Dissecting microbial communities with single-cell transcriptome analysis

Abstract

Revealing insights into the function of microbial communities requires moving beyond measuring bulk taxonomic composition to detecting interactions between subpopulations. Following the transformative impact of single-cell gene expression profiling techniques on numerous fields of human biology, recent years have seen increased application to microbes. We review progress in the development of these techniques and discuss challenges in applying them to microbial communities. We highlight applications for dissecting the microbiome in human health and disease that reveal functional heterogeneity within gut communities, antibiotic responses, and the dynamics of mobile genetic elements. As single-cell gene expression technologies continue to develop, they are becoming ever more essential for examining and modulating the role of microbial communities in clinical and wider environments.

The microbial communities that colonize our bodies – our microbiomes – are central to our development, metabolism, and prevention of infection (1–5). They comprise thousands of species, primarily prokaryotic, but also eukaryotes such as fungi. Our understanding of microbiomes has been revolutionized in recent decades by sequencing-based technologies, including 16S rRNA gene sequencing and untargeted metagenomics that determine the relative abundance and genomic content of microbes in a given niche (6, 7). These approaches have revealed rich diversity throughout human development and during disease (5, 8–10). However, it remains obscure how these complex communities shape human health. Determining the functional role of specific microbial populations is challenged by the fact that not all genes present in a microbiome community are expressed, and expression may be temporally dynamic or altered during environmental shifts (11). While techniques such as metatranscriptomics allow interrogation of gene expression in these populations (11), these measure bulk expression averaged across a whole microbial species, or higher taxonomic level, and cannot distinguish variation in transcriptional state across cellular subpopulations.

There has long been an appreciation in the microbiology community of the importance of single-cell variation in gene expression (12) and phenotype (13), and recently we have witnessed an increased focus on single-cell technologies to understand microbiome composition and function (14). These include single-cell DNA sequencing (15, 16), flow cytometry (17), and single-cell techniques for isolation and culture of rare microbiome constituents (14). Microscopy-based techniques, including fluorescence in situ hybridization (FISH)-based labeling of 16S rRNA probes, have helped us to understand the spatial structure of microbiome communities (18, 19).

Single-cell transcriptomic analysis is particularly powerful thanks to its unbiased and comprehensive nature. Here, we review technical developments that enable genome-wide, single-cell-resolved transcriptional analysis and the opportunities and challenges in applying these to microbial communities, with a primary focus on human microbiomes.

Single-cell transcriptomics in microbes

Single-cell transcriptomics methodologies have enabled seminal contributions across mammalian biology (20–22), but several challenges have prevented their extension to microbes. For instance, the rigid microbial cell wall prevents lysis by protocols designed for the membrane envelopes of mammalian cells. While mammalian messenger (m)RNAs can be targeted by their poly(A) tails, the lack of mRNA polyadenylation in prokaryotes requires tailored methods for mRNA capture and reverse transcription. Moreover, microbial mRNA content is 10²-10³-fold less than mammalian cells, and bacterial mRNAs have far shorter half-lives within the cell (23). To overcome these problems, numerous methods for microbial single-cell transcriptomics have emerged in recent years (Table 1). These techniques fall into two broad categories (i) sequencing-based approaches (single-cell RNA sequencing, scRNA-seq), and (ii) microscopy-based approaches derived from FISH.

Table 1:

Comparison of single cell transcriptomics methods for bacteria.

Approach	Method	mRNA capture	rRNA depletion	Magnitude (cells, log10)	Specialized equipment	Applications	Strengths	Weaknesses
Combinatorial indexing	PETRI-seq	Random	Cas9, hybridization	3–5	None	Persister cell state in E. coli (33); heterogeneity and cell cycle transcriptional analysis in E. coli and S. aureus (24, 68); biofilm heterogeneity in E. coli (36)	- Highly scalable without the need for specialized equipment - Can multiplex up to 96 samples	- Labor-intensive - High initial cost of barcoding oligonucleotides - High initial cell input (10^8–9 cells)
	micro SPLiT	Random/poly(dT)	Poly(A) polymerase	3–5	None	Heterogeneity in B. subtilis metabolism, prophage, sporulation (29); host-plasmid interactions in Pseudomonas putida (131); phage infection in Bacteroides fragilis (130)
	BaSSSh-seq	Random	Hybridization	3–5	None	Biofilm heterogeneity and immune interactions in S. aureus (25)
Droplet-based indexing	smRandom-seq	Random	Cas9	3–5	Custom microfluidics	E. coli heterogeneous responses to antibiotic stress (26); human stool microbiome (39); bovine rumen microbiome (40); tested in Serratia marcescens (150), B. subtilis, Acinetobacter baumanniii, Klebsiella pneumoniae, Pseudomonas aeruginosa, S. aureus (26)	- Validated for microbiome samples	- Requires custom microfluidics not yet widely accessible
	ProBac-seq	Targeted probes	None	3–4	10X Chromium controller	E. coli and B. subtilis cell states; drivers of Clostridium perfringens toxin expression heterogeneity (30)	- Targeted probes enable selective capture over rRNA and host transcripts - Rolling circle amplification to regenerate probes improves affordability	- Probe design for each species required, so unclear whether broad coverage of microbiomes could be achieved
Hybrid	BacDrop	Random	Rnase H (cell stage)	5–6	10X Chromium controller	Heterogeneous expression of MGEs and antibiotic responses in K. pneumoniae; tested in E. coli, P. aeruginosa and Enterococcus faecalis (27)	- Scale is considerably greater than other scRNA-seq methods - 10X Chromium controller is accessible at many institutions	- Need for 10X droplet microfluidics increases cost relative to combinatorial indexing methods
	M3-seq	Random	Rnase H (library stage)	5–6	10X Chromium controller	Heterogeneity in E. coli stationary phase, antibiotic responses, and phage infection; B. subtilis prophage induction in response to DNA-damaging antibiotics (35)
Flow sorting	MATQ-seq	Random	Cas9	2–3	FACS; optional i.DOT robotics	Stress responses, non-coding RNAs, and virulence gene expression in Salmonella enterica; validated in P. aeruginosa (28, 34)	- Potential for higher transcripts/cell captured - Flow sorting allows for enrichment of desired subpopulations	- Lower throughput than other methods - Scaling requires access to automation
FISH	par-seqFISH	Targeted probes	None	5–6	Automated microscopy and liquid handling	Spatial heterogeneity and subcellular localization of gene expression in planktonic and biofilm P. aeruginosa (46)	- Incorporates spatial information at population and subcellular levels - Comparable in scale to highest-throughput scRNA-seq methods	- Requires probe design for each species - Trade-off between transcripts/cell and number of cells - Substantial barriers to accessibility
	bacterial-MERFISH	Targeted probes	None	3–6	Automated microscopy and liquid handling	Single-cell variation in E. coli responses to carbon source switch; subcellular localization of E. coli mRNAs; spatial variation in Bacteroides thetaiotaomicron colonizing the mouse gut (47)

scRNA-seq techniques

Prior to scRNA-seq analysis, microbes in single-cell suspension are immediately fixed to prevent RNA degradation. Enzymatic digestion of cell walls can aid in cell permeabilization. While eukaryotic mRNAs can be captured using oligo(dT) primers that bind to their poly(A) tails, bacterial mRNAs are captured either using random priming (24–29), addition of poly(A) tails using RNA poly(A) polymerase (29), or design of gene-specific probes (30). Random priming results in a large fraction of ribosomal (r)RNA, which makes up the bulk of RNA in bacterial cells, often >90% of transcripts sequenced by this approach (31, 32). Several strategies tailored to scRNA-seq have emerged to overcome this, including targeted cleavage of rRNA-derived library fragments using Cas9 nuclease (26, 33, 34), RNase H cleavage of rRNA hybridized to targeted probes (27, 35), and pull-down of rRNA-derived cDNA (25, 36). One RNase H-based method (27) uses commercial universal rRNA probe sets, which may offer the broad coverage required for rRNA depletion in microbiome samples.

Cell-specific tagging of the captured mRNAs with oligonucleotide barcodes is universally used across the scRNA-seq methods, but how tagging is achieved varies. Three bacterial techniques for scRNA-seq, prokaryotic expression profiling by tagging RNA in situ and sequencing (PETRI-seq; 24), microbial split-pool ligation transcriptomics (microSPLiT; 29), and bacterial scRNA-seq with split-pool barcoding, second strand synthesis, and subtractive hybridization (BaSSSH-seq; 25), are based on combinatorial indexing, which is derived from a method used for eukaryotes called split-pool ligation-based transcriptome sequencing (SPLiT-seq; 37) (Fig. 1A). Cells are not physically separated, but cDNA synthesis happens in situ within fixed, permeabilized cells, and each cell acquires a unique oligonucleotide barcode through iterative splitting and pooling steps (Fig. 1A). This approach is scalable to hundreds of thousands of cells. Combinatorial indexing methods do not require specialized equipment, reducing the barriers to their adoption.

Fig. 1: Principles of microbial single-cell transcriptomics technologies.

A) Combinatorial indexing produces unique single-cell indices by iterative splitting and pooling of cells and introducing one of 96 different barcodes first by reverse transcription and then by ligation. This produces 96³ (884,736) possible barcode combinations, enabling unique barcoding of each cell. B) Droplet-based indexing relies on the generation of nanoliter droplet emulsions in which a cell is encapsulated with an oligonucleotide-containing gel bead. This bead introduces a barcode unique to transcripts of that cell, typically by DNA polymerase using a barcoded primer associated with that bead. C) Hybrid approaches introduce initial plate-based barcodes by reverse transcription, then use nanoliter droplet generation to add an additional cell barcode, increasing throughput compared with (B). D) Flow-sorting sorts individual cells into multiwell plates, where cells are lysed and a cell barcode is introduced during downstream processing. E) FISH-based transcriptomics analysis uses highly multiplexed probe capture and resolution to detect transcripts. Unlike scRNA-seq (A-D), no cell-specific barcodes are introduced but instead transcripts can be associated with individual cells by image analysis.

One early method (28) used fluorescence-activated cell sorting (FACS) to sort individual cells into wells, prior to lysis and barcode introduction (Fig. 1D). This approach is limited by the difficulty of physically sorting more than a few thousand cells into individual wells of multiwell plates. However, mRNA capture rates per cell appear to be favorable and so fewer cells are required for sequencing and further improvements have since been made (34). There have currently been no direct comparative studies for microbial scRNA-seq, impeding our ability to assess the transcript capture efficiency of different methodologies.

An common approach in mammalian biology is to physically isolate and index cells within nanoliter droplets (Fig. 1B). Probe-based bacterial sequencing (ProBac-seq; 30) uses targeted probe arrays for mRNA capture and is compatible with the widely available 10X Genomics Chromium system (38) for droplet generation. Such probe-based capture avoids sequencing undesired rRNA or host RNA, but the need for species-specific probe design limits its use for multi-species communities. smRandom-seq (26) is another droplet-based technology that enables high-throughput profiling of cells and good per-cell transcript capture efficiency. It requires custom microfluidics and gel beads but it has already been used to profile complex microbiomes (39, 40).

Two “hybrid” methods, BacDrop (27) and massively-parallel, multiplexed, microbial sequencing (M3-seq; 35), combine nanoliter droplet and combinatorial indexing. This approach introduces an initial barcode to fixed cells in 96-well or 384-well plate formats that is then combined with a second higher-complexity barcode by droplet-based indexing (Fig. 1C). These methods require access to the 10X Genomics Chromium system but produce barcode complexity that enables sequencing of millions of cells. The initial plate-based indexing further allows for high-throughput multiplexing across samples.

Fungi polyadenylate their mRNA like other eukaryotes but possess a β-glucan cell wall that requires permeabilization by enzymatic digestion. Several nanoliter droplet-based approaches (41–44) and a SPLiT-seq adaptation (45) have been developed. SPLiT-seq incorporating random hexamer priming may be of particular importance because it has close similarities with bacterial combinatorial indexing protocols (24, 25, 29), suggesting that cross-kingdom single-microbe transcriptional profiling is achievable.

FISH techniques

Microscopic methods based on FISH provide single-cell resolution and spatial information at scales ranging from microbial communities to subcellular localization of mRNAs (Fig. 1E). By contrast to scRNA-seq, these methods can be applied to individual fixed cells, as well as biofilms or host tissues (46, 47). For example, parallel sequential FISH (par-seqFISH; 46), built on seqFISH technology (48), has been used for expression analysis of 105 marker genes in planktonic and biofilm Pseudomonas aeruginosa cells. This approach allowed visualization of spatial heterogeneity within biofilms and the subcellular localization of motility- and piocin-related genes. The use of combinatorial barcodes can further expand the range of genes captured by FISH-based approaches (47), but one problem in dealing with bacteria is the high density of mRNA transcripts compared to mammalian cells. This hinders transcriptome-wide combinatorial barcoding due to insufficient resolution of fluorescent spots derived from different mRNA species (47). A technique called expansion microscopy overcomes this challenge by embedding fixed bacteria within a matrix and physically expanding them to a larger volume (47). When combined with a bacterial adaptation of multiplexed error robust FISH (MERFISH; 49), transcripts from 80 percent of genes could be quantified. There is currently a trade-off between the number of cells that can be captured and the breadth of transcriptome coverage (47), and the need to design species-specific probes will limit the applicability of bacterial-MERFISH to multispecies communities. Assessing probe specificity and off-target binding is especially challenging in this context. The need for automated microfluidics and microscopy may also pose technical barriers to adoption. Nevertheless, the ability to profile bacterial communities in situ while retaining spatial information is a major advantage of FISH-based approaches.

Challenges for single-cell transcriptomics

Applying single-cell transcriptomics broadly to microbial communities requires protocols for unbiased and representative sampling. The first application of scRNA-seq to the human microbiome was on stool samples and revealed heterogeneity in expression of succinate metabolism genes and MGEs in one species, Phascolarcobacterium succinatutens, and provided a broader insight into phage-microbe interactions (39). Stool samples are accessible and widely used in human microbiome metagenomic analysis, but the instability of fecal transcriptomes raises questions about the broader utility of these samples for gene expression profiling. Animal models, where samples can be harvested and rapidly preserved by flash freezing, fixation, or addition of RNA stabilizing solutions (11), capture transcriptional variation in situ. Other samples may be more readily accessible, including the bovine rumen microbiome (40) which can be sampled through oral gastric tubes. There are examples of direct metatranscriptional analysis of human gastrointestinal tissues such as the appendix (11, 50), and microbial communities on the skin (51), in the mouth (4), and vagina (52) are easier to sample from humans.

scRNA-seq methods require preparation of single-cell suspensions, yet most microbial communities do not exist in planktonic form but in aggregates such as biofilms. Although sonication, mild detergent washing, density gradient centrifugation, and filtration can be used to disaggregate and isolate single cells (16, 53–57), for scRNA-seq, these methods must minimize transcriptome perturbation by proceeding quickly or at low temperatures. Disaggregation can lead to sampling biases, as seen in studies on mammalian tissues; fibroblasts in particular are underrepresented in tissue disaggregation procedures, since they are often embedded in extracellular matrix (58). FISH-based methods do not require single-cell suspension recovery and hence enable assessment of populations otherwise missed by disaggregation processes. Variation in the efficiency of cell permeabilization can lead to further sampling bias (39). Currently, single-cell transcriptomic technologies have been applied to species whose cell walls can be digested by lysozyme, lysostaphin (for S. aureus) or zymolyase (for fungi). Other cell envelope types, including the S-layers of archaea (59), may require different methods. No procedures for permeabilization have been evaluated in fixed cells from species lacking cell walls (e.g., Mycoplasma (60)) or from species with unusual membrane lipid compositions (61).

Computational challenges arise for single-cell analysis of complex microbial communities, particularly for assignment of sequencing reads. Metagenomic sequencing analysis can be assisted by assembly of reads into larger contigs, including metagenome-assembled genomes (MAGs). This is not possible for metatranscriptomics where reads are derived from individual RNA transcripts. Hence, assignment relies on mapping to reference genome assemblies. This is a problem for diverse communities containing poorly characterized species, and MGEs in particular may be absent from reference assemblies. In applying scRNA-seq to the rumen microbiome (40), Jia et al. leveraged existing microbial genomes and MAGs to construct a non-redundant, pangenome database for mapping sequencing reads across species. While this is a promising strategy, it is difficult to assess the completeness of such databases, since sequencing reads derived from genes not present in the database will go unmapped. Strategies are needed that integrate the growing volume of public metagenomics data (62, 63) to maximize the breadth of pangenome mapping, and metagenomics analysis needs to be paired with single-cell transcriptomics to support assignment of reads derived from MGEs and other frequently-missed elements. Integration of more specialized databases, for example mapping scRNA-seq reads from human stool samples (39) to the Gut Phage Database (64), will also help.

A fundamental challenge is the low mRNA content of each microbial cell. Bacterial mRNAs are transcribed, simultaneously translated, and rapidly degraded within minutes. Capture rates of ~100–400 mRNAs per cell (24, 29, 34, 35) mean most genes are not detected in any individual cell. The mRNA content of bacterial cells is far lower than that of eukaryotes e.g., 10³-10⁴ mRNAs per E. coli cell (65, 66) and the average mRNA is present at less than one copy per cell (65, 67). Data sparsity is exacerbated in slow-growing populations where mRNA content per cell is even lower (33, 68). We have found that single-cell variational inference (scVI; 69), which can be used for data smoothing, is effective for inferring variation in transcript content in bacteria (68). Nevertheless, such techniques must be used cautiously (70), and inference of within-population diversity in low-abundance species will likely remain a challenge.

Drivers of intra-species variation in microbial communities

Variation in expression is inherent even to isogenic populations in homogeneous environments. Within microbiome communities, environment and genetics will have the most profound effect, but transcription is intrinsically noisy (Fig. 2A) owing to stochastic gene expression processes (71, 72). Other factors shape variation in expression, including gene regulation (71, 73, 74), and interactions between transcription and chromosomal replication (68) (Fig. 2B).

Fig. 2: Sources of intraspecies transcriptional variation in microbial populations.

A) Gene-intrinsic noise in expression variation. Transcript abundance at the single gene level is governed by stochastic processes including reversible protein binding (e.g., by RNA polymerase and transcription factors), transcription elongation, and transcript degradation. B) Transcriptional variation can be driven by cell-intrinsic factors such as chromosomal replication during the division cycle (left). This creates patterns of global covariance in gene expression based on chromosome position. The pattern shown (right) is based on a simulation of rapidly-growing E. coli as in (68). C) Environmental sources of heterogeneity include microbe–microbe interactions mediated by metabolic cross-talk or direct interactions based on physical microbe-microbe contact. Heterogeneity can also be driven by proximity to the gut mucosal layer, for example due to variation in the abundance of mucin, host cells, or other microbes. Different species are denoted by shape, whereas different transcriptional states are denoted by cell color. D) Variation in expression can also be driven by within-species genetic variation. Such variation can arise from carriage of multiple strains in a single individual (denoted by different cell colors), or by the diversification of a single clone through de novo mutations such as single nucleotide variants (SNVs) or genomic structural variants (SVs) (indicated by differently-colored marks on the circular chromosome).

An individual microbial cell is profoundly influenced by its local microenvironment (Fig. 2C), which may include interspecies competition, physico-chemical factors, and host interactions. scRNA-seq can capture the effects of such exogenous variables on gene expression but for spatially organized drivers such as contact-dependent interactions or environmental gradients, microscopy-based methods are particularly powerful. Sarfatis et al. applied bacterial-MERFISH to the colon of mice colonized with Bacteroides thetaiotamicron (47) and observed expression variation such that B. thetaiotamicron within the lumen showed elevated expression of several central metabolism genes, whereas those near the mucus tended to express transporters and metabolic genes associated with polysaccharide utilization loci (Fig. 2C).

Microscopy-based methods can also enable the design of probes that detect host and microbe transcripts in parallel. For example, in inflammatory bowel disease (IBD), spatial transcriptional profiling of the host revealed immune cell infiltration and stromal cell activation (75), with regional variation in inflammatory state associated with altered microbiome composition (76, 77). Conversely, transcriptional analysis could conceivably be used to investigate bacterial gene expression states locally associated with host inflammation. However, probe-based capture limits the range of microbial transcripts that can be profiled. Untargeted scRNA-seq, which captures global expression variation, could act as a guide for the targeted design of probe panels that dissect the spatial context of this variation. Alternatively, scRNA-seq can be complemented with spatial transcriptomics (78), spatial metagenomics (79, 80), and RNA-seq of cell aggregates (81) to capture localized microbe-microbe interactions.

Finally, heterogeneity can arise from within-species genetic variation (Fig. 2D). For example, Anaerostipes hadrus exhibits structural variants at a locus encoding genes for butyrate and inositol metabolism whose presence correlates with lower risk of metabolic disorders (82). Strains of Segatella copri, an abundant species in fecal samples from non-Western individuals, exhibits considerable intra-individual genetic variation associated with differences in the production of sphingolipids and in their ability to elicit inflammatory responses in intestinal epithelial cells (83, 84). This analysis was done by isolating strains and profiling phenotypes in vitro, but alternatively single-cell transcriptomics could further dissect the effects of such genetic variation in situ within the natural host environment.

Genetic variants arise within a population by mutation, transfer of MGEs, and reversible phase-variation events, such as promoter inversion or intragenic events which are observed in microbiome constituents such as B. thetaiotaomicron (85–87). Through scRNA-seq, we can simultaneously trace the genetic variation itself, through calling of single nucleotide variants directly from sequencing data, and map the transcriptional states of these variants within the host environment. A major challenge is large variation in gene content, because many species possess a limited core genome shared between strains and a much broader accessory genome that varies between isolates (88). Single-cell transcriptomics alone cannot distinguish between loss of expression of a gene and loss of the gene itself. To identify gene loss events, additional genomic assays, such as sequencing of isolates or integrated analysis with scDNA-seq (89), may be required.

Applications for single-cell transcriptional analysis

Division of labor within microbiome ecosystems

Even within a single bacterial species there may be subpopulations that differ in their phenotype and function (90, 91) This is often described as a division of labor, in which subpopulations within an assembly perform specialist functions (90, 92). Multicellular biofilm communities exhibit divisions of labor whereby individual cells may adopt heterogeneous metabolic states, differential extracellular matrix production, and capacity for motility (92–95) (Fig. 3A). For example, optimal production of biofilm matrix in B. subtilis is achieved by differential biosynthesis of biofilm components among different cellular populations (94). Biofilm-like communities have been observed within the mucus layer of the gut (96). Division of labor has also been used to explain heterogeneous expression of host-modulating factors in microbial pathogens, where expression of these factors incurs fitness costs for that subpopulation but supplies benefits for the broader population of that species (97). Likewise, division of labor may explain heterogeneous production of effector molecules, such as bacteriocins, secretion systems, or competence genes, that mediate microbe-microbe interactions (98–100). Already, scRNA-seq has linked fatty acid metabolism to heterogeneous toxin gene expression in Clostridium perfringens (30), demonstrating how this technique can be used to identify the drivers of this variation.

Fig. 3: Division of labor within microbial communities.

A) Subspecialization associated with biofilm formation. Specific subpopulations may produce biofilm components, such as matrix, exoproteases, or effector molecules (46, 92, 149), or may adopt phenotypes of motility (102) or dormancy. B) Modes of species-function relationships within a community. Left: No functional redundancy. Each species (S) expresses genes for a separate pathway (P). Middle: Functional redundancy with no intra-species variation. Each cell of each species adopts a generalist state, expressing genes for all pathways. Right: Functional redundancy with intra-species variation. Each species expresses genes for all pathways, but any given cell may express genes for only one pathway. Hence, within a species different, subspecialized expression states emerge. Transcriptional analysis can distinguish all three modes at the cellular level, but at the bulk level the two functionally redundant modes cannot be distinguished. Thus, single-cell transcriptional analysis is required for resolution of the distribution of functions among cellular populations of individual species.

Application of par-seqFISH to P. aeruginosa biofilms shows subpopulations associated with metabolic responses, motility, or effector molecules targeting the host or other microbes (46). Similarly, diversification of gene expression required for adhesion, metabolism, and virulence can be observed in scRNA-seq of S. aureus biofilms (25). Wang et al. employed a strategy based on fluorescently-labelled D-alanine probes to sort E. coli cells based on their spatial localization within biofilms, coupling this with bulk RNA-seq to reveal metabolic division of labor and nutrient cycling within these communities (101). Combination of such labelling techniques with scRNA-seq could similarly enable indirect reconstruction of spatial information in structured populations.

Nevertheless, subspecialization in microbiome communities is largely unexplored. Previous studies have tended to rely on methods, such as microscopy of reporter strains, that limit the number of genes that can be studied and can only be applied to genetically tractable organisms (94, 98, 102). What is apparent from metagenomics analysis is the functional redundancy seen in microbiomes, in which coexisting, taxonomically distinct organisms perform the same function within these communities (103–105). Functional redundancy has been suggested for metabolic functions, such as short chain fatty acid production by a range of genera (104, 106) and widespread bile acid modulation (103, 104, 107). Multiple species have been associated with reduced host inflammatory cytokine signaling in models of IBD (108, 109). The concept of functional redundancy is frequently based on overlap in gene functions across species and can be challenging to test and quantify (103, 105). Single-cell transcriptomics provides the resolution not only to observe which species express genes for a given function, but also how different functions might be distributed among subpopulations of individual species. Several scenarios can be distinguished. First, genes for a pathway or function may only be expressed in a single species (Fig. 3B, left). Alternatively, several species may express genes for the same pathways, but at the single-cell level they may do so either in a generalist fashion i.e., genes for metabolic pathways are expressed together in every individual cell of that species (Fig. 3B, middle), or as subspecialist functions i.e., these pathways are expressed separately in a series of discrete states (Fig. 3B, right) revealing within-species division of labor.

Applying scRNA-seq to the bovine rumen microbiome, Jia et al. generated an atlas of 174,531 microbial cells comprising 2,531 species (40). Rumen microbes have a major metabolic role in digesting plant products and converting these to nutrients available to the host. To analyze the distribution of metabolic functions among the microbial community, the authors performed cross-species analysis based on functional groups of orthologous genes and partitioned cells into clusters based on the functions of the genes they expressed. Individual cell clusters differentially expressed genes involved in carbon metabolism for lipids, starch, and pyruvate, and other functions such as sulfur metabolism. When analyzed at this functional level, not only did cells from individual species display functional heterogeneity, but subpopulations of cells from different species expressed genes for overlapping functions (Fig. 3B, right). It will be important to account for stochastic or technical noise, for example by comparison to in vitro samples, and to test whether these observations can be generalized to other systems, including in human microbiomes. While gene expression alone cannot confirm that any given pathway is active in a cell, it nevertheless serves as a crucial indicator for how functions are distributed among cellular subpopulations.

Stress response dynamics in heterogeneous populations

Microbial communities reshape in response to stresses including those imposed by antimicrobial therapy (10) and cancer chemotherapy (110, 111). Loss of diversity during stress and incomplete microbiome recovery post-treatment are associated with negative patient outcomes (10, 112). Antibiotic stress may modulate expression of antimicrobial resistance genes among commensal species (113) but measuring bulk transcriptional responses to stress is confounded by shifts in underlying microbiome composition, since an increase in an organism’s abundance also increases the expression of that organism’s genes (11, 114). Moreover, bulk transcriptional analysis cannot capture the heterogeneity in responses to stressors such as antimicrobials (115, 116) or reactive oxygen species (117).

Adoption of heterogeneous stress-tolerant states is widespread among microorganisms. Stress tolerance mechanisms range from antibiotic-tolerant persisters, which have been found in microbiome species including oral Streptococcus mutans (118) and gut-colonizing E. coli (119), to sporulation, which provides an ultra-tolerant state of extreme dormancy in some bacterial species (Fig. 4A). scRNA-seq has already identified a low-transcription persister state in E. coli (33), heterogeneous expression of sporulation gene modules in C. perfringens and B. subtilis (29, 30), and variable antifungal responses in C. albicans (44). scRNA-seq analysis also shows that antibiotic exposure increases transcriptional diversity by stimulating heterogeneous transcriptional signatures of stress and DNA damage responses, cell wall synthesis, ribosome biogenesis, and MGE mobilization (26, 27, 35, 44) (Fig. 4b). Stress-induced heterogeneity, coupled with pre-existing heterogeneity within microbiome environments, means that there is likely to be a large repertoire of immediate responses that can be selected at the single-cell level.

Fig. 4: Heterogeneity in transcriptional responses to stress.

A) Spontaneous adoption of stress-tolerant states. Spore-forming species undergo a developmental pathway to form ultra-resilient spores. These may be reactivated in response to signals such as nutrient availability. Many microbes transiently adopt stress-tolerant persister states with phenotypes exhibiting low growth or energy state (e.g., low ATP or cell membrane potential). Circles represent cells in specific transcriptional states, identified by color. For endospore-forming rods such as B. subtilis, these are accompanied by representative schematic figures of cellular morphology during spore formation. B) Stress-induced heterogeneity. Initially homogeneous populations are driven into multiple transcriptional states in response to antibiotics. Stress states shown are based on Klebsiella pneumoniae responses to meropenem treatment as in (27).

Single-cell transcriptomics is well-suited for exploring heterogeneous stress responses as the data can reveal the emergence of discrete transcriptional states but can also expose broader effects on physiology. Gene dosage effects on transcription allow direct inference of DNA replication patterns from the transcriptome (68), and thereby of replication rates within microbiome samples (120). Ribosomal biogenesis can be inferred from the presence of nascent pre-processing rRNA transcripts, providing a means to infer growth rate from scRNA-seq data in vivo (121).

Incorporation of fluorescently-labelled orthologs of amino acids (122) or cell wall precursors (101) could also enable flow sorting into growing and non-growing populations.

Mobile genetic element transmission and function

MGEs broadly include genetic material shared horizontally between microbes, including antimicrobial resistance cassettes (123), virulence factors (124), and bacteriocins used for microbe-microbe competition (125). MGEs mediate effects on gut metabolism and immunity and can even be transmitted vertically from mother to infant (126). Bacteriophages are MGEs that supply a major route for horizontal gene transfer but could also play therapeutic roles in modulating microbiome composition. MGEs can have a broad host range, and determining their distribution among individual members of multispecies communities is a challenge for bulk metagenomics analysis, particularly for uncultivated microbes. Computational methods may enable assignments for chromosome-integrated MGEs but not for elements such as plasmids that are not contiguous with chromosomal DNA (Fig. 5A, left). Association of DNA elements in the same cell can be probed by scDNA-seq, Hi-C (127, 128), and PCR-based approaches (129), but these cannot determine whether a particular genetic element can be transcribed within a given microbial species, nor how acquisition of that element influences cellular phenotype (Fig. 5A, middle). A major strength of microbial scRNA-seq is that it can reveal the transcriptional consequences of interactions of individual bacteria with MGEs (Fig. 5A, right).

Fig. 5: Mapping MGE dynamics.

A) Bulk sequencing methods cannot associate plasmids, which are not contiguous with the chromosome, to a particular genome. scDNA-seq can do this, but by itself cannot determine whether a plasmid is active (i.e., genes are being expressed), or the effect of the plasmid on the host microbe’s physiology. By capturing transcription from both the plasmid and the host chromosome, scRNA-seq enables dissection of plasmid uptake, utilization, and the effects on the host microbe. B) Sources of MGE-driven expression heterogeneity in microbial populations. 1) Spontaneous mobilization of transposable elements leads to reversible TE gene expression and possible host responses. 2) Reversible plasmid acquisition in a subpopulation of cells can shape the transcriptome by transcription from the plasmid, by stress responses caused by the burden of plasmid maintenance, and by the acquisition of new traits, such as metabolism or antibiotic resistance, that enable growth of that subpopulation. 3) Bacteriophage driven heterogeneity. Initial infection leads to transcription of phage genes as well as possible host responses. Outcomes of this initial infection depend on the type of bacteriophage but can involve lytic cell death, horizontal transmission of genes through transduction, or adoption of a lysogenic state in which the bacteriophage integrates into the bacterial chromosome as a prophage, resulting in continued transcription of bacteriophage genes. Subsequently, stochastic or environmentally-triggered induction leading to prophage excision can lead to lytic cell death, but can also lead to reversion back to the initial, non-lysogenic state.

Recent in vitro studies have demonstrated the versatility of scRNA-seq in dissecting MGE-microbe interactions. Spontaneous expression of MGEs following mobilization events generates a high degree of heterogeneity that results in distinct subpopulations of cells observable by scRNA-seq (Fig. 5B), including expression of endogenous prophages (chromosome-integrated bacteriophages) (24, 27, 29, 35, 68) and transposable elements (TEs) (27). Heterogeneous bacteriophage gene expression during lytic infection of E. coli has also been detected by scRNA-seq analysis (35). Although in this case the transcriptional response of infected E. coli cells was modest, a recent study of bacteriophage infection of a gastrointestinal pathobiont, Bacteroides fragilis, has shown a more complex interaction involving variable expression of anti-phage defence systems and other temporal host cell responses (130). Interestingly, scRNA-seq of interactions between the soil bacterium, Pseudomonas putida, and the broad host-range plasmid, pKJK5, showed that heterogeneous plasmid gene expression coordinated with several bacterial host genes (131). Genes for plasmid replication and transfer were also heterogeneously expressed, suggesting variable roles of different bacterial subpopulations in horizontal transmission.

Microbiomes beyond the mammalian host

Microbial community functions are ubiquitous, ranging from global nutrient cycles (132), soil fertility and crop health (133) to human disease transmission by insect vectors (134, 135). The questions scRNA-seq can address in human microbiomes can equally be applied to these communities. However, human microbiomes are often nutrient rich and high in microbial biomass, up to 10¹¹-10¹² cells/ml (136). Since bacterial mRNA is unstable, its presence generally depends on continuous synthesis and ongoing metabolic activity. In most environments, sparser nutrients will lead both to lower microbial density and lower mRNA content per microbe. However, the dependency of mRNA content on growth and metabolic activity also means that single-cell transcriptomics can be an important tool for analysis of intraspecies variation in cellular activities, complementary to labeling techniques such as bioorthogonal non-canonical amino acid tagging (BONCAT), which have been used to identify the metabolically and translationally active fractions of microbial communities in environments from soil, to seawater, to glacier ice (122, 137–139). Single-cell transcriptomics has the added advantage of identifying specific expression signatures correlated with these growth states.

Prospects and concluding perspectives

Despite rapid progress, there remains substantial scope for new methodologies for single-cell transcriptomics in microbes. Important observations in mammalian biology have been made when scRNA-seq has been combined with other methods probing the genome, proteome, and epigenome (140–143). Adaptation of these combined techniques to microbes will provide valuable context for understanding transcriptional heterogeneity, for example the role of DNA methylation in bacterial phenotypic heterogeneity (144) or the influence of chromatin structure in fungi (145). The Perturb-seq technique integrates scRNA-seq with pooled CRISPR-Cas9 libraries for high-throughput gene disruption (146). This enables functional screening of individual genes in shaping a cell’s transcriptome and phenotype. While direct use of CRISPR-Cas9 disruption in bacteria is frequently lethal (147), viable alternatives in genetically-tractable species include targeted gene knockdown with CRISPR interference (CRISPRi) (148).

Further development of specific computational techniques would also enhance the value of bacterial single-cell transcriptome datasets. We recently showed that expression heterogeneity in bacterial scRNA-seq data is shaped by the global influence of chromosomal replication on transcript abundance (68). Modeling transcription-replication interactions revealed replication dynamics within a population as well as signatures associated with molecular regulatory states such as repression at individual gene loci (68). Such analyses have the potential to make single-cell transcriptomics valuable for phenotyping microbial cells in their native environment.

Microbiomes comprise microbial cells with heterogeneous gene expression programs, existing in variable environments, with intra-population variation emerging by strain mixing and horizontal gene transfer. Although we are beginning to understand how the complex ecology of our microbiomes influences our health, predicting effective intervention strategies remains an elusive goal. Single-cell transcriptomics is poised to become a broadly transformative approach that can reconceptualize microbiome dynamics in terms of cellular populations navigating their complex ecological niche.