations summarizes autocorrelation of the displayed metric across the current ordered matrix.

Dot plot of mean expression

The dot plot uses the same two-dimensional grouping as the heatmap. Dot color represents mean expression, while dot area represents the weighted fraction of cells expressing the gene. Min. dot size and Max. dot size control the displayed size range without changing the underlying proportions; combinations with no data remain absent. The note below the plot reports separate sparse rook-neighbor Moran's I tests with 999 permutations for expression and fraction expressing.

Violin plot with observations

The distribution plot shows expression or fraction-expressing distributions across horizontal-axis groups and second-dimension facets.

Distribution display switches between violin and boxplot representations.
Show individual observations toggles the overlaid pseudobulk points; white points and error bars show means and standard errors.
Groups with fewer than two observations are omitted.
Facets per row, Height per facet, Facets per page, and Page Number limit the size of each rendered image while preserving the complete filtered analysis.
Each eligible facet receives a Kruskal-Wallis test across horizontal-axis groups using the displayed metric; facet p-values are BH corrected across all eligible facets, not only the current page.

Two-gene correlation

This plot compares two separately retrieved genes across matched sample_id values after applying active filters. The selected metric can be ln(CP10K+1) or fraction expressing.

Use Get 2nd gene data to retrieve the comparison gene through the same lazy reader and cache.
Points can be colored by metadata and are sized by log10(Number of nuclei) when cell counts are available.
Min. dot size and Max. dot size adjust the correlation-point size range.
Shared automatic or manual limits apply equally to both axes.
The plot is square with equal axis spacing.
An optional orthogonal total-least-squares line is invariant to swapping the x and y genes.
Expression mode reports Pearson R; fraction-expressing mode reports Spearman rho. The corresponding p-value appears in the plot title.

Plot-specific exploratory statistics

Trajectory: per-facet Spearman rho for monotonic progression in expression or fraction expressing, with BH correction across eligible facets.
Heatmap: sparse rook-neighbor Moran's I with a one-sided 999-permutation p-value for autocorrelation of the displayed metric.
Dot: separate sparse rook-neighbor Moran's I tests for mean expression and fraction expressing.
Violin or boxplot: per-facet Kruskal-Wallis tests of the displayed metric across horizontal-axis groups, with BH correction.
Two-gene correlation: Pearson correlation for expression and Spearman correlation for fraction expressing.

Interpretation: These statistics describe the currently selected gene, filters, dimensions, and displayed observations. Interactive testing across many genes and configurations is exploratory and introduces multiple-comparison considerations beyond within-plot BH corrections.

Running the app

Most users should use the hosted viewer. The repository also supports local development and reproducible deployment.

Requirements

R and the packages shiny, bslib, arrow, dplyr, ggplot2, scales, shinycssloaders, cachem, DT, and jsonlite;
S3 support in Arrow and internet access to the public data;
The required local table-definition and precomputed ABC Atlas URL files, plus the app’s www/ assets.

Use the project dependency records rather than the mouse viewer’s R version or another application’s lockfile.

Local setup

Download or clone the multiregion_gene_exploration repository.
Open the Shiny application directory containing app.R, table_column_definitions.csv, abc_atlas_gene_urls.csv.gz, and www/.
Restore the project’s dependencies from its renv.lock when using the renv-based environment.

install.packages("renv")
renv::restore()

Run the app

From the directory containing app.R:

shiny::runApp(".")

The hosted viewer is deployed to shinyapps.io. See the Rebuilding and Deployment Guide for manual deployment and GitHub Actions configuration.

Repository structure

The landing page follows the mouse repository’s code/shiny/ asset layout. The application bundle contains:

code/shiny/
    app.R
    table_column_definitions.csv
    abc_atlas_gene_urls.csv.gz
    renv.lock
    www/
        app_logo.png
        allen_institute_logo.svg
        tokens.csv (optional)
index.html
deployment_documentation.html
.github/workflows/deploy.yml

Large pseudobulk sources, gene partitions, metadata, CPS scores, value sets, and summary statistics are stored externally rather than bundled with the Shiny application.

Citation

If you use the viewer or underlying multiregion data, cite:

Travaglini, K. J., Gabitto, M. I., Ding, Y., et al. Multiregional single-cell profiling reveals shared and specialized cellular vulnerability in Alzheimer’s disease. bioRxiv (2026), preprint, not peer reviewed. View the publication DOI.

For the original SEA-AD atlas, see Gabitto, M. I., Travaglini, K. J., Rachleff, V. M., et al. Integrated multimodal cell atlas of Alzheimer’s disease. Nature Neuroscience (2024). Read the original atlas paper.

Please report the viewer version or Git commit and the data access date.

Feedback and support

Questions, bug reports, and suggestions are welcome. Open a GitHub issue or contact Jeremy Miller.

How the app was made

The Rebuilding and Deployment Guide documents the pseudobulk-to-Parquet workflow, detection counts, scaling factors, gene statistics, value sets, donor-level CPS joins, precomputed ABC Atlas gene links, AWS publication, local testing, shinyapps.io deployment, and GitHub Actions configuration.

Acknowledgements

We thank the research participants, tissue donors, families, and SEA-AD teams responsible for data generation, neuropathology, taxonomy, analysis, data access, and public resources. The app was developed by Jeremy Miller with coding assistance from Microsoft Copilot; beta testers include Irika Sinha and Kyle Travaglini.

SEA-AD Multiregion Viewer

SEA-AD Multiregion Gene Expression Viewer

Getting started

Examples from the multiregion paper (Travaglini et al., 2026)

THSD4: CA1 neurons (Figure 1g)

GHSR: CA2 neurons (Figure 1g)

RORB: EC IT_2 neurons (Figure 1h)

RELN: EC IT_3 neurons (Figure 1h)

MME: changes in vulnerable SST types (Figure 6e)

HCN1: baseline differences in SST types (Figure 6g)

SST: declining expression across regions (Figure 6i)

Examples from the MTG paper (Gabitto et al., 2024)

APOE: contrasting glial responses (Figure 6k)

OLIG2: region-specific changes in OPCs (Figure 7g)

Other examples

PTPRG: a strong microglial progression pattern

Data

Scientific source

Pseudobulk expression

Gene-level Parquet files

Metadata and value sets

AWS storage and access

Using the viewer

Left control panel

Gene

Filter and scale

Plot controls

Application tabs

Select gene of interest

Plot gene of interest

Trajectory across disease

Heatmap of mean expression

Dot plot of mean expression

Violin plot with observations

Two-gene correlation

Plot-specific exploratory statistics

Running the app

Requirements

Local setup

Run the app

Repository structure

Citation

Feedback and support

How the app was made

Acknowledgements