Differential Expression

The Differential expression tool compares gene expression between two populations of cells that you define — Group A versus Group B — and returns a ranked table of the genes that differ, with a volcano plot and per-gene violin plots. Unlike the quick violin plot for a selection, this runs a real statistical test (Wilcoxon rank-sum) on the server and reports fold changes, effect sizes, and adjusted p-values.

Opening the tool

The Differential expression builder in the Tools tab

Open the tool in either of two places:

The Tools menu at the top of the window has a Differential expression… item.
The Tools tab in the left sidebar has a Compare two populations… button under its Differential expression heading.

Either one opens the builder, which replaces the Tools tab’s normal contents while it is open; close it with the × in its header to return to the other tools.

Choosing what to compare

At the top of the builder, the Compare by menu picks the metadata field whose values define your groups. Any categorical field can be used, including custom annotations you have made yourself (see Custom Annotations and Accounts). Switching the field recolors the map by that field and clears any groups you had started.

Building the two groups

Below Compare by is a list of the chosen field’s values, each with its cell count. To assign a value to a group:

Choose the target group with the Add to: Group A / Group B toggle.
Click values in the list to add them to that group. Assigned values are tinted and marked A or B; click again to remove.

The list is sorted by cell count (largest first) and has a filter box, which is handy for fields with many values. Your current groups, with their cell counts, are shown in the Group A and Group B cards below the list.

Group B has two modes:

All other cells (the default) runs a one-vs-rest comparison: Group A against every other cell in the dataset. This is the classic “marker genes” test.
Pick cell types lets you choose specific values for Group B, the same way as Group A.

Each group needs at least 25 cells for a stable test.

Restricting a group with a filter

Each group can be narrowed by a second metadata field. For example, to compare male versus female cells within a single cell type, set Compare by to sex, put the sexes in Groups A and B, and set both groups’ filter to that cell type. The filter accepts more than one value, so you can also, say, restrict a comparison to a chosen set of donors.

Test settings

Expand Test settings to adjust:

Min. log₂ fold change and Adjusted p cutoff — the thresholds for calling a gene significant. They set which genes are counted in the results summary and listed in the table.
Min. fraction expressing — ignore genes detected in only a tiny fraction of both groups.

Mitochondrial, ribosomal, and hemoglobin genes, and genes detected in very few cells, are filtered out before the test, so the multiple-testing correction reflects the genes actually reported.

Running a comparison

Click Run comparison. A progress overlay shows the current step and, during the test itself, how many genes have been processed. Use Cancel to stop a run; this halts the job on the server, not just in your browser.

If no compute backend is configured for the site, the tool still runs with placeholder statistics over the dataset’s real genes, so the interface can be demonstrated.

Reading the results

The differential expression results pop-up

Results open in a pop-up over the plot with three parts:

A table of genes, sortable by any column: log₂ fold change, AUC (an effect size from 0.5 for no difference to 1 for always higher in Group A), adjusted p-value, and the mean expression and fraction expressing in each group. A search box and the All / Up in A / Up in B buttons narrow the list.
A volcano plot (fold change versus significance); the Volcano / MA toggle switches to an MA plot (fold change versus average expression).
A violin plot of the selected gene’s expression in Group A versus Group B.

Click any gene — in the table or on the plot — to highlight it and show its violin. Download CSV saves the full table; the file begins with a short header recording the groups, the filters applied, and the significance thresholds, so a downloaded result is self-describing.

Note

The test treats each cell as an independent observation. With thousands of cells this makes the p-values anti-conservative (an effect known as pseudoreplication), so nearly everything can look “significant.” Read the adjusted p-value as a ranking rather than an exact cutoff, and lean on the fold change and AUC to judge which differences matter.