Spatial Transcriptomics Technologies

Evolution of Spatial Transcriptomics (2016-2025)

Figure 1. Timeline of spatial transcriptomics technology development. Upper track (teal): sequencing-based methods from original ST (2016) through Visium, Slide-seq, Seq-Scope (2021), Stereo-seq (2022), to Visium HD (2023). Lower track (orange-coral): imaging-based methods from MERFISH foundation through STARmap, seqFISH+, MERSCOPE, Xenium, and CosMx. Gold stars indicate breakthroughs. Bottom scale shows resolution progression from 100μm to subcellular.

What is Spatial Transcriptomics?

Spatial transcriptomics refers to a collection of methods that measure gene expression while preserving the spatial location of transcripts within tissues. Unlike traditional RNA-seq that requires tissue homogenization (losing spatial information) or single-cell RNA-seq that requires cell dissociation (losing tissue context), spatial transcriptomics maintains the native tissue architecture while profiling the transcriptome.

Why Spatial Context Matters

• Tissue Architecture Understanding: Cells behave differently based on their location - a cancer cell at the tumor edge behaves differently than one in the center
• Cell-Cell Interactions: Understanding signaling between neighboring cells requires knowing their spatial relationships (ligand-receptor pairs)
• Microenvironment Mapping: Immune cells, stromal cells, and parenchymal cells form functional niches that can only be studied in spatial context
• Disease Heterogeneity: Tumors, inflammatory lesions, and developmental abnormalities show spatial patterns that inform diagnosis and treatment
• 3D Organization: Tissues are three-dimensional; understanding cell clustering and connectivity patterns requires volumetric analysis

Three Major Technology Categories

Figure 2. Platform classification by spatial resolution (y-axis) and gene coverage (x-axis). Upper left (coral): imaging-based platforms (Xenium, CosMx, MERSCOPE, STARmap) with subcellular resolution and targeted panels. Upper right (teal): high-resolution sequencing (Stereo-seq, Seq-Scope, Visium HD, HDST) with whole transcriptome. Middle right (golden): near single-cell sequencing (Slide-seqV2). Lower right (light teal): classic sequencing (Visium, GeoMx WTA). Lower left: ROI-based targeted (GeoMx panel mode). Stars denote subcellular + whole transcriptome breakthroughs.

Technology Comparison Guide

Figure 3. Visual comparison of spatial resolution from multi-cell regions to individual transcripts. Left to right: GeoMx DSP ROI (50-600μm, 50-100+ cells), Visium (55μm, 5-10 cells), Slide-seqV2 (10μm, 1-3 cells), Visium HD/HDST (2μm, single-cell), Stereo-seq/Seq-Scope (0.5μm, subcellular, ★), Xenium/CosMx/MERSCOPE (<0.1μm, individual transcripts). Yellow outlines: sequencing-based; coral outlines: imaging-based. Sequencing platforms capture ~20K genes; imaging platforms detect 500-6,000 genes.

Figure 4. Technical workflows across three approaches. Left (Array-based Capture): tissue permeabilization → mRNA capture on barcoded spots → reverse transcription → library prep → demultiplexing; output: Spot × Gene matrix. Center (In Situ Barcoding): spatial barcode sequencing → tissue overlay → mRNA capture → sequencing → high-resolution mapping; output: subcellular spatial matrix (★). Right (Imaging-based): probe hybridization → platform-specific detection (combinatorial/RCA/ISS) → cyclic imaging → barcode decoding → single-molecule localization; output: (x,y,gene_id) coordinates. Bottom table compares resolution, genes, output format, FFPE compatibility, and throughput.

Resolution vs Coverage Trade-off

Choosing the Right Technology

Figure 5. Decision flowchart for platform selection. Starting from sample type: FFPE samples → subcellular (Xenium, CosMx, MERSCOPE), multi-cell ROI (GeoMx), or whole transcriptome (Visium HD). Fresh frozen → Discovery path by resolution: subcellular (Stereo-seq, Seq-Scope ★), single-cell (Visium HD, Slide-seqV2), multi-cell (Visium); Hypothesis-driven path → imaging platforms or GeoMx; Protein+RNA → CosMx, Xenium, GeoMx; Rare cells → GeoMx DSP. Quick reference table summarizes recommendations by experimental need.

Landmark Papers Organized by Methodological Approach

1. Sequencing-Based Spatial Arrays (Capturing RNA onto surfaces)

Why these are grouped: These methods work by placing a tissue section onto a slide printed with barcoded capture spots (like beads or DNA nanoballs). The RNA is released from the tissue, captured on the surface, and then sequenced off-slide. The resolution is determined by the physical size and density of the capture spots on the array.

2. Imaging-Based In Situ Technologies (Visualizing RNA directly in tissue)

Why these are grouped: Instead of capturing RNA onto a slide, these methods leave the RNA inside the tissue cells. They use fluorescent probes and sophisticated microscopes to capture images of individual RNA molecules over multiple rounds of hybridization. They offer very high (subcellular) resolution but usually target a predefined panel of genes rather than the whole transcriptome.

3. Beyond 2D RNA: Multi-Modal, 3D, and Temporal Expansions

Why these are grouped: These papers represent the "next frontiers" of spatial biology. They push beyond standard 2D transcriptomics by adding new modalities (like proteins or active translation), analyzing thick 3D tissues, or adding the fourth dimension of time (live-cell tracking).

4. Benchmarking, Biological Application, and Future Outlook

Why these are grouped: These papers are less about inventing a new method and more about making sense of the existing ones. They include direct head-to-head comparisons of platforms, deep biological applications demonstrating real-world utility, and high-level perspectives on where the field is headed.

Common Analysis Workflows

Figure 6. Data output formats across platforms. Panel A (Sequencing-based): Spot × Gene count matrix with coordinates; high-res platforms require segmentation (⚠). Formats: MTX, H5AD, Parquet. Panel B (Imaging-based): transcript coordinate table (transcript_id, x, y, gene); requires cell segmentation. Formats: Parquet, CSV, Zarr. Panel C (Unified): all platforms converge to Cell × Gene matrix with spatial coordinates in AnnData/H5AD structure containing X (counts), obs (cell metadata), var (gene metadata), obsm (spatial coordinates), layers, and uns (images).

Figure 7. Platform-specific computational challenges. Low-res Sequencing (Visium): deconvolution, reference integration; tools: cell2location, RCTD; difficulty: medium. High-res Sequencing (Visium HD, Stereo-seq): cell segmentation (⚠), scalability, binning strategy; tools: Cellpose, Baysor; difficulty: high. Imaging-based: cell segmentation, assignment errors (~5-15%), panel limitations; tools: Cellpose, Baysor, StarDist; difficulty: high. ROI-based (GeoMx): low sample size, cell mixtures in AOIs; difficulty: low-medium (computation), high (statistics). Shared challenges (all platforms): batch effects, normalization, spatial statistics, cell-cell interactions.

1. Data Processing & Quality Control

• Image Registration: Align multi-round images using fiducial markers or amplicon positions
• Deconvolution: Enhance resolution using PSF-based methods (Huygens, deconvolution algorithms)
• Spot/Barcode Calling: Detect transcripts using local maxima detection, barcode decoding with error correction
• Cell Segmentation: Watershed, CellPose, or Baysor for single-cell boundaries

2. Cell Type Annotation

• Reference-Based: Transfer labels from scRNA-seq using FuseMap, Harmony, or Cell2location
• Unsupervised: Louvain/Leiden clustering on spatial features followed by marker gene validation
• Deconvolution: SpatialDecon, RCTD, or Cell2location for spot-level cell type mixtures

3. Spatial Analysis

• Neighborhood Analysis: Squidpy for cellular interaction scores, Delaunay triangulation for neighbors
• Spatial Domains: BayesSpace, HMRF, or SpaGCN for identifying tissue regions
• Ligand-Receptor: CellChat, CellPhoneDB for cell-cell communication inference
• 3D Analysis: Distance transform watershed, volumetric clustering, spatial statistics

Key Software Packages

🛠️ Hands-On Practice

The walkthrough below takes a Visium slide from raw spot counts to spatial statistics using Python/Scanpy/Squidpy — QC, normalization, clustering, plotting clusters back onto the tissue image, building a spatial neighbor graph, ranking spatially variable genes with Moran's I, and testing which cluster pairs sit next to each other. The same Squidpy calls apply to imaging-based data (Xenium, MERFISH, CosMx) once transcripts have been segmented into cells; the differences are noted at each step.

Environment & packages

Squidpy sits on top of Scanpy/AnnData and adds the spatial graph, spatial statistics, and image-aware plotting. For Xenium, MERFISH, or any multi-modal experiment (image + transcripts + polygons), spatialdata and its readers (spatialdata-io) are the better container because they keep coordinate transforms, segmentation shapes, and image pyramids aligned rather than flattening everything into adata.obsm["spatial"].

# conda / mamba recommended
conda create -n spatial python=3.10 -y
conda activate spatial

pip install scanpy squidpy leidenalg
# multi-modal / imaging platforms (Xenium, MERFISH, CosMx)
pip install spatialdata spatialdata-io spatialdata-plot
# spot deconvolution (GPU strongly recommended)
# pip install cell2location            # or run RCTD / spacexr in R
# transcript-aware segmentation
# pip install cellpose                 # Baysor is a separate Julia binary

Hardware. A single Visium section (~3–5k spots) runs comfortably on a laptop with 8 GB RAM. Visium HD (2 µm bins), Stereo-seq, and Xenium runs reach 10⁵–10⁶ cells or bins and need 32–128 GB RAM plus backed/Zarr-chunked I/O. Deconvolution with cell2location is a variational-inference fit and effectively needs a GPU; RCTD is CPU-only but slower.

Data structures & formats

• Spot-based (Visium, Visium HD, Slide-seq, Stereo-seq) — the observation is a barcoded area, not a cell. A Visium spot is ~55 µm across (100 µm center-to-center) and typically contains 1–10 cells, so adata.X rows are cell-type mixtures; recovering per-cell-type abundance requires deconvolution (cell2location, RCTD, SPOTlight) against a matched scRNA-seq reference. Slide-seq (10 µm beads), Stereo-seq (0.5 µm DNB, usually binned), and Visium HD (2 µm bins) are finer but still not cell objects until binned or segmented.
• Imaging-based (Xenium, MERFISH, CosMx) — the primary output is a transcript table (transcript_id, x, y, z, gene, qv), and cells only exist after segmentation. Resolution is genuinely single-cell, so segmentation quality is the limiting factor, not spot size. The trade-off is a targeted panel (~300–1,000 genes for Xenium/CosMx, ~150–500 for MERFISH) rather than the whole transcriptome.
• AnnData — the convergence point for both families: adata.X (counts), adata.obs (spot/cell metadata), adata.var (genes), adata.obsm["spatial"] (an n × 2 array of x/y coordinates in the full-resolution image frame), and adata.uns["spatial"] (H&E images, scale factors, spot diameter).
• Scale factors matter. scalefactors_json.json maps full-resolution pixel coordinates to the hires/lowres thumbnails; plotting with the wrong img_key/scale factor silently offsets every spot from the tissue.
• SpatialData / OME-Zarr — for Xenium and multi-sample studies: labels (segmentation masks), shapes (cell/nucleus polygons), points (transcripts), images, and tables held in one store with explicit coordinate transformations, so a re-segmentation does not desynchronize the table from the image.
• Spatial graph — sq.gr.spatial_neighbors() writes adata.obsp["spatial_connectivities"] and adata.obsp["spatial_distances"]. Every downstream spatial statistic (Moran's I, neighborhood enrichment, Ripley) reads from this graph, so its construction is a modeling choice, not boilerplate.

Minimal code walkthrough

Load a Visium H&E dataset, run standard QC and normalization, cluster with Leiden, project clusters back onto the tissue, build a hexagonal-grid neighbor graph, score spatially variable genes with Moran's I, and test which clusters are spatial neighbors more often than chance.

import scanpy as sc
import squidpy as sq
import numpy as np

# 1. Load data. The bundled demo (mouse brain, Visium H&E) ships coordinates,
#    the H&E image, and scale factors already wired into AnnData.
adata = sq.datasets.visium_hne_adata()
#    For your own SpaceRanger output instead:
#    adata = sc.read_visium("outs/")          # expects filtered_feature_bc_matrix.h5
#    adata.var_names_make_unique()
#    For Xenium / MERFISH, read via spatialdata_io and pull out the cell table:
#    import spatialdata_io as sdio
#    sdata = sdio.xenium("xenium_output/")    # keeps transcripts + polygons + image
#    adata = sdata.tables["table"]            # cells x panel genes, post-segmentation

print(adata)                                  # obsm["spatial"], uns["spatial"] present
print(adata.obsm["spatial"][:3])              # full-res pixel coords, NOT microns

# 2. QC. Same metrics as scRNA-seq, but interpret per SPOT, not per cell:
#    a low-count spot is often edge/low-cellularity tissue, not a dead cell.
adata.var["mt"] = adata.var_names.str.startswith(("MT-", "mt-"))
sc.pp.calculate_qc_metrics(adata, qc_vars=["mt"], inplace=True, log1p=True)
sc.pl.violin(adata, ["total_counts", "n_genes_by_counts", "pct_counts_mt"],
             jitter=0.4, multi_panel=True)

#    Filter gently. Aggressive thresholds carve holes in the tissue map and
#    break the neighbor graph. Inspect counts IN SPACE before choosing cutoffs.
sq.pl.spatial_scatter(adata, color="total_counts", img_alpha=0.5)
adata = adata[adata.obs["total_counts"] > 500].copy()
sc.pp.filter_genes(adata, min_cells=10)

# 3. Normalize + log1p. Keep raw counts around: deconvolution (cell2location,
#    RCTD) and most spatially-variable-gene models require raw integers.
adata.layers["counts"] = adata.X.copy()
sc.pp.normalize_total(adata, inplace=True)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, n_top_genes=2000)
#    NOTE: skip HVG selection for targeted panels (Xenium/MERFISH) — every gene
#    in a 300-plex panel was already chosen to be informative.

# 4. Expression-space clustering (spatially agnostic on purpose — this is the
#    baseline you compare spatial-domain methods like BayesSpace/SpaGCN against).
sc.pp.pca(adata, n_comps=50)
sc.pp.neighbors(adata)                        # EXPRESSION knn, not spatial
sc.tl.umap(adata)
sc.tl.leiden(adata, resolution=1.0, key_added="cluster")

# 5. Put the clusters back on the tissue. This is the plot that tells you
#    whether transcriptomic clusters correspond to real anatomy.
sq.pl.spatial_scatter(adata, color="cluster", size=1.3, img_alpha=0.6)

# 6. Spatial neighbor graph. Visium spots sit on a hex lattice -> 6 neighbors.
sq.gr.spatial_neighbors(adata, coord_type="grid", n_neighs=6)
#    Imaging platforms (irregular cell centroids) use a generic graph instead:
#    sq.gr.spatial_neighbors(adata, coord_type="generic", delaunay=True)
#    or radius-based, which is more interpretable in microns:
#    sq.gr.spatial_neighbors(adata, coord_type="generic", radius=30.0)
print(adata.obsp["spatial_connectivities"].shape)

# 7. Spatially variable genes via Moran's I. Positive I = neighboring spots
#    have similar expression, i.e. the gene forms spatial structure.
#    Run on the normalized layer; permutations give the empirical p-value.
sq.gr.spatial_autocorr(adata, mode="moran", genes=adata.var_names[adata.var.highly_variable],
                       n_perms=100, n_jobs=4)
svg = adata.uns["moranI"].sort_values("I", ascending=False)
print(svg.head(10)[["I", "pval_norm_fdr_bh"]])
sq.pl.spatial_scatter(adata, color=list(svg.index[:4]), img_alpha=0.4)

# 8. Neighborhood enrichment: which cluster pairs are adjacent more (or less)
#    often than expected under label permutation on the SAME graph?
sq.gr.nhood_enrichment(adata, cluster_key="cluster")
sq.pl.nhood_enrichment(adata, cluster_key="cluster", method="ward")
#    Complementary summaries on the same graph:
#    sq.gr.co_occurrence(adata, cluster_key="cluster")   # vs. distance, graph-free
#    sq.gr.ripley(adata, cluster_key="cluster", mode="L")

# 9. Deconvolution — REQUIRED for spot-based platforms before any statement
#    about cell types. A Leiden "cluster" on Visium is a spot NEIGHBORHOOD, not
#    a cell type. Fit cell2location (or RCTD in R) against a matched scRNA-seq
#    reference, then treat the per-spot abundances as the cell-type signal:
#    import cell2location
#    cell2location.models.RegressionModel.setup_anndata(ref_adata, labels_key="cell_type")
#    ... fit reference signatures, then Cell2location(adata_vis, ...).train()
#    -> adata.obsm["q05_cell_abundance_w_sf"]  # spots x cell types
#    Imaging platforms skip this: after segmentation, cells are annotated
#    directly with standard label transfer.

adata.write_h5ad("visium_spatial_processed.h5ad")

Common pitfalls & tips

• A Visium spot is not a cell. At ~55 µm each spot averages several cells, so a "cluster" is a tissue neighborhood and a co-expression pattern within a spot is not co-expression within a cell. Any cell-type claim on Visium/Slide-seq/Stereo-seq data needs deconvolution (cell2location, RCTD) or explicit binning/segmentation first — and the deconvolution is only as good as the scRNA-seq reference it is given, which must cover every cell type actually present in the section.
• Segmentation errors drive imaging-platform artifacts. On Xenium/MERFISH/CosMx, transcripts assigned to the wrong cell (bleed-over between touching cells, or nucleus-expansion masks swallowing neighboring cytoplasm) manufacture chimeric profiles — the classic symptom is "T cells expressing tumor markers." Typical assignment error rates are ~5–15%. Compare a nucleus-expansion baseline against a transcript-aware method (Baysor) or a membrane-stain-based Cellpose model before trusting rare hybrid populations.
• Absent ≠ not expressed on targeted panels. A gene missing from a Xenium/MERFISH/CosMx panel is undetected, not unexpressed. Never run whole-transcriptome tools (HVG selection, unfiltered GSEA, marker-based automated annotation trained on full transcriptomes) on panel data without accounting for the restricted gene universe, and never call a marker negative when it was simply never probed.
• Registration and autofluorescence artifacts masquerade as biology. Misaligned image-to-coordinate registration shifts spots off tissue and produces "expression" in empty regions; tissue folds, edges, and autofluorescent structures (elastin, lipofuscin, RBCs) generate hot spots that survive QC. Always overlay counts on the H&E (sq.pl.spatial_scatter with img_alpha) and check that high-signal regions correspond to real tissue before interpreting them.
• Batch effects live across slides, sections, and capture areas. Section thickness, permeabilization time, and slide-to-slide chemistry shift total counts substantially; on multi-sample designs integrate (Harmony/scVI) or model the slide as a covariate, and never build one spatial graph across physically separate sections — library_key in sq.gr.spatial_neighbors() keeps graphs per-section.
• Moran's I is a property of the graph, not just the gene. The same gene can be "significantly spatially variable" or not depending on n_neighs, radius, or Delaunay vs. grid. Fix the graph choice on biological grounds (a radius in microns matching the interaction scale you care about), report it, and sanity-check that the top hits look structured when plotted rather than trusting the ranking alone. The same caveat applies to nhood_enrichment, whose z-scores are permutation-based and scale with cluster size.
• Do not compare absolute counts across platforms. Detection efficiency differs by an order of magnitude between Visium, Slide-seq, Stereo-seq, and imaging platforms, and imaging platforms count individual transcript molecules rather than sequencing reads. Cross-platform comparisons should be made on relative/normalized quantities, rank correlations, or spatial pattern agreement — never on raw counts per unit area.