Entering edit mode
6 days ago
carolofharvest
▴
50
Hello ,
I want to use slingshot package for trajectory analysis but before doing that I need to subset my annotated data to remove some clusters. Should I re scale, run PCA and runUmap functions from Seurat package again ? I dont want to re cluster my data as some cells will definetely move to other clusters if I do clustering again.
Then don't do it? Not sure what you want to hear.
Should I use RunUMAP or RunPCA again after subsetting ? This is my question.
If you don't want to change embeddings, then no. If you do, then yes.
I am also not clear as to the root of your question.
The underlying problem is always the same in single-cell: After subsetting, do you ask yourself whether to re-featureSelect/PCA/UMAP/DimRed or not. If you do then things are more precise towards the remaining cells but coordinates and everything changes. Using variable genes from the previous analysis might introduce noise because genes that were variable before nolonger are, while others now might actually be more relevant. Sometimes celltype annotations might conflict because you see slightly different clustering that is not entirely in line with your previous annotation. However, if you keep the original coordinates and all you might see that in some parts of e.g. your UMAP things now look "odd" because the original continuum of cells is now broken and things like trajectories might not look convincing because the used embeddings (PCA or UMAP) were calculated with other cells in the mix that are nolonger present. I cannot give you a bulletproof answer. In scRNA-seq more than every other experiment type I've seen so far it "really depends" on the dataset and context. After all, it must look convincing and visually appealing in order to sell your narrative while not be scientifically or technically wrong. I know, not what you want to hear but "it depends". For me it always comes down with scRNA-seq that after all the typical analysis is done we sit and think about how we can present data and tailor analysis to emphasize the findings we think are supported by the data --- but tweak things to be most visually appealing. Paper writing after all is a lot of politics, so beside being scientifically correct and technically well done (many ways to do that) it must simply look good and catchy so the editor does not immediately desk-reject it.
Very true. Many high-impact scRNA studies conclude little but are presented beautifully. Everybody likes a pretty picture.