Functionality for STlist to accommodate multiple samples per file?

[edridgedsouza]

Hi,

I am looking to use the spatialGE package for some analysis of CosMx data. For cost effectiveness reasons, we have set up our experiments so that each slide contains multiple samples. Past experience has shown us that when computing spatial information, we get better interpretability when we analyze each sample separately, rather than doing the entire slide at once. However, looking through the documentation, it appears that the STlist assumes that there is only one sample per file.

I can load the slides into a Seurat object and it's relatively trivial to assign a metadata column for each sample ID (plus relevant clinical characteristics). Is there a convenient way to simply convert our existing Seurat object / metadata into a format compatible with STlist, where each sample is processed separately?

I suppose the most straightforward way to accomplish this would be simply writing new expression/metadata files for each sample after initial annotation, but if there's a more efficient way to convert existing annotated objects into a format compatible with spatialGE, it would be appreciated. Thanks.

[oospina]

Hi,

The easiest way to accomplish this would be to create one named list containing gene expression matrices, and one list containing spatial metadata. Each element of the list would be a sample, FOV, or whatever spatial unit that makes biological sense in your experimental setup. No need to generate files. Please see the last option in the rnacounts argument of the STlist function (https://fridleylab.github.io/spatialGE/reference/STlist.html).

Another less straightforward option would be to read the CosMx files directly with spatialGE, and then subset the STlist object to contain only the relevant spatial samples with the filter_data and itsrm_tissue argument. This implies you would have several STlist objects.

We have a way to import Seurat objects, but unfortunately, only V4 objects. We still need to devote time to developing code for importing V5 objects.

I hope this helps. Please feel free to keep the discussion here and I'll reply as soon as I can.

[edridgedsouza]

Thank you. With a bit of tinkering, I've managed to load it in like this:

# The internal function read_list_dfs expects the first column of the expression df to be the gene names
# Apparently process_lists() also expects this for the meta data

# The `spotcoords` argument accepts a df with ONLY barcode, x, and y
# So we'll have to manually add in spot metadata post-hoc

prep_input <- function(mat, type='expr') {
 if (type == 'expr') {
 firstcol <- 'gene'
 } else if (type == 'meta') {
 firstcol <- 'barcode'
 }
 out <- as.data.frame(mat) %>%
 rownames_to_column(firstcol) %>%
 relocate(!!sym(firstcol))

 if (type == 'meta') {
 out <- out %>% select(!!sym(firstcol), x_slide_mm, y_slide_mm)
 }
 return(out)
}

expr_mat_list <- list(
 'brain1' = prep_input(obj@assays$RNA@data[, obj@meta.data$tissue == 'brain1']),
 'brain2' = prep_input(obj@assays$RNA@data[, obj@meta.data$tissue == 'brain2']),
 'brain3' = prep_input(obj@assays$RNA@data[, obj@meta.data$tissue == 'brain3']),
 'trunk_shoulder' = prep_input(obj@assays$RNA@data[, obj@meta.data$tissue == 'trunk_shoulder']),
 'unknown' = prep_input(obj@assays$RNA@data[, obj@meta.data$tissue == 'unknown'])
)
meta_data_list <- list(
 'brain1' = prep_input(obj@meta.data[obj@meta.data$tissue == 'brain1', ], type='meta'),
 'brain2' = prep_input(obj@meta.data[obj@meta.data$tissue == 'brain2', ], type='meta'),
 'brain3' = prep_input(obj@meta.data[obj@meta.data$tissue == 'brain3', ], type='meta'),
 'trunk_shoulder' = prep_input(obj@meta.data[obj@meta.data$tissue == 'trunk_shoulder', ], type='meta'),
 'unknown' = prep_input(obj@meta.data[obj@meta.data$tissue == 'unknown', ], type='meta')
)


# Create STlist and add spot meta
mylist <- STlist(rnacounts = expr_mat_list, spotcoords = meta_data_list, cores=CORES)
for (samplename in names(mylist@spatial_meta)) {
 additional_meta <- obj@meta.data[obj@meta.data$tissue == samplename, ] %>%
 rownames_to_column('libname')

 mylist@spatial_meta[[samplename]] <- mylist@spatial_meta[[samplename]] %>% 
 left_join(additional_meta, by='libname')
}

Converting them all to dfs does raise the sparse -> dense coercion warning and allocated multiple GB of RAM per tissue; for smaller projects this isn't an issue, but I'm wondering if it may be possible to keep the expression data in [sparse] matrix form for efficiency?

I've managed to load it in and will be doing my further downstream processing from here. Will update if there are any issues. One question: you can see here that a given tissue may have multiple, often disjointed, FOVs. Are there any special considerations I should make in spatial analyses to ignore non-contiguous FOVs so that spatial analyses are not skewed by the empty spaces? Or in your experience, are these effects insignificant at the scale of the local interactions we care about? What is the best way to handle if it is an issue--analyze each FOV separately? Group FOVs into "contiguous zones"? Or just do nothing?

Thanks for the help

[daniannr]

Hi! Jumping in to say I'm also having the same issue since I'm working with V5 Seurat objects. Would be tremendously helpful if we can load the objects directly. Thanks for your work!