What's the point of integrating single cell data of different conditions?

Question

[ARP]

Hi all,

I have a conceptual question about this Seurat vignette.

In short, in the vignette they integrate single cell transcriptomics data from two different biological conditions stating that it will help with the following:

• Identify cell subpopulations that are present in both datasets
• Obtain cell type markers that are conserved in both control and stimulated cells
• Compare the datasets to find cell-type specific responses to stimulation

In my hands, celltype assignment works perfectly well using automated annotators such as SingleR no integration needed (at least with PBMC). And once I identify the cell-types, simply comparing the cells from the different conditions in each sub-populations would give me cell-type specific responses to stimulation. Especially for that last point, wouldn't integrating both conditions defeat the purpose of comparing them?

I totally get integrating different datasets to work around batch effect, but I'm really confused about integrating in this context. Can someone please explain the rationale of performing the analysis this way? Thanks!

Answer 1

[jared.andrews07]

Your thoughts are generally correct. There are a few other situations where integrating may be helpful - pseudotime and trajectory inference, RNA velocity (dubious as it may be), etc. Having consistent trajectories between conditions that you can then test between can be useful, see the tradeSeq vignette for some examples.

Mostly though, folks do it to annotate like with like based on cluster markers. Automated methods that function in a cell autonomous manner don't care about embeddings as you astutely point out. Those many of them also work on clusters (or over-clustered populations), and integration can be helpful there to again group like with like between samples.

Given that most integration methods function on the embeddings and don't adjust the underlying counts, they have no impact on differential expression between conditions.

So it just depends what you'd like to do downstream. For annotation, integration may not be necessary, though it can be convenient when automated methods don't do a great job and you're stuck doing it manually.