The evolving landscape of spatial proteomics technologies in the AI age

Abstract

Although single-cell technologies have provided deep insights into cellular heterogeneity and complexity, they fall short in explaining how cells form tissue structures, a crucial aspect for understanding the principles of complex tissues. Recently, spatial transcriptomics has begun to fill this gap, allowing in situ studies of tissues at cellular and subcellular resolution. However, these genomic-level methods primarily provide indirect measurements of cellular states, as most biological processes are controlled by proteins. Therefore, spatial proteomics has the potential to revolutionize our understanding of biological processes, with significant implications for both basic cell biology and clinical applications. In this review, we provide an overview of the recent technical achievements and remaining challenges in spatial proteomics. Specifically, we categorize the techniques into three main types: antibody-based, LC-MS/MS-based, and imaging mass spectrometry-based. We describe each method in detail and discuss its strengths and weaknesses. We also discuss the emerging opportunities of artificial intelligence for spatial proteomics. Finally, we review key issues and suggest future directions for the advancement of spatial proteomics.

1. Introduction

Dissecting cellular heterogeneity within tissues is fundamental to understanding tissue organization and tumor formation [[1], [2], [3]]. While single-cell technologies have made significant advancements in characterizing cellular heterogeneity and complexity, how these cells integrate to form coherent tissue structures remains unresolved [[4], [5], [6], [7], [8]]. Recently developed spatial genomics and transcriptomics techniques have partially addressed this gap by enabling in situ studies of tissues at cellular and subcellular resolution [[9], [10], [11], [12]]. However, these sequencing-based methods primarily provide indirect measurements of cellular states, as most biological processes are controlled by proteins [13]. The intricate relationship between proteins and transcripts is highly dependent on the tissue microenvironment, posing challenges for the comprehensive biological interpretation of tissue heterogeneity [14,15] Additionally, conventional nucleic acid measurements cannot reveal post-translational modifications of proteins, which play crucial roles in various biological processes. Therefore, high-throughput spatial proteomics holds the potential to revolutionize our understanding of biological processes, with significant implications for both fundamental cell biology and clinical applications [16].

Achieving high-throughput spatial proteomics has been regarded as one of the most important yet difficult-to-obtain spatial omics techniques. Researchers have developed multiplexed protein immunostaining methods by labeling antibodies with nucleotide chains, fluorescent groups, or metal tags, allowing for the simultaneous mapping of up to hundreds of proteins in the same tissue section [17,18]. Mass spectrometry (MS)-based proteomics enables label-free analysis with high specificity and deep proteome coverage and has recently been applied to enhance proteome resolution to the single-cell level [19,20]. Complementary to MS, laser capture microdissection (LCM) can isolate regions of interest within tissues and subsequently provide comprehensive proteomic abundance profiles [21,22]. In recent years, with the increasing attention on spatial omics technologies, more spatial proteomics techniques have been reported.

Despite the numerous spatial omics techniques reported in recent years, many reviews have primarily summarized recent advances in spatial transcriptomics and their applications, with fewer focusing on spatial proteomics. Therefore, in this review, we summarize recent developments in spatial proteomics technologies and their applications. We categorize current spatial omics technologies into three groups: antibody-based, LC-MS/MS-based, and imaging MS-based (Fig. 1; Table 1). We provide an overview of recent technological advancements, their advantages and disadvantages, and application progress for each category. Then, a summary of advances in artificial intelligence for spatial proteomics is provided. Finally, we discuss future directions for the development of spatial proteomics technology.

Fig. 1.

An overview of the main spatial proteomics methods. Current spatial proteomics methods can be categorized into three main types: imaging mass spectrometry (IMS) based, liquid chromatography with tandem mass spectrometry (LC-MS/MS) based, and antibody based. Each of these categories involves different strategies. The IMS-based approach encompasses both top-down and bottom-up proteomics strategies. The LC-MS/MS-based approach includes pixel-by-pixel analysis and region of interest (ROI) strategies. The antibody-based approach utilizes antibodies labeled with DNA oligonucleotides, metal molecules, or fluorophores.

Table 1.

Summary of different spatial proteomic technologies.

Category	Approaches	Label	Type	Resolution	Plex	Multiomics	Specialized requirement
Antibodies-based	t-CycIF	Fluoro-tag	Targeted	Subcellular	< 60	No	Fluorescence microscopes
	4i	Fluoro-tag	Targeted	Subcellular	> 40	No	Fluorescence microscopes
	IBEX	Fluoro-tag	Targeted	Subcellular	> 65	No	Fluorescence microscopes
	MELC	Fluoro-tag	Targeted	Subcellular	< 100	No	Fluorescence microscopes
	MxIF	Fluoro-tag	Targeted	Subcellular	61	No	Fluorescence microscopes
	MIBI	Metal-tag	Targeted	Subcellular	< 100	No	TOF mass spectrometer with ion beam
	CyTOF-IHC	Metal-tag	Targeted	Subcellular	32	No	Conventional microscopes, mass cytometer
	MIBI-TOF	Metal-tag	Targeted	Subcellular	36	No	Secondary ion mass spectrometer
	SABER-IMC	DNA-tag	Targeted	Subcellular	38	No	Fluorescence microscopes, mass cytometer
	CODEX	DNA-tag	Targeted	Subcellular	> 60	Yes	Fluorescence microscopes
	Immuno-SABER	DNA-tag	Targeted	Subcellular	38	No	Fluorescence microscopes
	DSP	DNA-tag	Targeted	Multicellular	55	Yes	Specialized equipment
	SPOT	DNA-tag	Targeted	Multicellular	32	Yes	Not required
	Spatial-CITE-seq	DNA-tag	Targeted	Multicellular	273	Yes	Not required
Imaging MS-based	Bottom-up	Label-free	Untargeted	Multi to subcellular	100 s	No	Mass spectrometer
	Top-down	Label-free	Untargeted	Multi to subcellular	100 s	No	Mass spectrometer
LC-MS/MS-based	nanoPOTS	Label-free	Untargeted	Multicellular	> 2000	No	Fluorescence microscopes, mass spectrometer
	MASP	Label-free	Untargeted	Multicellular	> 5000	No	3D-printed scaffold, mass spectrometer
	LCM-SISPROT	Label-free	Untargeted	Multicellular	> 500	No	Laser capture microdissection, mass spectrometer
	DVP	Label-free	Untargeted	Multicellular	1000 s	No	Fluorescence microscopes, mass spectrometer
	scDVP	Label-free	Untargeted	Cellular	1000 s	No	Fluorescence microscopes, mass spectrometer
	SCPro	Label-free	Untargeted	Cellular	1000 s	No	Fluorescence microscopes, mass spectrometer

2. Antibody-based high-plex spatial profiling

Antibodies are widely utilized to detect protein distribution by translating target information into detectable signals, such as fluorescence or chromogens, which is crucial for analyzing cell composition and state, tissue heterogeneity, and intercellular interactions. Conventional immunohistochemistry (IHC) or immunofluorescence (IF) methods achieve protein localization through various fluorophores or enzymes linked to antibodies. However, due to limitations in antibody species, the number of available fluorophores or chromogens, tissue autofluorescence, and spectral overlap, a single imaging session can typically detect no more than seven markers simultaneously even under optimized conditions [23]. Considering that more markers are required for precise characterization of cellular heterogeneity and functional mapping of complex tissues like tumors, many multiplexed antibody-based imaging methods have been developed [17,[24], [25], [26], [27]]. These methods vary significantly in their immunolabeling patterns, detection methods, and imaging strategies. Beyond the few methods utilizing enzyme-catalyzed chromogenic detection and Raman spectroscopy such as tyramide signal amplification-based methods [25,28] and stimulated Raman scattering microscopy [29], most approaches can be categorized based on the type of antibody labeling used: fluorophore-labeled, metal mass tag-labeled, or DNA oligonucleotide-labeled (Fig. 2; Table 1).

Fig. 2.

Schematic workflow of antibody-based proteomics. General workflow for antibody-based spatial proteomics by different labeled antibodies, including (A) fluorophore-labeled antibodies, (B) DNA oligonucleotide-labeled antibodies, and (C) metal-labeled antibodies. (D) Recently reported antibody staining-based proteomic approaches categorized by two parameters: antibody labeling strategies and antibody signal detection methods.

2.1. Fluorophore-labeled antibody for spatial profiling

Due to the widespread availability of fluorescent antibodies and the requirement for only conventional imaging systems, applications based on fluorescently labeled antibodies for multiplex protein imaging have been well established. While a few single-pass multiplexed imaging methods exist [29,30], most multiplexed protein imaging techniques rely on multiple rounds of antibody staining and imaging (Fig. 2A). This multi-step process involves immunostaining with primary or secondary antibodies conjugated to fluorophores, fluorescent image acquisition, and fluorescence signal quenching by chemical/photo inactivation or antibody removal. By repeating the staining-imaging cycles, these methods can leverage a limited number of antibody and fluorophore pairs to circumvent the limitations of spectral overlap, thus enabling comprehensive multiplexed imaging. Prominent examples of such techniques include Array Tomography (Array-Tomo) [31], multi-epitope-ligand-cartography (MELC) [32], multiplexed immunofluorescence (MxIF) [33], cyclic immunofluorescence (CycIF) [34], sequential immunohistochemistry (S-IHC) [35], tissue-based cyclic immunofluorescence (t-CyCIF) [36,37], iterative indirect immunofluorescence imaging (4i) [38], and iterative bleaching extends multiplexity (IBEX) [39,40]. The development and application of these methods have significantly expanded the number of targets that can be simultaneously detected on a single tissue section. These techniques offer extremely high resolution, providing subcellular and even organelle-level structural information [38], thereby deepening our understanding of the molecular architecture and function of different organs [31,41] and the tumor immune microenvironment [33,[35], [36], [37]].

Despite their theoretical capability to analyze an unlimited number of targets, several factors limit the number of proteins that can be analyzed or the quality of the data. Firstly, single rounds of staining typically take hours and can stain a maximum of 5–7 proteins, making the duration extremely long when many targets are involved. Secondly, prolonged tissue processing can alter antigenicity and cause tissue degradation, leading to signal attenuation. Additionally, incomplete fluorescence quenching in iterative staining cycles can increase background fluorescence and introduce artifacts, reducing the data quality. Therefore, incorporating appropriate controls during experiments is crucial for the reliability of conclusions. As the number of antibodies increases, cross-reactivity between antibodies imposes higher demands on antibody specificity and sample processing. Beyond these limitations, spatial proteomics based on fluorescently labeled antibodies also face sensitivity constraints, such as differences in protein abundance spanning seven orders of magnitude, which exceed the dynamic range of fluorescent antibody detection [26,42]. Furthermore, the analysis, integration, and interpretation of this high-dimensional information pose new challenges.

2.2. DNA-labeled antibody for spatial profiling

Unlike multi-round staining and imaging based on fluorescently labeled antibodies, DNA-labeled antibody methods offer robust encoding capabilities and scalability. The experimental workflow typically includes a single immunostaining step, followed by the hybridization and removal of complementary fluorophore-labeled oligonucleotides for rapid target readout by imaging (Fig. 2B). Notable examples of such techniques include multiplexed points accumulation for imaging in nanoscale topography (Exchange-PAINT) [43], DNA exchange imaging (DEI) [44], co-detection by indexing (CODEX) [45], and FFPE-CODEX [46]. Some methods enhance signal-to-noise ratio and increase the detection capability for low-abundance proteins through signal amplification. For instance, immunostaining with signal amplification by exchange reaction (immuno-SABER) achieves up to 180-fold signal amplification in diverse samples by hybridizing one or multiple rounds of concatemers based on DEI [47]. However, the length of the concatemers (300–700 bp) poses challenges in penetrating thicker tissues. Immunosignal hybridization chain reaction (isHCR) utilizes small-sized hairpins for easy penetration the tissues, and achieves signal amplification through the assembly of a linear DNA structure by iterative hairpin openings triggered by an initiator sequence [48]. HCR presents a novel staining protocol addressing the non-specific binding of DNA-conjugated antibody and expands the hairpin pool [49]. Multi-omics in situ pairwise sequencing (MiP-seq) amplifies signals via rolling circle amplification, increasing detection throughput through barcode encoding [50]. In addition to labeling DNA imager with fluorophores, metal isotopes can also be used. For example, Hosogane et al. combined immuno-SABER with imaging mass cytometry (SABER-IMC), utilizing 38-plex metal-labeled DNA imagers to image the tumor immune microenvironment in human melanoma, identifying previously undetected immune cell markers, such as T cell co-receptors and their ligands [51].

With the advancement of spatial sequencing technologies, new high-throughput methods for detecting protein distribution have been developed. These methods can obtain the spatial distribution of tens to hundreds of proteins with just one round of staining and detection. Typically, the DNA label on each antibody contains a poly(A) tail, a unique molecular identifier, and a barcode sequence unique to the corresponding antibody. This label is tagged with spatial information during subsequent reverse transcription or ligation [18,52] or photocleaved and collected based on region of interest (ROI) [53]. After sequencing, the DNA label and spatial barcode information are converted into the protein's spatial distribution. Ben-Chetrit et al. combined 32-plex protein detection with the commercial spatial transcriptomics platform 10x Visium to develop spatial protein and transcriptome sequencing (SPOTS), successfully applying it to parse two spatially distinct macrophages in the breast cancer microenvironment [52]. Liu et al. used spatial co-indexing of transcriptomes and epitopes (Spatial CITE-seq) to profile 189 proteins in mouse tissues and 273 proteins in human tissues, analyzing spatially specific germinal center reactions in tonsils and immune activation responses in the skin near COVID-19 mRNA vaccine injection sites [18]. Brady et al. utilized GeoMx DSP technology to spatially resolve gene and protein expression in metastatic prostate cancer samples, revealing intra- and inter-tumor heterogeneity [53]. Although sequencing-based methods require additional spatial barcoding or ROI selection which results in lower resolution compared to fluorescence imaging methods, their high throughput, low cost, and lack of need for specialized equipment, as well as their freedom from the limitations of fluorescent channels or metal tags, enable the effective acquisition of a large amount of target information in a single detection. Furthermore, these methods could integrate spatial protein distribution with gene expression, providing a more comprehensive molecular characterization of the cellular composition and functional maps within tissues.

2.3. Metal-labeled antibody for spatial profiling

Another antibody-based high-plex spatial profiling method involves conjugating antibodies with ionizable metal isotope tags and coupling with mass spectrometry, also referred to as next generation IHC [54]. After antibody staining, metal tags are released from the tissue sample and detected by a mass spectrometer. Then mass spectrometry signals are converted to protein abundance information (Fig. 2C). This approach requires only a single round of staining to detect over 40 markers, offering fast imaging speeds, low background noise, and high resolution. The two most common methods in this category are multiplex ion beam imaging (MIBI) and imaging mass cytometry (IMC). The primary difference between these methods lies in their ionization and detection mechanisms. MIBI uses a rasterized primary ion beam to strike the samples, liberating the metal adducts of the bound antibodies as secondary ions, which are then detected by a magnetic sector. With the dynamic range being approximately 1 log for IHC and ∼2.5 logs for quantitative IF [54], MIBI enables the analysis of up to 100 markers simultaneously over a five-log dynamic range and provides valuable insights into the pathogenesis of breast cancer [55]. The MIBI-TOF instrument was developed to increase channel multiplexing and reduce acquisition times by 50-fold, enabling the identification of structural features of the tumor-immune microenvironment in a retrospective cohort of 41 triple-negative breast cancer patients and resolutions down to 260 nm with near single-molecule detection sensitibity [56,57]. Rovira-Clavé et al. further advanced the resolution to 30 nm with high-definition MIBI (HD-MIBI), correlating the subcellular localization of the chemotherapeutic drug cisplatin with five distinct subnuclear structures in three dimensions (3D) [58]. IMC, on the other hand, uses high-resolution laser ablation to introduce metal tags into the mass spectrometer (usually Time-of-Flight, TOF) for analysis, similar to mass cytometry for single-cell analysis [59]. Giesen et al. coupled CyTOF mass cytometry with IHC (CyTOF-IHC) to simultaneously image 32 proteins and their modifications at subcellular resolution, applying IMC to human breast cancer samples to delineate cell subpopulations and interactions [60]. Schulz et al. combined RNAscope-based metal in situ hybridization with IMC detection of 16 proteins to analyze 70 breast cancer samples, revealing the variance of mRNA-to-protein correlations in different genes and patients [61].

Unlike other multiplexed methods, high-resolution metal-based mass spectrometry approaches require samples to be stable in a vacuum environment, at least during imaging. Additionally, the inability to amplify signals reduces the sensitivity for detecting low-abundance proteins, particularly in samples with target degradation due to harsh fixation. Since IMC or MIBI is a destructive technique, it prevents downstream analysis of the same section and the throughput is limited by the number of isotopically pure metals available. Compared to conventional fluorescence microscopy, mass spectrometers require more specialized operation, accurate calibration, and a stable environment. In addition, imaging speed and resolution limit the achievable imaging area, with higher resolutions requiring longer times.

Proteins exist and interact in a three-dimensional (3D) space within tissues and cells, making high-dimensional investigations of protein distribution more reflective of real biological conditions. 3D volume provides information that two-dimensional scans cannot capture, such as long-distance interactions or rare cell types. Imaging methods based on antibody staining are well-suited for 3D reconstruction through Z-axis stackings [39,43,44]. Additionally, MIBI-based methods can achieve continuous high-resolution 3D imaging by controlling the depth of ion beam bombardment [55]. As the thickness of the tissue increases, the penetration ability of antibodies, the efficiency of multiple rounds of staining removal or inactivation, and the imaging quality decline. To this end, various tissue clearing and preservation methods were developed to preserve antigen spatial distribution and permit effective antibody staining, such as clear lipid-exchanged anatomically rigid imaging/immunostaining-compatible tissue hydrogel (CLARITY) [62], clear, unobstructed brain imaging cocktails and computational analysis (CUBIC) [63], magnified analysis of the proteome (MAP) [64], clearing-enhanced 3D (Ce3D) [65], and system-wide control of interaction time and kinetics of chemicals (SWITCH) [41]. Despite these advancements, three-dimensional imaging of thicker tissues remains challenging. For instance, staining a 1 mm brain section might require several days [65].

In summary, while antibody-based high-plex spatial profiling methods has made significant strides (Fig. 2D), addressing its intrinsic limitations through technological advancements and methodological improvements will be key to unlocking its full potential. The integration of more robust antibodies, advanced imaging techniques, and comprehensive data analysis tools will pave the way for more accurate and high-resolution spatial proteomics, facilitating deeper insights into the molecular architecture and function of tissues.

3. LC-MS/MS-based spatial proteomics

Mass spectrometry (MS)-based proteomics has emerged as a pivotal tool for protein identification and biological functions investigation [66]. Advances in the sensitivity of liquid chromatography with tandem mass spectrometry (LC-MS/MS) systems have dramatically improved proteomic coverage and detection speed, enabling the analysis of thousands of proteins from a single cell [[67], [68], [69], [70], [71], [72]]. This progress has facilitated the development of spatially resolved proteomics by combining high-sensitivity MS with micro-manipulation techniques, such as laser capture microdissection (LCM). These approaches provide label-free analyses with high specificity, unbiased, and deep coverage for spatial proteomics [66]. The method can be categorized into two strategies: pixel-by-pixel-based spatial proteomics and ROI-based spatial proteomics (Fig. 3; Table 1). The following sections provide a detailed summary of the workflows, technological developments, and key features of these techniques.

Fig. 3.

Summary of LC-MS/MS-based spatial proteomics techniques. (A, B) Schematic diagrams of two spatial sampling strategies: pixel-by-pixel (A) and region of interest (ROI, B). (C) Typical sample preparation workflow for LC-MS/MS-based spatial proteomics.

3.1. Pixel-by-pixel strategy for LC-MS/MS-based proteomics

Pixel-by-pixel spatial proteomics involves dividing an entire tissue section into uniformly sized tissue voxels, and detecting the proteome in each voxel using LC-MS/MS (Fig. 3A) [73]. The workflow consists of four main steps (1) Tissue section segmentation: The tissue section is segmented into a regular grid shape using LCM or micromanipulation, with each tissue voxel individually collected in a microtube; (2) Proteomic sample processing: All voxels are processed using microprotein sample preparation workflows that digest the voxels into peptides that can be detected by LC-MS/MS; (3) LC-MS/MS detection: The highly sensitive mass spectrometer detects the proteome in each voxel sample; and (4) Data analysis: Bioinformatics analysis is used to visualize the spatial map of different proteins (Fig. 3C). This method has several advantages: it is unbiased, allowing non-targeted detection of protein spatial distribution and enabling de novo discovery of protein spatial variation; it has high coverage, with the high sensitivity of mass spectrometry allowing detection of thousands of proteins in each voxel; and it is applicable at the whole tissue level, allowing analysis of spatial proteome expression of whole tissue sections.

The pixel-by-pixel method dates back to 2007 when Petyuk et al. utilized microdissection technology to slice a mouse brain section into 1 mm × 1 mm voxels [74]. A 10 µm thick, approximately 7 mm × 11 mm coronal brain slice was divided into 71 voxels, with an average of 1000 proteins detected per voxel. This method visualized the spatial expression patterns of over a thousand proteins in the mouse brain, opening up new possibilities for studying the spatial brain proteome and its dynamics during disease progression. With the advent of high-sensitivity mass spectrometry and the development of microprotein sample preparation workflows, the pixel-by-pixel spatial proteomics technique has seen significant improvements in proteome coverage and spatial resolution. In 2020, Piehowski et al. utilized their newly developed micro-protein processing platform, NanoPOTS, in combination with laser capture microdissection technology to successfully detect over 2000 proteins in 100 µm × 100 µm voxel samples [75]. This method was applied to analyze the spatial proteome of uterine tissue by collecting 24 contiguous 100 µm × 100 µm voxels and analyzing proteins enriched in the stroma and luminal epithelium.

To improve the efficiency of micromanipulation, Ma et al. in 2022 developed a robust and reliable micro-sampling method using 3D-printed micro-scaffolds (MASP), enabling uniform compartmentalization of whole tissue slices while precisely preserving spatial information [76]. This method divided a 1-mm-thick mouse brain slice into hundreds of 400 µm × 400 µm voxels, with over 5000 proteins detected in each voxel. Using this method, the researchers explored the spatial distribution maps of key proteins involved in Alzheimer's disease in the mouse cerebral cortex. In 2023, Davis et al. further improved the spatial resolution to 40 µm [77]. They employed a three-length-scale workflow within serial sections. In three adjacent tissue sections, the first section was divided into 800 µm × 800 µm voxels, part of the second section was divided into 350 µm × 350 µm voxels, and part of the third section was divided into 40 µm × 40 µm voxels. This approach allowed to obtain spatial proteomics at the whole tissue level at a lower cost, while achieving high-resolution spatial proteomics in regions of interest. This method was applied to analyze the spatial proteome of a human brain tumor, and further examined protein expression in the vicinity of blood vessels at high resolution.

3.2. ROI strategy for LC-MS/MS-based proteomics

Although pixel-by-pixel spatial proteomics techniques provide spatial information at the whole tissue level, they are limited by the throughput and cost of mass spectrometry, requiring a trade-off between tissue area and resolution. ROI-based spatial proteomics offers a more economical approach, requiring significantly fewer protein samples than the pixel-by-pixel method and allowing analysis of spatial proteome differences at single cell level (Fig. 3B). The main workflow of this technique involves obtaining images of tissue sections using H&E staining, immunofluorescence staining, or imaging mass spectrometry to collect histological information. Next, ROIs for dissection are identified using artificial intelligence, big data analytics, or annotation by expert pathologists. These selected regions are then collected using a laser microdissection system. Finally, protein samples are prepared using microprotein sample preparation techniques and analyzed using high sensitivity mass spectrometry, followed by data processing and analysis (Fig. 3C).

In 2018, Xu et al. developed a spatial proteomics method called LCM-SISPROT, which allows for the selection of specific pathological regions based on H&E staining for spatial proteome analysis [78]. Specifically, tissue slides were initially subjected to H&E staining, and the stained slides were further evaluated for pathological and cellular characteristics by experienced pathologist. Using annotated morphological features, regions of interest were dissected using a Leica LMD system. H&E staining dyes could be selectively removed by the unique two-stage design of the Spintip device. This process enabled the successful acquisition of spatially resolved proteome distribution profiles for enterocytes, lymphocytes, and smooth muscle cells on the same tissue slices and across four consecutive sections separated by micrometer distances. In this report, they were able to identify more than 500 proteins from a colon cancer tissue section as small as 0.1 mm² and 10 µm thick. ROI-based strategies have also been successfully applied in fruit, where laser capture microdissection combined with nanodroplet sample preparation identified a total of 1870 protein groups and covered the different tissues [79].

Recently, Mund et al. have pushed ROI-based spatial proteomics to the single-cell level with a new method called “Deep Visual Proteomics” (DVP) [20]. This technique combines sub-micron resolution microscopy, artificial intelligence, and an ultra-sensitive proteomics workflow to characterize individual tumor cells. To accurately define single cell boundaries and cell classes, they introduce a software ‘BIAS’ to coordinate scanning and laser microdissection (LMD) microscopes. By quantifying thousands of proteins in tumor cells in an unbiased manner, DVP reveals the mechanisms that drive tumor evolution and identifies novel therapeutic targets. Specifically, DVP integrates four technological advances into a single workflow: first, advanced microscopy generates high-resolution tissue images; second, machine learning algorithms precisely classify cells before laser microdissection and single-cell collection; and third, mass spectrometry then analyzes the specific types of normal or diseased cells, creating protein maps that elucidate disease mechanisms. This technique allows researchers to effectively correlate the physiological characteristics of cells observed under the microscope with their protein functions, linking protein abundance to complex cellular or subcellular phenotypes while preserving spatial context.

Rosenberger et al. further developed the deep visual proteomics (DVP) technique to single-cell deep visual proteomics (scDVP) [19]. A modular and automated workflow was established to effectively capture single-cell proteomics. In this way, scDVP was able to identify a maximum of 2769 proteins in a single form cell. When scDVP was applied to murine hepatocytes, half of the proteome was found to be spatially differentially regulated around the central vein. Recently, Xu et al. developed a spatial and cell-type proteomics platform (SCPro) by combining image-guided spatial proteomics with flow cytometry-based cell-type proteomics to reveal cell-type heterogeneity within tissues [80]. Inspired by deconvolution algorithms in spatial transcriptomics, they used single cell proteomics data from FACS to deconvolute the cell type composition in LCM-based proteomics. In this manner, SCPro discovered a novel sub-cell type of immune cell, providing a promising capability for biological discovery.

3.3. Advances in the application of LC-MS/MS-based spatial proteomics

Although the development of spatial proteomics has lagged behind that of spatial transcriptomics, there have been numerous impressive biological applications of spatial proteomics. These applications primarily include the generation of protein atlases [81], the visualization of tissue microenvironments [82], the discovery of disease-related protein biomarkers [83], and the identification of novel cell subpopulations [80]. As an illustrative example, Eckert et al. developed an ROI-based spatial proteomics workflow to analyze as few as 5000 formalin-fixed, paraffin-embedded (FFPE) cells of ovarian cancer [84]. With this approach, they demonstrated the use of ultra-low-input proteomics to identify potential drivers of disease phenotypes and revealed that nicotinamide N-methyltransferase (NNMT) is a master metabolic regulator of cancer-associated fibroblasts. Another example of ROI-based spatial proteomics is its use in the study of sub-tumor microenvironments, which showed that these environments influence regional epithelial and immune phenotypes, thereby shaping key clinical metrics of disease progression [82]. In addition, mass spectrometry-based spatial proteomics has been applied to human tuberculosis granulomas, human skin, and COVID-19-related pulmonary injury [85].

An important application of spatial proteomics is the integrative analysis of spatial multi-omics. By combining spatial proteomics with other spatial omics techniques, such as spatial transcriptomics, spatial metabolomics, and spatial epigenomics, researchers can simultaneously explore their biological question from multiple dimensions. For instance, Fan et al. conducted a study integrating spatial transcriptomics, spatial proteomics, and single-cell sequencing to investigate cervical squamous cell carcinoma [86]. This approach successfully identified biologically and clinically relevant cellular ecosystems within the cancer. Similarly, Schweizer et al. utilized the spatial proteomics technique DVP alongside the spatial transcriptomics technology GeoMx to analyze ovarian cancer, revealing the molecular landscape of borderline ovarian tumors and their progression to invasiveness [87]. These examples demonstrate the power of combining spatial proteomics with other spatial omics techniques to provide a comprehensive understanding of complex biological systems and disease mechanisms.

4. Imaging mass spectrometry-based spatial proteomics

Imaging mass spectrometry (IMS)-based spatial proteomics is a powerful label-free approach for untargeted profiling of the spatial distribution of proteins within biological tissues [88,89]. Unlike targeted spatial proteomics techniques such as CyTOF, which rely on labeled antibodies for imaging mass spectrometry and are therefore costly and labor intensive, IMS provides a direct method for mass imaging of charged peptides or proteins [17]. In IMS, mass spectra of charged molecules are collected at specific x,y coordinates, creating signal intensity maps that represent ion images across the sample area [90]. This technique is useful for studying post-translational modifications (PTMs), protein structure, and splice isoforms [91]. IMS-based spatial proteomics can be divided into two main types: top-down proteomics and bottom-up proteomics (Fig. 4; Table 1). The primary difference between these methods is that bottom-up proteomics requires an additional in situ tryptic digestion step to break down larger proteins into smaller peptides prior to mass imaging. Although IMS-based spatial proteomics is a promising tool for exploring tissue heterogeneity, it faces challenges related to protein identification, proteome coverage, and spatial resolution. Recent advances aim to address these issues, and we will discuss these novel methods in detail.

Fig. 4.

Overview of imaging mass spectrometry-based spatial proteomics techniques. (A, B) Schematic diagrams of two mass spectrometry imaging strategies: spatially resolved top-down proteomics and spatially resolved bottom-up proteomics. (C) Principle of mass spectrometry imaging, including (i) workflow of spatial mass spectrometry imaging techniques and (ii) representative imaging results. Fig. 4C is reprinted with permission from Groseclose et al. Copyright 2007, John Wiley & Sons.

4.1. Spatially resolved top-down proteomics

Top-down proteomics is a mass spectrometry (MS) technique designed to analyze proteoforms, which are proteins characterized by specific PTMs and amino acid variants [92]. When integrated with matrix-assisted laser desorption ionization (MALDI) imaging mass spectrometry (IMS), spatial top-down proteomics enables the mapping of molecular distributions within tissues, providing detailed insights into the spatial functions of proteins. Advances in this field have largely been driven by IMS techniques, such as MALDI, nanospray desorption electrospray ionization (nanoDESI), and liquid extraction surface analysis (LESA) [[93], [94], [95]]. Although top-down spatial proteomics offers a unique solution for the elucidation of PTMs, an area where bottom-up proteomics and antibody-based methods often struggle, it still faces significant challenges in terms of protein identification, proteome coverage, and spatial resolution.

Several innovative strategies have been developed to address these challenges. To improve spatial mapping of protein, researchers have improved spatial mapping of histones by coupling a MALDI-IMS source to a high-field ultrahigh mass range (UHMR) Orbitrap [96]. In this manner, they could directly analyze proteoforms from a human kidney tissue section with capability up to 16.5 kDa at near-cellular resolution. Similarly, the combination of a nanoDESI platform with ion mobility spectrometry has resulted in improved proteome coverage [97]. In terms of increasing the maximum mass range, a novel caffeic acid matrix for MALDI-IMS has been introduced, enabling high-molecular-weight protein imaging up to 200 kDa [98]. Improving the signal-to-noise ratio of imaging is also critical [99]; improvements in ion transfer efficiency have been made, and a nanoDESI-based top-down proteomics method with higher ionization efficiency has enabled profiling of larger proteins up to 47 kDa [100].

Despite many advances in MALDI-IMS, protein identification and confirmation remain challenging [101]. To overcome this, several strategies have been developed. High mass resolution Fourier Transform (FT) mass spectrometers are used for IMS, providing mass accuracy higher than 5 ppm [102]. However, due to their low image acquisition rate, combination with faster imaging platforms such as MALDI-TOF and quadrupole-reflectron time-of-flight (qTOF) mass spectrometers offers a potential solution [103,104]. Another approach is to perform additional liquid chromatography (LC)-MS/MS measurements on an adjacent tissue section to tentatively identify intact proteins. Recently, microsampling technologies such as microLESA, liquid microjunction microextraction (LMJ), and laser capture microdissection (LCM) have been integrated to collect protein samples from the adjacent tissue section [105,106]. Then, the protein samples extracted from adjacent tissues can be analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS) to generate reference databases for MSI. Using this spatial top-down approach, a significant number of proteins can be identified in a single LC-MS/MS experiment.

4.2. Spatially resolved bottom-up proteomics

Spatially resolved bottom-up proteomics involves the enzymatic digestion of intact proteins into smaller peptides for downstream mass imaging. This method allows for more comprehensive sequence coverage by facilitating peptide fragmentation [107]. The workflow for spatially resolved bottom-up proteomics typically consists of four main steps: washing the tissue section to remove abundant low-molecular-weight molecules, applying trypsin to the section using a spray device, incubating for in situ digestion of the tissue, and performing mass imaging detection (Fig. 4C) [108]. This approach is more promising for FFPE samples than top-down spatial proteomics [109]. In FFPE samples, protein cross-linking due to PFA (paraformaldehyde) fixation and antigen retrieval can pose significant challenges for spatially resolved top-down proteomics. In contrast, the smaller size of digested peptides results in higher ion transmission and ionization efficiency compared to intact proteins. Therefore, spatially resolved bottom-up proteomics is a better choice for clinical FFPE samples [110].

Spatially resolved bottom-up proteomics holds great promise for acquiring spatial distribution information of proteins in biological tissue, but it is easily influenced by several key factors. Optimizing the on-tissue digestion workflow is critical for mass imaging, and several approaches have been proposed in recent years [111]. For example, the addition of detergent to the digestion workflow has been shown to increase peptide yield [[112], [113], [114], [115]]. Ensuring reproducibility of tryptic digestion in humidity chambers is critical, especially when comparing different samples within an experiment. In addition, the amount of trypsin per area of tissue is important, and due to tissue heterogeneity, the optimal amount of trypsin may vary. To increase proteome coverage, using a complementary proteases mixture of trypsin, Lys-C and Arg-C to digest tissue in situ has been reported [116]. Furthermore, this workflow also suitable for glycogen imaging by applying enzymes such as isoamylase and glycosidase F to digest glucose chains on tissue [117]. Thus, it is feasible to analyze the spatial distribution of both proteins and glycogen on the same tissue section [118].

Peptide identification is also a challenge for spatially resolved bottom-up proteomics because proteins cannot be identified by their m/z values alone. The lack of bioinformatic analysis methods further exacerbates this problem. Combining IMS with LC-MS/MS is a direct method to improve peptide identification [119]. In this method, two adjacent tissue sections are used: one for mass imaging and the other for LC-MS/MS analysis [120,121]. To minimize variation, it is preferable to use the same mass spectrometer to obtain IMS and LC-MS/MS data simultaneously. In addition, high-resolution mass spectrometry, such as MALDI FTICR imaging, has been used to increase confidence in peptide identification [[122], [123], [124]]. Ion mobility is another promising approach to increase confidence in IMS identifications [125,126]. Recently, some open-source bioinformatics tools, such as HIT-MAP and ImShot, have been developed to aid in peptide and protein annotation, providing valuable resources for spatially resolved bottom-up proteomics [127,128]. Collectively, these advances contribute significantly to the reliability and accuracy of spatially resolved bottom-up proteomics.

5. Artificial intelligence for spatial proteomics

Spatial proteomic technologies generate vast amounts of high-dimensional data, presenting significant challenges for data processing and interpretation. Key issues include variations in resolution, signal intensity, analytical protocols, and the need for specialized equipment and software, all of which complicate data integration across different studies. Additionally, the reliance on manual processing for iterative optimization and fine-tuning is not only time-consuming but also prone to subjective variability, potentially undermining the reproducibility and reliability of the conclusions drawn. In this context, artificial intelligence (AI) has emerged as a transformative force in spatial proteomics, providing robust solutions to these challenges through the automation, high-throughput capabilities, and standardization of analysis workflows. AI techniques are especially effective in handling high-dimensional data and encompass a wide range of applications, from data preprocessing to downstream analysis. These advancements are critical for enhancing disease diagnosis, prognosis prediction, and precision medicine, thereby enabling a deeper understanding of tumor microenvironments and cellular interactions. The following sections will provide a concise overview of deep learning (DL) applications in cell segmentation, cell type annotation, and the processing of proteomic data (Fig. 5).

Fig. 5.

Applications of artificial intelligence in spatial proteomics. Overview of AI applications in spatial proteomics, encompassing cell segmentation, cell annotation, and mass spectrum analyses.

Unlike single-cell technologies, most spatial proteomic experiments are conducted on intact tissues. To accurately comprehend cellular composition and intrinsic interactions, it is essential to assign each pixel to a specific cell. Accurate and automated cell segmentation is a foundational step, serving as the basis for subsequent analyses. Machine learning-based tools have proven invaluable for image segmentation, demonstrating superior performance in differentiating cell shapes, sizes, and densities [129,130]. These methods can be categorized into traditional machine learning [[131], [132], [133], [134]], convolutional neural network (CNN)-based approaches [[135], [136], [137], [138]], and R-CNN-enhanced DL methods [[139], [140], [141]]. Traditional machine learning techniques, such as ilastik [132], typically involve feature extraction from images to identify cell boundaries. However, these approaches often require extensive manual adjustments for accurate segmentation and tend to be less effective in complex environments. CNN-based DL methods, on the other hand, automatically learn features from raw imaging data and have shown remarkable success in delineating cellular structures within multiplexed fluorescence images. For example, DeepCell [136] enables robust identification of a cell's cytoplasm with single-cell resolution, though it necessitates prior training on new cell types. Both Mesmer [135] and Cellpose [138] are trained on diverse datasets, including images of cells and nuclei, to generalize more effectively. StarDist [137] and Cellpose [138] utilize U-Net, a specialized CNN architecture characterized by its U-shaped encoder-decoder structure and specifically tailored for image segmentation, to enable high effectiveness and precise feature localization. Based on Mask R-CNN, Cellseg[139] offers generalization, extreme accuracy, and sensitivity for large tumor cells, requiring no retraining. Additionally, methods such as UnMICST [140] and nucleAIzer [140] combine Mask R-CNN-based instance segmentation with semantic segmentation networks, demonstrating robust performance across various tissues and cell types.

Following cell segmentation and measurement, the precise delineation of cellular subtypes is crucial for elucidating the structural and functional intricacies of complex tissues. Various algorithms have been developed for dimensionality reduction, noise filtering, and assignment of annotations [129,130]. These methods can be categorized into unsupervised [[142], [143], [144], [145]], semi-supervised [45,146,147], and supervised approaches [[148], [149], [150]]. Unsupervised methods, such as Pixie [142], PhenoGraph [144], and X-shift [145], subset single cells based on similar marker expressions for cell type annotation, independent of prior knowledge, and are frequently utilized for clustering visualization. Although clusters generated by unsupervised algorithms exhibit similar features, they are label-free and may lack interpretability. This necessitates manual curation and parameter optimization, processes that can be subjective, labor-intensive, and require considerable expertise. To enhance the accuracy of cell annotation, promising automated approaches include probabilistic model-based and CNN-based semi-supervised or supervised algorithms, which leverage well-annotated datasets containing ground-truth labels for cell types. ASTIR [148] utilizes a probabilistic model that incorporates prior knowledge of marker proteins to categorize cells without accounting for spatial information provided by the training data. In contrast, CELESTA [146] employs a cell-type signature matrix to assign cell-type probabilities to each cell based on a combined marker- and spatial-scoring function in an iterative fashion. CellSighter [147], based on CNN, demonstrates promising classification performance but has been noted for its limitations in computational efficiency. In contrast, MAPS [150] enables accurate and rapid cell annotation with superior performance across multiple spatial proteomics platforms. The choice of method ultimately depends on specific user needs, such as the biological questions under investigation, desired annotation accuracy, dataset size, and available computational resources.

Mass spectrometry (MS) is a data-rich analytical tool capable of quantifying a wide range of analytes by measuring their mass-to-charge ratios (m/z). Given its strong data-mining capabilities, deep learning has now been integrated into nearly every stage of MS analysis, including spectral pre-processing, peptide identification, and protein quantification. In liquid chromatography-tandem mass spectrometry (LC-MS/MS) systems, accurate prediction of peptide retention times is essential for reliable protein analysis, with deep learning methods like Prosit [151] and DeepDIA [152] outperforming traditional approaches. Prosit, for example, utilizes a recurrent neural network (RNN) model and has demonstrated much higher prediction accuracy than traditional machine learning methods. Predicting peptide fragmentation is another complex process, influenced by factors such as MS instrument type, peptide sequence, and peptide charge state. Incorporating deep learning model has improved the predictability of peptide fragmentation, with tools like pDeep [153] and PredFull [154] showing excellent accuracy. Additionally, deep learning-based tools for de novo peptide sequencing, including DeepNovo [155], DIA-NN [156], MSBooster [157], and DeepPep [158], have achieved significant improvements. DIA-NN, for instance, uses multiple interference correction strategies and a dense feedforward neural network, greatly enhancing peptide identification and quantification [156]. Label-free quantification is a powerful tool in proteomics analysis, though traditional algorithms have struggled with limited proteome coverage and low sensitivity. Recently, AI-driven methods like PIQED [159], Quandenser [160], and IonQuant [161] have shown superior performance over traditional techniques by reducing search time, minimizing missing values, and increasing protein identifications.

6. Future perspective

Antibodies are pivotal for in situ detection of protein distribution, yet they come with inherent limitations. Effective staining demands rigorous validation, specific sample preparation, and optimized experimental conditions. In multi-round staining, changes in antigenicity and antigen loss can significantly compromise data quality. Furthermore, the utility of many antibodies is constrained by factors such as sequence accessibility, sensitivity, and cross-reactivity, thereby limiting multiplexed detection. The emergence of recombinant antibodies, nanobodies, and aptamers promises substantial improvements. Recent developments in narrow-spectrum fluorophores [162], efficient detection methods [29,163], and signal amplification techniques enhance sensitivity and throughput in protein detection. However, multi-round staining, imaging, and processing remain labor-intensive, time-consuming, and prone to introducing artifacts. Standardization and automation of these processes could greatly enhance reliability and reproducibility. Moreover, 3D volumetric reconstruction, combined with tissue preservation and clearing techniques, promises a more comprehensive characterization of tissue architectures and functions. The increase in detection throughput and resolution generates vast amounts of high-dimensional data, posing challenges in storage, analysis, and sharing. This also opens avenues for developing algorithms and software to standardize data processing, visualization, and compression. Advances in sequencing technology enable DNA-labeled antibodies to simultaneously capture protein and RNA information [18,52]. Scaling up protein capture and improving resolution to the cellular or even subcellular level, combined with multi-omics analysis, holds promise for deeper insights into cell types and interactions within complex tissues.

Liquid chromatography-tandem mass spectrometry (LC-MS/MS) is a powerful tool in proteomics, capable of unbiased detection of over 10,000 proteins from very small sample amounts (< 200 picograms). This extensive proteomic coverage underscores the significance of LC-MS/MS-based spatial proteomics. However, challenges remain in balancing resolution, throughput, and cost-effectiveness, presenting opportunities for technological advancements. First, more efficient proteomic preparation workflows are needed [164,165]. Automated and streamlined processes can increase throughput, reduce protein loss by minimizing manual intervention, enhance consistency and reproducibility, and reduce contamination risks. Second, the integration of advanced technologies such as artificial intelligence, single-cell omics, and microfluidics can achieve higher sensitivity, faster acquisition rates, and improved resolution. DVP technology is an example of the integration of AI with single-cell proteomics [20]. Third, advances in mass spectrometry are critical. Innovations such as trapped ion mobility time-of-flight instruments (timsTOF) and data-independent acquisition parallel accumulation-serial fragmentation (diaPASEF) have shown great potential in spatial omics. Finally, the study of PTMs in a spatial context is an exciting frontier. Techniques for spatial analysis of PTMs (e.g., phosphorylation, glycosylation, acetylation) offer deeper insights into proteome functionality, leveraging the robust detection capabilities of mass spectrometry.

Imaging MS-based spatial proteomics has advanced our understanding of disease by providing spatial context for biomolecules within biological systems [166]. This capability enables a spatially resolved understanding of biomolecules in pathological contexts. Despite these advances, challenges persist in protein identification and confirmation, with current techniques predominantly identifying high-abundance proteins. Addressing these challenges, particularly for low-abundance proteins crucial to biological functions, remains a priority. Integration of ion mobility as an additional separation step shows promise in enhancing protein identification complexity. Further improvements in mass spectrometry instrumentation are highly desirable. The integration of MALDI-IMS with spatial transcriptomics and routine histology holds potential to enhance our understanding of gene-to-protein expression patterns in health and disease [167]. In addition, there is a need for expanded open-source bioinformatics tools for mass imaging, akin to genomic field [168], to facilitate protein identification and statistical analysis.

While AI in spatial proteomics has achieved significant advancements, challenges remain. Accurate cell segmentation or localization still depends on high-quality marker staining, and variability in fluorescence intensity poses challenges for classification algorithms. Additionally, integrating multi-omics data to create comprehensive views of cellular and tissue states remains in its early stages and requires further development. The diversity of AI algorithms and the lack of standardized workflows can also hinder accessibility for non-computational scientists. Future research focusing on enhancing the robustness and scalability of AI algorithms, developing standardized protocols for data integration, and improving the interpretability of AI-driven insights will further advance our understanding of complex biological systems, enhancing the precision of medical diagnostics and therapies.

The development of single-cell proteomic atlases is critical for biological research [15]. Studying proteins directly at the single-cell level, rather than using transcripts as proxies, is very appealing because proteins directly perform biological functions. While numerous single-cell expression atlases exist, such as The Cancer Genome Atlas, Mouse Cell Atlas, and Fly Cell Atlas, corresponding single-cell proteomic atlases are limited [[169], [170], [171]]. Future development of robust and scalable single-cell proteomic techniques promises significant expansion of single-cell proteomic atlases. Spatial multi-omics analysis, integrating spatial proteomics with other spatial omics technologies, represents a promising avenue for comprehensive biological insights.

Declaration of competing interest

The authors declare that they have no conflicts of interest in this work.

Biography

Zhao Fangqing, a professor at the Institute of Zoology, Chinese Academy of Sciences, has been recognized with the National Science Fund for Distinguished Young Scholars and as a distinguished professor at the Chinese Academy of Sciences. His research focuses on developing efficient algorithm models and experimental techniques to investigate the dynamics and functions of human microbiota and non-coding RNAs, with the goal of understanding their impact on health and disease. Over the past few years, he has authored over 100 papers as corresponding author in prestigious journals including Cell, Gut, Nature Biotechnology, Nature Methods, Nature Genetics, Nature Cell Biology, Nature Computational Science, and Nature Communications.