Spatial transcriptomics is a family of methods that measures gene expression while retaining information about where the measured molecules occur within a biological specimen. Most approaches detect RNA, particularly messenger RNA, in tissue sections or intact preparations and associate those measurements with coordinates, microscopic images, or anatomically defined regions. The resulting data connect molecular profiles with the arrangement of cells, tissue compartments, and local environments.
Spatial context matters because tissues are organized systems rather than interchangeable collections of cells. A cell’s activity can depend on its position relative to blood vessels, tissue boundaries, neighboring immune cells, or developmental signaling centers. Conventional bulk RNA sequencing averages expression across a sample, while dissociation-based single-cell RNA sequencing separates cellular profiles from their original locations. Spatial methods retain part of this organization, enabling the study of molecular gradients, cellular neighborhoods, and localized disease processes.
The field encompasses sequencing-based capture methods, multiplexed imaging, in situ sequencing, and region-selective profiling. These technologies differ in transcript coverage, spatial resolution, sensitivity, tissue compatibility, and scale. Their interpretation combines experimental molecular biology, microscopy, computational image analysis, and spatial statistics; no single measurement format provides all of these capabilities equally well.
Historical development and the meaning of spatial resolution
Spatial transcriptomics grew from techniques that located specific nucleic acids in cells. In situ hybridization, developed in the late 1960s, used complementary nucleic-acid probes to identify selected sequences while preserving specimen structure. Later fluorescence-based methods made individual RNA molecules detectable as discrete spots. In 1998, Robert Singer and colleagues demonstrated single-molecule RNA visualization using multiple fluorescent probes, helping establish a quantitative approach to RNA localization. These methods showed that transcripts could be counted and positioned within cells rather than merely detected in tissue extracts.
A central challenge was multiplexing: measuring many genes without exhausting the limited number of distinguishable fluorescent colors. Sequential hybridization, molecular barcodes, and repeated imaging provided solutions. Eng and colleagues introduced seqFISH in 2014; Chen and colleagues introduced MERFISH in 2015. These approaches assigned genes optical codes read across successive imaging rounds. In parallel, Je Hyuk Lee and colleagues reported fluorescent in situ sequencing, or FISSEQ, in 2014, demonstrating highly multiplexed sequencing of RNA-derived products within fixed cells and tissues.
The term “spatial transcriptomics” became closely associated with a 2016 study by Patrik Ståhl and colleagues. Their method placed tissue sections on an array of positionally barcoded capture probes, allowing sequencing reads to be mapped back to locations on the slide. The original array used capture spots approximately 100 micrometres in diameter, with 200-micrometre center-to-center spacing. Subsequent technologies increased feature density and introduced bead arrays, patterned surfaces, and microfluidic delivery of spatial barcodes. Slide-seq, reported in 2019, used approximately 10-micrometre beads; other methods subsequently reached smaller nominal capture features.
“Resolution” nevertheless has several meanings. A platform may specify the diameter of a capture spot, the spacing between features, the optical localization precision of a fluorescent molecule, or the accuracy with which transcripts can be assigned to individual cells. These are not equivalent. A small capture feature may contain very few detected molecules, while several neighboring features may need to be combined for reliable expression analysis. Conversely, accurate molecular coordinates do not guarantee accurate cell boundaries.
The scope of the field is similarly broader than a literal reading of “transcriptome” might suggest. Some assays survey RNA without a predetermined gene panel, whereas others measure a selected set of tens, hundreds, or thousands of genes. Both are commonly included because they retain spatial information and support expression-based tissue analysis. Distinguishing targeted from transcriptome-wide measurement, and nominal from effective resolution, is therefore essential when comparing studies.
Sequencing-based capture and positional barcoding
Sequencing-based spatial methods generally attach a location identifier to RNA-derived molecules before sequencing. A common design uses a surface coated with DNA oligonucleotides containing a spatial barcode, a unique molecular identifier, and an oligo-dT sequence that binds the polyadenylated tails of many messenger RNAs. A tissue section is positioned on the surface, imaged, and permeabilized. Released transcripts bind nearby probes and are copied into complementary DNA. Sequencing then identifies both the transcript and its associated location.
The spatial barcode distinguishes one capture feature from another. The molecular identifier distinguishes captured molecules from copies generated during amplification: reads sharing a gene assignment, spatial barcode, and molecular identifier can often be collapsed into a single molecular count. This reduces amplification-related distortion, although barcode errors, identifier collisions, and incomplete capture remain relevant. These counts estimate the molecules successfully captured and processed, not the total number originally present in the tissue.
Array construction determines how positions are encoded. Printed or lithographically patterned surfaces have predefined coordinates. Randomly deposited bead arrays require decoding so that each bead’s barcode can be matched to its physical location. Slide-seq uses this latter principle. High-density patterned technologies can provide much smaller capture features; Stereo-seq, reported in 2022, uses DNA nanoball arrays with submicrometre feature spacing. The original Visium assay used 55-micrometre spots, whereas Visium HD introduced a continuous array of 2-micrometre barcoded squares. These dimensions describe capture geometry, not automatically the size of an independently resolved biological unit.
Other strategies assign spatial information within the tissue. Deterministic barcoding in tissue, or DBiT-seq, reported in 2020, delivers two sets of barcodes through perpendicular microfluidic channels. Their intersections define a two-dimensional coordinate system. Region-selective approaches instead collect molecules from selected compartments. In digital spatial profiling, for example, oligonucleotide tags associated with probes can be released from optically selected regions and subsequently counted. Such measurements may preserve regional organization without supplying the position of every individual transcript.
Not all sequencing-based assays capture native polyadenylated RNA directly. Some use pairs of probes that hybridize to adjacent sequences in a selected transcript and are then joined into a product suitable for capture and sequencing. This design is particularly useful for fragmented RNA in archival tissue, but the measured gene set is determined by the probe panel.
The main advantage of sequencing-based capture is the ability to combine extensive gene coverage with familiar sequencing workflows. Its main constraints include incomplete molecular recovery, possible transcript displacement during processing, and mixtures of cells within capture areas. Increasing sequencing depth improves sampling only until the available library is substantially exhausted; it cannot recover molecules that were never captured.
Imaging-based transcript detection and in situ sequencing
Imaging-based spatial transcriptomics identifies RNA within preserved specimens using repeated cycles of probe binding, fluorescence measurement, or sequence readout. Rather than collecting molecules onto an external surface, these methods localize signals directly in the tissue. A suitable microscope records transcript-associated spots, and computational decoding assigns each spot to a gene. The resulting dataset can contain a gene label and two- or three-dimensional coordinates for every accepted detection, often accompanied by information about signal quality and cell assignment.
Single-molecule fluorescence in situ hybridization commonly uses multiple short probes that bind along a target RNA. Their combined fluorescence makes an individual molecule distinguishable from background. Multiplexed methods extend this principle by assigning each gene a pattern across imaging rounds and channels. If a transcript is detected in particular rounds but not others, the resulting sequence of positive and negative observations forms an optical barcode. Probe architecture and decoding schemes vary, but repeated measurement is the central mechanism.
MERFISH—multiplexed error-robust fluorescence in situ hybridization—uses codebooks designed to tolerate some detection errors. Appropriate spacing between valid codes can allow certain missing or spurious signals to be recognized or corrected. seqFISH and its extensions also use sequential imaging to distinguish many genes; seqFISH+, reported in 2019, demonstrated imaging of approximately 10,000 genes in individual cells. Commercial systems implement related multiplexed strategies with different probe chemistry, amplification, imaging, and gene-panel constraints. Their gene counts alone do not establish equivalent sensitivity or accuracy.
In situ sequencing reads short nucleotide sequences from localized nucleic-acid products. Some workflows use padlock probes whose ends hybridize to adjacent target sequences and are ligated to form circles. Rolling-circle amplification then produces localized copies that generate bright sequencing signals. Other methods construct broader libraries from RNA-derived complementary DNA. Targeted probe designs can provide high specificity, whereas broader approaches face additional challenges in capture efficiency, sequence identification, and optical crowding.
Imaging has distinctive advantages for studying subcellular organization. Transcript coordinates can distinguish nuclear from cytoplasmic distributions or reveal RNA in elongated neuronal processes. However, a bright spot is not automatically an independent RNA molecule: amplification chemistry, overlapping signals, background fluorescence, and decoding errors influence counting. At high transcript density, optical crowding makes nearby molecules difficult to separate. Repeated imaging also requires accurate registration despite stage drift and specimen deformation.
Targeted imaging depends strongly on panel design. A panel selected to distinguish known cell classes may identify those classes well while missing unexpected biological programs. Unmeasured genes cannot generally be recovered from the images, even if computational models predict their likely expression. Imaging-based assays therefore exchange some discovery breadth for direct localization and often high detection efficiency for selected targets.
Specimen preparation, histology, and cell assignment
The specimen is a major determinant of spatial transcriptomic data quality. RNA begins to degrade after tissue collection, and delays, temperature changes, freezing, fixation, and storage can alter molecular recovery. Fresh-frozen tissue and formalin-fixed, paraffin-embedded tissue require different chemistries. Formalin preserves morphology through crosslinking but also modifies and fragments nucleic acids, complicating extraction and direct capture. Probe-based assays can accommodate some of this damage by interrogating relatively short target regions.
Section thickness sets another important constraint. Many slide-based assays use sections on the order of several to roughly ten micrometres thick, although exact requirements vary. A thin section contains only part of many cells and excludes much of their cytoplasm or processes. A thicker section can increase molecular content but may superimpose cells along the depth axis and make probe penetration or imaging more difficult. Consequently, expression measured from a section is not necessarily the expression of an intact cell.
Permeabilization must balance release and positional retention. Insufficient permeabilization can leave RNA inaccessible; excessive treatment can damage tissue or allow molecules to move before capture. Diffusion is especially consequential when a platform’s features are smaller than the distance transcripts can travel during processing. Tissue composition also matters: fibrous, fatty, calcified, or necrotic material can behave differently from more uniform soft tissues. A protocol that works for one organ cannot be assumed to produce equivalent recovery in another.
Histological images provide the anatomical framework for interpretation. Hematoxylin and eosin staining can reveal tissue compartments, while fluorescent nuclear stains, membrane markers, or immunohistochemistry can assist cell identification. Registration aligns these images with capture coordinates or transcript images. Misalignment can associate expression with the wrong tissue boundary, particularly near narrow structures such as small vessels, epithelial layers, or densely packed cell interfaces.
Assigning molecules to cells requires image segmentation. Nuclear segmentation is often easier than identifying complete cell boundaries, but expanding a nuclear outline by a fixed distance is unreliable for irregular cells. A neuron, for example, may extend RNA-containing processes far beyond its nucleus, while adjacent immune cells may have very little intervening space. Algorithms that combine nuclear staining, membrane information, morphology, and transcript distributions can improve assignment, but their assumptions still affect the final cell-by-gene matrix.
Quality assessment must therefore examine tissue and molecular measurements together. Relevant indicators include tissue integrity, detected molecules per area, genes per feature, background outside the tissue, positive and negative probe controls, and consistency across replicate sections. A high total transcript count can coexist with poor localization or segmentation. Conversely, a small, genuinely low-RNA cell should not be rejected solely because it contains fewer detected molecules than a large neighboring cell.
Data structures, preprocessing, and reference integration
Most spatial datasets combine an expression matrix, coordinates, images, and experimental metadata. Rows may represent capture spots, bins, segmented cells, or selected regions; columns generally represent genes. Imaging assays may additionally retain a much larger table of individual transcript coordinates. These distinctions matter because a “sample” in a statistical analysis can mean a molecule, a cell, a capture feature, a tissue section, or an organism, each with a different relationship to biological replication.
For sequencing-based data, preprocessing includes read alignment or transcript assignment, spatial-barcode correction, molecular-identifier deduplication, and the construction of count matrices. For imaging data, it includes image registration, spot detection, optical-code decoding, background rejection, and transcript-to-cell assignment. Errors introduced at these stages propagate into downstream results. Retaining raw or minimally processed information makes it possible to reassess decisions such as bin size, segmentation boundaries, or decoding thresholds.
Normalization attempts to distinguish biological expression differences from variation in measured material and assay efficiency. A common method scales counts according to the total detected molecules per observation, followed by a transformation that reduces the dominance of abundant genes. However, total RNA content can itself vary biologically with cell size, activity, or identity. Normalizing every cell to the same total may therefore remove genuine differences as well as technical variation. Spatial changes in tissue thickness, permeability, or probe accessibility create additional confounding that simple scaling does not necessarily resolve.
Reference integration often uses single-cell or single-nucleus RNA sequencing to identify cell types. In cell-type deconvolution, the expression of a capture spot is modeled as a mixture of reference cell-type profiles. A simplified expectation is
where is the observed count for gene in spot , represents sampling intensity, represents the contribution of cell type , is its reference expression, and represents background. Real methods differ in their distributional assumptions, platform corrections, and constraints on these quantities.
A fitted contribution is not automatically a literal cell fraction. Cell types differ in RNA content, and reference data may omit rare populations or contain states absent from the spatial specimen. Closely related cell types can have nearly indistinguishable profiles, especially when only a small gene panel is measured. Dissociation can also induce stress-associated expression or preferentially lose fragile cells.
Label transfer and computational prediction are similarly distinct from direct observation. A model may infer the likely identity of a spatially measured cell or predict unmeasured genes from a reference atlas. Those outputs are useful hypotheses, but their reliability depends on reference coverage and model assumptions. Predicted expression should remain distinguishable from experimentally detected RNA in both analysis and visualization.
Spatial statistics, tissue domains, and inferred interactions
Spatial transcriptomics makes location part of the analytical model. An initial task is often to identify genes whose expression varies systematically across tissue rather than fluctuating independently among observations. Tests may use neighboring locations, spatial covariance functions, or explicit anatomical coordinates. Spatial autocorrelation describes the tendency for nearby observations to be more similar—or less similar—than expected under a specified null model. Moran’s , for example, compares deviations from the mean across a weighted neighborhood structure.
The definition of “neighbor” has biological consequences. A fixed-radius graph connects observations within a chosen distance, whereas a nearest-neighbor graph gives each observation a similar number of connections. Both may connect cells across an empty lumen or a sharp tissue boundary unless anatomy is incorporated. Distances measured in a two-dimensional section also omit separations along the third dimension. Apparent adjacency in an image is therefore an approximation to physical proximity in the original tissue.
Spatial-domain detection identifies regions with related molecular patterns, often integrating expression and coordinates. Methods may combine dimensionality reduction, graph-based clustering, probabilistic models, or machine learning. Including location can reduce scattered classifications and reveal coherent layers or compartments. Excessive spatial smoothing, however, can erase rare cells, narrow boundaries, or abrupt transitions. A domain recovered by an algorithm may reflect cell-type composition, a shared cell state, an experimental artifact, or some combination of these.
Cellular-neighborhood analysis asks whether particular cell types occur together more often than expected. A suitable null model must preserve relevant structure. Randomly shuffling labels across an entire organ can make ordinary anatomical compartmentalization appear to be a specific cellular attraction. Restricted permutations, comparisons within compartments, and replication across specimens can help distinguish local association from broader tissue organization. Neighborhood radius also changes interpretation: immediate cell contact differs from membership in the same larger niche.
Expression of a ligand in one cell population and its receptor in another is frequently used to infer potential communication. Spatial proximity strengthens the plausibility of some interactions, but does not demonstrate signaling. RNA abundance may not predict protein abundance; the ligand may require processing or secretion; a receptor may need several subunits; and pathway activation may depend on other conditions. Contact-dependent interactions require different distance assumptions from signals carried through extracellular fluid.
Statistical dependence is especially important when comparing experimental groups. Thousands of cells from one individual do not constitute thousands of independent biological replicates. Models must account for cells nested within regions, sections, and individuals, as well as spatial dependence. Multiple-testing correction is necessary when evaluating many genes or interactions, and effect sizes and uncertainty should accompany significance measures. Spatial association establishes a pattern; causal interpretation generally requires additional experimental evidence.
Tissue atlases, development, and nervous-system organization
One major use of spatial transcriptomics is constructing molecular tissue atlases: reference descriptions that connect cell identities and states to anatomical positions. Dissociation-based surveys can reveal extensive cellular diversity, but many populations are difficult to interpret without knowing their location. Spatial mapping can distinguish related cells occupying different layers, identify transitional boundaries, and determine whether a transcriptional state is dispersed or restricted to a particular structure.
The brain provides a prominent example because its function depends strongly on anatomical organization. In the cerebral cortex, excitatory neurons occupy layers with characteristic molecular and projection properties, while inhibitory neurons and glial cells show additional regional patterns. Spatial measurements can relate expression-based classifications to cortical depth, white matter, nuclei, and other structures. Large mouse-brain studies published in 2023 combined extensive single-cell profiling with spatial mapping, demonstrating how molecular cell classifications can be placed within a common anatomical framework.
Subcellular localization adds another dimension in the nervous system. RNA in dendrites and axons can support local protein synthesis far from the cell nucleus. Imaging-based methods can reveal distributions that would be obscured by whole-cell averaging. Interpretation requires care because a process crossing a tissue section may belong to a cell whose body lies elsewhere. Assigning all nearby transcripts to the closest nucleus can therefore misrepresent both cell identity and RNA localization.
Developmental studies use spatial data to investigate how cellular states become organized into tissues. Sections collected at successive embryonic stages can reveal changing domains of transcription-factor expression, boundaries between developing structures, and localized signaling programs. Tomo-seq, introduced in 2014, combined RNA sequencing with ordered tissue sectioning to reconstruct expression patterns along anatomical axes. Later capture- and imaging-based methods provided finer-scale views of developing organs and embryos.
These measurements can connect developmental expression programs to physical position, but they do not directly show cell movement or ancestry. A trajectory inferred from expression similarity is not equivalent to a lineage established by tracking or genetic labeling. Likewise, a set of embryos collected at different stages gives a population-level time series rather than a continuous record of one organism. Spatial observations become more informative when combined with lineage tracing, controlled perturbations, and established embryological landmarks.
Three-dimensional atlases can be assembled from serial sections or generated with methods that image thicker specimens. Serial reconstruction requires alignment across sections despite cutting distortion, missing material, and changing anatomy. Comparisons across individuals require additional registration to a shared coordinate system. A reference atlas is consequently both a measurement resource and a model of correspondence: it must represent genuine anatomical variation rather than forcing every specimen into an unrealistically identical arrangement.
Cancer, inflammation, and localized tissue injury
In cancer research, spatial transcriptomics examines how malignant cells, stromal cells, blood vessels, and immune populations are arranged within tumors. A tumor may contain molecularly distinct regions associated with invasive margins, necrosis, fibrous stroma, or differing immune infiltration. Bulk expression can mix these components, while dissociation-based measurements lose their relationships. Spatial measurements can separate changes in cell composition from changes in the activity of a particular population, provided the assay and analysis resolve the relevant cells.
Studies of breast, pancreatic, colorectal, and other cancers have mapped localized malignant-cell programs and their surrounding microenvironments. In pancreatic ductal adenocarcinoma, for example, spatial profiling has been used to investigate relationships among tumor epithelium, fibroblast-rich stroma, and immune compartments. Fibroblasts can express different combinations of inflammatory and extracellular-matrix genes, and their distribution can vary relative to tumor structures. Such patterns describe tissue organization without establishing that one population caused a neighboring state.
Immune organization is another central application. A tumor can contain many T cells overall yet have relatively few within malignant epithelial regions. Spatial assays can distinguish infiltration into tumor compartments from accumulation at their edges. They can also characterize organized aggregates containing B cells, T cells, and other immune populations. Identification of a tertiary lymphoid structure generally requires appropriate morphological and cellular evidence; a cluster of lymphocyte-associated transcripts alone is not sufficient to establish its full organization.
In inflammation and tissue injury, spatial analysis can map where immune responses overlap with damaged parenchyma, vascular structures, or repair-associated cells. Kidney studies, for example, can distinguish expression patterns associated with glomeruli, tubules, interstitium, and vessels. Lung tissue presents a different architecture involving airways, alveolar regions, and vascular compartments. These examples illustrate why anatomical annotation is integral to biological interpretation rather than a decorative layer added to an expression map.
Pathogen-associated RNA can sometimes be detected alongside host transcripts, but detection depends on assay design. A targeted host-only panel will not identify an unrepresented pathogen. Polyadenylated-RNA capture does not sample all microbial transcripts equally, and many bacterial RNAs are poorly represented by such chemistry. Apparent pathogen reads also require controls for contamination, mapping ambiguity, and environmental background before they can be assigned biological meaning.
Clinical translation introduces requirements beyond exploratory research. A candidate biomarker must be evaluated for reproducibility, specimen compatibility, analytical performance, and relevance in independent patient groups. Differences in tissue processing, sampled region, disease stage, and treatment history can alter spatial patterns. An association with an outcome in a research cohort does not by itself establish a validated diagnostic test or a basis for treatment selection.
Multimodal integration, benchmarking, and unresolved limits
Spatial transcriptomics is increasingly combined with measurements of proteins, chromatin accessibility, metabolites, and tissue morphology. RNA describes one layer of cellular regulation, but it does not directly measure enzyme activity, receptor occupancy, protein modification, or metabolic flux. A cell may contain abundant RNA for a protein that is poorly translated, rapidly degraded, or inactive. Combining transcript measurements with antibody-based protein detection can therefore clarify cell identity and reveal disagreements between molecular layers.
Multimodal data can be collected from the same section, adjacent sections, or separate specimens registered to a common anatomy. These designs provide different strengths of correspondence. Same-section assays can associate molecular layers with the same cells, but their chemistries must be mutually compatible. Adjacent sections preserve approximate anatomy while sampling different portions of cells and sometimes different small structures. Computational alignment cannot eliminate those physical differences, and agreement between modalities should not be assumed simply because their images overlap.
Benchmarking is difficult because there is no complete, error-free transcript map of most tissues. Independent single-molecule assays, standardized samples, synthetic controls, and carefully annotated histology can test specific aspects of performance. Relevant metrics include sensitivity, false-positive detection, gene coverage, molecular localization, cell-assignment accuracy, reproducibility, tissue area, processing time, and cost. A method that performs well on abundant markers may not perform equally well on rare transcripts or densely packed tissue.
Several trade-offs remain fundamental. Smaller spatial units usually contain fewer captured molecules, making their expression profiles noisier. Larger units increase molecular counts but mix cells and blur boundaries. Broad gene coverage supports discovery, while targeted panels can devote greater detection efficiency to selected transcripts. Large-area imaging increases anatomical coverage but produces substantial image data and demands more computation. Consequently, a single ranking based on nominal resolution or number of genes conceals important differences in experimental purpose.
Data standards and provenance affect whether results can be compared or reused. Expression matrices alone are insufficient when conclusions depend on segmentation, image registration, or tissue annotation. Reusable datasets benefit from physical coordinate units, assay and panel descriptions, specimen metadata, image-processing details, and explicit distinctions between observed and predicted values. Human datasets also require appropriate protection of participant information, particularly when expression data are linked to clinical records or other potentially identifying data.
A deeper limitation is temporal. Most spatial assays measure a fixed specimen at one point in time. Observed patterns combine developmental history, prior exposures, cell movement, current state, and technical sampling. They do not directly reveal the order in which those influences acted. Perturbation experiments, longitudinal designs where feasible, lineage information, and independent functional measurements are needed to test proposed mechanisms. Spatial transcriptomics can locate the molecular consequences of a process with considerable detail; establishing how that process unfolded remains a separate experimental problem.