Introduction

The concept of spatial biology is not new. For decades, researchers have used microscopy to obtain spatial information on tumours, relying on targeted protein-based techniques such as immunohistochemistry (IHC) and immunofluorescence, and thus characterize tumours for clinical and research purposes. These studies have revealed that certain histological patterns have predictive and/or prognostic value. For example, stromal infiltration of T cells is linked to a favourable prognosis1 and is predictive of response to neoadjuvant chemotherapy2 in patients with triple-negative breast cancer (TNBC). Limitations of these techniques include: the requirement for preselection of target molecules (and reagents), with a limited representation of complex tumour-associated states not easily defined by a small number of markers; suboptimal resolution; and issues related to bias in sample selection.

Advances in genomics have demonstrated increasing utility to improve precision in cancer diagnosis and prognostication, and guide novel therapeutic approaches3. A technological foundation for many of these genomic applications is high-throughput sequencing, which enables extensive readout of nucleic acid sequences. However, many genomic approaches and open-access resources, such as The Cancer Genome Atlas and the International Cancer Genome Consortium, rely on information generated from measuring the average genomic profiles in biological specimens containing large numbers of cells (bulk approaches), which do not fully capture the underlying cellular heterogeneity of those samples. In the past decade, advances in single-cell genomic approaches, including single-cell RNA sequencing (scRNA-seq), have provided extensive insights into the complex cellular and molecular landscape of tumours, and hold potential to improve application of genomic-based precision medicine.

scRNA-seq, a technique that has been comprehensively reviewed elsewhere4,5, provides transcriptome-wide information but requires the dissociation of tissue samples, consequently leading to the absence of spatial context and loss of fragile cell and tissue types during dissociation. As a result, essential information on tissue architecture and cellular interactions is lost. Over the past decade, a range of spatial profiling techniques have been developed to characterize the genome, transcriptome and proteome. Spatial transcriptomic technologies, which provide spatially resolved transcriptomic (SRT) data, are powerful tools to explore the spatial architecture of tumours and the tumour microenvironment (TME). These technologies can resolve gene expression at subcellular spatial resolution. Other complementary techniques, such as multiplexed imaging using proteomics, can overcome certain technical challenges faced by current SRT technologies to reveal novel biological insights with clinical implications. In this Review, we summarize the landscape of spatial technologies with a focus on SRT, provide an overview of key results from the application of spatial profiling techniques in cancer research, and discuss opportunities and challenges to integrate these findings into clinical practice.

Overview of spatial technologies

Spatial transcriptomic technologies

scRNA-seq versus SRT

For two decades, bulk RNA-seq was used extensively in comparative studies and large atlas projects to classify and characterize biological samples, yielding a considerable amount of data that is a valuable resource for biological and clinical investigations. In 2009, a new type of technology emerged that made it possible to study individual cell transcriptomes, known as scRNA-seq6. scRNA-seq enabled the identification of cell-to-cell variation in gene expression and subsequent transcriptomic classification and characterization of known and novel cell types. The introduction of the first scRNA-seq assays for tens to hundreds of cells was followed by the rapid development of more-scalable alternatives to profile hundreds of thousands of cells. scRNA-seq quickly became an invaluable tool for biological research and was selected as the Method of the Year by Nature Methods in 2013 (refs. 7,8,9,10). Although the utility of scRNA-seq in describing the cellular components of tumours is unquestionable, this technology has the considerable challenge that it cannot preserve information about the physical organization of cells, a vital aspect of many biological processes.

A few years after the first scRNA-seq technologies were introduced, several new methods capable of gene expression profiling of tissue sections were developed. These technologies, known as SRT or spatial transcriptomic technologies, complemented the scRNA-seq methods by placing biological phenomena in their physical context. SRT technologies have demonstrated potential to uncover biological and molecular processes in neuroscience11,12,13, developmental biology14,15,16, immunology17,18, cancer19,20,21 and many other fields. As a result of its utility in biological research, SRT was selected as the Method of the Year by Nature Methods in 2020 (refs. 22,23,24,25). SRT technologies overcome many limitations of scRNA-seq by profiling gene expression levels in situ. Different experimental methods have been devised to achieve this goal, which differ in crucial aspects such as resolution, instrumentation and complexity (the number of transcripts that can be targeted simultaneously) (Fig. 1). SRT technologies can be broadly classified into imaging-based and in situ capture-based methods.

Fig. 1: Overview of single-cell and spatial transcriptomics.
figure 1

Single-cell RNA sequencing (scRNA-seq) protocols involve isolating individual cells and their RNA, typically into separate wells or droplets using microfluidic devices. After reverse transcription of RNA into cDNA, a cell-specific barcode is added to the cDNAs, enabling pooling of cDNA from multiple cells, which can then be analysed using next-generation sequencing (NGS) technologies. Computational analysis has a pivotal role in constructing cellular taxonomies and characterizing cell types. In contrast to scRNA-seq, spatially resolved transcriptomic (SRT) technologies facilitate spatially resolved profiling of gene expression. In methods using in situ capture, RNA molecules are applied onto arrays of spatially barcoded spots that capture mRNA by hybridization. Subsequently, RNA molecules are converted into cDNA directly on the array surface, which integrates the spatial barcode into the cDNA molecule. Similar to scRNA-seq, these cDNA molecules are then sequenced using NGS. Imaging-based SRT technologies use fluorescent probes to target selected RNA molecules, the precise locations of which can be detected through microscopy by measuring fluorescent signals.

Full size image

Imaging-based SRT

Imaging-based techniques provide the highest spatial resolution and can resolve the precise position of individual RNA molecules in tissue sections. Many of these technologies are based on single-molecule fluorescence in situ hybridization principles, whereby individual transcripts are detected using complementary DNA probes attached to a fluorophore. Once the probes have bound to their target molecules, the intracellular location of the transcripts can be determined by measuring light emitted from the fluorophore with microscopic imaging. Using various barcoding strategies, this strategy can be multiplexed by running several rounds of hybridization and imaging, detecting different transcripts in each round. One of the major limitations of imaging-based methods is the reduced multiplexing capacity, often limited to a few hundred targets owing to optical limitations in the imaging process26. Nonetheless, in the past few years, instrument manufacturers have achieved advances to create panels encompassing thousands of genes, such as Xenium and MERSCOPE.

In situ capture-based SRT

In situ capture-based SRT technologies are an untargeted alternative to imaging-based techniques because they sequence the entire protein-coding transcriptome. Achieving transcriptome-wide coverage has the advantage of providing a more comprehensive understanding of cellular mechanisms. In situ capture-based technologies use various innovative strategies to capture the transcripts in situ onto microscopic capture elements, or ‘spots’, arranged in symmetrical patterns on a glass slide. A tissue section is positioned on top of the patterned array and permeabilized to release RNA molecules, allowing them to diffuse downward onto the spots. The spots are coated with DNA probes that bind to the poly(A) tail of mRNA molecules and, consequently, these technologies can capture and quantify virtually the entire protein-coding transcriptome. A study published in 2023 described the near transcriptome-wide spatial detection of gene expression using a number of gene-specific probes tiled across target transcripts rather than poly(A)-based capture, enabling the detection of non-polyadenylated molecules and analysis of fragmented RNA such as that commonly found in formalin-fixed paraffin-embedded (FFPE) samples27. Although this approach enables substantially wider transcriptome coverage than imaging-based techniques, the size of the spots sets constraints on the type of biological processes that can be investigated. The dimensions of the spots range from a few hundred nanometres to several hundred micrometres depending on the platform. Given that the average diameter of a mammalian cell is ~10 μm, the spot sizes range between subcellular to supracellular (Table 1). Moreover, spatial resolution, defined as the centre-to-centre distance between spots, sets limitations on the range at which cellular biological processes can be investigated. Another important limitation of in situ capture-based technologies is their low capture efficiency (Table 1). Most of these technologies rely on vertical diffusion of transcripts onto the spots, which is less efficient than direct analysis of transcripts in tissue. Consequently, when measuring transcripts with low abundance, imaging-based techniques provide a distinct advantage. The clear trade-off between spot size, spatial resolution, capture efficiency and transcriptome coverage must be considered when choosing which SRT technology to use for an experiment, and this choice will determine the physical scale and transcriptomic complexity at which biological processes can be studied. Apart from resolution and complexity, we note that other aspects of SRT technologies influence their popularity, including the cost of equipment and reagents, turnaround time per sample, field-of-view, user-friendly analysis tools, options for multimodality, training required to use them and commercial availability.

Table 1 Comparison of different spatial transcriptomic technologies
Full size table

Analysis of SRT data

Data generated with SRT technologies has provided new opportunities for the comprehensive exploration of tissue landscapes. This innovative approach enables researchers to delve into data-driven methodologies, facilitating the construction of maps that reveal the intricate spatial distribution of genetic information within tissues. Nevertheless, harnessing the insights from SRT data necessitates a nuanced approach. Additional processing steps come into play when analysing cellular information derived from SRT data. The gene expression measurements obtained at subcellular spatial resolution using imaging-based methods are typically partitioned to generate cellular profiles through a process known as cell segmentation. Cell segmentation involves identifying and delineating individual cells within an image, then allocating specific genetic transcripts to these identified cells. Cell segmentation algorithms often leverage computer vision and machine learning techniques to automate tasks because manual segmentation can be time-consuming. Most algorithms operate on images of cells with DAPI-stained nuclei, which can be used for a broad range of cell types but tend to oversimplify their morphology. To mitigate this issue, some methods leverage protein stains for specific cell types of interest to improve the partitioning28. Although cell segmentation is the most common approach to partition SRT data prior to downstream analysis, segmentation-free methods can assign cell type labels to individual pixels of an image29.

Gene expression measurements aiming to achieve supracellular spatial resolution present a different set of challenges. These measurements encompass a mixture of cells, requiring computational deconvolution methods to infer the abundance of distinct cell types. Whether dealing with subcellular or supracellular spatial resolution, SRT data analysis often benefits from incorporating insights from scRNA-seq or single-nucleus RNA sequencing (snRNA-seq) datasets. scRNA-seq remains the gold standard for in-depth characterization of cellular states, and a substantial collection of annotated scRNA-seq datasets are open access, such as CELLxGENE and Human Cell Atlas. Cell signatures identified from scRNA-seq data are frequently used to estimate the abundance of distinct cell types from mixed gene expression profiles.

Integration of scRNA-seq and SRT data

Several computational tools have been developed to deconvolve SRT data using scRNA-seq cell signatures, facilitating the mapping and exploration of the cellular composition in tissue sections30,31. These tools are commonly used in SRT data analysis, although several challenges must be considered. Importantly, achieving a meaningful integration of scRNA-seq and SRT data hinges on ensuring comparable coverage of cell types between datasets; however, this is often difficult to achieve owing to variations in sample input, sample handling and technological discrepancies. Furthermore, within complex scRNA-seq datasets, cell type populations are often delineated hierarchically. This classification typically begins with broad categories, such as cancer, stromal and immune cells, at the highest level and gradually progresses to more specific distinctions (such as specific cell types or cell states) at the lowest level. Deconvolving the SRT data at each hierarchical level is feasible, yet the task becomes progressively more challenging when discerning between cell states with increasingly similar gene expression profiles. Sample-matched scRNA-seq and/or snRNA-seq and SRT data could offer a tailored solution for heterogeneous cancer samples that could capture some patient-specific nuances in gene expression.

Despite these challenges, the integration of scRNA-seq and SRT data unlocks a multitude of possibilities, empowering researchers to tap into the extensive repertoire of single-cell data and thereby simplifying the comprehension of cell-type compositions within tissues. Additionally, this integration facilitates the discovery of co-localization patterns and provides insights into intricate intercellular communication events, such as putative receptor–ligand interactions. In cancer research, this knowledge can be used to tailor therapeutic strategies, predict response to treatment and develop targeted interventions based on spatially informed molecular signatures.

Navigating the complexities of single-cell transcriptomics and SRT data successfully requires a multifaceted expertise encompassing biological knowledge, proficiency in bioinformatics, statistical methods and programming skills. These large datasets demand meticulous processing and analysis, typically using specialized software packages and libraries, many of which are written in the R or Python programming languages. Notably, Seurat and Scanpy are recognized as two of the most widely used frameworks for developing end-to-end data analysis workflows32,33. A wide range of data analysis tools are available (Table 2), and their applications in clinical and biomedical research have been reviewed elsewhere34,35.

Table 2 Computational tools commonly used for analysis of SRT data
Full size table

Spatial proteomic technologies

For decades, conventional protein-based methods, such as IHC and immunofluorescence, have been the primary tools used to characterize tumours. These methods are, however, limited by their ability to capture the full complexity of cellular diversity within tumours and the TME because they can only target a small number of spatially resolved proteins. In the past few years, developments in spatial proteomic technologies have focused on expanding this palette, using multiplexed proteomic detection to provide a more comprehensive picture of tumours and the TME. Most of these platforms use either fluorophore or metal tags. Fluorophore-based methods commonly use iterative imaging acquisition, which involves sequential rounds of antibody-based detection to increase the number of protein targets. Examples include multiplexed IHC, tissue-based cyclic immunofluorescence, InSituPlex and MACSima imaging cyclic staining36,37,38,39. Other methods such as co-detection by indexing (CODEX) use primary antibodies labelled with DNA barcodes to reduce the time associated with multiplexed imaging, enabling the detection of >50 proteins40,41. Fluorescence-based methods provide spatial information with high resolution, but the number of fluorophores that can be multiplexed is limited by spectral overlap. These assays are also affected by autofluorescence that results in decreased sensitivity. Metal-based imaging mass cytometry and multiplex ion beam imaging can potentially overcome this problem using detection of metal isotope-labelled antibodies with laser desorption mass spectrometry, enabling the measurement of proteins without interference42,43. Finally, new advances in mass spectroscopy imaging are enabling the unbiased assessment of diverse biomolecules from intact tissues at increasingly fine resolution44,45. The technical aspects of spatial profiling methodologies involving proteomics have been comprehensively reviewed elsewhere46,47,48,49.

Cellular architecture

Over the past few years, increasing awareness of tumour ecosystems has driven the emergence of new paradigms in oncology, such as leveraging stromal remodelling to enhance drug perfusion50, proposing new immunotherapy approaches51 or targeting malignant cell heterogeneity52. These advances rely on high-resolution multimodal analyses of the tumour and the TME. Multimodal analysis of dissociated single cells (for example, integrating gene expression and chromatin architecture) remains the best method to define specific and cell type-related signatures. The growing amount of single-cell transcriptomic data released over the past few years has led to a more refined taxonomy of cancer cells and tumour-associated cells, and a deeper understanding of their cellular hierarchies. These results form an integral part of high-resolution spatial profiling to capture multidimensional data on cellular architecture within the TME.

Cancer cells

Single-cell analyses can systematically resolve the differences in neoplastic phenotypes within a tumour. Profiling of non-malignant, premalignant and malignant samples has elucidated transcriptional and epigenetic changes that occur during malignant transformation, such as increased numbers of stem-like cells and chromatin accessibility associated with HNF4A motifs in non-malignant colon epithelial cells in their transformation to colorectal cancer (CRC) cells53. In patients with breast cancer, spatially resolved single-cell methods have demonstrated that genomic evolution from in situ to invasive disease occurs within ducts before neoplastic cells escape from the basement membrane, and the invasive tumour is established by comigration of multiple clones through the basement membrane into adjacent tissues54. Across several solid tumour types, remarkable intratumour heterogeneity is a common feature, revealing that neoplastic cells within established tumours can adopt highly variable states characterized by large-scale differences in gene expression55,56,57,58,59. This diversity probably affects the expression of biomarkers, cellular behaviour and response to treatment60,61. To systematically characterize the spectrum of gene expression programmes activated within a tumour, analyses of large datasets have identified numerous ‘meta-programmes’ activated by cancer cells, reflecting diverse cellular processes and lineage-specific patterns60. These cellular patterns indicate that distinct functions are performed by unique subsets of cells within a tumour and that specific programmes are linked. For example, in head and neck squamous cell carcinomas (SCCs), an individual cancer cell rarely expresses gene signatures associated with proliferation and epithelial-to-mesenchymal transition (EMT) states, even though these features often co-occur within a given tumour55. By contrast, EMT and an interferon-activated state are often correlated in the same cancer cell across several solid tumour types61,62, suggesting that similar regulatory programmes control the adoption of these states. A deeper understanding of the ‘rules’ governing different phenotypic states could help to develop a cellular model for drug synergy. The diverse states existing within tumours could partially arise in a stochastic manner but also as a result of genomic diversity, which can rapidly evolve even in the absence of strong selective pressure, particularly during therapy63.

Immune cells

The diverse responses of different tumour types to immunotherapies have prompted numerous investigations into the tissue-specific features of the immune TME. Traditionally, CD8+ and CD4+ T cells along with macrophages have been considered the most abundant immune cell populations in tumours, whereas other cell types, such as B cells and natural killer cells, have received less attention. Single-cell analysis has enabled researchers to classify immune cells into subpopulations at a higher resolution, revealing a remarkable diversity of cell states and unveiling novel clusters with well-defined functions64,65,66,67. For example, tumour-associated macrophages (TAMs) were conventionally classified as having either an M1 pro-inflammatory or an M2 anti-inflammatory phenotype. However, clustering of scRNA-seq data has uncovered a spectrum of TAM states with biological and clinical relevance that do not simply resemble M1 or M2 cells68,69. Furthermore, the integration of specific techniques such as single-cell T cell receptor (TCR) sequencing with scRNA-seq has linked T cell phenotypes with TCR clonotypes, enabling a detailed examination of roles of clonal expansion and replacement, and how phenotypic states are shaped by TCR stimulation and environmental stimuli70. Trajectory analysis has further dissected T cell state transitions, providing insights into the dynamic changes of these cells in the TME71. For example, a study revealed that CD8+ T cell exhaustion, a common consequence of chronic exposure to antigens, arises through more than one mechanism72.

Cancer-associated fibroblasts

Cancer-associated fibroblasts (CAFs) modulate tumour progression through synthesis and remodelling of the extracellular matrix, and influence angiogenesis, invasion, metabolic reprogramming and immune regulation73. Single-cell studies have helped to unravel the heterogeneity of CAFs74. Novel subsets of CAFs with distinct transcriptional, functional and spatial properties have been reported across multiple solid tumours55,75,76. Some of these studies found unique subsets of CAFs that had not been well-characterized previously. For example, a novel subpopulation of antigen-presenting CAFs (apCAFs) that express MHC class II and CD74 was found in pancreatic ductal adenocarcinomas (PDACs) from mouse models and from patients76. These apCAFs exhibit characteristics of professional antigen-presenting cells and can cross-present antigens to T cells. A similar subpopulation of apCAFs has also been observed in CRC77 but not in breast cancer75.

The metastatic TME

Single-cell profiling also enables the exploration of intertumour heterogeneity and characteristics of the metastatic TME. In various tumour types, although the expression profile of malignant cells in metastases largely matches that of cancer cells from the corresponding primary tumours, the stromal and immune cell populations have distinct differences55,78. For example, the composition of myeloid cells changes remarkably from primary to metastatic tumours in oesophageal SCCs, with a transition from monocyte predominance to macrophage predominance79. Similarly, across several specimens from patients with melanoma, nodal metastases and mucosal melanomas had the greatest abundance of TAMs and monocytes, whereas naevi and acral melanomas and brain metastases had reduced immune cell infiltration, albeit the latter were enriched in myeloid cells80. Correlative spatial transcriptomic studies of liver metastases from CRC have further revealed that the metastatic TME undergoes spatial reprogramming of immunosuppressive cells such as MRC1+CCL18+ TAMs81. These findings might explain the common clinical observation of differential treatment responses between different organ sites. Furthermore, tracking spatiotemporal evolution from primary to metastatic tumours can provide insights into which subclones are more likely to metastasize and, in particular, into the underlying mechanisms that could potentially be therapeutically targeted82.

Spatial architecture

One of the major limitations of single-cell technologies based on the analysis of cells in suspension is the lack of information on the structural organization of cells within a tissue, which has increasingly been recognized as a factor that contributes to the heterogeneity and variable therapeutic response observed in most cancers. Spatial profiling provides a more comprehensive view of the intricate cellular makeup of solid tumours, enabling annotation of different cell types and their locations, and examination of the various layers of spatial organization (Fig. 2). These layers include cell co-localization and cellular interactions, microenvironmental niches containing cellular neighbourhoods, and specialized microanatomy structures, such as tertiary lymphoid structures (TLS).

Fig. 2: Layers of information extracted using spatial profiling.
figure 2

Spatially resolved technologies enable the extraction of different layers of information from a tumour specimen. a, Cellular composition can be inferred from single-cell RNA sequencing (scRNA-seq) data using cell-type mapping techniques. b, Analysis of cell co-localization can be used to assess co-expression of genes encoding ligand–receptor pairs. c, Gene expression patterns can be directly associated with morphological data from histology images. d, Gene expression patterns can be analysed across large tissue structures, for example, to delineate cellular composition from a tumour core towards the leading edge. e, Guided by data on cell morphology, a tissue section can be segmented into spatial neighbourhoods, facilitating the extraction of unique gene expression profiles from each segment.

Full size image

Cell localization and co-localization

Researchers can investigate the physical location and co-occurrence of diverse cell populations in the TME by using spatial techniques to map the cell composition of tumours. This approach provides valuable insights into the cellular interactions and communication networks that drive tumour progression and immune responses within the TME. Spatial deconvolution has revealed that cell subpopulations have distinct spatial enrichment and co-enrichment. In breast cancer, CAF subsets such as myofibroblastic CAFs (myCAFs) and inflammatory CAFs (iCAFs) have differing spatial distributions: myCAFs are typically found in invasive cancer regions whereas iCAFs are dispersed across invasive cancer, stromal and tumour-infiltrating lymphocyte-aggregated regions62,83. Consistent with their immunoregulatory properties, iCAFs co-localize with different lymphocyte and B cell populations, whereas myCAFs correlate with CD8+ T cells, suggesting a functional relevance to invasive cancers with high tumoural infiltration of lymphocytes62. In another study of breast cancer, vascular CAFs were exclusively found around endothelial cells in vessel-like structures in line with their pro-angiogenic features, and reticular-like CAFs surrounded aggregated immune cells, consistent with their proposed functions of regulating naive T cells in lymphoid tissues84. In cutaneous SCCs, CAFs, TAMs and regulatory T (Treg) cells were found at the cancer cell–stroma interface. By contrast, CD8+ T cells and neutrophils were largely excluded from tumours, suggesting that cells at the cancer cell–stroma interface might influence the access of effector T cells to tumours85. In PDAC, ducts were enriched in a subpopulation of TAMs expressing CD163 and MS4A4, whereas regions containing cancer cells and stromal cells were preferentially enriched in TAMs expressing IL1B, consistent with their functions as tissue-resident and inflammatory TAMs in these regions, respectively86. In head and neck SCCs, stromal cell subsets were found in close proximity to cancer cells expressing a partial EMT-related gene signature55. Subsequent ligand–receptor analyses indicated a potential crosstalk mechanism between these adjacent populations that promoted EMT55. A pan-cancer study reinforced this notion by showing that cancer cells undergoing EMT tend to co-localize with CAFs and endothelial cells at the interface of the tumour61. In the melanoma immune TME, the presence of less common cell lineages, such as B cells and Treg cells, was always accompanied by that of more common cell lineages, such as cytotoxic and T helper cells, but not vice versa80. A similar hierarchical organization of immune cells was also noted in TNBC87. These hierarchical orderings suggest a coordinated antitumour immune response, whereby immune cells are recruited to tumour sites in a context-dependent manner80. A deeper understanding of the ‘rules’ governing the organization of the TME could lead to the development of novel strategies for the stratification of patients for treatment with immunotherapies.

Cellular interactions

In addition to co-localization, ligand–receptor interactions can provide indirect evidence for cellular interactions. Spatial mapping of breast cancer specimens showed co-localization of iCAFs with several lymphocyte subpopulations, and interactions were supported by ligand–receptor analysis revealing an enrichment in immunoregulatory ligand–receptor pairs expressed by adjacent iCAFs and lymphocytes62. In cutaneous SCCs, a population of tumour-specific keratinocytes found within a fibrovascular niche at the leading edges of tumours was identified as a hub for intercellular communications, with extensive autocrine and paracrine interactions with CAFs, endothelial cells and myeloid cells85. In PDAC, inflammatory fibroblasts co-localized with a cancer cell subcluster expressing stress response-related genes, which suggests that these cancer cells could have a role in eliciting a response involving stromal and immune cells86. In CRC, CAFs expressing FAP and TAMs expressing SPP1 co-localize, contributing to a desmoplastic microenvironment through mechanisms promoting extracellular matrix and collagen fibril deposition and organization, and activation of TGFβ signalling88. In hepatocellular carcinoma, interactions between CAFs expressing POSTN, TAMs expressing FOLR2 and endothelial cells expressing PLVAP create a so-called oncofetal niche that is associated with increased EMT, early recurrence and improved response to immune-checkpoint inhibitors (ICIs)89.

Several studies of the immune TME have identified immunosuppressive niches characterized by heterotypic interactions between TAMs and other cell types. In CRC liver metastases, TAMs are educated by malignant cells to shift towards immunosuppressive states via a mechanism mediated by a crosstalk between cancer cells expressing CD47 and TAMs expressing SIRPA81. In breast cancer, lipid-associated macrophages and CXCL10high TAMs, all expressing high levels of PD-L1 and PD-L2, were seen juxtaposed to PD-1+ lymphocytes, creating an immunosuppressive niche62. An independent study of breast cancer specimens similarly revealed co-localization of CXCL10high TAMs and T cell subsets, often in the presence of a type I interferon signal, suggesting that a spatially restricted type I interferon response could be related to macrophage-induced recruitment of certain T cell subsets83. In cutaneous melanomas, a spatially restricted immunosuppressive TME forms along the invasive cancer cell–stroma interface, shaped by PD-1–PD-L1-mediated cell contacts involving TAMs, dendritic cells (DCs) and T cells90. This milieu is juxtaposed with cytotoxic T cell synapses with melanoma cells in fields of tumour regression located a few millimetres away in the same specimen, demonstrating the co-occurrence of invasion and immunoediting90.

Cellular neighbourhoods

On a larger scale, heterotypic cells can be organized in multicellular neighbourhoods or compartments with distinct cellular compositions and gene expression features, suggesting that cellular phenotypes are shaped by their local niches. These multicellular communities have distinct malignant, stromal and immune features91,92,93. Investigators performed spatial mapping of immune cells across 13 different cancer types and consistently identified four distinct regions: two cancer regions enriched, respectively, in proliferating cancer cells and SPP1+ TAMs, a stromal region harbouring Treg cells and terminally exhausted CD8+ T cells, and an immune region with aggregates of B cells and DCs94. Analysis of communication pathways revealed that malignant, immune and stromal cells work in concert to recruit specific cell subpopulations through secretion of cytokines and chemokines92. Inhibitory receptors are also expressed, highlighting the presence of negative feedback within these hubs92.

At the tissue level, analysis of the centre and periphery of tumours has revealed differences in gene expression patterns between these regions. Pathways activated in the centre of tumours are mainly linked to altered cellular metabolism, endocytosis and hypoxia, whereas those activated in the tumour periphery tend to be related to stress and inflammation95. A comparison of primary brain tumours (glioblastomas) and brain metastases revealed distinct cellular topographies and interactions, suggesting that the interactions of glioblastoma and brain metastases with the surrounding brain tissue are fundamentally different despite sharing a common tissue niche96.

TME landscapes and clinical correlates

Treatment response and resistance

A major challenge for the efficacy of cancer treatment is heterogeneity of the TME, which is coordinated by both tumour-intrinsic regulations and reprogramming in response to external stimuli, including anticancer drugs. The functional heterogeneity of tumours can be assessed by longitudinal profiling of samples obtained before and after treatment as well as by performing comparisons of patients with and without a response to treatment (responders and non-responders, respectively). Both approaches can reveal potentially targetable pathways associated with sensitivity or resistance to treatment as well as predictive biomarkers to guide clinical management. For example, spatial analysis of samples from patients with HER2+ breast cancer before and after receiving neoadjuvant HER2-targeted therapy revealed that treatment leads to decreased activation of HER2-dependent signalling and increased immune infiltration, and that these changes are more pronounced in tumours from patients with a pathological complete response (pCR)97. Moreover, on-treatment gain of expression of the pan-leukocyte marker CD45 predicted pCR with high accuracy97. Transcriptional analysis of samples from patients with castration-resistant prostate cancer before and after androgen-deprivation therapy identified the presence of treatment-resistant subclones in pretreatment samples98. Stromal cells adjacent to these resistant cell clusters lacked nuclear androgen receptor expression observed in normal stroma. Further analyses of these adjacent regions revealed upregulation of pathways such as TGFβ, which is associated with stromal cell activation and tumour progression98. Spatial mapping of samples from patients with PDAC before and after receiving neoadjuvant treatment using digital spatial profiling identified distinct multicellular communities91. Post-treatment tumours were enriched in programmes associated with neural-like progenitors, neuroendocrine-like malignant cells, neurotropic CAFs and CD8+ T cells91. Further validation of these findings could reveal important biomarkers linked to cell phenotypes associated with the development of treatment resistance.

With the widespread use of ICIs across multiple cancer types, comparative serial transcriptional profiling has shed important light on T cell activation and expansion in response to treatment with these agents99,100 as well as on specific subsets of immune and/or stromal cells associated with response or resistance100,101,102. However, data from these studies lacked crucial information on cell–cell interactions within the context of the original tissue. In samples from patients with melanoma who had received ICIs, spatial profiling showed that cytotoxic T cells evolve towards an exhausted phenotype as they approach and infiltrate the tumour, and particularly areas enriched in PD-L1+ TAMs103. Mapping of the spatial distributions of cytotoxic T cells and PD-L1+ macrophages predicted response to ICIs103. In the context of TNBC, researchers used imaging mass cytometry to profile 43 proteins in samples obtained from 243 patients before and after neoadjuvant treatment with atezolizumab104. Their results demonstrate that cell composition, activation state and spatial localization are closely linked and influence response to ICIs, differing between sensitive and resistant tumours104. In this study, proliferating CD8+ TCF1+ T cells and MHCII-expressing cancer cells were the main predictors of response, followed by interactions with B cells and granzyme B+ T cells104. In a cohort of patients with TNBC who had received neoadjuvant pembrolizumab and radiotherapy, spatial analyses using CODEX identified 12 distinct spatial neighbourhoods that defined immune–non-immune cell relationships105. Enrichment in the B cell-dominant neighbourhood at baseline and in four other immune cell-rich neighbourhoods after treatment were predictive of response105. Spatial analysis of samples from patients with ovarian cancer who had received an ICI plus a poly(ADP-ribose) polymerase (PARP) inhibitor identified interactions of exhausted CD8+ T cells with either PD-L1+ TAMs or PD-L1+ cancer cells as potential biomarkers of response to treatment106.

Another approach to gain insights into the response to ICIs involves profiling of tumour subtypes with known differences in response. In CRC, a comparison of mismatch repair (MMR)-deficient and MMR-proficient tumours revealed a difference in their T cell compartments, with MMR-deficient tumours — which tend to be more sensitive to ICIs — being enriched in CXCL13+ T cells92. These CXCL13+ T cells were mainly localized outside lymphoid aggregates but in close proximity to cancer cells, consistent with their effector role92. In some tumour types in which the currently available ICIs have not demonstrated clinical benefit, spatial profiling has revealed novel potentially actionable targets. For example, a study of castrate-resistant prostate cancer revealed that metastatic tumours tend to be depleted of immune cell infiltrates but harbour high expression of the immune checkpoints B7-H3 and TIM3 (ref. 107).

Prediction of clinical outcomes

Integration of single-cell spatial data with clinical outcomes from large trial cohorts enables the exploration of spatially defined TME components as predictors of treatment response, disease progression and survival. Given the cohort sizes required for adequate power to predict survival end points, most current studies are using a combination of scRNA-seq with multiplexed proteomic methodologies, such as multiplexed IHC or immunofluorescence. Multiple studies have demonstrated that incorporating spatial information improves outcome prediction relative to integrating data on cell phenotype proportions and/or clinical classifications108,109,110, highlighting the clinical relevance of spatial organization within the TME. An analysis of specimens from patients with TNBC using multiplexed ion beam imaging showed that neither immune composition nor single-cell expression of functional proteins are associated with recurrence or survival, whereas immunoregulatory and functional protein interactions predict patient outcomes111. Spatial features in the TME can predict prognosis and/or response to treatment in multiple cancer types. For example, co-localization of TAM subsets correlates with progression-free survival in patients with renal cell carcinoma112 and co-localization of Treg cells and dysfunctional T cells predicts a poor outcome in patients with oestrogen receptor-positive breast cancer113. The distance between immune subsets and cancer cells can also be predictive. For example, closer proximity of antigen-experienced cytotoxic T cells to melanoma cells intratumourally is correlated with an improved response to ICIs80, and high infiltration of cancer-adjacent cytotoxic T cells and proximity of M2-polarized TAMs to cancer cells are correlated with favourable survival outcomes in patients with PDAC114,115. In TNBC, a tumour compartmentalization pattern in which immune cells are spatially separated from cancer cells is associated with improved survival, independent of pathology scoring of tumour-infiltrating lymphocytes87. Taking this concept a step further, spatial parameters derived from multiplexed imaging were incorporated into a machine-learning model to train predictions of progression from in situ to invasive cancer116. This study demonstrated that coordinated changes in the TME enable tracking of the invasive transition of in situ disease and, thus, could be used to predict invasive disease relapse after initial diagnosis116.

Tertiary lymphoid structures

TLS are specialized microanatomical structures found in inflamed tissues resembling secondary lymphoid organs and are particularly amenable to spatial studies owing to their complex composition and structured organization. They consist of a central zone enriched with B cells surrounded by T cells, DCs and macrophages117. TLS have been detected in various cancer types, and are believed to promote adaptive immune responses and facilitate lymphocyte recruitment. Extensive evidence supports their potential as prognostic and predictive biomarkers for immunotherapy response118,119,120,121, although their precise cellular composition, distribution within tumours, and presence across different cancer types and patients across the same cancer type are considerably heterogeneous122.

TLS have established predictive and prognostic values, with multiple studies demonstrating their correlation with response to ICIs118,119,121 and survival119,120,121,123,124,125,126. In a study of specimens from patients with various solid tumour types who had received ICIs121, the presence of mature TLS had predictive and prognostic relevance independent of PD-L1 status, CD8+ T cell infiltration and tumour type. In another study, involving samples from patients with urothelial carcinoma who had received a dual ICI combination127, TLS correlated with response to therapy whereas PD-L1 status and tumour mutational burden did not127. These studies suggest that TLS have potential to outperform the biomarkers of response to ICIs currently used in the clinic.

To gain insights into the involvement of TLS in antitumour immunity, researchers have turned to single-cell spatial methodologies, which enable a more detailed characterization of the components of TLS and their precise functions in the context of the TME (Fig. 3). For example, the application of spatial profiling has shed light on the distinct characteristics of TLS in metastatic melanoma119. Melanomas with a high abundance of B cells within TLS had a concomitant influx of TCF7+ naive and/or memory-like T cells. Conversely, tumours lacking TLS tended to have T cells with an exhausted-like dysfunctional phenotype, characterized by increased expression of TIM3, PD-1 and granzyme B119. In renal cell carcinoma, TLS+ tumours had a greater abundance of IgG-producing plasma cells around apoptotic tumour cells coated with IgG antibodies, suggestive of antitumour effector activity128. In another study, researchers used multiplexed imaging to reconstruct a 3D atlas of CRC, demonstrating that 2D TLS domains are interconnected within larger 3D networks128. Such networks were present along invasive fronts, inside tumours and in layers of the muscularis and subserosa tissues129. The precise function and organization of TLS remain to be better characterized. Such insights could refine their use as biomarkers and lead to therapeutic strategies targeting TLS, such as those altering their function and abundance122.

Fig. 3: Characterization of tertiary lymphoid structures using spatially resolved approaches.
figure 3

a, Image of a haematoxylin and eosin (H&E)-stained section of a tertiary lymphoid structure (TLS) from a breast tumour (A.S., unpublished data). b, Overlay of the H&E-stained section with spatially resolved data (obtained using a Xenium platform) for improved molecular characterization of the TLS. c, Magnified area of image shown in part b (red box). Coloured spots identify detected transcripts encoding CD3 (CD3E; orange), CD20 (MS4A1; green) and CD23 (FCER2; red) for detection of T cells, B cells and germinal centre B cells, respectively.

Full size image

Other TME niches with prognostic relevance

An analysis of samples from patients with CRC who had received ICIs130 identified nine distinct cellular neighbourhoods that were defined by a collection of components characteristic of the CRC immune TME. Enrichment in PD-1+CD4+ T cells within a granulocyte cellular neighbourhood was associated with improved survival, although the overall frequency of these cells did not predict survival130. Similar findings in several other studies have been reported, suggesting that spatially defined multicellular subgroups can be used to predict patient outcomes110,113,131,132,133. Furthermore, the classification of tumours into ecosystems consisting of cancer cells and their associated cells using single-cell and spatially resolved methodologies has shown correlation with clinical outcomes62,134, suggesting that ecosystem-based cancer classification might be a promising approach for developing precision medicine approaches targeting the TME.

Opportunities for clinical application

Over the past few years, advances in spatial profiling technologies have shown potential to capture the complexity of the cellular architecture of tumours, providing a more complete understanding of their biology. This information has the potential to improve the precision of cancer diagnosis and treatment (Fig. 4). Intratumoural heterogeneity across cancer and tumour-associated cells within the TME underlies many of the challenges we face in the clinical setting, such as disease relapse, treatment resistance and toxicity135. By delineating spatially resolved gene expression patterns associated with specific tumour subtypes, molecular signatures or patterns of response to treatment, spatial omics are ideal tools to dissect the multilayered complexity of cancers. Integrating spatially resolved data with clinical outcomes could guide the discovery of novel biomarkers for prognosis, prediction of response to therapy and patient stratification, subsequently informing personalized treatment decision-making, and monitoring disease progression and response to therapy. These data can also be used to determine spatially resolved genomic or molecular changes before and after treatment.

Fig. 4: Overview of applications of spatial profiling in cancer research.
figure 4

Tumours are dissected for detailed studies of their cellular composition and spatial profiling to yield high-dimensional information on cellular neighbourhoods and heterogeneity. Tumour sampling before and after treatment, particularly if the patient can be stratified based on likelihood of response to treatment, can enable a detailed understanding of disease subsets, prognosis and mechanisms underlying treatment response and resistance. CAF, cancer-associated fibroblast; TAM, tumour-associated macrophage; TLS, tertiary lymphoid structure.

Full size image

Immuno-oncology is an area in which single-cell and spatial profiling have substantial clinical potential. ICIs have transformed the treatment paradigm for several solid tumour types, such as melanoma and non-small-cell lung cancer, although currently only a subset of patients with these cancers have a durable response, and clinical benefit remains limited for those with many other cancer types. Resistance mechanisms are poorly understood and the biomarkers currently used for patient stratification, such as IHC-assessed PD-L1, have demonstrated low sensitivity and specificity136. Using spatial profiling to interrogate intratumoural and TME heterogeneity, researchers could identify more effective predictive biomarkers and resistance mechanisms potentially associated with novel therapeutic targets. Spatial mapping of the TME across a range of tumour types has already provided valuable insights into the heterogeneity and diversity of antitumour immune responses100,104. The next step will be to transform these useful insights into clinically relevant applications. For example, TLS are both predictive of response to ICIs and prognostic of improved outcomes. Hence, therapeutic induction of TLS could be a way to boost responses to ICIs, although experimental induction of TLS using either CD40 agonists or the chemokines CXCL13 and CCL21 in preclinical models have thus far yielded mixed results137,138, suggesting that more knowledge is required to enable targeting of TLS in a therapeutically meaningful way.

Spatial profiling technologies might also pave the way for precision pathology. Using machine learning or deep learning algorithms, incorporating spatial gene expression profiles with pathology slides is now feasible139,140. Studies have also shown that multiplex spatial profiling in the clinical setting is feasible, with an increasing number of studies incorporating these techniques into large clinical cohorts104. With the discovery of new cell markers and antibodies, and the automation of computational analyses, traditional pathology methods can yield outputs of higher resolution, improving the accuracy of tumour classification. Advances in such techniques might also enable prediction of transcriptomic profiles using haematoxylin and eosin-stained slides to facilitate a scalable and affordable implementation of these concepts in clinical settings, and thus refining the biomarkers used for diagnosis and prognostication.

Challenges and future directions

Clinical translation

Despite the promising potential of novel spatial techniques, many challenges remain to be addressed before these tools can be incorporated into routine clinical practice. Most of the studies performed thus far using spatial profiling techniques, particularly spatial transcriptomics, are exploratory in nature and used a small number of samples and/or patients, with potential biases in tissue selection. Much larger studies are required to obtain clinically meaningful results that overcome the challenges of intratumoural and intertumoural heterogeneity. To effectively leverage spatial technologies to improve our understanding of treatment response and resistance, and to aid the development of novel therapies, clinical trials will have to be designed carefully to incorporate spatial profiling in a feasible way. Cautious consideration should be given to the type and format of samples to be collected, optimal time points for collection (before, on and/or after treatment) and the most suitable technique to be used that balances research questions and practical considerations. For example, a discovery-based research question looking at tumour characterization might favour the use of techniques such as SRT, which enable untargeted spatial analyses to uncover unknown interactions. By contrast, a targeted research question, such as validating a novel biomarker, might be better addressed using multiplexed imaging methodologies with a defined protein panel. Integrating different techniques to overcome their shortcomings might be another option to increase clinical applicability, such as mining high-dimensional SRT data to define lower-dimensional panels of proteins or genes for clinical use.

Technical challenges and the prohibitive costs involved are current major barriers to expanding the size (and hence power) of clinical studies involving spatial analysis. Researchers have, however, made some progress in these areas over the past few years. Up until 2021, transcriptomic studies required the collection and isolation of fresh, live cells for sequencing, thus posing substantial logistical challenges that precluded their routine use in clinical settings. One of the major developments in spatial transcriptomics in the past few years is the feasibility of using FFPE tissue specimens, making profiling studies more compatible with current clinical practices and workflows. This advance expands the potential for applying spatial transcriptomics to large prospective and retrospective clinical cohorts, albeit with concerns including poorer data quality with the use of FFPE tissue specimens rather than fresh frozen samples owing to increased RNA degradation as a result of sample preparation and storage time141. This challenge could be overcome with better processing techniques and quality control142. Cost has been another major prohibitive factor in the wider clinical application of single-cell and spatial technologies, although the cost of sequencing a human genome has decreased from about US $1 million in the early 2000s to less than $1,000 now, and could soon approach $100 per sample143. Similarly, a substantial reduction in the cost of next-generation sequencing has facilitated its incorporation into routine clinical practice over the past decade. Advances in technology, methods and other fields, such as artificial intelligence (AI), will further reduce costs over time, and thus facilitate the incorporation of spatial technologies into the clinic. For example, mining high-dimensional single-cell transcriptional data to define lower-dimensional panels of proteins or genes for clinical use might be another way to overcome the financial barriers of translating these into the clinic.

Technical aspects

The representation of spatially resolved data as a 2D snapshot is an important consideration when using such data. Spatial methods measure the abundance of molecules (such as transcripts or proteins) on thin tissue slices taken from a larger tissue sample and these measurements characterize the molecular anatomy of the tissue in 2D. However, tissue anatomy is inherently 3D, and valuable information can be lost when the total tissue volume is not considered144. Reconstructing a 3D volume using data from consecutive spatial experiments is challenging for the same reasons that make aligning consecutive histology sections challenging (for example, tissue folds, cracks or missing information), although deep learning might provide support to facilitate and scale up 3D reconstructions145. Furthermore, characterizing a 1-cm3 tissue cube using 5-μm thick tissue sections would require ~2,000 spatial experiments, which would be prohibitively costly and time-consuming. The resulting data would probably have strong batch effects and other artefacts that would make the reconstruction even more challenging.

Ongoing work using AI is showing promise to address these challenges by providing the next generation of improved data integration and imputation methods, ultimately enabling the reconstruction of full 3D volumes of molecular tissue anatomy in a cost-efficient manner (Fig. 5). Studies using AI have already shown the possibility of imputing molecular data from histology images alone146,147,148, although to fully capitalize on AI for 3D reconstruction, other omics information from the same tissue section should be considered. Methodological advances have already demonstrated that data on histology, gene expression, chromatin accessibility and protein expression can be combined in different ways48,149,150. These combinations can also include spatial mass spectrometry imaging, overcoming the anticipated destructive exposure to laser radiation used in conventional mass spectrometry151. Multimodal generative AI models can, therefore, be used for robust normalization and batch correction, spatial imputation and cross-modality data integration. Together, these approaches could enable spatial modelling of structures in 3D, at least at the millimetre scale, to provide essential insights on tumour biology.

Fig. 5: Generation of 3D view of biopsy samples using deep learning.
figure 5

Producing a stack of images from haematoxylin and eosin-stained tissue is a straightforward approach for establishing a 3D view of tumour morphology. A fraction of these tissue sections can subsequently be processed for molecular analysis using spatially resolved technologies. The obtained spatial data on gene expression and morphology are then used to train a deep learning (DL) model that facilitates downstream imputation of gene expression in all of the stained tissue sections in the stack. Using this approach, a 3D image of histology and gene expression can be obtained in an affordable way.

Full size image

Extensive bulk genome sequencing has revealed numerous recurrent genomic copy number variant (CNV) gains and losses, along with frequent mutations60. CNVs and single nucleotide variants (SNVs) can be enriched in specific molecular subtypes, suggesting that each genomic event might drive cells into distinct functional states. The genomic landscape of several cancer types (including breast and prostate cancers) is dominated by alterations in CNVs over those in SNVs. A preclinical study has provided evidence that breast cancers undergo continuous branching evolution via ongoing copy number and mutational events, generating multiple genomic clones per tumour61. Nevertheless, the effect of subclonal CNVs and their spatial context on functional heterogeneity remains under-explored, and how many and by what mechanisms subclonal genomic events affect the phenotype of cancer cells remains poorly understood.

Spatial genomic methods investigating DNA are being developed to circumvent the aforementioned issues. The current approaches have primarily been designed to be restricted to shorter stretches of the genome, such as targeted assays using laser-capture microdissection principles152,153 or open chromatin patterns149,150, owing to cost. Newer methods are emerging to spatially resolve the genomic features of cancer cells, revealing, for example, the presence of spatially segregated genomic clones within tumours and premalignant lesions, such as breast ductal carcinoma in situ154,155. These findings suggest that genomic clonal growth patterns might in part explain observations describing spatially segregated gene expression programmes in tumours61,62. Other studies from the past few years have shown that untargeted sequenced-based SRT can also capture cancer clone heterogeneity156 through CNV inference using methods developed for scRNA-seq analysis157. Thousands of measurements across a tissue section can be analysed using a barcoded surface to predict genomic gains and losses. Importantly, this approach has provided a data-driven phylogenetic tree of tumour evolution and led to the description of benign areas of the prostate that harbour the same genomic mutations as the tumour156. Thus, the possibility of an unbiased view over large tissue areas enables the identification of early genetic events not evident when assessing morphology, opening opportunities for new approaches to diagnostics and monitoring, and will be of particular interest for identifying early disease biomarkers.

Furthermore, how genomic variants contribute to predominant patterns of interaction of cancer cells with stromal and immune cells remains poorly understood. Experimental methods described in 2023 (ref. 158) offer a first glimpse into the immune repertoire in the context of tumour subclones with distinct spatial patterns of B cell and T cell clonotypes in the breast cancer microenvironment, suggesting that this could become a relevant area to elucidate binding specificity between immune cells and cancer cells. Mutations in stromal cells might also have a role in defining the tumour ecosystem, which is another aspect that future research should address159.

Conclusions

The field of spatial profiling has undergone transformative growth, from being used to simultaneously characterize a few transcripts or proteins to seamlessly providing spatially resolved maps of whole organs. Although spatially resolved profiling approaches remain in the early stages of development, they are revolutionizing how we examine cellular interactions and the architectural relationships between cancer cells and the TME. These interactions have a pivotal role in defining tumour biology and, consequently, patient outcomes. The clinical needs of the cancer community will set the stage for the best use of the available palette of multimodal spatial and computational tools to provide new ways to diagnose and treat cancer.