Understanding UMAP

[ckaur]

Hi,

I am fairly new to the single-cell transcriptomics domain and have recently started analyzing my dataset of erythroid precursor cells. The cells that I am working with are in the terminal differentiation phase. The differentiation phases go like this: Proerythroblasts --> Basophilic erythroblasts --> Polychromatic erythroblasts --> orthochromatic erythroblasts. I have obtained a UMAP and roughly annotated it with these cell stages (see the figure). However, I am concerned about the Polychromatic erythroblasts not falling adjacent to the orthochromatic in the UMAP. Is it usual to get such results in a UMAP where cells that should be similar to each other are poles apart in the UMAP?

[crazyhottommy]

The distances between points on UMAP mean little. It is a non-linear dimension reduction. on a PCA plot, you can say the further the points the more different they are. Search "UMAP" https://github.com/crazyhottommy/scR...seq-clustering I have some notes.

[ckaur]

Thanks for your reply! I have been told that distance is meaningless in UMAPs, but global topology, structure, and connectivity of the data are conserved. In my case, the Poly and Ortho cells are similar to each other, so should be connected (I assumed). I have read many articles on UMAP preserving the global structure of the data and my data does not show that. Is there a way I can try to understand this through an example? Thanks!

[fchatonnet]

Since you're studying a differentiation process where some kind of progression between cell stages is expected, I'm a bit concerned that this progression is not apparent on the UMAP. It should be conserved, because as you wrote, UMAP is supposed to keep the connectivity between related populations and differentiation generates a continuum of cells with intermediate states between de defined subtypes.
Could you try to represent the data on a PCA plot to better see the relationships between the cells?
Maybe it's also worth checking the different quality parameters of your data and verify that you work only on high quality cells with no doublets, low rRNA / mtRNA content, no dead cells and so on.
Good luck!