Seurat Find Marker output

[rasing02]

Query.on.Seurat.Find.Markers.Output.docx

[rharao]

For the benefit of other readers, I have OCR'd your attachment into Markdown.

My questions are based on Seurat output pasted below:

How is the average expression calculated?
In this case, we observe that the percentage of cells expressing the disease marker is 27%, while the healthy group shows only 0.14%. Does the average expression reflect only the cells that are expressing the marker (i.e., the 27% and 0.14% of cells), or is it calculated based on all cells—both expressing and non-expressing—and then averaged over the total number of cells in the group?

Is uneven cell distribution accounted for in these calculations?
Does the analysis consider potential variations in cell distribution across the groups, or is this factor ignored in the calculation of average expression?

Are Log2 fold changes (Log2FC) driven by differences in the percentage of cells expressing the marker?
If so, then wouldn't the observed changes in fold change (FC) essentially reflect differences in the percentage of cells expressing the marker? In that case, why would we need to focus on FC values instead of directly examining the %Cell values?

[rharao]

How is the average expression calculated?
In this case, we observe that the percentage of cells expressing the disease marker is 27%, while the healthy group shows only 0.14%. Does the average expression reflect only the cells that are expressing the marker (i.e., the 27% and 0.14% of cells), or is it calculated based on all cells—both expressing and non-expressing—and then averaged over the total number of cells in the group?

This depends on how you obtained the "Average Expression" values. They don't seem to match the log2FC provided. (Note: 27% and 14%, not 27% and 0.14%)

Is uneven cell distribution accounted for in these calculations?

By cell distribution, do you mean differences in cell number between groups? If so, then yes, these are averages and percents, not sums and counts.

Are Log2 fold changes (Log2FC) driven by differences in the percentage of cells expressing the marker?

Yes, definitely, if those differences exist. In this case, the difference in non-zero expression rate between the two groups is not as extreme as you seem to have believed.

[rasing02]

Thank you for your response.

1. Regarding How is the average expression calculated? - Apologies for typing error, its 27% and 14%. I just used this output as an example.
   In general I wanted to know when we follow seurat pipeline (https://satijalab.org/seurat/articles/sctransform_vignette.html) for analysis with default settings, then what does average expression represent? Is it based on expression profile of only the cells that express the genes? or all cells? do we include all zero values?

2. Regarding uneven cell distribution accounted for in these calculations? - Yes I mean differences in cell numbers between groups.

3. Regarding Are Log2 fold changes (Log2FC) driven by differences in the percentage of cells expressing the marker? -
   a) If fold change (FC) essentially reflect differences in the percentage of cells expressing the gene, then why we calculate FC values instead of directly examining the %Cell values.
   b) Also is it possible that for gene X - in group A only 20% cells express the gene, and in group B has 60% cells express the gene, but average expression level for group A is higher than group B

Thank you for your patience in addressing these questions.

[rharao]

what does average expression represent? Is it based on expression profile of only the cells that express the genes?

Absent further information, I will assume that you're referring to average expression values from Seurat::AverageExpression. All cells in the group, including those with zero expression for a feature, are pseudobulked for the average expression. This makes sense if you believe that a group of cells with almost no transcripts of a particular feature should have an average expression for that feature close to zero.

If fold change (FC) essentially reflect differences in the percentage of cells expressing the gene, then why we calculate FC values instead of directly examining the %Cell values.

It's the same reason that we incorporate UMIs in cell barcode oligos. Some cells can have relatively less or more of a transcript than other comparable cells, and these differences can be biologically informative.

Also is it possible that for gene X - in group A only 20% cells express the gene, and in group B has 60% cells express the gene, but average expression level for group A is higher than group B

Try it: can you think of two positive values $X_A$ and $X_B$ such that $0.2 \cdot X_A > 0.6 \cdot X_B$?

[rasing02]

Thank you for taking the time to address my queries. I appreciate your help.

To clarify the last question, here's the situation:

Group A has the following expression values: 5, 5, 5, 5, 5, 0, 0, 0, 0, 0.
In this group, 50% of the cells express the gene (5 cells).
The average expression level for the cells that express the gene is 2.5.

Group B has the following expression values: 10, 10, 10, 0, 0, 0, 0, 0, 0, 0.
In this group, 30% of the cells express the gene (3 cells).
The average expression level for the cells that express the gene is 3.

Maybe I oversimplified this, but I just wanted to know if this scenario is possible, where a smaller percentage of cells express the gene at a higher level. Usually, we don’t see this.

Anyway, I truly appreciate your time and effort in answering all of my questions.

[rharao]

Yes, it is possible.

The average expression level for the cells that express the gene is 2.5.
The average expression level for the cells that express the gene is 3.

No, the average expression levels for all cells in the groups are 2.5 for A and 3 for B in the scenario you've described. The averages only for cells with non-zero expression are 5 for A and 10 for B. $0.5 \cdot 5 = 2.5 < 0.3 \cdot 10 = 3$.

Where the groups have specific and useful biological definition, the whole-group averages are generally more informative than the averages among only the cells that have non-zero expression.

Usually, we don't see this.

It depends on your definition of "usually" and how much consideration you give that. I wouldn't take it as an impossibility. Try to imagine the following cellular development trajectory:

• In State 1, cells transcribe a moderate amount of a feature X.
• In State 2, downstream of a microenvironmental change, a certain transcription factor is activated in the cells, dramatically increasing each cell's transcription of X as well as another feature Y. This state lasts only for a short time.
• In State 3, the protein product Y inhibits the transcription of X. Transcription of Y is not moderated by much. Cells in State 3 outnumber those in State 2.

Now imagine that, in a certain clustering scheme, Group A contains State 1 and Group B contains State 2 and State 3. You'll see that it's possible that, compared to Group A, fewer cells in Group B have X transcripts, but the average expression of X within all of Group B is greater.

[rasing02]

Thank you for your response; it has been very helpful. So, Seurat pipeline includes all cells (zero and non-zero values) when calculating fold changes and average expression.
Thank you again for your time and assistance!