Using ecological network theory to evaluate the causes and consequences of arbuscular mycorrhizal community structure



See also the Commentary by Öpik and Moora


Arbuscular mycorrhizal fungi (AMF) are widespread and their symbiotic interactions involve the majority of terrestrial plant species (Wang & Qiu, 2006). These obligate biotrophs generally improve the nutrition and vigor of the host, thereby affecting individual plant traits (van der Heijden et al., 1998) as well as the composition and functioning of entire plant communities (Moora & Zobel, 1996; Hartnett & Wilson, 1999; Bever, 2002). Studies on individual plant traits are useful in determining fitness benefits to the plant (e.g. increased growth, resistance to pathogens, etc.), whereas studies on community-level interactions can potentially explain constraints on host–symbiont web architecture (e.g. Bluthgen et al., 2007). Community-level studies have been limited, however, to small subsets of natural plant communities, because processing and identifying AMF species associated with numerous plant root systems have proven costly and painstaking. Recent advances in next-generation sequencing technologies (Margulies et al., 2005) have removed this hurdle and improved the detection of rare AMF species (Opik et al., 2009). This increased capacity in describing whole plant–AMF networks provides an opportunity to identify the causes, and assess the functional consequences, of symbiotic network architectures (i.e. topology).

Network theory, originally developed to describe the flow of information within computational and social networks (Emerson, 1972), has more recently been applied to ecological studies of various mutualistic systems (Jordano et al., 2003; Olesen et al., 2007; Joppa et al., 2010). The major advantage of an ecological network approach is that topological metrics can be quantified for any given network involving two or more groups of interacting organisms (e.g. plants and pollinators, food webs, etc.). For example, ecological networks may be described in terms of their ‘nestedness’. High nestedness occurs when specialist species interact with a subset of partners with which generalist species also interact. For example, a specialist pollinator would tend to specialize on a generalist plant, and vice versa (Fig. 1a). This absence of reciprocal specialization was shown to be a pervasive feature of pollination networks (Bascompte et al., 2003; Joppa et al., 2009, 2010) that potentially favors diversity and stability of ecological communities (Memmott et al., 2004; Burgos et al., 2007; Bastolla et al., 2009; Thébault & Fontaine, 2010). Ecological networks can also be described according to their ‘modularity’, that is, the tendency of species to be grouped into modules in which interactions are more frequent than with the rest of the community (Fig. 1b). Thompson (2005) suggested that communities may assemble into distinct modules based on the functional complementarity of their traits, and this may offer some insight into coevolutionary dynamics between symbiotic species (Guimarães et al., 2007).

Figure 1.

Hypothetical interaction matrices sorted so as to depict the maximal nested state (a), or the maximal modular state (b), of a plant–arbuscular mycorrhizal fungal (AMF) network. Filled cells represent an interaction between a given plant and AMF species.

In this Letter, we argue that an ecological network approach could provide a framework by which to characterize and compare plant–AMF communities from different environments or at different successional stages. This, in turn, could improve our understanding of mechanisms structuring mycorrhizal communities and bring mycorrhizal science to a more predictive level (Johnson et al., 2006). In a recent study, Opik et al. (2009) used pyrosequencing to describe AMF communities associated with 10 plant species in a forest understory community. Here, we have used their published data set to demonstrate the applicability of ecological network theory to characterize plant–AMF communities. Our exercise revealed that this particular plant–AMF network was both highly nested and modular. We discuss possible reasons and implications for such topological features, and stress the potential for ecological network theory to direct future research on plant–AMF communities. In concluding, we argue that there may be a reciprocal advantage for advancing ecological network theory using plant–AMF communities as a model experimental system.

Data set

Opik et al. (2009) sampled individual root systems from 10 plant species in a 100 m2 plot established in a hemiboreal forest in Estonia. A total of 458 root systems were sampled, from which DNA was individually extracted. DNA extracts were pooled by plant species and PCR-amplified using the AMF specific primer AM1 and the general eukaryotic primer NS31. The exact conditions for PCR are described in Opik et al. (2009). Amplicons were pyrosequenced, yielding the number of sequence reads of each AMF taxon associated with each plant species.

Plant–AMF network topology

An interaction matrix was drawn using the data published by Opik et al. (2009) (Fig. 2). The matrix nestedness was calculated using the bipartite package of R statistical software (R Development Core Team, 2007). This metric varies from 0 to 100 (perfectly nested matrix). To assess the statistical significance of this nested structure, random matrices were generated using three different null models. These models use constrained randomizations of the original interaction matrix and, according to the degree of constraints, can be prone to either type I or type II errors. The first model, originally developed by Atmar & Patterson (1993), is a full randomization of the filled cells across the matrix. Here, the probability (ρij) of each cell (ij) to be filled in the random matrices is equal to 1/N, where N is the total number of filled cells in the original matrix. This model has been criticized (e.g. Ulrich et al., 2009) for overestimating the statistical significance of nestedness (i.e. type I error). The second model, proposed by Bascompte et al. (2003), partially controls for row and column totals, so that the probability (ρij) of cell (ij) to be filled is equal to (ρi + ρj)/2, where ρi and ρj are, respectively, the proportion of filled cells in row i and column j. This second null model is more conservative than the first in estimating the statistical significance of nestedness. The third null model fully controls for row and column totals, so that the probability (ρij) of cell (ij) to be filled is equal to (ρiρj). This third model is the most conservative of the three (i.e. most prone to type II error), as the total number of filled cells for each row and each column in each random matrix is equal to the corresponding total in the original data matrix from Opik et al. (2009). The third model thus controls for the effects of a species’ abundance on its level of generalism in partner choice (Vazquez, 2005). For each null model, 100 randomizations were performed and the nestedness of each outcome was calculated as described earlier. We considered nestedness to be significant if 95% or more of the random matrices of a given null model were less nested than the original data matrix.

Figure 2.

An interaction matrix in its maximal nested state, drawn from the published data set of Opik et al. (2009). Rows and columns represent, respectively, plant and arbuscular mycorrhizal fungal (AMF) species sampled in a 100 m2 forest plot. Abbreviations for plant species: oxa, Oxalis acetosella; gal, Galeobdolon luteum; vio, Viola mirabilis; par, Paris quadrifolia; hep, Hepatica nobilis; fra, Fragaria vesca; hyp, Hypericum maculatum; geu, Geum rivale; ver, Veronica chamaedrys; ger, Geranium pratense. Abbreviations for AMF taxa: Acau, genus Acaulospora; Scut, genus Scutellospora; GlomA, genus Glomus group A; GlomB, genus Glomus group B; GlomC, genus Glomus group C.

To analyze the modularity of the mycorrhizal network described by Opik et al. (2009), we implemented an algorithm in R software that was developed by Guimerà & Amaral (2005). The algorithm uses a simulated annealing procedure to distribute the species of the community in different modules in order to reach maximal modularity (Mmax). For more details about the algorithm, see Guimerà & Amaral (2005). After determining Mmax for the original interaction matrix, we assessed its statistical significance by performing 100 randomizations while controlling for row and column totals (i.e. using the third null model described above), and recalculating Mmax. We considered modularity to be significant if 95% or more of the random matrices were less modular than the original data matrix. We then performed a chi-squared test to assess the nonrandomness of AMF taxa among the modules identified by the algorithm.


The interaction matrix drawn from the mycorrhizal community described by Opik et al. (2009) demonstrated significantly higher nestedness (N) than the randomly generated matrices under the first two null models (original matrix, N = 82.6; null model I, N = 40.5 ± 3.4 (1 SD); null model II, N = 55.2 ± 4.3). Under the third null model, six out of 100 random matrices were more nested than the original data matrix, and 25 had a nestedness value above 80 (N = 76.9 ± 4.6).

The original mycorrhizal network was also found to be significantly modular (< 0.01), as all of the randomized matrices had lower Mmax values (0.204 ± 0.02 SD) than the original data matrix (0.264). Fig. 3 shows the original network divided into distinct modules according to the modularity algorithm. AMF taxa were not randomly distributed across these modules (χ2 = 66.6, df = 36, < 0.01). More specifically, members of the genera Acaulospora and Scutellospora were mostly confined to a single module associated with Viola mirabilis, a ‘forest specialist plant’ (sensuOpik et al., 2009). On the other hand, members of the Glomus group A clade were the most generalist in their partner choice and mainly found in the module comprising the most plant species.

Figure 3.

An interaction matrix in its maximal modular state, drawn from the published data set of Opik et al.(2009). Black cells represent recorded interactions found within one of the three modules identified by the modularity algorithm. Gray cells are interactions not included into any module, and white cells indicate that no interactions were observed between the corresponding species. Module affiliation is shown for AMF (above the matrix) and for plants (numbers in the left). Abbreviations for plant species: Oxa, Oxalis acetosella; Gal, Galeobdolon luteum; Par, Paris quadrifolia; Fra, Fragaria vesca; hyp, Hypericum maculatum; Geu, Geum rivale; Ver, Veronica chamaedrys; Ger, Geranium pretense; Hep, Hepatica nobilis, Vio, Viola mirabilis. Abbreviations for arbuscular mycorrhizal fungal (AMF) taxa: acau, genus Acaulospora; scut, genus Scutellospora; glomA, genus Glomus group A; glomB, genus Glomus group B; glomC, genus Glomus group C.

Plant–AMF network structure

Historically, all AMF species were considered broad generalists (Smith & Read, 2008), as laboratory assays demonstrated nearly complete compatibility between a range of host plants and cultured AMF species (Klironomos, 2000). This belief may have arisen from experimental artifact, as compatibility assessments can only be conducted with cultured fungi that are likely to exclude specialist and unculturable species (Sýkorováet al., 2007). Hence, both plants and AMF seemed to have a broad fundamental niche regarding their partner choice. It was later observed, however, that neighboring plants under field conditions can differ widely in their rootborne AMF communities (Vandenkoornhuyse et al., 2003; Alguacil et al., 2009). Here, we suggest that ecological network analysis can provide a valuable platform to evaluate the relative contribution of niche-based vs neutral mechanisms involved in plant–AMF community assembly.

A highly nested structure, as we found in the interaction matrix drawn from data by Opik et al. (2009), suggests that some AMF taxa specialize for only a few plant species. We thus argue that niche-based processes, driven by specific functional traits, may play a key role in the assembly of plant–AMF communities. For example, recent studies have suggested that the development of distinct AMF communities in the rhizosphere of different plant species (e.g. Johnson et al., 1992; Bever et al., 1996) may be driven by preferential allocation of plant carbon to the most beneficial fungal partner (Bever et al., 2009; Kiers et al., 2011). Also, both plants and AMF species have distinct seasonal peaks in their activities (Daniell et al., 2001; Pringle & Bever, 2002; Oehl et al., 2009), implying that phenological compatibility may be another niche-based mechanism driving partner choice. Likewise, our modularity analysis revealed a phylogenetic trend in the distribution of AMF taxa into different network modules. That important functional traits are conserved across major AMF lineages (Powell et al., 2009) lends more support to the notion that plant–AMF communities are constructed so as to maximize functional matching among partners (Thompson, 2005).

Ecological network analysis may also evoke neutral mechanisms for the plant–AMF community assembly, based on the abundance and spatial distribution of each species. For example, the nestedness of the interaction matrix drawn from data by Opik et al. (2009) was significant only when compared with null models that did not control for the observed abundance of each species. Given the correlation that should exist between the abundance of a species and its degree of generalism in partner choice, our results suggest that the plant–AMF network studied by Opik et al. (2009) relied at least partly on neutral assembly processes. One such process was proposed by Dumbrell et al. (2010), who found that a single fungal species displayed strong dominance in many AMF communities, with a disproportionately high number of subordinate AMF species. They suggested that fungal dominance was likely the result of a positive feedback occurring during the build-up of the plant–AMF community. A ‘founder AMF’ species colonizing plant roots earlier during ecological succession would benefit from more plant-derived carbon than ‘latecomers’, which would favor its growth and spread through the soil, and increase its probability of colonizing newly formed roots. This positive feedback, termed ‘preferential attachment’ in the network theory literature (Barabasi & Albert, 1999), has been found to cause nestedness in other types of mutualistic networks (Medan et al., 2007).

Our network analysis thus allowed us to conjure the existence of both niche-based and neutral mechanisms involved in structuring plant–AMF communities. From these, we may hypothesize a general assembly process based on successive filters (sensuDiamond, 1975), the first one being neutral and determined by overlapping spatial patterns, the second one being niche-based and determined by functional traits. Hence, during community build-up, AMF communities randomly associated to different plant species may gradually differentiate, subdividing the network into distinct functional modules. This is corroborated by data from Davison et al. (2011), who found that AMF communities associated with different plant species were more differentiated later in the growing season. To further verify this hypothesis, we suggest that more work be done to characterize the functional traits of AMF species belonging to same modules. This could be done by establishing pure cultures of AMF collected from a given site, growing them in standardized conditions and measuring ecologically relevant traits such as mycelial structure, hyphal life span, nutrient uptake, and C acquisition (van der Heijden & Scheublin, 2007).

Functional consequences of plant–AMF network topology

Besides providing insights into the mechanisms that may be responsible for plant–AMF community assembly, a network approach could also help us to understand the functional consequences of community structure. For example, high nestedness should limit interspecific competition among plants for AMF symbionts, thus favoring a higher diversity of coexisting species (Bastolla et al., 2009). Nested networks have also been shown to be more resistant to species extinction than randomly assembled communities (Thébault & Fontaine, 2010), thus conferring a greater stability to disturbance. However, those results arose from modeling studies that assumed only niche-based processes. In other words, two noninteracting species were assumed to be fundamentally incompatible. The recent demonstration that neutral mechanisms can also produce nested structures as those observed in nature (Krishna et al., 2008) calls for more work incorporating neutral mechanisms and their functional consequences.

Even though less work has been conducted to explore the ecological consequences of modularity, this topological metric may be important from an evolutionary viewpoint. Coevolution between plant and AMF species has naturally been studied on a pairwise basis, where a plant is inoculated with a ‘home’ or ‘away’ mycorrhizal community (e.g. Johnson et al., 2010; Callaway et al., 2011). Modularity analysis may allow us to refine our understanding by predicting that species belonging to the same modules should be better coadapted to each other (Guimarães et al., 2007). Yet another consideration in modularity analysis is the turnover of species within and across modules. As AMF community structure may vary over time (Dumbrell et al., 2011), it is likely that some species change modules and perhaps even alter between being a specialist or generalist species, suggesting that reciprocal selective pressures exerted between plants and fungi may be themselves fluctuating over time. Such temporal variability in the generalism of a species has been reported in other mutualistic networks (Diaz-Castelazo et al., 2010; Lazaro et al., 2010) and should be investigated in plant–AMF communities.

Advancing ecological network theory using plant–AMF communities

As we have discussed, ecological network theory is a promising approach to test the relative importance of niche-based vs neutral mechanisms involved in structuring plant–AMF communities, as well as to provide insights into the functional consequences of these structures. To face these challenges, there needs to be an empirical platform for testing various hypotheses. Most ecological networks that have been studied, however, do not easily lend themselves to experimental manipulation of community interaction patterns. For example, most data on mutualistic networks come from studies on pollination systems, because this mutualism is widespread and data are readily available. Those systems are, however, rather unsuitable for manipulative experiments, as it is hard to control which organisms will interact. For this reason, recent advances in ecological network theory have relied on modeling work to depict the causes and consequences of divergent network topologies (Dunne et al., 2002; Thébault & Fontaine, 2010). Inevitably, these models simplify the interactive complexity of real communities. For example, community simulations have mainly used fixed interaction matrices depicting constant species interactions through time, a scenario that is unlikely to occur in nature (e.g. Petanidou et al., 2008; Diaz-Castelazo et al., 2010; Lazaro et al., 2010). There is thus a need to design manipulative experiments that will test predictions made by these models.

The plant–AMF symbiosis may comprise a model experimental system for understanding the causes and consequences of different network topologies. It is possible to inoculate individual plants with specific AMF species and to grow these in a common garden. In other words, it is possible to build specific plant–AMF communities knowing the identity and initial abundance of each species, and the structure of the network. Such a model system could be used, for example, to test the importance of neutral mechanisms in structuring mutualistic communities, by testing the relationship between a species’ initial relative abundance and its level of generalism following the build-up of the community. Conversely, experimental manipulations of plant–AMF networks could be used to test the importance of niche-based mechanisms. For example, it was shown that phylogenetically distant AMF species have more distinct and complementary niches (Maherali & Klironomos, 2007; Powell et al., 2009) than closely related AMF species. This provides the opportunity to test whether roots colonized by phylogenetically overdispersed AMF assemblages are more apt to limit de novo root colonization (i.e. invasion) than those with phylogenetically clustered assemblages, as it is assumed that fewer empty niches would be left available in the former group (Elton, 1958). Finally, artificially constructed plant–AMF communities could be used to evaluate the functional consequences (in terms of species persistence, plant diversity, productivity, etc.) of different network structures. For example, it would be possible to test various hypotheses regarding the correlation between network nestedness and the rate of species extinction (Thébault & Fontaine, 2010), using experimental protocols designed to test soil microbial stability following stress and disturbance (e.g. Lacombe et al., 2009; Royer-Tardif et al., 2010).

Limitations of our analysis

In other mutualistic systems, it is generally accepted that interaction frequency is a good proxy for the functional impact of one species on its partner (Vazquez et al., 2005). On the other hand, we lack evidence that the number of AMF sequence reads in plant roots is indicative of the functional impact of the fungus on its host, especially when considering the wide variation in biomass allocation inside vs outside the roots (Powell et al., 2009). For this reason, we restricted our present analysis to two topological metrics that are quantified from binary (i.e. presence/absence) data matrices. Future work should strive, however, to find appropriate quantitative measures of interaction strengths in plant–AMF systems. For example, the number of independent interactions recorded in replicated data sets could be one way of corroborating results like those presented by Opik et al. (2009).

The fact that Opik et al. (2009) may not have sampled all potential host plants in their plot could bias our estimate of nestedness. Nevertheless, Nielsen & Bascompte (2007) showed that estimates of nestedness were generally robust against incomplete sampling designs. Moreover, Bascompte et al. (2003) showed that larger networks were consistently more nested than smaller ones, which implies that our estimation of nestedness was probably conservative.


New molecular tools, such as pyrosequencing technology, have increased our capacity to thoroughly describe plant–AMF communities in natural settings. Ecological network theory provides quantitative tools to study such data sets and to generate hypotheses related to selective partnering and community-level functional attributes. Conversely, the plant–AMF symbiosis, or perhaps mycorrhizal symbioses in general, comprise a model experimental system for advancing ecological network theory, such as testing hypotheses related to neutral vs niche-based mechanisms controlling community structure, and to the functional consequences of different network topologies.


P-L.C. was financially supported by an NSERC Postgraduate Scholarship as well as by a Vanier Canada Graduate Scholarship. The authors wish to thank three anonymous referees for their valuable comments to a previous draft of the manuscript.