Spatial Omics Enters the Microscopic Realm: Opportunities & Challenges

Abstract

Spatial transcriptomics (ST) enables systematic profiling of whole-transcriptome gene expression in tissues while preserving spatial context. Recent advances in sequencing- and imaging-based ST technologies have ushered in the era of microscopic-resolution ST (μST), allowing transcriptome mapping at cellular and even subcellular scales with unprecedented precision. Despite these advances, μST faces substantial challenges, including sparse transcript discovery per submicron (or micron)-sized spatial units and data fragmentation across platforms hindering integration and analysis. There is also a growing demand for scalable, segmentation-free, and universally applicable analysis methods, as well as strategies for 3D mapping, multi-omics integration, and artificial intelligence (AI)-driven spatial analysis. This review highlights recent breakthroughs, outlines key challenges, and discusses emerging experimental and computational solutions shaping the future of μST.

1. Why Microscopic Resolution Matters

The structural organization of biological systems fundamentally determines their function, from cellular interactions to tissue and organ architecture. Microscopy has long been instrumental in visualizing this spatial organization, driving advances in microbiology, histology, cytology, and clinical pathology. The integration of molecular biology techniques has expanded the capabilities of microscopy, enabling the visualization of specific proteins, RNAs, and even DNA sequences within cells and tissues. While chromogenic and fluorescent labeling have provided valuable insights into gene expression and cellular function, these approaches are limited by the small number of molecular targets that can be analyzed simultaneously.

Spatial transcriptomics (ST) has emerged as a transformative technology enabling genome-wide transcriptome profiling while preserving spatial context [1]. However, early ST approaches, termed coarse ST (cST), were limited by relatively low spatial resolution (>100 μm, center-to-center distance)—comparable to or worse than human vision—restricting their ability to resolve transcriptomic features at the cellular and subcellular levels (Fig. 1A, left). Successive technological developments improved spatial resolution to a near-cellular scale (~10 μm), analogous to the resolution achieved by conventional magnification methods (magnified ST or mST; Fig. 1A, center). Most recently, microscopic-resolution spatial transcriptomics (μST, <1 μm resolution; see Glossary) has been realized (Fig. 1A, right), providing whole-transcriptome coverage at subcellular precision and uncovering spatial features undetectable by earlier cST technologies (Fig. 1B). μST facilitates detailed spatial mapping of gene expression at resolutions previously unattainable, opening new opportunities for spatially resolved molecular analysis at unprecedented scales.

Figure 1.

Exponential Advances in Microscopic Spatial Transcriptomics (μST) Technologies

(A) Why resolution matters: Early spatial transcriptomics (ST) had coarse resolution (~40 μm), similar to human vision (coarse ST, cST; left). Advances in microbead- and microfluidic-based arrays improved resolution to 10–20 μm, akin to magnification tools (magnified ST, mST; center). Recent adaptation of next-generation sequencing (NGS) arrays has enabled microscopic ST (μST; right) with micron to submicron resolution. The array schematics (middle row) depict differences in spot/pixel size and spacing. The text boxes (bottom row) illustrate how micron-scale features captured by μST may be obscured or lost at the lower resolutions of cST and mST.

(B) cST vs μST in biological applications: In a study of acne pathology, 10x Visium (cST) failed to resolve TREM2 macrophages identified by scRNA-seq, whereas Seq-Scope (μST) visualized inflamed hair follicles and associated immune cells, including TREM2 macrophages (cyan), at high resolution (adapted from [9] with permission from AAAS).

(C) sST resolution improves towards μST: Resolution (represented as center-to-center distance of pixels) of sequencing-based ST (sST) has improved exponentially via deterministic (blue), microbead-based (green), NGS-based (orange), and tissue expansion (red center) methods, roughly halving every year. Many platforms are now commercialized (red outline) or shared as open-protocol platforms (e.g., Seq-Scope, Nova-ST, Open-ST). Despite lower sensitivity for rare transcripts compared to targeted iST, the scalable, untargeted nature of sST enables broad genetic analyses beyond the reach of iST. Grey arrow shows the trend of halving center-to-center distances each year.

(D) iST scaling for the whole transcriptome: Imaging-based ST (iST), inherently μST, achieves high resolution via FISH (green)- or ISS (blue)-derived methods. However, practical time and cost constraints limit the number of detectable genes, and most commercial platforms (red outline) profile fewer targets. iST also lacks access to additional transcriptomic modalities (e.g., splicing, allele-specific expression). The gray arrow shows the trend of doubling target coverage per year.

This review examines recent breakthroughs in μST, highlighting key technological advances (Table 1). We will then examine computational challenges and solutions for μST data, including segmentation-dependent and segmentation-free analyses (see Glossary), cross-platform integration, and scalable data processing. Finally, we will discuss emerging trends in AI-driven spatial analysis, multi-omics integration, and clinical applications, envisioning a unified framework for the next generation of spatial omics research.

Table 1. Overview of Microscopic Spatial Transcriptomics (μST) Technologies.

This table summarizes μST technologies achieving ≤2 μm resolution, released in the past five years through academic publications, commercial platforms, or press release. Reported resolution reflects optical resolution (iST) or spatial array resolution (sST), adjusted for tissue expansion where applicable. Capture rate refers to untargeted transcriptome profiling and varies by tissue type and protocol and is provided as approximation only. Imaging areas are variable; representative published or commercially supported areas are provided for reference.

Technology	Class	XY Resolution	Year	Availability (Mar-1, 2025)	Number of Genes	UMI / μm2	Analysis Area	Ref
Seq-Scope-X	sST	0.18 μm	2025	Academic	Whole transcriptome	~5	2 mm x 2 mm	[19]
Illumina Spatial	sST	1 μm	2025	Pre-Market	Whole transcriptome	>30	50 mm x 15 mm	[15]
Deep-STARmap (Stellaromics Pyxa)	iST	0.32 μm	2024	Academic/Pre-Market	Up to 1,017 genes	Targeted	4.5 mm x 4.5 mm (0.1 mm Z axis)	[39]
10x Xenium 5K	iST	0.2–0.3 μm	2024	Commercial	Up to 5,000 genes	Targeted	12 mm x 24 mm	[37]
Nanostring CosMx 6K	iST	0.2–0.3 μm	2024	Commercial	Up to 6,000 genes	Targeted	20 mm x 15 mm	[38]
Singular G4X	iST	0.2–0.3 μm	2024	Pre-Market	Up to 300 genes	Targeted	Up to 40 cm2	[26]
10x Visium HD	sST	2 μm	2023	Commercial	Whole transcriptome	~2.16	6.5 mm x 6.5 mm	[4]
GenePS	iST	~0.1 μm	2023	Commercial	Up to 1,000 genes	Targeted	Up to 1 cm2	[32]
Stellaromics Plexa	iST	0.32 μm	2023	Pre-Market	Up to 4,000 genes	Targeted	Up to 225 mm2	[25]
Open-ST	sST	0.6 μm	2023	Academic	Whole transcriptome	Up to 10	3 mm x 4 mm	[13]
Nova-ST	sST	0.6 μm	2023	Academic	Whole transcriptome	~2.94	10 mm x 8 mm	[14]
10x Xenium	iST	0.2–0.3 μm	2022	Commercial	Up to ~400 genes	Targeted	12 mm x 24 mm	[24]
NanoString CosMx	iST	0.2–0.3 μm	2022	Commercial	Up to 1,000 genes	Targeted	20 mm x 15 mm	[30]
Resolve Molecular Cartography	iST	0.2–0.3 μm	2022	Commercial	Up to 1,000 genes	Targeted	Up to 26 mm2	[31]
Stereo-seq (STOmics)	sST	0.5–0.7 μm	2022	Commercial	Whole transcriptome	Up to 14.5	Up to 174.24 cm2	[11]
Pixel-seq	sST	1 μm	2022	Academic	Whole transcriptome	Up to 10	6 mm x 30 mm	[12]
Rebus Esper	iST	0.26 μm	2021	Commercial	~30 (high-fidelity) or ~1,230 (high-plex)	Targeted	Up to 300 mm2	[44]
Vizgen MERSCOPE	iST	0.2–0.3 μm	2021	Commercial	Up to 1,000 genes	Targeted	Up to 300 mm2	[27]
Seq-Scope	sST	0.5–0.7 μm	2021	Academic	Whole transcriptome	Up to 23	MiSeq 0.6 mm2 HiSeq >3 mm2 NovaSeq >50 mm2	[8] [9] [10]
Ex-Seq	iST	0.07 μm	2021	Academic	Whole transcriptome or up to 42 genes	<0.5 (untargeted)	1.7 mm x 1.0 mm	[35]

2. Advances in Spatial Omics: Emergence of μST

2.1. Sequencing-Based Spatial Transcriptomics (sST): Advancing Resolution to Microscopic Scale

The first functional prototype of sST was introduced in 2016, which repurposed a coarse microarray chip (200 μm center-to-center distance) as a spatially barcoded slide [2], demonstrating that untargeted spatial transcriptome analysis is possible. Since then, many efforts have been made to improve the resolution of sST through various strategies.

A straightforward way to improve resolution is to fabricate the microarray with finer features. Microprinting has improved resolution down to 100 μm and has been successfully commercialized as 10x Visium. Microfluidic delivery of spatial barcodes (see Glossary) in DBiT-Seq further reduced the resolution to 10 μm [3], and the recent 10x Visium HD, which incorporates improved microarray fabrication technology, achieved high-quality sST at 2 μm resolution [4]. These techniques utilize pre-assigned deterministic barcodes that are matched to known spatial coordinates (deterministic barcoding; see Glossary).

Independent of deterministic barcoding, microbead-based approaches have been developed to address the resolution challenge. In these efforts, two-dimensional microbead arrays containing random spatial barcodes were generated, and either sequencing-by-ligation (Slide-Seq [5]) or sequencing-by-hybridization (HDST [6]), was used to determine and register the spatial barcode sequences and corresponding spatial coordinates of each microbead. Initial implementations of Slide-Seq and HDST suffered from low RNA capture efficiency compared to other sST methods; however, subsequent improvements in Slide-Seq chemistry (Slide-SeqV2 [7]) enabled its commercialization as Curio Seeker.

The third approach, repurposing next-generation sequencing (NGS) instruments for sST, has been the most successful in producing high-resolution data at the microscopic scale so far. Using Illumina sequencing instruments, such as MiSeq [8], HiSeq [9] and NovaSeq [10], Seq-Scope has been able to achieve submicron resolution (0.5-0.6 μm), which can enable analysis at the single-cell and even subcellular levels. Using high-resolution (0.5-0.7 μm) BGI DNA nanoball (DNB) arrays, Stereo-Seq was also able to undertake single-cell-level analysis [11]. At a slightly lower resolution (1 μm), Pixel-Seq allows sequencing-free replication of high-resolution sequencing arrays through bridge amplification [12], which can further reduce the array synthesis costs. Stereo-Seq is now commercialized as STOmics by BGI, while the Seq-Scope protocol has been openly shared with academic investigators [8–10], facilitating the development of derivatives such as Open-ST [13] and Nova-ST [14], as well as the commercial development of a spatially optimized Illumina sequencing system [15]. By analyzing serial sections and reconstructing them along the Z-axis, some of these technologies have been adapted for three-dimensional analysis [13,16], although their Z-axis resolution (~10 μm) is much coarser than their XY resolution (<1 μm).

All of these approaches aim to increase array density to minimize pixel-to-pixel distances. Recently, an alternative strategy has been introduced: tissue expansion [17], which physically enlarges the tissue itself. Using this technique, Ex-ST reduced the effective resolution of Visium from 100 μm to 40 μm [18] without compromising the efficiency of transcriptome capture. More recently, in Seq-Scope-X, tissue expansion was combined with a high-resolution Seq-Scope array to achieve nanoscale resolution (180 nm), enabling tissue-wide subcellular analysis that revealed widespread phenotypic differences between the nuclear and cytoplasmic transcriptomes of hepatocytes, suggesting that hepatocytes could switch their transcriptomic profiles over time [19]. Notably, sST platforms have followed a trend of halving center-to-center distances each year over the past decade, resulting in a >1000-fold improvement in resolution (Fig. 1C), underscoring the rapid pace of technological advancement in spatial transcriptomics.

2.2. Imaging-Based Spatial Transcriptomics (iST): Expanding Scale and Coverage

Because iST is based on tissue imaging by either fluorescence in situ hybridization (FISH) or in situ sequencing (ISS), its resolution is inherently microscopic, at the optical limit of 200-300 nm. ISS was first described in 2013 [20], and since then, its precision and throughput have been massively improved through multiple adaptations, such as FISSEQ [21], STARmap+ [22] and Bar-Seq [23], among many others, and subsequently implemented in commercial products, such as Stellaromics Plexa/Pyxa [25] and Singular G4X [26], as well as in ISS-derived technologies incorporating FISH-based barcode detection, such as 10x Xenium [24]. Most ISS methods—and ISS-derived platforms like Xenium—share a general workflow involving mRNA-dependent probe ligation, rolling circle amplification (RCA), and cyclic fluorescence-based barcode readout, with platform-specific differences in probe design, amplification schemes, barcode encoding, sequencing chemistry, and imaging strategies.

FISH methods also have achieved massive multiplexing through the development of error-robust barcoding strategies, sequential hybridization cycles, and improved imaging instrumentation—culminating in technologies such as MERFISH [27] and seqFISH [28], among others, and subsequently to commercial products such as Vizgen MERSCOPE [29], Nanostring CosMx SMI [30], Resolve Molecular Cartography [31] and GenePS [32]. In addition, several iST methods have been further combined with tissue expansion to achieve nanoscale resolution of transcript identification, as demonstrated in expansion MERFISH [33], seqFISH+ [34] and Ex-Seq [35].

The throughput of iST technologies has also improved massively over this period (Fig. 1B), and several technologies, such as seqFISH+ [34] and MERFISH [36], could potentially cover up to 10,000 genes. However, achieving such extensive gene coverage would be prohibitively expensive and time-consuming, requiring multiple rounds of imaging, generation of gene-specific probes for each target, and challenges in precise image registration (see Glossary) across imaging cycles. Consequently, most current iST platforms offer readouts of several hundred genes, balancing resolution, cost, and time efficiency. Commercial solutions also continue to increase throughput, as exemplified by the 10x Xenium 5K and Nanostring CosMx 6K platforms, which profiles 5,000 – 6,000 genes selected as highly distinguishable markers of various cell types identified in scRNA-seq studies [37,38]. Another challenge is limited scalability, as the imaging machine can process only one sample at a time, with each session typically taking several days for most iST platforms. This scalability issue has been partially mitigated by the availability of a massively large imaging area in platforms such as G4X where 40 cm² of area is available [26], allowing multiple samples to be processed simultaneously in a single batch of experiments. Additionally, many iST platforms support 3D spatial analysis in thin sections (5–20 μm), and when combined with tissue-clearing methods, platforms such as Deep-STARmap allow deeper 3D analysis in thicker sections (up to 200 μm) [25,39].

2.3. Emergence of μST Era: Biological and Clinical Utility

Traditionally, sST has been considered stronger in unbiased, transcriptome-wide coverage, while iST has been regarded as superior in resolution. However, as described above, sST has improved its resolution at an astonishing pace (Fig. 1C), while iST has steadily increased its gene throughput to cover most of the transcriptome (Fig. 1D). As a result, the gap between the two technologies has significantly narrowed, opening the era of μST—where both platforms now offer high-resolution, high-content data that were previously inaccessible across diverse tissue types.

One of the earliest demonstrations of clinical utility for sequencing-based μST came from a study using Seq-Scope, which precisely mapped TREM2 macrophages in human acne lesions (Fig. 1B) [9]. Similarly, a preclinical study using Pixel-Seq uncovered both neuronal and glial heterogeneity as well as chronic pain–induced alterations in the mouse parabrachial nucleus [12]. Various μST techniques have been also used to generate an atlas-level characterization of whole embryogenesis or brain morphology [11,16,35,39–44], and also applied to human clinical studies across diverse tissues and disease states. For instance, in liver cancer, Stereo-seq delineated a tumor-border “invasive zone” characterized by metabolic reprogramming, hepatocyte–tumor cell crosstalk, and local immunosuppression [45]. Other Stereo-seq studies have defined pathological transcriptome features in Alzheimer’s disease [46], immune–vascular interactions in breast tumors [47] and metabolic barriers to tertiary lymphoid structure maturation in hepatocellular carcinoma [48]. The CosMx platform has offered single-cell spatial atlases of brain-resident immune cells and tumor microenvironments [49], and has been used in early-phase clinical trials to evaluate in situ drug activity [50]. Xenium, meanwhile, has revealed niche-level remodeling dynamics in fibrotic lung tissue [51], highlighting the potential of μST to dissect spatial pathology in complex human diseases.

2.4. Beyond Transcriptomics – Rapidly Emerging Area of High-Resolution Multi-Omics

Compared to iST, which is practically limited as a targeted assay, sST captures the transcriptome with intact sequence information, enabling multimodal analyses beyond simple transcript counts, including splicing status, allele-specific expression, and somatic alterations such as gene fusions. High-resolution sST technologies such as Seq-Scope, Stereo-Seq and Seq-Scope-X have indeed utilized splicing status to infer nuclear regions [8,11,19] or to estimate cell type trajectories by distinguishing current (spliced) and future (unspliced) transcripts [11,52].

Moreover, there is potential for ST technologies to examine additional genomic and epigenomic modalities beyond the transcriptome. iST, with its high resolution, could approach subnuclear chromatin structure and visualize the organization of genomic information [53,54]. In situ chromatin tagmentation (CUT&TAG) has also been combined with iST to spatially profile epigenomic marks at single cell and subnuclear scales [55]. iST could also be coupled with proximal labeling strategy to spatially profile the ribosome-associated translatome [39,56]. Low-resolution sST has also been applied to capture genomic sequences directly from tissue, leading to the recovery of spatially identifiable DNA sequences and the identification of copy number variations in cancer tissue [57]. In addition to capturing DNA information, sST could also be used to assay chromatin accessibility [58,59] or epigenome information [60]. Somatic mutations leading to immune repertoire diversity in the TCR and BCR loci have also been spatially profiled by low- [61–63] and high-resolution [64] sST. Nevertheless, iST approaches to genome and epigenome have still not achieved genome-wide coverage, while current sST still suffers from low resolution in addition to the issues associated with low genome capture rate and shallow data depth, which obscures effective analysis outcome at high resolution.

Since iST is basically a highly multiplexed imaging method, it could also be integrated with protein staining methods. Commercial iST solutions have already been combined with a selected set of protein markers [24], sometimes offering massive multiplexing capabilities [30]. There are also several commercial solutions that specifically target the detection of protein modalities through multiplexed imaging [65,66]. Using oligonucleotide-tagged antibodies, sST has also been modified to detect protein modalities in low- [67–69] and high-resolution [19,70] analyses. The inclusion of protein modalities could provide a more comprehensive view of cellular phenotypes that are sometimes difficult to detect by examining genomic or transcriptomic information alone.

3. Analysis of μST Data at Its Original Microscopic Resolution

Lower-resolution ST technologies, such as cST and mST, spatially resolve transcripts into 10-200 μm resolution “spots” (see Glossary), each typically covering multiple cells (Fig. 2A). Each spot is typically modeled as a mixture of cell types, characterized using scRNA-seq or identified through unsupervised learning [71–74]. In a typical tissue section, lower-resolution technologies produce thousands of spots, each containing hundreds to thousands of transcripts (Fig. 2B, Visium and Slide-Seq). In contrast, μST technologies resolve transcripts with much finer-grained spatial coordinates or “pixels” (see Glossary). A typical tissue section contains millions to billions of pixels, each containing only a handful of transcripts (Fig. 2B, Stereo-Seq, Xenium, MERSCOPE, CosMx). The extreme sparsity—referring to the small number of transcripts per pixel at submicron resolution—and scale of μST data present unique computational challenges that require new data formats and methods to analyze at its original resolution. Earlier μST methods have often employed grid-based segmentation to alleviate the computational challenges in platforms like Seq-Scope, Stereo-seq or Visium HD, when histology-based cell segmentation (see Glossary) is challenging or inaccurate [4,8,9,11,16]. Grid segmentation allows the application of methods designed for lower-resolution technologies at the expense of compromised spatial resolution, which can be partially mitigated by sliding window heuristics [75].

Figure 2.

Challenges in Microscopic Spatial Transcriptomics (μST) Analysis

(A) Conceptual diagram illustrating how spatial transcriptomics (ST) techniques capture gene expression using spatial features—spots or pixels—at varying resolutions. Coarse ST (cST) and magnified ST (mST) use larger spots, while μST captures data at the fine-grained pixel level.

(B) Comparative analysis of mouse brain data generated by cST (Visium), mST (Slide-SeqV2), and μST (Visium HD, Stereo-Seq, Xenium, MERSCOPE, CosMx SMI), assessing resolution (red, μm), spot number (orange), spot density (yellow, spots/mm²), average UMIs per spot (green), average genes per spot (blue), and tissue area (black, mm²). μST datasets have millions of spots with highly sparse information (few UMIs/spot), while cST and mST produce fewer spots with denser information (>100 UMIs) per spot.

(C) Schematic of the challenges in histology-based cell segmentation due to 2D analysis of 3D tissues. Issues include missing nuclei, multinucleate or anucleate cells, and variabilities in cell size and shape, making it difficult to accurately define cell boundaries.

(D) Hypothetical comparison of analysis strategies: near-cellular binning (left), nucleus-based segmentation (center), and pixel-level analysis (right). Pixel-level analysis preserves the native resolution and provides a robust alternative where cell segmentation is unreliable.

When histology images are available, segmentation of the tissue section into individual cells is currently the most common practice for μST analysis, especially for iST. Cell segmentation methods for microscopy images adapt neural network architectures from computer vision to predict cell or nucleus positions from DAPI, PolyT, H&E, or cell membrane staining [76,77]. For example, StarDist [78] and Cellpose [79] both choose U-Net as their base model, while μSAM [80] fine-tunes the SAM (Segment Anything Model) vision foundational model [81]. These methods are supervised and often provide pre-training of multiple models for different histological modalities while allowing users to fine-tune the model with their own labeled data. In these cell segmentation methods, the transcript-level μST data are typically aggregated to the single-cell level to apply existing spatially-agnostic single-cell analysis tools [82,83] or newer spatially-aware single-cell analysis tools [84,85] to characterize cell identities and tissue organizations. Although histology-based segmentation is widely used, it struggles in complex tissue regions with crowded cells, in identifying cells with irregular shapes, in identifying multinucleate or anucleate cells, or in identifying cells where a 2D slice does not contain their nuclei (Fig. 2C and 2D). Artifacts from cell segmentation errors propagate to common downstream tasks and confound the results of differential expression and cell-cell interaction analyses [86].

With submicron-resolution ST technologies, the transcript-level data itself can also be used for cell segmentation, either without or together with histology images. For example, Baysor [87] and Proseg [88] use unsupervised probabilistic models that can be initialized with nuclear positions or cell boundaries but do not require histological staining to segment μST data into individual cells. Although both methods are initially designed for iST, their approaches may be particularly useful for sST where histology images are often unavailable. However, the richer and noisier information from >10,000 genes poses additional challenges, making these methods difficult to apply to large-scale sST datasets. Alternatively, segmentation-free methods designed for μST directly profile cell types or functions at the pixel level, avoiding the artifacts of cell segmentation. For example, TopACT [89] applies cell type classifiers constructed from scRNA-seq references across multiple scales of aggregation and adaptively chooses the resolution at local spots [90]. FICTURE [91] uses a multilayered Dirichlet model to assign individual pixels to spatial factors (see Glossary) obtained from unsupervised learning or pseudo-bulk cell type expression profiles, revealing microscopic spatial features that are undetectable by near-cellular binning or conventional cell segmentation algorithms in various datasets (Fig. 3). Future development of segmentation-free analysis tools will lead to a more comprehensive and unbiased characterization of ST including small and irregularly shaped cells and extracellular RNA (Fig. 2D).

Figure 3.

Comparison of different computational strategies for the analysis of real-world μST data.

10X Xenium mouse brain (top) and human breast cancer (bottom) data, and Vizgen MERFISH mouse liver (middle) data were analyzed using near-cellular binning (left), image-based cell segmentation (center) and pixel-level analysis (right) methodologies. Near-cellular binning obscures high-resolution spatial information, while image-based cell segmentation often fails in regions where nuclear or membrane markers are poorly stained or ambiguous. In contrast, pixel-level analysis enables clear visualization of tissue microstructures across the entire histological area. Near-cellular binning and pixel-level analyses were performed using Latent Dirichlet Allocation (LDA) and stochastic variational inference, respectively, both implemented in FICTURE [91]. Cell segmentation analyses for 10x Xenium datasets are provided by 10x Genomics in the standard output generated by Xenium Ranger. Cell segmentation analyses for the Vizgen dataset are from the Squidpy [85] vignette (https://squidpy.readthedocs.io/en/stable/notebooks/tutorials/tutorial_vizgen_mouse_liver.htm), based on the cell segmentation provided within the released dataset.

In addition, data fragmentation—arising from cross-platform differences in resolution, modality, and file structure—poses barriers to integrative analysis and tool development. Thus, efficient and interoperable data formats for storing μST data across heterogeneous technologies are getting increasingly important to facilitate the development of scalable cross-platform analysis tools. Because μST data are extremely large and sparse, containing millions to billions of pixels per tissue section, it is challenging to store and access them efficiently with existing formats [92]. For these reasons, many publicly available resources for μST data do not release the data in their original resolution [16,93,94]. Instead, they publish data aggregated by coarse resolution grids (e.g. 10 μm x 10 μm or 25 μm x 25 μm) or segmented cells in AnnData [82] or h5Seurat [83] format, limiting the full potential of the original submicron resolution. New data formats such as SpatialData [95], OME-Zarr [96], OME-NGFF [97], and SpatialFeatureExperiment [98] have been proposed to address these challenges, but it remains unclear whether these formats are sufficiently efficient and interoperable to handle various μST analysis and visualization tasks across heterogeneous μST platforms. On the other hand, repurposing many existing data formats originally designed for geospatial data [99], such as Cloud-Optimized GeoTIFF (COG), Mapbox Vector Tiles (MVT), MBTiles, and PMTiles, is an attractive option for leveraging existing software infrastructure to develop tools for μST data analysis, as they are designed to efficiently store and access large spatial data in cloud-based environments.

4. New Opportunities Created by μST

With the accumulation of μST data, new biological questions become addressable while new computational challenges arise. RNAs and proteins are known to be differentially localized to cellular compartments and organelles, including nuclei, mitochondria, and endoplasmic reticulum [100]. With subcellular resolution transcriptomic and proteomic data, it is possible to characterize the localization of individual genes with respect to landmarks such as cell and nuclear boundaries [101,102], quantify the colocalization of gene pairs [101,103], jointly model genes and cells to identify subcellular localization patterns shared by multiple genes [104], or reconstruct subcellular RNA kinetic landscapes by pulse-chase labeling coupled with in situ sequencing [105]. In addition, the spatial whole transcriptome assayed by sST technologies can identify transcriptional dynamics such as splicing, polyadenylation, and RNA degradation to enhance the sensitivity of subcellular analysis [19].

The subcellular resolution of spatial omics also provides new opportunities for cell-cell interaction (CCI) analysis. Existing CCI methods developed for scRNA-seq data evaluate interactions only as enrichment of known ligand-receptor pairs [106–108]. Methods applied to low resolution ST are additionally informed by spatial proximity [109,110], while μST allows direct inference of CCI [111–113] and ligand-receptor relationships [114–116] at the single cell level. However, all existing methods start from segmented cells although CCI analysis is sensitive to segmentation errors [86]. Future developments that fully leverage μST to model interactions at micron resolution will empower hypothesis generation for CCI including ligand-receptor interactions and long-range molecular signaling.

Increased resolution of μST also allows subcellular RNA analysis, which may reveal intricacies of polarized RNA distribution, RNA granulation and RNA trafficking into organelles. Fine-grained separation of nuclear and cytoplasmic transcriptomes could also improve spatial single cell trajectory analysis by distinguishing between current (cytoplasmic) and future (nuclear) transcriptomes, complementing previous RNA splicing-based analysis, which have notable limitations [117].

Interactive visualization and exploration of μST data is extremely important to make the data accessible to investigators and clinicians to identify regions of interest relevant to disease-related features, improve diagnosis with molecular markers, and generate new hypotheses. Many existing visualization tools such as Loupe, Xenium Explorer, MERSCOPE Vizualizer are platform specific, and it is important to develop cross-platform visualization tools that can handle interoperable data formats. Existing low-resolution tools such as Giotto [118] and Squidpy [85] can now handle cell-segmented μST data, and new tools such as vitessce [119], Voyager [98], WebAtlas [120], BelleVista [121], and Celldega [https://broadinstitute.github.io/celldega/] are emerging to handle μST data at its original resolution.

While many μST technologies can generate 3D spatial omics data by serial sectioning [13,16,40,41], the analysis of such data at microscopic resolution remains in its infancy. Aligning multiple 2D sections at submicron resolution presents an additional challenge, and several methods, such as feature-based image alignment [13], probabilistic alignment [122] and non-grid alignment [16], have been proposed. Although many existing analysis methods for 2D data could, in principle, be extended to 3D, most require substantial modifications to accommodate the extra dimension, with only a few exceptions [16]. Finally, existing spatial data formats either lack native support for 3D data or become inefficient for 3D spatial queries. As 3D spatial omics, including 3D μST data, continue to grow in both prevalence and scale, new analysis methods, data structures, and file formats optimized for 3D data will be essential.

With the increasing availability of high-resolution spatial omics data coupled with histological images, it is becoming increasingly important to integrate these two types of data with deep learning approaches. Large foundational models (see Glossary) built from hundreds of thousands of histopathological images have demonstrated their utility in various tasks such as cancer subtyping, mutation prediction, and prognosis prediction [123–126]. In contrast, spatial omics data paired with histological images remain scarce, especially at microscopic resolution [127,128]. Despite the limited number of available ST datasets, studies have shown that spatial transcriptomics can be predicted from histological images [129–131]. For example, ST-Net [129], employs convolutional neural networks to extract morphological features, while BLEEP [130] incorporates reference ST datasets and bimodal contrastive learning. TRIPLEX [131] builds on these by integrating multi-resolution image features to capture both global tissue architecture and local cellular details. Prediction accuracy is expected to improve with more advanced models and higher-resolution ST data. Extending this to three dimensions, recent efforts have used bimodal contrastive learning [132] and multi-resolution image features [133] to predict 3D spatial transcriptomics from 3D tissue images and 2D ST data. Although these methods have yet to be applied to μST, similar approaches may offer cost-effective, scalable solutions for approximating 3D spatial omics, supporting broader adoption of 3D μST in research and clinical settings. Ultimately, both 2D and 3D multi-modal analyses of histopathology and spatial omics are poised to drive future advances in spatial biology.

5. Future Directions in μST Technology and Analysis

To date, the μST field has focused primarily on gene expression analysis, with limited integration of protein-level measurements. We anticipate that multimodal strategies established for lower-resolution techniques—such as spatial proteomics, epigenomics, and genomics [19,57–60,67–70]—will soon be adapted to the high-resolution domain of μST. This expansion will require advances in multimodal analytical approaches to effectively leverage microscopic resolution. However, existing limitations in coverage and data depth pose challenges. For example, while spatial genomic applications can perform low-depth analyses like copy number variation (CNV) detection [57], their sensitivity remains insufficient for tracking somatic mutations or other small-scale genomic aberrations. The challenges from low coverage can be exacerbated at high resolution, requiring improvements in experimental throughput and statistical methods. In addition to profiling molecular status of the target organism, μST could also interrogate commensal and pathogenic microorganisms associated with the host and systematically profile their content and activities [134–136].

The clinical utility of μST and spatial multi-omics (see Glossary) remains nascent, despite exhibiting considerable promise for translational applications. μST enables spatially resolved gene expression profiling at single-cell resolution, offering insights into cellular heterogeneity, spatial context, and tissue microenvironments that are inaccessible to bulk sequencing or low-resolution ST. As highlighted in Section 2.3, μST has already demonstrated its value in precision oncology and neuroscience by advancing our understanding of tumor architecture, neural circuits, and disease pathology. Looking ahead, μST could play a transformative role in these and other fields by uncovering localized disease mechanisms and informing targeted interventions. Moreover, its integration with genomic, proteomic, metabolomic, or spatial epigenomic data may further refine patient stratification, reveal complex regulatory networks, and accelerate the development of precision diagnostics and therapeutics.

These applications highlight the promise of μST in biomedical research and translational medicine. However, realizing its full clinical impact will require robust computational frameworks and standardized analytical pipelines to ensure reproducibility and regulatory approval. Wide clinical adaptation of μST would also require continued methodological development, including more straightforward and reproducible sample processing procedures, streamlined and standardized analysis pipelines, improved cost efficiency, and clear demonstration of clinical benefit in patient care (see Outstanding questions).

Outstanding Questions.

• Can sequencing-based ST (sST) be made more sensitive to capture deeper transcriptome information, and can imaging-based ST (iST) be scaled for broader accessibility? Could these approaches be effectively combined to leverage the strengths of both?

• How can 3D spatial omics with improved z-axis resolution be achieved, and what tools are needed for effective 3D alignment, analysis, and visualization at the microscopic scale?

• What are the best strategies for harmonizing cross-platform spatial data while preserving resolution and enabling seamless integration?

• Can we develop a user-friendly, accessible data repository that allows non-computational biologists and physicians to explore μST datasets without requiring computational expertise?

• Can segmentation-free approaches improve spatial analysis and bypass the challenges of conventional segmentation?

• What advances are needed to enable comprehensive spatial multimodal analysis— genomic, transcriptomic, proteomic, epigenomic—at microscopic resolution?

Highlights.

• Sequencing-based ST (sST) is approaching and surpassing microscopic resolution scale, while imaging-based ST (iST) is scaling up to near-transcriptome-wide coverage.

• Microscopic-resolution ST (μST) presents unique computational challenges due to the vast number of spatial features and extreme data sparsity.

• Persistent challenges in cell segmentation and cross-platform integration require the development of segmentation-free approaches and harmonized, scalable computational frameworks.

• New opportunities in AI integration, 3D spatial omics, and clinical applications are expanding the impact of μST in both research and translational medicine.

Glossary

Cell Segmentation

The process of delineating individual cells within histological images to aggregate spatial transcriptomics data at the single cell level

Deterministic Barcoding

A barcoding approach in which the spatial position of each barcode is predefined and fixed during array fabrication, allowing direct alignment to known spatial coordinates

Foundational Models (in AI)

Large-scale neural networks pre-trained on large datasets, often adaptable to different tasks including spatial data analysis and histology interpretation

Pixel (in ST)

The smallest spatial unit in a μST dataset, typically submicron in size, each associated with a small number of transcripts

Registration (in spatial omics)

The process of aligning or mapping spatial barcodes to their precise physical locations within a tissue or array, enabling accurate reconstruction of spatially resolved molecular data

Segmentation-Free Analysis

Computational approaches that analyze spatial transcriptomics data at the pixel or transcript level without relying on predefined cell boundaries

Spatial Barcodes

Short DNA sequences used to tag molecular information (e.g., RNA transcripts or proteins) with spatial coordinates in tissue samples, allowing reconstruction of spatially resolved molecular maps during data analysis

Spatial Factors

Latent features or patterns inferred from spatial transcriptomics data that describe gene expression variation across spatial coordinates

Spatial Multi-Omics

The integrated analysis of multiple molecular layers (e.g., transcriptomics, proteomics, epigenomics) in a spatial context

Spot (in ST)

The spatial unit in cST or mST dataset, typically 10 – 100 μm in size, each associated with hundreds to thousands of transcripts

μST (Microscopic-Resolution Spatial Transcriptomics)

The latest generation of spatial transcriptomics that achieves subcellular resolution (<1 μm), enabling highly detailed spatial gene expression profiling. Earlier generations with lower resolution are referred to as coarse (cST) or magnified ST (mST)