Gene vs Molecule Cell Filter Best Practice

[katkatrach]

Hello!

I am using Pagoda2 in combination with Seurat for clustering and some preprocessing. I have a question regarding the point at which to use the Gene vs Molecule filter in a 10X Multiome pipeline of multiple samples (20+)

Option 1. Apply filter after mt, er, gene count for each individual sample, then merge, sctransform, pca, integrate, etc

Option 2. Apply the filter to the merged object, with layers merged as well (new to Seurat v5), so application on one combined matrix. Then splitting the layers again for sctransform, pca, integration.

It is my understanding that applying the filter to each sample individually would help preserve quality variability between samples, but may lead to imperfect integration. On the flip side, merged object application of the filter would maintain that the filter cutoffs are consistent across each object (such as my seurat mt er etc filters), but could possibly remove important cells in the lower quality samples.

I am curious of anyone has any thoughts to what the better practice may be. Any insight is appreciated! Thanks.

[evanbiederstedt]

I would do Option 1 myself. That strikes me more conceptually clear.

Option 2 is part of this school of thought to try to use sequencing artifacts etc. in order to "improve integration", but it's never clear what precisely that means. At some point, these discussions began pretty isolated from the original goals of integration imho. Please refer to the walkthrough here: https://github.com/kharchenkolab/conos

I'll leave this GitHub issue open for others to discuss.