Unveiling the role of spatial transcriptomics in the analysis of the tumor immune microenvironment (Review)

Introduction

The tumor immune microenvironment is a key factor determining tumor initiation, progression and therapeutic response (1,2). Its spatial heterogeneity and complex intercellular interactions have long been challenging to decipher. Although traditional single-cell sequencing technologies can reveal cellular diversity, they lose vital spatial location information. The emergence of spatial transcriptomics (ST) has provided a revolutionary tool for analyzing the tumor immune microenvironment in spatial dimensions by preserving in situ spatial coordinates while detecting whole-transcriptome or specific gene set expression (3,4).

ST technologies have been widely applied across diverse cancer types, systematically revealing the spatial organization and functional states of different cell populations within the tumor immune microenvironment. These technologies enable the precise localization of key cell subsets and the effective identification of tumor niches with key biological functions, including tertiary lymphoid structures (TLS), immune exclusion niches and tumor invasion front niches (5,6). The present review systematically summarizes the core advances of ST in elucidating the composition of the tumor microenvironment (TME), the spatiotemporal logic of immune escape and the evaluation of clinical significance across distinct stages of tumor progression, from premalignant lesions and locally advanced disease to distant metastasis.

The fine-resolution molecular information provided by ST is driving disease subtyping, prognostic prediction and treatment response assessment toward greater precision, a progress particularly pronounced in oncology. Studies (7-10) have shown that the spatial molecular landscape of tumors is closely associated with their growth behavior, metastatic potential, and response to immunotherapy, highlighting the value of this technology in two clinically translational directions: The discovery of spatial context-associated biomarkers and the elucidation of molecular mechanisms underlying therapeutic response. However, the practical application of ST in clinical diagnosis and decision-making still faces a series of key challenges, including insufficient resolution, limited sample representativeness, risk of algorithmic bias and lack of cross-platform standardization. The rapid advancement of machine learning has provided new opportunities to address these technical bottlenecks (11). As multi-omics technologies and machine learning become increasingly integrated, with continued breakthroughs in multimodal data integration and three-dimensional tissue reconstruction, the understanding of tumor evolution will continue to deepen.

Overview of spatial transcriptomics

The emergence of spatial transcriptomics

Single-cell RNA sequencing (scRNA-seq) technologies have evolved from initial transcriptome profiling to encompass multiple omics dimensions, including three-dimensional genomics, epigenomics and proteomics, and have even enabled high-throughput single-cell multi-omics sequencing (12-16). These advancements have become key drivers across a wide range of research areas (17). ScRNA-seq is used to profile RNA expression in individual cells, providing insights into cell-type-specific functions under specific conditions. However, scRNA-seq requires dissociating cells from their native tissue microenvironment. During this dissociation process, cells lose their original spatial context, thereby precluding the detection of influences from neighboring cells or the surrounding environment, as well as the coordinated relationship between spatial organization and cellular function. ST integrates molecular biology with histology techniques, enabling direct mapping of RNA molecule distributions on tissue sections and resolving the spatial features of gene expression. This approach thus overcomes a fundamental limitation of scRNA-seq (18). Based on their underlying molecular detection principles, ST technologies can be broadly classified into two main categories: Sequencing-based and imaging-based ST technologies.

Technical classifications and platform comparisons

Sequencing-based ST technologies

Sequencing-based ST technologies originated from single-cell sequencing. In 2016, Stahl et al (19) first proposed a ST method based on spatial barcoding, which has since been widely adopted. This technique enables whole-transcriptome capture by converting the cell barcodes used in single-cell sequencing into spatial location barcodes and replacing bead-based single-cell capture with fixed-size, spatially defined spots on a chip. It then employs amplification and sample indexing principles similar to those of single-cell transcriptome sequencing to achieve high-throughput spatial transcriptome sequencing. Several improved, novel sequencing-based ST technologies have emerged, including 10x Genomics Visium (20), Slide-Seq (21) and Stereo-seq (22). Visium offers enhanced resolution and sensitivity, is widely used in neuroscience, cancer research, and developmental biology, and benefits from a mature commercial platform and standardized experimental protocols, including support for formalin-fixed and paraffin-embedded (FFPE) samples. Slide-Seq uses randomly deposited barcoded beads on a slide for mRNA capture, providing higher resolution suitable for applications such as nervous system mapping or liver lobule zonation; however, the stochastic distribution of beads may affect data consistency. Stereo-seq achieves nanoscale resolution, enabling subcellular-level analysis, which makes it suitable for tissue mapping in contexts such as embryonic development, although its demanding experimental requirements pose certain limitations (summarized in Table I).

Table I Comparison of sequencing-based spatial transcriptomics technologies.

Table I

Comparison of sequencing-based spatial transcriptomics technologies.

Core performance parameters	10x Genomics visium	Slide-seq	Stereo-seq
Transcript capture strategy	Oligonucleotides containing spatial barcodes and capture sequences are pre-synthesized on defined spots (55 μm in diameter) on a glass slide, with the positional barcode of each spot decoded via sequencing. After tissue section attachment, mRNA diffuses to the capture spots through permeation for binding.	A monolayer of 10 μm beads coated with random barcode DNA oligonucleotides is arrayed on a glass slide. The spatial position of each bead is decoded via SOLiD sequencing, followed by tissue section attachment for mRNA capture.	A patterned array of DNBs serves as the capture substrate, with each DNB carrying a spatial barcode and a poly-T capture sequence. The high-density DNB array generated by rolling circle amplification achieves ultra-high resolution.
Resolution	55 μm	10 μm	500 nm
Detection sensitivity	~500 UMIs per 100 μm2	~550 UMIs per 100 μm2 for the V2 version.	~1,450 UMIs per 100 μm2
Throughput and capture area	Moderate. One slide contains four capture areas (6.5×6.5 mm2 each), allowing 1-4 samples to be processed simultaneously. Gene read throughput depends on sequencing depth.	Low to moderate. A single slide typically accommodates only one sample (6.2×6.2 mm2), resulting in low sample throughput. However, the gene detection throughput at the single-cell/per-spot level is very high.	High. A single chip can accommodate multiple samples (up to 13.2 cm×13.2 cm). The large capture area combined with high resolution yields exceptionally high data output.
Tissue compatibility	FFPE (strong)/FF	FF (main)	FF (main)
Primary applications	Large-scale tissue gene expression analysis, suitable for liver, tumor, neural tissues.	Studies of nervous system, liver lobule zonation.	Ultra-high-resolution tissue mapping, for example, embryonic development, tumors
Advantages	Mature commercial platform, standardized workflow, supports FFPE samples	Higher resolution than 10x Visium, supports whole transcriptome sequencing	Extremely high resolution, allows subcellular analysis, suitable for large tissues
Limitations	Low resolution, insufficient for single-cell level analysis	Limited capture area, random bead distribution affects data consistency	High computational demand, stringent experimental requirements

[i] FFPE, formalin-fixed paraffin-embedded; seq, sequencing; DNB, DNA nanoballs.

Notably, differences in sensitivity, throughput, resolution and other parameters across ST platforms directly determine the robustness of the same biological conclusion under different technological settings. Taking resolution and sensitivity as an example, Visium's 55 μm resolution mixes an average of 3-10 cells per spot, smoothing out cellular heterogeneity. As a result, conclusions that depend on rare cell types or low-abundance genes may be lost due to signal dilution. By contrast, Slide-Seq (10 μm) and Stereo-seq (500 nm) achieve single-cell to subcellular resolution. However, when the analytical unit is reduced to the pixel level in Stereo-seq, the robustness of its conclusions becomes highly dependent on the accuracy of subcellular pixel clustering algorithms, different algorithms may lead to radically different definitions of cell boundaries and divergent conclusions regarding gene localization.

More importantly, differences in analytical assumptions fundamentally determine whether the aforementioned hardware metrics can be translated into reliable biological conclusions. Specifically: Visium operates on the assumption that 'gene expression within each 55 μm spot represents the average state of that region'. This assumption holds in tissues with uniformly distributed cell types. However, in highly heterogeneous microenvironments (for example, tumor-immune boundaries), the mixed signal within a spot can lead to a conclusion of 'high gene expression' that is actually a summation artifact of multiple cell types rather than a true feature of any single cell type. Slide-Seq assumes that 'mRNA captured by a single 10 μm bead originates primarily from a single cell'. This assumption approximately holds in tissues with large, loosely packed cells (for example, cerebral cortex). But in cell-dense tissues (for example, embryonic tissues) or samples with abundant cell debris, it is common for a single bead to simultaneously capture mRNA from multiple cells, rendering the 'one bead=one cell' analytical premise invalid and leading to cell type annotation errors. Stereo-seq assumes that 'signals from 500 nm DNA nanoball (DNB) spots can be accurately reconstructed into true cell boundaries through post hoc clustering (binning)'. However, this assumption faces fundamental challenges in practice, the choice of clustering bin size (for example, bin50 vs. bin100) lacks a gold standard, and different sizes may yield mutually contradictory subcellular localization conclusions. More importantly, mRNA captured by DNB spots may originate from cellular protrusions (for example, neuronal axons or tumor pseudopodia) located hundreds of nanometers away, in which case the 'clustering reconstructs cell boundaries' assumption is biologically invalid. Therefore, any subcellular localization conclusion requires independent validation (for example, by spatial in situ hybridization) to confirm its robustness.

Imaging-based spatial transcriptomics technologies

Imaging-based ST technologies function by designing fluorescently labeled oligonucleotide probes that hybridize in situ with target RNAs, directly converting RNA molecular information into localizable and quantifiable fluorescence signals. These primarily include in situ sequencing (ISS) and in situ hybridization (ISH). Specifically, in ISS, RNA is reverse transcribed, amplified via rolling circle amplification (RCA), and sequenced in situ; this method has been applied in research related to cancer, tuberculosis and other diseases (23-26). Classic ISH techniques include multiplexed error-robust fluorescence in situ hybridization (MERFISH) (27) and sequential fluorescence ISH (seqFISH) (28), among others. In MERFISH, multiple rounds of hybridization and imaging are performed to detect the presence or absence of fluorescently labeled probes. Subsequently, sequential images are decoded using error-robust barcodes associated with each transcript's identity (27,29,30). MERFISH offers extremely high resolution and has been widely used for gene expression studies at single-cell and subcellular levels. seqFISH (31,32) is similar to MERFISH and is suitable for subcellular-level research, particularly in immunology. Both techniques can detect up to 10,000 target genes with subcellular resolution (28,33).

However, spatial resolution has largely been confined to two-dimensional tissue sections, even though the structure and distribution of cells and molecules within tissues are three-dimensional. The recent emergence of tissue clearing techniques, which render intact tissues or even whole organisms optically transparent, has gradually made three-dimensional immunofluorescence imaging a reality. Existing tissue clearing-based three-dimensional immunostaining techniques fall into three categories: Solvent-based clearing (34-36), aqueous-based clearing (37-40) and hydrogel-based clearing (41-44). The rapid development of tissue clearing technologies enables the application of imaging-based ST to three-dimensional tissues and organs, as demonstrated by techniques such as ExSeq combined with CLARITY and smHCR combined with PACT.

The aforementioned imaging-based technologies also exhibit notable differences in core performance parameters, which directly determine the robustness of biological conclusions across different research contexts (Table II). First, regarding transcript capture strategies, both MERFISH and SeqFISH rely on multi-round specific probe hybridization, whereas ISS depends on signal amplification via RCA. This makes ISS more sensitive to single-base mismatches (suitable for mutation detection), but RCA efficiency varies considerably among different RNAs, potentially introducing quantitative bias (45-48). Second, in terms of resolution, although all three claim subcellular resolution, the actual effective resolution is influenced by the diffraction limit of the imaging system and probe density: MERFISH typically requires an optical resolution of ~200 nm to resolve single molecules, whereas the RCA products of ISS can cause signal overlap in densely expressed regions, leading to gene misassignment errors (27,49,50). Third, regarding detection sensitivity, single-molecule FISH (smFISH) directly labels each RNA molecule and offers the highest theoretical sensitivity; ISS achieves signal amplification via RCA, enabling detection of low-abundance transcripts, but at the cost of higher background noise (51-53). Fourth, the trade-off between throughput and capture area is critical: MERFISH can cover ~1 cm2 per experiment detecting thousands of genes, while the high-throughput version seqFISH+ can detect tens of thousands of genes but requires days of imaging time, limiting sample throughput (28). Fifth, in terms of tissue compatibility, multi-round hybridization technologies (MERFISH, SeqFISH) have stringent requirements for tissue autofluorescence and RNA preservation, performing poorly on FFPE samples; by contrast, the RCA products of ISS are more resistant to photobleaching and, when combined with tissue clearing, are better suited for three-dimensional imaging (for example, ExSeq+CLARITY) (54,55). Finally, differences in analytical assumptions are often overlooked: MERFISH relies on Hamming distance for barcode error correction, assuming that the error probability of each hybridization round is independent and known; whereas ISS decoding depends on sequential reading of fluorescent signals across sequencing rounds, assuming that amplification products remain intact throughout the entire sequencing process. If these assumptions are violated, MERFISH's error correction mechanism can still partially recover data, while ISS may produce false negatives due to base deletions (27,56,57).

Table II Comparison of imaging-based spatial transcriptomics technologies.

Table II

Comparison of imaging-based spatial transcriptomics technologies.

Core performance parameters	MERFISH	seqFISH/seqFISH+	ISS
Transcript capture strategy	A combinatorial binary encoding strategy is used. Each mRNA is assigned an N-bit binary barcode, which is read out through the presence or absence (1/0) of fluorescent probes across multiple imaging rounds. Error-correcting codes (for example, Hamming distance) are employed to ensure detection accuracy.	Through multiple rounds of hybridization-imaging-probe stripping cycles, each round uses a set of fluorescently labeled probes to generate a unique temporal barcode for each mRNA molecule.	reverse transcription → rolling circle amplification → in situ sequencing
Resolution	Subcellular (~200 nm)	Subcellular (~200-300 nm)	Subcellular (~300-500 nm)
Detection sensitivity	Single-molecule level (directly labeling each RNA molecule)	Single-molecule level (direct labeling)	Ssingle-molecule level with signal amplification
Throughput and capture area	Capable of detecting 10,000-20,000 genes.	seqFISH+: ~24,000 genes, but with imaging time spanning several days	Capable of detecting hundreds to a thousand genes.
Tissue compatibility	Best suited for fresh-frozen tissue; poor performance on FFPE; 3D imaging requires combination with tissue clearing.	Similar to MERFISH	Compatible with both fresh-frozen and fixed tissues; when combined with tissue clearing (for example, ExSeq + CLARITY), enables high-quality 3D imaging.
Analytical assumptions	Hamming distance-based error. correction, assuming that the error probability in each hybridization round is independent and known	Early seqFISH had no error correction; seqFISH+ employs combinatorial barcoding and enables partial error correction.	It relies on the sequential reading of fluorescent signals across sequencing rounds, assuming that the RCA products remain intact and the signal does not attenuate throughout the entire sequencing process.

[i] MERFISH, multiplexed error-robust fluorescence in situ hybridization; seqFISH, sequential fluorescence in situ hybridization; ISS, in situ sequencing; RCA, rolling circle amplification; FFPE, formalin-fixed and paraffin-embedded.

In summary, the choice of ST technology depends on the specific research objectives. Sequencing-based technologies enable whole-transcriptome profiling and are suitable for studying complex tissues such as tumors. Imaging-based technologies generally offer higher resolution and are primarily applied at the single-cell or subcellular level for predefined gene panels rather than whole-transcriptome analysis (Table III). Notably, with continuous technological advancements in recent years, novel technologies capable of combining large fields of view with high resolution have become feasible.

Table III Comparison of sequencing-based and imaging-based spatial transcriptomics technologies.

Table III

Comparison of sequencing-based and imaging-based spatial transcriptomics technologies.

Core performance parameters	Sequencing-based spatial transcriptomics technologies	Imaging-based spatial transcriptomics technologies
Basic principle	Encodes positional information from tissue sections into 'spatial barcodes', acquiring transcriptomic data via high-throughput sequencing.	Uses fluorescently labeled probes for in situ hybridization with target RNAs, directly localizing RNA molecules through microscopic imaging.
Representative technologies	10x Genomics Visium, Slide-Se, Stereo-seq	MERFISH, SeqFISH, STARmap
Resolution	Typically tens to hundreds of micrometers; Stereo-seq can achieve 500 nm.	Generally higher, often at subcellular level.
Detection area	Larger	Smaller
Spatial dimensions	Primarily focuses on 2D tissue sections.	Enables 3D imaging via tissue clearing techniques (for example, CLARITY, PACT).
Primary applications	Tumors, developmental biology, whole-transcriptome mapping of complex tissues.	Neuroscience, immune microenvironment, subcellular localization, validation of cellular communication.
Experimental workflow	Relatively standardized, highly commercialized, suitable for FFPE samples.	Relatively complex, requires multiple rounds of hybridization and imaging, with stringent equipment and operational requirements.
Scalability	Easily integrated with multi-omics approaches.	Can be combined with multimodal imaging, such as immunofluorescence.

[i] STARmap, spatially-resolved transcript amplicon readout mapping; FFPE, formalin-fixed and paraffin-embedded.

In practice, ST is often used in combination with scRNA-seq. ScRNA-seq sequences individual cells from dissociated tissues, allowing for high-precision identification of all cell types and subtypes present within a tissue through unsupervised clustering and the expression of known marker genes (58). Computational deconvolution algorithms (such as Cell2location, SPOTlight and Tangram) (59-61) can integrate scRNA-seq and ST data, enabling high-throughput mapping of cell types and signaling within tissues, thus generating comprehensive cellular atlases. For instance, this approach allows visualization of the spatial enrichment of different immune and stromal cells within tumor tissues (62). ScRNA-seq data can also be used with computational tools (such as CellPhoneDB, NicheNet and ICELLNET) to predict potential cell-cell communication (CCC) mediated by ligand-receptor (L-R) interactions (63-66). ST can then validate whether these predicted interactions indeed occur between physically adjacent cells. Integrating both approaches facilitates the robust reconstruction of cell communication networks within their spatial context (67).

From initial bulk sequencing techniques, which could only determine average molecular signals from large, heterogeneous cell populations, to single-cell sequencing, which enabled the revelation of unique molecular profiles of individual cells and further to spatial single-cell omics, which allows the resolution of gene expression programs and cell-cell interactions among heterogeneous cells in different regions of a tissue or organ, or among homogeneous cells executing biological processes, the rapid advancement of spatial omics technologies has led to a more comprehensive and profound understanding of cellular functions in physiological, pathological, aging and apoptotic processes (68,69). Over the years, the resolution of ST technologies has continuously improved, evolving from the initial hundred-micrometer scale (Visium) to single-cell and even subcellular levels (Stereo-seq, MERFISH). Concurrently, the continuous development and optimization of data analysis methods, along with the integration of artificial intelligence technologies such as deep learning models, have enabled ST to integrate multi-level information from single-cell sequencing, proteomics and other modalities, providing richer data support for investigating tissue microenvironments, cell-cell interactions and disease progression. The TME is key for tumor initiation and progression. ST offers unique advantages for studying the TME, and research in this area has advanced rapidly (70,71).

Spatial dissection of the immune microenvironment during tumor progression

The tumor immune microenvironment is a dynamic and complex ecosystem composed of tumor cells, immune cells, cancer-associated fibroblasts (CAFs), vascular endothelial cells and extracellular matrix, among other components (2). During tumor initiation and progression, the tumor immune microenvironment serves not only as the supportive niche for tumor cell survival but also as a key site for immune evasion. Its composition is characterized not merely by cellular diversity, but more importantly by pronounced spatial heterogeneity in anatomical distribution (72). The emergence of ST has enabled in situ dissection of this spatial heterogeneity, revealing how stromal components restrict immune cell infiltration by constructing physical and chemical barriers, how immune cells undergo functional exhaustion within specific niches and why TLSs play diametrically opposite roles in different contexts (5,73). Such spatial information is pivotal in transforming the tumor immune microenvironment from a mere 'cell inventory' into a 'functional map', and it provides direct spatial targets for precision intervention.

Premalignant niches

ST enables the identification, at the very earliest stages of tumorigenesis, of spatial niche features that portend future immune escape and malignant progression and it reveals the spatiotemporal evolutionary patterns of premalignant lesions across multiple cancer types. In esophageal squamous cell carcinoma, single-cell ST has generated the first single-cell-resolution spatiotemporal dynamic atlas of multistage esophageal carcinogenesis (74), thereby surpassing the traditional linear evolution model and giving rise to a theoretical framework of 'spatial evolution of carcinogenesis'. To the best of our knowledge, this previous study (74) was the first to uncover that CAFs and invasive epithelial cells can form a key CAF-epithelial interactive niche (CAF-Epi niche) at the premalignant stage, which may serve as an important predictor of premalignant progression risk across several squamous cell carcinomas. In lung adenocarcinoma (LUAD), ST reveals that KRT8+ alveolar intermediate cells represent temporal and clonal precursors to recognizable premalignant lesions; these cells are highly sensitive to IL-1β and other inflammatory signals and drive tumor initiation at the premalignant stage. This niche is highly active in early lesions but attenuates in invasive LUAD, suggesting that targeting therapy in conjunction with immunotherapy may offer an effective strategy for early-stage LUAD (75). Moreover, the interplay between KRAS mutation and p53 inactivation can sculpt a 'progenitor-like' microenvironment. ST confirms that this state interacts with surrounding stromal and immune cells to form a self-reinforcing 'premalignant niche', and that targeting KRAS or restoring p53 function can dismantle this niche (76). Temporal analyses further demonstrate a sequential process of 'fibrosis first, then immune remodeling' during early LUAD carcinogenesis: fibroblast reprogramming precedes macrophage expansion, with fibrotic signaling biased toward the tumor core and inflammatory fibroblasts enriched at the periphery (77). In gastric cancer, ST combined with single-cell sequencing constructs a high-resolution spatiotemporal atlas spanning normal gastric mucosa, gastritis and gastric cancer. This reveals a trajectory of mature chief cells transitioning into cancer stem cells (CSCs) accompanied by activation of EGFR and WNT pathways; CSCs, together with CXCL13+ T cells, CCL18+ M2 macrophages and inflammatory CAFs (iCAFs), constitute a CSC niche that drives tumor progression (78). Another study further identifies an NAD-dependent immunosuppressive microenvironment as a key driver of early intestinal-type gastric carcinogenesis: the AREG/NAMPT dual pathway promotes malignant transformation by activating macrophages and fibroblasts, and blockade of this pathway can inhibit the progression of premalignant lesions (79). Collectively, these studies demonstrate that, the premalignant niches exhibit context-dependent variation in cellular composition and driving signals across cancer types and ST enables researchers to directly localize and validate these rare and dynamic premalignant niches within native tissue, thereby providing precise spatial targets for early intervention.

Locally advanced stage

The locally advanced stage of tumors is a key phase in which tumors transition from in situ lesions to an invasive phenotype, accompanied by spatial remodeling and functional dysregulation of the immune microenvironment (80,81). During this stage, tumor cells, infiltrating immune cells and stromal cells form complex interaction networks within specific spatial compartments, collectively driving key events such as immune evasion, T cell exhaustion and immune exclusion (81,82). ST overcomes the limitations of conventional dissociation-based techniques and resolves in situ, within native tissue, the dual functionality of TLSs, the spatially heterogeneous niches of CAFs, the local drivers of T cell exhaustion and the molecular composition of immune exclusion barriers.

TLS

TLSs are ectopic lymphoid aggregates that form in non-lymphoid tissues under pathological conditions such as tumors, and they are primarily composed of B cells, T cells and dendritic cells (83). The application of ST enables precise dissection of the spatial localization, internal cellular composition of TLSs and their interactions with the surrounding microenvironment, thus revealing for the first time the mechanisms by which TLSs exert a dual role in antitumor immunity.

The antitumor function of TLSs is closely associated with their maturation status and spatial location. A recent pan-cancer spatial atlas analysis of TLSs across 12 cancer types reveals that TLS maturation is accompanied by B cell differentiation, plasma cell dissemination, Tfh cell enrichment and enhanced interferon signaling. Moreover, a distance-dependent tumor signaling gradient exists around intratumoral TLSs: Tumor regions proximal to TLSs are enriched in immune activation pathways, whereas regions distal to TLSs are enriched in proliferation and invasion pathways (84). In prostate cancer, a previous study (85) clearly showed that intratumoral TLSs are not only more abundant but also more prone to form mature structures with germinal centers, and their presence is an independent protective factor for disease-free survival. By contrast, peritumoral TLSs are predominantly immature aggregates with loose architecture. In hepatocellular carcinoma (HCC), high-resolution ST combined with pseudotime analysis enables precise classification of TLSs into mature, maturation-prone and functionally deviated subtypes, among which mature and maturation-prone TLSs are notably associated with favorable prognosis and positive response to immune checkpoint blockade (ICB) (86). In renal cell carcinoma (RCC), TLSs are found to serve as sites where B cells undergo clonal diversification, selection and expansion, and the resulting plasma cells secrete IgG antibodies that directly induce tumor cell apoptosis. The B cell maturation trajectories within TLSs revealed by ST suggest that strategies aimed at inducing TLS formation or enhancing TLS function may improve ICB efficacy (87). In head and neck squamous cell carcinoma (HNSCC), mature TLSs are also associated with superior response to ICB therapy (87). Collectively, these cross-cancer studies consistently demonstrate that the antitumor function of TLSs is broadly conserved, and targeting TLS maturation pathways may represent a universal strategy for potentiating antitumor immunity.

Specific stromal cell subsets can actively promote TLS formation. In nasopharyngeal carcinoma, a population of CXCL13+ CAFs has been shown to actively recruit B cells and promote their activation and antibody production by secreting CXCL13 and providing TNFSF13B-mediated maturation signals (88). In colorectal cancer (CRC) liver metastases, CCL19+ fibroblasts are identified as key players that facilitate TLS formation and enhance antitumor IgG responses (89). These findings suggest that the TLS-promoting function of CAFs may be broadly conserved across cancer types, and that targeting CAFs to reconstitute a TLS-supportive microenvironment could emerge as a novel adjunct to immunotherapy.

The immunosuppressive functions of TLSs exhibit pronounced cancer-type specificity and can be driven by distinct metabolic or transcriptional programs. In high-grade serous ovarian cancer (HGSOC), Xu et al (90) explicitly associates 'dysmorphic TLS-like aggregates' with early patient relapse. Additionally, structurally mature TLSs (mTLSs) are associated with poor prognosis in HGSOC. Mechanistically, the study demonstrates that the presence of mTLSs associates with a reduction in follicular helper T cells within the TME, which biases CD8+ T cell differentiation toward a terminally exhausted (TIM-3+PD-1+) phenotype rather than a stem-like exhausted (TCF-1+PD-1+) phenotype (87). This finding cautions that simply increasing TLS abundance in HGSOC may be insufficient to improve prognosis and could even be detrimental; combined strategies that target Tfh cells or employ anti-TIM-3 antibodies may be required to reverse terminal exhaustion. Similar phenomena have not been reported in HCC or RCC, suggesting that the immunosuppressive roles of TLSs may be highly dependent on the unique immune microenvironmental context of ovarian cancer. In a study of HCC (86), ST identified a subset of 'functionally deviated TLSs' in which intratumoral B cells display aberrant activation of tryptophan metabolism (high IDO1 and TDO2 expression), thereby creating a localized immunosuppressive milieu within the TLS. In clear cell RCC (ccRCC), the mechanism underlying TLS dysfunction is the aberrant enrichment of the transcription factor IRF4, which is associated with an immature TLS phenotype (91). Collectively, these findings indicate that although functionally deviated TLSs may exist across multiple tumor types, the underlying molecular mechanisms remain context-dependent.

ST has elevated TLS from a static, binary biomarker to a complex immune entity that must be functionally evaluated across multiple dimensions, structural, spatial, metabolic, transcriptional and cellular composition, thereby guiding the design of targeted intervention strategies.

CAF associated niches

CAFs represent a key stromal cell population within the TME characterized by pronounced heterogeneity. The rise of ST makes it possible, for the first time, to redefine their functional subtypes and identify targetable CAF subpopulations based on their spatial localization, molecular signatures and interactions with neighboring cells in native tissue contexts.

The spatial compartmentalization of CAF functional subtypes appears to be a conserved feature across cancer types. In HCC, VEGFA+ CAFs are enriched in the tumor core, where they promote immunosuppressive angiogenesis through spatially adjacent vascular endothelial cells (92). In pancreatic ductal adenocarcinoma (PDAC), myofibroblastic CAFs (myCAFs) are closely juxtaposed with tumor cells and directly sequester immune cells, whereas iCAFs reside farther from the tumor center and shape an immunosuppressive environment (93). In oral squamous cell carcinoma, epithelial cells in highly metabolic tumor regions secrete lactate, which induces iCAF conversion and subsequently activates the HIF1α-CXCL12 axis to recruit regulatory T cells (Tregs) (94). In HNSCC, MHC-Ihigh Galectin-9+ CAFs construct a spatially defined 'barrier zone' that excludes and exhausts CD8+ T cells through the upregulation of chemokines and Galectin-9 (95). Although these studies span different cancer types, ST analyses collectively reveal a general principle: the spatial distribution of CAF subtypes is not random; rather, they form specific spatial organizational patterns with the invasive front, perivascular regions or immune cell-enriched areas.

The upstream signals and functional outputs that drive CAF spatial heterogeneity vary considerably across cancer types. In glioblastoma (GBM), Jain et al (96) integrate scRNA-seq and ST to identify a CAF subpopulation highly expressing PDGF and TGF-β, whose primary function is to enhance immunosuppression by promoting the enrichment of CSCs and M2 polarization of macrophages. This mechanism is particularly prominent in GBM, but whether it plays an equally key role in other solid tumors remains unclear. In melanoma, GDF15+ CAFs activate the GFRAL/RET axis on tumor cells through paracrine signaling, thereby inducing macrophage M2 polarization (97); in cervical cancer, a COL14A1+VIM+ CAF subpopulation interacts with tumor cells via the MDK-SDC1 pathway and is notably associated with T cell exclusion (4). These distinct ligand-receptor pairs, uncovered through ST combined with single-cell sequencing and other approaches, indicate that the molecular pathways by which CAFs mediate immune escape are highly context-dependent.

Through ST, the association between CAF spatial distribution and immunotherapy response is reported in multiple cancer types, yet its generalizability as a biomarker remains to be validated. In non-small cell lung cancer (NSCLC), patients who respond to neoadjuvant immunotherapy exhibit spatially organized units at the tumor margin comprising antigen-presenting CAFs, SELENOP+ macrophages and CD4+ T cells that promote immune activation (98). In HGSOC, the spatial organization pattern of tumor-associated CAFs with tertiary lymphoid structure-like aggregates is associated with early relapse (90). In cervical squamous cell carcinoma, POSTN expression, a functional indicator of myCAFs, is markedly lower in the drug-sensitive group than in the drug-resistant group (99). These findings suggest that CAF spatial configurations may serve as predictive biomarkers; however, the specific marker combinations vary across cancer types and require independent validation in each.

ST leads to the realization that CAFs, functioning as key orchestrators within the TME, display extensive spatial heterogeneity and functional plasticity. However, the exact molecular mechanisms and biomarkers are strongly dependent on the cancer type, which necessitates independent validation and the selection of tailored CAF-targeted strategies for each cancer context.

T cell exhaustion niches

T cell exhaustion represents one of the core mechanisms of immune evasion in locally advanced tumors. The application of ST enables the localization of exhausted T cell aggregates and the identification of their spatially proximal relationships with specific tumor cell subpopulations or stromal cells.

Spatial associations between exhausted T cells and specific myeloid or malignant cell subpopulations are reported across multiple tumor types. In primary central nervous system lymphoma, T cell exhaustion markers are spatially associated with a malignant B cell subpopulation that highly expresses multiple chemokines and immune checkpoint ligands (100). In HCC, regions enriched with exhausted T cells concurrently exhibit high expression of the M2 macrophage marker CD163 and the CAF marker VIM (101). In ccRCC, exhausted T cells are preferentially located within the tumor interior, and the uneven distribution of the same T cell clones across different tumor regions suggests that the local microenvironment shapes T cell status and expansion (102). In breast cancer, exhausted T cells predominantly aggregate in regions with immature TLS and are associated with the spatial distribution of CXCL13+ T cells (103). These cross-cancer ST studies collectively support a general conclusion: T cell exhaustion is not a cell-autonomous process but is driven by the spatial architecture of the local microenvironment.

Studies find that the key stromal cell types driving T cell exhaustion may differ across cancer types. Xun et al (104) performed ST analyses across multiple cancer types (breast cancer, CRC, ovarian cancer and ccRCC) and found that exhausted CD8+ T cells and ALCAMhigh macrophages are highly enriched at the tumor boundary, with co-localized Tex cells exhibiting higher exhaustion scores. This phenomenon is observed across multiple types of cancer, suggesting that ALCAMhigh macrophages may represent a cross-cancer driver of exhaustion. However, in a study of ovarian cancer chemotherapy (105), T cell exhaustion was shown to be driven primarily by chemotherapy-induced myeloid cell networks (Myelonets) rather than by pre-existing macrophage subpopulations, indicating that the driving mechanisms of T cell exhaustion can be reshaped in the context of treatment. In GBM, the spatial distribution of T cell exhaustion is also closely associated with hypoxia signaling (106); this feature may be particularly prominent in the highly hypoxic GBM microenvironment, while it may not dominate in other tumors with improved oxygenation.

These ST-based findings suggest that reversing T cell exhaustion represents a key strategy in immunotherapy; however, the selection of spatial targets must be tailored to the cancer type and the treatment context. In ovarian cancer undergoing chemotherapy, ST reveals that chemotherapy-induced myeloid cell networks form an immunosuppressive microenvironment that drives functional exhaustion of CD8+ T cells, thereby promoting chemoresistance (105). This finding indicates that co-targeting myeloid cells may be an effective strategy to reverse exhaustion in ovarian cancer, whereas in other cancer types, targeting CAFs or specific macrophage subpopulations may be required. ST enables the identification of spatial drivers of T cell exhaustion across various cancer types and treatment settings, thus providing a logical basis for precision combination therapy.

Immune exclusion niches

Among the most direct contributions of ST is the elucidation and definition of 'immune exclusion niches', the phenomenon by which cytotoxic T cells are actively excluded from the tumor core.

CAF subpopulations enriched in extracellular matrix components (ECM-CAFs) mediate immune exclusion is conserved across multiple cancer types. A pan-cancer study integrating 16 cancer types reveals that ECM-CAFs are specifically enriched at the tumor boundary. These ECM-CAFs not only form a physical barrier that impedes T cell infiltration but also directly induce T cell exhaustion by highly expressing LGALS9, TIM-3 and other immune checkpoint ligands. In patients with HCC who are non-responsive to immunotherapy, the enrichment of ECM-CAFs is notably associated with a spatial distribution pattern in which T cells are confined to the tumor periphery, thereby providing solid clinical evidence for the concept of a 'spatial immune barrier' (107). Similar immune exclusion patterns are also observed in triple-negative breast cancer and high-grade serous ovarian cancer (108,109). This cross-cancer and cross-platform evidence suggests that ECM-CAF-mediated immune exclusion represents a prevalent immune evasion mechanism.

The specific molecular features and spatial organization patterns of immune exclusion are modulated by local microenvironmental factors. In GBM, integrated ST and spatial proteomic analyses reveal that mesenchymal-like tumor cell states are tightly organized with macrophages and T cells into specific multicellular units, and this spatial organization is governed predominantly by hypoxic signaling (106). In GBM, hypoxia is a key driver of immune exclusion; however, in well-oxygenated tumors such as renal cancer, hypoxia may not be a major contributing factor. Furthermore, in triple-negative breast cancer, immune exclusion is associated with the spatially restricted activation of the TGF-β signaling pathway (109), while in ovarian cancer, immune exclusion is associated with specific CAF subpopulations (108). These differences indicate that, while immune exclusion is a prevalent phenomenon, its upstream driving signals and specific spatial organization patterns are highly context-dependent.

ST is what converts immune exclusion from a descriptive concept into a targetable entity, jointly characterized by specific cell subsets, particular molecular pathways and defined spatial locations.

Distant metastasis stage

Distant metastasis is one of the leading causes of cancer mortality. Its occurrence depends on the ability of tumor cells, after detaching from the primary site, to successfully adapt to and remodel the microenvironment of distant organs (81,110,111). The application of ST enables the direct in situ investigation of two key steps in the metastatic process: The pre-conditioning of the pre-metastatic niche and the dynamic remodeling of the tumor invasion front (5,75). ST allows researchers to visualize, across different cancer types and metastatic organs, the microenvironmental differences between primary and metastatic lesions, the spatial propagation characteristics of immunosuppressive cellular neighborhoods and the spatial coupling mechanisms of epithelial-mesenchymal transition (EMT) and immune escape at the invasion front (94,112-114), thereby elevating metastasis research from a 'molecular inventory' to the level of a 'spatial evolutionary atlas'.

Pre-metastatic niche

In recent years, advances in ST have opened new avenues for understanding tumor evolution and metastatic mechanisms from a spatial dimension. In CRC liver metastasis, ST reveals that CCL19+ fibroblasts form specific spatial structures in the liver that promote TLS formation and antitumor IgG responses, and the absence of this niche is associated with metastatic progression (89). In PDAC, ST comparison of primary and metastatic lesions reveals marked alterations in the spatial distribution patterns of iCAFs and myCAFs at metastatic sites relative to primary tumors (115). Spatial proteomic analysis of lymph node metastases further demonstrates that, although lymph node metastases retain lymphoid structures composed of B cells, CD4+ T cells and other immune cells, they exhibit pronounced abnormalities in functional markers: T cells highly express PD-1 and LAG-3, Tregs display an activated state with elevated Ki67, ICOS and IL-10, and myeloid cells and CAFs show upregulated PD-L1 and IL-10 expression. Importantly, spatial neighborhood analysis identifies a 'myeloid/CAF-enriched cellular neighborhood' that is enriched not only within metastases but can also be traced into adjacent uninvolved tumor regions, suggesting that immunosuppression can propagate spatially (116). In a multi-organ metastasis study of PDAC, integrated ST and single-cell imaging reveals that metastases do not simply recapitulate primary tumor features; instead, they adapt to the local microenvironment through organ-specific clonal selection and lineage remodeling. Liver metastases manifest an immune exclusion niche dominated by a basal-like phenotype, whereas lung metastases display a classical phenotype with a relatively active immune microenvironment, highlighting the substantial transcriptomic heterogeneity and dynamic microenvironmental changes during pancreatic cancer metastasis (93). A study using high-resolution single-cell ST (CosMx SMI) on SCLC systematically constructs a spatial cell atlas of primary tumors and lymph node metastases, revealing from a spatial dimension the heterogeneous evolution of the TME and the remodeling of cellular interactions during SCLC lymph node metastasis (117).

These studies collectively demonstrate that ST enables visualization of the microenvironmental differences between primary and metastatic sites across different cancer types and metastatic organs, as well as the spatial propagation characteristics of immunosuppressive cellular neighborhoods. However, it remains unclear whether these findings are generalizable to other types of pre-metastatic niches. The molecular features of pre-metastatic niches are highly dependent on the primary tumor type and the tissue context of the metastatic target organ. Therefore, further cross-cancer and cross-metastatic organ comparative studies are required to extract universally instructive principles of metastatic evolution from this context dependency. Such cross-organ and cross-cancer comparisons are made possible by ST, which in turn provides spatial targets for precision intervention against the metastatic microenvironment.

Tumor invasion front

The tumor invasion front is a dynamic zone at the interface between the primary tumor and the surrounding stroma, representing the most active site of EMT, tumor cell invasion and the establishment of an immunosuppressive microenvironment.

An EMT signature is enriched at the tumor-stroma boundary across multiple cancer types rather than being uniformly distributed throughout the tumor. In HCC, high-resolution Stereo-seq technology precisely delineates an invasion zone ~500 μm wide at the tumor boundary, within which tumor cells exhibit pronounced EMT signatures, metabolic reprogramming and molecular hallmarks of immune escape (118). In gastric cancer, spatial atlases reveal that the tumor margin (G2-enriched region) displays higher activity of EMT and angiogenesis pathways compared with the tumor core (G1-enriched region) (119). In CRC, the deep learning algorithm GASTON identifies type II genes that predominantly describe the upregulation of EMT genes outside the tumor boundary (120). Collectively, these independent studies across diverse cancer types consistently demonstrate that the tumor-stroma interface serves as a spatial hub for EMT, a broadly conserved principle.

The specific stromal cell subsets and signaling pathways that drive EMT vary across cancer types. In CRC, studies reveal that distinct CAF subpopulations spatially promote EMT through different signaling axes. Yang et al (121) find that malignant epithelial cells in metastatic CRC can undergo EMT to transdifferentiate into CXCL1+ CAFs, a process driven by the transcription factor BHLHE40; Lu et al (122) identify a distinct population of CTHRC1+ fibroblasts that secrete WNT5A to form a spatially adjacent pro-EMT niche with malignant epithelial cell. In GBM, M2-like tumor-associated macrophages (TAMs) are predominantly distributed at the tumor invasion front and exhibit a significant spatial association with mesenchymal-like tumor cell states, implicating them in the construction of an immunosuppressive barrier at the tumor boundary (96). In acral melanoma, tumor cells with high EMT signatures are spatially co-enriched with APOE+/CD163+ TAMs and promote tumor invasion through IGF1-IGF1R signaling (123). In ccRCC, the EMT signatures in metabolically active regions are associated with sphingolipid metabolic reprogramming in M2 macrophages (124). These differences indicate that, although the spatial localization of EMT is universal, the specific modes of interaction with microenvironmental cells are highly context-dependent.

ST enables the redefinition of EMT, transforming it from a purely tumor-cell-intrinsic phenotypic program into an 'invasion front niche', a niche that is jointly orchestrated by a specific spatial location, specific stromal cell subsets and a specific metabolic microenvironment, and that functions simultaneously in invasion and immune exclusion. Fig. 1 illustrates the typical cellular composition and organization of five biologically important spatial niches within the TME across the course of tumor progression.

Figure 1 Typical cellular composition and organization of five biologically important spatial niches within the tumor microenvironment across the course of tumor progression. The figure summarizes the major spatially defined niches in the tumor immune microenvironment across premalignant, locally advanced and metastatic stages. In the locally advanced stage: TLSs exhibit dual functions; mTLSs promote antitumor immunity via B cell differentiation and plasma cell dissemination (for example, CXCL13+ CAFs recruit B cells and enhance antibody production), while functionally deviated TLSs (for example, those with IDO1/TDO2 upregulation in HCC) create immunosuppressive milieus. CAF-associated niches are spatially compartmentalized into distinct subtypes. The myCAFs localize adjacent to tumor cells, whereas iCAFs are positioned distantly; specific CAF subsets further express distinct ligand-receptor pairs. T cell exhaustion niches are characterized by spatial association of exhausted CD8+ T cells with M2-like TAMs. Immune exclusion niches are defined by ECM-CAFs that form physical barriers. In the distant metastasis stage: At the tumor invasion front, EMT is enriched and driven by distinct CAF subsets (for example, CXCL1+ CAFs via BHLHE40 or CTHRC1+ fibroblasts via WNT5A). Spatial transcriptomics enables the in situ resolution of these niches, transforming descriptive patterns into testable spatial hypotheses and potential therapeutic targets. TLS, tertiary lymphoid structures; mTLS, mature TLS; CAF, cancer-associated fibroblast; iCAF, inflammatory CAF; myCAF, myofibroblastic CAF; HCC, hepatocellular carcinoma; TAM, tumor-associated macrophage; ECM, extracellular matrix; EMT, epithelial-mesenchymal transition.

Toward clinical application: Limitations and break-throughs in spatial transcriptomics technology

Conventional clinical diagnosis and treatment rely on imaging and histopathology as adjunct tools, yet often overlook the spatial cellular phenotypes underlying histopathological findings. The finer-resolution molecular insights offered by ST enable more precise disease stratification, outcome prediction and assessment of therapeutic response. Notable progress has been made in oncology, where spatial molecular landscapes have been shown to associate with tumor growth, susceptibility to metastasis and efficacy of immunotherapy. These discoveries highlight the clinical potential of spatial biomarkers, providing actionable insights for precision medicine. However, before ST can be effectively integrated into routine clinical diagnosis and decision-making, several key technical challenges must be addressed, including limited resolution, sample constraints, algorithmic bias and lack of cross-platform standardization. There is an urgent need to establish widely accepted standardized operating procedures covering sample collection, platform testing and data analysis. The rapid development and maturation of machine learning in recent years have created new opportunities to address these technical bottlenecks (11). Looking ahead, as omics technologies and machine learning become increasingly integrated, advances in multimodal data integration and three-dimensional tissue reconstruction will provide even more powerful technical support for deepening the understanding of tumor evolution and overcoming challenges in cancer therapy.

Clinical translational potential of spatial transcriptomics technology

The breakthroughs of ST in revealing cellular heterogeneity and decoding cell-cell interactions are driving the translation of related research toward clinical applications. In this context, the present review focuses on two key translational paths: First, the identification of spatially defined biomarkers for use in clinical diagnosis or subtyping; and second, the elucidation of mechanisms underlying therapeutic response to guide the discovery of new therapeutic targets (see Fig. 2).

Figure 2 Spatial transcriptomics play multiple roles in advancing the precision and personalization of tumor immunotherapy. For patients resistant to immunotherapy, these technologies enable the exploration of tumor heterogeneity and the immunosuppressive microenvironment, identify distinct gene expression features and facilitate the efficient development of novel drug targets. They also aid in the development and validation of tumor biomarkers, guiding personalized treatment strategies. Furthermore, by integrating multi-omics analytical platforms and machine learning techniques, spatial transcriptomics are driving a shift in biomarker screening from single-marker development to the establishment of multi-parameter predictive models. Simultaneously, they contribute to the construction of more comprehensive tumor atlases, reshaping the scientific understanding of tumor biology.

First, ST technology, with its unique spatial resolution capability, holds a notable advantage in discovering novel spatially defined candidate biomarkers. From an ST perspective, the TME can be partitioned into multiple spatial domains, each characterized by consistent functional states or morphological components. Molecular expression profiles differ substantially across these domains, and CCC within the same domain are mediated by key L-R pairs. Biomarkers associated with tumor prognosis are often derived from differentially expressed molecules across distinct regions of the TME, reflecting divergent clinical outcomes. Chauvin et al (125) used the Xenium platform to perform ST analysis on tissue samples from human HGSOC. They found that, compared with tumor cell regions, stromal regions exhibited elevated FST and FSTL3 signaling in cancer-associated mesothelial cells and CAFs. Tumor models overexpressing FSTL3 displayed a more fibrotic microenvironment with reduced immune cell infiltration (125). Moreover, FSTL3 similarly influences tumor infiltration in CRC, gastric cancer and LUAD, underscoring its potential as a prognostic biomarker across multiple cancer types (126-128). Expanding biomarker screening beyond specific tumor types, the combination of single-cell RNA-seq and ST enables the resolution of single-cell gene expression within complex, highly structured tissues. It is well established that fibroblasts and macrophages are ubiquitous in the TME and engage in intensive CCC. The transferase SLFN11 is significantly downregulated in both cell types across most tumors, and multi-omics analyses suggest its association with DNA repair, the P53 pathway and tumor immunity. SLFN11 expression level thus holds great promise as a diagnostic biomarker and a predictive indicator of treatment efficacy (129). Furthermore, identified biomarker molecules can also guide the selection of clinical therapeutic agents in oncology. Wang et al (130) through ST sequencing and bulk RNA-seq analysis, provided the first evidence of the prognostic significance of an ADME-related prognostic signature in ccRCC and its potential application in guiding personalized treatment strategies. In summary, the integration of ST technology on one hand provides direction for biomarker development, screening for differentially expressed molecules within specific spatial domains may improve discovery efficiency, and on the other hand, the wealth of omics data facilitates the combined development of multi-marker panels, further enhancing the predictive and discriminative power of biomarkers, supporting the construction of tumor prediction models and aligning with the advancement of precision medicine.

Second, ST can resolve the spatial context of disease initiation and progression, an unique perspective that facilitates the discovery of previously overlooked, spatially dependent novel therapeutic targets. Tumor development is characterized by multiple stages and steps, and intratumoral heterogeneity (ITH) leads to regional differences in immune status and drug sensitivity. Conventional bulk sequencing masks this spatial complexity, whereas ST preserves tissue architecture while revealing tumor heterogeneity patterns, positioning it as a key technology to address this challenge. Zhang et al (131) used ST analysis to show that in regions of HCC responsive to cabozantinib and nivolumab (CABO/NIVO) combination therapy, PAX5 is localized to immune areas adjacent to tumor clusters, indicating a B cell-driven active immune state. By contrast, within tumor-CAF interaction zones, ECM remodeling genes are activated, forming a drug-resistant niche with sparse immune infiltration. Qin et al (132) identified four tumor cell subpopulations in pancreatic cancer and found that crosstalk between RPS4Y1+ tumor cells, fibroblasts, and CD8+ T cells promotes immunosuppression. Song et al (133) demonstrated in non-SCLC that PD-L1 expression in the stromal region has stronger predictive power for bispecific antibody response, and that integrating DSP with multiplex immunofluorescence improved research efficiency. ST-based dissection of ITH can guide the development of prognostic indicators and novel therapeutic targets. Mao et al (134) found that regions with low MC content colocalize with CAF genes and exhibit minimal immune infiltration, and that MC heterogeneity promotes immune evasion in HCC, with relevant gene signatures serving as prognostic markers. Guo et al (135) observed in MMRd endometrial cancer that immune features in tumor-enriched regions align with CD8+ T cell infiltration and identified potential biomarkers including HLA class I and DNMT3A. ST technology provides direct visualization of cellular distribution and communication across distinct niches, offering tissue- and molecular-level evidence for understanding immunotherapy sensitivity and resistance, and holds promise for identifying new therapeutic targets.

The highly immunosuppressive TME is a major contributor to drug resistance in solid tumors. In recent years, researchers have increasingly combined ST with technologies such as CODEX and single-cell sequencing to dynamically monitor immune cell infiltration and changes in gene expression. In the context of radiotherapy, which can promote antitumor immunity, Oyoshi et al (136) found that following radiotherapy, the infiltration of immune cells including T cells and B cells in the TME increased, with distinct expression patterns observed at different stages. In muscle-invasive bladder cancer, patients who achieved pathological complete response after neoadjuvant chemotherapy exhibited lower neutrophil counts and lower neutrophil-to-lymphocyte ratios, along with increased infiltration of B and T cells (137). Current research focuses on the contribution of specific immune cell subtypes to drug resistance, including macrophages, Tregs, CAFs and plasma cells (138). In GBM, the roles of distinct macrophage subtypes have received considerable attention. Liu et al (139) demonstrated that VSIG4+S100A10+ TAMs colocalize with T cells and suppress antitumor immunity. Gao et al (140) identified a novel MDM subtype with high TREM-1 expression in perinecrotic hypoxic regions and showed that targeting this subtype enhances chemosensitivity. In recurrent GBM, oligodendrocyte enrichment and microglial activation have been observed (141). In low-grade glioma, FN1+ TAMs are enriched in hypoxic regions and associated with recurrence (142). Similar phenomena have also been reported in CRC and metabolic dysfunction-associated steatohepatitis-related hepatocellular carcinoma, where such macrophages are typically localized to hypoxic niches and participate in immunosuppression. In HCC research, distinct CAF subtypes exert opposing effects: A Treg-CAF cluster at the tumor margin forms an immunosuppressive niche considered to drive immunotherapy resistance in steatohepatitic HCC (143), whereas the FMO2+ CAF subset is associated with improved prognosis following anti-PD-1 therapy (144). Collectively, ST compensates for the inability of conventional transcriptomics to resolve cellular spatial localization, providing a powerful tool for elucidating the mechanisms by which stromal cell subsets mediate immune evasion.

Key technical challenges constraining clinical application

Since the advent of ST technology, continuous iterative optimization to meet research and other application needs, along with advances in bioinformatics capabilities for spatial data analysis, has revealed complex intercellular spatial relationships within the TME. These insights are key for understanding tumor progression and treatment resistance, and demonstrate strong translational potential in the field of oncologic clinical therapy. However, before ST can be truly implemented in clinical practice, connecting spatial biomarkers to patient treatment outcomes and demonstrating their validity across different therapeutic settings will require overcoming a series of obstacles. The present review discusses key considerations that constrain the use of ST in clinical diagnosis and treatment, encompassing technical algorithms, sample selection and data validity validation.

Trade-off between resolution and coverage

The fundamental principles underlying sequencing-based and imaging-based methods inherently limit the resolution and coverage of sample detection. Next generation sequencing (NGS)-based ST platforms, represented by 10x Genomics Visium, rely on fixed spots with spatial coordinates to capture mRNA released from all cells within the corresponding region. Due to the physical size of individual spots, typically 10 to 100 mm, the data captured by each spot represents mixed transcriptional signals from multiple cells, precluding precise assignment to single cells and thus yielding lower resolution. By contrast, imaging-based ST platforms, represented by Xenium, CosMx and CODEX, are designed to acquire in situ data at single-molecule resolution, enabling RNA localization at the subcellular level (down to 100-200 nm) and retaining more intracellular information. However, these methods are limited to fluorescence detection of predefined target genes, and detection efficiency decreases as panel size increases.

In clinical research, specific diagnostic and therapeutic needs place different emphasis on resolution vs. coverage, necessitating a deliberate trade-off. Sequencing-based methods enable unbiased capture of polyadenylated transcripts, allowing genome-wide investigation of gene expression. They are suitable for gaining a global understanding of tissue architecture, identifying novel cell subtypes, and revealing overall expression differences across distinct functional regions of tumor tissue. Additionally, the sequence information obtained through NGS can be used to detect somatic variants, transcript isoforms, non-coding genes and TCR/BCR clonotypes. Imaging-based methods, constrained by reliance on pre-designed probes, limit the number of target genes that can be measured on a single panel and cannot retrieve sequence information. However, subcellular-resolution platforms offer higher resolution, shorter turnaround times and greater interpretability of results, making them particularly well-suited for hypothesis validation and clinical testing. The high-resolution microscopy required for multi-round fluorescence capture renders imaging-based ST technologies more expensive than sequencing-based approaches, a significant factor limiting their clinical adoption. With ongoing technological iterations, sequencing-based methods are progressively moving toward higher resolution. The Stereo-seq spatial omics platform from BGI achieves nanoscale resolution through chip-arrayed DNBs. Future ST platforms are expected to combine both high resolution and broad coverage, making them more suitable for clinical applications.

Sample limitations

FFPE samples are a common method of tissue processing in clinical practice. Formaldehyde fixation preserves tissue and cellular morphology, and after dehydration and paraffin embedding, the tissue can be sectioned into slices of 5 μm or thinner for histologic and immunopathologic assessments required for diagnosis, while also allowing sample storage under non-frozen conditions. In medical practice worldwide, standardized protocols for the collection, preservation, sectioning and storage of FFPE samples have been systematically established. Due to the routine need for solid tumor diagnosis in clinical settings, hospitals have accumulated large repositories of such specimens, often accompanied by detailed clinical annotations, providing valuable resources for investigating tumor development and therapy (5,145). The ability of ST technologies to obtain accurate spatial expression information depends critically on the preservation of tissue morphology and RNA integrity. Because RNA is unstable at room temperature and highly prone to degradation, ST analysis requires sample fixation. Compared with other sample types, FFPE samples offer distinct advantages, including well-preserved spatial context and cellular architecture with minimal distortion, making them the current sample of choice for ST analysis. However, it is important to note that during the fixation process, formaldehyde crosslinks free molecular chains, which simultaneously leads to nucleic acid fragmentation and protein-nucleic acid crosslinking, ultimately resulting in samples that cannot provide high-quality RNA molecules for ST analysis (146). This represents a major factor limiting the application of FFPE samples in clinical ST analysis. By contrast, fresh frozen (FF) samples allow the acquisition of intact RNA molecules, but their preparation, transport, storage and processing require stringent conditions, including maintaining subzero temperatures to prevent RNA degradation (147,148). Moreover, clinical practice lacks standardized operating procedures to ensure consistency and quality control of FF samples. To improve RNA quality from such samples, Wulf et al (149) developed a novel miRNA sequencing strategy called CapTS-seq, which combines chemical capping and template switching to enhance miRNA library quality; however, this study did not address validation of detection efficiency related to ST technologies (149). Therefore, obtaining FFPE samples with high-quality RNA remains a challenge that must be addressed when advancing ST technologies for clinical tumor diagnosis and therapy.

Bioinformatics algorithmic bias

Regardless of whether spot-based or imaging-based platforms are used, the transcriptomic landscape that contains spatial information is often not directly visualized. This inevitably requires computational methods to model and annotate spatial data, indirectly reconstructing the complex distribution and expression states of tumor cells, immune cells and stromal cells within the TME under in vivo conditions (150-152). These computational approaches, such as cell segmentation and deconvolution, are essentially fitting procedures that approximate true spatial adjacency relationships rather than direct reflections of the real world. Consequently, algorithmic bias can readily lead to ST results with low reliability, thereby affecting downstream analyses.

Segmentation errors

RNA hybridization-based ST captures molecular and cellular features at subcellular resolution, yielding more accurate transcriptional mapping and cell phenotyping compared with sequencing-based platforms. This, however, places higher demands on the accuracy of segmenting image volumes into individual cells. Existing cell segmentation methods rely on structural markers (image staining-based segmentation) and transcript clustering to delineate individual cells. Approaches such as StarDist, CosMx and Xenium achieve segmentation through specific staining of cellular components (StarDist uses nuclear staining and expansion to generate pseudo-cell boundaries) (153). Segger models the spatial relationship between nuclei and surrounding transcripts to refine initial boundaries (154). These methods are limited by manual annotation, staining quality, sensitivity to cell density and mismatch with true transcriptomic profiles, making it still challenging to correctly assign individual molecules to their cells of origin. On one hand, tumor cells, immune cells and stromal cells within the TME exhibit distinct morphologies and highly variable spatial compositions, particularly in regions with high cell density, irregular morphology or tightly packed tissues such as the epithelial layer and tumor parenchyma, where under-segmentation or over-segmentation frequently occurs. On the other hand, parameter settings in segmentation algorithms also constrain performance; for example, the optimal parameters of DBSCAN are closely related to target cell size and transcript density (155). Erroneous assignment of RNA molecules from neighboring cells to the target cell leads to data contamination, which has been shown across multiple tissues and platforms to affect downstream analyses such as false positives in differential expression and misinterpretation of cell-cell interactions. These effects often dominate analytical outcomes (156,157). To further improve the accuracy of segmentation-based computational methods, functional annotation of transcriptomes can be leveraged. JSTA incorporates prior knowledge of cell type-specific gene expression to achieve cell segmentation, improving RNA assignment accuracy by >45% (158). RedeFISH aligns segmented cell expression profiles with single-cell expression data to achieve optimal segmentation. cellAdmix effectively identifies and separates confounding data through matrix factorization of local molecular neighborhoods, thereby reducing its impact on downstream analyses (156). Furthermore, manual annotation and review remain necessary at this stage to address segmentation errors caused by tumor cell atypia (155).

Deconvolution uncertainty

Sequencing-based platforms provide spatially resolved gene expression data by sequencing RNA molecules captured within spots, but typically lack cellular resolution. Therefore, cell type deconvolution is required to infer the cellular composition of each spot (159,160). Current computational methods fall into two major categories: Supervised deconvolution and unsupervised deconvolution. Supervised deconvolution methods, represented by RCTD, SPOTlight, cell2Location, Tangram, CytoSpace and Spotiphy, utilize single-cell RNA-seq reference data and employ strategies such as regression models, Bayesian models and optimal transport mapping to infer the compositional proportions, counts or even single-cell localization of different cell types from the mixed expression signals of each spot. For such methods, the degree of alignment between the single-cell RNA-seq reference dataset and the ST data significantly affects deconvolution results. If a particular cell type is absent from the reference data, the deconvolution model will erroneously assign expression signals from the missing type to other related cell types, which is a major source of result bias. Furthermore, deconvolution outputs are typically cell type proportions, which are compositional data where the components sum to a constant (for example, 100%). This violates the independence assumption of conventional statistical methods such as linear regression, leading to spurious associations. Some Bayesian models address this by setting probabilistic priors for variables such as cell type proportions and gene signatures from reference data, updating these considerations using spatial data. The resulting posterior distributions provide interval estimates rather than point estimates for parameters, thereby effectively quantifying uncertainty. Unsupervised deconvolution methods, represented by CARD and SpiceMix, do not rely on external reference data. Instead, they automatically discover latent 'factors' from the statistical structure of the ST data alone using dimensionality reduction techniques such as matrix factorization, with the factors then manually annotated as cell types. In this process, the number of factors and their annotation depend on the researcher's subjective judgment. Matrix factorization algorithms are sensitive to initial values and data noise and different runs may yield different results. The resulting 'factors' may represent genuine biological cell types or may be technical noise subject to overinterpretation, representing a biological hypothesis that requires validation (101,161).

Batch effects

The clinical application of ST technology in oncology still requires overcoming challenges posed by batch effects. Batch effects arise from multiple factors including sample collection, processing conditions, sequencing platforms and variations in section geometry. If not properly adjusted, they may obscure true biological variation, leading to inaccurate downstream analyses. This issue is particularly pronounced in clinical data involving multiple patients, multiple institutions and diverse treatment histories. As a precursor to ST data, single-cell RNA-seq also suffers from batch effects and data variability. Correction strategies developed to date, such as Harmony, ComBat, fastMNN, Scanorama and Seurat-CCA, are based on shared embedding space analysis and neural network decoding, reducing noise and batch-specific artifacts while maximizing retention of true biological variation. However, directly applying batch correction tools designed for single-cell RNA-seq to ST data ignores spatial location information of cells, resulting in suboptimal performance (162). Although existing algorithms have begun to incorporate spatial features of ST data during batch effect removal, the complexity and non-specificity of spatial characteristics mean that no method has yet explicitly addressed this problem by integrating spatial information from gene distributions with gene expression features during batch effect correction (163,164).

Overinterpretation of L-R interactions

Deciphering CCC between adjacent cells is fundamental to understanding the coordination of cellular community behavior, tissue architecture and tumor-immune interactions within the TME. L-R interactions represent a key aspect of CCC research, as receptor activation can trigger intracellular signaling cascades that subsequently regulate cellular behavior. Advances in high-resolution sequencing technologies have greatly propelled the study of CCC using ST data. Nearly 50 L-R databases have emerged in this field, and these resources differ substantially in annotation quantity, species coverage, and signaling types. The choice of database largely influences analytical outcomes (165). The current mainstream strategy uses high-quality L-R databases as prior knowledge to align ST expression profiles with these databases for inferring potential cell-cell communication events. Tools such as Giotto, Squidpy, Spateo and CytoSignal provide modules that integrate L-R expression with spatial information to infer and quantify L-R communication scores (166-169), while COMMOT employs optimal transport theory to model cost-effective signaling molecule flow among spatially adjacent cell types (170). However, computational inference based on database alignment cannot be equated with actual biological processes. Even when ligand-expressing cells and receptor-expressing cells are spatially adjacent, this only indicates the possibility of CCC rather than established functional signal transmission, due to potential physical barriers or high RNA expression but low protein abundance (171). Furthermore, a strong L-R colocalization signal at a single time point cannot distinguish whether the signaling event has yet to initiate or has already terminated. In computational inference, these distinct scenarios are often uniformly classified as 'communication present', leading to overinterpretation of L-R interactions (171). Bridging the gap between computational inference and actual biological processes ultimately depends on rigorous experimental validation.

Platform standardization and the translational gap

The rapid development of ST technologies has given rise to multiple detection platforms. However, technical heterogeneity across platforms poses serious challenges for data integration and comparison. Addressing cross-platform standardization is particularly key for advancing ST toward clinical diagnosis and precision medicine applications.

In recent years, several teams worldwide have attempted to establish standardized evaluation frameworks for ST platforms and have made considerable progress. You et al (172) systematically compared 11 sequencing-based platforms and, through the development of the cadasSTre dataset, identified molecular diffusion as a key parameter affecting effective resolution across platforms. Plummer et al (173) developed SpatialQM software for imaging platforms to support quality assessment and annotation of cross-platform data. These two studies have laid the foundation for benchmark testing of platform performance required for clinical applications. Nevertheless, multiple rounds of large-scale benchmarking have consistently shown that the standardization process for ST platforms remains constrained by three core challenges. First, there is a lack of quantitative comparability. Platforms exhibit substantial differences in key parameters such as molecular capture efficiency, spatial resolution and downstream analytical capabilities, making direct comparison of gene expression levels across platforms difficult. Second, the reliability of spatial coordinates cannot be compared across platforms. Third, spatial barcode-based technologies suffer from spot-swapping issues, while imaging methods are limited by optical diffraction and image registration errors. Therefore, future standardization efforts urgently need to proceed along three fronts: Establishing cross-platform quantitative reference systems, promoting standardized operating procedures and conducting clinically value-driven multicenter cohort studies to evaluate and continuously calibrate platform outputs. Furthermore, as new technologies continue to emerge, standardized assessments must remain dynamically updated to enable researchers and clinicians to make informed decisions based on the latest evidence (174).

When conducting multicenter cohort studies, cross-cohort validation difficulty represents another core bottleneck for the clinical translation of ST. In addition to differences in platform technologies and batch effects, patient phenotypic heterogeneity and variations in clinical sample processing collectively make direct cross-cohort comparison of gene expression nearly impossible. Even under identical technical conditions, the TME and tissue characteristics across cohorts of patients with the same cancer type remain highly variable, preventing the validation of transferable spatial signatures (175). For testing centers that have not yet established standardized collection protocols for FF/FFPE samples, discrepancies in any step, collection, storage or sectioning, can amplify cross-cohort inconsistency by affecting RNA quality and spatial feature integrity. The breakthrough for achieving cross-cohort validation may lie in establishing standardized quality control metrics applicable to clinical samples, developing more interpretable and robust integration algorithms and, on this basis, conducting large-scale, multicenter prospective cohort studies (173).

Integrating machine learning: A key path to overcoming clinical translation barriers in spatial omics

As a key branch of artificial intelligence, machine learning identifies data patterns through mathematical algorithms to make predictions (176,177) and can efficiently process large datasets while automating decision-making. In recent years, machine learning has been increasingly applied to oncology research. Leveraging advances in omics technologies, multi-level tumor-related data can now be obtained. Large-scale data analysis, while illuminating the path to cancer therapy, requires substantial human effort. The introduction of machine learning offers a new solution for faster and more accurate assessment of immunotherapy efficacy and prognosis in patients with cancer. Trained on large-scale datasets, machine learning algorithms can help improve the accuracy of image segmentation and deconvolution in ST, as well as correct batch effects, positioning them as a powerful tool to overcome clinical translation barriers in ST technology. Furthermore, the deep integration of machine learning with spatial omics has given rise to multimodal data integration and 3D tissue reconstruction, facilitating the construction of human tumor molecular atlases and deepening the understanding of tumor progression.

The iterative evolution of machine learning tools

Machine learning methods are primarily divided into three categories: supervised learning, unsupervised learning and reinforcement learning. In supervised learning, each sample is labeled with an outcome or diagnosis by an expert; these labeled datasets are used to train models by identifying complex patterns within the labeled training data, which can then be applied to make decisions or predictions on new, unseen data. Applications include regression for predicting continuous variables, classification for identifying categories and image segmentation for automatic detection of relevant regions. Unsupervised learning does not rely on labels and aims to discover latent patterns within data, which can be used to identify drug interactions. Reinforcement learning is an autonomous system that learns through trial and error (178-180). Although most machine learning algorithms were developed as early as the 1950s, advances in big data and computational power have driven a profound transformation over the past two decades, evolving the field from task-specific tools to general-purpose, robust and highly transferable foundation models. At the core of this evolution are models such as Nicheformer (181) and Novae (182), which are pre-trained on massive datasets and capable of handling diverse tasks including cell annotation, spatial domain identification and batch correction. Concurrently, model architectures have evolved toward more refined feature learning. Graph contrastive learning approaches such as AugGCL (183) and SpaMGCL have enabled more robust handling of data noise and sparsity. At the application level, collaboration among models has become increasingly close; models such as STransfer (184) and spRefine (185) can efficiently transfer knowledge, considerably reducing manual effort.

In the medical field, machine learning holds substantial promise as an adjunct clinical tool for cancer detection and disease prediction. Deep learning, a subdiscipline of machine learning, automatically extracts features from low to high dimensions through neural networks trained on large amounts of data, learning complex input-output mappings. Tools in this field have evolved from early proof-of-concept models such as spaGCN and ST-Net, which were used for spatial domain detection and gene imputation but had limited cross-cohort validation and generalizability, to more sophisticated architectures. Subsequent models, including XFuse and DeepSpacE, have introduced probabilistic inference and validation across multiple cancer types, advancing the development of generalizable frameworks (6).

Addressing the limitations of traditional ST technologies

As discussed in previous sections, ST technology still faces numerous challenges in areas such as cell segmentation accuracy, RNA detection efficiency, batch effect removal and modeling of intercellular communication. These technical limitations have, to some extent, hindered the clinical translation of ST. In recent years, machine learning methods, particularly deep learning and graph neural networks, have garnered attention for their powerful feature extraction and pattern recognition capabilities and have been successfully applied to address the aforementioned shortcomings of ST technology.

In the area of cell segmentation and type annotation, traditional methods struggle to distinguish between spatially adjacent cells that are morphologically similar but functionally heterogeneous. To address this, the JSTA framework proposes a computational strategy that jointly performs cell segmentation and type annotation. By incorporating cell type-specific gene expression as prior knowledge, it effectively constrains the consistency between segmentation boundaries and annotation results. Simulation experiments have shown that this approach improves the accuracy of RNA molecule assignment by >45% (158).

In improving deconvolution accuracy, existing methods typically rely on single-cell RNA-seq data to construct reference matrices. However, due to data noise, batch effects and spatial heterogeneity, deconvolution results often contain biases. To overcome these limitations, recent trends have focused on multimodal data fusion. One approach integrates H&E-stained tissue images, extracting cellular morphological features using segmentation algorithms or deep convolutional networks. Another approach integrates spatial coordinate information, employing spatial autocorrelation models or graph neural networks to enhance the spatial smoothness of deconvolution results. Representative tools have demonstrated notable success: Spatial-ID significantly improves cell type mapping accuracy through transfer learning and spatial embedding strategies. STANN uses deep neural networks to achieve cross-modal mapping from single-cell data to ST data. BANKSY combines molecular features of individual cells with information from their local microenvironment, orienting cell phenotype inference around tissue spatial architecture (159). The deep neural network-based spatial cell typing method incorporates an enhanced gene selection strategy and lightweight deep neural network for data training, providing a faster and more accurate solution for ST data analysis, with high accuracy in cell type identification across different brain regions, species and ST platforms (186).

In terms of spatial feature extraction and spatial domain identification, various machine learning models have been integrated into ST analysis pipelines. STLearn provides a deep learning method that extracts tissue section image features as additional model inputs. SpaGCN constructs a graph network based on spatial proximity relationships among target spots and employs a graph convolutional network to learn their spatial distribution features. STAGATE introduces a cell type-aware attention mechanism that expands the structure of the neighborhood graph network to enhance local resolution, with a particular focus on improving boundary detection between distinct spatial domains. GraphST integrates similarity information from gene expression profiles during graph construction as an additional spatial criterion for edge connectivity. SEDR, as the first analytical tool designed for integrated ST datasets, employs a graph autoencoder to represent spatial features while minimizing batch effects induced by multi-sample integration through latent space constraints.

In batch effect removal, the gene spatial integration (GSI) pipeline employs an autoencoder network to extract the native spatial distribution information of gene expression and project it into a common feature space alongside expression features, thereby achieving seamless integration of multiple spatial datasets while preserving biological variation. Application of GSI to the human dorsolateral prefrontal cortex dataset demonstrated that this method stably improves the performance of mainstream clustering tools such as Seurat. Gene spatial distribution, as an independent information dimension, holds significant value in integrative analysis and batch effect removal is necessary for achieving fine-grained tissue feature characterization (163).

In modeling intercellular communication, systematic dissection of L-R interactions depends on experimentally validated L-R databases, yet current resources remain incomplete. Recently, breakthroughs in deep learning-based protein structure prediction tools, such as AlphaFold and related derivatives, have made it possible to predict protein complex structures, particularly high-confidence configurations of L-R complexes, at scale. However, it must be noted that these predictions are computational inferences rather than experimentally validated facts and thus carry inherent uncertainty and risk of false positives. Therefore, careful evaluation of predicted data, clear labeling of computationally inferred results and increased transparency in database construction are key prerequisites for ensuring the scientific and reliable use of L-R datasets now and in the future (165).

In summary, machine learning methods hold potential for addressing the limitations of traditional ST technologies across multiple levels, including cell segmentation, deconvolution, spatial feature extraction, batch effect correction and intercellular communication modeling. With the continued development of new paradigms such as multimodal fusion and self-supervised learning, machine learning is expected to propel ST into a new phase characterized by higher accuracy, enhanced interpretability and improved cross-platform integration capability.

Future prospects of deep integration

This shift has enabled larger-scale data training and more diverse functional annotations. The rapid development of various omics technologies, including ST, has provided machine learning with massive amounts of high-quality analytical data. The effective integration of the two holds promise for constructing more comprehensive maps of the TME, thereby enabling a more systematic understanding and analysis of tumor progression.

Classical ST studies have revealed spatial features of the TME that limit tumor prognosis. However, due to the limited sample sizes amenable to manual analysis, sequencing results often lack broad representativeness, impeding subsequent translational research. The introduction of machine learning has made large-scale ST data analysis feasible (187) thereby facilitating the construction of high-resolution TME molecular atlases. For example, Roller et al (188) developed a gene expression-based precision immunophenotyping tool to reflect the spatial infiltration patterns of CD8+ lymphocytes in the TME based on CD8 immunohistochemistry analysis of 2,023 tumor samples from patients. Similarly, Sun et al (189) constructed a neoantigen-reactive T cell atlas from single-cell transcriptomic data of thousands of tumor-infiltrating lymphocytes and built a gene-based machine learning model to accelerate the identification of neoantigen T cell receptors (Neo TCRs) and the engineering of Neo T cells for therapeutic purposes.

Multi-omics joint analysis is increasingly becoming a new trend in tumor immunology research (190-193). Machine learning models, by learning shared representations across different molecular modalities such as transcriptomics, proteomics and metabolomics, can effectively integrate these data while preserving their multi-dimensional spatial information. For example, tools including MUSE, SpatialGlue and MISO employ neural network architectures to align molecular information of varying resolutions, dynamic ranges and biological origins, enabling integrative analysis of multimodal data. Wu et al (194) developed an interpretable multimodal deep learning framework based on ST and spatial proteomics inputs, using a graph neural network approach, and achieved excellent performance in predicting immunotherapy response in patients with HCC. Stupichev et al (195) adopted an artificial intelligence-based multimodal approach to combine multiple TME-related biomarkers into a composite score for patients with ccRCC who respond favorably to immunotherapy, thereby overcoming the low predictive power of single biomarkers. Even in the face of highly heterogeneous molecular and morphological data, machine learning algorithms can achieve effective integration. He et al (196) developed a deep learning algorithm trained on 30,612 spatially resolved gene expression data points matched with histopathological images from 23 patients with breast cancer, enabling direct prediction of spatially resolved transcriptomes from tissue images and offering the possibility of image-based screening for spatial variation in molecular biomarkers. Bell et al (197) established a semi-supervised spatial multi-omics integration pipeline combining imaging, spatial ST, and single-cell RNA-seq. By performing ST and single-cell analysis on FFPE samples from pancreatic intraepithelial neoplasia (PanINs), they found that as lesions progress, CAFs transition from an inflammatory to a proliferative signaling state. This analytical framework has broad applicability and can be used to dissect spatiotemporal evolutionary dynamics across multiple cancer types.

In the future, the deep integration of machine learning with spatial omics will further extend into three-dimensional spatial modeling, digital twin construction and spatiotemporal dynamic analysis. Three-dimensional reconstruction can restore the authentic structure of tissues within their native environment. Digital twins integrate multimodal data to build patient-specific virtual tissue models, enabling real-time simulation of disease progression and therapeutic response. Spatiotemporal modeling helps reveal the dynamic principles underlying cell state transitions and microenvironment evolution. The synergistic development of these cutting-edge technologies will drive tumor research from static description toward dynamic prediction, providing a solid theoretical foundation and technical support for precision medicine and early intervention.

Outlook

Considerable progress has been made in ST, yet transforming it into a core driving force for tumor immunology research still requires confronting the most pressing unresolved questions in the field and defining clear priorities accordingly. At the level of experimental design, the greatest bottleneck is the 'spatial snapshot' mindset, the vast majority of studies describe static maps at a single time point, failing to distinguish causality from correlation. Therefore, the top priority is to establish standardized time-series spatial sampling models that incorporate key time points including baseline, early treatment response and resistance/relapse in organoids, animal models and clinical specimens, thereby capturing the dynamic evolution of the microenvironment. In terms of computational benchmarking, differences across platforms in resolution, sensitivity and analytical assumptions lead to biological conclusions that are highly dependent on the technological approach. The field critically lacks cross-platform comparable benchmark datasets. There is an urgent need to initiate systematic inter-platform benchmarking studies that apply mainstream technologies in parallel to the same set of tissue samples and make raw data along with standardized analysis pipelines publicly available. At the clinical validation level, although spatial biomarkers are being identified at a rapid pace, the vast majority lack validation in prospective cohorts and cross-platform definition of prognostic thresholds. The core priority is to establish standardized quality control metrics applicable to clinical samples and to launch multicenter prospective clinical trials to evaluate the independent predictive value of spatial parameters compared with conventional biomarkers.

The deep integration of high-resolution ST with spatial proteomics, spatial metabolomics and high-precision imaging technologies holds promise for constructing a multi-dimensional spatiotemporal atlas of the tumor immune ecosystem that encompasses gene expression, protein function, metabolic activity and intercellular interactions. From the perspective of technological development, the core direction of ST lies in the deep fusion of sequencing-based and imaging-based approaches: Sequencing-based methods excel at whole-transcriptome discovery but have limited resolution, whereas imaging-based methods achieve subcellular precision but are constrained by predefined gene panels. Future breakthroughs include the development of spatial multi-omics to enable multimodal detection on the same tissue section, the extension of analysis into three-dimensional space through tissue clearing technologies, the incorporation of machine learning algorithms and the advancement of standardized workflows compatible with FFPE samples and cross-platform validation. On this basis, future research needs to extend into the temporal dimension, establishing time-series ST atlases that cover the entire process of tumor initiation, progression and treatment, thereby revealing the dynamic evolutionary trajectory of the microenvironment under therapeutic pressure and elucidating the formation and dissolution of immune exclusion barriers, the spatiotemporal sequence of T cell exhaustion and the spatial expansion patterns of drug-resistant clones. Concurrently, the deep integration of multimodal data will establish a complete functional chain from gene expression to protein function to metabolic activity. Combined with three-dimensional imaging technologies, this approach holds promise for constructing a four-dimensional tumor immune ecosystem. Furthermore, structure-based large language models or generative artificial intelligence methods should be developed to predict novel L-R binding pairs directly from protein sequences or spatial structures without relying on existing databases, followed by in situ validation incorporating spatial proximity relationships, thereby providing new molecular targets for spatially guided intervention.

Currently, the tension between the highly organized, cancer type-specific nature of microenvironment remodeling mechanisms and cross-cancer universal principles, as well as the gap between physical proximity and actual intercellular interactions, remain core challenges in the field. In the future, integrating technologies such as spatial lineage tracing, in situ perturbation and proximity labeling holds promise for transforming these static associations into direct validation of dynamic functional interactions. In summary, ST is driving tumor immunology research from spatial atlas mapping toward mechanistic dissection and clinical translation. By defining three major priorities, experimental design, computational benchmarking and clinical validation, and introducing time-series sampling, multi-omics integration and AI-driven L-R prediction, the field will achieve comprehensive and dynamic dissection of the tumor immune microenvironment, laying a solid foundation for spatially guided precision immunotherapy.

Availability of data and materials

Not applicable.

Authors' contributions

JJ, XT, YC, QY and BS contributed to writing-original draft, writing-review and editing. JM contributed to writing-review and editing, and funding acquisition. Data authentication not applicable. All authors read and approved the final manuscript.

Ethics approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Acknowledgements

Not applicable.

Funding

The present review was supported by National Natural Science Foundation of China (grant no. 82472701).

References