Spatial Transcriptomics: Assigning Spots to Cell Types

Jonathan Alles

EVOBYTE Digital Biology

Introduction

If you’ve ever looked at a spatial transcriptomics map and tried to color each spot as if it were a single cell, you’ve felt the friction between biology and measurement. Most sequencing-based assays sample RNA from capture “spots” laid down on a slide. Those spots are often bigger than a cell and sometimes overlap multiple cells. The result is intuitive but easy to forget: each spot’s gene expression can be a mixture of cell types and partial cells. That’s why assigning spots to cell types or to fine tissue regions feels slippery at first.

The good news is that we now have a mature workflow to reason about mixtures, improve effective resolution, and borrow strength from single‑cell atlases. In this post, we’ll unpack how spot geometry shapes resolution, why mixtures are the norm rather than the exception, and how reference mapping and deconvolution turn fuzzy signals into cell-type and region maps you can trust. Along the way, we’ll point to pragmatic tools and show small, copy‑pasteable snippets to help you get started.

What a “spot” really captures: resolution and mixtures in sequencing‑based ST

Sequencing‑based spatial transcriptomics (ST) measures RNA captured by spatially barcoded features on a slide. Classic whole‑transcriptome platforms like 10x Genomics Visium use circular spots roughly the size of small tissue features; a typical Visium spot is about 55 µm in diameter with 100 µm center‑to‑center spacing. A single spot therefore often covers several cells, sometimes blending epithelial, stromal, or immune neighbors into one measurement. That blending is not an artifact; it is the measurement. Treating a spot as a single pure cell overstates certainty.

Resolution is improving, but mixtures still matter. High‑density bead arrays like Slide‑seqV2 push to around 10 µm beads, which gets closer to single‑cell scale in some tissues while boosting sensitivity compared with earlier iterations. Still, even a 10 µm bead can straddle a cell boundary or capture stray transcripts from a neighbor, especially in crowded tissue.

At the other end of the spectrum, imaging‑based in situ platforms such as MERFISH or CosMx SMI directly localize transcripts at single‑cell or subcellular resolution using targeted probe panels. These approaches reduce mixture issues dramatically, but tradeoffs include targeted gene sets and specialized instrumentation; sequencing‑based assays remain attractive for whole‑transcriptome discovery and broad accessibility.

A major shift arrived with Visium HD. Instead of 55 µm circles, Visium HD arranges millions of 2 × 2 µm barcoded squares without gaps across the capture area. You can keep data at that fine scale or bin squares into larger “pixels” that match cell size or segmentation boundaries. While this brings single‑cell‑scale sampling, remember that even 2 µm features can cover cytoplasm from adjacent cells; a “bin” may still mix RNA if cells interdigitate. The technology elevates spatial detail, but does not abolish mixtures by itself.

The practical takeaway: plan analyses as if your spot or pixel likely contains contributions from multiple cells. Model mixtures first; assign crisp single‑cell labels only when evidence supports them.

From spots to cell types and tissue regions: how deconvolution and segmentation work

Two complementary tasks dominate early spatial analysis. First, you want to estimate which cell types contribute to each spot (deconvolution). Second, you want to delineate tissue regions or microenvironments that explain spatial structure (segmentation and spatial clustering).

Deconvolution reframes each spot’s expression as a weighted mixture of cell‑type profiles learned from a single‑cell RNA‑seq (scRNA‑seq) reference. Probabilistic methods like Stereoscope and RCTD estimate cell‑type proportions per spot using negative‑binomial or related count models. Others such as SPOTlight and SpatialDWLS use matrix factorization or weighted least squares with marker gene signatures. Bayesian and deep generative approaches, notably cell2location and DestVI, borrow information across spots, capture overdispersion, and even model within‑cell‑type states. In practice, these models agree on the big picture and differ at the margins, especially for rare or colocalized cell types.

Region identification leans on spatial priors and images. Algorithms like BayesSpace increase effective resolution by creating “subspots” that honor neighborhood structure, then cluster them to reveal finer tissue domains. This is particularly helpful on classic Visium data where each spot blends multiple niches. With Visium HD, you can segment nuclei on the histology image, bin 2 µm squares within each boundary, and derive cell‑centric profiles before deconvolution—shrinking the gap between sequencing‑based and imaging‑based granularity. Even so, recent benchmarking reminds us that “single‑cell resolution” is not automatically “single‑cell level”; careful segmentation and QC remain vital.

A useful mental model is to treat spots as pixels in a satellite photo. Deconvolution estimates the land‑cover mix in each pixel; spatial clustering draws region boundaries across pixels that share composition and gene programs. Neither step is perfect alone, but together they reconstruct cellular neighborhoods and tissue architecture with surprising fidelity.

Bring in single‑cell atlases: picking and using references that actually match your tissue

Because deconvolution depends on a reference, picking the right scRNA‑seq datasets matters as much as choosing an algorithm. Start with species, tissue, and compartment. Match adult versus developmental stage, healthy versus disease, and if possible, assay chemistry and processing (nuclei vs cells) to the spatial prep. When you control for these, the model spends less effort reconciling batch effects and more on biology.

Community resources help you find references fast. The Human Cell Atlas (HCA) Data Portal curates large, open scRNA‑seq collections and provides guides for exploring and downloading datasets by tissue and network. You can also tap CELLxGENE Discover via the cellxgene‑census Python API to filter for organism, tissue_general, and disease, and pull an AnnData reference in a few lines. These catalogs change quickly, so pin a census version and record dataset accession IDs for reproducibility.

Once you have a reference, two workflows dominate. Reference mapping aligns your new data to a labeled atlas and transfers annotations—think Azimuth in the Seurat ecosystem or Tangram for aligning sc/snRNA‑seq to spatial data. Deconvolution uses the same or similar references but outputs mixtures instead of one label per spot. Many teams run both: map to an atlas to standardize labels, then deconvolve spatial spots using those labels as constraints or priors.

Here’s a lightweight example of each idea.

First, a tiny Python sketch using scvi‑tools to run Stereoscope‑style deconvolution. It assumes you have an scRNA‑seq AnnData with cell‑type labels and a Visium/HD AnnData with raw counts.

import scanpy as sc
import scvi

# 1) Load reference scRNA-seq and train a negative binomial model
sc_ref = sc.read_h5ad("lung_scrna_ref.h5ad")            # cells × genes; sc_ref.obs["cell_type"] exists
scvi.model.SCVI.setup_anndata(sc_ref, labels_key="cell_type")
sc_model = scvi.model.SCVI(sc_ref)
sc_model.train(max_epochs=50)

# 2) Load spatial data and infer cell-type proportions per spot (Stereoscope)
st = sc.read_h5ad("patientA_visium.h5ad")               # spots × genes (raw counts)
st_model = scvi.external.SpatialStereoscope.from_rna_model(st, sc_model)
st_model.train(max_epochs=250)
props = st_model.get_proportions()                      # DataFrame: spots × cell types

# 3) Store in .obs for plotting
st.obsm["celltype_props"] = props.to_numpy()
st.uns["celltype_names"] = props.columns.tolist()

Second, a minimal snippet to discover an HCA‑like reference with CELLxGENE Census. This pulls healthy human lung cells from public atlases, returning an AnnData you can save or feed into your mapping/deconvolution toolchain.

import cellxgene_census as census
import anndata as ad

with census.open_soma(census_version="latest") as c:
    adata = census.get_anndata(
        c, organism="Homo sapiens",
        obs_value_filter="tissue_general == 'lung' and disease == 'normal'",
        column_names={"obs": ["cell_type", "assay", "tissue_general"]}
    )
adata.write_h5ad("hca_lung_reference.h5ad")

These are deliberately compact. In real projects you’ll harmonize gene nomenclature, perform QC, and evaluate mapping confidence scores before trusting any labels.

The state of the toolbox: what works well today

If your assay is classic Visium, expect supra‑cellular spots. Model mixtures explicitly and consider resolution enhancement. BayesSpace’s subspot strategy is a dependable first pass for sharpening boundaries before or after deconvolution, especially in layered tissues like cortex or stratified epithelia. The method’s Bayesian prior makes spatial coherence a feature, not a bug, and often clarifies crypts, ducts, and immune aggregates that blur at spot scale.

If you’re using Visium HD, lean into its fine grid. Segment nuclei on the histology, bin 2 µm squares within each cell polygon, and generate cell‑centric count tables. That makes downstream models behave more like scRNA‑seq with coordinates. Even then, be conservative: edge cells and tightly packed regions may still pull signal from neighbors, so proportion estimates remain useful for boundary cells. 10x documentation and community guides describe typical bin sizes and show how to fold segmentation into standard pipelines.

For deconvolution on sequencing‑based data, pick one of the widely benchmarked methods and keep the setup simple. Stereoscope and RCTD provide robust baselines and scale well; cell2location adds hierarchical Bayesian structure that improves estimates when you have many related cell states; DestVI models within‑cell‑type gradients, which helps when a single “macrophage” label masks inflamed versus resolving programs. Regardless of the choice, validate with orthogonal cues: region‑specific markers, known layer boundaries, or independent stains.

When mapping single‑cell references onto spatial data, Tangram remains a versatile option because it handles both low‑resolution Visium and high‑resolution targeted imaging like MERFISH. If you live in the Seurat/Azimuth world, reference mapping and label transfer give you fast, standardized annotations and well‑tested diagnostics to spot mismatches between your tissue and the atlas. Use those diagnostics. A perfect algorithm cannot rescue a poor reference.

Finally, keep an eye on sequencing‑based platforms beyond Visium. Slide‑seqV2 continues to mature in sensitivity and bead packing. Stereo‑seq pushes feature sizes down to the submicron range using DNA‑nanoball patterning; while not in every core facility, it illustrates how quickly resolution is improving on the sequencing side. Higher resolution reduces mixing but does not remove it; computational deconvolution and careful region modeling remain first‑class citizens.

Summary / Takeaways

Spots are not cells, and that’s okay. Sequencing‑based spatial assays measure mixtures by design; your analysis should reflect that reality. Start by understanding your platform’s geometry and what it implies for mixing. Use deconvolution to estimate cell‑type proportions per spot and spatial clustering or segmentation to define regions that explain variation. Bring in single‑cell references that truly match your tissue, and don’t be shy about validating with known markers and histology. As resolution climbs—from 55 µm spots to 2 µm squares and submicron arrays—the line between “spot” and “cell” gets thinner, but mixtures persist at boundaries and in dense tissue. The winning strategy is to combine better data with better models and keep biological plausibility front and center.

If you’re about to start a new analysis, pick one dataset today and try both flows: reference mapping for quick annotations, and deconvolution for mixture‑aware proportions. Then ask yourself one question on the resulting map: where would a pathologist draw the boundary? If your regions and proportions agree with anatomy and markers, you’re on the right track.