Support for spatial transcriptomics (Visium)

[lgaspardboulinc31]

Dear developers,

Thanks for the amazing tool!
I was wondering If you were planning to add a module for spatial transcriptomics data as most packages have difficulties for conversion.

I leave here a snippet of code I have written to do so:

# Prepare the custom conversion
  ## Obs - Metadata
  obs_mapping <- colnames(spatial_obj@meta.data)
  names(obs_mapping) <- colnames(spatial_obj@meta.data)
  print(length(obs_mapping))
  print(paste0("Nrow metadata:", nrow(spatial_obj@meta.data)))
  
  ## var mapping - Seurat does not often compute var statistics like in sc.pp
  
  ## uns mapping - image identidities
  ### Image data
  images = list(lowres=spatial_obj@images$tissue_lowres_image.png@image)
  
  ### Scale factors
  scalefactors_seurat <- spatial_obj@images$tissue_lowres_image.png@scale.factors
  scalefactors = list(fiducial_diameter_fullres=scalefactors_seurat$fiducial, spot_diameter_fullres=scalefactors_seurat$spot,
                      tissue_hires_scalef=scalefactors_seurat$hires, tissue_lowres_scalef=scalefactors_seurat$lowres)
  ### Metadata
  metadata = list(chemistry_description = chemistry_description, software_version = software_version)
  spatial_list <- list(list(images=images,metadata=metadata, scalefactors=scalefactors))
  names(spatial_list) <- sample_name
  
  spatial_obj@misc[["spatial"]] <-spatial_list
  
  # Obsm mapping - Coordinates
  obsm_spatial <- as.matrix(spatial_obj@images$tissue_lowres_image.png@boundaries$centroids@coords[,c("y","x")])
  colnames(obsm_spatial) <- NULL
  rownames(obsm_spatial) <- colnames(spatial_obj)
  print(paste0("Obsm dim:", dim(obsm_spatial)))
  ## Add it to dim reduction
  spatial_obj[["spatial_adata"]] <- CreateDimReducObject(embeddings = obsm_spatial, assay="Spatial",key="spatialadata_")
  
  
  print(paste0("Count dim",dim(spatial_obj@assays$Spatial$counts)))
  print(names(spatial_obj@assays$Spatial@layers))
  print(paste0("Objet dim:",dim(spatial_obj)))
  # Trying to use custom annotation
  adata <- as_AnnData(
    spatial_obj,
    assay_name = "Spatial",#spatial assay
    x_mapping = "counts",
    layers_mapping = c(counts="counts"),
    obs_mapping = obs_mapping,
    var_mapping = TRUE,
    obsm_mapping = list(X_pca = "pca", X_umap = "umap", spatial = "spatial_adata"),
    obsp_mapping = TRUE,
    uns_mapping = c("spatial")
  )
  print(paste0("Anndata dim ", dim(adata$X)))

However, I had an error when writing this object. But weirdly, the object was still written and seems functional in python environment as I was able to access .X and plot features & metadata.

> write_h5ad(adata,paste0(save_path, sample_name,".h5ad"), mode="a")
Error:
! Unexpected shape for X
ℹ Expected [36601], got [642, 36601]

> rlang::last_trace(drop = FALSE)
<error/rlang_error>
Error:
! Unexpected shape for X
ℹ Expected [36601], got [642, 36601]
---
Backtrace:
    ▆
 1. ├─anndataR::write_h5ad(...)
 2. │ └─object$as_HDF5AnnData(path, compression = compression, mode = mode)
 3. │   └─anndataR:::as_HDF5AnnData(...)
 4. │     └─HDF5AnnData$new(...)
 5. │       └─anndataR (local) initialize(...)
 6. └─anndataR (local) `<fn>`(base::quote(`<dgRMatrx[,36601]>`))
 7.   └─private$.validate_aligned_array(...)
 8.     └─cli::cli_abort(...)
 9.       └─rlang::abort(...)

Have you already experimented such issue? Do you see any way I could change my code?

Thanks a lot!
Best,
Lucie

[lazappi]

I think there are a couple of things here, I've created a sub-issue for the error.

This is something we should look into but it might not be straightforward. AnnData isn't really designed for spatial data so there aren't default slots to store some of the information. Hopefully, the flexibility in the mapping helps to handle this but we might be able to come up with some way to make it easier to use.