Single-Cell and Spatial Omics: Methods and Applications.

Single‐cell and spatial omics have contributed to multiple aspects of applications in human biology and disease, ranging from developmental biology, evolutionary biology, disease mechanism identification, and clinical translation. These applications are divided into atlas‐based, algorithms‐based, and clinical translation‐based, which are summarized in the following subsections and Figure  4 . Applications of single‐cell and spatial omics analysis in biology and disease. The applications of single‐cell and spatial omics analysis in biology and disease can be classified into three types, including atlas driven, algorithm driven, and clinical driven. Constructing atlases at different levels provided a foundational map for biology and medicine. Developing algorithms specific to biomedical problems enabled AI‐driven biomedical discovery. Analysis according to clinical need facilitated the biomarker discovery, subtype identification, and drug discovery, some of which were validated by preclinical experiments and clinical trials. Single‐cell and spatial atlases constitute a foundational resource for biomedical discovery, providing reference maps of human cell types, tissue architecture, and disease‐associated cellular ecosystems. These resources span species‐level, organ‐level, disease‐level, and longitudinal datasets, collectively enabling systematic characterization of human biology in both health and disease (Table  3 ). Selected single‐cell/spatial atlas for all organs and systems. https://atlas.kpmp.org/explorer/ Note : Some of the atlases were named based on their description in the corresponding references. In 2016, the launch of The Human Cell Atlas (HCA) opened up the analysis of human biology at the single‐cell level. The current version of HCA project has constructed a single‐cell dataset covering 62 million cells from around 9100 human donors, providing a comprehensive reference for human cell biology investigation [ 314 ]. Complementing HCA, Single Cell Atlas [ 315 ] integrated eight kinds of multiomics data, including single‐cell and spatial omics across different healthy human tissues. The Tabula Sapiens constructed a multiple‐organ single‐cell transcriptomic atlas of humans, which identified nearly 500 types of cells [ 316 ]. These species‐level single‐cell atlases serve for a foundation that would facilitate disease mechanism discovery and treatment design. An organ is a structural unit composed of multiple tissues capable of performing specific physiological functions, characterized by specialized cell types, defined size, and complex morphological structures (e.g., buds, tubes, branches). Organ‐level single‐cell or spatial atlas facilitated the understanding of organ development and function in fine‐grained resolution, and identified key cellular components during disease development process. The integration of single‐cell atlas data from organ‐level depicted the cell interactions in brain, kidney, and lung, among others. The Brain Cell Atlas [ 317 ] integrated human and mouse single‐cell transcriptomic data to chart regional brain development and neuronal diversity. Multimodal spatial cell atlas of human kidneys [ 318 ] and intestine [ 319 ] are as part of the Human BioMolecular Atlas Program (HuBMAP), which serve as reference for understanding the organs. Human Lung Cell Atlas was constructed on over 2.4 million cells from 486 individuals with healthy or diseased status, which identified shared cell states across multiple lung diseases [ 320 ]. Single‐cell and spatial transcriptome atlas of eight regions of human heart revealed cellular niches and potential drug targets for human cardiac [ 321 ]. Moreover, kidney atlases comparing human and mouse highlighted conserved pathways and species‐specific disease mechanisms [ 322 ]. Collectively, these reference atlases standardize annotation frameworks, support evolutionary and developmental analyses, and serve as essential baselines for disease‐focused investigations. Disease‐focused atlases extend these foundations to pathological contexts. Many studies constructed single‐cell atlas by performing single‐cell sequencing on clinically diagnosed patient samples to reveal the disease mechanisms, especially for cancer. Early disease atlases were limited in scale due to technological and cost barriers, such as a Pancreatic ductal adenocarcinomas (PDAC) atlas profiling 57,530 cells from 35 tumors and controls, which revealed extensive intratumoral heterogeneity [ 323 ]. With technological maturation, cohort sizes have expanded dramatically, enabling cellular‐resolution characterization across hundreds of patient samples. Pan‐cancer single‐cell atlases encompassing more than 700 cases have delineated conserved and divergent immune, stromal, and malignant programs across tumors, identifying recurrent T cell exhaustion states [ 324 ], natural killer cell phenotypes [ 325 ], and cancer‐associated B cell populations [ 326 ]. Gavish et al. integrated scRNA‐seq data of 1163 tumour samples covering 24 cancer types to construct a pan‐cancer single‐cell transcriptome atlas, identifying 11 intratumor transcriptomic patterns [ 327 ]. Moreover, the analyses of disease biology mechanism have been deepened by collecting samples across different locations, tissues, stages, and clinical phenotypes. A study analyzed the single‐cell transcriptome data from different sites of collected 10 HCC patients with primary and/or metastatic tumors samples [ 328 ], which identified microenvironment‐based subtypes for predicting the prognosis of HCC patients. These resources collectively reshape our understanding of tumor microenvironment composition and intercellular communication. Beyond cancer, disease‐level single‐cell atlases have illuminated key mechanisms in infectious and degenerative diseases. During the COVID‐19 pandemic, large‐scale single‐cell atlases of patient lungs and bronchoalveolar immune cells revealed the immune landscape and molecular pathology of SARS‐CoV‐2 infection [ 329 , 330 , 331 ]. Integrating single‐cell transcriptomics with bulk RNA‐seq and proteomics further uncovered transcriptional regulators such as FOXO3A in COVID‐19‐induced fibrosis [ 332 ]. For neurodegenerative diseases, single‐cell cohort studies have similarly provided population‐level insights into Alzheimer's disease (AD). A single‐nucleus transcriptomic atlas of the brain vasculature identified diverse cellular mediators that contribute to AD risk [ 333 ]. A multiregion atlas across 48 individuals with and without AD further delineated region‐specific vulnerable cell states associated with AD pathology [ 334 ]. In a larger cohort of 437 aged individuals, a comprehensive snRNA‐seq atlas of the prefrontal cortex revealed two major cellular community trajectories distinguishing progression toward AD dementia from normative aging [ 335 ]. Complementary single‐cell QTL analyses identified cell‐subtype specific regulatory variants and causal genes underlying AD susceptibility [ 336 ]. Moreover, cross‐tissue single‐cell transcriptomic atlases have also uncovered common pathogenic mechanisms shared across distinct diseases. In particular, conserved fibroblast phenotypes have been identified across inflammatory conditions such as inflammatory bowel disease (IBD), rheumatoid arthritis (RA), and pulmonary fibrosis, highlighting shared cellular programs across tissues [ 337 ]. These disease‐level single‐cell atlases elucidate molecular circuits, immune dysregulation, and microenvironmental remodeling underlying diverse pathologies. Spatial omics provides anatomical context by localizing cell states and interactions within tissue microenvironments. Integrative single‐cell/spatial analyses have mapped disease‐specific niches, such as invasive zones in liver cancer [ 338 ], and spatially recurrent malignant states across pan‐cancer cohorts [ 339 ], illustrating how spatial architecture shapes tumor progression. Moreover, Chang et al. combined ST with scRNA‐seq to construct a multistage spatial evolution map of esophageal squamous cell carcinoma (ESCC), showing that cellular heterogeneity increased progressively and that invasive cells gradually breached the epithelial–mesenchymal interface to form a heterogeneous spatial landscape [ 340 ]. In glioblastoma, Greenwald et al. integrated ST and spatial proteomics to define 14 spatial modules (including eight malignant cell states) and confirming the coexistence of structured and unordered regions within the tumor [ 341 ]. Likewise, multimodal integration of single‐cell and ST profiling in metastatic breast cancer delineated three distinct spatial phenotypes across diverse clinicopathological backgrounds [ 342 ]. Integrative ST and scRNA‐seq also identified pathologic niches in idiopathic pulmonary fibrosis (IPF) [ 343 ] and SARS‐CoV‐2‐associated niches in the human placenta [ 344 ]. Longitudinal data construction and single‐cell atlas depiction identified dynamic molecule changes along disease progression or therapeutic interventions. A longitudinal single‐cell atlas of pediatric acute myeloid leukemia profiled diagnosis, remission, and relapse trajectories, mapping shifting cellular hierarchies [ 345 ]. Spatial proteomic profiling before and after glioma treatment revealed remodeling of tumor‐specific neoantigen regions and immune infiltration, providing direct evidence for evaluating the reshaping effect of treatment on the tumor microenvironment (TME) [ 346 ]. Furthermore, longitudinal single‐cell multiomics atlases characterized patients with COVID‐19 and tuberculosis, identifying temporal molecular changes associated with severity and recovery [ 347 ]. Cardiovascular atlases further demonstrated how cardiac cell states and intercellular communication remodel during heart failure or injury [ 348 , 349 , 350 ]. Together, these multiscale disease cohorts provide dynamic, single‐cell and spatially resolved reference frameworks that illuminate pathological mechanisms and establish essential resources for downstream computational modeling and translational research. While numerous algorithms have been developed to address general challenges in single‐cell and spatial omics analysis (summarized in Section  3 ), a subset of computational methods has been specifically designed to tackle biological and disease‐related questions. These specialized algorithms exploit intrinsic data properties and bridge the gap between methodological innovation and biomedical discovery, enabling insights that are directly relevant to understanding physiology, evolution, and pathology. Understanding the evolution of life, encompassing both interspecies and cross‐species dynamics, has long been a research focus in biology. Recent advances in single‐cell omics have facilitated the construction of cross‐species cellular atlases, revealing both conserved and divergent regulatory landscapes across organisms. For instance, GeneCompass, leveraging 120 million human and mouse single‐cell transcriptomes, uncovered conserved gene regulatory mechanisms between these species [ 302 ]. Similarly, TranscriptFormer, an AI‐driven virtual cell model trained on single‐cell data from twelve species spanning 1.5 billion years of evolutionary history, provided a scalable computational platform for systematic cross‐species evolutionary studies [ 351 ]. Together, these approaches highlight how large‐scale single‐cell datasets combined with evolutionary modeling can illuminate fundamental principles of life history and organismal diversification. In developmental biology, spatiotemporal single‐cell and spatial omics have become essential for reconstructing embryogenesis and organogenesis with cellular precision. A key analytical challenge lies in aligning spatial datasets across developmental time and integrating 3D morphodynamic information [ 352 ]. Spateo addressed this by enabling 3D reconstruction and digitization of whole‐embryo ST, revealing spatial gradients and signaling networks that drive mouse embryogenesis and asymmetrical cardiac organogenesis [ 353 ]. By integrating multisample ST data, Stereopy provided a general framework to dissect spatiotemporal cellular heterogeneity and niche‐mediated regulation, exemplified by studies of cortical and cardiac development [ 354 ]. Aging represents another biological process with inherent evolutionary significance. Single‐cell omics allows for the high‐resolution dissection of cellular and molecular changes associated with aging. Based on constructed aging clocks, Tarkhov et al. revealed that epigenetic aging is driven by both stochastic and coregulated changes through scDNA methylation and transcriptome joint analysis [ 355 ]. Single‐cell transcriptome‐based aging clocks can also reveal organ‐level aging properties. Matthew et al. revealed that heterochronic parabiosis and exercise reverse aging process in neurogenic regions with different ways [ 356 ]. Aging also characterizes the heterogeneity of human immune system function. The constructed sc‐ImmuAging based on the single‐cell transcriptome profiles of peripheral blood mononuclear cells has revealed interindividual heterogeneity during infection and vaccination [ 357 ]. Collectively, these tools exemplify how single‐cell and spatial algorithms can quantify evolutionary and developmental processes at unprecedented resolution. Deciphering disease mechanisms and translating single‐cell and spatial discoveries into clinically meaningful insights have become major drivers of computational innovation. A growing class of algorithms now connect the single‐cell or spatial profiles with clinical phenotypes, enabling the identification of phenotype‐specific molecular patterns for diseases. Methods such as Scissor [ 183 ], PENCIL [ 184 ], and ScRAT [ 185 ] enabled the discovery of cell subpopulations specific to disease prognosis and classification, including the COVID‐19 severity‐specific biomarkers and immunotherapy response‐specific cell subpopulations. Bulk and single‐cell integrative tools such as scBGDL extended these capabilities by predicting clinical subtypes and risk groups from heterogeneous data, while identifying candidate driver genes underlying these classifications [ 358 ]. SingleDeep enabled disease classification across diverse conditions, including lupus, AD, and COVID‐19 [ 359 ]. Cancer‐Finder performed accurate cancer cell annotation on single‐cell and ST, identifying gene programs enriched at the tumor–normal interface associated with patient outcome [ 360 ]. Moreover, in spatial proteomics, geometric deep learning approaches uncovered spatial motifs within the tumor microenvironment that stratify cancer prognosis [ 237 ]. Spatial multiomics derived immune scoring systems further demonstrated the translational potential of these approaches, providing robust predictors of hepatocellular carcinoma recurrence [ 361 ].HEX provided a deep learning model to generate spatial proteomics from H&E images, and by combining original H&E image with virtual spatial proteomics for patient outcome prediction, the method identified spatial niches for therapeutic response prediction of lung cancer in a low‐cost way [ 362 ]. Individualized treatment was also facilitated by emerging single‐cell algorithms. Computational frameworks were developed to optimize personalized combination therapy design [ 363 ], prioritize multitargeting treatment for coinhibition of malignant clones [ 364 ], and predict patient‐specific responses to treatment [ 365 ]. Multiscale PHATE extracted multimodal cellular signatures predictive of clinical outcome, as shown in COVID‐19 datasets [ 366 ]. Evolutionary modeling tools reveal dynamic clonal behavior during disease progression. For example, SPRINTER [ 367 ] revealed widespread clone proliferation heterogeneity when applied to a longitudinal, primary‐metastasis‐matched dataset of non‐small cell lung cancer (NSCLC), while SCEVAN [ 368 ] reconstructed clonal substructure and spatial evolutionary trajectories in brain tumors. An emerging frontier involves quantifying causal effects and treatment perturbations directly at single‐cell resolution. Perturbation‐aware models such as CINEMA‐OT employed causal optimal transport to reveal mechanisms of environmental exposure, for example, diminished antiviral responses following cigarette‐smoke stimulation [ 369 ]. scCausalVI employed causality‐aware generative modeling to identify treatment‐responsive cell populations and their underlying molecular programs in infectious disease [ 370 ]. Together, these computational frameworks link single‐cell and spatial data to concrete biomedical applications, spanning fundamental biology to clinical translation. Single‐cell and spatial omics have facilitated the identification of cellular states, disease‐associated niches, and molecular circuits that directly inform clinical decision‐making. By linking high‐resolution molecular phenotypes to patient‐level phenotypes and outcomes, these approaches enable biomarker discovery, precise disease classification, therapy response prediction, and personalized treatment strategies across diverse disease contexts (Table  4 ). Beyond cancer and infectious diseases, recent studies across autoimmune disorders, cardiovascular disease, kidney disease, and chronic inflammatory conditions further demonstrate the translational potential of single‐cell and spatial omics profiling. Main discoveries of single‐cell and spatial omics for clinical translation. Traditional bulk‐level biomarkers mask critical cellular heterogeneity, whereas single‐cell and spatial omics technologies can mine not only gene markers, but also cell states or spatial structural features for diagnosis and prognosis prediction or providing references for clinical treatment plan selection. Single‐cell omics is an effective tool in identifying immuno‐oncology biomarkers for predicting the therapeutic response of cancers [ 392 ], suggesting important clinical values [ 393 ]. In cancer, spatial niches were identified as potential biomarkers for clinical diagnosis and prognosis. Chen et al. discovered an immune niche consisting of antitumoral macrophages, CD8 + T cells and natural killer T cells (MT 2 ) for predicting survival in patients with small cell lung cancer [ 394 ]. Liu et al. revealed that the abundance of specific cancer‐associated fibroblast (CAF) subtypes correlates with patient survival in a pan‐cancer analysis, highlighting their potential as prognostic biomarkers [ 384 ]. Chang et al. identified a “CAF‐Epi” niche in ESCC for predicting the survival periods of patients in different risk groups [ 340 ]. Spatial biomarkers can also predict the immunotherapy response of cancer. For example, topographical differences between immune cells and tumor cells predicted PD‐1 blockade response of cutaneous T cell lymphomas (CTCL) [ 382 ], and fractions of proliferating CD8 + TCF1 + T cells and MHCII + cancer cells distinguished the difference in immune checkpoint blockade (ICB) response of triple‐negative breast cancer [ 383 ]. These spatial level or single‐cell level information identified potential biomarkers for cancer, going beyond the molecule‐based biomarkers defined by bulk‐level omics. Furthermore, by predicting cell types from histopathology image and performing cell phenotyping analysis, study suggested the value of histopathology image as a cost‐effective alternative in characterizing tumor microenvironments and identifying biomarkers for immunotherapy response of NSCLC, which significantly motivated the clinical translation of single‐cell spatial biomarkers [ 395 ]. Beyond cancer, single‐cell or spatial biomarkers have shown strong clinical potential across diverse diseases. In IBD, inflammatory fibroblast populations predicted both disease severity and anti‐TNF therapy response [ 371 ], while a unique cellular module, GIMATS consisted of cell subpopulations was associated specifically with resistance to anti‐TNF treatment [ 372 ]. In systemic lupus erythematosus (SLE), the interferon signaling (IFN) signature and plasmablast signature were associated with flare risk [ 373 ]. In kidney diseases, the gene signatures of a fibrotic microenvironment characterized by injured tubule cells predicted severe disease status [ 389 ]. Similarly, in IPF, spatial omics identified distinct fibrotic niches in diseased lungs [ 390 ]. In cardiovascular disease, single‐cell profiling of dilated cardiomyopathy and normal individuals identified disease‐associated cell states, which are potential biomarkers for heart failure [ 386 ]. Single‐cell and spatial analyses also identified immune cell populations as important markers for chronic and infectious respiratory diseases [ 396 ]. Despite these advancements, the translation of single‐cell biomarkers into routine clinical assays requires standardization, large‐cohort validation, and platform harmonization. Nevertheless, these biomarkers establish a foundation for next‐generation diagnostics that incorporate cellular context and spatial architecture. By resolving cellular ecosystems and tissue architecture, single‐cell and spatial omics overcome the limitations of traditional bulk sequencing and enable precise differentiation of disease subtypes and tailored matching of treatment plans. In cancer, this approach enabled the stratification of patients into different immune subtypes [ 385 ] and immune archetypes based on the composition of tumor immune microenvironment, guiding immunotherapy decision‐making. Single‐cell profiling also improved subtype classification in NSCLC [ 397 ], papillary thyroid carcinoma [ 398 ] and nasopharyngeal carcinoma [ 399 ], enabling tailored treatment strategies. Moreover, analysis of single‐cell transcriptome has identified ecotypes with different cellular states to stratify patients in breast cancer [ 400 ] and even pan‐cancer datasets [ 401 ]. Importantly, this paradigm extends to noncancer diseases, including autoimmune diseases. For example, multiplexed scRNA‐seq (mux‐seq) identified two molecular subtypes occupying different components of cell populations for SLE [ 374 ]. The accumulation of fragmented single‐cell studies and bulk omics studies further refined molecular subtypes for autoimmune diseases, which were discussed in this review [ 402 ]. These molecular subtypes are relevant to distinct treatment strategies, for example, corticosteroids would be a potential treatment for the neutrophil subtype of SLE, while the infliximab may target the inflammation subtype [ 402 ]. Spatial omics further refines subtype classification by mapping microenvironmental interactions. Spatial multiomics analysis revealed four distinct spatial CAF subtypes across cancer types and omics platforms, whose abundance and composition were associated with specific TME features and clinical outcomes [ 384 ]. Spatial omics analysis also provided evidence supporting existing cancer subtyping schemes. Some ST analysis identified the heterogeneity among histologic subtypes and molecular subtypes for cancer [ 403 , 404 ], suggesting the potential role of spatial omics in refining molecular subtyping of cancer. As a result, single‐cell and spatial omics provide actionable frameworks for patient stratification and personalized therapy selection across malignant, infectious, and autoimmune diseases. Single‐cell and spatial omics are accelerating therapeutic discovery by identifying targetable cell states, reconstructing drug–response landscapes, and enabling high‐resolution pharmacology. The development of targets targeting abnormal activated pathways or mutated genes in tumor cells has high specificity with experimental validation. Gao et al. screened out tumor neoantigens by single‐cell analysis on circulating tumor cells, which can significantly inhibit the growth of mouse transplanted tumors [ 381 ]. Marks et al. discovered that specific signaling pathways were activated in drug‐resistant AML cells [ 405 ]. Targeting intercellular signaling pathways in the TME can shape TME for therapeutic values. Chen et al. discovered that specific signaling axes in PDAC promoted neural infiltration, and inhibiting this signaling axis could reduce the proportion of related cells and the rate of neural invasion [ 406 ]. By integrating GWAS and single‐cell transcriptome data, the apoptosis‐related gene STK24 was identified as a therapeutic target for lung adenocarcinoma [ 378 ]. Beyond cancer, this pattern was also extended to other diseases. For example, one study utilized spatial proteomics to identify JAK/STAT and IFN signaling pathways for driving a lethal skin disease, which was validated by in vivo experiments involving JAKi inhibition treatment on diseased mouse models [ 391 ]. Single‐cell and spatial omics technologies demonstrate unique advantages in deconstructing host immune responses to inform vaccine and therapeutic antibody development for infectious diseases. For instance, in RSV research, single‐cell transcriptomic profiling of infected cotton rats systematically mapped the landscape of host factors and inflammatory pathways involved in infection and recovery, providing a foundational resource for identifying novel targets for antiviral intervention [ 387 ]. In the study of SARS‐CoV‐2, researchers identified multiple effective neutralizing antibodies from a large number of antigen‐binding IgG1+ clonotypes by performing HT scRNA and VDJ sequencing on antigen‐enriched B cells from recovered patients [ 388 ]. Single‐cell and spatial omics analyses also played important roles in validation of treatment efficacy and biomarkers development in preclinical models and clinical trials (Table  5 ). Relevant clinical trials including observational, interventional Phase I and Phase II studies, have been employed to evaluate the efficacy of new therapies or identify cellular biomarkers for disease classification. Specifically, scRNA‐seq was used to identify changes in the immune microenvironment in patients following therapeutic interventions [ 17 ] and was applied to evaluate the efficacy of adjuvant multiple myeloma vaccines [ 377 ], BRAF inhibitors and immunotherapy for colon cancer [ 375 ], chemotherapy combined with immune‐checkpoint inhibitors for lung cancer [ 379 ], and neoadjuvant immunotherapy for cutaneous squamous cell carcinoma [ 407 ]. Beyond cancer, single‐cell TCR/BCR combined with scRNA‐seq also revealed treatment mechanism for autoimmune diseases, including RA [ 408 ] and autoimmune hemolytic anemia [ 409 ]. scRNA‐seq was also applied in observational clinical trials, where identified biomarkers for disease diagnosis [ 410 ] and treatment efficacy [ 411 ]. Applications of single‐cell and spatial omics in preclinical animal experiments and clinical trials. In addition to these clinical trials, some studies remain at the preclinical stage. In preclinical animal models, analyses using scRNA‐seq, spatial RNA‐seq, and multiplex proteomics have identified potential mechanisms of disease treatment, including chemotherapy in neuroblastoma [ 412 ], anti‐PD‐1 therapy for HCC [ 413 , 414 ], TIGIT inhibition combined with BCMA‐CAR‐T therapy in multiple myeloma [ 415 ], prolactin receptor‐targeted therapy in adenomyosis [ 416 ], and claudin‐1 based targeted therapy in primary sclerosing cholangitis [ 417 ]. While performing the preclinical animal experiments is laborious and costly, AI‐driven virtual cell was proposed as another avenue to intelligently perform preclinical testing and evaluation on drug response and patient progression [ 418 ]. The realization of this concept would greatly decrease the cost in experimental design and accelerate the clinical translation of novel treatments. Moreover, ongoing efforts on developing algorithms to predict drug response from single‐cell profiles are promoting the drug discovery and screening process [ 182 , 419 , 420 ]. Technologies include multiplex scRNA‐seq pharmacotranscriptomics [ 421 ] and multimodal pharmacogenomics such as single‐cell EpiChem [ 422 ] have facilitated the measurement of drug–target engagement on single‐cell level. However, the role of spatial omics in drug discovery is still limited, which required further methodological developments. Collectively, these advances position single‐cell and spatial omics as indispensable tools for therapeutic development, enabling rational drug design, combination therapy optimization, and precision pharmacology across a wide spectrum of diseases.

Xiaoping Cen: data curation and writing – original draft. Xiaolan Huang: data curation and writing – original draft. Enjin Deng: data curation and writing – original draft. Xue Gong: investigation and writing – review and editing. Na Tan: investigation and writing – review and editing. Ye Jifeng: investigation. Yin Wang: writing – review and editing. Roland Eils: writing – review and editing. Qun Luo: writing – review and editing. Yixue Li: supervision, project administration, and funding acquisition. Fangfang Qu: supervision, conceptualization, writing – review and editing, and funding acquisition. All authors approved the final manuscript.

The authors have nothing to report.

This work was supported in part by the National Key R&D Program of China (2023YFF1204701 and 2025YFE0126600), Prevention and Control of Emerging and Major Infectious Diseases‐National Science and Technology Major Project (2025ZD01901900), National Natural Science Foundation of China (82400622 and 12371485), Guangdong Basic and Applied Basic Research Foundation (2025A1515011597), Major Project of Guangzhou National Laboratory (GZNL2025C01013 and SRPG22007), Young Scientists Program of Guangzhou Laboratory (QNPG23‐01), and Startup Program of Guangzhou National Laboratory (YW‐YFYJ0101).

Single‐cell and spatial omics have rapidly evolved into transformative technologies for dissecting developmental processes, tissue organization, and disease mechanisms at cellular and subcellular resolution. Despite these advances, several key challenges remain before these platforms can be deployed broadly across biological research and translational medicine. Technological limitations remain a major bottleneck. Despite substantial improvements, current platforms still face constraints in cost, throughput, and comprehensive molecular detection. Long‐read technologies offer a promising avenue by enabling full‐length transcript and epitranscriptome profiling at single‐molecule resolution [ 423 ]. These capabilities are crucial for resolving RNA editing, isoform diversity, and regulatory variation in health and disease. Meanwhile, next‐generation spatial multiomics technologies must balance affordability with sensitivity; innovations such as low‐cost barcoding strategies (e.g., MiP‐seq) point toward more scalable platforms [ 424 ]. Continued development of integrative assays that coprofile genome, epigenome, transcriptome, and proteome within the same cell or tissue context will be essential for achieving a mechanistic understanding of molecular regulation. Equally important are emerging pharmacogenomics [ 425 ] and single‐cell pharmacotranscriptomics frameworks [ 421 ], which will facilitate cell‐type resolved drug discovery and accelerate precision therapeutics. Methodological challenges persist on the computational side. Although a variety of computational algorithms and tools now exist for single‐cell and spatial data, including clustering, denoising, multimodal integration, spatial domain analysis, and phenotype prediction, their performance across diverse tissues, diseases, and experimental conditions remains uneven. Benchmarking studies often provide discordant recommendations, reflecting rapid algorithmic turnover and inconsistent evaluation standards [ 426 ]. The limited interpretability of deep learning‐based approaches further complicates biological insight. Recent advances in interpretable architectures have begun to distill biologically meaningful features from single‐cell and spatial gene profiles [ 427 , 428 ], yet broader validation across disease systems is required. In the future, these methods should be applied on more datasets and diverse biomedical contexts to drive the biomedical discovery. Foundation models trained on massive single‐cell and spatial atlases introduce a powerful data‐driven paradigm by enabling unified embeddings, cross‐modal prediction, and improved generalization [ 144 ]. Quantum computing is also emerging as a potential future direction for spatiotemporal single‐cell analysis and cell‐based therapeutics [ 429 ], but its practical utility remains to be established. Future work, however, must determine how these models can be adapted and fine‐tuned for specific biological questions, disease settings, and clinical decision‐making. Importantly, algorithms on associating single‐cell or spatial profiles with disease or biology phenotypes have improved our understanding of genotype‐phenotype relationships and helped identify phenotypic‐specific molecular alterations. While more and more algorithms are developing, computational advancements in single‐cell and spatial omics analysis should serve for biomedical applications in the future. Biological interpretation and translational application constitute the next frontier. Although extensive atlases have been generated across organs, developmental stages, and species, most atlas efforts remain dominated by transcriptomics, with limited incorporation of epigenomic, proteomic, metabolic, or functional perturbation data. Approaches such as single‐cell/spatial CRISPR screening provide mechanistic insight into genotype–phenotype relationships, enabling causal inference rather than correlation [ 430 , 431 , 432 ]. Future studies should focus on integration of existing atlas to enable cross‐organ, cross‐system, and cross‐species analyses, thereby generating a more comprehensive view of human biology system and providing insights for disease treatment. Translational application has lagged behind basic discovery; few molecular insights have progressed into clinical trials. More in‐depth analysis should be performed to identify effective biomarkers, disease subtypes and treatment strategies for registering clinical trials. Besides, histopathology image is becoming a promising cost‐effective alternative for spatial omics in some scenarios, including biomarker discovery, disease subtypes and even performing personalized treatment, which would accelerate the clinical translation of single‐cell omics or spatial omics [ 268 ]. However, methodological improvements and clinical cohort validation are required to further establish the reliability of histopathology as a surrogate for spatial omics. With the accumulation of data and methods, the development of cross‐scale multimodal foundation models is highly promising. These models can effectively link single‐cell spatial molecular information with variable clinical phenotypes, thereby providing high‐resolution disease genotype‐phenotype associations and promoting the development of personalized medicine. Together, addressing these technological, computational, biological, and translational challenges will determine how single‐cell and spatial omics evolve from powerful discovery tools into foundational components of clinical research, precision diagnostics, and targeted therapeutics.

Recent years have witnessed major advances of single‐cell and spatial omics technologies and computational methods, which have greatly enriched the understanding of human biology and diseases of all organs and systems. Nevertheless, the successful integration of massive, heterogeneous cross‐modal datasets and the development of robust, scalable models to interpret them remain critical bottlenecks for the field. Future efforts on developing biology‐guided multiomics, cross‐modal sequencing technologies, and advanced data analysis are thus essential to provide in‐depth understandings of human biology at multiple resolutions, further accelerating clinical translation in personalized disease treatment.

Single‐cell and spatial omics technologies have transformed the study of human biology by enabling molecular profiling at single‐cell resolution and within intact tissue architecture. Unlike conventional bulk sequencing techniques that measure averaged signals across heterogeneous populations, single‐cell approaches characterize distinct cell types, transitional states, lineage relationships, and regulatory programs across the genome, epigenome, transcriptome, proteome, and metabolome [ 1 ]. Spatial omics technologies further address the limitations of single‐cell sequencing by integrating molecular profiling with spatial localization, providing a map for molecular data directed to tissue architectures [ 2 ]. Together, these technologies provide a multidimensional view of complex biological systems and have become essential tools for understanding development, homeostasis, and the cellular basis of disease. Technological advances over the past decade have rapidly expanded the modalities, throughput, and resolution achievable in single‐cell and spatial profiling [ 3 ]. Improvements in microfluidics, barcoding chemistry, imaging, sequencing, and mass spectrometry (MS) now support deep transcriptomic, genomic, epigenomic, proteomic, and metabolomic characterization at scale [ 4 ]. Spatial platforms have progressed from multicellular spot‐level measurements to subcellular transcriptomic mapping, high‐plex protein imaging, and emerging spatial epigenomic assays [ 5 , 6 ]. Joint‐profiling technologies increasingly allow simultaneous measurement of multiple molecular layers within the same cell or tissue region, enabling the reconstruction of regulatory relationships that cannot be inferred from single‐modality data alone [ 1 , 2 ]. These innovations collectively generate complex, high‐dimensional datasets that require dedicated computational frameworks for accurate interpretation. In parallel with experimental progress, computational analysis methods for single‐cell and spatial omics have expanded substantially. Classical tasks, including quality control (QC), normalization, dimensionality reduction, and clustering, remain foundational, but analytical demands now extend to tasks such as multimodal data integration, trajectory inference, gene regulatory network inference, cell–cell communication modeling, and cell niche identification [ 7 , 8 , 9 , 10 , 11 ]. Significantly, machine learning and deep learning approaches, ranging from matrix factorization and graph neural networks to variational autoencoders and transformer‐based architectures, have improved robustness to noise and enhanced cross‐modality inference [ 12 , 13 ]. Moreover, recently developed foundation models trained on large‐scale cellular atlases provide unified representations across tissues and platforms, enabling transfer learning and cross‐modal prediction and offering new opportunities for biological discovery and translational research [ 14 , 15 ]. These combined technological and computational developments have accelerated applications across multiple species, organ systems, and biomedical domains, spanning the development of life, oncology, immunology, neuroscience, cardiometabolic diseases, infectious diseases, and related fields, revealing disease‐associated cellular phenotypes, regulatory circuits, and microenvironmental structures inaccessible to bulk assays [ 16 ]. Such insights facilitate biomarker discovery, therapeutic target identification, and mechanistic understanding of treatment response or resistance, and are increasingly being explored in preclinical and clinical contexts [ 17 ]. Despite these advances, challenges remain in data sparsity, integration across platforms, achieving higher spatial resolution for additional modalities, and establishing standards for clinical translation. In this review, we synthesize recent advances in single‐cell and spatial omics technologies, as well as computational and artificial intelligence (AI)‐driven analytical frameworks. We highlight their applications in human biology and disease across diverse organ systems and discuss key challenges and future directions for enhancing clinical and translational impact. The review is structured in four sections. First, we summarize breakthrough technologies for single‐cell and spatial sequencing across multiple molecular layers, including genome, epigenome, transcriptome, proteome, and metabolome. Second, we discuss computational advancements, reviewing state‐of‐the‐art methods for mono‐omics analysis, multi‐omics integration, and cross‐modal/cross‐scale data integration, with a focus on AI‐based algorithms and foundation models. Third, we summarize representative applications of these technologies in human biology and disease, illustrating how they drive biomedical discoveries. Finally, we critically discuss remaining challenges and future directions for advancing biomedical discovery and translation using single‐cell and spatial omics.

The authors declare no conflicts of interest.

The generation of massive single‐cell and spatial multi‐omics data necessitates the development of a variety of algorithms to tackle the sparsity and high dimensional issues of data. Compared with traditional methods that rely on experience and parameter tuning, AI especially deep learning suggests powerful methods in dealing with these data issues, providing efficient tools for data imputation, dimension reduction, and clustering [ 12 ]. AI can also overcome the heterogeneity among different layers of data, facilitating the effective integration of multi‐omics, cross‐modal data, and linking cell‐level or spatial‐level omics data with patient‐level phenotypes. Recent foundation modeling techniques and agents further extended the power of AI, facilitating the analysis of multiple tasks in an effective, efficient and automated way [ 144 ]. These recent breakthroughs in single‐cell and spatial omics analysis have been summarized in the following subsections, along with Figure  3 and Tables  1 and 2 . AI‐based algorithms for single‐cell and spatial omics analysis. Existing AI‐based algorithms can be classified into six types according to their function. (1) Data denoising and imputation: some algorithms performed data denoising directly on the single‐cell omics data, while others performed imputation from other types of single‐cell omics data. (2) Dimension reduction and clustering: autoencoder also facilitated the cell type clustering or spatial domain identification by learning latent representation from single‐cell or spatial omics data. (3) Multi‐omics integration and Inference: multimodal autoencoders enabled multi‐omics integration and identified cross‐modal interaction. (4) Cross‐modal integration and inference: integrated spatial omics with histopathology image using graph‐based neural network improved spatial domain identification. Transformer provided cross‐modal inference from histopathology images to spatial omics. (5) Phenotype prediction: transformer linked single‐cell omics with clinical phenotypes and facilitated the phenotype‐specific subpopulation identification. (6) Foundation model: developing foundation models on large‐scale single‐cell or spatial omics data facilitated multiple downstream tasks by pretraining and fine‐tuning paradigm. Algorithms for single‐cell and spatial omics analysis. Foundation models for single‐cell and spatial omics analysis. Single‐cell genome analysis provides a direct view of genomic alterations at cellular resolution, for understanding clonal dynamics, genomic instability, and evolutionary trajectory, yet remains constrained by low sequencing coverage. This challenge has motivated the development of two major categories of methods. The first reconstructs copy‐number alterations (CNAs) directly from scWGS profiles. For example, HiScanner [ 145 ] enables CNA detection from shallow scDNA data, while SCICoNE reconstructs CNA histories in tumor samples using statistical modeling and MCMC sampling [ 146 ]. MEDALT further infers lineage trajectories to identify fitness‐associated alterations and genes [ 147 ]. A second group of methods infers SNPs or CNAs from nongenomic single‐cell modalities, including transcriptomic [ 148 ] and chromatin accessibility profiles [ 149 ]. SComatic [ 150 ] and Cellsnp‐lite [ 151 ] are common methods for detecting SNPs from scRNA‐seq or ATAC‐seq data. Approaches such as Numbat [ 148 ] and CopyKAT [ 152 ] yield better performance in extracting CNA‐like signals from scRNA‐seq data in a benchmarking study [ 153 ], with CopyVAE [ 154 ] extending this strategy using variational autoencoders to improve robustness. As scDNA‐seq protocols continue to evolve but coverage remains limited, these inference and lineage‐reconstruction models provide practical strategies for characterizing genomic heterogeneity. Continued benchmarking and methodological refinement will be crucial for single‐cell genome analysis, especially for reliably reconstructing clonal evolution reliably at cellular resolution. Single‐cell transcriptome analysis has become the most extensively employed approach for dissecting cellular heterogeneity. A comprehensive ecosystem of software packages and pipelines, including Seurat [ 155 ], Scanpy [ 156 ], Bioconductor‐based workflows, and command‐line toolkits, provides end‐to‐end analysis of scRNA‐seq data, covering QC, normalization, dimensionality reduction, clustering, and visualization of scRNA‐seq data [ 157 ]. These frameworks have become standard platforms for single‐cell analysis and are continuously extended to support multimodal assays and large‐scale atlas datasets. A central challenge in scRNA‐seq analysis is the high sparsity and noise of count matrices, caused by limited mRNA capture efficiency, amplification bias, and stochastic gene expression. Numerous methods have been proposed to denoise or impute scRNA‐seq data based on deep generative models and graph‐based modeling, such as SmartImpute [ 158 ], scVGAMF [ 159 ], DCA [ 160 ], and GraphSCI [ 161 ]. Deep generative models like DCA adopt variational autoencoder architectures with count‐based likelihoods (e.g., ZINB) to jointly model noise and biological variation. Meanwhile, GraphSCI [ 161 ] simultaneously learns gene–gene relationships and cell–cell correlations, enabling improved imputation by integrating gene co‐expression networks and cell‐specific expression profiling into a stacked graph convolutional networks (GCNs)‐autoencoder model. Besides, SAVER‐X [ 162 ] used a transfer learning scheme to learn transferable gene–gene relationships from different datasets to achieve data denoising for noisy scRNA‐seq data while realize ample usage of existing data. Dimensionality reduction and clustering are crucial steps for identifying cell populations and transcriptional programs [ 163 ]. Classical methods such as PCA and graph‐based clustering have been complemented by deep learning and manifold‐learning algorithms, which can better capture continuous trajectories and non‐linear structures in the data. Methods such as scVI [ 164 ], VASC [ 165 ], scvis [ 166 ], and scDHA [ 167 ] used autoencoder‐based models to learn low‐dimensional representations that preserve complex manifolds while mitigating technical noise and batch effects. While traditional clustering methods cannot flexibly regulate the process to achieve optimal results, scDeepCluster [ 168 ] was proposed by applying an autoencoder model, which integrated clustering loss with ZINB loss to perform clustering, and scDSC [ 169 ] incorporated GNN‐based structural information among cells to improve clustering quality. Knowledge‐guided models such as d‐scIGM incorporated domain constraints to generate biologically meaningful clusters [ 170 ] and another unsupervised deep model jointly optimize clustering and batch correction [ 171 ]. Cell‐type annotation, which is traditionally performed by manually matching cluster‐specific markers to known cell types, is increasingly automated by supervised and reference‐based methods. Tools such as BERMUDA [ 172 ], scScope [ 173 ], MARS [ 174 ], and scArches [ 175 ] employed transfer learning or deep metric learning (DML) to map new datasets onto existing reference atlases while correcting for batch and platform variation. Specifically, scScope supported scalable analysis of millions of noisy scRNA‐seq profiles and facilitates rapid identification of cell composition [ 173 ]. To discover novel cell types, MARS [ 174 ] used a deep learning scheme to learn from different datasets to perform cell‐type annotation on unannotated scRNA‐seq data. mtSC also leveraged multiple reference datasets by a multitask DML model to improve annotation reliability across studies [ 176 ]. While these methods are still time‐consuming and required reference datasets, GPTCelltype enabled efficient and automated cell type annotation using chain‐of‐thought prompt learning of GPT‐4 with comparable performance [ 177 ]. Trajectory and state‐transition inference remain essential for understanding dynamic biological processes. RNA velocity [ 178 ] and improved frameworks such as scVelo [ 179 ] estimated future transcriptional states by leveraging spliced and unspliced transcripts. For longer‐timescale developmental or lineage trajectories, probabilistic and deep‐learning approaches were developed. VITAE [ 180 ] combined a hierarchical mixture model with variational autoencoders to infer branching trajectories, whereas PRESCIENT [ 181 ] applied a deep generative framework to reconstruct cell‐state transitions and lineage structure. Besides, there are other tools for perturbation response prediction and disease state prediction for single‐cell data. For example, CellOT leveraged optimal transport theory and convex neural architectures to develop a framework for predicting the response of cells to given perturbation [ 182 ]. Some algorithms were designed to associate the scRNA‐seq data with phenotypic information. Scissor was developed by using the similarity between individual single cells and bulk samples as a constraint in a regression model to identify relevant subpopulations [ 183 ]. PENCIL was developed to perform simultaneous function of selecting informative features and identifying cell subpopulations by deploying a Learning with rejection strategy [ 184 ]. ScRAT applied attention‐based neural network to quantify the cell–cell interaction and the contributions of each cell in phenotype prediction [ 185 ]. scPhase utilized an attention‐based multiple instance learning to learn a patient‐level representation, which predicted clinical phenotypes across multiple patient cohorts [ 186 ]. Finally, quantum computing has been explored as an alternative computational paradigm. Quantum annealing strategies have been applied to optimize clustering cost functions [ 187 ], capture non‐linear gene interaction patterns for feature selection [ 188 ], and simulate gene regulatory interactions for network inference [ 189 ]. While these approaches are still exploratory, they suggest potential directions for future research alongside classical algorithms. The wide range of computational methodologies developed for single‐cell transcriptome have not been expansively described in this section due to the limited scope. More detailed description for these methods can be referred to the specialized reviews [ 190 , 191 , 192 ]. Single‐cell epigenomic assays, including ATAC‐seq, DNA methylation, and single‐cell Hi‐C, capture chromatin accessibility, regulatory state, and 3D genome structure. Computational workflows typically involve peak calling or bin‐based quantification, followed by dimensionality reduction, clustering, and inference of regulatory motifs or transcription factor activity. As these datasets are markedly sparser and noisier than scRNA‐seq, their analysis requires methods tailored to modality‐specific sparsity and coverage constraints. EpiScanpy extended the Scanpy framework to single‐cell epigenomic data, addressing modality‐specific issues via multiple feature‐space constructions and nearest‐neighbor graph building [ 193 ]. For chromatin accessibility, SnapATAC supported large‐scale single‐cell ATAC‐seq analysis and trajectory reconstruction [ 194 ], and SnapATAC2 further generalized to multiple single‐cell modalities [ 195 ]. Benchmarking studies suggested strong performance of SnapATAC and SnapATAC2, particularly for large‐scale chromatin datasets with complex cell‐type structure [ 196 ]. Single‐cell Hi‐C data measure cell‐level 3D chromatin organization, contributing to the analysis of transcription process varying between different cell types. The imputation and embedding of single‐cell Hi‐C data is challenging as its 2D contact map adds complexity in data structure [ 197 ]. Studies have applied advanced deep learning methods in embedding single‐cell Hi‐C data. Higashi used hypergraph representation learning to capture the latent structure in 2D contact map [ 198 ], suggesting a practical tool for imputation and embedding of single‐cell Hi‐C data. scDEC‐Hi‐C combined a chromosome‐wise convolutional autoencoder with a deep embedding and clustering model to improve imputation and cell clustering [ 199 ]. Comparative analyses indicated that BandNorm, scHiCTools, Higashi, FastHigashi, and scVI‐3D all showed competitive performance in imputation and embedding tasks [ 200 ]. To address limited scDNA methylation data, scHiMe inferred methylation profiles from single‐cell Hi‐C using joint modeling [ 201 ]. Integrative modeling of multiple epigenomic modalities is an active area of research. Approaches such as ConvNet‐VAEs attempted to combine HM and chromatin accessibility data using one‐dimensional convolutional variational autoencoders [ 202 ]. Continued development is required to integrate diverse epigenomic modalities into unified regulatory representations. Single‐cell proteomics provides direct measurements of protein abundance and posttranslational states. Computational analysis of these data involves preprocessing (including imputation, transformation, and normalization), feature selection, dimensionality reduction, and clustering, often using graph‐based or density‐based algorithms adapted from scRNA‐seq analysis. Tools such as SCPline provide streamlined preprocessing pipelines tailored to low‐input proteomic datasets [ 203 ]. Deep learning frameworks, including scPROTEIN, used contrastive learning to derive robust embeddings from multiple single‐cell proteomics datasets [ 204 ]. PINNACLE used a geometric deep learning approach, which aimed to model protein interactions by generating context‐aware protein representations based on cell‐type information from scRNA‐seq data [ 205 ]. These methods offer potential for dissecting disease mechanisms and protein‐level regulation at cellular resolution. SCM identifies the small molecules involved in the metabolic activity of cells, providing supplements for single‐cell proteomics. Methods such as SCMeTA provide preprocessing and analytical pipelines for MS‐based SCM data [ 206 ]. Given the limited availability of single‐cell proteomics and metabolomics data, several methods infer protein or metabolic states directly from scRNA‐seq. BABEL was among the first to perform cross‐modal translation between single‐cell omics using multiple autoencoders [ 207 ]. However, benchmarking suggests that TransferData in Seurat v3 and v4 often performs competitively for predicting surface protein levels from transcriptomes [ 208 ]. For metabolism, scMetabolism quantifies metabolic pathway activity from scRNA‐seq profiles, while GEFMAP used geometric deep learning to predict metabolic states from gene expression, enabling approximation of metabolomic profiles when direct assays are unavailable [ 209 ]. Spatial genomic analysis aims to reconstruct allele‐specific CNAs, subclonal structure, and tumor evolutionary dynamics directly from spatially resolved tissue sections. Although current spatial DNA assays remain limited by low coverage and technical constraints, several computational tools have attempted to infer CNAs from Spatially resolved transcriptomics (SRT) or imaging‐based genomic measurements. CalicoST reconstructed allele‐specific CNAs and spatially informed evolutionary trajectories by integrating spatial coordinates with transcript‐derived genomic signals [ 210 ]. SlideCNA applied expression‐aware spatial binning for overcoming sparsity limitations of CNAs data while maintaining spatial signal to recover CNA patterns and demonstrates its potential for spatial subclone detection [ 211 ]. However, these inference methods have not been fairly benchmarked, which requires further investigation. Methods developed for spatial epigenome analysis are limited. Most of the analyses of spatial ATAC data are performed by methods in manipulating single‐cell ATAC data, including ArchR [ 212 ] and SnapATAC2 [ 195 ]. spaPeakVAE was extended from spaVAE, which captured spatial dependencies of signals and enabled effective analysis in multiple tasks [ 213 ]. The analysis of spatial epigenome is more common in integrative analysis of multiple spatial omics layers. Spatial transcriptome analysis encompasses the computational workflows used to interpret spatially resolved gene expression profiles and to delineate tissue organization at high resolution. Standard analytical pipelines typically include data preprocessing, spatial feature construction by integrating gene expression with spot coordinates and histology, dimensionality reduction, and downstream tasks such as spatial domain identification, detection of spatially variable genes (SVGs), and inference of spatial trajectories or cell–cell interactions [ 214 ]. A number of general‐purpose tools, such as Seurat, Scanpy, Giotto, and Squidpy, provide integrated frameworks for preprocessing, visualization, and exploratory analysis of ST datasets. Spatial domain identification is an essential task of ST analysis, aiming to organize molecular observations together with their spatial context into coherent tissue regions. Numerous specialized methods have been developed to enhance the resolution and biological interpretability of these spatial partitions, including BayesSpace [ 215 ], stLearn [ 216 ], SpaGCN [ 217 ], STAGATE [ 218 ], and GraphST [ 219 ]. BayesSpace applied a Bayesian statistical method that used spatial neighborhood information for clustering analysis [ 215 ]. stLearn learned robust spatial clusters of gene profiling based on a spatial graph‐based imputation method, and inferred spatial trajectory by diffusion pseudotime‐based method [ 216 ]. SpaGCN integrated gene expression, spatial proximity, and histology using GCNs to identify spatially enriched patterns [ 217 ]. STAGATE employed a graph attention autoencoder to learn latent spatial embeddings [ 218 ], whereas GraphST leveraged self‐supervised contrastive learning to derive highly discriminative spot representations [ 219 ]. Benchmarking studies across 13 methods and 34 datasets highlighted persistent challenges in detecting discontinuous spatial domains and regions with subtle transcriptional gradients [ 220 ]. Notably, algorithms were also developed to perform SVGs identification, among which SPARK‐X, a non‐parametric method for effectively detecting spatially expressed genes from large‐scale SRT data, outstood among all the benchmarking methods [ 221 ]. Moreover, integrating multi‐slice SRT enhanced the generation of biologically relevant domains by combining multi‐view graph network, contrastive learning, and attention mechanisms [ 222 ]. In subcellular‐resolution ST, cell segmentation is a prerequisite for cell‐level analyses. Methods such as SCS combined imaging data and sequencing information using transformer architectures [ 223 ], whereas Bering used graph deep learning to incorporate transcript colocalization patterns [ 224 ]. UCS integrated nuclei segmentation from staining images and spatial transcriptome data to define cell boundaries [ 225 ]. These methods have facilitated cell‐level analyses in subcellular ST. A key objective of spatial omics is the characterization of cellular niches, the localized microenvironments that shape cell states, interactions, and tissue‐level functions. CellCharter applied Gaussian mixture modeling to identify, characterize, and compare cellular niches across spatial datasets [ 226 ]. scNiche resolved niches at single‐cell resolution by extracting multi‐view cellular features [ 227 ]. NicheCompass learned interpretable cell embeddings using graph‐based neural networks to map niche structure and organization [ 228 ]. Together, these approaches provided effective frameworks for dissecting microenvironmental architecture and its functional relevance within tissues. More detailed information about spatial transcriptome analysis methods can be referred to these specialized reviews [ 13 , 229 , 230 , 231 ]. Spatial proteomics analysis aims to quantify cell phenotypes, signaling states, and microenvironmental organization from multiplexed protein measurements in tissue sections. Cell segmentation and spatial clustering are also important tasks in spatial proteomics analysis. For MSI data, HIT‐MAP [ 232 ] and SubCellBarCode [ 233 ] provided integrated pipelines for spatial proteomics analysis. For multiplexed IF (mIF) imaging data, AnnoSpat identified cell types and quantified cell–cell proximity relationships in spatial proteomics data [ 234 ]. STARLING applied a probabilistic machine learning model to quantify cell populations and cellular phenotypes from spatial proteomics data [ 235 ]. NPF was proposed to improve the resolution of sequencing‐based spatial proteomics methods [ 236 ]. Algorithms also enabled the connection of spatial proteomics with clinical phenotypes. Study has designed a geometric deep learning method and constructed spatial cellular‐graph to model tumour microenvironments from multiplexed spatial proteomics profiles [ 237 ], which facilitate spatial biomarker discovery. Computational methods for spatial metabolomics are limited. SmartGate enabled iterative peak selection and spatial structure identification for imaging MS‐based spatial metabolomics data analysis [ 238 ]. MetaVision3D facilitated the transformation of serial 2D imaging sections of brain into 3D spatial metabolome with high resolution [ 239 ]. Moreover, these analyses are limited to single‐omics levels, while integrative analysis methods should be developed to gain biological insights. While the analysis of different single‐cell omics data provides different perspective of cellular activities, the integrative analysis of single‐cell multi‐omics data offers a more comprehensive picture of cellular states and regulatory programs. Many algorithms were developed to integrate single‐cell epigenome data with single‐cell transcriptome data, to reveal how regulation in chromatin and methylation links to RNA expression. Some earlier algorithms including MOFA+ [ 240 ] projected single‐cell multio‐mics data into a shared low‐dimensional latent space by matrix factorization to capture major sources of variation across modalities. However, the integration of single‐cell multiomics data meets challenges. First, the datasets may not be measured in matched cells, bringing obstacles to single‐cell multi‐omics analysis. Some studies developed algorithms specific to the measurement of unpaired single‐cell multi‐omics data. UnpairReg performed regression analysis on unpaired single‐cell multio‐mics data by considering the biological associations between regulatory elements and between target genes, providing an accurate estimation of scRNA‐seq data where only scATAC‐seq data are available [ 241 ]. GLUE constructed modality‐guided regulatory graphs based on prior knowledge and aligns omics‐specific embeddings through adversarial learning, enabling cross‐omics regulatory inference, which even works for triple omics (RNA, ATAC, methylation) integration [ 242 ]. Monae designed multiple graph‐based autoencoders for learning cross‐omics interactions, facilitating imputation and integration of unpaired single‐cell multi‐omics data [ 243 ]. Recent work aimed to construct interpretable latent spaces by disentangling shared and modality‐specific components from single‐cell multio‐mics data. UnitedNet provided an explainable VAE‐based framework by grouping cell embeddings and performing contrastive learning on group‐level embeddings, facilitating the interpretation of group‐specific multi‐omics relevance, including RNA–protein correlation at single‐cell level [ 244 ]. scTFBridge has innovatively disentangled learned latent spaces of single‐cell multio‐mics data into shared and specific components, where the shared components are suggested to be related to specific TF regulatory activities [ 245 ]. While the complexity of single‐cell multi‐omics data requires further disentanglement, these methods have enriched the interpretation of complex molecular activities within biological systems, suggesting potential tools for biomedical discoveries. Additional methods for single‐cell multi‐omics analysis can be found in these reviews [ 246 , 247 ]. Integrative analysis of spatial multi‐omics data contributes to a more complete view of molecular regulation and biological pathways. SpatialGlue developed a graph neural network model with a dual‐attention mechanism that deciphers spatial domains by intraomics integration of spatial location and omics measurement followed by cross‐omics integration [ 248 ]. SWITCH is a deep generative model for cross‐modal translation of spatial multiomics data by graph attention networks (GATs) [ 249 ]. SpaFusion designed graph‐based autoencoders with high‐order cell graph of spatial transcriptome and spatial proteome as input to enable accurate clustering [ 250 ]. The model integrated multi‐level information, capturing both omic‐specific features and unified consensus representations across different omics data. The integration of ST and SM data meets challenges in different data distribution, while SpatialMETA constructed a conditional variational autoencoder to overcome the difficulties and to facilitate interpretation of spatially correlated ST–SM patterns [ 251 ]. In addition to improving the performance of common spatial omics tasks, spatial multi‐omics integration can also be used for downstream biomedical discovery. For example, stClinic designed a dynamic graph model that integrates spatial multi‐slice multiomics and clinical phenotype data to uncover clinically relevant niches [ 252 ]. Some other algorithms are developed to integrate spatial omics across different platforms [ 253 ] or different samples. PASTE provided cross‐slice alignment by Gromov–Wasserstein optimal transport algorithm [ 254 ]. CAST is a deep graph neural network‐based method for cross sample alignment of spatial omics at the single‐cell level [ 255 ]. These approaches lay the groundwork for comparative spatial analyses across tissues, conditions, and technologies. Moreover, additional representation strategies and comprehensive summaries are available in this recent review [ 256 ]. Integrative frameworks that link single‐cell, spatial, and bulk‐level omics enable cross‐scale characterization of molecular and cellular phenotypes in complex tissues. Among these integrative approaches, methods that characterize regulatory variation, especially single‐cell eQTL (sc‐eQTL) analyses, offer a way to resolve cell‐type specific genetic effects. For example, IBSEP improved cell‐type‐specific eQTL (ct‐eQTL) prioritization by combining bulk RNA‐seq and scRNA‐seq data [ 257 ]. EXPRESSO analyzed sc‐eQTL summary statistics by integrating 3D genomic data and epigenomic annotation to prioritize causal variants [ 258 ]. gsMap integrated ST data with Genome‐wide association studies (GWAS) summary statistics to map spatially resolved cell activity to human complex traits [ 259 ]. Together, these methods facilitated cross‐scale data analysis and provided insights for the underlying mechanism of human complex traits. ST often suffers from limited resolution, as each spot typically captures transcripts from multiple cells. To address this, deconvolution methods aim to infer cell‐type compositions and cell‐type specific expression profiles, by leveraging scRNA‐seq reference datasets. Representative methods, including RCTD, SPOTlight, cell2location, and Tangram have been systematically reviewed elsewhere [ 214 ]. More recently, graph‐based and deep learning models such as STdGCN, ST‐deconv, SDePER, and EnDecon extend these ideas by incorporating spatial neighborhood structure, contrastive representation learning, or ensemble strategies to improve accuracy and robustness in heterogeneous tissues [ 260 , 261 , 262 , 263 ]. Meanwhile, STged moved beyond traditional deconvolution by reconstructing cell‐type‐specific expression profiles directly from mixed spots, enabling fine‐grained expression modeling from low‐resolution spatial data [ 264 ]. Cross‐modal integration of spatial omics and histopathology images is another important direction. Some studies are integrating spatial omics and histopathology images to improve spatial domain identification, including MISO [ 265 ], SpaCell [ 266 ], and StereoMM [ 267 ]. Typically, MISO extracted graphs from spatial multi‐omics and histopathology image, and designed autoencoders for interpretable spatial clustering. Other studies performed cross‐modal prediction from histopathology images to spatial profilings. HistoST and THItoGene applied transformer framework for extracting morphology features from histopathology images and generate spatial transcriptomes, while BLEEP and mclSTExp performed contrastive learning for aligning the histopathology image with spatial transcriptome data [ 268 ]. The prediction can also reach single‐cell level [ 269 ]. ROSIE performed translation of histopathology to mIF data using ConvNet framework [ 270 ]. HistoPlex designed a GAN network to translate histopathology image into multiplex spatial proteomics images [ 271 ]. GigaTIME applied nested U‐Net framework to perform the similar task, which was validated on 14,256 patients with only histology data to generate virtual mIF data [ 272 ]. These cross‐modal inference methods not only enhanced the understandings of genotype‐phenotype associations, but also provided potential tools for clinical translation of spatial omics. Single‐cell and spatial omics datasets are high‐dimensional and noisy, making them well suited to representation learning. Many AI‐based analysis methods rely on transformers, graph neural networks, generative adversarial networks, or variational autoencoders to derive cell‐ and gene‐level embeddings [ 273 , 274 , 275 ]. Foundation models extend this idea by pretraining on large unlabeled datasets and then adapting to downstream tasks, potentially offering reusable representations across studies [ 144 ]. These methods not only facilitate efficient analysis of data, but also enable downstream analysis including clinical applications. For single‐cell omics, several transformer‐based foundation models have been proposed, including scBERT [ 276 ], scFoundation [ 277 ], GeneFormer [ 278 ], and scGPT [ 279 ]. These models have shown improved performance in tasks such as clustering and batch correction compared with conventional methods like Seurat. Among the models, scGPT was able to perform single‐cell multi‐omics integration, and the updated version of GeneFormer [ 280 ] was pretrained on the largest single‐cell datasets. Single‐cell foundation models are suggested to be robust tools for diverse applications, while machine learning models are efficiently adapted to specific datasets [ 281 ]. scTranslator facilitated the cross‐modal translation from single‐cell transcriptome to single‐cell proteome by pretraining on large‐scale single‐cell multi‐omics data, yielding better performance than previous methods [ 282 ]. Epiagent [ 283 ] is a foundation model for single‐cell epigenome analysis. SCimilarity [ 284 ] facilitated scalable search for similar cells by metric learning and pretraining on single‐cell atlas datasets. CellWhisperer [ 285 ] enabled interactive single‐cell data exploration based on dialogues. Besides, agents for semi‐automated analysis pipelines are also emerging. Foundation models on spatial omics are still limited. scGPT‐spatial [ 286 ] presented a foundation model for ST, which was designed to enhance SRT data analysis through continual pretraining of scGPT. KRONOS is a foundation model built for spatial proteomics, which was trained in a self‐supervised way on over 47 million image patches covering 175 protein markers, 16 tissue types, and 8 imaging platforms [ 287 ] and achieved state‐of‐the‐art performance across cell phenotyping, treatment response prediction, and retrieval tasks. Nicheformer learned cell representations that captures spatial context from spatial and single‐cell omics data based on a transformer framework, allowing the transfer of rich spatial information to scRNA‐seq data and the characterization of multi‐scale spatial niche for disease mechanism analysis [ 288 ]. OmiCLIP [ 289 ] enabled the cross‐modal inference of spatial transcriptome from histopathology images. Recently, the development of SpatialAgent [ 290 ] integrated large language models to form agent tools, which facilitated the automatic investigation of spatial omics data and experimental preparation, which innovating the spatial biology investigation. Overall, foundation models offer a promising direction for standardizing representations and accelerating downstream analyses, but questions remain regarding data curation, cross‐cohort generalizability, interpretability, and fair evaluation across tasks and platforms.