Single-cell dimensionality reduction and clustering

Question

[xingyu]

During the single-cell clustering process, the number of subpopulations varies at different resolutions. HOW to ensure that some cell subpopulations remain unchanged within a certain resolution while further dividing other subpopulations. This is because, after annotating with markers, I found that some subpopulations can be annotated, while others still exhibit significant heterogeneity.

[ATpoint]

Simply subcluster only a subset of cells or clusters, but not all. Nothign wrong with that. For example, you have a sample with stroma and immune cells, so you do a crude clustering first to separate both populations. Then you subcluster immune cells, then maybe split into myeloid and lymphoid, then only myeloid, then only monocytes etcetc. Just take the population you feel is underclustered and keep clustering this until you're satisfied. There is no rule or imperative to always take all cells. You take what you need to answer your scientifi question.

[Bastien Hervé]

Sometimes reviewers get picky when it comes to clustering relatively homogeneous population. If the sub population are already in the literature with specific marker genes then it will be OK, but fine tuning clustering on some variation of the same cell population (disease/control of the same cell type) can be troublesome to explain. They will think one is cherry picking.

[ATpoint]

Neither OP nor my example talks about homogeneous populations though.

[xingyu]

Thanks for your suggestion. I am trying now!

Initially, I used the clustree package to assist with cell clustering, but I found that a high resolution poses the risk of over-clustering. Therefore, I employed the ROGUE tool developed by Zhang Zemin's team to evaluate the purity of each cell subpopulation. Currently, I am trying it according to your suggestion.

Answer 1

[Bastien Hervé]

I usually set a high clustering resolution until I consider all populations have split, then I aggregate following a hierarchical clustering based of marker genes from the different populations. You can also get input from Silhouette scoring and Adjusted Rank Index (ARI)

[xingyu]

Thank you for your comments.

I have a small question: Does Seurat include functions for Silhouette scoring and Adjusted Rand Index (ARI)

[Bastien Hervé]

Not that I know of. The only mention I remember is this one

[yura.grabovska]

Do you have some dummy code you would be happy to share? I'm not OP but this sounds like a fairly good approach and I'd be curious to find out more

Answer 2

[Michael]

I am not sure if that is exactly what you want to investigate, but it sound like a Cluster tree plot could help with this. In R, run findClusters in Seurat at different resolutions, then do this:

library(clustree)
clustree([e-mail], prefix="RNA_snn30_res.")

[xingyu]

Thank for your reply.

Initially, I used the clustree package to assist with cell clustering, but I found that a high resolution poses the risk of over-clustering. Therefore, I employed the ROGUE tool developed by Zhang Zemin's team to evaluate the purity of each cell subpopulation. Extracting impure cell clusters separately might be a good approach to address the issues in my dataset. Currently, I am trying it according to ATpoint's suggestion.