Recommendations for multi-modal, multi-slice spatial data

[Thapeachydude]

Hello,

I would love to run your tool on multi-modal, multi-slice spatial RNA/Protein data.
I would have the following questions:

1). Which functions would I need to run for multi-modal, multi-slice data?
I saw that in the multi-modal tutorial the sn.pp.cal_spatial_exp is run for each modality, all of which is then passed to sn.pp.prepare_data. Where does sn.pp.process_multi_slices from the multi-slice tutorial fit into all of this? Would this be instead used for multi-slice data?

2). Is there any documentation you could point me to that explain a bit in detail what the functions are doing and what parameters they take?
i.e. what is k_cutoff, is_pca, is X_rna the raw or normalized count matrix?

3). Where would batch-effect correction happen? Is this addressed by running the multi-slice call? Or should it run before?

Apologies for the many questions, if there's documentation that I've missed, feel free to point me to it : )
Cheers!

[ForwardYang98]

Hi,
Sorry for the delayed response. For your questions:

1. Sorry to say that sn.pp.process_multi_slices is currently not suitable for ​​multi-modal, multi-slice data​. In practice, sn.pp.process_multi_slices integrates sn.pp.cal_spatial_neighbors and sn.pp.cal_spatial_exp​​, which extract the three-view features of cells (molecular profiles of cells, molecular profiles of neighborhoods, and cellular compositions of neighborhoods) within a predefined neighborhood range. scNiche was initially developed primarily for ​​single-modal spatial omics data​​. Although we have conducted preliminary testing on spatial multi-modal data, there is currently a lack of well-established interfaces to batch-process ​​multi-modal, multi-slice data​​. We will update the corresponding functions in future releases, but this will take some time. Alternatively, you can run ​​sn.pp.cal_spatial_exp​​​​ separately on each slice, then follow the multi-modal tutorial to construct new AnnData objects, merge them into a combined AnnData object via ​​sc.concat​​, and finally run ​​sn.pp.prepare_data​​ or ​​sn.pp. prepare_data_batch​​ (depending on your data size). An example workflow is as follows:

cutoff = 30
adata_rna1 = sn.pp.cal_spatial_exp(adata=adata_rna1, mode='KNN', k_cutoff=cutoff, is_pca=False, verbose=False, layer_key='X_pca')
adata_cuttag1 = sn.pp.cal_spatial_exp(adata=adata_cuttag1, mode='KNN', k_cutoff=cutoff, is_pca=False, verbose=False, layer_key='X_lsi')
adata_rna2 = sn.pp.cal_spatial_exp(adata=adata_rna2, mode='KNN', k_cutoff=cutoff, is_pca=False, verbose=False, layer_key='X_pca')
adata_cuttag2 = sn.pp.cal_spatial_exp(adata=adata_cuttag2, mode='KNN', k_cutoff=cutoff, is_pca=False, verbose=False, layer_key='X_lsi')
...

# new AnnData
adata1 = sc.AnnData(adata_rna1.X.copy())
adata1.obs = adata_rna1.obs.copy()
adata1.obsm['X_rna'] = adata_rna1.obsm['X_data'].copy()
adata1.obsm['X_rna_nbr'] = adata_rna1.obsm['X_data_nbr'].copy()
adata1.obsm['X_cuttag'] = adata_cuttag1.obsm['X_data'].copy()
adata1.obsm['X_cuttag_nbr'] = adata_cuttag1.obsm['X_data_nbr'].copy()
adata1.obsm['spatial'] = adata_rna1.obsm['spatial'].copy()

adata2 = ...

# combined AnnData object
adata_combined = sc.concat([adata1, adata2, ...], keys=None)
adata_combined.obsm['X_rna'] = np.nan_to_num(adata_combined.obsm['X_rna'])
adata_combined.obsm['X_rna_nbr'] = np.nan_to_num(adata_combined.obsm['X_rna_nbr'])
adata_combined.obsm['X_cuttag'] = np.nan_to_num(adata_combined.obsm['X_cuttag'])
adata_combined.obsm['X_cuttag_nbr'] = np.nan_to_num(adata_combined.obsm['X_cuttag_nbr'])
adata_combined.obsm['spatial'] = np.nan_to_num(adata_combined.obsm['spatial'])
...

# prepare
choose_views = ['X_rna', 'X_rna_nbr', 'X_cuttag', 'X_cuttag_nbr']
adata_combined = sn.pp.prepare_data(adata=adata_combined, verbose=True, choose_views=choose_views)
...

2. we will update the function files with a description for each parameter after a while. Specifically, k_cutoff represents the number of cells selected as neighbors, X_rna is the normalized count matrix, and is_pca indicates whether to perform PCA dimensionality reduction on X_rna. In general, if the batch removal step has already been performed, the latent representations output by the batch effect removal methods (such as X_scVI) can be used directly.
3. It is recommended to perform the batch removal step before running scNiche.

[Thapeachydude]

Hi,
thanks for the detailed reply and apologies for not getting back to you. I've had to deal with various pre-processing steps on additional batches for the dataset before I can now get back to the niche detection part.

I will stick with single-modal spatial data for now then. Fortunately, we don't see much of a batch effect for the transcriptomics data. I would have two follow-up questions if I may:

1). Is there a way to output the loss over epochs i.e. like you did in Supplementary Figure 4?

2). Would you have any recommendations for the selection of the number of clusters? Does it make sense to test various values and then settle for one? With ≈ 10 million cells the dataset is quite large, so testing many values on the full dataset will likely not be feasible.

Best

[ForwardYang98]

Hi,
For your questions:

1. The training information stored in adata.uns['loss'].
2. To be honest, the selection of the number of clusters usually depends on the specific dataset, and I don’t have a very definite recommendation. You might try testing different values (e.g., 5–20) on the embeddings trained by scNiche. Since this only involves repeating the unsupervised clustering step, the computational cost should be acceptable. :)

[Thapeachydude]

Hi, thanks for the reply. A quick follow-up regarding the training. Would you have any recommendations for training very large datasets?

As mentioned I have roughly 10 million cells with 5000 features. I tried a moderate number of 250 batches, but even with 48GB of GPU RAM it didn't finish a single epoch in 12h. So, I increased the number of batches massively. With 40'000 batches (I guess this would mean ≈ 250 cells per batch?) an epoch takes roughly 2h. Would you recommend even more batches or is there a performance drop if too few cells are in each batch?

Also I found that with 250 batches model = sn.tr.Runner_batch(adata = adata, verbose = True, device = "cuda:0") uses primarily GPU RAM, whereas with 40'000 it will use mostly CPU, even with cuda:0.

[ProDong0512]

Hi, thanks for the reply. A quick follow-up regarding the training. Would you have any recommendations for training very large datasets?

As mentioned I have roughly 10 million cells with 5000 features. I tried a moderate number of 250 batches, but even with 48GB of GPU RAM it didn't finish a single epoch in 12h. So, I increased the number of batches massively. With 40'000 batches (I guess this would mean ≈ 250 cells per batch?) an epoch takes roughly 2h. Would you recommend even more batches or is there a performance drop if too few cells are in each batch?

Also I found that with 250 batches model = sn.tr.Runner_batch(adata = adata, verbose = True, device = "cuda:0") uses primarily GPU RAM, whereas with 40'000 it will use mostly CPU, even with cuda:0.

In general, scNiche could maintain a relatively stable performance under different batch number settings, as shown in the Supplementary Fig.S4 of our article.

Additionally, based on our experience, the following approaches may be also worth noting in practice:

epoch number setting. For large datasets, starting with a smaller epochs initially is recommended. And you can also evaluate the convergence of the model by visualizing the training information stored in adata.uns['loss'].
Dimensionality reduction and batch effect removal. Considering that the number of genes measured in CosMx data usually far exceeds the number of cell types that exist, dimensionality reduction (scVI, scArches, or PCA) can help balance the dimensionality of features across different views, allowing for more accurate niche identification.

Overall, it may take some time to find the optimal parameter configuration. If you have any results to share, I would be most interested in seeing them!