Problem with sctransform after subset of integrated dataset

[dartagnan323]

Hi I am using seurat 5.2. (I did find some similar requests but did not seem to find a solution)
I have an object with 8 samples of single cell that I first merged then integrated using SCTransform and harmonyintegration parameter which I found to be best in my case.
After I did subset from clusters:

Endo <- subset(x = GCharm, idents = c("Endo1", "Endo2", "Endo3", "Endo4", "Endo5", "Endo6", "EndoAdipo", "EndoLymph")) DefaultAssay(object = Endo) <- "RNA" Endo <- PercentageFeatureSet(Endo, pattern = "^mt-", col.name = "percent.mt") Endo <- SCTransform(Endo, vars.to.regress = "percent.mt", verbose = FALSE)

for Sctransform, all the scaling that I need to redo after subsestting should be integrated so I guess it should be fine.
I did the same for another type of cells.
So in every case I get a message saying that SCT is different than the previous one it contained which apparently from another post I found is not important.
But for one of these 2 subsests I get 4 times:
Warning message: In size + sum(size_args, na. rm = FALSE) : NAs produced by integer overflow
is that a problem?

In addition for each one, I get an sct model list of only 1 whereas I had 8 satnding for my 8 conditions before in my original object. I don't know why it is not separate anymore. I did try to split layers in the RNA assay but it says they already are and I have 10 layers (including the 8 samples + counts and scale data).
Should I re-do my scaling and all on the RNA assay too? I don 't think it matters for SCTransform?
I could do PCA, UMAP, neighbors, clusters. It's fine for one subset but for the other where I have NA errors, I get a few dots (cells) spread out in the UMAP and the main clusters with 99% of cells are then in a small corner. It does not happen with the subset without NA errors.
Then, the fact that I have only 1 layer in SCTransform then generates (I think) an error in the step of integration. (I still have information of integration from the main object before subset, but I suppose it will be overwritten...?)

Note in harmony::HarmonyMatrix(data_mat = Embeddings(object = orig), :
HarmonyMatrix is deprecated and will be removed in the future from the API in the future
Error in contrasts<-(*tmp*, value = contr.funs[1 + isOF[nn]]) :
Contrasts can not be applied to factors with less than 2 levels

So I tried to apply split layers:
Endo[["SCT"]] <- split(Endo[["SCT"]], f = Endo$Sample)
I get:
Note : Input is a v3 assay and split() only works for v5 assays; converting to a v5 assay
Note : Assay SCT changing from SCTAssay to Assay5
Also I note there are 24 layers which is not the 8 I expect (data x8 counts x8, scaledata x8)
then I try to integrate again
Endo <- IntegrateLayers(object = Endo, method = HarmonyIntegration, normalization.method = "SCT", verbose = F)
and I get another error message:
Error in IntegrateLayers():
! None of the features provided are found in this assay

Could you advise as to what I am doing wrong as all this code worked fine on the main object but does not work on subset object. I guess I don t subset properly or there is a step I miss after subsetting.

1. Why do I get NA errors in only one of my subsets
2. Why I can't integrate anymore (how to get my normal layers back?)

thanks a lot for any help,

[dartagnan323]

issues resolved:
insisting on splitting layers after I saw another post, I resolved this issue:
First, define RNA assay as defaultassay
Then join layers of RNA assay as only the data but not the counts are still split in my subset object after integration on the whole object with all cell types.
Then split again the RNA assay: I get the same number of data and counts layers.
Now when I run my integration on the subset it works.
This also resolved my NA issues in the other subset of cells. and I have a nice UMAP.