Cell cycle regression and n_embedding tuning

[BiotechPedro]

Hello Constantin,

I've been playing a bit with lemur on the following dataset. Basically, a cell line treated under multiple conditions (~10) and I have two replicates of the experiment. However, the cell cycle is a clear confounding factor. How should I regress it? Is the next way the correct one to follow?
fit <- lemur(sce, design = ~ conditions + experiment, n_embedding = 15, test_fraction = 0.25)
fit <- align_harmony(fit, design = ~ Phase)

I've already tried this and the cell cycle effect disappears, but I don't know if this is technically a good procedure. On the other hand, fit <- align_by_grouping(fit, grouping = sce$Phase) does almost nothing. What would be your way to proceed?

Also, would you mind to share some ideas on how to tune the n_embedding? I've read that it follows the same logic as PCs, but maybe you've found a normal range of use (e.g. 15-20) or you put it larger for very heterogeneous data.

Thanks a lot!

Pedro

[BiotechPedro]

Hi Constantin! Here I am again :)

I've read LEMUR's recent version of the preprint and it is greatly explained. I really like the new applications.

Although I ended up not analysing the dataset I commented about above, I am again dealing with a cell lines dataset. In here, the cell cycle effect is big, so I am wondering how would you proceed. Which code from the above, or another, would you use? Would you subset the data to work on a concrete cell cycle phase? Would you just remove the cell cycle genes from the HVGs and continue as normal?

Regarding the number of latent dimensions, I see in your new preprint version that it varies depending on the dataset. Would the general recommendation be to compute the cumulative variance explained as you do in Suppl. Fig. 7A and then to look for the 'elbow'?

Thank you!

Pedro