Spatial transcriptomics technology in cancer research

In recent years, spatial transcriptomics (ST) technologies have developed rapidly and have been widely used in constructing spatial tissue atlases and characterizing spatiotemporal heterogeneity of cancers. Currently, ST has been used to profile spatial heterogeneity in multiple cancer types. Besides, ST is a benefit for identifying and comprehensively understanding special spatial areas such as tumor interface and tertiary lymphoid structures (TLSs), which exhibit unique tumor microenvironments (TMEs). Therefore, ST has also shown great potential to improve pathological diagnosis and identify novel prognostic factors in cancer. This review presents recent advances and prospects of applications on cancer research based on ST technologies as well as the challenges.


Introduction
Cancer is the leading cause of death worldwide (1). A series of studies have shown that high cellular heterogeneity is one of the main causes that make cancer difficult to cure (2). Therefore, it is of vital importance to characterize the heterogeneity of tumor. Single-cell sequencing technologies, especially single-cell RNA sequencing (scRNA-seq), provides useful tools for uncovering cellular heterogeneity (3)(4)(5)(6). However, they always lose spatial histology information while cell dissociation, hence difficult to characterize spatial cellular interactions and organization of tumors.
The spatial structure of tumors is closely related to tumorigenesis, progression, and treatment response. Histologically similar tumors from different regions are molecularly distinct and have different tumorigenicities, showing the association of tumor initiation and the spatial location (7). In addition, spatially adjacent cells have stronger interactions, which could functionally remodel the TME and promote tumor progression (8,9) and evolution (10). The tumor cells at different positions may exhibit heterogeneous metastatic potential (5), while the site of lymphocyte infiltration usually indicates specific treatment response (11). Scientists have used immunohistochemistry (IHC) and immunofluorescence (IF) staining etc. technologies to locate the cells and proteins in flash-frozen or formalin fixed paraffin-embedded (FFPE) tissue sections. However, these approaches are low-throughput and can only label limited number of pre-selected proteins in a specific section, thus incapable of discovering the distribution of novel proteins or cell types. By contrast, ST technologies have much higher throughput and can capture the whole transcriptome, showing its power in constructing spatial cell atlas in embryo, brain, heart, etc. (12). In the meanwhile, an increasing number of studies used ST to profile spatial heterogeneity of cancers (13,14). Currently, ST has been used to distinguish tumor and non-tumor tissues, special spatial areas such as tumor interface and tertiary lymphoid structures (TLSs) (15)(16)(17), and identify spatial-specific prognostic factors in cancer (18,19).
In this article, we first made a brief introduction of ST techniques, and reviewed recent studies on cancer research using ST. Then, we proposed our insights into the challenges and prospects of applying ST into cancer research.

Categories of ST
ST technology could be divided into two main categories according to detection methods, including imaging-based methods and sequencing-based methods (20). Imaging-based ST methods consist of in situ hybridization (ISH) and in situ sequencing (ISS) ( Table 1). In ISH, RNA molecules from individual parts (or cells) within the tissue were achieved by hybridizing a labeled probe complementary to the target of interest. This technique was first used for visualizing gene expression in 1982 (21). Single-molecule RNA fluorescence in situ hybridization (smFISH) has a stronger and more robust signal compared with initial ISH (22). Afterwards, seqFISH (23), seqFISH+ (24), multiplexed error-robust FISH (MERFISH) (25), MERFISH+ (26) et al. further improved smFISH in the aspect of target throughput. Nevertheless, ISH-based methods are not transcriptome-wide, which always need prior knowledge to design probes and obstruct comprehensive expression analysis in a single experiment (27)(28)(29). In ISS, RNA molecules from a cell are sequenced directly in its tissue context. The first ISS technique was published in 2013, using padlock probes to target known genes (30). Later, BaristaSeq (31) and STARmap (32) improved sensitivity and/or number of detected genes. In short, most ISS-based ST techniques have subcellular resolution. However, they usually have a limited number of targeted genes or low detection efficiency (33), thus restricting their applications in the specific scenarios ( Table 1).
The sequencing-based ST techniques include laser capture microdissection (LCM)-based methods and in situ barcoding (ISB)-based methods (Table 1). LCM-based methods, such as geographical position sequencing (Geo-seq) (34), TIVA (35) and NICHE-seq (36), utilize a laser beam to cut out specific tissue regions identified under a microscope (37, 38). Compared with initial LCM method, Geo-seq improved sensitivity but with lower resolution. TIVA can be performed on live cells but with low throughput, while NICHE-seq has higher throughput but not applicable to human samples. Generally, LCM-based ST techniques are labor-intensive and low-throughput, thus inapplicable of processing samples in batches. ISB-based ST techniques capture RNA molecules in situ, then perform cDNA sequencing ex situ. In terms of barcoding, it can be subdivided into two groups. The first group uses solid phase-based capture (SPBC) methods (13), and the tissue is transferred to a substrate bearing a pre-arranged set of DNA barcodes, which includes 10x Genomics Visium (39), Slide-seq (40), Slide-seq2 (41), HDST (42), Stereo-seq (43) etc. The second group, including NanoString digital spatial profiling (DSP) (44) and ZipSeq (45), uses selective barcoding methods, which means DNA barcodes are either collected from or delivered to selected tissue locations. ISB-based ST techniques have been used to study mouse olfactory bulb, gingival tissue, adult human heart tissue as well as multiple cancers (12). Most ISB-based techniques are transcriptome-wide with relative higher throughput, and some of them have subcellular resolution including Seq-Scope (46), HDST (42), APEX-seq (47), PIXELseq (48) and Stereo-seq (43) etc.
Apart from the above-mentioned two ST categories, there also exists bioinformatic methods for reconstructing spatial positions of cells using scRNA-seq data. For example, novoSpaRc allows for de novo spatial reconstruction of singlecell gene expression with no inherent dependence on any prior information (49). Though many tools are available for the reconstruction of spatial positions of cells currently, their effectiveness remains to be validated in the future (50).

ST provides new insights in cancer research
Spatial heterogeneity of the tumor cell Tumor tissues roughly consist of various cell types, including tumor cells, stromal cells, and immune cells (51) ( Figure 1A). The differential cell composition induces the diversity and heterogeneity of TME. The heterogeneity confers different abilities of proliferation, immune resistance, immune escape, and survival etc. on tumor cells, which is reflected in many aspects, such as cell composition, gene expression pattern and cell spatial positions. Thus, tumor cells can be divided into several subpopulations in terms of genotype, phenotype, or spatial position.
Currently, neoplastic spatial heterogeneity has been reported in multiple malignant tumors, such as invasive micropapillary carcinoma (IMPC), gastric cancer (GC), glioblastoma (GBM), primary liver cancer and melanoma (19,(52)(53)(54)(55) (Table 2). For instance, ERBB2, the receptor of HER2, differentially expressed in  (15). Further analysis showed these differentially expressed genes (DEGs) were associated with immune response, mitogenic programs and tumor invasion pathways, suggesting different areas may have differential abilities on tumor progression, invasion and immune resistance (39). Additionally, certain stress may induce certain new cell populations, which adapt to the specific TME. For instance, a functional subgroup was found in mouse pancreatic cancer model in a hypoxic microenvironment. Furthermore, the hypoxia-induced tumor tissues had less subpopulations and simplified functions (53). Intriguingly, though tumor cell subpopulations are spatially heterogeneous on the same slide, the transcriptional profile of each subpopulation is probably recurrent across tumor slides. Further analysis revealed these transcriptional programs were independent of cell-cycle states, thereinto reactive-hypoxia was associated with chromosomal alterations, indicating potential connection with genome instability (56). Moreover, tumor epithelial cells also exhibited specific colocalization with immune cell subpopulations (57, 58).
In bladder cancer, CDH12+ epithelial cells colocalized with exhausted CD8+ T cells (57) ( Figure 1B). In breast cancer (BC), tumor epithelial cells were negatively correlated with cancer- associated fibroblasts (CAFs), endothelial cells, B cells etc. (15). In sum, the diverse colocalization of tumor cell and other cell types inflects potential cell-cell interactions and the complex TME in tumors.
Tumor cells also showed different gene expression profiles in tumor core and periphery. Compared with tumor core, the tumor peripheral area is nearer to adjacent normal tissues, thus it has different TME ( Figure 1A). In a neuroblastoma mouse model, a tumor cell cluster, which was most enriched at the tumor core, expressed more cancer-associated genes than that dispersed all over the tumor area (81). In prostatic cancer (PC), TAGLN (tumor suppressor) and HLA had higher expression in the periphery, whereas NUPR1 and KLK4 etc. were expressed higher in the tumor core (82) ( Figure 1C). These findings jointly indicate the tumor core probably is more malignant than the periphery.
A series of ST-based studies showed tumor subpopulations are spatially mutually exclusive on the same slide ( Figure 1A). In BC, the tumor area with high signatures of epithelialmesenchymal transition (EMT), interferon (IFN) and major histocompatibility complex (MHC) was negatively correlated with that with high signature of proliferation (11). Another study revealed HER2+ and ESR1+ breast tumor clones had mutually exclusive localizations (83) ( Figure 1D). In primary pancreatic ductal adenocarcinomas (PDAC), TM4SF1+ tumor cells had mutually exclusive spatial locations with S100A4+ tumor cells (52). Another study identified three recurrent sub-TME phenotypes within the same tumor tissue in PDAC, which also showed clear boundaries with each other (84). The mutually exclusive locations of tumor cell subpopulations suggest different clonal origins and potential competition across tumor areas (39).

Spatial heterogeneity of microenvironment
The stromal cell is an important component of TME, which is associated with tumor growth, progression, immunosuppression and metastasis (85)(86)(87)(88)(89). ST has uncovered a series of spatial distribution preference of stromal cells in TME. For example, in a lung cancer mouse model, loss of Tgfbr2 resulted in a remodeling of the stroma and induced tumor development (89).
Among stromal cells, fibroblasts showed the most prominent spatial colocalization features in multiple cancer types. In bladder cancer, fibroblasts were observed resided in close proximity to CDH12+ epithelial cells (57). In colorectal cancer (CRC), tumor-specific FAP+ fibroblasts and SPP1+ macrophages colocalized in the tumor area, which were proved contributed to desmoplastic TME (90) ( Figure 1E). In PDAC, fibroblasts and terminal ductal cell populations were significant enrichment in tumor areas, suggesting ductal cells in the cancer region may express hypoxia-response genes due to low oxygen content. In diffuse-type GC, CCL2+ fibroblasts and endothelial cells were enriched in the deep invasive layer of GC compared with the superficial layer, suggesting a greater ability of tumor invasion (91). In addition, CAFs also exhibited wellpreserved colocalization patterns with endothelial cells and perivascular cells (15), which is consistent with a previous study in cutaneous squamous cell carcinoma SCC (92). Interestingly, in cervical SCC, CAFs were enriched around certain tumor areas. Compared with tumors without surrounded by CAFs, CAF-surrounded tumors were more active in metabolism and cell growth and downregulated cellular adhesion, apoptosis, and immune response, suggesting a supportive TME for tumor progression and metastasis (93). On the other hand, fibroblast subpopulations may have mutually exclusive locations. For example, Wu et al. found spatially negative correlation between myofibroblast-like CAFs and inflammatory-like CAFs in estrogen receptor positive and triple-negative breast cancer (TNBC) samples (11). This finding was also observed in a HER2+ breast tumor dataset, suggesting a conserved relationship across BC subtypes (15).
The immune cell infiltration is frequently observed in tumor tissues, which is closely associated with treatment response (94). However, not all immune cells can infiltrate into tumors (81), especially in metastases (95). ST-based studies have shown the immune cell has a specific preference on the spatial distribution as well. For instance, in a lung cancer mouse model, immune cells were more concentrated in the outer area of tumors, particularly CD4+ and CD11c+ cells. Intriguingly, when Irf1 or Socs1 was knocked out, immune cells were scattered throughout the tumor (89). As we know, macrophages can be roughly divided into M1 macrophage and M2 macrophage in terms of encouraging or decreasing inflammation (96). M1-like macrophage and M2-like macrophage appeared to have mutually exclusive locations in PDAC. Of note, a M2-like macrophage subpopulation was observed colocalized with proliferating cancer cells (83,90). The M2-like macrophages were most enriched in the ducts, whereas the M1 macrophage were enriched in the stroma and cancer regions (52). Similarly, in BC, two macrophage subpopulations, which were outside of the conventional M1/M2 classification, displayed a modest negative spatial correlation (11). In addition, macrophage subpopulations also colocalized with T cell subpopulations in BC and neuroblastoma (15,81). Further analysis revealed the colocalization might be associated with immune activation and tumorigenesis. In GBM, the colocalization of immune cells and stromal cell subpopulation was associated with immunosuppressive microenvironment (54,97). Additionally, the presence of certain tissues may also influence the distribution of T cell subtypes. In liver cancer, intact continuous fibrous capsule indicated significant decrease of exhausted T cells and downregulated immune checkpoint genes, suggesting it may act as a barrier preventing the infiltration of immune cells (16). In addition, SPP1+ tumorassociated macrophages (TAMs) were colocalized with proliferating cancer cells (83), which is consistent with a recent study in CRC (90).

Special spatial structures in TME Tertiary lymphoid structures
TLSs, sometimes also known as tertiary lymphoid organs or ectopic lymphoid structures, are organized aggregates of lymphoid cells that arise postnatally in nonlymphoid tissues, such as tissues subjected to chronic inflammation and cancers (98). They are characterized by an inner zone of CD20+ follicular B cells surrounded by CD3+ T cells. Recently, more evidence supported the important function of TLS in delaying or promoting cancer progression (98)(99)(100).
TLS was mostly found in para-tumor tissues (16). However, it also exists in tumor, leading-edge and para-tumor areas (19,98) ( Figure 1A). Previous studies generally identify TLS by multiple IHC or multiplex IF staining (101)(102)(103), which is accurate but labor-consuming and low-throughput. Mature TLSs are characterized by the presence of a germinal center containing T follicular helper (Tfh) cells and follicular dendritic cells in close contact with B cells (104). TLS could be identified by certain gene signatures or cell population aggregates using ST ( Figure 1F; Table 2). For example, in BC, Andersson et al. evaluated the degree of B-and T-cell colocalization by spot deconvolution, ergo potentially constituting parts of a TLS which they called TL-like structure. At last, they proposed a TLS signature with 171 genes for TLS prediction (15). In renal cell carcinoma (RCC), Meylan et al. proposed a TLS signature with 29 genes (17). A similar gene signature was also proposed in liver cancer (16). Of note, in BC, researchers found a negative correlation between B cells and plasma cells, and colocalization of B cells and T cells (15), which probably suggested the presence of TLS.  Tumor interface has unique ecosystem The tumor interface, the invasion front for tumor cells to expand, is a critical region to uncover the tumor invasion and progression (5, 105-107). The intermediate zone between the tumor tissue and the non-tumor tissue has unique microenvironment, thus existing unique characteristics in cellcell interaction, cell composition and immune states ( Table 2).
The tumor interface exhibits a transitional state between the tumor and adjacent non-tumor tissues. In a zebrafish model of melanoma, a distinct cell cluster at the interface was found histologically resembled the microenvironment, but transcriptionally resembled tumor, and may function as a bridge for interaction between tumor and neighboring tissues (107). This unique interface was also observed in human melanoma (107). In primary liver cancer, PROM1+ and CD47 + cancer stem cell (CSC), which was scattered in tumor areas, was found to be gradually increased from leading-edge to tumor to portal vein tumor thrombus, and closely related to TME remodeling and tumor metastasis. The unique metabolic features at the interface, where may occur sudden decrease or increase of certain hallmark pathways, suggest it may serve as a buffer between tumor and normal regions (16).
The tumor interface has specific cell compositions. In cutaneous SCC, a tumor-specific keratinocyte (TSK) subpopulation located at the leading edges functioned as a hub for intercellular communication, which is probably due to short distance between the interface and tumor or normal tissues (92). In BC, LumA tumor cells scarcely existed in the interface of tumor, whereas LumB cells were scattered throughout the tumor tissue, suggesting heterogeneous spatial distribution of tumor cell subpopulations (108). In a mouse pancreatic cancer model, a tumor cell cluster upregulated LDHA (a hypoxia-related gene) in the interface area compared with that in the tumor center and had stronger ability for survival and invasion (53).
The tumor interface has unique immunosuppressive TME as well. In intrahepatic cholangiocarcinoma (ICC), macrophage and NK/T cells were enriched close to the boundaries. Of note, immune cells were recruited to the borderline from the tumor side, and the region nearer to the borderline enriched more immune cells, showing the heterogeneous distribution of immune cells in both the axial and lateral directions ( Figure 1G). Additionally, immune checkpoint genes such as BTLA, CTLA4, CD96 and IDO1 were enriched on the tumor side of the interface (19,113). Similar aggregation of macrophage subpopulations and T cells was also reported in ICC metastases (113). In GBM, a subset of IL-10-releasing HMOX1+ myeloid cells were identified in the direct neighborhood of mesenchymal-like tumor regions, which drove T cell exhaustion and thus facilitated the immunosuppressive TME (97). In TNBC, a CD8+ T cell population was identified at the interface instead of tumor core, but their functions need further elucidation (109).

Treatment response
It is very important to predict the response of cancer patients for specific cancer therapy clinically. However, it is quite difficult due to complex tumor heterogeneity (2,123). ScRNA-seq has revealed specific cell components could influence treatment response in diverse cancers (127). ST further revealed the specific spatial aggregation and cell-cell interactions of certain cell subpopulations could induce differential treatment responses. For instance, the TLS, which aggregates immune cells in or near the tumor, provides a niche that promotes in situ B cells maturation toward plasma cells in RCC (17). These plasma cells disseminate into the tumor tissue and can produce IgG. The tumors with high percentages of IgG-labeled tumor cells were more infiltrated with CD68+ macrophages, which is one of the main effectors of antibody-dependent cellular cytotoxicity (17). This process may promote immunoreactivity and results in better response to ICI treatment. Additionally, the tumor invasive front was found enriched with specific cell subpopulations, which differ in metabolic states and functions from those elsewhere. SAA+ hepatocyte subpopulations near the invasive front are associated with recruitment of M2-like macrophages in ICC, which may form a niche with impaired immune response and promote further tumor invasion and a worse treatment response (19). Moreover, ST also uncovered the connection between specific cell-cell interaction and treatment response. Most FAP+ fibroblasts and SPP1+ macrophages were colocalized in CRC (90,121). These FAP + fibroblasts enhanced the recruitment and the proinflammatory activity of SPP1+ macrophages through WNT5A-FZD2 pairs and the expression of TGF-b superfamily genes (immunosuppressive molecules) etc., respectively. In addition, FAP+ fibroblasts and SPP1 + macrophages upregulated genes of extracellular matrix (ECM)related pathways, suggesting their role in facilitating the generation of desmoplastic structures, which further limited the immune cell infiltration and induced diminished ICI treatment response (90,121). Interestingly, the interaction of macrophage subpopulations also affects the treatment response in colorectal cancer liver metastasis (116).

Clinical application of ST Diagnosis
The IHC technique has long been used by physicians and pathologist to diagnose a tissue as benign or malignant, determine the grade and stage of a tumor, identify the cancer cell types, and find the origin of metastasis (128)(129)(130). Compared with IHC-and IF-based methods, ST has comparable resolution and is almost transcriptome wide, indicating an enormous potential in cancer pathology (20,131). Through spot deconvolution and prior cell marker genes, researchers can estimate the cell composition of spots and further divide the ST sections into several spatially different areas i.e., tumor area, tumor leading-edge area and paratumor area. Several pilot studies have shown ST-based pathological annotations displayed comparable or even higher accuracy than that from pathologists (38, 82,110,112,123) ( Table 2)

Prognosis associated factors
Prognostic factors can indicate the clinical outcomes of various kinds of diseases (132) ( Table 2). Currently, three kinds of prognostic factors identified by ST have been reported in multiple cancer types. The first is gene markers. In melanoma, in the tumor compartment, high expression levels of CD8, CD3, TIM3, IDO1 etc. suggested longer progression free survival (PFS), whereas high levels of B2M and PD-L1 in macrophage compartment were associated with longer overall survival (OS) (18). In ICC, hepatocytes close to invasive fronts with high expression level of SAA1 and SAA2 were correlated with worse prognoses (19). Moreover, many gene markers were identified in PDAC, liver cancer, BC, bladder cancer, SCC and GC (53, 92,112,114,121,126,133). Cell subpopulations were also potential markers of prognosis. For example, CCL2+ endothelial cells and fibroblasts in the deep invasive layer were associated with poor clinical outcomes (91). Similarly, in bladder cancer, CDH12+ epithelial cell was associated with poor prognosis (57). Additionally, FAP+ and INHBA+ CAFs, and high expression of HSF1 were reported negatively correlated with survival in GC (58,126). In glioma, blood-derived TAMs indicated poor prognosis (54). A gene signature usually consists of tens to hundreds of genes, which can serve as a prognostic factor as well. For instance, in PC, Taavitsainen et al. proposed a gene signature called PROSGenesis score and further demonstrated high PROSGenesis score was associated with good prognosis (123). In HCC, Wu et al. proposed high TLS-50 signature score was associated with good prognosis (16). Unsurprisingly, higher CAF signature predicted unfavourable PFS and OS in cervical SCC (93) and GC (126).

Challenges of ST
ST has displayed its advantages in characterizing tumor heterogeneity in gene expression patterns and cell compositions, and potential in clinical studies (12,134). Though powerful, the cost will limit its wide application. More importantly, two aspects of technical issues also need to concern.
On the one hand, the performance and applicability of ST need to be further improved. First, the detection efficiency of ST is relatively low compared with scRNA-seq (29), which can hardly capture RNA molecules with low expression levels, leading to missing of potential genes that play a role in tumor progression, metastasis, or relapse. Since for most high-resolution SPBC techniques, the area with barcode to capture molecules is less than 30%, thus more than 70% of the mRNA molecules will be missing. Second, most ST approaches only obtain single-end transcripts instead of the full-length transcripts. Thus, it is hard to investigate immune cell receptor repertoires and alternative splicing events, which are important for cancer research. Furthermore, full-length transcripts will make it possible for variant calling across the transcriptome and differential expression analysis at the isoform level as well (135). Third, ST is not exactly suitable for FFPE samples. FFPE tissue blocks, which are the gold standard method of preservation of human tissue for diagnosis, are usually stored for a long time. Consequently, RNA molecules in these blocks are often degraded seriously, hence it is a great challenge to utilize these samples. NanoString DSP and 10x Genomics Visium have shown their compatibility for FFPE tissue blocks (17,18,136). However, the quality of data probably varies with specific samples and is much lower than using the fresh one.
On the other hand, the bioinformatic tools for ST data analysis do not meet current needs. Generally, extracting valuable information from raw ST data requires several steps of data processing, including imaging processing, reads mapping, gene expression mapping, followed by downstream analysis such as spot deconvolution, clustering, detection of spatially variable genes, cellular interaction inference, and trajectory inference et al. (137) (Supplementary Table 1). Some of these steps, such as deconvolution and clustering, are also frequently used in bulk RNA-seq and/or scRNA-seq data processing, hence many tools initially designed for RNA-seq or scRNA-seq are compatible with ST data. However, their power will unavoidably decrease due to neglect of spatial location and data structure features (50). Thus, it is essential to take features exclusive to ST data into consideration. First, for ST with multicellular resolution, there are plenty of tools for deconvolution of spots, such as SPOTLight (138), SpatialDWLS (139), CARD (140) and Stdeconvolve (141) etc., yet it is difficult to affirm the precise number of cells within a spot but for H&E-stained images. In addition, for ST with subcellular resolution such as HDST (42) and Stereo-seq (43), few tools could identify the individual cell by merging multiple adjacent spots. To distinguish an individual cell from ST is quite important if we want to make full use of this ultra-high-resolution technology. Second, tools for integrating spatial data from multiple batches, platforms, omics, and species are scanty. In this aspect, Spacemake (142), BASS (143) and MAPLE (144) claimed their capacity of multi-platform data integration of ST, but it is essential to further confirm their effectiveness.
Previous studies have shown the prevalent existence of heterogeneity in genomics, transcriptomics, proteomics, and epigenomics (145). However, single-omics technology can only profile the tumor heterogeneity from a specific angle, which ineluctably loses a large amount of information about the other omics. To handle this problem, scientists have developed a series of multi-omics techniques parallel sequencing of single-cell genomes and transcriptomes (146)(147)(148)(149)(150)(151)(152). At present, spatial multi-omics technology such as MOSAICA (153), DbiT-Seq (154), SM-Omics (155), NanoString CosMx ™ SMI platform (156) and spatial protein and transcriptome sequencing (SPOTS) (157) etc., can quantify the transcriptome and multiple proteins and retain the spatial coordinates. The multi-modal spatial genomics approach provides a promising platform for studying the intrinsic and extrinsic factors contributing to spatial heterogeneity in gene expression and genomic variants (158). Currently, spatial multi-omics has been used in cancer research (41). Another study tried to integrate protein subcellular localization, affinity proteomics, mass spectrometry (MS) data sets and RNA-seq information in a human lymphoma cell line. Their work helps to deepen the knowledge on the architecture of the cells and the complexity of cancer heterogeneity (159). In sum, multi-omics technology has shown its unprecedented value in cancer biology.

Prospects of ST in cancer research and clinical application
ST techniques have ultrahigh resolution and retain the spatial information of genes and cells, which make them capable of solving many outstanding biological questions. As we know, most current ST can only display gene expression profile and organization of cells on the plane, which is two-dimensional, and does not truly recapitulate the spatial, cellular, and chemical environment of highly complex tumors and their stroma (160). The real spatial atlas should be three-dimensional (3D), which can restore the most realistic spatial environment of cells in tumors, and further provides a more precise atlas and solid basis for research on the mechanism of tumorigenesis and progression, cancer heterogeneity, and clinical applications. Currently, there are bioinformatic tools for reconstruction of 3D ST atlas using multiple sequential adjacent slices and the 3D architecture of human heart (161), cardiac organoid (162), and mouse brain (163,164) etc. have been reconstructed (Supplementary Table 1). However, few studies characterized the 3D structure of tumor tissues (165). We believe the reconstruction and characterization of the 3D architecture of tumors based on ST will revolutionize our knowledge of cancer. In addition, ST may shed light on the role of organelles in cancer.
Compared with normal cells, cancer cells display alterations in energy metabolism, which are closely associated with mitochondrial activities (166). Subcellular ST techniques could help to precisely locate the genes that involve in the abnormal energy metabolism and further reveal the mechanism of how they contribute to cancer growth. Previous research also revealed lysosomal related to the dysregulation of tumorigenesis-associated pathways in cancer (166). Interestingly, the intracellular positioning of lysosomes has close connection with the function of cells. For example, the lysosomal subpopulations at the edges of the cancer cell could regulate cell adhesion, exocytosis, and invasion (167). However, it is not clear how the spatial position affects the functions of lysosomes. In this respect, ST techniques such as APEX-seq (47), show great potential for discovering the genes and pathways involved in the functional transition within the organelle. Moreover, ST also shows great potential in premalignant disease research. Currently, the aggressiveness of a premalignant lesion is primarily evaluated by cell morphology (168). However, cell morphology alone is not always sufficient for predicting the evolutionary trajectory of a premalignant lesion. ST remains spatial information as morphology, and it has higher resolution and could predict the differentiation directions of tumor cells based on gene expression profiles and splicing events (169,170). The additional information provided by ST allows it to identify the precancerous lesions more accurately.
Tumorigenesis, tumor progression and metastasis usually accompany with genomic, transcriptomic, and metabolic changes, which can serve as a basis for diagnosis and identification of subtypes of cancer (171,172). Recent studies proved ST could obtain these variations (16,56,(172)(173)(174)(175), suggesting ST has great potential in cancer clinical application, especially for digital pathology (DP). In addition, the combination of ST and other information, i.e. pathological imageology and spatial proteasome, may bring new insights into DP (176).

Concluding remarks
Currently, there are two major kinds of ST techniques in term of the experimental principle, thereinto ISB-based ST technique is among the most promising one due to high resolution and throughput, as well as the ability of whole transcriptome profiling. ST has shown its power in characterizing spatial heterogeneity and clinical applications in cancer. Researchers have profiled the TME in PDAC, BC, CRC, GC, PC, lung cancer, liver cancer, skin cancer, gliomas etc. based on ST. In addition, an increasing number of studies have shown the strength of ST in pathological diagnosis. Many novel potential prognostic markers were also discovered recently. These results jointly suggest ST is a promising technology for comprehensively elucidating spatial heterogeneity, discovering specific spatial structures in tumor tissues as well as for applications in clinical such as pathological diagnosis and prognostic prediction.
With the development of ST techniques and matched bioinformatic tools, the challenge in cost, sensitivity, and automation will be overcome in a few years. In addition, spatial multi-omics technology further integrates transcriptome with other omics such as proteomics, providing more comprehensive landscape of cancer, which may revolutionize our knowledge on cancer. In conclusion, ST technology is progressing rapidly, and it is promising for cancer research and clinical application.

Author contributions
QY collected the related papers, drafted, and revised the manuscript. LW designed the review and helped to draft and revise the manuscript. MJ prepared the figure and the tables, and participated in the discussion of revision. All authors contributed to the article and approved the submitted version.