Hi there,

I am working on several publicly available single cell datasets. I tipically have several samples per dataset and I am using the latest Seurat version, v5, for preprocessing, QC and data integration. During some test analyses, I noticed that the UMAP clustering can change quite significantly, depending on whether I run the normalisation and variable feature selection on

1. a Seurat object split into multiple layers (1 layer per sample) compared to
2. the scenario where all cells have been previously merged into a single layer.

I am mostly concerned about the following steps:

obj <- NormalizeData(obj)

obj <- FindVariableFeatures(obj)

obj <- ScaleData(obj)

obj <- RunPCA(obj)

which in Seurat v5 are run independently for each layer you define in the Seurat object obj.

I wonder how much of an effect this potentially has on downstream analyses, especially data integration. Based on the Seurat's vignette (https://satijalab.org/seurat/articles/seurat5_integration), keeping samples separated in different layers would be the way to go, and I think it makes sense. I could imagine that the variable feature selection, in particular, can be affected by the layer structure, and that splitting samples allows for a sample specific search of the most important genes, rather than potentially diluting the sample-specific effect in a single global search. I am not 100% sure how much the normalization can be affected though.

I would really like to hear any feedback or opinions regarding this topic. I find it a bit worrying that results and interpretation would be affected by a fairly small change in the workflow.

Thanks a lot!