Scaling performed within IntegrateLayers and RPCAIntegration and issue with regressing out variables

[JTumulty]

I have noticed that when performing integration with IntegrateLayers using the RPCAIntegration method, that many of the features I attempted to regress out in my PCA (e.g. nCount RNA, and percent mito), are now strongly enriched in certain regions of the umap after integration with RPCA. This totally defeats the purpose of regressing them out in my PCA!

I suspect that this has to do with the step in the RPCAIntegration method where each layer is rescaled before FindIntegrationAnchors is run. I believe that, no matter what, this rescaling step should be more clearly specified somewhere. Even better, the RPCAIntegration function should either fetch the variable used in regression from the original ScaleData run, or you should be able to specify them so that these same variables are used in any scaling performed within the integration functions.

Seeing as I have seen a lot of people having trouble with other integration methods and cell cycle regression I think this is an important topic to discuss. Let me know if anyone else has observed this effect or has thoughts!

[JTumulty]

After running FindIntegrationAnchors independently, I don't think this issue is having as large of an effect as I previously thought. However, I still wonder if it is worth it to use the scaled.data that the user actually provides as opposed to one recalculated within the integration function.

I also think it is worth discussing this topic in general. I am not sure there is a clear answer for dealing with feature regression and integration, especially for a method like RPCA where you do not get back corrected data on which you could run subsequent scaling.