Introduction

Microbial communities play critical roles in ecosystem functioning (e.g., nitrogen and carbon cycling) and in the persistence and health of many plant and animal species important to restoration, management, and agronomy [1,2,3]. However, as human activity in the Anthropocene continues to disrupt natural environments and microbial processes through increased environmental stress (e.g., higher temperatures [4]), there remains an urgent need to understand the factors that undergird microbiome stability in order to mitigate potential changes to ecosystem services. Previous work has shown that stress affects microbiome composition [5,6,7,8], microbe–microbe interactions [9, 10], and microbe–host interactions [11, 12]. However, the link between environmental stress and the stability of these microbial systems is less clear.

One promising avenue for understanding how stress affects microbial community stability is network analysis. When applied to ecology, networks are mathematical representations of communities in which nodes represent individual taxa and edges represent observed correlations in abundances among taxa from which interactions may be inferred [13, 14]. Network properties (e.g., modularity, sparsity, etc.) have been used to successfully predict the stability of networks of macroorganisms such as plant-pollinator networks [15] and food webs [16], and have recently been applied to microbiomes [6, 17]. In particular, communities with certain network characteristics—greater modularity, reduced positive associations among taxa, and greater negative associations among taxa—are more stable, meaning these communities: (1) have more limited shifts in composition in response to environmental perturbations and/or (2) are more likely to return to their equilibrium composition after a perturbation [17,18,19].

The first of these properties, modularity, quantifies how strongly taxa are compartmentalized into groups of interacting/co-occurring taxa (i.e., a module). Modularity can reflect biological processes such as shared ecological functions among taxa in a module [20,21,22], spatial compartmentalization [23], or similar habitat preferences [24, 25], and can impact community stability. For instance, high modularity in food webs stabilizes communities by restricting the impact of losing a taxon to its own module, thereby preventing the effects of that taxon’s extinction from propagating to affect the rest of the network [16].

The second property that can predict stability in co-occurrence networks is the ratio of negative to positive associations between taxa. Positive relationships represent high niche overlap and/or positive interactions between taxa, while negative relationships indicate divergent niches and/or negative interactions. The relative fraction of negative:positive associations can impact community stability [17, 18, 26, 27]. For example, positive interactions can destabilize microbial communities by creating positive feedback loops between taxa supporting each others’ fitness; thus, when one member in the positive feedback loop decreases in abundance it will negatively impact the fitness of other taxa reliant on the feedback loop’s services [18].

Here, we combine network theory and high throughput sequencing across 270 soil microbial communities from 40 replicate stress gradients to provide the first investigation of how the stability of microbial networks changes with stress. To do this, we characterized soil microbial communities from the Florida scrub, an ecosystem whose topography generates a series of naturally replicated stress gradients of decreasing nutrient and water availability [11, 28] within just a few square kilometers of area and a few meters of elevation gain [29]. We then assessed prokaryotic and fungal community properties and network structure across these replicate stress gradients to: (1) relate environmental stress and microbial diversity and (2) relate environmental stress and network stability using modularity and network cohesion (i.e., the strength and number of positive and negative associations). To gain insight into whether changes in network properties could be attributed to changes in species interactions, we (3) examined how the abundance of facilitative, stress tolerant, and pathogenic taxa change along the stress gradient.

Methods

Study system and field soil surveys

We conducted this study at Archbold Biological Station (27°11′ N, 81°21′ W) in the imperiled and endemic Florida Scrub (~15% remaining). This ecosystem, one of the oldest in Florida, is located along the Lake Wales Ridge in central Florida (Fig. 1). We focused on soil microbiomes from three scrub habitats (flatwoods, scrubby flatwoods, and rosemary scrub) that occur at increasing elevation relative to the water table (Supplementary Fig. 1; F2,132 = 35.45, P < 0.0001), and experience corresponding decreases in soil moisture [30] resulting in increasing stress [29, 31, 32]. The stress gradient associated with elevation is a strong determinant of plant community assemblages in this system [29], and recent work has shown that facilitative microbial effects on plant population dynamics increase in higher stress, higher elevation sites in the Florida Scrub [11]. Flatwoods occupy the lowest elevation sites and have the highest vegetation density followed by scrubby flatwoods (intermediate elevation and vegetation density) and rosemary scrub (highest elevation and lowest vegetation density). Rosemary scrub occurs as isolated patches across the landscape surrounded by scrubby flatwoods and flatwoods, providing naturally replicated stress gradients to test the relationship between stress and network structure.

Fig. 1: Schematic of sampling sites and stress gradients.
figure 1

(Left) Map of replicate stress gradient sampling sites at Archbold Biological Station, Venus, FL, USA (scale bar, 500 m). Dots indicate locations of the 40 replicate stress gradients from which soil samples were collected in the six environments (2 soil depths × 3 habitat types), and triangles represent 24 additional rosemary scrub sites that were sampled. (Right) Diagram of environmental gradient from low stress (flatwoods) to intermediate stress (scrubby flatwoods) to high stress (rosemary scrub). Images are examples of each habitat and have colored dots that represent habitat IDs used in the rest of this work (Photo Credits: D. Revillini, B. Almeida).

Full size image

In July 2017, soil samples were collected from the three habitat types. We sampled from 71 of the 106 mapped rosemary scrub patches at Archbold Biological Station, as well as from the neighboring scrubby flatwoods and flatwoods for each of the 71 replicate focal locations (Fig. 1). Sampling points for each soil core were randomly selected within the target habitats using ArcGIS (Version 10.5.1). Cores were divided into two depths: crust (0–2.5 cm depth) and subterranean (9–11.5 cm depth) for a total of 426 samples (71 locations × 3 habitat types × 2 soil depths). Each sample was the aggregation of 3 cores within ~2 m of the sampling point. Samples were flash frozen in 50 mL conical tubes in the field using liquid nitrogen and stored at −80 °C.

Microbiome DNA extraction and sequencing

Total genomic DNA was extracted from 1 g soil samples using the E.Z.N.A. Soil DNA Kit (OMEGA Bio-Tek, Norcross, GA, USA) following manufacturer’s instructions with minor modifications. To increase the amount of DNA obtained from sandy soils with few microbes, homogenization was performed in 15 mL conical vials instead of microcentrifuge tubes, then the supernatant was transferred to microcentrifuge tubes for heat incubation and alkaline lysis. We subsampled available extractions, selecting 40 sextuplet (3 habitat type × 2 soil depth) sample combinations that passed our initial quality control checks using endpoint PCR and gel electrophoresis. Because of our previous work showing strong interactions between the soil microbiome and endangered rosemary scrub plants [11, 33, 34], we sequenced additional samples from the rosemary scrub (24 crust and six subterranean samples). We also sequenced six negative controls in which 1 mL of Ultrapure Water (G-Biosciences; 786-293) was used in place of soil during DNA extractions to test for potential contamination during the extraction process. In total, 276 samples were sequenced (270 soil extractions and six negative controls). Fungal and prokaryotic genomic libraries were prepared and sequenced at the University of Minnesota Genomics Center (UMGC) following the two-step dual-indexing approach in [35]. 16S_V4 and ITS1 amplicons (Earth Microbiome Project primers [36,37,38,39]) were sequenced on two Illumina MiSeq lanes (v3, 300 bp paired end) at UMGC to characterize prokaryotic and fungal communities, respectively.

Bioinformatic processing

Reads were demultiplexed using bcl2fastq (Illumina). We merged reads from both lanes to counterpart fastq files, built directory structures, and created paired-end manifest files using Supplementary Files 1–4. We processed sequencing data with associated metadata (Supplementary Table 1) through QIIME2 (v2018.8; Supplementary File 5; ref. [40]) to remove sequencing adapters and chimeras, join paired-end reads, classify operational taxonomic units (OTUs), and calculate diversity metrics. OTUs, which were Exact Sequence Variants after denoising with Dada2 in QIIME2, were grouped into “species” based on 97% similarity to references in the Greengenes (v13_8; [41, 42]) and UNITE (v7_01.12.2017; [43]) databases for prokaryotic and fungal classification, respectively. Higher levels of taxonomic organization such as order and phylum were attached to species assignments. All six negative controls had <40 (20.9 ± 0.9) reads after denoising, which is consistent with a lack of contamination in our samples. Samples were rarefied to 2000 reads to calculate OTU richness, diversity (Shannon index), and evenness (Pielou’s index). Samples with <2000 reads (10 of 270 prokaryotic and 9 of 270 fungal samples) were excluded from analyses because their rarefactions curves did not reach saturation. Singletons were not removed, unless otherwise indicated, as their removal (OTUs with <2 or <20 reads) did not affect alpha diversity differences among habitats (Supplementary Fig. 2). We used combined relative abundances (i.e., taxon read count/total read counts in sample) of prokaryotes and fungi to construct co-occurrence networks [44, 45] and calculating cohesion as suggested by Herren and McMahon [17].

Microbial abundance and diversity analyses

We quantified the environmental stress experienced by microbial communities by measuring prokaryotic abundance (natural log of the number of genomic target sites for 16S_V4 primers per microliter in a sample) using a qPCR dual-indexing approach [35]. We used prokaryotic abundance as a proxy for microbial stress as it is analogous to using plant biomass to estimate stress in plant communities [46,47,48]. To quantify relative levels of stress between different soil depth × habitat combinations, we first performed an analysis of deviance in which a mixed model factorial ANOVA (fixed effects: habitat, depth, habitat  ×  depth; random effect: replicate focal location) was compared to a null model containing only the random effect followed by a Tukey’s HSD test. The resulting stress gradient from least (1) to most stressful (6) environment was: flatwoods crust < scrubby flatwoods crust < rosemary crust < flatwoods subterranean < scrubby flatwoods subterranean < rosemary subterranean. We then regressed prokaryotic abundance against the stress ranking using a Spearman’s Correlation test to confirm that prokaryotic abundance decreases (i.e., stress increases) along this ranked gradient. After this confirmation, we used Spearman’s Correlations to test how richness, diversity, and evenness varied along this stress gradient.

Microbiome network construction

In order to understand how the structure of microbial communities changed along the stress gradient, we constructed co-occurrence networks using SparCC (Version 2016_10_17; ref. [49]) with OTUs present in ≥10% of all samples [17] and default parameters. Significance of correlations between taxa abundances was calculated using 1000 permutations. SparCC was used in place of shuffled Pearson correlation coefficients because it accounts for sparse, compositional data. We constructed six networks (one for each habitat  ×  depth combination) rather than one large network. Building individual co-occurrence networks enabled comparisons of network structure, like modularity, across the stress gradient, and minimized the effects of broad habitat preferences for one of the six habitat × depth combinations (e.g., microbes forming a flatwoods crust module). This also allowed for comparing how co-occurrences between taxa change along the stress gradient. For example, if pairs of taxa switch from negative to positive associations when transitioning from low to high-stress environments, this relationship would not be visible in a network in which communities from all stress levels are combined. Essentially, strong negative and positive associations at low and high-stress extremes, respectively, would cancel each other out (i.e., taxa would have weak correlations), which would be indistinguishable from weak associations between taxa that remain constant along the gradient.

To provide insight into the role of stress in microbial networks, we then quantified two network properties that have been associated with stability of ecological communities in perturbation studies: (1) the number and strength of positive correlations, and (2) how compartmentalized the network is, via cohesion and modularity analyses, respectively [6, 17, 18].

Modularity

To understand how modularity changed across the stress gradient, we first identified modules (groups of taxa whose abundances are more correlated/anti-correlated with each other than the rest of the community) using the Clauset–Newman–Moore algorithm (greedy_modularity_communities from the networkx Python package; 50). We then calculated modularity, a measure of whether connections tend to occur within or between modules, using the quality function in the networkx Python package. The Clauset–Newman–Moore algorithm identifies modules by starting with each node in its own “module” and sequentially joining pairs of these modules that increase the modularity metric (Q) the most until additional pairing no longer increases modularity [50]. Modularity (Q) is calculated using the following equation from [51]:

$$Q \,=\, \frac{1}{{2m}}\mathop {\sum}\limits_{ij} {\left( {A_{ij} \,-\, \frac{{k_ik_j}}{{2m}}} \right)\delta (c_i,c_j)},$$

Where m is the number of significant pairwise correlations between taxa abundances in the network, Aij is 1 if OTUs i and j are connected and 0 if they are not, ki and kj are the number of taxa that have significant correlations with taxa i and j, respectively, and δ(ci, cj) equals 1 if i and j are in the same module and zero if they are not. The \(( {A_{ij} \,-\, \frac{{k_ik_j}}{{2m}}} )\) component of modularity calculates the probability that a connection exists between i and j given how well-connected i and j are to the rest of the network (i.e., Aij is the observed value and \(\frac{{k_ik_j}}{{2m}}\) is the expected value). This probability is used to weight connections between nodes within the same module (i.e., when δ(ci, cj) is equal to 1). Large, positive modularity values (i.e., close to 1) indicate that more connections occur within, rather than between, modules compared to random chance [52]. Communities with high modularity tend to be more stable because effects from changes in abundance of one species are more strongly limited to that species’ module [16]. We calculated one value of modularity for each environment’s network for a total of six modularity values. To determine if stress can explain variation in modularity, we used a Spearman’s Correlation to test the relationship between modularity and the stress rankings of our six environments.

Cohesion

By characterizing positive and negative co-occurrences separately, cohesion provides insights into associations among taxa caused by both positive and negative species interactions and/or by both similarity and differences in the niches of microbial taxa [17]. Two cohesion values (positive and negative) are calculated for each sample (j) as the sum of the significant positive or negative correlations between taxa weighted by taxa abundances:

$$C_j^{pos} \,=\, \mathop {\sum }\limits_{i \,=\, 1}^n a_i \cdot \bar r_{i,r \,> \, 0} \,\,\left( {\rm{Positive}}\,\, {\rm{Cohesion}} \right).$$

and

$$C_j^{neg} \,=\, \mathop {\sum }\limits_{i \,=\, 1}^n a_i \cdot \bar r_{i,r \,<\, 0}\,\,\left( {\rm{Negative}}\,\,{\rm{Cohesion}} \right),$$

where ai is the abundance of taxon i in sample j and \(\bar r_{i,r \,> \, 0}\) and \(\bar r_{i,r \,<\, 0}\) are positive and negative connectedness, respectively [17]. Within a given network (i.e., one of the six habitat × depth combinations), the positive (\(\bar r_{i,r \,> \, 0}\)) and negative connectedness (\(\bar r_{i,r \,<\, 0}\)) for a given taxon i were calculated as the average of all its significant positive or negative correlations with all other microbial taxa found in the network. Negative and positive cohesion range from −1 to 0 and 0 to 1, respectively, with higher absolute values signifying more and/or stronger correlations. Communities with larger negative cohesion values tend to be more stable, as their composition is less variable over time (i.e., have lower beta diversity between sampling times; 17).

We used the two cohesion values to evaluate two metrics of the microbial networks. First, we calculated the proportion of negative to positive co-occurrences as the absolute value of negative:positive cohesion, which allowed us to determine if more stressful environments are better characterized by processes driving negative associations, which might include competition and niche divergence, than less stressful environments. We then tested the relationship between stress and negative:positive cohesion using a Spearman’s Correlation test with the stress rankings of our six environments as the explanatory variable. Second, we tested the relationship between stress and total cohesion—a metric of network complexity calculated as the sum of positive cohesion and absolute value of negative cohesion—using a Spearman’s Correlation test with the six environmental stress rankings.

Next, to determine which cohesion value (positive or negative) was more sensitive to the stress gradient, we regressed both positive cohesion and the absolute value of negative cohesion against the six ranked environments (Spearman’s Correlation test) and compared these two correlations to each other using the paired.r function (a Fisher z-transformation followed by a z-test) in the R package psych. We also calculated the slopes of the relationships of positive and negative cohesion with stress using a nonparametric Siegel regression (mblm function in the mblm R package).

Microbial composition and function analyses

To gain additional insight into how the microbial community changed across the stress gradient, we further assessed abundances of different prokaryotic and fungal taxa as well as fungal functional guilds across the stress gradient. Counts for OTUs were normalized separately for prokaryotic and fungal communities by dividing by the total number of counts within a sample. We tested if taxa abundances were affected by the stress gradient by correlating abundance with the stress rankings of our six habitats using Spearman’s Correlation tests and correcting for multiple comparisons (Benjamini–Hochberg). We also assigned fungi to functional guilds by associating “species”, as identified with 97% similarity to representative sequences in the UNITE fungal database, with guilds in the FunGuild database [53] and filtering the OTU abundance tables for guild assignments with confidence rankings “Highly Probable” and “Probable” [54, 55]. FunGuild associates taxonomic groups to functional guilds such as “wood saprotroph”, “plant pathogen”, and “ectomycorrhizal fungi”, according to an online community annotated database. Using the same procedure as for taxonomic groups, we then tested if fungal functional guild abundances were affected by the stress gradient.

Results

Quantification of the stress gradient

The qPCR data measuring prokaryotic abundance established a clear stress gradient across all six habitat/depth combinations in which stress was inversely proportional to prokaryotic abundance (Fig. 2). Subterranean soils contained lower estimated prokaryotic abundance than crust soils (Depth: F1,187 = 269.4, P < 0.0001) and prokaryotic abundance decreased across our habitat stress gradient from flatwoods to scrubby flatwoods to rosemary scrub (Habitat: F2,187 = 58.2, P < 0.0001). Expanding on a priori knowledge of the study system’s gradient of habitat types, our data supported the interpretation of relative levels of stress across this ecosystem as: flatwoods crust < scrubby flatwoods crust < rosemary crust < flatwoods subterranean < scrubby flatwoods subterranean < rosemary subterranean. This sequence of habitat and depth combinations is strongly correlated with prokaryotic abundance (Fig. 2, Spearman’s ⍴ = −0.748, P < 0.0001). Thus, hereafter, we discuss the stress gradient following this sequence of six environments (assigned values 1–6).

Fig. 2: Quantifying microbial stress.
figure 2

Estimation of microbial stress by prokaryotic abundance in different habitat/depth combinations. Graph depicts natural log of the number of 16S_V4 primer target sites (i.e., number of prokaryotes) in soil DNA extractions. X-axis represents the categorical stress rankings of the six environments along the stress gradient. P value is derived from a Spearman’s Correlation test. Detailed statistics are available in Supplementary Table 3.

Full size image

Microbiome diversity and network properties across the stress gradient

To assess how microbial community properties change along the stress gradient, we compared richness and diversity across these six environments. Overall, fungal and prokaryotic OTU richness decreased with increasing stress across the six environments (Fig. 3a, b; Prokaryotes: Spearman’s ⍴ = −0.517, P < 0.0001; Fungi: Spearman’s ⍴ = −0.807, P < 0.0001). This decreasing relationship with the stress gradient was significant, but weaker for Shannon diversity (Supplementary Fig. 3a, b; Prokaryotes: Spearman’s ⍴ = −0.479, P < 0.0001, Fungi: Spearman’s ⍴ = −0.652, P < 0.0001), because species evenness remained fairly constant across habitats (Supplementary Fig. 3c, d; Prokaryotes: Spearman’s ⍴ = 0.015, P = 0.805, Fungi: Spearman’s ⍴ = −0.165, P = 0.008). The taxa in our co-occurrence networks that are present in our high stress (low diversity) sites are largely a subset of those present in our low stress (high diversity) sites (Supplementary Fig. 4). Only 4.6% (33/720) of taxa whose abundances were used to construct the six co-occurrence networks are unique to the high-stress subterranean samples (Supplementary Fig. 4a).

Fig. 3: Characterizing microbial diversity and network properties across the stress gradient.
figure 3

a, b Decrease in prokaryotic (a) and fungal (b) community richness (observed OTUs) across the stress gradient. Error bars depict standard error of the mean. c Decrease in modularity of microbiome networks across the stress gradient. A single value of modularity was calculated for each environment. d Decrease in ratio of negative:positive cohesion across the stress gradient. Box plots show inner quartiles and median negative:positive cohesion. Both negative correlations of modularity and negative:positive cohesion with stress are consistent with decreasing microbiome stability across the stress gradient. X-axis represents the categorical stress rankings of the six environments along the stress gradient. Detailed statistics are available in Supplementary Table 3.

Full size image

Furthermore, stressed microbial communities had lower network stability as evidenced by our analyses of modularity and cohesion. Microbial community networks decreased in modularity along the stress gradient (Fig. 3c; Spearman’s ⍴ = −0.886, P = 0.033), indicating that microbial communities in high-stress environments are less compartmentalized than in low stress environments. We also found that the ratio of negative:positive cohesion decreased with stress (Fig. 3d; Spearman’s ⍴ = −0.712, P < 0.0001), indicating that positive associations between taxa, rather than negative associations, dominate in high-stress environments. Similarly, total cohesion (sum of positive cohesion and the absolute value of negative cohesion) decreased with increasing stress (Fig. 4a, Spearman’s ⍴ = −0.896, P < 0.0001), indicating that community networks become less complex across the stress gradient.

Fig. 4: Changes in cohesion across the stress gradient.
figure 4

a Decrease in total cohesion (sum of positive cohesion and the absolute value of negative cohesion) across the stress gradient. b Decrease in positive cohesion across the stress gradient. c Decrease in the absolute value of negative cohesion across the stress gradient. Negative cohesion decreases more strongly across the stress gradient than does positive cohesion (P < 0.0001), driving the decrease in negative:positive cohesion seen in Fig. 3d. Box plots show inner quartiles and median values. X-axis represents the categorical stress rankings of the six environments along the stress gradient. P values are derived from Spearman’s Correlation tests. Detailed statistics are available in Supplementary Table 3.

Full size image

This reduced complexity of microbial networks (i.e., decreased total cohesion) was attributable to both positive (Fig. 4b, Spearman’s ⍴ = −0.658, P < 0.0001) and negative (Fig. 4c, Spearman’s ⍴ = −0.924, P < 0.0001) cohesion decreasing in magnitude across the stress gradient; however, negative co-occurrences were a much stronger driver of decreasing total cohesion than positive co-occurrences (Z = 9.25, P < 0.0001). In fact, the decrease in negative cohesion (Siegel slope = −0.0303; 95% CI (−0.0296, −0.0320)) was twofold higher than that of positive cohesion (Siegel slope = −0.0141; 95% CI (−0.0138, −0.0158)), thereby driving the decrease in negative:positive cohesion. This amounted to negative cohesion declining in magnitude by 53% between the lowest stress environment (flatwoods crust) and the highest stress environment (rosemary subterranean).

Characterization of taxonomic and functional groups across the stress gradient

We then quantified changes in prokaryotic taxa, fungal taxa, and fungal functional guilds across the stress gradient to understand which groups are impacted by the stress gradient. In general, we found that several oligotrophic (i.e., stress tolerant) taxa and plant mutualists increased in relative abundance along the stress gradient (Supplementary File 7). Oligotrophs positively associated with the stress gradient included the bacterial phylum Acidobacteria (Spearman’s ⍴ = 0.686, PFDR < 0.0001) and the bacterial order Xanthomonadales (Fig. 5a, Spearman’s ⍴ = 0.695, PFDR < 0.0001; most of our OTUs belonged to the oligotrophic family Sinobacteraceae and not the plant pathogen genera Xanthomonas or Xylella). Plant–mutualist groups positively associated with stress included the nitrogen-fixing Rhizobiales (Fig. 5a, Spearman’s ⍴ = 0.650, PFDR < 0.0001), the ectomycorrhizal fungal orders Russulales (Fig. 5b, Spearman’s ⍴ = 0.385, PFDR < 0.0001) and Boletales (Fig. 5b, Spearman’s ⍴ = 0.453, PFDR < 0.0001), and the arbuscular mycorrhizal fungal order Glomerales (Spearman’s ⍴ = 0.162, PFDR = 0.016). Conversely, plant pathogens tended to decrease with stress. The fungal order Capnodiales, of which our OTUs were predominantly from the pathogenic family Mycosphaerellacea, had the strongest negative correlation with the stress gradient (Fig. 5b, Spearman’s ⍴ = −0.700, PFDR < 0.0001). Abundance of Capnodiales decreased from making up a substantial portion of communities in the least stressful environment (7.7% ± 0.2%, mean ± s.e.m.) to 0.2% (±0.009% s.e.m.) in the most stressful habitats. Consistent with the taxonomic assessment above, we found that the plant pathogenic functional guild assigned using FunGuild decreased in abundance across the stress gradient (Spearman’s ⍴ = −0.701, PFDR < 0.0001), while ectomycorrhizae were the fungal guild most positively correlated with the stress gradient (Spearman’s ⍴ = 0.495, PFDR < 0.0001). In fact, the percentage of fungal species classified as members of mutualistic guilds that are positively associated with the stress gradient was ~4.5 times greater than among randomly drawn subsets of fungal species (P < 0.0001, see Supplementary Fig. 5 for details).

Fig. 5: Microbial orders most positively and negatively impacted by the stress gradient.
figure 5

a, b Top ten prokaryotic (a) and fungal (b) orders whose abundances are most positively or negatively correlated with the stress gradient (all correlations with P < 0.0001 after FDR correction). Abundances of taxa from kingdom to species identification are included in Supplementary File 6.

Full size image

Discussion

In this study, we found support for environmental stress destabilizing microbial community networks: microbes in higher stress habitats were less diverse, less modular, and dominated by positive co-occurrences (i.e., having lower negative:positive cohesion) compared to communities in low stress habitats. We also found that negative cohesion (i.e., strength/number of negative associations) decreased more strongly along the gradient than positive cohesion (i.e., strength/number of positive associations). To determine if the increased representation of positive correlations in high stress could be due, at least in part, to more positive interactions as opposed to greater niche overlap, we compared changes in abundance of mutualists between low and high-stress sites, finding that mutualistic taxa have strong positive relationships with the stress gradient while pathogenic fungi have strong negative relationships. Given that microbial communities often experience more species turnover (i.e., are less stable) when negative correlations are less frequent [17, 18, 56] and when network modularity is reduced by removal of taxa important to modular structure [19, 57], our results indicate that stressed microbial communities have lower network stability. This outcome is concerning, as it suggests that increased anthropogenic stress could select for microbial network topologies that are more prone to rapid fluctuations and regime shifts. Below we discuss ecological mechanisms through which environmental stress can destabilize microbial networks: (1) how changes in taxonomic and functional group composition can lead to increasing representation of positive correlations in high-stress sites, and (2) how decreasing modularity and negative:positive cohesion (i.e., fewer and/or weaker negative correlations relative to positive correlations) in stressed environments can combine to destabilize microbial communities.

Loss of antagonistic interactions and increased environmental filtering as drivers of decreasing negative:positive cohesion in stressful environments

Within microbiome networks, correlations between taxa can be driven by two non-mutually exclusive mechanisms: species interactions and environmental filtering [13, 58]. In the case of species interactions, positive correlations can be caused by facilitation/mutualism among taxa and negative correlations can result from competition [44, 59]. In the case of environmental filtering, positive correlations among taxa reflect ecological or functional similarity [60], while taxa with divergent niche requirements are negatively correlated. While both ecological mechanisms could explain the dominance of positive cohesion in high-stress habitats, our data support a role for species interactions. We observe decreases in relative abundance of pathogens and increases in relative abundance of microbial mutualists along our stress gradient, which agrees with previous complementary in vitro work demonstrating increased facilitation among microbes in stressful environments [9, 10]. We discuss both alternatives below.

The Stress Gradient Hypothesis [61], primarily developed to describe plant-plant interactions, provides general predictions about how species interactions change with increasing stress. It posits that as stress intensifies, the frequency of competitive interactions decreases and that of facilitative interactions increases, which is consistent with the decreasing negative:positive cohesion along our stress gradient. This outcome can occur through two mechanisms: (1) “competitive” taxa that engage in many antagonistic interspecific interactions being replaced by slow-growing, stress-tolerant species (e.g., oligotrophic microbes) as stress increases [62, 63], and (2) species that provide direct benefits to other community members increasing as stress increases [64]. Our data support the former. For example, we found that the oligotrophic phylum Acidobacteria [65] showed the strongest positive correlation with the stress gradient.

It is also likely that facilitative taxa increased in relative abundance along the stress gradient while antagonistic taxa decreased. In our study, plant–microbial mutualists are some of the taxa most positively correlated with the stress gradient across both fungi and prokaryotes. For example, ectomycorrhizal fungi were the fungal guild most positively correlated with the stress gradient, increasing from only 21% (±3.3% s.e.m.) in the least stressful environment to 64% (±4.7% s.e.m.) in the highest stress environment. We found a similar relationship between stress and plant–bacterial mutualists, such as the order Rhizobiales and family Sinobacteraceae [66]. We also saw concomitant decreases in species that engage in antagonistic interactions. For instance, the plant fungal pathogen guild showed one of the strongest negative relationships with environmental stress (Spearman’s ⍴ = −0.701). These results demonstrate that relative abundances of facilitative microbes increase and antagonistic microbes decrease with stress for at least one major group in the community, plants. Whether this extends to how microbes interact with each other is less clear because there is far less information about how different microbial taxa affect one another. However, experimental exposure of microbial communities to stress has demonstrated that microbes switch from competition to facilitative interactions when placed in toxic media [9] and engage in metabolic cross-feeding when a common nutrient is depleted [10]. As a result, it is likely that the increase in facilitative potential seen between microbes and plants in our system [33, 34] also extends to interactions between microbes themselves.

The alternative ecological mechanism, environmental filtering, may also drive the decrease in negative:positive cohesion in high-stress conditions. For example, the dominance of ectomycorrhizal fungi in high-stress environments may reflect that in xeric conditions very few niches are available for fungi, and that plant roots constitute a major proportion of available plant habitat in these circumstances. However, given that our data support both ecological mechanisms underlying the Stress Gradient Hypothesis (i.e., greater representation of stress-tolerant taxa and increasing facilitation), a decrease in the ratio of negative:positive biotic interactions is likely to be at least part of the microbial response to stress, even if it is not entirely independent of abiotic factors selecting for higher niche overlap.

Low modularity and positive associations as processes destabilizing stressed communities

While increasing dominance of positive correlations or reduction in negative associations in microbiomes can destabilize communities alone [17, 18, 56], the lower modularity in high-stress sites may exacerbate these consequences. For example, Agler et al. [19] found that removal of taxa important for structuring community subnetworks (i.e., modules) led to greater variation in microbiome composition than when those taxa are present (i.e., when modular structure is likely maintained). Modular organization of microbial communities can reflect: (1) groups of taxa that interact with each other more than with other groups, or (2) groups of taxa that share a niche distinct from other groups. In either scenario, a modular organization insulates these groupings from disturbances by ensuring that fluctuations in abundances of sensitive taxa within one module are unlikely to spread to taxa in other modules due to limited linkages between them [16, 67].

A major argument for the destabilizing effects of positive co-occurrence/interactions in microbial communities centers around the role of positive feedback loops in which microbes that support each others’ fitness outcompete other microbes [18]. As a result, the community as a whole becomes dependent on maintaining those feedback loops, and perturbations experienced by any of these loops’ members can rapidly propagate through the entire system [18]. In contrast, negative co-occurrences/interactions can lead to negative feedback loops that dampen not only perturbations that their own members experience, but also those experienced by linked positive feedback loops [68]. Low modularity in high-stress environments indicates that cross-module associations among taxa are likely to be more prevalent than in low stress environments; thus, environmental perturbations impacting taxa in one module would be more likely to propagate to other modules because there are relatively more associations (i.e., interactions or niche similarities) between taxa in different modules at high-stress sites (Fig. 6). For example, in our system, plant–microbial mutualists such as ectomycorrhizal fungi, Rhizobiales, and Sinobacteraceae could support each other by facilitating their shared plant partners, and the accompanying low modularity would mean that any perturbations experienced by this dominant positive feedback loop would be easily propagated to the rest of the microbial community. The increased connectivity between modules could be due to increased interactions with the plant–mutualist module (if linkages in the microbial network represent species interactions), or reliance on the same environmental factors (if linkages in the microbial network represent high niche overlap). For example, in our system, the importance of plant roots as a microbial niche may increase with increasing stress. As a result, environmental effects on their shared plant root niche (e.g., replacement of compatible plants with incompatible plants, destruction of the plant community, etc.) will impact more taxa under high-stress than low stress, because more microbes share plant roots as a niche requirement.

Fig. 6: Role for environmental stress in destabilization of microbial networks.
figure 6

Communities in high-stress environments (right) compared to communities in low stress environments (left) have less stable network properties. Communities under high stress are less species rich (a), less modular (b), and are less dominated by negative associations (lower negative:positive cohesion, c). The consequences for microbiomes (d) are that the reduction or loss of a taxon in response to an environmental disturbance (lightning bolt) can more easily propagate to the rest of the community (impacted taxa in black, unimpacted in gray). Positive associations (blue) are important pathways through which the effects on one taxon can cascade to impact other taxa in the community because: (1) the loss of a facilitator reduces the facilitated taxon’s fitness and/or (2) positively associated taxa are likely to be impacted by the same environmental factors (i.e., their niches strongly overlap).

Full size image

Conclusions and implications

Our study provides the first evidence that naturally-occurring microbiomes display network properties characteristic of unstable communities when under persistent stress. Because microbes can explain variation in traits important to plant distributions (e.g., growth rates, seedling establishment, secondary compound synthesis, etc. [69,70,71,72,73,74]) and undergird a plethora of ecosystem services (e.g., nutrient cycling; [3, 75, 76]), understanding how the stability of microbiomes is impacted by stress has important implications for modeling species’ distributions [77], maintaining ecosystem services [3, 77], and improving agriculture [1]. However, the development of general and testable principles for how these important, microscopic communities respond to even common environmental stress gradients is still in its early stages [3]. By characterizing microbiome network structure along a highly replicated, natural stress gradient, we demonstrate that, in high-stress environments, microbial taxa co-occur more frequently and the modularity of their communities breaks down.

As anthropogenic disturbances increasingly stress ecosystems, microbial communities may become destabilized. Our study system is one example of a naturally-occurring stress gradient that can serve as a model for environmental stressors that are predicted to increase in the Anthropocene (e.g., lower water availability; 4). While we find dramatic relationships between environmental stress and stable network properties (78% and 51% of variation in modularity and negative:positive cohesion was explained by our stress gradient, respectively), we encourage testing these relationships in a wider variety of ecosystems, including soil communities of other terrestrial biomes and microbiomes from aquatic habitats. We hope this study will provide a roadmap for characterizing microbiome network structures in a diverse range of study systems to generate broad insight into whether a negative relationship between stress and microbiome stability is a unifying principle across microbial communities. We also advocate for future studies elucidating how the consequences of unstable microbial communities cascade upwards to impact socially important goals such as restoration efforts and agriculture.