Spatial Transcriptomics Explained (2026)

Spatial transcriptomics is the set of technologies that measure which genes are switched on inside a tissue without throwing away the one piece of information older RNA methods discarded: where each cell sat. For decades, measuring gene expression meant grinding tissue into a slurry, sequencing the RNA, and reporting averages. You learned what a sample expressed; you lost the map. A tumor boundary, a cortical layer, a developing limb bud — all of it collapsed into a single column of numbers. Spatial transcriptomics keeps the slide intact, so you can ask not just what is expressed but where, next to what, and in what neighborhood. That shift — from a list to a map — is why pathologists, neuroscientists, and developmental biologists adopted it so fast. This article is a practitioner’s walkthrough: how the methods actually work, where they break, and how to read the data.

What this covers: the context against bulk and single-cell RNA-seq, the two technology families and their core trade-off, the wet-lab-to-analysis pipeline, real applications, the failure modes, and practical takeaways.

Context: bulk vs single-cell vs spatial

To see why location matters, line up the three generations of RNA measurement and ask what each one keeps and what it throws away.

Bulk RNA-seq takes a chunk of tissue, lyses every cell, and sequences the pooled RNA. The output is one expression profile per sample — an average over potentially millions of cells of dozens of types. Bulk is cheap, deep, and statistically clean, and it remains the right tool for many questions. But the average is a liar in heterogeneous tissue. If a tumor sample shows moderate expression of an immune gene, you cannot tell whether every cell expresses it a little or a few infiltrating T cells express it a lot. Those two scenarios mean completely different things clinically, and bulk cannot distinguish them.

Single-cell RNA-seq (scRNA-seq) solved the averaging problem. Tissue is dissociated into a suspension of individual cells, each cell is barcoded, and you get a separate expression profile for every cell — typically thousands of genes across tens of thousands of cells. This was transformative: cell-type atlases, rare-population discovery, trajectory inference. But dissociation has a hidden cost. The moment you break the tissue apart to make a single-cell suspension, every cell loses its address. You know a cell is a macrophage; you no longer know whether it was deep in a tumor core or sitting in healthy margin. Dissociation also stresses cells, biasing which ones survive and distorting stress-response genes.

Spatial transcriptomics is the third generation. It measures expression in situ — RNA is captured or imaged while the tissue architecture is preserved on the slide. You trade some depth or some gene-panel breadth (depending on the method) for the coordinate of every measurement. The payoff is enormous for any biology where position is the point: you can finally see that the immune gene is concentrated at the invasive tumor edge, that a signaling molecule forms a gradient across a developmental axis, or that two cell types are physically interacting rather than merely co-present in the same dissociated tube.

A useful mental model: bulk gives you the average, single-cell gives you the parts list, and spatial gives you the assembly diagram. The three are complementary, and the strongest studies in 2026 routinely combine them — using a deep scRNA-seq atlas to interpret a lower-depth spatial map, for instance.

How it works: the two families

Almost every spatial transcriptomics platform falls into one of two families, and the split runs all the way down to the physics of how RNA is read. Understanding which family a method belongs to tells you most of what you need to know about its resolution, gene coverage, and failure modes.

Both families start from a tissue section on a slide; sequencing-based methods route to whole-transcriptome readout at spot resolution while imaging-based methods route to targeted gene panels at subcellular resolution.

Sequencing-based methods

Sequencing-based spatial transcriptomics captures RNA on a spatially barcoded surface, then sequences it like any other library. The slide is tiled with a grid of capture features, each carrying a unique positional barcode. RNA released from the overlying tissue binds to the nearest feature, picks up that feature’s barcode, and gets reverse-transcribed. After sequencing, you decode each read’s positional barcode to place it back on the tissue. Because you are sequencing released RNA rather than imaging chosen targets, these methods are effectively unbiased and whole-transcriptome — you can detect any expressed gene, including ones you never thought to look for.

The trade-off is spatial granularity, which depends on how small and tightly packed the capture features are.

• 10x Visium is the most widely deployed platform. Its capture spots are 55 micrometers across, arranged on a hexagonal grid. At that size each spot overlaps several cells, so a Visium spot is a small neighborhood, not a single cell. You get genome-wide expression per spot but must computationally deconvolve mixed cell types. The Visium HD product moved to a continuous 2-micrometer grid that approaches single-cell scale.
• Slide-seq (and Slide-seqV2) packs randomly deposited 10-micrometer barcoded beads onto a slide, pushing resolution down to roughly cell scale at the cost of more involved bead-location decoding.
• Stereo-seq uses DNA nanoball arrays with spot-to-spot distances down to the sub-micrometer range and can cover very large fields, which made it the workhorse for whole-embryo spatial atlases. Raw nanoball resolution is subcellular, though capture efficiency means practical analysis often bins spots together.

The common signature of the family: broad gene coverage, resolution set by feature geometry, and a deconvolution problem when features are larger than cells.

Imaging-based methods

Imaging-based spatial transcriptomics never sequences anything. Instead it lights up individual RNA molecules directly in the tissue using fluorescent probes, then reads them out under a microscope. The dominant approach is multiplexed single-molecule fluorescence in situ hybridization. You design probes against a chosen panel of genes, each gene assigned a unique combinatorial barcode encoded across multiple rounds of imaging. Over successive rounds of hybridization, imaging, and probe stripping, each RNA molecule blinks out its barcode color-by-color, and software decodes the sequence back to a gene identity and an exact xy position.

Because you are imaging individual transcripts, resolution is subcellular — you can often see which transcripts are nuclear versus cytoplasmic. The catch is that you can only measure genes you put probes against. Panels in 2026 typically run from a few hundred to a few thousand genes, far short of the whole transcriptome.

• MERFISH (multiplexed error-robust FISH) pioneered the error-correcting combinatorial barcode scheme that lets hundreds to thousands of genes be distinguished despite imaging noise.
• 10x Xenium is an instrument-integrated, padlock-probe and rolling-circle-amplification system that delivers high sensitivity on panels typically in the hundreds of genes, with expanding whole-transcriptome offerings.
• CosMx (NanoString) targets large panels and offers protein co-detection, useful when you want transcripts and surface markers on the same section.

The common signature: pinpoint spatial and even subcellular resolution, high per-transcript sensitivity, but a fixed, pre-selected gene panel and longer imaging time per slide.

The core trade-off

Put the two families on the same axes and a clean tension appears: resolution versus gene-panel breadth. Sequencing-based methods give you the whole transcriptome but at coarser spatial features; imaging-based methods give you subcellular precision but only on a curated panel. The upper-right “broad and sharp” corner — every gene at single-molecule resolution across a whole slide — is the goal everyone is converging on, and 2026 has narrowed the gap (Visium HD pushing sequencing toward cell scale, whole-transcriptome imaging panels expanding), but no single platform fully owns that corner yet.

Sequencing-based platforms cluster toward broad gene coverage at coarser resolution while imaging-based platforms cluster toward subcellular resolution on narrower panels.

The practical consequence: your choice of platform should follow your biological question. If you are doing discovery and do not know which genes matter, you want the unbiased breadth of a sequencing method. If you have a defined panel and need to resolve individual cells in dense tissue — say, distinguishing tumor cells from infiltrating lymphocytes touching them — you want the imaging family. Treating the two as interchangeable is the most common planning mistake.

The data pipeline

Whichever family you pick, the journey from tissue block to biological insight follows a recognizable arc. Each stage has its own pitfalls, and errors propagate downstream, so the pipeline rewards care at the front end.

The pipeline runs left to right from wet-lab tissue prep through capture or imaging into the computational stages of segmentation, expression mapping, cell typing, and neighborhood analysis.

Tissue preparation and fixation. Everything depends on tissue quality. Sections are cut thin (commonly around 10 micrometers) and either fresh-frozen or formalin-fixed and paraffin-embedded (FFPE). RNA is fragile, so RNA integrity is the gatekeeping metric — degraded RNA yields sparse, noisy data no algorithm can rescue. FFPE compatibility, now standard on the major platforms, was a major unlock because it opened the world’s archived clinical blocks to spatial profiling.

Capture or imaging. This is where the families diverge. Sequencing methods permeabilize the tissue to release RNA onto the barcoded surface, then build and sequence a library. Imaging methods run their multi-round hybridize-image-strip cycles, accumulating a stack of fluorescence images. Either way, the wet-lab output is raw — reads with barcodes, or terabytes of microscopy images.

Cell segmentation. To assign expression to cells, you must first draw the cell boundaries. Segmentation uses a nuclear stain (DAPI) and often membrane markers, fed to algorithms such as Cellpose or platform-native tools. This is one of the hardest and most error-prone steps: in dense tissue, cells overlap, nuclei touch, and a boundary drawn wrong assigns a transcript to the neighbor. Segmentation errors are a leading source of artifactual “cell types” that are really two real cells merged or one cell split.

Mapping to an expression matrix. Now you collapse the raw data into the familiar cells-by-genes (or spots-by-genes) matrix, each entry a count. For sequencing methods you decode positional barcodes and tally unique molecular identifiers per feature; for imaging methods you assign each decoded transcript to a segmented cell. The output looks like a single-cell matrix with coordinates attached, which lets you reuse much of the scRNA-seq software ecosystem.

Quality control and normalization. Filter out low-count cells, likely doublets, and features in tissue tears or folds. Normalize for differences in capture efficiency. Spatial data adds wrinkles single-cell does not have — for instance, edge effects where tissue meets the slide border, and “spillover” where transcripts bleed between adjacent cells.

Cell typing and annotation. Cluster cells by expression and assign identities, usually by transferring labels from a matched scRNA-seq reference atlas. For coarse-feature methods like standard Visium, this step includes deconvolution: statistically estimating the mix of cell types contributing to each multi-cell spot rather than assigning one label.

Neighborhood and niche analysis. This is the stage that justifies going spatial in the first place. With every cell typed and placed, you can quantify which cell types sit next to which, identify recurrent multicellular “niches,” map ligand-receptor signaling between neighboring cells, and detect spatially variable genes whose expression depends on position. The biology that bulk and single-cell could never see lives here.

Biological interpretation. Finally, you tie the spatial patterns back to the question — a tumor’s immune-excluded versus immune-infiltrated regions, a brain region’s layered organization, a developmental gradient. Tools like Squidpy, Giotto, and the spatial extensions of Seurat and Scanpy support this end-to-end workflow.

Applications

Spatial transcriptomics earns its cost in problems where position carries the biology.

Tumor microenvironment. Cancer is the killer app. Tumors are ecosystems — malignant cells, immune cells, fibroblasts, blood vessels — and their spatial arrangement predicts how the tumor behaves and responds to therapy. Spatial methods reveal whether T cells penetrate the tumor or are stranded at its margin (“immune-excluded” versus “inflamed” tumors), a distinction that bulk averaging erases but that strongly tracks immunotherapy response. Researchers map the invasive front, find immunosuppressive niches that protect the tumor, and locate exactly where a drug target is expressed relative to the cells you want to kill.

Neuroscience. The brain is the most spatially organized tissue there is — cortical layers, nuclei, and circuits are defined by position. Spatial transcriptomics has driven large-scale brain cell atlases that place molecularly defined neuron types into their anatomical context, something dissociation-based methods structurally cannot do because they scramble the very architecture that defines neural identity. It also enables in-situ mapping of disease pathology, such as gene-expression changes in cells surrounding amyloid plaques.

Developmental biology. Development is spatial pattern formation — gradients, boundaries, and territories emerging over time. Whole-embryo spatial atlases, enabled especially by large-field methods, let researchers watch organ territories specify and trace how signaling gradients organize the body plan. Capturing the spatial map at successive timepoints turns a static snapshot into something close to a developmental movie.

Other fast-growing areas include immunology (germinal centers and lymphoid architecture), tissue regeneration, and infectious disease (host-pathogen interactions in situ).

For deeper molecular-engineering context on the tools reshaping this field, see our explainers on base editing and single-base CRISPR therapeutics and prime editing, both of which increasingly rely on spatial readouts to validate where edits land in tissue.

Trade-offs, limitations, and what goes wrong

Spatial transcriptomics is powerful but far from a solved instrument. The honest failure modes:

Resolution versus coverage, again. It bears repeating because it is the constraint everything else flows from. You cannot yet buy whole-transcriptome, single-cell, whole-slide data off the shelf at low cost. Every project picks a corner of the trade-off and lives with the consequences. Plan the analysis around what your platform actually resolves, not what you wish it resolved.

Segmentation is the silent error source. Because so much downstream analysis depends on correctly assigned cell boundaries, segmentation mistakes quietly corrupt cell typing and neighborhood statistics. In dense tissue, expect a meaningful fraction of “cells” to be merges or splits. Always sanity-check segmentation visually before trusting niche-level conclusions.

Sensitivity and dropout. Capture and detection efficiency is well below 100 percent. Many transcripts present in a cell are simply never recorded, producing zeros that are technical, not biological. This is worse for lowly expressed genes and means absence of signal is weak evidence of absence of expression.

Cost, throughput, and data volume. These are expensive experiments. Imaging-based runs can take a long time per slide and generate terabytes of images; sequencing-based runs cost serious money per sample for whole-transcriptome depth. Sample sizes in spatial studies are correspondingly small, which limits statistical power — be cautious about overgeneralizing from a handful of sections.

Two-dimensional slices of three-dimensional tissue. Most spatial transcriptomics is done on thin sections. A 2D slice through a 3D structure can mislead about true neighborhoods and connectivity. Serial sectioning and 3D reconstruction help but add cost and registration headaches.

Batch effects and reference dependence. Cell typing often leans on transferring labels from a separate scRNA-seq atlas; if that reference does not match your tissue or condition, annotations inherit its blind spots. And like all high-throughput assays, spatial data carries batch effects across slides, runs, and operators that must be modeled, not ignored.

Analysis is still maturing. Methods for spatially variable gene detection, niche identification, and cross-sample comparison are active research, and different tools can disagree. Treat any single algorithm’s output as a hypothesis to validate, not a verdict.

The related challenge of editing and then confirming molecular changes in context connects spatial methods to the gene-regulation toolkit — see our piece on CRISPR epigenetic editing and gene silencing without cutting DNA, where spatial readouts help show whether silencing actually took hold in the right cells.

Practical takeaways

• Let the question pick the platform. Discovery without a gene list points to a sequencing-based, whole-transcriptome method; resolving individual cells in dense tissue on a known panel points to an imaging-based method. This single decision shapes everything downstream.
• Budget for the matched single-cell atlas. Most spatial analyses are far stronger when paired with a deep scRNA-seq reference for cell-type annotation and, for coarse methods, deconvolution. Plan for both.
• Treat segmentation as a first-class result, not a preprocessing detail. Inspect it visually. Bad boundaries silently invalidate niche-level conclusions.
• Interpret zeros cautiously. Dropout is pervasive; missing signal rarely means a gene is truly off.
• Validate spatial patterns with a second method. The field’s analysis tools still disagree, so confirm key niches or spatially variable genes before building a story on them.
• Mind the dimensionality. A 2D section is not the whole tissue; be careful generalizing neighborhood claims to 3D.

FAQ

What is the difference between single-cell and spatial transcriptomics?
Single-cell RNA-seq dissociates tissue into individual cells and profiles each one deeply, but loses every cell’s location in the process. Spatial transcriptomics measures expression while the tissue stays intact, preserving each cell’s coordinates and neighbors. Single-cell gives you the parts list; spatial gives you the assembly map. They are complementary, and many studies use a single-cell atlas to interpret a spatial dataset.

Which is better, Visium or Xenium?
Neither is universally better; they sit on opposite sides of the core trade-off. Visium (sequencing-based) measures the whole transcriptome but at coarser, multi-cell spots. Xenium (imaging-based) resolves individual transcripts at subcellular scale but only on a pre-selected gene panel. Choose Visium for unbiased discovery and Xenium when you need single-cell resolution on known genes.

Does spatial transcriptomics work on FFPE samples?
Yes. FFPE compatibility is standard on the major 2026 platforms and was a pivotal advance, because it opened the vast archives of formalin-fixed clinical tissue blocks to spatial profiling. RNA integrity still matters — heavily degraded archival material yields sparse, noisy data — so sample quality assessment remains an essential first step.

What spatial resolution can these methods achieve?
It depends on the family. Imaging-based methods like MERFISH and Xenium reach subcellular resolution, distinguishing individual transcripts. Sequencing-based methods vary: standard Visium spots span about 55 micrometers (several cells), while Visium HD, Slide-seq, and Stereo-seq push toward single-cell or even sub-micrometer feature spacing, though practical resolution is also limited by capture efficiency.

Why not just use bulk RNA-seq if it is cheaper?
Bulk reports an average over all cells, which hides where and in which cells genes are expressed. In heterogeneous tissue like tumors, that average can be actively misleading — you cannot tell uniform low expression from a few cells expressing strongly. When position or cellular composition is part of the biological question, the spatial context that bulk discards is exactly the information you need.

Further reading

• 10x Genomics — Spatial gene expression overview — vendor documentation for the Visium and Xenium families.
• Nature Methods — Method of the Year 2020: spatially resolved transcriptomics — the field-defining editorial and primer.
• Moffitt, Lundberg & Heyn, Nature Reviews Genetics — “Spatial transcriptomics” review — a thorough technical review of the platform families and analysis.
• Related: base editing and single-base CRISPR therapeutics, prime editing explainer, and CRISPR epigenetic editing.