Spatial omics illuminates tumor heterogeneity

Summary

Intratumoral heterogeneity (ITH) is the coexistence of diverse cancer cell states, genotypes, and microenvironmental niches within a single tumor and is a major driver of therapeutic resistance and disease progression. While the clinical implications of ITH are well appreciated, conventional methods and models have been limited in resolving the spatial and functional complexity underlying ITH. Recent advances in single-cell and spatial omics, high-plex imaging, and in vivo CRISPR perturbation and lineage tracing now enable unprecedented, mechanistic dissection of tumor ecosystems. These technologies are beginning to reveal how cell-cell interactions, spatial organization, and clonal evolution collectively shape tumor behavior and treatment response. We review emerging methods for spatial proteomics and transcriptomics, functional genomics, and clonal tracing and highlight how their integration is redefining the study of ITH. We also discuss current challenges, including scalability, accessibility, and multimodal data integration, and opportunities as cancer biology enters the spatial era.

Emerging spatial omics, in vivo CRISPR, and lineage-tracing technologies are transforming the study of intratumoral heterogeneity. Srinivas et al. review how integrating spatial proteomics, transcriptomics, functional genomics, and clonal tracing is uncovering the cellular interactions in heterogeneous tumor ecosystems that drive tumor progression, evolution, and treatment response.

Introduction

What differentiates cancer from other diseases and makes it such a monumental challenge to biologically understand and clinically treat can be explained by tumor heterogeneity. A tumor is the result of a lengthy evolutionary process that involves accumulating mutations and the selection of the fittest cells, often taking years, if not decades, before diagnosis. In most cases, the tumor starts from a single mutant cell that escapes many checkpoints in place, such as those in the cell cycle and immune surveillance. But, the resulting tumor is not simply a mass of many copies of the initial mutant cell, but rather many derivatives that share certain genetic and histological features but can differ significantly otherwise.^1,2,3 In addition, a tumor is a cellular ecosystem, not just composed of cancer cells, but involving many types of immune and stromal cells. This cellular ecosystem is dynamic and plastic, where the cancer cell plays a tug-of-war to overcome homeostatic constraints, usually derailing non-cancer components of the tumor from their normal behavior along the way.^4,5 The dynamic nature of tumor ecosystems often limits the long-term efficacy of targeted treatments, frequently resulting in transient responses rather than lasting remissions. To improve this clinical outlook and extend curative therapies to many more cancer patients, it is essential to understand tumor heterogeneity better.

Classical pathology revealed the first evidence of tumor heterogeneity, as different histological features are often found within the same tumor sample.⁶ DNA sequencing provided the first molecular insights into intratumoral heterogeneity (ITH), where tracking mutations across tumor regions and metastases revealed that clonal tumorigenesis underlies tumoral heterogeneity.⁷ Single-cell genomics technologies, especially single-cell transcriptomics, enabled an unprecedented level of granularity in studying ITH and the tumor microenvironment (TME), leading to the identification of new cell subtypes and states.⁸ However, sequencing methods, both bulk and single-cell, isolate molecules and cells from their tissue context as they require tissue dissociation and cell lysis, which leaves out essential information regarding cellular location in tissue or tumor. Cellular behavior is fundamentally controlled by spatial biology, where the organization of cells and the distribution of molecules dictate almost everything from gene expression to cell signaling to metabolism, and to immune reactions.⁹ So, it should come as no surprise that the emergence of spatial biology methods is revolutionizing the many domains of biological research, especially cancer research. Here, we will introduce the leading spatial biology methods and highlight exemplary studies to showcase how spatial biology is expanding our understanding of tumor heterogeneity at a head-spinning pace.

Spatial proteomics

Spatial proteomics (SP) approaches rely on the detection of protein markers with antibodies, much like classic immunohistochemistry (IHC) and immunofluorescence (IF) staining. What differentiates SP methods from IHC and IF is their multiplexing capabilities. While IHC and IF are limited to 2–3 markers per sample, it is possible to stain around 50–100 markers on a single tissue section with leading SP methods. SP technologies can be methodologically categorized into three main groups: multiplexed IHC (mIHC), multiplexed IF, and mass spectrometry (MS) based (Figure 1).

Figure 1.

SP technologies

mIHC relies on repeated cycles of staining, imaging, and signal removal, typically detecting one marker per cycle. mIF enables the detection of multiple markers per cycle through iterative staining and imaging using fluorophore-conjugated antibodies. In contrast, MS-based imaging approaches use metal-conjugated antibodies and allow highly multiplexed detection of many markers in a single acquisition; however, some implementations involve tissue ablation or destruction during the measurement process.

Multiplexed immunohistochemical consecutive staining on a single slide (MICSSS) is an exemplary mIHC method, where the tissue sample is iteratively stained with one antibody per cycle, and the chromogenic substrate is chemically removed after imaging.¹⁰ While its multiplexing capacity is 10–20 markers, MICSSS does not require expensive setup costs and is particularly easy to adopt for labs with IHC experience.

Since IF microscopy utilizes multiple channels enabling co-detection of multiple (usually 2–3) markers at a time, multiplex immunofluorescence (mIF)-based SP methods are advantageous in their multiplexing capabilities and experimental speed. There are several methods in this category, which differ in their approach for antibody signal detection and signal removal between cycles: cyclic immunofluorescence (CyCIF),¹¹ multi-epitope ligand cartography (MELC),¹² iterative indirect immunofluorescence imaging (4i),¹³ and iterative bleaching extends multiplexity (IBEX).¹⁴ There are commercially available mIF platforms that automate iterative staining and imaging, such as Lunaphore COMET,¹⁵ Miltenyi MACSima,¹⁶ Rarecyte Orion,¹⁷ and Leica Cell DIVE.¹⁸ mIF methods and platforms detect fluorophores conjugated to either primary or secondary antibodies. Co-detection by indexing (CODEX) uses fluorophore-conjugated probe hybridization to antibody-conjugated oligonucleotides, which allows staining with all antibodies in a single reaction.¹⁹ This method is commercially available as Quanterix PhenoCycler.

MS-based methods include imaging mass cytometry (IMC, commercially available as Standard BioTools Hyperion)²⁰ and multiplexed ion beam imaging (MIBI),²¹ where tissue is stained with a panel of metal-conjugated antibodies, followed by laser ablation or ion beam sputtering of tissue in a scanner-like approach, and detection of the ions by time-of-flight mass cytometer (CyTOF) or secondary ion mass spectrometry (TOF-SIMS).

While a given lab’s prior experience with adjacent methods and instrument accessibility are likely to determine which of these SP methods a lab prefers to use in real life, it is important to consider the strengths and limitations of each approach. mIHC has sub-cellular resolution (i.e., membrane, cytoplasm, and nucleus), while MS-based methods have 0.5–1 μm resolution, and mIF has submicron resolution, limited by light diffraction. MS-based methods offer up to 40–50 antibodies per panel (limited by the number of lanthanide metals available for conjugation), more than doubling the multiplexing of mIHC. In current practice, mIF matches the multiplexing capacity of MS and, in theory, can expand to several hundred markers as long as tissue integrity is preserved. Relatively cheaper instrumentation (i.e., microscope/slide scanner versus mass cytometer/spectrometer), a wide selection of ready-to-use antibodies (fluorophore-versus ion-conjugated), and faster imaging time (as in area per hour) favor mIF over MS-based methods.

Spatial transcriptomics

Antibody detection-based approaches used in SP are inherently amenable to spatial biology, as staining is conducted on intact two-dimensional tissue and cell location data can easily be extracted from microscope images for mIHC and mIF-based methods. For MS-based methods, there is not a tissue image per se, but the cell/molecule location information is readily available as the laser/ion beam scans tissue, generating a pseudo-image area by area. Unlike proteins, bulk and single-cell RNA detection methods rely on sequencing, which necessitates RNA isolation, hence tissue destruction. While multiplexing was the main challenge in moving from traditional tissue staining methods to emerging SP methods, preservation of spatial data itself has been the challenge to devise spatial transcriptomics (ST) methods.

Two traditional methods led the way to the invention of ST. Dissection approaches copy from the playbook of traditional laser microdissection, where a selected piece of tissue region, or region of interest (ROI), as commonly referred to, is isolated for sequencing (Figure 2). For example, Nanostring GeoMx, which is currently the most widely adopted method in this sector, uses a photocleavable linker.²² RNA molecules are bound with barcoded in situ hybridization (ISH) probes, while selected proteins are recognized by barcoded antibodies. Upon illumination of the ROI with ultraviolet light, the linker is cleaved to release barcode oligonucleotides, which are then sequenced. In a similar fashion, light-controlled cross-linking (Light-seq)²³ or hybridization (ZipSeq)²⁴ of area barcodes can be used to incorporate spatial location data into sequencing. Transcriptome in vivo analysis (TIVA)²⁵ tag and photo isolation chemistry (PIC)²⁶ methods use photosensitive molecules to control reverse transcription by light activation to geotag mRNA molecules. While light-controlled methods provide spatially tagged whole transcriptome data, either as bulk or single-cell, the spatial data corresponds to an ROI. Hence, these methods are not designed to provide truly single-cell spatial data.

Figure 2.

ST technologies

Microdissection-based approaches isolate spatially defined tissue regions for downstream bulk RNA sequencing. FISH and ISS use probes to detect predefined mRNA targets at subcellular resolution; FISH relies on fluorescent labeling for detection, whereas ISS determines sequences directly in situ from probes or cDNA derived from transcripts. Spatial barcoding methods use arrays of capture probes with positional barcodes that are incorporated during reverse transcription, allowing transcripts to be mapped back to their original spatial coordinates following sequencing.

The second traditional method upon which novel ST methods have been developed is fluorescence in situ hybridization (FISH) (Figure 2). Sequential FISH (seqFISH) employs a combination of fluorescent probes, where in each cycle, four or five probes, each detectable by a different fluorescent channel, are hybridized to mRNA molecules, imaged, and stripped.²⁷ Debarcoding of the composite image identifies each mRNA based on fluorescent color code. seqFISH+²⁸ and multiplexed error-robust fluorescence in situ hybridization (MERFISH)²⁹ use intermediatory probes known as saddle probes, each of which is further hybridized sequentially by a combination of fluorescent readout probes. Use of saddle probes extends the barcode capacity, which improves mRNA identification specificity. MERFISH differs from seqFISH+ in its barcoding strategy, which utilizes a binary code rather than the color combination code. Commercial platforms descending from these methods include Vizgen MERSCOPE and Nanostring CosMX.³⁰ While multiplexed FISH methods provide spatial transcriptome readout at the subcellular level (resolution limited to light diffraction) with extremely high sensitivity (single RNA per cell), their multiplexing capacity is currently limited to a few thousand genes due to optical crowding. Because these platforms target a predefined subset of transcripts rather than the whole transcriptome, experiments require prior selection of gene panels, which may introduce bias or constrain hypothesis generation. To overcome this challenge, Cheng et al. developed reverse-padlock amplicon-encoding fluorescence in situ hybridization (RAEFISH).³¹ This method uses a pair of probes (reverse padlock and priming probes) that recognize a specific region of target mRNA and amplify the sequence by rolling circle amplification (RCA) reaction. The many copies of the target are then recognized by fluorescent dye-labeled probes with cycles of binding, imaging, and elution on the MERFISH platform. Imaging a single probe target sequence in RAEFISH, rather than an array of sequences as in MERFISH, appears to mitigate the signal crowding problem, as the authors demonstrated whole transcriptome sequencing of both human and mouse samples with this method. The RCA-amplified single-target read on RNA also enables RAEFISH to detect shorter RNA species, such as long non-coding RNA, more efficiently than MERFISH can. However, the concordance of mRNA reads of tested samples was higher between MERFISH and bulk RNA-seq compared to the concordance between RAEFISH and bulk RNA-seq, suggesting a trade-off of sensitivity to achieve whole transcriptome coverage. It should be noted that, as we write this review, imaging-based ST platforms are expanding RNA coverage to the whole-transcriptome scale either by major expansions on existing platforms like CosMx or via the introduction of new systems like Atera.

In an effort to make FISH-based spatial RNA sequencing more affordable, the profiling of RNA in situ through single-round imaging (PRISM) devises a unique approach to detect multiple unique barcodes per imaging channel in a single round.³² Each padlock probe contains 4 bits, each for a fluorescence channel (e.g., Texas Red, Alexa Fluor 488, Cy3, Cy5). Imaging probes per channel are mixed in a defined ratio so that multiple barcodes per bit (i.e., fluorescence channel) can be distinguished based on the fluorescence intensity. A mixture of pooled probes for all channels is hybridized to padlock probes bound to their target mRNAs and amplified by RCA in a single step. Chang et al. demonstrated PRISM on human hepatocellular carcinoma tissue samples using a 31-gene panel, uncovering heterogeneities in the TMEs, and demonstrating cancer-associated fibroblasts restricting immune cell infiltration by establishing physical barriers. While PRISM does not require special equipment other than a conventional fluorescence microscope and does not need image registration, it is currently limited to the identification of a small selection of RNA targets. Other multiplexed FISH ST methods have been developed, mostly to increase the signal-to-noise ratio, with much limited multiplexing capacity, such as Saber-FISH,³³ Clamp-FISH,^34,35 EEL-FISH,³⁶ and SCRINSHOT.³⁷

The third category of ST methods is based on in situ sequencing (ISS) (Figure 2). In targeted approaches, complementary oligonucleotide probes bind to target mRNAs. Methods like STARmap³⁸ employ a set of probes (e.g., snail and barcode probes), while others like Barista Seq,³⁹ targeted-ExSeq,⁴⁰ and Xenium use padlock probes. In comparison, methods such as FISSEQ⁴¹ and untargeted ExSeq⁴⁰ use poly-T primers to reverse transcribe mRNAs in an untargeted fashion (i.e., no pre-designed probes). The next step after mRNA hybridization involves rolling circle DNA amplification of circular probes (for targeted methods) or circular cDNA (for untargeted methods), generating multiple copies of the barcode/cDNA sequence to amplify signal detection in the following step, where the sequencing is performed either by sequencing-by-synthesis or sequencing-by-ligation. ISS methods, similar to FISH methods, generate truly spatial data in the form of high-resolution microscope images of intact tissue. However, ISS methods also suffer from the optical crowding issue, and hence, the number of mRNAs detected per cell is limited to a few thousand at best with the current technologies.

The fourth category of ST methods is based on spatial barcoding (Figure 2). Unlike FISH and ISS-based methods, there is no actual microscopy imaging involved. Therefore, methods in this category need to incorporate spatial data into sequencing data. In essence, all these methods use poly-A binding probes with unique molecular identifiers (UMI) and spatial barcodes, which hybridize to mRNAs in tissue, serving as primers for in situ reverse transcription to generate cDNAs. Then, cDNAs are collected and bulk sequenced. Each cDNA contains a snippet from its complementary mRNA, enabling mRNA identification, the spatial barcode, which is shared among all cDNAs collected from the same spatial location, and the UMI, which is used to correct for polymerase chain reaction (PCR) skewing to accurately quantify the mRNA. Spatial barcoding-based ST methods differ in their means of incorporating spatial probes. Spatial Transcriptomics (the method, not to be confused with the umbrella term) and Visium print a microarray of probes directly on a slide, where the exact location of each probe/spatial barcode is known a priori.⁴² Slide-seq⁴³ and high-density spatial transcriptomics (HDST)⁴⁴ employ probe-conjugated beads that are then placed in an array on the slide. Stereo-seq,^45,46 Seq-Scope,⁴⁷ and Pixel-seq⁴⁸ chemically modify a silicon wafer, an Illumina sequencing flow cell, or a hydrogel surface to incorporate probes. Unlike microarray printing, beads and chemical modifications randomly distribute probes on the surface, necessitating imaging of the spatial barcodes before tissue is placed and sequenced so that mRNA reads can be assigned to their spatial location. Deterministic barcoding in tissue (DBIT-seq) uses an alternative approach, where probes are ligated onto tissue that is already on a slide.⁴⁹ A microfluidic chip applies a predefined set of spatial probes to the tissue twice, in perpendicular directions, creating a molecular grid. Spatial barcoding-based ST methods provide whole transcriptome coverage and plug in to existing bulk sequencing pipelines, hence not necessitating expensive dedicated equipment. While newer methods and commercial solutions (e.g., Visium HD) continue to increase resolution, they cannot truly capture single-cell data because cell distribution on spatial grids/spots is probabilistic. Which means one spatial location (as marked by shared spatial barcodes) can house more than one cell.

Spatial functional genomics

Spatial biology methods are becoming more widely available thanks to commercially available platforms (Table 1) and have generated previously unattainable insights into cancer biology. The methods and example studies discussed so far identify the genetic bases of observed phenotypes in tumors, hence useful for forward genetics (Table 2). Emerging technologies combine the aforementioned spatial methods with arrayed or pooled CRISPR screens to enable direct testing of the functional roles of individual genes/mutations, therefore, allowing for reverse genetics. In essence, spatial functional genomics methods use means of identifying individual CRISPR mutations alongside measuring phenotypical readouts using a spatial biology method. This allows studying the functional consequences of targeted CRISPR mutations in a screen in situ (Figure 3).

Table 1.

Commercial spatial biology platforms

Platform	Class	Method	Resolution	Plexity	Capture area/throughput	Tissue compatibility	Relative cost	Best use cases
COMET	protein	cyclic mIF	subcellular	∼40 proteins	∼9 × 9 mm (per chip)	FFPE, fresh frozen, fixed frozen	medium	high-plex protein mapping with moderate area
MACSima	protein	cyclic mIF	subcellular	100+ proteins	up to ∼47 × 17 mm (max scan area)	FFPE, fixed frozen	medium	ultra-high plex protein profiling over large areas
Orion	protein	one-shot mIF	subcellular	∼20 proteins (simultaneous)	whole slide	FFPE, fresh frozen	medium	fast whole-slide protein imaging (lower plex)
Cell DIVE	protein	cyclic mIF	subcellular	60+ proteins	up to ∼4 × 2 cm	FFPE	medium	deep protein profiling in defined regions
PhenoCycler (CODEX)	protein	cyclic mIF	subcellular	100+ proteins	whole slide (tile-based)	FFPE, fresh frozen	medium	high-plex SP
Hyperion (IMC)	protein	imaging mass cytometry	subcellular	∼40 proteins	ROI-based (low throughput)	FFPE, fresh frozen	high	high-dimensional protein analysis with metal tags
GeoMx DSP	ROI-based	barcoded probe + ROI selection	multicellular (10–600 μm ROIs)	whole transcriptome; 570+ proteins	∼12 ROIs/slide (variable)	FFPE, fresh frozen	medium	region-specific profiling (not single-cell)
MERSCOPE Ultra	RNA	MERFISH	subcellular	up to ∼1,000 genes	∼1–3 cm2	FFPE, fresh frozen, fixed frozen	medium	high-resolution ST
CosMx SMI	RNA + protein	multiplexed smFISH	subcellular	∼1,000–6,000 genes; ∼64 proteins	∼20 × 15 mm	FFPE, fresh frozen	high	multiomic spatial profiling (RNA + protein)
Xenium	RNA	ISS	subcellular	up to ∼5,000 genes	∼10 × 22 mm	FFPE, fresh frozen	high	high-plex RNA imaging with strong sensitivity
Visium v1	sequencing-based	spatial barcoding	1–10 cells/spot (∼55 μm)	whole transcriptome	6.5 × 6.5 mm (4 capture areas/slide)	fresh frozen	medium	discovery ST
Visium HD (WT)	sequencing-based	high-density spatial barcoding	near single-cell (∼2 μm)	whole transcriptome	6.5 × 6.5 mm (2/slide) or 11 × 11 mm (2/slide)	FFPE, fresh frozen, fixed frozen	medium	high-resolution transcriptomics
Visium HD (3′)	sequencing-based	high-density spatial barcoding	near single-cell (∼2 μm)	whole transcriptome (3′)	6.5 × 6.5 mm (2/slide)	fresh frozen	medium	cost-efficient high-res transcriptomics

Table 2.

Summary of selected spatial studies of patient tumors

Methods	Cancer	Key finding	References
IMC	BRCA	spatially defined tumor subgroups based on multiplex protein profiling outperform traditional clinical subtypes in predicting patient survival	Jackson et al.50
MIBI	BRCA	in situ breast tumors with preserved myoepithelial layers and high E-cadherin expression are associated with increased risk of progression to invasive disease	Risom et al.51
Slide-seq, MERFISH, ExSeq, CODEX	BRCA	tumors exhibit patient-specific copy number alterations alongside distinct macrophage states (e.g., CD163+ vs. proliferative) and spatial EMT-associated niches linked to immune exclusion	Klughammer et al.52
Visium, Xenium	BRCA	a rare CX3CL1+/PROM1+ tumor-associated myoepithelial cell population localizes to the tumor boundary, suggesting a role in tumor-stroma interactions	Janesick et al.53
CyCIF	BRCA	PARP inhibitor treatment increases tumor-associated macrophages via SREBP-1 signaling, which can be counteracted by CSF1R blockade to restore anti-tumor immunity	Mehta et al.54
MIBI	BRCA	tumors segregate into spatial immune states (cold, mixed, compartmentalized), where PD-1/PD-L1/IDO/LAG3 expression patterns and immune organization correlate with survival	Keren et al.55
CODEX	BRCA	in TNBC treated with immunotherapy, patients stratify into immune-infiltrated (TLS-rich), delayed responders, and persistently immune-excluded groups based on spatial TIL dynamics	Shiao et al.56
IMC	BRCA	enrichment of CD8+TCF1+ T cells and MHC-II+ tumor cells predicts response to checkpoint blockade, while CD15+ tumor cells are associated with resistance	Wang et al.57
mIF	BRCA	distinct immune phenotypes (excluded [collagen-rich, TGFβ/VEGF-driven], ignored [CD163+ myeloid/WNT-driven], and inflamed [CLEC9A+ DC-rich]) predict immunotherapy response	Hammerl et al.58
Visium, CODEX	BRCA, CRC, PDAC, RCC, UCEC, CHOL	tumors are composed of spatially distinct microregions harboring genetically diverse subclones with unique metabolic and immune interactions	Mo et al.59
CyCIF, GeoMx	CRC	large-scale spatial gradients in chromatin marks (H3K27ac/me3) and signaling (EGFR, p53) are observed, with immune evasion linked to PD-1+ T cells interacting with PD-L1+ myeloid cells	Lin et al.60
CODEX	CRC	distinct cellular neighborhoods, including granulocyte-enriched regions with PD-1+CD4+ T cells, are associated with improved patient survival	Schurch et al.61
GeoMx	CRC	structured tumor-immune hubs include myeloid-rich regions at the tumor boundary, while mismatch repair-deficient tumors show intratumoral hubs enriched for activated T cells	Pelka et al.62
IBEX	FL	DC-SIGN+ dendritic cells form supportive niches that enhance BCR signaling in malignant B cells, with stromal desmoplasia linked to early relapse risk	Radtke et al.63
IMC	GBM and BrM	specific cellular neighborhoods enriched in T cells and macrophages, including MPO+ macrophages, are associated with long-term survival in brain tumors	Karimi et al.64
MICSSS, MERFISH	HCC	intratumoral niches containing mregDCs and CXCL13+ helper T cells regulate differentiation of progenitor-exhausted (PD-1+TCF1+) CD8+ T cells during immunotherapy	Magen et al.65
GeoMx, mIF	HCC	distinct tumor subclones interact with CAF-rich microenvironments to promote EMT and CD8+ T cell exhaustion via HLA-E-NKG2A checkpoint signaling	Sun et al.66
Stereo-seq	HCC and liver CHOL	CXCL6+ tumor cells activate JAK-STAT3 signaling in hepatocytes, inducing SAA1/2 expression that recruits and polarizes macrophages toward an immunosuppressive M2 state	Wu et al.67
Visium, CosMx	HGSOC	spatially distinct tumor subclones with unique CNAs engage in specific ligand-receptor interactions (e.g., FGF, WNT) and recruit different immune and stromal cell populations	Denisenko et al.68
CosMx, Xenium, MERFISH	HGSOC	a malignant cell-intrinsic gene program (MTIL) predicts T/NK cell infiltration and is associated with improved survival and increased sensitivity to immune killing	Yeh et al.69
Visium, Xenium, CODEX	LUAD	early and invasive tumors show distinct regions, including NKX2-1+ alveolar-like and CAF-rich invasive zones where IDO1 expression and CCL22+ immune cells contribute to immune suppression	Takano et al.70
IMC	LUAD	interactions among CD163+ macrophages, FOXP3+ Tregs, and B cells are prognostic, and a defined immune-stromal marker panel predicts post-surgical progression	Sorin et al.71
MICSSS	LUAD	distinct CAF subsets (FAP+ vs. MYH11+) organize collagen-rich extracellular matrix barriers that physically exclude T cells from tumor regions	Grout et al.72
Slide-seq, CODEX, MIBI, mIF	NB	neuroblastomas contain spatially segregated malignant cell states (sympathoblast-like vs. adrenergic) that are largely separated from immune and stromal compartments	Patel et al.73
Visium	OSCC	tumor core regions are enriched for keratinization markers, while the invasive edge shows ECM remodeling and EMT programs linked to disease progression.	Arora et al.74
IMC	PDAC	distinct cellular neighborhoods include a CD44-high macrophage-rich vascular niche that expresses pro-angiogenic factors such as VEGFA and fibronectin.	Sussman et al.75
Spatial Transcriptomics (method)	PDAC	integration with single-cell data maps tumor and stromal populations (e.g., inflammatory fibroblasts and stress-response cancer cells) to spatially organized niches	Moncada et al.76
GeoMx	PRAD	metastatic tumors show spatial heterogeneity in androgen receptor and neuroendocrine programs, with AR-high regions displaying low immune infiltration and high B7-H3 expression	Brady et al.77
Visium, mIF	RCC	tertiary lymphoid structures support plasma cell maturation and antibody production that targets tumor cells, correlating with improved immunotherapy outcomes	Meylan et al.78
Visium	RCC	IL1B+ macrophages accumulate at tumor-normal boundaries and are associated with nearby tumor cells undergoing EMT-like transitions	Li et al.79
Slide-seq	RCC and SKCM	distinct immune niches, including T cell zones and germinal center-like structures, reveal spatially organized immune responses and cell-cell interactions	Liu et al.80
mIF, GeoMx	SKCM	successful adoptive T cell therapy is associated with tumor-reactive CD8+ T cells engaging myeloid cells through IFN signaling and chemokine pathways (e.g., CXCL9/10)	Barras et al.81
IMC	SKCM	proliferating CD8+CD45RO+Ki67+ T cells located near SOX10+ tumor cells are predictive of improved response to checkpoint blockade	Moldoveanu et al.82
CyCIF, GeoMx	SKCM	melanoma progression involves localized immunosuppressive niches marked by IFNγ signaling, IDO1 induction, and PD-1-PD-L1 interactions primarily between T cells and myeloid cells rather than cancer cells	Nirmal et al.83
Spatial Transcriptomics (method), MIBI	SSCC	a tumor-specific keratinocyte population expressing EMT and integrin signaling genes localizes to fibrovascular niches and is associated with immunosuppression and poor immunotherapy response	Ji et al.84

Cancer-type abbreviations: BRCA, breast invasive carcinoma; CHOL, cholangiocarcinoma; CRC, colorectal cancer; FL, follicular lymphoma; GBM, glioblastoma multiforme; HCC, hepatocellular carcinoma; HGSOC, high-grade serous ovarian carcinoma; LUAD, lung adenocarcinoma; NB, neuroblastoma; OSCC, oral squamous cell carcinoma; PDAC, pancreatic ductal adenocarcinoma; PRAD, prostate adenocarcinoma; RCC, renal cell carcinoma; SKCM, skin cutaneous melanoma; SSCC, skin squamous cell carcinoma; UCEC, uterine corpus endometrial carcinoma; BrM, brain metastases.

Figure 3.

Spatial functional genomics methods

Proteomics-based approaches use genetically encoded protein barcodes linked to perturbations (e.g., sgRNAs), typically composed of combinatorial epitope tags fused to engineered protein constructs. Antibody-based detection of these epitope tags enables identification of barcodes at single-cell resolution. Transcriptomics-based methods instead readout perturbations by detecting associated oligonucleotide barcodes or sgRNA sequences, usually captured as RNA. ISH and ISS enable targeted, subcellular detection of barcodes or sgRNAs alongside selected transcripts, whereas spatial barcoding approaches provide broader transcriptome-wide profiling with barcodes but typically at lower spatial resolution.

Spatial functional genomics methods can be categorized into two groups based on whether they use SP or ST for the readout. For SP readout, protein barcodes are utilized. Pro-Codes are protein-based barcodes composed of linear epitope combinations designed to label each CRISPR mutation in a screen. Matched Pro-Code and sgRNA pairs are stably introduced into cells via lentiviral transduction. Pro-Codes are then expressed in the cells fused to a protein scaffold, which can be engineered to direct the barcode to specific cellular compartments, such as the cell surface via a mutant NGFR or the nucleus via mCherry with a nuclear localization signal. The resulting barcodes are detected through antibody staining of the epitopes in parallel with phenotypic markers, thus enabling high-dimensional protein-level phenotyping of cells in a CRISPR screen.⁸⁵ Perturb-map employs Pro-Codes to resolve in vivo arrayed CRISPR screens at single-cell resolution using tissue staining such as MICSSS and CyCIF.⁸⁶ Consecutive tissue sections can be analyzed with an ST method, such as Visium, to enrich the optical data (i.e., perturbation quantification, morphology, and immune phenotype) with spatially resolved gene expression data. In lung adenocarcinoma, Perturb-map identified that loss of TGFBR2 in cancer cells leads to stromal TGFB signaling, promoting aggressive clonal growth with immune exclusion.⁸⁶ Perturb-map led to the discovery of cancer cell-secreted IL-4 as a potent regulator of immune checkpoint therapy response and CCL7 as a master regulator of immune TME phenotype in ovarian cancer.⁸⁷ In pancreatic cancer, Perturb-map identified extracellular matrix modifiers PAI1 and PAI2 as regulators of immune evasion.⁸⁸ These studies demonstrate that cancer cell mutations shape immune TME and drive clonal immune selection in heterogeneous solid tumors.

In a similar approach to that of Pro-Codes and Perturb-map, EpicTag fuses combinatorial epitope tags with fluorescent reporters and uses MIBI to track tumor clones. EpicMIBI demonstrated clonal growth dynamics between endocrine and non-endocrine cells within the same small cell lung tumor, finding that PTEN mutant clones can promote the growth of neighboring clones with wild-type PTEN.⁸⁹ Another method that employs combinatorial epitope tags is EPICode, which expresses similar tag combinations with fluorescent reporters fused to either histone H2B or translocase of outer mitochondrial membrane 20 (TOMM20) to direct their cellular localization to the nucleus or mitochondria, respectively.⁹⁰

Several recently developed technologies spatially map in vivo genetic perturbation screens in conjunction with phenotyping using ST methods. Perturb-FISH uses a chimeric U6/T7 promoter to express sgRNAs. The U6 promoter is commonly used to drive the production of functional sgRNA to guide CRISPR-associated protein 9 (Cas9) in CRISPR experiments. T7 polymerase enables in situ transcription of the sgRNA to produce many copies of it. Consequently, MERFISH is employed to detect both sgRNAs and mRNAs simultaneously. Using this method, Binan et al. screened 35 genes in the NF-kB pathway in a xenograft model of a human melanoma to identify how these mutations affected gene expression in cancer cells as well as in neighboring T cells.⁹¹

Several CRISPR-ST methods take advantage of CROPseq design, which was originally developed to detect single guide RNA (sgRNA) sequences by single-cell RNA sequencing (scRNA-seq). CROPseq takes advantage of the duplication of long-terminal repeats (LTRs) during lentiviral genomic integration, which results in duplicated U6-sgRNA sequences, one independent and one downstream of an mRNA coding sequence. The independent U6-sgRNA sequence is transcribed to sgRNA by RNA polymerase III, while the other copy downstream of the mRNA coding sequence (e.g., EF1a promoter-driven Puromycin resistance gene) is incorporated into the transcribed mRNA, where the sgRNA sequence is located at the end of the mRNA sequence, right before the poly-A tail, enabling sequencing by poly-A capture as in 10X Chromium scRNA-seq.⁹²

Perturb-Multi integrates a 185-mer barcode downstream of the sgRNA sequence on CROPseq. Padlock probes targeting the barcode sequences, as well as endogenous mRNAs, are amplified with RCA and detected with MERFISH. Saunders et al. injected a lentiviral CRISPR library targeting 202 genes and 50 controls into LSL-Cas9-GFP (i.e., Cre recombination-induced expression of Cas9 and green fluorescent protein) mouse pups, followed by injection of AAV-Cre (i.e., Cre recombinase delivered via adeno-associated virus) into the same mice when they reached adulthood, which created mosaic perturbations in their livers. This screen identified several genes regulating hepatocyte zonation, stress response, and lipid accumulation.⁹³ While this study did not include any tumor models, Perturb-Multi can be used for spatial CRISPR screens in cancer models.

CRISPRmap is another method that uses the CROPseq plasmid with a barcode sequence adjacent to the sgRNA sequence. Each barcode is bound by a set of two detection probes (padlock and primer), which are combined by T4 ligation and amplified with RCA. The combined detection probe set contains four readout sequences and is imaged with seqFISH alongside endogenous mRNAs. This unique debarcoding approach, which distributes four readout sequences to two probes, increases specificity as hybridization of two correct probes is needed to identify each sgRNA. Gu et al. tested a CRISPR library containing 364 sgRNAs targeting 27 DNA damage repair genes in a xenotransplant model of human esophageal carcinoma, which lacked Cas9. In this model, the library was used for clonal tracing rather than perturbation. They found that tumors in these mice grew predominantly clonally despite subcutaneous injection, which was previously shown to establish mixed tumor growth in subcutaneous breast tumor transplants.⁹⁴

PerturbView also utilizes the CROPseq vector modified with a chimeric U6/T7 promoter to create multiple copies of sgRNAs, which are then sequenced with ISS. Kudo et al. introduced a non-targeting library composed of 100 sgRNAs in a xenograft model of human colon cancer and sequentially analyzed tumors with Xenium for targeted mRNA detection and ISS for sgRNA detection. They identified several cancer genes whose expression depended on tumor clonal diversity.⁹⁵ While this study did not demonstrate a CRISPR perturbation screen in vivo, PerturbView is poised to be used for such spatial functional genomics studies.

CROPseq-multi incorporates barcodes within the synthetic loop joining the CRISPR RNA (crRNA) and trans-activating CRISPR RNA (tracrRNA) into sgRNA. This placement minimizes the distance between a barcode and its matching sgRNA in order to minimize the chance of recombination-mediated mismatches between pooled plasmids during cloning. In addition, two sgRNA-barcode sequences are co-transcribed from a single U6 promoter and separated by tRNA sequences, which leads to cleavage of the transcript into two independent sgRNAs and separator tRNAs by endogenous RNase enzymes. As in the original CROPseq plasmid, sgRNA-barcodes are also transcribed as mRNA from the EF1a promoter, which enables poly-A capture for untargeted sequencing. An additional T7 promoter placed in the opposite orientation at the 3′ end of sgRNA-barcodes mediates in vitro transcription of sgRNA-barcodes, which can be detected using probe-targeted optical sequencing.⁹⁶ While CROPseq-multi is a promising, versatile tool, its use in spatial functional genomics is pending demonstration.

Perturb-CAST employs a barcode design that enables perturbation detection by the 10X Visium platform. Each barcode contains a combination of probe target sequences for three genes that are known to be not expressed in the tissue of interest. These barcodes are sequenced alongside endogenous mRNAs using Visium’s targeted sequencing method developed for formalin-fixed paraffin-embedded (FFPE) tissue, which utilizes probes rather than poly-A/T hybridization to capture RNA molecules. Barcodes are integrated into plasmids for gene overexpression or short hairpin RNA (shRNA)-mediated gene knockdown. Breinig and colleagues introduced a library of eight plasmids that overexpress select oncogenes and knock down select tumor suppressor genes alongside sleeping beauty transposon plasmid via hydrodynamic tail vein injection to generate heterogeneous liver tumors in mice. This combinatorial screen found that VEGFA and CTNNB1 oncogenes were mutually exclusive with distinct liver tumor histology and TME compositions.⁹⁷

Perturb-DBIT does not require any specialized plasmid or barcode design, as the sgRNAs are directly sequenced after either polyadenylated capture (PAC) or direct capture (DC). In PAC, RNA molecules are enzymatically polyadenylated and then reverse transcribed with a biotinylated poly-T primer carrying UMI and linker sequences. These reactions produce cDNAs for all RNAs, including sgRNA, mRNA, and other small RNA species. In the DC version, the same poly-T primer is used to synthesize cDNAs for mRNAs, while a direct capture primer recognizing the sgRNA scaffold (constant for all sgRNAs, similar to poly-A constant for all mRNAs) is used for reverse transcription of sgRNAs. In both approaches, cDNAs are spatially barcoded and then sequenced with DBIT-seq. Baysoy et al. performed a whole-genome CRISPR screen in a xenograft metastatic lung model and identified top-enriched gene knockouts (e.g., S100A4 and MT1E) as well as transcriptional and non-coding RNA (e.g., tRNA and miRNA) alterations in these metastatic clones. The researchers performed another genome-wide CRISPR screen in a mouse model for metastatic breast cancer in the lung, identifying enriched knockouts in spatial clusters with distinctive TMEs and associated ligand-receptor signaling. They confirmed immune phenotypes on a select number of tumors using CODEX on an adjacent tissue section, which allows spatial overlaying of CODEX immune staining with the DBIT-sequencing of the same sample.⁹⁸

Spatial functional genomics is an exciting frontier where rapid technological developments have led to unprecedented discoveries in tumor biology. The power of these methods is their ability to establish causative molecular mechanisms between cancer genetic events and phenotypes (morphology, protein markers, gene expression, and non-coding RNA) across cell types (cancer, immune, stromal, etc.) and spatially (clonal dynamics, cell-cell interactions, immune TME archetypes, etc.). In addition to generating immensely rich (i.e., high-content) data, these methods rapidly test multiple hypotheses in parallel (i.e., high-throughput), expediting biological discoveries. As described above, different methods require different cloning, barcoding, and imaging/sequencing strategies. Some of these techniques are more laborious, time-consuming, expensive, or simply more restricted due to limited access to specialized equipment. Hence, each method comes with its pros and cons (Table 3).

Table 3.

Spatial functional and clonal tracing methods

	Category	Method	Detection	Scale (theoretical)	Scale (demonstrated)	Tissue mapped	Resolutionc	Accessibility (materials)b	Accessibility (instrument)b
Spatial functional	proteomics	Perturb-map	mIHC/mIF	100s	364 barcodes	mouse lung adenocarcinoma	single cell	moderate	moderate
	–	EpicMIBI	MIBI	10s	20 barcodes	mouse small cell lung cancer	single cell	moderate	difficult
	–	EPICodea	mIF	100s	500 barcodes	in vitro	single cell	moderate	moderate
	ISH	Perturb-FISH	MERFISH	1,000s	35 barcodes	human melanoma in mouse xenograft	single cell	easy	moderate
	–	Perturb-Multi	RCA-MERFISH	1,000s	252 barcodes	mouse liver	single cell	moderate	difficult
	–	CRISPRmap	seqFISH	1,000s	364 barcodes	human esophageal carcinoma in mouse xenograft	single cell	moderate	difficult
	ISS	PerturbView	ISS	1,000s	100 barcodes	human colorectal carcinoma in mouse xenograft	single cell	moderate	difficult
	–	CROPseq-multia	ISS	1,000s	NA	NA	na	moderate	moderate
	spatial barcoding	Perturb-CAST	Visium	100s	256 barcodes	mouse hepatocellular carcinoma	near single-cell	difficult	moderate
	–	Perturb-DBIT	DBIT-Seq	genome-wide	genome-wide	lung metastases of human colorectal carcinoma or mouse breast cancer cell lines transplanted into mice	near single-cell	easy	difficult
Clonal tracking	scRNA	scGESTALTa	scRNA-seq	1,000s	376 barcodes	zebrafish brain	single cell	moderate	moderate
	–	ScarTracea	scRNA-seq	1,000s	8 clones	zebrafish hematopoietic cells	single cell	moderate	moderate
	–	LINNAEUSa	scRNA-seq	1,000s	748 barcodes	zebrafish larva	single cell	moderate	moderate
	–	macsGESTALTa	scRNA-seq	1,000s	95 clones	mouse pancreatic cancer metastases	single cell	moderate	moderate
	ISH	intMEMOIR	FISH	10,000s	29 clones	Drosophila brain	single cell	difficult	moderate
	–	Spacebar	seqFISH	10,000s	96 clones	human melanoma in mouse xenograft	single cell	moderate	moderate
	–	Zombie	FISH	10s	39 barcodes	chick embryo brain	single cell	moderate	moderate
	ISS	BaSISS	ISS	10s	6 subclones	human breast cancer	single cell	difficult	moderate
	spatial barcoding	Slide-DNA-seq	Slide-Seq	10s	2–3 subclones	human colorectal carcinoma; liver metastases of mouse lung adenocarcinoma syngeneic transplants	near single-cell	easy	difficult
	–	Slide-tags	Slide-Seq	10s	2 subclones	human melanoma lymph node metastasis	near single-cell	easy	difficult
	–	Space-TREX	Visium	10,000s	162 clones	mouse brain	near single-cell	moderate	moderate
	–	Spatial iTracer	Visium	1,000s	81 barcodes	mouse brain organoid	near single-cell	moderate	moderate
	–	KP-Tracer	Slide-Seq & Slide-tags	1,000s	10s subclones	mouse lung adenocarcinoma	near single-cell	moderate	difficult

^aTissue mapping with spatial resolution not formally demonstrated.

^bRelative and subjective comparison.

^cWhile ISS and ISH methods can detect barcode signals at subcellular resolutions, barcodes are used to track individual cells.

Besides these practical considerations, perturbation scale, phenotyping capacity, resolution, and dimension differ for each method. The perturbation scale is determined by barcode capacity. SP-based methods tend to be more limited due to inherent limitations with the number of available epitope tags or antibodies. On the other hand, some ST-based methods offer genome-scale perturbation. However, it is critical to appreciate that tumor tissue in mouse models, where these CRISPR screens are performed, can only contain a certain number of tumor clones. An analyzable tissue section is even further limited in how many tumor clones it can contain. It is simply mathematically not feasible to perform a genome-wide spatial CRISPR screen, unless the focus is to quantify perturbation enrichment/depletion, which simply defies the purpose of performing a spatial screen as opposed to a bulk sequencing-based screen. Therefore, independent of the theoretical limits of barcoding capacity of the chosen method, spatial CRISPR screens require prioritization of perturbation targets to allow sufficient observation of each perturbation for statistically meaningful conclusions.

Phenotyping capacity (i.e., number of markers, transcripts, etc.), resolution (i.e., single cell versus bin/spot), and dimension (i.e., actual physical tissue area) are all determined by the spatial analysis method chosen to analyze the screen. As discussed in the previous sections, SP methods provide single-cell resolution in larger tissue areas at the expense of more limited phenotyping. ISH and ISS methods provide sub-cellular resolution and moderate phenotyping capacity (hundreds to a few thousand transcripts). Spatial barcoding methods reach the whole transcriptome scale phenotyping but lack true single-cell resolution.

Beyond standard proteomic or transcriptomic mapping, spatial functional genomics introduces a critical layer of complexity: the high-fidelity assignment of perturbations to specific cells. Each detection modality faces distinct challenges in this regard. SP-based methods are prone to epitope loss, signal masking, and cumulative tissue damage caused by repetitive staining cycles. Furthermore, the limited dynamic range and varying signal-to-noise ratios of protein-based barcodes can skew debarcoding accuracy. In contrast, ISH and ISS-based methods often struggle with optical crowding and signal overlap in dense tissues. Because sgRNAs provide limited binding space for probes, maintaining high specificity without sacrificing detection efficiency remains a hurdle, even with signal amplification. Spatially barcoded sequencing faces a different set of “collision” issues, where multiple perturbations may coexist within a single spatial bin. Furthermore, lateral diffusion of transcripts during tissue lysis and the inherent capture bias associated with engineered poly-A scaffolds can lead to misassignments. Consequently, rigorous optimizations are mandatory. This includes analyzing barcode libraries in the absence of perturbations to ensure uniform detection and using non-targeting sgRNAs to establish a baseline. Ultimately, significant “hits” identified in these screens must be validated through low-plex follow-up experiments to confirm the observed phenotypes.

Spatial clonal tracing

Spatial functional genomics methods described above are capable of spatially tracing tumor clones when uncoupled from Cas9-mediated mutagenesis, where sgRNA sequences or associated barcodes are used to track individual clones on a tumor section. In fact, some of the methods described above have only been demonstrated for clonal tracing rather than perturbation mapping. The current frontier in clonal tracing is the combination of both space and time. The emerging methods combine tools from temporal clonal tracing, historically analyzed by bulk sequencing and more recently by scRNA-seq, with spatial sequencing methods. In essence, spatiotemporal clonal tracing methods use either spatial DNA sequencing of polymorphic DNA sequences or spatial RNA sequencing of modifiable barcodes that record the ancestry of cells. Reconstruction of polymorphic DNA sequences or backtracking of recorded barcode alterations reveals a phylogenetic tree of the clones (Figure 4).

Figure 4.

Spatiotemporal omics methods

Slide-based approaches infer lineage relationships retrospectively by detecting naturally occurring DNA variants. IntMEMOIR leverages integrase-mediated recombination to generate stochastic edits across arrays of engineered target sites, which are cumulatively recorded over time. Lentiviral approaches introduce stable genetic barcodes into individual cells that are heritably maintained across cell divisions. In contrast, CRISPR-based methods use barcodes targeted by sgRNA-Cas9 complexes to accumulate stochastic mutations over time, enabling prospective recording of lineage history across generations.

A DNA sequencing-based method that does not need barcodes, hence better aligned for analysis of human samples, is slide-DNA-seq. This method uses bead-based spatially barcoded slides, such as those used for slide-Seq described above, and employs tagmentation with hyperactive Tn5 transposase to link DNA fragments to spatially indexed bead oligonucleotides. Zhao et al. clonally mapped human colorectal cancer samples as well as mouse lung adenocarcinoma metastases in the liver with slide-DNA-seq, where principal component analysis (PCA) of the sequencing data distinguished clones.⁹⁹ The same group also developed Slide-tags, where a tissue sample is placed on a slide with DNA-barcoded beads, as in slide-DNA-seq, followed by photocleavage of spatial barcodes to tag single nuclei, which are then dissociated and subjected to single-nucleus (sn) profiling assays. Russell et al. applied this method to a human melanoma sample and analyzed the cells with droplet-based combinatorial snRNA-seq and snATAC-seq. These analyses revealed two spatially distinct tumor clones that also differ in their cancer cell differentiation states, antigen presentation, and T cell infiltration phenotypes.¹⁰⁰

Base-specific in situ sequencing (BaSISS) requires serial tissue sections, where whole genome sequencing of the first section informs the design of padlock probes that recognize clone-defining somatic variants. ISS of the consecutive tissue section then reveals clones, in addition to measuring gene expression. Using BaSISS on human breast cancer samples, Lomakin et al. showed that spatially distinct but genetically similar clones maintain their transcriptional and histological features.¹⁰¹

Integrase-editable memory by engineered mutagenesis with optical in situ readout (intMEMOIR) uses integrase-editable transcribed barcodes for clonal tracing. Each barcode consists of 10 three-state memory elements (i.e., trits), which are short sequences flanked by orthogonal serine integrase target sites (two attP and one attB). Depending on which of the attP sequences is used by the enzyme Bxb1, the trit sequence gets either deleted or inverted. The recombination is irreversible and continues to record enzymatic activity over generations of cells. Chow et al. demonstrated intMEMOIR in Drosophila brains, where they used heat shock-inducible Bxb1 barcoded neuroblasts in embryos to track neuronal progeny in adults.¹⁰² While intMEMOIR has not been published for tumor studies yet, it stands as a viable option, as both the barcode and inducible enzyme (e.g., doxy-inducible) can be stably expressed in a cancer cell line, and the clones can be identified on tissue samples using FISH.

TRacking and gene EXpression profiling (TREX) uses a lentiviral barcode library, where random 30 bp barcode sequences are incorporated at the 3′ end of nuclear-localized fluorescent protein (H2B-EGFP) expressed under EF1a promoter. Ratz et al. demonstrated this library in the developing mouse brain and employed ST to identify clones concomitant with transcriptional profiling.¹⁰³ A similar, lentiviral and GFP-fused barcode, called SpaceBar demonstrated the feasibility of tracing tumor clones with simultaneous ST in a xenograft mouse model for melanoma.¹⁰⁴

Several methods use CRISPR mutagenesis to record clonal development. In essence, all these approaches target inducible Cas9 to a defined barcode sequence to introduce random insertions/deletions (i.e., “scars”) and then sequence the transcribed barcode. Single-cell genome editing of synthetic target arrays for lineage tracing (scGESTALT) uses an array of 10 CRISPR targets transcribed at the 3′ UTR of dsRed reporter.¹⁰⁵ ScarTrace uses eight tandem copies of H2A-GFP as its transcribed barcode.¹⁰⁶ Lineage tracing by nuclease-activated editing of ubiquitous sequences (LINNAEUS) uses CRISPR to introduce random insertions/deletions at RFP expression cassettes that are present in 16–32 independent genomic loci of a zebrafish model.¹⁰⁷ While these single-cell clonal tracing methods were demonstrated in zebrafish development models, a modified version of scGESTALT, called multiplexed, activatable, clonal, and subclonal GESTALT (macsGESTALT) was used in the KPCY (Kras^LSL-G12D/+; Trp53^LSL-R172H/+; Pdx1-cre; LSL-Rosa26^YFP/YFP) pancreatic cancer mouse model. macsGESTALT uses five sgRNA target sites tandemly expressed at the 3′ UTR of the puromycin resistance gene under the EF1a promoter. U6 promoter controls expression of five tandem sgRNAs separated by tRNA spacers. Simeonov et al. used doxycycline-inducible Cas9 to “scar” primary PDAC and performed scRNA-seq of the primary and metastatic tumors to create a lineage map of PDAC metastasis. They found that despite sharing the same genetic background, most clones did not metastasize. Acquisition of a late-hybrid EMT state and expression of the S100 gene marked the rare metastatic clones.¹⁰⁸

Another CRISPR scar approach is iTracer, where the barcode is located at the 3′ UTR of GFP or RFP and delivered with the Sleeping Beauty plasmid, which also expresses a sgRNA targeting the barcode sequence. Doxy-induced Cas9 creates random insertions/deletions (i.e., scars) in the target sequence, creating unique barcodes. An additional sgRNA cassette enables simultaneous genetic perturbation and lineage tracing. He et al. demonstrated iTracer in human cerebral organoids and used Visium for ST.¹⁰⁹

Zombie uses CRISPR adenine base editing to precisely mutate a single nucleotide in each of a series of barcodes, which are encoded on a lentiviral plasmid and downstream of phage promoters (T3, T7, SP6). Hence, the barcodes are not expressed in live cells but can be transcribed in fixed cells with phage RNA polymerases (e.g., T3) and detected by targeted (i.e., probe-based) ST methods, as demonstrated in mouse brains.¹¹⁰

Spatial tracing models that combine multiple approaches described above are emerging. For example, the KP-Tracer mouse model was recently described, which utilizes scRNA-seq reading of CRISPR scarred barcodes placed downstream of mCherry reporter, which was introduced to the KP (Kras^LSL-G12D/+; Trp53^fl/fl) Cas9 (Rosa26^{LSL-Cas9-P2A-mNeonGreen}) mouse model. Intratracheal injection of adenoviruses or lentiviruses that deliver Cre recombinase initiates multiple non-small cell lung tumors arising from normal lung epithelium in these mice.¹¹¹ Jones et al. analyzed KP-Tracer tumors using Slide-seq and Slide-tags, explained above, to spatially trace lung tumors in this mouse model. They identified spatial communities associated with tumor progression, such as a hypoxic environment that leads to fibrosis, immune cell reprogramming, and the emergence of pro-metastatic clones.¹¹²

Each spatial clonal tracing method presents a distinct set of technical trade-offs (Table 3). Slide-based approaches are particularly advantageous for human clinical samples, as they can leverage natural somatic variants, circumventing the need for live-cell genetic engineering. However, these methods are susceptible to tissue fragility, barcode bleeding across cell boundaries, and the inherent transcript biases of single-nucleus sequencing. Lentiviral barcoding offers a robust alternative for prospective lineage tracking, yet achieving a precise multiplicity of infection (MOI) remains challenging, particularly in vivo; a high MOI leads to ambiguous lineage assignments, while an overly low MOI results in sparse labeling. Furthermore, lentiviral integration sites can induce transcriptional silencing or disrupt endogenous cellular physiology. Because these barcodes are static, they offer limited resolution into the iterative branching of a lineage tree. CRISPR-based recorders address this by enabling continuous evolution of the barcode; however, their recording capacity is finite. In systems with high proliferative activity, the barcode site may reach saturation, where no further mutations can be stored. Additionally, stochastic CRISPR activity can result in homoplasy, where identical mutations occur independently in unrelated cells, leading to erroneous clonal merging. Ultimately, all spatial methods are constrained by the limited area problem. A 2D section provides only a partial snapshot of a 3D phylogeny. While multi-section analysis can reconstruct these volumes, it imposes significant financial and computational burdens. Consequently, researchers must weigh these practical limitations to choose a suitable method for their studies and remain mindful of the current limitations in this space.

Conclusions and future perspectives

Spatial biology is the newest frontier in cancer research, and this rapidly developing domain has already provided a previously unattainable depth of knowledge regarding tumor architecture, heterogeneity, intercellular communication, and how all these define the tumor immune cycle. Although the spatial oncology field is in its early stages, the precipitously increasing number of publications, which we could only highlight a select set of here, is a sign of the excitement in the field. However, there are several challenges ahead that need to be addressed to fully utilize spatial biology methods for basic science discoveries as well as clinical applications.

In an ideal scenario, there should be a standardized method that can capture both transcript and protein levels genome-wide/proteome-wide at single-cell or subcellular resolution. This method should be highly specific and sensitive to detect mRNAs and proteins, and relatively fast to analyze large tissue sections. In reality, the current methods offer trade-offs between such qualities. While it is exciting that more methods and tools are being developed to address these limitations and offer new possibilities, it is becoming harder to navigate the ever-evolving landscape of spatial biology methods. Often, new methods and tools are not adopted because they require extensive investment in terms of both budget and time. Hence, for most newcomers to the field, the commercial solutions present the only viable options. However, these methods are also prohibitively expensive, especially if optimization rounds or large numbers of samples are involved. Handling and analysis of large-scale datasets is another hurdle before the democratization of spatial biology.

In addition to improvements in sensitivity, specificity, scale, resolution, and speed, integration of multiple omics measurements will expand how much we can learn in a single experiment. In addition to RNA and protein measurements, spatial analyses of DNA, epigenome, and metabolome are emerging. While the use of serial tissue sections can aid in combining any given method, the gold standard would be the ability to combine selected spatial omics on the same tissue section. Emerging technologies provide exciting spatial multi-omics capabilities, such as spatial ATAC-RNA-Seq and spatial Cut&Tag-RNA-Seq for integrated epigenomics and transcriptomics¹¹³; spatial co-indexing of transcriptomes and epitopes (spatial CITE-seq)¹¹⁴ and spatial protein and transcriptome sequencing (SPOTS)¹¹⁵ for combined protein and transcriptomics; spatial multimodal analysis (SMA)¹¹⁶ for transcriptomics and metabolomes. Other methods, such as multi-omics in situ pairwise sequencing (MiP-seq),¹¹⁷ spatial assay for transposase-accessible chromatin (ATAC)-RNA-protein-seq (spatial ARP-seq),¹¹⁸ and CUT&Tag-RNA-protein-seq (spatial CTRP-seq),¹¹⁸ are pushing the limits of multi-omics with three modality combinations.

The majority of spatial studies are performed on two-dimensional tissue samples that are often a few micrometers thick. Hence, we are blind to cells and molecules present on any other angular plane other than the one present on the slide. One way to mitigate this limitation is computational stitching of multiple serial sections to create a quasi-3D map. In essence, this approach is viable with most of the spatial biology methods but requires significantly more experimental and computational time. On the other hand, there are emerging 3D ST and SP methods that take advantage of tissue clearing, such as STARmap and three-dimensional imaging of solvent-cleared organs profiled by mass spectrometry (DISCO-MS), or expansion microscopy, such as ExSeq. STARmap and ExSeq are limited to a few hundred (around 200–300) microns of tissue thickness. Whereas DISCO-MS can work on entire organs, it lacks single-cell resolution.^38,40,119 3D CyCIF employs a laser scanning confocal microscope to image 30–50 μm thick tissue samples stained with the CyCIF method to capture more intact cells and cellular neighborhoods compared to 2D SP methods usually performed on 5 μm tissue sections.¹²⁰

Compared to the wider adoption of ST and SP, spatial functional genomics and spatiotemporal tracking remain niche, partly because these methods are even more technically sophisticated and expensive. Any technical advancement in the spatial omics technologies is poised to improve spatial functional genomics and tracking since ST and SP methods are used for the detection of mutants and clones. While cell transplantation enables in vivo tumor experiments for many cancer types, integration of spatial functional genomics and tracking with genetically engineered mouse models will capture immune selection and evolution of tumors developing from healthy cells all the way to metastases in their native microenvironments. Organoids and humanized mouse models can open the way to perform forward genetics and clonality studies with patient-derived samples.

Spatial data analysis remains a primary bottleneck in realizing the full potential of spatial biology. While computational workflows vary by modality, several demanding challenges are shared across platforms. For imaging-based methods, pre-processing requires tiling of fields of view and accurate image registration across iterative detection cycles. This process often relies on fiducial landmarks, such as fluorescent beads, or increasingly on feature-based and deep learning-driven registration approaches to correct for tissue warping or damage incurred during staining cycles. A more fundamental challenge is cell segmentation; in the absence of universal membrane markers, algorithms frequently infer cell boundaries from nuclear stains using heuristic expansion or machine learning methods, leading to inaccuracies in dense tissues or complex morphologies. Imaging-based transcriptomics methods are also constrained by limited transcriptome coverage; to address this, imputation methods leverage scRNA-seq reference data to predict missing expression profiles, although this can introduce biases toward reference atlases. These approaches also facilitate cell-type deconvolution in sequencing-based spatial methods that lack single-cell resolution. While extensions of scRNA-seq analytical frameworks can be applied to the resulting count matrices, the true value of spatial datasets lies in their spatial context. Downstream analyses, therefore, include identification of spatially variable genes, cellular neighborhoods and niches, intercellular signaling (e.g., ligand-receptor interactions), and multimodal integration with histological features. Although commercial platforms for spatial data analysis exist, they are often platform-specific and inflexible. Consequently, the field has gravitated toward transparent, open-source ecosystems such as SquidPy,¹²¹ Scimap,¹²² Giotto,¹²³ SpatialData,¹²⁴ and Seurat,^125,126 which provide scalable and customizable frameworks for this rapidly evolving bioinformatics frontier.

Ongoing tissue atlas projects such as the Human Tumor Atlas Network¹²⁷ and Multi-Omic Spatial Atlas in Cancer¹²⁸ are large-scale coordinated efforts to spatially profile human tumors. Open-access datasets from such atlas projects can help cancer researchers around the world refine their hypotheses or experimental designs and explore the patient relevance of their findings. These projects and many other individual ones are starting to integrate artificial intelligence (AI), which can significantly expedite inferring meaningful observations from vast spatial datasets.^{121,129,130,131,132,133,134,135} Another front is the use of AI to extrapolate ST data from simple hematoxylin and eosin (H&E) stained patient samples.^{136,137,138,139,140,141} A realistic prediction would be that such AI analysis of H&E, combined with low-plex spatial profiling, can be regularly used in near-future clinics to guide treatment decisions for targeted and immunotherapies.

Patients are routinely stratified for immune checkpoint inhibitors (ICIs) and targeted therapies primarily based on marker expression assessed by IHC, often complemented by molecular assays. For example, IHC for PD-L1 is commonly used to guide anti-PD-1/PD-L1 therapies in addition to biomarkers such as microsatellite instability and, less frequently, tumor mutational burden. However, the predictive power of these assays remains limited, motivating the search for improved biomarkers. Recent spatial studies have identified candidate spatial biomarkers with enhanced prognostic and predictive associations. For instance, GeoMx analysis of lung cancer samples demonstrated that PD-L1 expression in macrophages, rather than tumor cells, was more strongly associated with ICI response and overall survival.¹⁴² Spatial profiling of ICI-responsive and resistant liver tumors using MERFISH and mIHC revealed coordinated cellular triads of CXCL13⁺ T helper cells, PD-1-high progenitor CD8 T cells, and dendritic cells enriched in maturation and regulatory molecules in responsive tumors.⁶⁵ Similarly, mIHC analysis of pancreatic tumors showed that spatial proximity between IL10⁺ myelomonocytes, PD-1⁺ CD4 T cells, and granzyme B⁺ CD8 T cells correlates with patient survival.¹⁴³ CODEX analysis of cutaneous T cell lymphoma further demonstrated that differences in tumor microenvironment topology, including regulatory T cells and effector PD-1⁺ CD4 T cells, are associated with response to ICI therapy.¹⁴⁴ Across multiple spatial platforms, including CODEX and CosMx, distinct macrophage states have also emerged as important predictors, with IL4I1⁺ macrophages correlating with favorable outcomes and SPP1⁺ macrophages associated with hypoxic niches and poor prognosis.¹⁴⁵ In addition, tertiary lymphoid structures, which are spatially organized immune aggregates enriched in T, B, and dendritic cells, consistently predict survival and immunotherapy response across multiple cancer types.^146,147,148

While these studies highlight the promise of spatial biology for improving clinical outcome prediction, translation into routine clinical practice remains limited by several challenges. Given that FFPE tissue samples and IHC are the mainstay of clinical pathology, early clinical adoption will likely focus on low-plex spatial proteomic approaches compatible with FFPE workflows. However, instrument and reagent costs, data storage and computational requirements, and the need for specialized bioinformatics expertise represent significant barriers. Moreover, trade-offs between throughput and spatial resolution must be carefully managed: although approaches such as tissue microarrays enable high-throughput analysis, reduced tissue area limits the ability to capture ITH. Standardization of protocols and reproducibility across platforms and pathology laboratories will also be critical, and initial efforts toward harmonization are already underway.^149,150

What we have seen in a few short years since the inception of spatial biology has demonstrated its immense potential, and the years ahead should capitalize on and realize its full capabilities. There is an exciting future ahead, when we will witness advancements and refinements in spatial biology methods, which will become more affordable and accessible. Integrated with AI, spatial biology will continue to transform our understanding of tumor biology in laboratories and inform diagnosis and treatment decisions in the clinics.

Acknowledgments

This work was supported in part by the University of Alabama at Birmingham start-up funds provided to G.M.

Author contributions

N.S., investigation, visualization, writing – review & editing; S.E., writing – review & editing; G.M., investigation, writing – original draft, writing – review & editing, and funding acquisition.

Declaration of interests

The authors declare no competing interests.