Using Slingshot on a Cell Sub-population about slingshot HOT 1 CLOSED

Hey @hkarakurt8742 ,

That's a very good question and I'm not sure how well qualified I am to answer it, but I'll give it my best.

Regarding normalization, I don't think you would want to start from scratch here. The kinds of technical effects you generally seek to remove by normalization would probably be more apparent in all the data than it would in a subset. And for some normalization methods (such as global scaling methods, like CPM) there wouldn't be any difference.

Dimensionality reduction is a bit murkier. Thinking about PCA, you might get better results if you rerun on the subset because there may be large sources of variation in the overall data that aren't relevant in the subset and only serve to introduce noise. By that same token, however, there may be biologically important sources of variation that are less obvious in a subset of the data. I think I would lean toward re-running it, but I can't call that a strong recommendation.

Similarly, clustering is a bit unclear, but if you already have clusters on the full data that make sense to you, then I would say you probably don't need to re-run it, particularly because this would most likely lead to smaller clusters. Clustering for the purposes of trajectory inference is a bit different from the standard clustering for cell type inference. Specifically, I think you don't need as high a level of resolution for trajectory inference; if two cell types are adjacent along a lineage (and there's no branching point between them), then there is no real need for them to form separate clusters.

Hope this helps and I would be happy to hear your thoughts!

from slingshot.