Microbial Pathogenesis in the Era of Spatial Omics

ABSTRACT

The biology of a cell, whether it is a unicellular organism or part of a multicellular network, is influenced by cell type, temporal changes in cell state, and the cell’s environment. Spatial cues play a critical role in the regulation of microbial pathogenesis strategies. Information about where the pathogen is—in a tissue or in proximity to a host cell—regulates gene expression and the compartmentalization of gene products in the microbe and the host. Our understanding of host and pathogen identity has bloomed with the accessibility of transcriptomics and proteomics techniques. A missing piece of the puzzle has been our ability to evaluate global transcript and protein expression in the context of the subcellular niche, primary cell, or native tissue environment during infection. This barrier is now lower with the advent of new spatial omics techniques to understand how location regulates cellular functions. This review will discuss how recent advances in spatial proteomics and transcriptomics approaches can address outstanding questions in microbial pathogenesis.

INTRODUCTION

Location is a critical regulatory node of cell biology during microbial pathogenesis at every “order of magnitude” from molecular interactions to tissue microenvironments. For instance, Gram-negative bacterial lipopolysaccharide induces a distinct Toll-like receptor 4 (TLR4) signaling complex at the cell surface compared to the endosome (1, 2). Apicomplexan parasites use unique secretory organelles to deliver virulence effectors, forming novel compartments for intracellular growth (3). Shigella and Listeria co-opt host actin, polymerizing “tails” that mediate host cell escape (4). In biofilms, bacterial gene expression is a function of their position within a colony (5). Moreover, many pathogens exploit specialized tissue environments with unique cell architectures that cannot be replicated in vitro, such as gastric infection by Helicobacter pylori or fibrotic remodeling of the lung during helminth infection (5, 6).

In the last decade, spatial proteomics tools have become accessible, followed closely by an expansion of spatial transcriptomics methods. These tools were primarily developed to study eukaryotic cell biology, and repurposing them to study microbial pathogenesis often requires unique adaptations, described here. Comparatively mature methods, including cell fractionation and proximity-labeling techniques are being creatively applied to address molecular, organellar, and cell type-specific changes during infection. However, most pathogenic microbes are not genetically tractable and infect non-model organisms or tissues that are difficult to access or reconstitute in vitro (Fig. 1). Exploring the biology of microbial pathogens with limited tools for experimental manipulation has required new tools to study microbes in their native environment, a need that has been met with the emergence of imaging-based transcriptomics and proteomics tools.

FIG 1.

Questions in microbial pathogenesis to approach with spatial omics tools. (A) Pathogens that lack model organisms or tools for experimental manipulation (e.g., a virus recently transferred to a zoonotic host); (B) studying the biology of infected primary cell types or tissues with complex architecture that are poorly modeled in vitro (e.g., infected neurons or bacteria encroaching on intestinal mucus [green]); (C) understanding the constituents of novel structures that form during infection like a bacterial pedestal or granuloma (green core); (D) evaluating subpopulations of microorganisms (e.g., bacteria deep within versus peripheral to a biofilm [green], a constituent of the gut commensal population, and tissue-resident versus circulating parasites). Red indicates a population of interest, and purple indicates nontargeted. The figure was generated by BioRender.

CELL FRACTIONATION APPROACHES TO STUDY PATHOGEN ORGANELLES AND MICROBIAL MANIPULATION OF HOST CELL COMPARTMENTS

Cell fractionation by differential centrifugation was introduced in the 1930s by Bensley and Hoerr, who first separated mitochondria and microsomes from insoluble cell debris (7). In the 1940s, pairing fractionation with electron microscopy allowed biochemists to associate the biochemical functions of organelles with the structure and location in the cell (8). These approaches continue to be important to define the unique organelle biology of parasitic protozoa (9), study bacterial outer membrane, inner membrane, cytoplasm, and periplasmic compartments (10, 11), and enrich stress granules as well as viral inclusion bodies from infected host cells (12). This approach coupled with liquid chromatography-mass spectrometry (LC-MS) has been applied to understand how macrophage and dendritic cell phagosomes are modified by Mycobacterium infection and how Legionella manipulates Dictyostelium phagocytosis (9, 13–15). However, the main drawbacks of differential centrifugation are that endolysosomal system components can elute across many density fractions, and it is impossible to separate small transport vesicles of similar density, which necessitates secondary validation of host protein recruitment.

Magnetic purification of microbe compartments.

Magnetic purification has been a notable solution to vesicle contamination issues inherent to density centrifugation. David Russell’s lab identified Mycobacterium effectors that arrest phagosome maturation by loading the phagolysosome pathway with iron dextran and magnetically isolating the vacuole after gentle permeabilization of the plasma membrane (16). Streptococcus pyrogens have been magnetically labeled prior to uptake by human neutrophils and used to isolate bacterium-containing phagosomes (17). Robust protocols have been developed to isolate magnetotactic bacteria based on their membrane-encased ferromagnetic nanoparticles, known as magnetosomes, and Plasmodium based on hemozoin in the feeding organelle (18–21). Hemozoin is sufficiently magnetic to separate infected and uninfected red blood cells. This approach was combined with phospho-proteomics analysis to identify 18 FIKK serine/threonine kinases used by Plasmodium falciparum but not by Plasmodium knowlesi to phosphorylate erythrocyte target proteins and secreted parasite virulence factors (22). Despite the robustness of magnetic isolation approaches, there has not been widespread use of these tools to isolate pathogen-containing vacuoles for omics analysis, an outstanding question for many obligate intracellular protozoan and bacterial pathogens.

LOPIT-mediated cell fractionation.

To circumvent fraction contamination issues, the Lilley Lab developed an experimental pipeline and informatics tool to infer high-likelihood organelle resident proteins. In LOPIT (localization of organelle proteins by isotope tagging) subcellular fractions are isotopically labeled and pooled into a single LC-MS sample, limiting run-to-run variability. Organelle proteomes are mapped by comparing the abundance of known organelle-resident proteins in each fraction to the abundance of every protein across each fraction in the sample using Bayesian analysis (23, 24). LOPIT has been used to quantify human cytomegalovirus (CMV) proteins localized to fibroblast organelles over a 120-h infection (25). However, the number of fractions that can be reliably measured by LOPIT is limited by the number of stable isotopes. As an alternative, “label-free” quantification approaches, where run-to-run MS variability is normalized informatically, have been developed (26–28).

In 2017, HyperLOPIT was released, where 10 fractions can be pooled and proteins are identified with higher-sensitivity MS and mapped to organelles using a multivariable machine learning tool (29). This technique is particularly well suited to understand the cellular architecture of apicomplexan parasites, which have a mixture of organelles used by euthermic eukaryotes (e.g., mitochondria, nucleus, endoplasmic reticulum [ER]), red algae (plastids), and evolutionarily distinct secretory organelles that deliver effectors into the host cytosol. For example, HyperLOPIT was used to map a high-coverage Toxoplasma gondii proteome to 25 functionally distinct parasite organelles or organelle regions (30). HyperLOPIT protein localization data sets are also proving exceptionally useful to predict whether proteins identified by coprecipitation or proximity labeling belong to a specific subcellular region, as described in the following section (29, 31).

ENZYMATIC PROXIMITY LABELING TOOLS FOR LOCAL PROTEIN AND TRANSCRIPT ENRICHMENT

Proximity labeling techniques use a label-targeting (bait) protein conjugated to an enzyme to decorate near-neighbor molecules with affinity purification tags. Decorated proteins or nucleic acids are then purified and identified by LC-MS or sequencing, respectively. These tools were initially developed to assess protein localization by immunofluorescence microscopy, enhance contrast in electron microscopy, and mediate click chemistry in live cells (32, 33), but are now proving robust techniques for spatial omics analysis. Each enzyme has unique considerations that necessitate evaluation during the experimental design phase (Table 1).

TABLE 1.

Current methods for spatial proteomics

Tool (reference)	Isolation method	Labeling radius	Labeling time	Sample type	Applications
BioID (35–38, 41, 42)	Bait protein-conjugated BirA biotinylation	40 nm	18 h	Live cells	Protein-protein interactions, organelle proteomics
BioID2 (48, 49, 54, 55)
TurboID (48, 53, 54)		10 nm	10 min
MiniTurbo (54, 55)
Considerations
Labeling kinetics and cytotoxicity: biotin-phenol labeling is less cytotoxic than H2O2 labeling used for APEX; BioID2 requires 10× less biotin-phenol than BioID, reducing cytotoxicity but is most efficient at 37°C; TurboID is 2× faster than miniTurbo, but associated with high background and endogenous biotin depletion toxicity
Steric hinderance and stability: steric hinderance is an issue for BioID (36 kDa) and minimal for BioID2 (26 kDa), TurboID (35 kDa), and Miniturbo (28 kDa); miniTurbo is unstable
Protein targeting: protein targeting requires genetic modification of a “bait” protein; best suited for cell lines
APEX (56, 61)	Bait protein-conjugated ascorbate peroxidase monoavidin	120 nm	0.5–1 h	Live/fixed cells	Protein-protein interactions, organelle proteomics, in vivo proteomics
APEX2 (48, 51–53)		20 nm		Live/fixed cells, in vivo labeling
Considerations
Labeling kinetics and cytotoxicity: H2O2 labeling is more cytotoxic than biotin-phenol labeling (bioID); APEX2 has high activity at lower H2O2; APEX2 can be used for in vivo proteomics and transcriptomics
Protein targeting: protein targeting requires genetic modification of “bait” protein; best suited for cell lines
TSA-BAR (70, 71, 73)	Antibody-conjugated tyramide biotinylation	500 nm	10 min	Live/fixed cells, tissue	Protein-protein interactions, organelle proteomics
Considerations
Pros: antibody-based tagging is ideal for studying primary cells, tissues, and organisms with limited genetic tools
Cons: fixation and permeabilization required for intracellular structures
HRP (62)	Bait protein-conjugated HRP	300 nm	60 min	In situ	Cell surface, in vivo proteomics
Considerations
Pros: optimized for surface antigens, live-cell labeling, and in vivo labeling
Cons: HRP labeling is incompatible with acidic or highly reducing organelle environments
HyperLOPIT (29, 30)	Subcellular fractionation	NAa	NA	Lysed cells	Organelle proteomics
Considerations
Pros: fractionation links protein sequence information with biochemical properties of organelle
Cons: limited to 10 isotopically labeled fractions
LCM (81, 83, 84, 89)	Image-targeted laser ablation/polymerization	10 μm	3 h	Live/fixed tissue	In situ cell proteomics
Considerations
Pros: primary tissue compatible; user validates target region by imaging; compatible with fixed samples
Cons: low protein yield/coverage; cannot resolve single-cell borders; labor- and time-intensive
DVP (94)	AI image-targeted laser ablation	10 μm	≥3 h	Fixed tissue	In situ cell proteomics
Considerations
Pros: utilizes AI to reduce human imaging bias; automation increases throughput
Cons: labor-intensive; requires custom equipment
autoSTOMP (95–97)	Imaging-guided UV biotinylation	~1 μm	72 h	Fixed cells and tissue	Organelle, cell, in situ proteomics
Considerations
Pros: primary cell/tissue compatible; can resolve organelles; user validates target region by imaging
Cons: requires fixation; low protein yield/coverage; time-intensive

^aNA, not applicable.

Promiscuous biotin ligases: BioID, TurboID, and miniTurbo.

Biotin identification (BioID) evolved from a technique to evaluate protein-genome interactions using DNA adenine methyltransferase called DAM-ID (34). In BioID the “bait” protein of interest is fused to BirA, a promiscuous biotin ligase derived from Escherichia coli. BirA biotinylates proteins within a 10-nm radius upon the introduction of exogenous biotin (35). Fusing BirA to the host SNARE syntaxin-6 showed that the signal sequence YGRL, which regulates plasma membrane protein recycling to the trans-Golgi network, is manipulated by Chlamydia for vesicle delivery to the inclusion body (36) (Fig. 2A). The nuclear localization of the Chlamydia type II secretion system effector SINC (secreted inclusion protein-C) and its proximity to the nuclear lamin protein emirin were also determined by BioID and confirmed by coprecipitation (37) (Fig. 2B, panel i). In this study, 22 biotinylated proteins were conserved across BioID replicate experiments; however, over 260 proteins were unique to each replicate, consistent with high background labeling in this approach.

FIG 2.

Predicting molecular interactions and subcellular localization by enzyme-mediated proximity ligation with the APEX, BioID, and TSA-BAR approaches. (A) The labeling enzyme (teal Pac-Man) is conjugated to a bait protein (gray) or signal sequence that is abundant and robustly localized to a target organelle to label resident proteins or nucleic acids (light aqua region), here depicting a microbial (red) vacuole. (B) A bait protein is enzyme conjugated to label any nearby molecules (i), such as a target protein with an engineered acceptor peptide (ii), using split GFP (green) where the larger N terminus of GFP is conjugated to the enzyme under inducible control and the smaller C terminus is conjugated to the bait protein (iii). (C) In protein mapping approaches, cell lines are generated where each expresses a label-targeting protein (described in panel A) and comparative proteomics is used to map host and/or microbial protein abundance in each compartment. (D) In TSA-BAR and nanobody-APEX, a label-targeting enzyme is targeted via an antibody (blue) raised against a specific bait antigen. The figure was generated by BioRender.

BioID was used to understand the functional significance of evolutionarily divergent protozoan cytoskeletal structures. The hook complex (or bilobe) is a cytoskeletal structure at the distal flagellar pocket in kinetoplastids. By conjugating BirA to the Trypanosoma brucei protein TbMORN, 7 components of the hook complex were identified, including two flagellum attachment points (38). In apicomplexan parasites, the inner membrane complex (IMC) is a series of flattened membrane sacks between the inner leaflet of the parasite plasma membrane and a network of intermediate filaments, which is essential for motility, invasion, and division (39). Fusing BirA to the T. gondii IMC protein ISP3 identified 17 proteins partitioned to distinct regions of the IMC, including two apical cap proteins, AC1 and AC2 (40). BirA-AC2 was generated to identify a family of apical cap proteins that bridge the IMC with the plasma membrane and intersect with mitogen-activated-protein-kinase signaling, nicely demonstrating how proximity labeling can be used in iterative cycles to refine protein interaction maps (41–43).

BioID has helped elucidate transient interactions controlling viral entry and budding. Using the human immunodeficiency virus (HIV) protein Vpu as bait led to the discovery that Vpu promotes viral shedding by interacting with AP1 and the lipid raft protein tetherin (44). To assess whether the cooperation of herpes simplex virus glycoproteins was necessary for entry, researchers coexpressed gD-BirA with gH linked to a biotin acceptor peptide (45) (Fig. 2B, panel ii). The gH acceptor peptide was effectively biotinylated; however, a negative-control glycoprotein from Epstein-Bar virus (EBV) was too. This result highlights limitations of BioID related to the overexpression of the bait protein and 18-h labeling (Table 1), which can report artifact interaction at sites of protein maturation (ER) and turnover (lysosome). A third caveat of BioID is labeling steric hindrance imparted by the 321 amino acid BirA enzyme. For example, when BirA was fused to discrete sites on HIV group-specific antigen (Gag) or the Gag precursor, AEG-1 (astrocyte-elevated-gene-1) was the only protein conserved between experiments (46, 47).

Steric hindrance was addressed with BioID2, a 233-amino-acid BirA homolog from the thermophile Aquifex aeolicus. BioID2 requires a log less biotin but is limited to temperatures above 37°C and susceptible to background labeling from endogenous biotin (48). A split-green fluorescent protein (GFP)-BioID2 system (Fig. 2B, panel iii), in which an inducible BirA conjugated to the GFP N terminus is coexpressed with a C-terminal GFP-bait protein (49), has been used to discover innate immune signaling components activated by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) protease nonstructural protein-5 (NSP5) in compared to an inactive NSP5 allele (50). In this report, over 2,000 labeled host proteins were prioritized by cross-referencing genome-wide-association studies’ data sets and applying Significance Analysis of INTeractome (SAINT) to calculate the probability that a candidate protein interacts with the bait based on confidence intervals generated from the abundance of known interaction partners in the data set (51). This approach may be useful to determine the localization and interaction partners of secreted microbial effectors.

Using a directed-evolution approach, the Ting lab has addressed the lengthy labeling kinetics of BirA with two additional tools. TurboID is a 35-kDa BirA protein with 15 amino acid modifications, and miniTurbo is an N-terminal truncation of TurboID (28 kDa) with 13 amino acid changes (52). TurboID and miniTurbo have 10-min minimal labeling times with 3- and 6-fold-higher labeling efficiency than BioID after 6 h. However, turboID is so efficient that it can cause cytotoxic sequestration of endogenous biotin, high background labeling, and bait protein instability (53) (Table 1). Some bait proteins are not equally amenable to tagging by both enzymes, as reported in a study that used miniTurbo or turboID to label all but three SARS-CoV-2 proteins, enabling network analysis of signaling cascades cooperatively regulated by multiple viral proteins (54). miniTurbo facilitated the first high-coverage proteome of the apicomplexan vacuole membrane, a novel compartment formed independently of the phago/endosomal system and which cannot be isolated by fractionation. To achieve this, miniTurbo was fused to the arginine-rich amphipathic helix (RAH) domain of the secreted T. gondii effector rop17. Remarkably, when ectopically expressed in human cells, BirA-RAH was selectively localized to the parasite vacuole membrane following infection, providing a high-coverage proteome of host and parasite proteins at the vacuole (3).

Ascorbate peroxidases: APEX and APEX2.

The second major genetically encoded proximity labeling technique uses ascorbate peroxidase (APEX). In the presence of hydrogen peroxide (H₂O₂), APEX oxidizes supplemented biotin-phenols to phenoxyl radicals, which label nearby nucleic acids or proteins until quenched. Biotin-phenol and H₂O₂ are both cytotoxic, but the APEX reaction time is brief (Table 1). H₂O₂ also inhibits APEX function; however, this was circumvented in APEX2 by two activity-enhancing mutations that also reduced the labeling radius to 20 nm (55, 56). APEX2 has revealed that SUMO regulates stress granule disassembly when conjugated to three bait proteins: G3BP1, HMR1, and FXR1 (57, 58). Repurposing this tool to study viral elementary bodies may be possible, given the structural similarities between these membrane-free organelles. For example, to identify proteins recruited to the poliovirus replication site the C terminus of enterovirus Golgi apparatus-specific-BFA-resistance-factor-1 tagged with APEX2 recruited to poliovirus inclusion bodies and labeled glycolytic enzymes that regulate de novo RNA synthesis, as well as RNA binding inhibitors of viral replication (59).

APEX2 labeling can be limited by protein tertiary and quaternary structures and transmembrane topology. However, membrane steric inhibition can be exploited to control spatial resolution (Fig. 2C) to identify structures like the inner mitochondrial membrane proteome. An analogous approach with horseradish peroxidase (HRP) chemistry identified the neuronal cell surface proteome (60, 61). Linking APEX2 to minimal localization motifs is proving a valuable tool to understand trafficking of secreted microbial effectors (Fig. 2A). APEX2 expressed in the host nucleus labeled a T. gondii modulator of the NCoR/SMRT repressor complex called TgNSM, which cooperates with a second effector, TgIST, to limit protein kinase R activation of cell death (62). The Seeliger lab adapted APEX2 labeling to visualize and isolate proteins enriched in the Mycobacterium tuberculosis cytoplasm or periplasmic space using click chemistry with tyramide-alkyne and tyramide-azide probes rather than biotin-phenol (63) (Fig. 2B, panel i).

Using a conceptually similar approach, RNA-GPS has been developed to isolate and sequence transcripts labeled by APEX2 targeted to discrete areas of the cell (64, 65) (Fig. 2C). This approach was used to evaluate regional amplification of the SARS-CoV-2 genome to eight regions, demonstrating an unexpected localization to the host mitochondria in addition to the nucleolus (66). The Ingolia lab paired APEX2-eIF4A1 RNA sequencing with LC-MS to show that this dead box helicase interacts with the ribosomal 43S preinitiation complex and stress granules when mTORC is inhibited (65).

In another advance of APEX, E. coli has been engineered to express APEX-nanobody fusion proteins with antigenic specificity for Ebola and Marburg viruses (Fig. 2D). A proof-of-concept study demonstrated that APEX-nanobodies can amplify the immunofluorescence signal from infected cells (67). This approach could be applied to a range of primary cells and non-model pathogens with limited tools for genetic manipulation, similar to TSA-BAR (tyramide signal amplification-biotinylation by antibody recognition) described below.

TSA-BAR.

In tyramide signal amplification-biotinylation by antibody recognition (TSA-BAR), a primary antibody specific to the target (molecule, organelle, or cell) is detected by a secondary antibody conjugated to HRP (Fig. 2D). In the presence of H₂O₂, tyramide-biotin covalently attaches to nucleic acids and proteins within a 0.5-μm radius of the target antigen (68, 69). Tyramide chemistry is reported to label nucleic acids with less steric hindrance than proteins. TSA sequencing approaches have identified chromosome and transcript sequences and calculated the distance of these targets from nuclear speckles (using SON as bait) or the inner nuclear envelope (using laminins as bait) (70–73). While this technique has not been applied to viral infection, there is potential to address long-standing questions regarding the location of viral integration and regulators of viral latency and reactivation.

Pros and cons of enzyme-mediated proximity labeling approaches.

There is tremendous potential for this suite of host labeling-targeting proteins to advance our understanding of microbial pathogenesis in a range of infection settings. APEX2, TurboID, and miniTurbo have a similar labeling range (10 to 20 nm) and the highest spatial resolution of all available tools (Table 1). A central limitation of BioID- and APEX-derived techniques is that they require molecular engineering of a label targeting protein and cell transduction. This is a significant barrier for many primary cell types, non-model host organisms, and pathogens that lack robust tools for genetic manipulation (Fig. 1A and B). The labeling times required for APEX, BioID1 and -2, and HRP approaches are too long to capture the dynamics of processes like microbial entry, vacuole escape, or cell-to-cell spread. APEX2 has addressed this issue by reducing the labeling time, followed by TurboID and miniTurbo, which can label in as few as 10 min. Protocols to conditionally express these tools in Drosophila and mice are opening the door to evaluating host-pathogen interactions in primary cells and native tissue environments (74–76). However, these experiments are extremely cost- and labor-intensive. TSA-BAR ameliorates this limitation at a cost of spatial resolution.

A caveat of these approaches is that bait proteins require robust targeting to the organelle of interest. Proteins with alternative pools in the cell or that relocalize in response to signals like cell stress, inflammatory cytokines, and secretion system expression can confound the interpretation of results because labeling will occur at all sites. Bulky enzyme adducts may be particularly problematic in host-pathogen interaction: for example, to identify effectors injected via secretion systems. Additional controls may be required to evaluate infection biology, such as comparison with attenuated microbes or host cells deficient in immune sensing or cell death pathways. Finally, these approaches pool thousands of cells so additional approaches to isolate infected from infected cells may be necessary to reduce the noise of irrelevant interactions.

MICROSCOPY-MEDIATED SPATIAL OMICS TOOLS TO STUDY ORGANELLES AND CELL BIOLOGY IN SITU

Single-cell RNA sequencing and proteomics techniques are rapidly changing our understanding of host and microbial cell type specificity and functional heterogeneity (77–79). However, these tools fail to capture important spatial information about cells within their native tissue or colony microenvironment. Moreover, harsh enzymatic and physical disruption methods used for single-cell omics can lose sensitive cell types and alter mechanical signaling and gene expression. To address these issues, new technology at the interface of microscopic tissue evaluation and transcription or protein expression is under development.

In situ proteomics and transcriptomics with laser capture microdissection and deep visual proteomics.

Laser capture microdissection (LCM) uses a fluorescence microscope to image cell types or structures of interest and isolate them from a larger sample (80). Ablative LCM employs a UV laser to remove unwanted cells, while enrichment LCM uses the infrared (IR) light to selectively activate a thermolabile polymer in target cells, after which the cell composite is lifted off the slide for genomic, transcriptomic, and proteomic analyses (81–84) (Fig. 3A). Depending on target abundance, LCM collection time can exceed 2 h per sample, limiting the number of cells that can be acquired from fresh tissue slices (85, 86). Fixation extends the sample collection window but reduces the number of identifiable peptides. For these reasons, read depth and proteome/transcriptome coverage remain major challenges in LCM.

FIG 3.

Microscopy-mediated omics tools. (A) In LCM, structures of interest are imaged and stabilized or non-target structures are ablated in a photopolymer matrix. Cells are processed for RNA-seq or MS. (B) In DVP, an artificial intelligence platform identifies structures of interest and ablative LCM is used to isolate them for MS analysis. (C) In autoSTOMP, photo-biotin tags are conjugated to proteins in target structures using confocal microscopy and then purified for LC-MS. (D) In FISSEQ, cDNA synthesis and rolling circle amplification are used to identify transcripts in situ. (E and F) Slides containing a barcoded capture oligonucleotide array (10×/Visium [E]) or a bead array (SlideSeq [F]) are overlaid with tissue. Upon permeabilization, RNAs are captured for library prep in vitro and next-generation sequencing (NGS). (G) In HDST, a hexagonal well array containing capture beads is overlaid with tissue. RNAs are captured for in vitro library prep and NGS; however, the resolution is enhanced by binning reads from neighboring hexagonal wells. (H) In Light-seq, barcoded oligonucleotides are hybridized to RNAs for a cross-junctional synthesis reaction that facilitates cDNA amplification, library preparation and NGS. The figure was generated by BioRender.

Despite these caveats, LCM is an important approach to study host-microbe interactions, particularly for pathogens that infect non-model organisms or have tropisms that cannot be accurately recapitulated in vitro. For example, LCM-seq was used to determine that pig enterocytes infected by the swine pathogen Lawsonia intracellularis upregulate Rho GTPases compared to uninfected enterocytes in the same intestine (87). The Koshy lab employed LCM-seq to evaluate neurons that interacted directly with T. gondii by using a strain of parasite engineered to inject Cre-recombinase into zsGreen reporter mice (88). However, the dominant transcriptional signature identified belonged to CD8 T cells. This is in keeping with other reports showing LCM has insufficient resolution to capture complex cellular architecture and is best suited to study multicellular tissue structures (89–92).

Deep visual proteomics (DVP) combines LCM and machine learning to identify and isolate phenotypically distinct cells (Fig. 3B). This technique uses a custom scanning microscope with an LCM interface to image and excise cells of interest at a rate of 50 to 100 per hour. High-resolution MS and label-free quantitation are used to evaluate protein abundance (93). DVP has been used to analyze tumor cell heterogeneity in melanoma and salivary gland carcinoma, a methodology that could easily be extended to examine infected cell heterogeneity. Artificial intelligence (AI) image analysis has already expanded the capacity to discover and score infection-induced cell phenotypes beyond what a typical user can quantify (94). DVP could dramatically enhance spatial protein discovery, but the need for specialized microscopy hardware may hinder its rapid adoption for microbial pathogenesis research.

Organelle and cell enrichment in situ by autoSTOMP.

autoSTOMP (automated spatially targeted optical microproteomics) standard immunofluorescence staining and a confocal microscope have been used to visualize structures of interest and then selectively biotinylate proteins in those structures with a photochemical biotin tag for purification and identification by LC-MS (95) (Fig. 3C). The imaging and cross-linking can take 72 h per sample, but the process is automated. autoSTOMP has identified host and parasite proteins enriched on the T. gondii vacuole membrane and protein expression in tissue-resident immune cells (96, 97) (Table 1). Similar to LCM, image-based targeting means that autoSTOMP can be applied to primary cells and tissues. However, the throughput and resolution (~1 μm) are higher, so it is straightforward to modify the target region, based on colocalization stains, to capture subpopulations of a structure or control areas of the cell/tissue for comparative analysis. Although early in development, STOMP has the potential to evaluate difficult-to-isolate structures arising from microbial interactions, like viral inclusion bodies or actin tails used by Shigella and Listeria to escape host cells.

Sources of bias in spatial proteomics.

Spatial proteomics can be challenging to analyze with standard informatics tools. Differential enrichment pipelines generally assume full proteome coverage, but subcellular proteomics studies only capture a subset of the proteome by design. Moreover, low proteome coverage in LCM, DVP, and autoSTOMP can make it difficult to distinguish between proteins that are not expressed (true negatives) versus low-abundance proteins falling below the detection limit. For these reasons, imputing “missing values” with a standard value close to 0 is more likely to lead to false positives or invert fold change in spatial proteomic experiments compared to full-coverage proteomic experiments, and controls embedded in the experiment should be carefully evaluated early in the analysis.

Low coverage may be particularly problematic for the detection of pathogen proteins, which are often a minor fraction of a total host-pathogen proteome, and to resolve paralogous genes like the tandem repeat surface proteins used by Trypansosoma and Plasmodium species to evade antibody detection. Additionally, low-coverage compounds can have alignment errors due to posttranslational modifications, irreversible biochemical alterations (oxidative stress), and fixatives or detergents. It is likely that emerging enrichment strategies for single-cell proteomics (isotopically labeled carrier proteins) will also benefit the resolution of spatial proteomics.

Single-cell transcriptomics by FISSEQ.

Fluorescent in situ sequencing (FISSEQ) combines the spatial resolution of RNA-fluorescent in situ hybridization (FISH) with single-cell transcriptomics (98) (Fig. 3D). Samples are fixed and cDNAs are reverse transcribed in situ with rolling circle amplification. FISH probes are hybridized to the cDNA “nanoballs” to initiate SOLiD (sequencing by oligonucleotide ligation and detection) sequencing by ligation up to 27 times. Each sequencing reaction is aligned with the pixel coordinates of the FISH probe to map the transcriptional signature. When fibroblasts were targeted in a Drosophila embryo wound-healing model, 4,171 genes mapped to a 5-pixel area (pixel size in nanometers not reported) with 90% strand alignment. However, over 40% of reads aligned to rRNA. Most mRNA transcripts detected are highly expressed extracellular matrix proteins from fibroblasts, suggesting less abundant transcripts are likely missed due to limited read depth or signal crowding (99) (Table 2).

TABLE 2.

Current methods for spatial transcriptomics

Method (reference)	RNA capture method	Labeling radius	Isolation time	Sample type	Application(s)
APEX2-seq (68)	In vitro RNA isolation and library prep	20 nm	1 h	Cell culture, in vivo labeling	Organelles, cells, in vivo transcriptomics
Considerations
Pros: excellent spatial resolution, labeling time, and read depth
Cons: requires genetic modification of “bait” protein; subcellular pools of bait protein cannot be resolved
TSA-seq (78)	In vitro RNA isolation and library prep	0.5 μm	24 h	Live/fixed cells and tissue	Organelles, cells, in situ transcriptomics
Considerations
Pros: good spatial resolution and excellent read depth; applicable to primary cells/tissues
Cons: subcellular pools of bait cannot be resolved; steric hindrance from antibody-enzyme
LCM (81, 83, 84, 89)	Image-targeted laser ablation/polymerization	10 μm	3 h	Live/fixed tissue	In situ cell proteomics
Considerations
Pros: user validates target region by imaging
Cons: low RNA yield/coverage; cannot resolve single cells borders; labor- and time-intensive
FISSEQ (99)	In situ rolling circle amplification and SOLiD	0.4 μm	48 h	Fresh/frozen/fixed	In situ cell transcriptomics
Considerations
Pros: user validates target region by imaging; good labeling radius with in situ cDNA synthesis
Cons: very low mRNA coverage; reliably reports only most abundant transcripts
10×/Visium (100)	On-slide primer array and cDNA amplification	55 μm	≥2 h	Fresh/frozen tissue	In situ cell transcriptomics
Considerations
Pros: user validates target region by imaging; commercial platform
Cons: low coverage confounded by weak spatial resolution (5–20 cells) of on-slide barcode array; expensive
Slide-seq (101)	On-slide bead array and SOLiD	20 μm	≥3 h	Fresh/frozen tissue	In situ cell transcriptomics
Considerations
Pros: reasonable spatial resolution (1–2 cells); 62% of beads in array map to 1 cell; cost-effective
Cons: no imaging; cell type position inferred from transcripts recovered; low mRNA coverage at 100–1,000 transcripts/bead
HDST (102)	On-slide bead array and cDNA amplification	13 μm	≥3 h	Fresh/frozen tissue	In situ cell transcriptomics
Considerations
Pros: good spatial resolution due to binning reads in hexagonal bead array (Lightseq > HDS > Slide-seq >> 10×/Visium); 86% of transcriptome identified across entire sample
Cons: no imaging; cell type position inferred from transcripts recovered; low mRNA coverage at 1.3% of transcripts/bead
Light-seq (103)	Light-directed barcoding and spatial indexing	2 μm	≥3 h	Fixed tissue	In situ cell transcriptomics
Considerations
Pros: good spatial resolution due to light-conjugated barcoding (Light-seq > HDS > Slide-seq >> 10×/Visium); pooling 25 similar cells yields 85% coverage of transcriptome ~3,500 transcripts
Cons: specialized microscopy and photolithography reagents needed

Barcode and bead array sequencing using 10×/Visium, Slide-seq, and high-definition spatial transcriptomics.

10×/Visium has commercialized a competing technology in which a glass slide is arrayed with barcoded reverse transcription primers (100) (Fig. 3E). The tissue is overlaid, imaged, and permeabilized, and local RNAs hybridize to the primers. After cDNA synthesis, the tissue is removed, and a library is prepared in vitro for and next-generation sequencing (NGS). Despite its low spatial resolution, 55 μm in diameter or 5 to 20 cells, the MacLeod lab used the 10×/VisiumST (spatial transcriptomics) to show that brains chronically infected with T. brucei contained tissue-remodeling macrophages, follicular-like CD4⁺ T cells, and plasma cells in reticular network formations resembling tertiary lymphoid structures (101). Moreover, T. brucei expressing “slender” or “stumpy” phenotype transcripts were in discrete brain regions, supporting a model in which T. brucei exploits the circumventricular organs to enter the central nervous system. In Mycobacterium leprae infection, reversal reactions are associated with tuberculoid leprosy and bacterial clearance rather than lepromatous lesions, and the biology is poorly modeled in animals. 10×/VisiumST showed that skin biopsy specimens from patients undergoing reversal reactions contained granuloma macrophages expressing antimicrobial peptides, whereas biopsy specimens from disseminated lepromatous patients did not (102). In reovirus-infected hearts, this approach showed cytotoxic T cells undergoing pyroptotic cell death contribute to myocardial pathology (103). 10×/VisiumST has been used to identify tumor-resident bacterial families and develop a GeoMX spatial profiling antibody panel that showed that intratumor bacteria reside in a poorly vascularize microenvironment, containing immune cells with tissue repair markers and low T cell activation markers (104). These studies highlight the utility of spatial transcriptomics to evaluate host-pathogen interactions that depend on tissue architecture or occur in non-model organisms as well as the importance of secondary studies to validate the function of differentially enriched transcripts.

Bead arrays, initially developed to mediate single-cell sequencing, have emerged as a high-resolution solution for spatial transcriptomics at a near-single-cell level (Fig. 3F). In Slide-seq, 1.5 million barcoded 10-μm beads are arrayed on silicon-coated slides and then overlaid with a tissue section (105, 106). RNA capture, primer hybridization, and imaging are performed on the slide array using SOLiD chemistry like FISSEQ. In a pilot paper, 65.8% of beads were clearly identifiable as a single cell type, consistent with 10- to 20-μm resolution, where 100 to 1,000 transcripts mapped to beads, depending on the tissue assayed. This technique has been applied to assess immune cell identity and localization in damaged kidneys (107).

High-definition spatial transcriptomics (HDST) captures RNAs on ~3 million barcoded beads arrayed across 1.4 million hexagonal wells placed 2 μm apart on a silicon wafer (108) (Fig. 3G). Tissue sections are laid on the array, and RNAs are captured by barcoded primers on the bead surface. Sequential RNA hybridization with label decoder oligonucleotides generates a unique fluorescence “address” for each barcode prior to in vitro cDNA amplification and library preparation. While the coverage for each spatial barcode is low, at 1.3% of transcripts per bead, 86% of all genes were detected across the entire array (similar to 10×/Visium and FISSEQ). Each cluster of 24 wells containing the bead array is 13 μm (~one cell). However, the functional resolution of HDST is higher because transcripts belonging to every possible group of 24 wells are compared (binned) to informatically assess the transcriptionally similar and dissimilar wells representing a cell border (a concept borrowed from hexagonal confocal microscope detector arrays like Zeiss Airyscan). Vickovic and colleagues reported that HDST has a 1,400-fold-higher resolution than 10×/Visium but a 25-fold lower resolution than Slide-seq. This superior spatial resolution may be tremendously valuable for host interactions with extracellular pathogens.

Photoactivated barcoding for in situ transcriptomics by Light-seq.

Light-seq was developed to circumvent the limitations of bead array density and deconvolution by labeling cDNAs synthesized in situ with UV cross-linkable barcoded DNA oligonucleotides (109) (Fig. 3H). An advantage of this approach is that unique barcode sets can be used to study of multiple cell types in the same sample, a distinct advantage for microbial pathogenesis experiments comparing infected and uninfected cell types or even immune cells and extracellular eukaryotic pathogens in the same microenvironment. In mixed-cell-line experiments, as few as 25 cells per phenotype (barcode) could be differentially enriched at 85% coverage of the transcriptome, another advantage to study rare cell types in a mixed population or tissue. Cell-type-specific targeting in the retina led to the differential enrichment of 3,400 transcripts profiling as few as 91 to 1,112 cells per phenotype. Although this technique has not been directly compared to array approaches, it is four times more sensitive than single cell sequencing and has a resolution of 2 μm, approaching HDST.

Sources of bias in spatial transcriptomics.

Limited read depth is a central caveat for spatial transcriptomics that leads to a reporting bias toward abundant transcripts. APEX-seq and TSA-BAR are less susceptible to these issues because spatial transcripts from millions of cells are pooled. Confident evaluation of transcripts belonging to a specific cell type generally requires a median read depth of 1 million reads, depending on the frequency of a cell type in a population (110). Infection diversifies the cell phenotypes in the tissue niche, further compounding coverage per cell type and downstream analysis. Simply increasing read depth may not be sufficient to enhance spatial transcriptome coverage because “drop-out” events, where a transcript is not captured or amplified, can be caused by mRNAs or cDNAs competing for a limited number of hybridization events in a constrained space (111). The resolution of the hybridization array also influences transcriptome coverage. For example, the limited number of barcoded amplicons in FISSEQ is partially compensated for by the high spatial resolution of in situ rolling circle amplification, whereas the coverage of the 10×/Visium platform is impaired by the poor spatial resolution of the array. An additional consideration for bacterial pathogenesis is that these spatial transcriptomic tools have not been optimized to label and enrich prokaryotic transcripts.

CONCLUSIONS

Spatial omics techniques are opening the door to understanding molecular interactions between microbes and their host cells, novel structures formed during infection, cell complexity in situ, and pathogens with limited experimental tools. Adaptations to the proof-of-principle protocols, originally designed to study eukaryotic biology, will be necessary to isolate unique microbial structures (e.g., hyphal wall), resolve highly paralogous microbial gene products (e.g., Var genes), and detect low-abundance microbial sequences in infected samples. Moreover, dramatic changes in gene expression that occur during infection may require additional controls, such as attenuated microbial strains or host genetic controls, to narrow candidates to the relevant biology. The rapid advance of single-cell transcriptomics and proteomics will likely accelerate the resolution of spatial techniques that similarly suffer from limited material and coverage. Despite these caveats, spatial omics techniques have enormous potential to impact our understanding of the relationships between pathogens and their primary cell or native tissue environment. Addressing these questions in molecular pathogenesis will likely refine these emerging technologies as well.