Re-integrate after removing low-quality cells?

First of all, thank you for this excellent family of packages. Our lab uses many on a near-daily basis. This is less of a technical help question and more of a discussion question:

After integration with SCVI and standard clustering procedures, I almost always find 1-3 “bad” clusters of high mito, low complexity cells and remove them. My question is: is it better to completely re-integrate the subset, or just re-cluster using the same integration vectors as before? My instinct says it’s probably fine to do the latter, but the vectors that defined those original clusters are now somewhat meaningless / have little impact on remaining cells.

Any insight you may have would be greatly appreciated, thanks!

1 Like

For small, technically-driven clusters, we don’t see much of a difference in rerunning scvi. If it was something major, you might consider rerunning (and reselecting HVG).

2 Likes

Thank you! This has largely been my experience as well, having now tried both approaches on the same dataset