Cluster sub-set analysis using Seurat

Question

[singcell]

I am trying to dig deeper into my Seurat single-cell data analysis. If I want to further sub-cluster a big cluster then what would be the best way to do it:

1) Decreasing the resolution at FindClusters stage

or

2) extracting the individual cell index and re-clustering and then further analysis.

What would be the major difference in outcome between two approaches? How to perfrom the second approach in Seurat? using Whichcells function to extract and then I am not sure about next step please

Thank you.

[shoujun.gu]

It doesn't matter, since the raw count table are the same either way. The real important thing is the algo you use for sub-clustering. Modularity based clustering may not work good. Maybe gaussian mixture will be better?

[igor]

the raw count table are the same either way

Clustering is not based on the raw counts table (in the default Seurat workflow, at least). It's based on the reduced dimensions matrix which will change when you remove some of the sources of variation.

Answer 1

[swbarnes2]

Subsetting and rerunning the umap coords and clustering will separate out different things better.

Personally, I cheat, and use Loupe. It's easy to use the mouse to draw categories, then export your categories into the metadata of your seurat object. But you have to have your data go through the 10xGenomics pipeline for that.

[shoujun.gu]

Never perform the analysis geometrically.

[igor]

What is a geometric analysis?

[shoujun.gu]

Like "use the mouse to draw categories", or make conclusions based on your eyes or based on dimension reduced visualizations, etc. In my opinion, all analysis should be done analytically (mathmatically).

[igor]

For scRNA-seq, there has to be substantial manual supervision to guide the analysis and make biologically meaningful conclusions. An exploratory analysis just to help you understand your data does not need to be perfect.

[shoujun.gu]

In general, I don't quite agree with "manual supervision". All judgement should be based on some mathematical algo. If the results are not "biologically meaningful", it means that algo used is problematic here. Theoretically, researcher should improve or replace algo instead of applying "manual supervision" on the same algo.

Answer 2

[igor]

Both approaches are valid. It really depends on your question. You can increase the resolution to get more clusters. You can just select the cluster of interest and analyze it more closely or in different ways. For example, if your population of interest is only 10% of the total cells, it would be difficult to analyze it in the full dataset. If you just keep those cells, that becomes easier.

If you want to extract the cells of a specific cluster, you can use subset(). The Interaction Tips vignette has a few examples such as:

# Can I create a Seurat object of just the NK cells and B cells?
subset(pbmc, idents = c("NK", "B"))

[singcell]

Thank you for your reply. I have got integrated Seurat object containing KO and WT samples. Can I still use subset() function? Can I still differentiate between WT and KO samples during downstream analysis after subsetting.

[igor]

Subsetting just keeps/removes specific cells. It does not alter their metadata.

Answer 3

[j.gleixner]

There now is also Seurat::FindSubCluster

Answer 4

[ineslagobaldaia]

I have a follow-up question:

How can I define a new cluster based on a feature's expression level, but maintain it in the same object, along side the other clusters that were already defined by clustering?

Thanks!