Questions on Spatial Coordinates and Large ST Data in gsMap

[Cravit998]

Dear Dr. Song,
First, I would like to express my sincere respect for your work.
In my view, gsMap is a major methodological advance that elegantly overcomes several intrinsic limitations of traditional GWAS interpretation. I truly appreciate this contribution to the field.
While using gsMap for large-scale ST datasets, I encountered several conceptual and practical questions. I would be very grateful if you could provide clarification or suggestions:

1. How is “spatially connected” defined in gsMap? Does it rely strictly on XY coordinates from the ST file?
   In the Supplementary Note, gsMap states:

“we leverage the graph attention (GAT) layer to aggregate information from spatially connected spots.”

May I confirm whether “spatially connected” refers specifically to adjacency based on the XY coordinates in the original ST dataset? This question becomes important because some ST datasets, such as the human dataset from Science (https://www.science.org/doi/10.1126/science.add7046), do not provide true physical XY coordinates, but rather pseudo-layout coordinates for visualization. If gsMap builds the graph strictly from these XY coordinates, would this lead to inaccuracies in the aggregation step or in the resulting GSS/PCC values?

2. How should gsMap handle datasets with true 3D coordinates (XYZ)?
   For example, the ST dataset from Nature (https://www.nature.com/articles/s41586-023-06812-z) provides 3D spatial coordinates (XYZ). gsMap currently takes only XY coordinates: If I discard the Z dimension, will the loss of 3D spatial neighborhood information distort the spatial graph? Would you recommend projecting XYZ into 2D, or using another method to preserve true spatial relationships?

3. The Nature dataset (https://www.nature.com/articles/s41586-023-06812-z) contains ~4 million cells, and generating the GSS requires extremely large memory. To solve this, I considered splitting the dataset into many parts and running gsMap separately. However, two problems arise:

   3a. Splitting by XY coordinates risks losing entire brain regions in some partitions. So, if splitting by cell type would be better? Or is there a recommended or optimal splitting strategy for gsMap?

   3b. After splitting, each part will compute its own GSS and SNP annotations. This may lead to the same gene having different annotations in different parts, even though it is biologically identical. How should one deal with this issue? Should annotation consistency be enforced across all parts?

4. I plan to integrate PCC results from several ST datasets. However, different datasets may assign different spatial or cell-type GSS patterns to the same gene (e.g., a gene appears glutamatergic-specific in dataset A but GABAergic-specific in dataset B). How should such discrepancies be interpreted or handled?

I don't know if these concerns make sense within the gsMap, or would you recommend alternative approaches to address them?

Thank you very much for your time and for developing such an impactful tool. I would deeply appreciate any insight, clarification, or recommendations you could provide.

Best regards,
Yubin

[LeonSong1995]

Sorry for the late reply — I’ve been quite busy recently and only just found the time to get back to you. Thanks for your patience.

Thanks for your questions!

1. Requirement for spatial coordinates
   gsMap requires spatial coordinates that faithfully reflect the underlying tissue spatial architecture. We are not aware of the exact procedure used to generate the pseudo-coordinates you mentioned. If these coordinates are derived from measurements or reconstructions that preserve true spatial relationships between cells/spots, they may be usable. However, purely arbitrary or artificial layouts that do not correspond to real spatial organization are not supported, as they would invalidate the embedding.

2. Support for 3D spatial data
   The current gsMap version does not support 3D spatial transcriptomics data. However, we are actively developing a new version that will explicitly support 3D ST analysis. Please stay tuned for future updates.

3. Whether to split large datasets
   In general, we do not recommend splitting the dataset. Instead, running gsMap on CPU is usually a good option, as CPUs typically provide much larger memory capacity, and the GAT-based model used in gsMap runs efficiently on CPU in most cases. If splitting is absolutely necessary due to memory constraints, we recommend splitting the data by cell type, rather than arbitrarily.

4. How to combine results after splitting
   After splitting, results should be aggregated using slice mean. This ensures that each gene is compared against the same baseline across splits, which is critical for maintaining statistical consistency.

5. Multiple datasets
   Similarly, when analyzing multiple datasets, we recommend using slice mean to ensures that each gene is compared against the same baseline.

Best,
Liyang