Closed rbutleriii closed 3 months ago
Hi @rbutleriii ,
Thanks for your question!
Re 1. The low number of anchors might indicate the reference data you used did not match well with your query data. Re 2. As for "integrated" assay, Seurat no longer generates such assay/layer in v5, so you will not see it in the object (you may consider generate it by yourself with other packages or Seurat v4, if you truly want to access it). But you can access the dimReduc object from the integration by using command like obj$integrated.cca. Re 3. SCTransform is recommended but please feel free to use the standard NormalizeDate() to normalize your datasets.
In addition, in an effort to more promptly address user issues, we’ve started asking users to direct questions like this to our Discussions board, where community members and developers can provide more targeted assistance.
We’re going to close this issue for now but strongly encourage you to repost your question in the other forum.
We look forward to seeing your post there - thanks for using Seurat!
Hi, I am trying to use my matched snRNAseq and CosMx data to predict cell types for the CosMx data from the snRNAseq. I generally have the code together (See #8660 for a related issue), but I want to make sure I am taking the correct approach. It briefly notes that this was done on the tutorial CosMx data, but the steps were not laid out. Edit: I updated the scripts to reflect further understanding of Seurat v5 integration. It gets worse....
The query data is from 5 slides, so it is stored in the seurat object as 5 layers:
So, that means I should probably integrate them and join layers, correct?
Is SCTransform mandatory for this assignment, or could you compare the standard normalized/scaled CosMx data? Would you modify the settings for RNA method to handle CosMx similar to the
clip.range = c(-10,10)
argument in the CosMx tutorial?Presumably, I would want to feed the 950 genes available for CosMx data in to the reference dataset as features for scaling and pca calculation. However, it would be preferable to also have these appear in the reference's
scale.data
layer, as well as include them for the Reference'sRunPCA
. I included them by using theunion
of my CosMx features and the ones found withFindVariableFeatures
. If anything the additional variable features not in the CosMx data seem like they would be useful for generating the reference UMAP correctly, and likely wouldn't impact the transfer anchors at all?PCs I have a 30, as that was informative for the snRNAseq, though perhaps scale back the PCs considering how few genes are represented in the CosMx set.
When I did run the
FindTransferAnchors
without integrating the layers, specifying only the 950 features in the CosMx set for the refScaleData
andRunPCA
, I get remarkably few anchors (25) for over 140k cells in CosMx, ~100k in snRNAseq. So something seems odd. If I integrate with v5, I think it should be generating a singlescale.data
layer, but the assay will still beNanostring
? In v4 it seemed like it generated aintegrated
Assay with it's ownscale.data
slot. This way it looks like it is basically not using theharmony
orharmony.umap
reductions, only the combinedscale.data
layer and potentially thepca
reduction has been recalculated on the integrated layers (though i didn't callRunPCA
again)? Is there any reason to bother with the integration, or will it by default transfer label predictions to each of the original 5 layers (assuming INormalize
%>%FindVariableFeatures
%>%ScaleData
%>%RunPCA
each of them)?