Enrichmap on cell type assignment using gene sets - Cell segmented Visium HD data

[Rafael-Silva-Oliveira]

Hi, and thank you for the great work on Enrichmap, it has been my go‑to tool for identifying patterns in my Visium HD data.

I’m currently using Enrichmap for cell-type assignment across multiple spatial domains, and I would like to confirm that my workflow is appropriate for cell‑segmented Visium HD data (i.e., using actual cell boundaries rather than bins). Based on the paper, it seems feasible, but I want to make sure what I am doing actually makes sense.

Summary of My Workflow

Data setup

• Three Visium HD samples (lung cancer), concatenated into a single AnnData object.
• Cell segmentation performed → each observation corresponds to a cell, not a bin.
• 12 spatial domains (clusters) - batch corrected and comparable across the 3 samples.
• 21 curated cell-type gene sets.

Enrichmap scoring

• I compute Enrichmap scores for each cell across all gene sets across all 250k segmented cells.

        em.tl.score(
            adata=self.adata,
            gene_set=self.marker_genes,
            smoothing=True,
            correct_spatial_covariates=True,
            batch_key=self.batch_key, # in this case, is sample_id (3 samples)
        )

• For each domain, I perform a one‑vs‑rest Wilcoxon test on Enrichmap scores for each cell type (with padjusting).
• Compute mean change (domain vs. rest).
• Use an “elbow” type of approach to identify the largest gap in mean change → assign the most likely cell type per domain.

Additional diagnostics

For each domain, I also compute:

• Moran’s I (using spatial graphs built per sample)
• Cell-type proportions (argmax of Enrichmap scores per cell)

This would give something like:

Scenario	Interpretation
High mean change + high proportion + high Moran's I	Strong, spatially coherent cell-type identity; this domain is clearly X cell type
High mean change + low proportion + high Moran's I	Minority but spatially clustered cell type defining the domain (e.g., immune infiltrate hotspot, I am seeing in my data a few of fibroblasts + b cell for example)
High mean change + high proportion + low Moran's I	Cell type is common and enriched but spatially scattered (no spatial substructure)
Low mean change + high proportion	Cell type is common in this domain but also common elsewhere (not discriminative)

Here is an example of one of the plots:

We see that domain would most likely be assigned to be a Fibroblast, but seems to be enriched (light blue) with b cells and smooth muscle. Cell type proportions and Morans I then shows a bit more detail on this (Essentially, is that domain guaranteed to be mostly Fibroblast or could we assign it as something like stormal-imune, etc)

Main Question

Is Enrichmap suitable and recommended for cell-type assignment in cell-segmented Visium HD data?
In other words, is it valid to use Enrichmap scores at the single‑cell level (rather than bin level) for domain‑level cell-type inference? Overall I am getting really strong cell typing results, and they make sense. I see pericytes cells where they should be present comparing to the HE image, same thing with smooth muscle/connective tissue, and so on, so I would assume this is indeed possible with Enrichmap

I guess the results will mostly depend on how good my gene sets are, and not so much on Enrichmap approach (which is very solid as is). I was using AUCell, and enrichmap gives far better discrimination/granulation of my data.

If there are aspects of this workflow that you would advise modifying or validating differently, I would be very interested to hear your thoughts! Thanks again for the great work done with Enrichmap!

[cenk-celik]

Hi Rafael, thanks for the kind words and for such a detailed write-up of your workflow — really helpful to see how you're using EnrichMap in practice.

To answer your main question directly: yes, EnrichMap is absolutely suitable for cell-segmented Visium HD data. The scoring framework is agnostic to whether observations represent bins or cells. What matters is that each observation has a gene expression profile and spatial coordinates, both of which you have. So there is nothing conceptually wrong with applying it at the single-cell level.

A few thoughts on your workflow:

• One thing worth noting is that with cell-segmented data, the spatial graph used for smoothing will reflect actual cell neighbourhoods rather than bin adjacency, which is arguably more biologically meaningful. Just make sure the spatial connectivity graph is constructed per sample before scoring (i.e., batch_key="batch").

• The Wilcoxon one-vs-rest approach on EnrichMap scores per domain is sensible and a natural extension of how we use the scores. Your interpretation table is well thought out, in particular, the "high mean change + low proportion + high Moran's I" scenario is a useful distinction that captures spatially coherent minority populations (immune infiltrates, as you note). Nevertheless, in my hands, I found pseudobulk DEGs (one-vs-rest) more profound when using EnrichMap for refining cell type annotations.

• You're right that the discrimination you get is largely driven by the quality and specificity of your gene sets. With 21 curated cell-type signatures across lung cancer tissue, I'd recommend checking for any overlap between your gene sets (e.g. fibroblast vs smooth muscle markers can share ECM-related genes) since shared genes can dilute the contrast between closely related stromal populations. If you haven't already, visualising the pairwise Jaccard similarity between your gene sets can help flag this.

• For the domain assignment step, as spatial domains approach addresses a different problem, I can't provide a good answer for that. Since those tools usually find domains with mixed cell populations, assigning a single cell type to a domain would not be appropriate in my opinion.

Overall, your workflow looks solid and the fact that your assignments are concordant with histology is a good sign. It's great to see EnrichMap being applied to cell-segmented Visium HD data, and your use case is a nice demonstration of how the spatial correction gives you better granularity than non-spatial enrichment methods.

Cenk

[Rafael-Silva-Oliveira]

Hi Rafael, thanks for the kind words and for such a detailed write-up of your workflow — really helpful to see how you're using EnrichMap in practice.

To answer your main question directly: yes, EnrichMap is absolutely suitable for cell-segmented Visium HD data. The scoring framework is agnostic to whether observations represent bins or cells. What matters is that each observation has a gene expression profile and spatial coordinates, both of which you have. So there is nothing conceptually wrong with applying it at the single-cell level.

A few thoughts on your workflow:

• One thing worth noting is that with cell-segmented data, the spatial graph used for smoothing will reflect actual cell neighbourhoods rather than bin adjacency, which is arguably more biologically meaningful. Just make sure the spatial connectivity graph is constructed per sample before scoring (i.e., batch_key="batch").
• The Wilcoxon one-vs-rest approach on EnrichMap scores per domain is sensible and a natural extension of how we use the scores. Your interpretation table is well thought out, in particular, the "high mean change + low proportion + high Moran's I" scenario is a useful distinction that captures spatially coherent minority populations (immune infiltrates, as you note). Nevertheless, in my hands, I found pseudobulk DEGs (one-vs-rest) more profound when using EnrichMap for refining cell type annotations.
• You're right that the discrimination you get is largely driven by the quality and specificity of your gene sets. With 21 curated cell-type signatures across lung cancer tissue, I'd recommend checking for any overlap between your gene sets (e.g. fibroblast vs smooth muscle markers can share ECM-related genes) since shared genes can dilute the contrast between closely related stromal populations. If you haven't already, visualising the pairwise Jaccard similarity between your gene sets can help flag this.
• For the domain assignment step, as spatial domains approach addresses a different problem, I can't provide a good answer for that. Since those tools usually find domains with mixed cell populations, assigning a single cell type to a domain would not be appropriate in my opinion.

Overall, your workflow looks solid and the fact that your assignments are concordant with histology is a good sign. It's great to see EnrichMap being applied to cell-segmented Visium HD data, and your use case is a nice demonstration of how the spatial correction gives you better granularity than non-spatial enrichment methods.

Cenk

Thank you very much for your feedback, Cenk! Super helpful thoughts and regarding the 2nd bullet point about pseudobulk:

• Would this be a normal pseudobulk based on the gene expression for each sample and check the top markers to see which cell type to assign (using e.g. pydeseq2)? This would thus be a complement to Enrichmap approach, correct? Or were you referring to some approach of using DEG with Enrichmap?

I will check the pairwise Jaccard similarity index as well! Great idea to check those overlapping genes.

Regarding cell annotation, would you say that taking the maximum value of a score from Enrichmap for a given cell is an appropriate approach for individual cell annotation? It may not be the best, but just looking to see if this would be sound if the gene sets are appropriate for that annotation task.

Thanks again for your time on this!