Spatial Transcriptomics of Adipose Tissue: Technologies, Applications, and Challenges

Abstract

Adipose tissue is a complex metabolic and endocrine organ that plays a central role in systemic energy homeostasis. While single-cell and single-nucleus RNA sequencing have revealed remarkable cellular heterogeneity within adipose tissue depots, these approaches lack spatial context, limiting the ability to understand how cellular organization and microenvironmental cues shape adipose tissue biology. Spatial transcriptomics (ST) has emerged as a powerful technology to overcome this barrier by allowing one to map gene expression directly within intact tissue sections. Recent advances in ST platforms now permit analysis at a high resolution, enabling interrogation of adipocyte subpopulations, stromal progenitors, immune cell infiltration, and tissue remodeling. In this review, we provide an overview of current ST technologies, computational strategies for analysis, and recent applications for understanding adipose tissue biology. We further highlight key opportunities for ST to address unanswered questions surrounding adipogenic niches, depot-specific remodeling, and immune cell interactions. Together, these advances position ST as a transformative tool for dissecting the architecture and function of adipose tissue in health and metabolic disease.

INTRODUCTION

Adipose tissue is crucial for storing and regulating lipids, playing a vital role in maintaining metabolic homeostasis.^1-4 With the concerning rise in obesity, diabetes, and cardiometabolic diseases worldwide, adipose tissue has been the focus of extensive research.^5-7 In humans and other mammals, adipose tissue is organized into several depots with differences in function, cytokine/adipokine profiles, cellular composition, and gene expression patterns.^5,8-10 The major white adipose depots are broadly classified as visceral or subcutaneous, with visceral fat accumulation more strongly associated with metabolic disease.¹¹ Distinct, potentially specialized depots can also be found in the dermis, face, mammary glands, joints, bone marrow, cardiovascular system, and intramuscularly.^4,5,12

Adipocytes are broadly categorized as white, beige, or brown.^13,14 White adipocytes (WAs) store lipids as triacylglycerol that can later be mobilized as fatty acids. WAs contain a single, large (50 to 150 μm diameter) lipid droplet that comprises >90% of the cellular volume. In contrast, brown adipose tissue (BAT) is a thermogenic organ that generates heat by uncoupling adenosine triphosphate production from fatty acid oxidation, a process mediated by uncoupling protein 1.15 Brown adipocytes (BAs) are smaller (approximately 20 μm) and contain many lipid droplets and numerous mitochondria, which give them their eponymous brown coloration.¹⁶ While WAs and BAs arise from distinct developmental lineages,^17-19 beige adipocytes, which have an intermediary phenotype, are thought to arise mainly from interconversion of WAs in response to adrenergic stimulation.^13,19-21 Adipose tissue also contains a diverse array of non-adipocyte cell types, including adipose stromal cells (ASC), vascular endothelial cells, smooth muscle cells, pericytes, dendritic cells, macrophages, and other immune cells.^8,22,23 Importantly, the organization and composition of adipose tissue can shift dramatically in response to environmental or metabolic cues.^24-26

Single-cell RNA sequencing (scRNA-seq) of adipose tissue has been performed by several groups, revealing complex patterns of gene expression (Table 1).^22,27-32 However, due to their large size, fragility, and buoyancy, mature adipocytes cannot be captured efficiently using conventional scRNA-seq approaches, which rely on enzymatic digestion and microfluidics. Thus, single-nucleus RNA sequencing (snRNA-seq) has become popular to capture expression information from all nuclei present in adipose tissue.^33,34 Using both techniques, multiple subpopulations of ASC have been identified that differ in location, transcriptomic profile, and propensity for adipogenesis.^23,35-38 Both murine white adipose tissue (WAT) and BAT contain three or more subpopulations of ASC that are differentially primed for adipogenesis.^23,37,38 Similarly, human ASCs show a range of adipogenic potential, and proportions of subpopulations differ among depots.^8,29,36

Table 1.

Summary of major transcriptomic studies of adipose tissue

Author (year)	Method	Tissue	Key findings in adipose tissue
Burl et al. (2018)23	scRNA-seq	Mouse eWAT and iWAT	Adrenergic stimulation results in expansion of proliferating ASCs which may display quiescent or adipogenic trajectories.
Schwalie et al. (2018)35	scRNA-seq	Mouse WAT	Uncovered a population of Cd142+ ASCs that suppress adipogenesis through paracrine signaling.
Hepler et al. (2018)38	scRNA-seq	Mouse visceral WAT	Identified FIPs as a fibrogenic, pro-inflammatory mural cell population that likely contributes to obesity-associated WAT dysfunction.
Hill et al. (2018)39	scRNA-seq	Mouse eWAT, human VAT	Cd9+Ly6c- macrophages accumulate during obesity and are associated with CLS.
Jaitin et al. (2019)40	scRNA-seq	Mouse eWAT, human VAT	Trem2+ macrophages are important for initiating CLS formation and may regulate adipocyte size.
Merrick et al. (2019)36	scRNA-seq	Mouse iWAT	ASCs reside in the interstitial tissue, and Dpp4+ ASCs give rise to committed progenitors.
Vijay et al. (2020)8	scRNA-seq	Human VAT and SAT	Found dysfunction in multiple cell types in WAT of obese and T2DM patients.
Sárvári et al. (2021)42	snRNA-seq	Mouse eWAT	Obesity leads to an increase in abundance of LAMs and upregulation of lipid-handling genes across all ATMs.
Bäckdahl et al. (2021)107	ST (Visium)	Human SAT	Found three spatially and transcriptionally distinct subpopulations of adipocytes that respond differently to insulin.
Shan et al. (2022)37	scRNA-seq	Mouse iWAT and visceral WAT	Identified differences in ASCs between male and female mice and between depots, with females enriched for mitochondrial genes.
Emont et al. (2022)29	scRNA-seq	Human and mouse VAT and SAT	Created single-cell resolution cellular atlases of human and mouse WAT.
Burl et al. (2022)41	scRNA-seq	Mouse BAT	Three subpopulations of ASC were identified, one of which is responsible for cold-induced proliferation which correlates with immune cell recruitment.
Stansbury et al. (2023)109	ST (Visium)	Mouse eWAT	Analyzed immune cell composition in early and late obesity, found evidence for LAMs that facilitate CLS formation.
Lundgren et al. (2023)111	ST (Visium)	Mouse BAT	Identified spatially organized brown adipocytes that contribute to cold tolerance after repeated cold exposure.
Nigro et al. (2023)112	ST (Visium)	Mouse iWAT	Exercise training reduces collagen deposition, increases vascularization and neurite density, and promotes beiging in WAT.
Kohda et al. (2025)110	ST (Visium)	Mouse eWAT	Clec4e+ macrophages influence ECM remodeling and CLS formation during HFD.

scRNA-seq, single-cell RNA sequencing; eWAT, epididymal white adipose tissue; iWAT, inguinal white adipose tissue; ASC, adipose stromal cell; WAT, white adipose tissue; FIP, fibro-inflammatory progenitors; Ly6c, lymphocyte antigen 6 family member C1; CLS, crown-like structure; Trem2, triggering receptor expressed on myeloid cells 2; Dpp4, dipeptidylpeptidase-4; VAT, visceral adipose tissue; SAT, subcutaneous adipose tissue; T2DM, type 2 diabetes mellitus; snRNA-seq, single-nucleus RNA sequencing; LAM, lipid-associated macrophage; ATM, adipose tissue macrophage; ST, spatial transcriptomic; BAT, brown adipose tissue; Clec4e, C-type lectin domain family 4 member E; ECM, extracellular matrix; HFD, high fat diet.

scRNA-seq and snRNA-seq have further illuminated the importance of immune cell dynamics in adipose biology and tissue remodeling.^23,39-42 For example, genetic or diet-induced adipocyte hypertrophy results in the recruitment of diverse immune cells to WAT.²⁶ Recruited immunotypes are enriched in macrophage subtypes specialized in lipid handling and extracellular matrix (ECM) remodeling.^41,43 Recruited macrophages often participate in the phagocytic removal of defective adipocytes, creating crown-like structures (CLS).⁴⁴ Jaitin et al.⁴⁰ identified that triggering receptor expressed on myeloid cells 2 (Trem2)+ macrophages (lipid-associated macrophages [LAMs]) are important for initiating CLS formation, and Trem2 signaling induces the LAM phenotype that may help regulate adipocyte size. Likewise, Hill et al.³⁹ found Cd9+ lymphocyte antigen 6 family member C1 (Ly6c)- macrophages accumulate during obesity and are associated with CLS, but do not have a classical M1 or M2 phenotype. Burl et al.⁴¹ also found an increase in macrophage activation and proliferation during catabolic tissue remodeling in the epididymal WAT (eWAT). Further, local proliferation of macrophages and dendritic cells strongly correlated with brown adipogenesis in BAT during cold stress.^23,41 Lastly, Sárvári et al.⁴² demonstrated that obesity leads to an increase in abundance of LAMs and upregulation of lipid-handling genes across all adipose tissue macrophage subsets in WAT.

Despite the insights gained from scRNA-seq and snRNA-seq, spatial transcriptomics (ST) can provide important contextual information not available using these techniques. A key limitation of scRNA-seq is the loss of critical spatial context due to tissue dissociation. This is particularly important in adipose tissue, where cell–cell interactions strongly influence tissue remodeling and metabolic function. ST allows for the identification of spatially restricted cell states that would be lost with dissociation-based methods. Using ST, molecular programs can be linked to pathology, further elucidating tissue dynamics. Emerging technologies in ST offer the ability to access transcriptomic information within intact tissues at subcellular resolution.^45-48 Evidence from low-plex immunofluorescence studies has highlighted the importance of spatially organized immune–adipocyte interactions, such as the formation of CLSs and the immune/progenitor interactions that shape proliferation and differentiation of ASCs (Fig. 1).^41,49,50 Here, we will outline the most recent advances in ST and how it can be applied to address additional questions in adipose tissue biology.

Figure 1.

Graphical representation of spatial phenomena in white (left) and brown (right) adipose tissue. A, lipid-associated macrophages (red) surrounding a dying adipocyte in a CLS; B, white adipocyte undergoing beiging process; C, heterogeneous gene expression in mature adipocytes; D, receptor/ligand interactions; E, interstitial adipose stromal cell (ASC); F, perivascular ASC; and G, immune/ASC interactions.

Spatial transcriptomic technologies

ST approaches fall into two main categories: sequencing and imaging-based (Fig. 2).^51,52 Sequencing-based approaches are generally unbiased and involve capturing RNA directly from tissue sections onto a barcoded matrix or slide, then creating a library for sequencing and further deconvolution. Current limitations of these methods include the size of the capture area, spacing of the barcoded ‘spots’ for single-cell resolution, capture efficiency, and ability to detect low-abundance transcripts.^52,53 Imaging-based approaches utilize fluorescent in situ hybridization (FISH)-based technology, in which RNA molecules are directly or indirectly tagged with fluorescent probes or subject to in situ sequencing coupled with fluorescence imaging readout.^51,52 Image-based methods generally require prior knowledge of tissue composition to select probes for capturing transcripts of interest. Limitations of image-based platforms include the risk of optical crowding and limited number of genes that can be profiled.^45,52

Figure 2.

A simplified comparison of sequencing versus imaging-based spatial transcriptomics. In general, sequencing-based techniques (left) entail applying sectioned tissue to a scaffold/slide containing barcoded ‘spots’ so that the RNA is captured with a unique barcode corresponding to the area of tissue it came from. cDNA is synthesized from these barcoded oligos and subsequentially sequenced, resulting in a gene x location matrix. Imaging-based (right) techniques typically involve applying fluorescent probes to sectioned tissue in a targeted manner to directly target genes of interest. Several rounds of imaging take place and the fluorescent signal and location of probes is then compiled into a gene×location matrix. 3D, three dimensional; 2D, two dimensional.

In addition to choosing between imaging or sequenced-based ST, one must also consider differences in sensitivity, spatial resolution, capture size, cost, and commercial availability (Table 2). ST platforms also vary in compatibility with tissue preservation methods and section thickness. This can present problems especially for WAT, which is notoriously difficult to cryosection. ST technologies have advanced rapidly in the past couple years, with substantial improvements in both single-cell resolution and analytical methods available.

Table 2.

Summary of spatial transcriptomic technologies

	Sequencing or imaging-based?	Spot size/resolution (μm)	Distance between spots (μm)	Capture area (mm)	Tissue type	Genes (n)	Reference
Visium	Sequencing	55	100	6.5 × 6.5	F	Whole transcriptome	54
Visium HD WT panel	Sequencing	2	-	6.5 × 6.5	F, FF, FFPE	Protein coding only	54
Visium HD 3’	Sequencing	2	-	6.5 × 6.5	F	Whole transcriptome	54
Slide-seq	Sequencing	10	10	3 × 3–10 × 10	F	Whole transcriptome	55
DBiT-Seq	Sequencing	10	20–100	5× 5	F, FF, FFPE	Whole transcriptome	57
Stereo-seq	Sequencing	0.22	0.5	10 × 10–132 × 132	F, FFPE	Whole transcriptome	59
Seq-Scope	Sequencing	~0.6	~0.63	7× 7	F	Whole transcriptome	60
GeoMx	Sequencing	Variable	Variable	36 × 14	F, FF, FFPE	Whole transcriptome	61
Xenium	Imaging	0.2	-	12 × 24	F, FFPE	5,000+	63
CosMx	Imaging	0.12	-	15 × 20	F, FFPE	6,000+	66
seqFISH+	Imaging	0.103	-	Variable	F, FFPE	10,000+	67,68
MERSCOPE	Imaging	0.1	-	10 × 10	F, FF, FFPE	1,000+	69
Molecular Cartography	Imaging	0.3	-	10 × 10	F, FFPE	100
Spatial CITE-seq	Other	25	50	2× 2	F	Whole transcriptome	70
EEL FISH	Other	0.2–0.4	-	Variable	F	2,000+	71

F, fresh frozen; HD, high definition; WT, whole transcriptome; FF, fixed frozen; FFPE, formalin-fixed paraffin-embedded.

SEQUENCING-BASED SPATIAL TRANSCRIPTOMIC METHODS

Visium⁵⁴ is a commercial sequencing-based technology developed by 10x Genomics (https://www.10xgenomics.com/). In this method, tissue is affixed onto a glass slide patterned with barcoded oligonucleotide probes. The first-generation Visium slide featured a spot diameter of approximately 55 µm and a center-to-center distance of 100 µm. This relatively coarse resolution left substantial areas of tissue between capture spots. The most recent generation, Visium high definition (HD), has significantly improved spatial resolution with a spot size of approximately 2 µm, enabling subcellular transcript localization. 10x Genomics currently offers first-generation Visium panels as well as two Visium HD assay types: one targeting the whole transcriptome, which is compatible with fresh frozen, fixed frozen, and formalin-fixed paraffin-embedded (FFPE) tissues, and another focusing on protein-coding genes which is compatible only with fresh frozen tissues.

Slide-seq⁵⁵ was commercialized by Curio Biosciences (https://curiobioscience.com). This technology incorporates 10 µm spatially indexed, barcoded beads packed onto a coated coverslip to create a high-density array. In the original design, spatial positions of the barcodes were pre-mapped using a separate sequencing-by-synthesis step, allowing transcriptomic reads to be assigned back to their original spatial locations. Slide-seqV2 incorporates technical improvement to improve mRNA capture efficiency and library preparation.⁵⁶

Deterministic barcoding in tissue for spatial omics sequencing (DBiT-seq)⁵⁷ uses microfluidic chips with parallel channels to first introduce a set of oligoDT-tagged barcodes containing a linker onto one axis (e.g., the X-axis) of a tissue section that hybridizes to mRNA. A second chip is placed perpendicular to the first and applies a second set of barcodes that ligate to the first via linkers. The second set of barcodes contains a linker, a unique molecular identifier (UMI), and a polymerase chain reaction handle to assist with downstream purification steps. The results are a grid of cDNA uniquely barcoded to indicate the two dimensional (2D) location of the transcripts. The channel width may vary from 10 to 50 µm. When compared to other platforms, DBiT-seq has relatively low capture rates.⁵⁸

SpaTial Enhanced REsolution Omics-sequencing (Stereo-seq)⁵⁹ is a next-generation ST platform, notable for its ultra-high spatial resolution and large-area scalability and commercially available through Complete Genomics (https://completegenomics.com). Stereo-seq has an extremely high resolution with a spot size of just 0.22 µm and a center-to-center distance of 0.5 µm between spots, enabling subcellular spatial resolution. The capture area can range from 5×5 mm to 13.2×13.2 cm, making it one of the few platforms that can simultaneously offer nanometer-scale resolution and centimeter-scale coverage. Stereo-seq uses DNA nanoball (DNB) technology on patterned array chips. Each spot has a randomly barcoded DNB template (coordinate ID), UMI, and oligo dT probe to capture polyadenylated mRNA. Tissue sections are placed onto the array, processed, then undergo in situ reverse transcription. The resulting cDNA is recovered and used to create a library for high-throughput sequencing. Stereo-seq has been shown to exhibit high gene detection sensitivity and transcript capture rates, outperforming methods like Slide-seqV2 and DBiT-seq.⁵⁸

Seq-Scope⁶⁰ is a high-resolution sequencing-based ST platform that repurposes standard Illumina sequencing flow cells as spatially barcoded RNA-capture arrays. Flow cells are loaded with a custom oligonucleotide library containing random spatial barcodes, poly(dT) sequences for mRNA capture, and Illumina sequencing adapters. These are clonally amplified into dense, spatially discrete clusters directly on the flow cell surface through solid-phase bridge amplification. Initial sequencing is performed to create a high-resolution spatial map linking each barcode to a unique physical location. The mean distance between neighboring clusters is 0.6 µm, giving Seq-Scope subcellular resolution.

GeoMx⁶¹ Digital Spatial Profiler (NanoString Technologies/Bruker Spatial Biology) enables combined transcriptomic and proteomic profiling from user-defined regions of interest (ROIs). In this approach, in situ hybridization probes and/or antibodies are conjugated to unique indexing oligonucleotides via ultraviolet (UV)-cleavable linkers. The tissue is stained and imaged to guide ROI selection. ROIs, typically encompassing 50 to 100 cells, are then selected by the user based on tissue morphology or marker expression. UV light is precisely directed to each ROI to cleave the indexing oligos, which are subsequently collected. Released oligonucleotides are used to generate a sequencing library. Spatial localization of transcripts or proteins is inferred by mapping the collected oligos back to their respective ROIs. Using GeoMx, hundreds of proteins and whole transcriptomes can be simultaneously interrogated in the same tissue.⁶²

IMAGING-BASED SPATIAL TRANSCRIPTOMIC METHODS

Xenium,⁶³ developed by 10x Genomics, is a highly multiplexed image-based ST platform that uses FISH to detect and localize hundreds to thousands of RNA species. The system supports both pre-designed gene panels targeting up to 5,000 genes and custom panels, making it adaptable to a wide range of tissue types. Each gene of interest is targeted by multiple padlock probes (typically 5–8), which hybridize to the RNA in a highly specific interaction involving dual target recognition sites. If both arms of the probe successfully hybridize to the target site, the probe arms are ligated together, and the probes undergo rolling circle amplification. Transcripts are then labeled with fluorescent probes in a series of cycles in which four fluorophores are added, imaged, and cleaved. This creates a transcript-specific pattern of fluorescence that can be decoded to identify the optical signature and location of each gene. In comparison with similar technologies Xenium offers a lower false positive rate and more robust detection of transcripts across diverse tissue type.^64,65 Recently, 10x Genomics has begun to offer limited protein immunofluorescent (IF) subpanels in conjunction with Xenium, allowing for interrogation of protein along with RNA, though they are currently only available for human tissues.

CosMx⁶⁶ Spatial Molecular Imaging (NanoString Technologies/Bruker Spatial Biology) is a highly multiplexed image-based ST platform that enables simultaneous detection of up to 6,000 RNA targets. Each gene of interest is targeted by up to five probes that bind to different 30 to 50 bp segments of the RNA. Each probe also contains a readout domain that is unique to each target RNA. During each cycle, secondary fluorescent probes hybridize to the probe with a UV-cleavable linker. After each cycle, UV light cleaves the link connecting the fluorophore to the probe. Through 16 rounds of fluorescent labeling with four fluorophores, a unique optical barcode for each RNA species is generated based on its readout sequence. CosMx uses a Cellpose algorithm to perform cell segmentation based on nuclear and cell membrane stains. The system also supports protein co-detection of up to 64+ targets allowing for multimodal spatial profiling. One comparative study suggests that it may have lower transcript sensitivity compared to platforms like Xenium, possibly due to differences in hybridization chemistry and probe design.⁶⁵

Sequential fluorescence in situ hybridization (seqFISH)^67,68 involves cycles of probing mRNA with fluorescent probes and imaging to generate a barcode readout. Each RNA species is assigned a unique barcode that is read out over multiple rounds of hybridization and imaging. The order of fluorophore signals across these rounds forms a sequence that is unique to the target mRNA. Between imaging cycles, the probes are removed to allow for the next round of hybridization. The same probe sequences are reused throughout the process, but they are conjugated to different fluorophores in each round. seqFISH+ further increases the number of possible barcode combinations by using 60 ‘pseudocolors’ generated from three fluorophore channels: Alexa Fluor 488, Cy3b, and Alexa Fluor 647. By reusing these fluorophores across 24 imaging rounds, a unique pseudocolor barcode is produced for each transcript, denoting spatial position. Barcoding and imaging are carried out in sequential steps to minimize signal crowding and decoding errors.

MERSCOPE⁶⁹ is a commercial ST platform developed by Vizgen (https://vizgen.com/) that implements Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH). In this method, each RNA species is targeted by a set of 30 to 50 oligonucleotide probes that encode a binary barcode unique to that transcript. During each imaging round, fluorescent labels are hybridized and then removed, with each RNA molecule fluorescing or not in each cycle producing a binary pattern used for transcript identification. The barcodes are decoded across imaging rounds to determine both the identity and spatial coordinates of each transcript within the tissue. MERSCOPE employs error-robust Hamming distance-based codes (typically with a Hamming distance of 4), ensuring that even with signal dropout or noise, each barcode remains distinguishable from others—minimizing decoding errors. This enables the simultaneous profiling of thousands of RNA species with high specificity. In addition, MERSCOPE can detect up to five protein targets alongside RNA molecules, allowing for integrated ST and proteomic analyses.

Molecular Cartography is a high-plex ST platform developed by Resolve Biosciences (https://resolvebiosciences.com/). For each target transcript, a set of primary probes are hybridized directly to the RNA molecule to detect up to 100 transcripts. These are followed by secondary probes containing unique barcodes, which serve as docking sites for tertiary probes labeled with fluorescent reporters. The tissue is then imaged, and the fluorescent probes are removed. This process is repeated across eight imaging cycles, with different combinations of fluorophores applied in each round. The resulting sequence of fluorescence signals generates a unique optical barcode for each target gene, enabling precise identification and localization of transcripts within the tissue.

ADDITIONAL SPATIAL TRANSCRIPTOMIC METHODS

Spatial CITE-seq⁷⁰ is a multimodal spatial omics technique that simultaneously profiles proteins and the transcriptome within intact tissue sections. It utilizes up to 300 antibody-derived tags (ADTs), each consisting of a polyA tail, a unique DNA barcode specific to the target protein, and a UMI. These ADTs bind to their corresponding protein targets within the tissue. Spatial barcoding is achieved using a microfluidic system similar to DBiT-seq, creating a 2D grid of uniquely barcoded locations. Both mRNA and ADTs within the tissue are captured and reverse transcribed into cDNA in situ. The resulting cDNA is then amplified and sequenced, allowing for the integrated analysis of target proteins and whole transcriptome.

Enhanced ELectric Fluorescence in situ Hybridization (EEL FISH)⁷¹ is a method that enhances RNA capture efficiency and spatial resolution through electrophoretic transfer of RNA from tissue onto a capture surface. The surface of the capture slide is coated in indium tin oxide, oligo(dT), and positively charged poly-d-lysine. After permeabilizing the tissue section, an electric field is applied, causing RNA molecules from the tissue to migrate onto the slide surface where they bind stably. The tissue is removed, and probes are added in 16 imaging cycles, with each RNA receiving a unique 16 bit binary barcode allowing for localization of mRNA transcripts. By analyzing the RNA on a glass surface rather than the tissue itself, lateral diffusion artifacts are greatly reduced, improving spatial fidelity and resolution.

Spatial transcriptomic analysis

Analysis of ST data generally aims to map gene expression profiles to defined spatial regions within a tissue. Unlike scRNA-seq, where each cell is uniquely barcoded, ST assigns transcripts to capture regions of varying size, making individual-cell resolution challenging. Many scRNA-seq analytical tools can be adapted for ST analysis. The number of methods for analyzing ST is substantial and growing, with several comprehensive reviews available.^51,72-74 Initial preprocessing differs somewhat between image- and sequencing-based platforms but ultimately generates gene-by-spot matrices containing transcript counts and spatial coordinates.^46,51 Subsequent analysis steps include cell segmentation, quality control, normalization, and dimensionality reduction and clustering. Spatial data can then be integrated with reference scRNA-seq datasets or other modalities for cell/spot identification. Final steps can involve spatial expression pattern identification, and analysis of cell–cell interactions.

Cell segmentation can infer single-cell transcriptomes, especially when the spatial platform has subcellular resolution, such as in imaging-based ST.⁵¹ Some commercial platforms (e.g., Xenium) provide pre-segmented data using IF staining. Many methods perform cell segmentation from cell stains, while segmentation-free methods can distinguish cell types with variable morphologies without staining and may bypass the need for explicit cell boundary detection.⁷⁵

Cellpose⁷⁶ is a cell segmentation method that uses a neural network trained on a large, diverse dataset of manually segmented images. The network predicts spatial gradients within cells, effectively creating a ‘flow field’ that directs pixels toward the cell center, enabling accurate segmentation of cells with complex morphologies. It requires prior staining and supports user-provided annotations to fine-tune the model for specific tissues. Baysor⁷⁷ identifies cells from transcript coordinates and can be implemented without prior staining. A neighborhood composition vector is calculated for each transcript based on the proximity of nearby transcripts. These neighborhoods are then embedded in a three dimensional (3D) color space, such that transcripts with similar neighborhoods display similar colors. To enhance accuracy, Baysor uses a Markov Random Field prior, assuming spatially proximate transcripts share cell labels. Factor Inference of Cartographic Transcriptome at Ultra-high Resolution (FICTURE)⁷⁸ is a segmentation-free approach that extracts factors from spatial data by first learning spatial expression patterns then assigning a factor type identity to each pixel based on the proportional identity of nearby anchor points. Each factor may correspond to a cell, a cell state, or a subcellular feature. This enables fine-scale analysis of complex tissues containing non-uniform cells. Spot-based Spatial cell-type Analysis by Multidimensional mRNA density estimation (SSAM)⁷⁹ estimates gene-specific transcript density across the tissue to construct pixel-level expression vectors. These vectors are down-sampled and clustered to define cell-type signatures, after which each pixel is classified to generate a cell-type map. Tissue domains are then identified by clustering local cell-type compositions.

Quality control and normalization methods in ST vary by platform and may involve excluding extreme values, such as transcripts below a certain threshold. Yet, this filtering can unintentionally remove biologically relevant signals that occur at low abundance.⁸⁰ Normalization is important in sequencing-based technologies to account for spot-to-spot variation in capture efficiency and/or sequencing depth across spots.⁵³ However, normalization by library size can also introduce bias, particularly in image-based platforms where gene panels are designed to target specific marker genes.⁸¹ While this effect is diminished with larger datasets, SpaNorm is a method specifically developed to account for region-specific library size effects in ST data without removing true spatial domain signals.⁸²

Dimensionality reduction (e.g., principal component analysis [PCA], uniform manifold approximation and projection [UMAP], etc.) is used to transform complex multidimensional data into a lower-dimensional space while preserving as much information as possible, enabling intuitive 2D or 3D visualization. Many tools developed for scRNA-seq dimensionality reduction can be repurposed for ST and have been summarized elsewhere.⁸³ However, specialized methods have also been developed that incorporate both spatial information and gene expression. Spatial probabilistic Principal Components Analysis (SpatialPCA) and Spatial Transcriptomics Analysis with topic Modeling to uncover spatial Patterns (STAMP) are two such approaches that explicitly integrate spatial context into dimensionality reduction, allowing for more accurate downstream analyses.^84,85

Clustering is a key step in both scRNA-seq and ST analysis, allowing users to group cells or spots based on similar gene expression patterns. A benchmarking study of clustering methods for ST found that clustering methods that include spatial information (e.g., spatial Graph Convolutional Network [SpaGCN]) can produce more accurate clustering compared to methods that do not incorporate spatial information (e.g., Leiden and Louvain).^86,87

Multiple cells often occupy the same pixel/space depending on the spatial resolution, cell morphology, and section thickness. Integration with scRNA-seq improves ST resolution, enabling clearer inference of gene expression patterns.⁸⁸ Integration is also valuable for imaging-based methods, which typically have higher resolution, but fewer genes are measured. Whole-transcriptome coverage from scRNA-seq data can be integrated with image-based ST and gene imputation can infer expression of genes not directly measured, based on shared patterns across modalities. Integration typically involves spot deconvolution to estimate cell-type composition within each capture location (spot, pixel, or ROI). Cell2location⁸⁹ uses cell-type signatures from scRNA-seq reference data to decompose spatial mRNA counts into cell-type abundance at each location. Similarly, Robust Cell Type Decomposition (RCTD)⁹⁰ calculates the mean gene expression profile for each annotated cell type in the scRNA-seq reference dataset, then fits each spot in the ST dataset as a combination of these cell types. In one benchmarking study, Cell2location, spatial dampened weighted least squares (SpatialDWLS), and RCTD outperformed other methods in accurately deconvolving spots compared to ground truth.⁹¹

Beyond scRNA-seq integration, additional tools have been developed to combine ST with other data types. These approaches often employ deep learning to integrate multiple modalities, including hematoxylin and eosin (H&E) images, spatial Assay for Transposase-Accessible Chromatin RNA sequencing (ATAC–RNA-seq), protein expression, metabolomics, and more, aiding in spatial domain identification and cellular deconvolution.⁹² MultI-modal Spatial Omics (MISO)⁹³ is one such method, designed to integrate diverse modalities with a particular emphasis on extracting information from H&E images, which are inexpensive to generate yet contain rich biological information that can be mined with machine learning to augment ST. MultiMAP⁹⁴ is another approach, developed for dimensionality reduction across multimodal datasets. It constructs a nonlinear manifold that captures relationships among data types, builds a joint neighborhood graph (MultiGraph), and projects the data into a shared low-dimensional embedding, while allowing users to assign weights to different modalities.

Spatial expression pattern identification involves detecting genes or gene programs with structured, non-random expression variation across tissue coordinates, which can be linked to tissue organization and function. Some comprehensive toolboxes, such as Giotto, include this functionality, while several specialized methods have been developed.⁹⁵ Trendsceek⁹⁶ is a nonparametric method using marked point processes to test whether a gene’s expression is randomly or non-randomly structured across a tissue. For each gene, it evaluates whether expression depends on spatial position by comparing pairwise relationships between nearby cells at different distances. SpatialDE⁹⁷ applies Gaussian process regression to compare spatial variance with spatial noise and a likelihood-ratio test determines whether including spatial structure improves the model and provides information on the magnitude of spatial expression changes. nnSVG,⁹⁸ a more recently developed method designed for large datasets, differs from previous methods by using gene-specific length scale parameters, which account for variation in smoothness and scale across genes, enabling detection of both fine-grained local patterns and broad tissue-wide trends.

Cell/cell interactions can be inferred from spatial proximity of transcripts, most often through ligand-receptor analysis or neighborhood/niche analysis. Ligand-receptor analysis can be challenging for image-based datasets with a limited number of probes (e.g., Xenium) and may be better suited for whole-transcriptome data sets. Tools such as Giotto can perform multiple steps in ST analysis, including neighborhood analysis.⁹⁵ Other programs have been specifically developed to examine cell–cell interactions. hoodscanR⁹⁹ uses the Bioconductor SpatialExperiment framework and calculates each cell’s k-nearest neighbors to compute distances. A SoftMax-based algorithm assigns probabilities of neighborhood membership, enabling visualization and analysis of cellular neighborhoods. Copulacci¹⁰⁰ models ligand-receptor co-expression between neighboring spots and is well-suited for datasets with sparse count data. It assumes a Poisson distribution for each ligand or receptor (marginals) and applies a flexible correlation model (Gaussian copula) to capture co-variation of ligand-receptor pairs. The method can also be applied in a cell type–specific manner by restricting analysis to edges connecting two annotated cell types. Niche Interactions and Cellular Heterogeneity in Extracellular Signaling (NICHES)¹⁰¹ infers cellular niches by multiplying the expression of a ligand in a sending cell with that of its receptor in nearby receiving cells. A matrix of niches is then generated by combining expression of sets of ligand-receptor pairs which can be further analyzed using Seurat or similar frameworks.

Applications in adipose tissue

Applying ST to adipose tissue is not without its challenges. Mature adipocytes are dominated by large lipid droplets that can occupy >90% of the cell volume, yet ST experiments rely on approximately 5 to 10 µm tissue sections, which can exclude the perinuclear region where most mRNA resides.¹⁰² Additionally, the cytoplasm of typical WAs is <200 nm thick, making it difficult to distinguish adipocyte signals from those of closely associated stromal cells, which, in the case of poised preadipocytes, have overlapping expression patterns.¹⁰² Furthermore, some stomal cells are interstitially located and have a dendritic-type morphology, potentially contacting multiple adipocytes over tens of microns.²⁸ Because mRNA localization is concentrated near nuclei, spatial profiles may not adequately capture these distal cell–cell interactions.

Despite these challenges, adipose tissue offers unique opportunities for spatial analysis. For example, localized sites of cell proliferation and differentiation termed ‘adipogenic niches’ are known to host complex interactions between adipose progenitors, immune cells, and adipocytes. However, the dynamics of these niches remain incompletely understood.^28,49,103 Further, there are structurally defined areas within adipose depots where neogenesis is initiated both during development and tissue remodeling. For example, cold-induced neogenesis occurs at the dorsal edge of BAT, within distinct niches, and involves close interactions between ASC and different immune subtypes.¹⁹ During obesity, CLS accumulation in WAT is associated with inflammation, fibrosis, diminished adipogenesis, and insulin resistance.¹⁰⁴ Thus, understanding the temporal organization of these structures and localized signaling may lead to a better understanding of ‘healthy’ WAT remodeling. Finally, there is clear spatial heterogeneity among adipocytes within each depot; however, the extent of this heterogeneity and implications for metabolic health remain largely unresolved.^105,106 Collectively, when coupled with advanced computational tools, these phenomena represent key opportunities for investigation using ST to produce a more comprehensive view of adipose tissue architecture (Fig. 1).

With the growing popularity of ST technology, recent studies have started to address different aspects of adipose biology, primarily utilizing the Visium platform (Table 1). For example, Bäckdahl et al.¹⁰⁷ performed ST on human subcutaneous abdominal WAT and identified 18 cell classes, including three transcriptionally distinct adipocyte populations (Adipo^PLIN, Adipo^LEP, and Adipo^SAA) that were spatially organized within the tissue. Among these, Adipo^PLIN displayed a strong transcriptional response to insulin and its calculated abundance was positively associated with metabolic indicators of insulin sensitivity. This dataset was subsequently included in a meta-analysis of human WAT that combined scRNA-seq, snRNA-seq, and ST data.¹⁰⁸ Using deconvolution and integration of multiple single-cell datasets, this study generated a comprehensive spatially resolved transcriptomic map of human WAT.

Two studies explored spatial organization of and cell signaling within CLS in the eWAT of mice fed a high fat diet (HFD).^109,110 Stansbury et al.¹⁰⁹ used ST to analyze immune cell types arising during early and late obesity (8 weeks vs. 14 weeks of HFD). They identified a subset of LAMs (‘pre-LAMs’) that arise from differentiating monocytes, co-localize with CLS early in obesity, and express multiple ligands that may further drive LAM development and CLS formation. Kohda et al.¹¹⁰ examined WAT using ST after 8 to 36 weeks of HFD to determine cell interactions within CLS during tissue fibrosis. Using scRNA-seq and histological methods, they characterized a subset of macrophages transcriptionally distinct from LAMs that express high C-type lectin domain family 4 member E (Clec4e) during the fibrotic phase. Using ST, they identified clusters that aligned with CLS and co-express markers for Clec4e, LAMs, and fibroblasts. They conclude Clec4e+ macrophages may influence ECM remodeling by suppressing collagen production in fibroblasts via secretion of oncostatin M.

Lundgren et al.¹¹¹ evaluated thermogenic responses in BAT following one or two transient cold exposures (8 hours at 4 °C) in mice acclimatized to thermoneutrality (30 °C) using snRNA-seq, metabolomics, and ST. They identified spatially organized subpopulations of BAs that contribute to ‘thermogenic memory.’ These adipocytes co-localize to the BAT periphery and induce de novo lipogenic genes (fatty acid synthase [Fasn] and stearoyl-CoA desaturase [Scd]) upon return to thermoneutrality. This lipogenic program contributes to elevated acylcarnitine levels, thereby improving cold tolerance. Lastly, Nigro et al.¹¹² used a multi-omics approach to evaluate adaptations to exercise training in the inguinal WAT of mice. They found exercise training modulates ECM-remodeling genes, reduces collagen deposition, increases vascularization and neurite density, and promotes beiging in WAT. These exercise-dependent effects were dependent on PR/SET domain 16 (Prdm16). Collectively, these studies demonstrate the potential of ST to uncover depot-specific and context-dependent remodeling of adipose tissue architecture and microenvironment interactions.

CONCLUSION

The field of ST is rapidly evolving, with new and transformative technologies emerging frequently. Many achieve high resolution and large capture areas, addressing limitations of current methods. Spatial analysis in 3D is of particular interest and can be achieved by interrogating serial tissue sections. For example, Open-ST spatially profiles sections and integrates them with H&E and/or IF staining to reconstruct tissue in 3D. Probabilistic alignment of ST experiments (PASTE)¹¹³ aligns serial sections based on expression similarities between adjacent spots and is compatible across different ST platforms. STARmap¹¹⁴ uses in situ sequencing of hydrogel-embedded tissues to generate a 3D spatial map with single-cell resolution, assigning three coordinates (x, y, and z) to each transcript. Its capabilities have been expanded with Deep-STARmap,¹¹⁵ enabling higher multiplexing. Other innovations include RIBOmap,¹¹⁶ which captures all ribosome-bound mRNAs in spatial context, and Deep-RIBOmap¹¹⁵ which provides 3D capabilities.

By localizing transcripts in their native context, ST enables the discovery of cell- and niche-specific dynamics and therapeutic targets in various tissues, providing new avenues for the treatment of metabolic diseases.^117-119 Looking forward, integrating ST with proteomics, epigenomics, metabolomics, and advanced imaging will provide a multi-level view of adipose tissue architecture, advancing our understanding of depot-specific function and plasticity and uncovering therapeutic opportunities for obesity and cardiometabolic disease.