How to account for cell type composition in differential expression analysis?

Question

[a.e.c]

Dear all,

I used xCELL to obtain cell type enrichments from my bulk RNA-seq samples but now I do not know how to handle this information for a differential expression analysis. I have been investigating and one option is using CellCODE but then I do not use the information extracted from xCELL. Any other suggestion?

Thanks,

Answer 1

[Samuel Brady]

I would take the standard deviation of each xCELL signature score (across samples) and plot a clustered heatmap of the 20-100 signatures with the highest standard deviation. Those signatures with low standard deviation represent cell types that don't vary between your samples and thus are not of interest to you. Those with the highest standard deviation vary dramatically across your samples and are of interest.

We did this with ssGSEA enrichment scores of C2 signatures (not xCELL) but it should work similarly for xCELL scores. We used this approach on a metastatic cancer patient whose samples we obtained from different parts of the body (RNA-Seq) and it correctly showed which samples came from which tissue and which samples were enriched in tumor vs. normal cells.

Answer 2

[a.e.c]

Thanks Samuel, but once I have the most variable cell types across my samples, how do I get to the gene level and find DE genes? You mention to plot a clustered heatmap with the top most variable signatures (=20-100 genes?) and then what? How do I know a gene is consistently DE within my groups?

[Samuel Brady]

Maybe I misunderstood your question. I thought you were look for signature scores (cell type scores) in order to determine the cell type composition of your RNA-Seq samples. What is the scientific question you are trying to address by doing a differential expression analysis of individual genes? Are you trying to see which individual genes are driving the differences in your cell type composition scores?

[a.e.c]

No, I would like to adjust for different cell type composition (enrichments) in my DE analysis. I have 3 groups to compare: controls, case1, case2.

[Samuel Brady]

I don't think one can adjust RNA-Seq data for cell type composition. For example, if you have tumor RNA-Seq data and you would like to subtract out the normal cell types, this is not reliably possible, as far as I know.