Annotation general question

[synatkeamsk]

Hi Stereopy Team,

Thank you for your continued support with my Stereo-seq analysis. Your guidance has been extremely helpful, and I have been able to complete many downstream analyses successfully.

I have a question regarding cell type annotation of square bins (Bin50). After identifying marker genes for each cluster, I noticed that many canonical markers from different cell types appear together within the same cluster. For example, some clusters express markers associated with T cells, B cells, and macrophages simultaneously. This makes me think that each Bin50 spot likely contains transcripts from multiple neighboring cells rather than representing a single cell type.

Given this, I am wondering how cluster annotation is typically performed for Bin50 data. Is it appropriate to interpret a cluster as being enriched for a particular cell type (e.g., macrophage-enriched) based on the dominant markers, even though the cluster may contain a mixture of cell types? In other words, should these clusters be considered spatial regions or microenvironments rather than pure cell populations?

Additionally, does Stereopy provide any tools for cell type annotation beyond marker-gene analysis and SingleR? For example, are there built-in methods for cell type deconvolution, label transfer from scRNA-seq references, or probabilistic annotation that can estimate the contribution of multiple cell types within each bin?

I would appreciate any recommendations on best practices for annotating Bin50 data.

Thank you for your help.

Best regards,
Synat

[wanruiwen-genomics-cn]

Dear @synatkeamsk,

Thank you for reaching out and for your thoughtful question!

Short answer: Yes, Bin50 spots inherently capture transcripts from multiple cells, so clusters should be interpreted as spatially coherent regions enriched for certain cell types rather than pure populations. Stereopy offers several annotation tools beyond marker genes and SingleR, and you can also integrate external deconvolution methods via AnnData interoperability.

---

Assessment

• Type: usage/question
• Severity: low
• Affected module: stereo/tools/cell_type_anno.py, stereo/algorithm/single_r/single_r.py

What Is Happening

When working with square bins (e.g., Bin50), each spot spans a 50×50 DNB (DNA nanoball) area, which typically covers multiple cells. As a result, gene expression profiles within each bin are a composite of all cells captured within that spatial footprint. This is why you observe canonical markers from different cell types co-expressed within the same cluster — it is an expected consequence of the spatial resolution, not an analysis error.

Clustering on Bin50 data groups spatially adjacent spots with similar composite expression profiles, so each cluster effectively represents a spatial region or microenvironment that is enriched for a dominant cell type, but may also contain contributions from neighboring cell types.

Is This Expected?

Yes. This is inherent to bin-based spatial transcriptomics data. Unlike single-cell RNA-seq where each observation represents one cell, each Bin50 spot aggregates signal from a tissue area containing multiple cells. The mixed-marker signature you observe is a natural consequence of this spatial averaging.

Recommended Workflow

Step 1: Interpret clusters as spatial domains / microenvironments

It is standard practice to label Bin50 clusters by their dominant cell type signature (e.g., "macrophage-enriched region") rather than assigning a single cell type identity. You can use data.tl.annotation() to assign these labels to your clusters:

# After clustering, annotate clusters with descriptive labels
data.tl.annotation(
    annotation_information={
        '0': 'T-cell-enriched',
        '1': 'Macrophage-enriched',
        '2': 'B-cell-enriched',
        # ... etc.
    },
    cluster_res_key='leiden'
)

Step 2: Use SingleR for reference-based annotation

Stereopy includes a built-in SingleR implementation (stereo/algorithm/single_r/single_r.py) that performs reference-based cell type annotation using Spearman correlation against a labeled reference dataset. While originally designed for single-cell resolution, it can still help identify the dominant cell type per bin:

# Assuming ref_data is your reference StereoExpData with known cell types
data.tl.single_r(
    ref_exp_data=ref_data,
    ref_use_col='celltype',
    cluster_res_key='leiden'
)

Step 3: Use correlation-based annotation (CellTypeAnno)

Stereopy also provides CellTypeAnno (stereo/tools/cell_type_anno.py), a correlation-based annotation tool using the FANTOM5 reference database (or a custom reference). It computes Spearman or Pearson correlation between bin expression profiles and reference cell type profiles with bootstrap resampling for robustness.

Step 4: Explore spatial microenvironment analysis

For understanding the cellular composition within spatial regions, consider using:

• Co-occurrence analysis (stereo/algorithm/co_occurrence.py): Quantifies how frequently different cell types co-localize within spatial neighborhoods.
• Niche extraction (stereo/algorithm/get_niche.py): Identifies spatial niches based on cell type proximity and clustering results.
• Cell-cell communication (stereo/algorithm/cell_cell_communication/): Analyzes ligand-receptor interactions within spatial niches.

Step 5: Consider cell-bin (single-cell resolution) or external deconvolution

For true single-cell-level annotation, the recommended approach is:

1. Use cell segmentation: If you have corresponding image data, Stereopy supports cell segmentation to generate cell_bins data (bin_size=1), which gives single-cell resolution.
2. Use external deconvolution via AnnData: Stereopy data can be converted to AnnData format, enabling you to use popular deconvolution tools from the scverse ecosystem:

from stereo.io.reader import stereo_to_anndata

# Convert to AnnData for use with external tools
adata = stereo_to_anndata(data)

# Then use tools like:
# - cell2location (probabilistic deconvolution)
# - SPOTlight (NMF-based deconvolution)  
# - RCTD (robust cell type decomposition)
# - Tangram (deep learning-based mapping)

Alternative Approaches

If you want to avoid the mixed-signal issue entirely:

• Option A: Cell segmentation — If image data is available, use Stereopy's cell segmentation pipeline to generate cell-level bins (cell_bins), which will give you single-cell resolution annotation.
• Option B: Smaller bin sizes — Use a smaller bin size (e.g., Bin20 or Bin10) to reduce multi-cell mixing, though this trades off against data sparsity.
• Option C: Multi-resolution analysis — Perform clustering at multiple bin sizes and compare to gain insight into spatial organization at different scales.

Notes

• When interpreting Bin50 clusters, focus on the relative enrichment of marker genes rather than expecting exclusive expression of a single cell type's markers.
• The data.tl.find_marker_genes() results for Bin50 data should be interpreted as identifying spatially variable genes that distinguish regions, not cell-type-exclusive markers.
• For deconvolution approaches, having a well-matched scRNA-seq reference from the same tissue type is critical for accurate results.
• Stereopy's totalVI integration (stereo/algorithm/total_vi.py) is available for multi-modal data analysis if you have both RNA and protein measurements.

Please let us know if you have further questions!

Best regards,
Stereopy Maintainer