An embedding in obsm (e.g. X_umap, X_tsne, spatial).
Annotations in obs (cell types, clusters, samples…).
Optional: marker genes from sc.tl.rank_genes_groups (uns), colours as uns[<col>_colors].
Very large files (multi-GB): pack them first with scbpack and open the resulting .scb.
AnnData Browser
3
12
✕
What everything does
Two linked embeddings side by side. The left pane is coloured by an annotation,
the right by a gene (or a second annotation). Everything is computed in your browser.
Choosing what to look at
Colour left by
Annotation shown in the left pane — cell types, clusters, sample, any obs column.
Right pane
Switch the right pane between the selected gene and a second annotation.
Gene
Type to search; the right pane recolours. Expression comes from the current layer.
Highlight
Type a group name or value to dim everything else.
Layer
Pick X, raw or any stored layer. The second menu can apply a
log1p CP10k transform on the fly when the layer holds raw counts.
Colours
Colormap, an optional clip of the top 0.5% (stops one bright cell flattening the scale),
and the tissue image toggle for spatial data.
Moving around
Pan / Box / Lasso
Drag to pan and scroll to zoom, or draw a box or freehand loop to select cells.
Clicking the active tool turns it off: with none of the three lit the plot still pans, the selection is
cleared and you are back to the whole dataset. Shortcuts: p pan, b box,
l lasso.
Selecting
Shift-drag selects with any tool. ⌘/Ctrl-drag adds to the
selection, Alt-drag removes, Esc clears.
Touch
Drag with one finger to pan, pinch with two to zoom, double-tap to reset.
A single tap inspects the nearest cell, standing in for the hover tooltip. Box and lasso
selection work by dragging once the matching tool is active.
Sync zoom
Keeps both panes on the same view. Double-click a plot, or press Reset view, to zoom back out.
Z / Labels / Point size
Draw order for overplotted points, on-plot group labels, and point and label sizing.
Legend
Shows cell counts per group. Click a label to hide or show it, double-click to show only
that group, click the swatch to change its colour. Clicking the last group still visible — or
double-clicking the one already on its own — brings every group back, so you can always get to the
full dataset. A ▸ marks groups with marker genes — click it for the top ones.
Hover
Point tooltips give the cell's annotations and the current gene's value.
Panels
Selection & differential expression
One-vs-rest or group-vs-group tests — selected cells
or any annotation group, against the rest or another group. Ranked table plus a volcano plot, with all
genes downloadable as CSV. Selections can be exported as cell IDs or saved as a new annotation column.
Running a test reads every gene, so it is offered for files you open yourself; on a dataset served over
the network the rest of the panel still works and the test is disabled.
Composition
Stacked bars of one annotation within another, as proportions or counts.
Dot plot / Heatmap
Top markers per group or your own gene list. Many datasets already carry
marker genes, computed when the file was packed. Compute markers recomputes them in the browser
for any annotation — available when you open a local file.
Violin
Gene expression or any numeric column, split by an annotation.
Cell counts
Bar chart of cells per group, sorted by count or name.
Gene vs gene
Scatter of two genes with a correlation, points coloured by the left pane.
Gene set score
Mean expression of a list of genes, shown in the right pane.
Group notes
Free text per group — useful while annotating. Exportable.
How the statistics work
Differential expression
A Wilcoxon rank-sum test of group A against group B, or against
every other cell when no second group is given. Non-detected cells are kept rather than filtered, and
all share the lowest ranks; the variance is tie-corrected. The p-value is two-sided from the normal
approximation and is adjusted across genes by the Benjamini–Hochberg procedure. A Welch
t statistic is reported beside it for reference.
Log fold change
log2 of the ratio of group means taken on the de-logged scale
(expm1), with a small floor added to each mean so that genes absent from one group stay
finite. Computed from the values as currently transformed.
% detected
Fraction of cells in the group with a non-zero value for that gene.
Top markers
Per group, genes with a positive log fold change ranked by the Wilcoxon statistic,
keeping the strongest few. Precomputed markers stored in the file use this same definition, so they
match what Compute markers would produce.
Anything that ranks all genes — differential expression, marker finding —
has to read the entire matrix. On a file opened from your own disk that is a quick pass over memory. For a
dataset served over the network it would mean one request per gene, so those operations are offered only for
local files; hosted datasets use the markers computed when they were packed. Everything driven by a gene you
name — plots, violins, dot plots, gene set scores — reads only those genes and works either way.
Getting things out
⬇ Left / ⬇ Right
Download either embedding. Every panel chart also exports as SVG.
Save view / Load view
Writes a small JSON file with the colours, gene, zoom and panel state so you
can return to exactly this view, or send it to someone with the same dataset.
Export annotations
Annotations you create from selections come out as JSON. Merge them back into
the original file with scbpack apply-annotations.
Nothing you open is uploaded — a file you choose is read locally in your
browser, and a hosted dataset is streamed read-only from its server.