UMAP or TSNE for single cells?

Question

[synat.keam]

Dear all,

I am just slowly getting into single cells. I have been working with my data for 2.5 months now and tried to learn all the tips about single cell good practices.

I personally find UMAP is tricky especially when we are trying to fine-tune to get a good looking UMAP. I find TNSE is easier to produce consistent 2D visualization. Just would like to ask for some opinion whether you guys think TSNE is now popular for cluster visualization compared with UMAP. I am thinking of going forward with TSNE. Hope you can give some opinion.

Thanks,
Synat

[cfos4698]

Some food for thought: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011288

[Mensur Dlakic]

They talk in that paper as if it is unclear whether t-SNE and UMAP preserve global distances. It is well known that they don't - only local distances are preserved. Global distances MAY be preserved, but I have seen simple embeddings (< 1000 data points) where they were not.

More food for thought:

https://www.nature.com/articles/s41587-020-00809-z

[dsull]

The twitter thread associated with that paper says "they don't preserve local or global structure & are misleading".

Btw, there are some cool discussions about UMAPs today on twitter by multiple computational biology professors (give them all a follow!).

Answer 1

[jamesxli2007]

The t-SNE maps do enable some interesting analysis. Like the size, density, shape and relative positions of clusters all tell something about the underlying data. For instance, a cluster shaped like a queue is most likely dominated by a single feature/gene. You just need to play with the method a little more.

[dsull]

Hard disagree. One should not go around making biological conclusions based on the fact that you see a "queue" shape. Different computational biologists diverge on whether such techniques are useful as visualization, but interpreting them in that way (i.e. a downstream analysis in the embedding space) is not the way to go. Many computational biologists, at the very least, caution against such interpretations.

[jamesxli2007]

Comparing t-SNE and UMAP, our experience is similar to what you have said: the latter is way too instable and it produces too many fake clusters. Unfortunately, the latter has been widely promoted on the net by so-called computational biologists.