merging SCTransformed objects removes cell clusters

[daniquebax]

Dear Seurat team/users,

Thanks for the great package! I'm just running into an issue when I try to merge a couple of different datasets:

I am trying to merge 4 seuratobjects, of a timeline experiment. In which each timepoint should contain different cell types.
When I analyze (SCTransform, runPCA, find neighbors, find clusters, runUMAP) the objects by themselves (before merging) I get the clusters which I expect. Only one cluster at timepoint 0 and more cell types emerging as the timeline continues.

However, when I try to merge the 4 objects (according to https://satijalab.org/seurat/articles/integration_introduction.html#perform-integration-with-sctransform-normalized-datasets), I get a completely overlapping UMAP, with no biologically logical cell clusters:

#I load all my objects like this:
ESC_matrix = "matrix_url"
ESC_cells = "barcodes_url"
ESC_features = "features_url"
ESC_data <- ReadMtx(ESC_matrix, ESC_cells, ESC_features)
ESC <- CreateSeuratObject(ESC_data, names.delim = "-", project = "ESC")
ESC$percent.mt <- PercentageFeatureSet(ESC, pattern = "^mt-")
ESC <- subset(ESC, subset = nFeature_RNA > 500 & nFeature_RNA < 2500 & percent.mt < 13)

merged <- merge(ESC, y = c(g048, g072, g120), add.cell.ids = c("ESC", "g048", "g072", "g120"), project = "merged_obj")
merged <- JoinLayers(merged)
#split for normalization
merged[["RNA"]] <- split(
  merged[["RNA"]],
  f = merged$orig.ident
)
#normalize on each sample seperately
merged <- SCTransform(merged, vst.flavor = "v2")
#RunPCA
merged <- RunPCA(merged)
merged<- FindNeighbors(merged, dims = 1:30, reduction = "pca")
merged<- FindClusters(merged, resolution = 0.8, algorithm = 1)
merged<- RunUMAP(merged, dims = 1:30, reduction = "pca", reduction.name = "umap.unintegrated")
DimPlot(merged, reduction = "umap.unintegrated", label = FALSE, group.by = "orig.ident")

Then my UMAP looks like this:

The 120h timepoint should be completely different than the ESC. And the ESC should not contain any different cell types. So, biologically, this doesn't make any sense

Things I've tried already:
-removing the JoinLayers, and split steps
-running SCTransform on individual objects before merging and using scale.data as variable features, as described in #2814
-running SCTransform on individual objects before merging and using SelectIntegrationFeatures as variable features, as described in #6185
-In RunUMAP adjusting the number of: dimensions, n.neighbors, min.dist or spread

Probably I am doing something wrong, but I cannot figure out what. Any help would be greatly appreciated!!

Best,
Danique