Clusters moving for same dataset and code using tSNE

Question

[chilifan]

I am clustering data using Satija's Seurat package. Link: https://satijalab.org/seurat/pbmc3k_tutorial.html I am using this exact code and their data, and I generate the same plots as them. However, visualiizing my clusters with tSNE, they are positioned differently compared to the example on satijas website. I haven't though much about this, I've clustered some data after that and had the feeling that the cells in the separate clusters are still the same, and it doesn't really matter how they are positioned. But now I'm presenting this for a group of biologists, and they are going to ask why they are positioned differently. How should I explain this? And more important: does it matter at all to the end result?

Answer 1

[i.sudbery]

tSNE (or t-Distributed Stochastic Neighbour Embedding) is a stochastic technique (i.e based on randomness). We would no expect to get exactly the same result multiple times, unless we fix the random seed. Actaully, what you'll notice here is that mostly the bottom plot is pretty close to just being a rotation of the top one.

For a myriad of reasons, you shouldn't use tSNE for clustering. You derive your clusters by some other method, and then visualise the clusters using tSNE. Because most clustering algos are more stable than tSNE on repeated runs, you would expect the memberships of the clusters to be the same.

[chilifan]

Thank you, this makes perfect sense:) I am indeed using another method for clustering, so I conclude that I can trust the cell identities of my clusters then.