Differential expression analysis in Seurat

Question

[Lucy]

Hi,

I would like to get some clarification on whether I should be using the RNA assay or the SCT assay for differential expression analysis in Seurat. I have seen several issues on the GitHub and FAQ 4, however these usually refer to data that has been integrated using the Seurat workflow. In my case, I have not performed integration so have an RNA and SCT assay only. I have used Harmony for batch correction.

The choice of assay seems to make a large difference to the number of differentially expressed genes. I am using the logistic regression (LR) option of FindMarkers() with "donor" as a latent variable. In one example, using the SCT assay I get 372 positive cluster markers, compared with 690 positive cluster markers using the RNA assay. The genes identified using the SCT assay are also identified by the RNA assay, but I just get extra genes with the RNA assay.

Why would the RNA assay give me so many more genes? And which assay should I trust?

When plotting gene expression on the UMAP and in violin plots, am I ok to use the SCT assay, or should I again be using RNA?

Many thanks for helping to clear up this confusion!

Best wishes,

Lucy

Answer 1

[rpolicastro]

In the Seurat FAQs section 4 they recommend running differential expression on the RNA assay after using the older normalization workflow. Emphasis mine.

1. I’ve run an integration analysis and now want to perform a differential expression analysis. Which Assay should I use?

   We recommend running your differential expression tests on the “unintegrated” data. By default this is stored in the “RNA” Assay. There are several reasons for this.

   The integration procedure inherently introduces dependencies between data points. This violates the assumptions of the statistical tests used for differential expression.

   We recommend performing differential expression on the RNA assay, after normalization.

   If you have used the SCTransform workflow for normalization and integration, we still suggest performing differential expression on the RNA assay.

For plotting I would still use the normalized RNA assay slot.

[Lucy]

Thanks, I saw this, but their question refers to integration analysis and discusses other effects specific to the integration workflow, so I wasn't sure whether this was also relevant for non-integrated data

[rpolicastro]

This applies for any usage of SCTransform.