Unveiling the complexity and ecological function of aquatic macrophyte–animal networks in coastal ecosystems

(1) Macrophyte communities in a changing world

Macrophyte communities are exposed to multiple local and global stressors, including nutrient enrichment, pollution, outbreaks of grazers caused by overfishing of predatory fish, invasive species, increasing sedimentation and temperature rise (Strain et al., 2014; Takolander, Cabeza & Leskinen, 2017). These major disturbance factors are driving extensive regime shifts in seaweed and seagrass assemblages worldwide, often resulting in the loss of complex, persistent canopy-forming macrophytes in favour of simpler, transient mat-forming species with lower ecological value (Benedetti-Cecchi et al., 2001; Airoldi et al., 2014; Mineur et al., 2015; Unsworth et al., 2015; Gorman et al., 2020).

Changes in macrophyte community composition inevitably alter the range of ecological services that they provide to their associated fauna via trophic and non-trophic relationships, with consequent shifts in animal diversity and abundance (Benedetti-Cecchi et al., 2001; Edgar et al., 2004; Lilley & Schiel, 2006). The mechanisms through which disturbance propagates from macrophyte to animal communities are complex and elusive: as macrophyte–animal systems are embedded in larger littoral food webs, perturbations at the level of macrophyte–animal links are likely to propagate to higher trophic levels through cascading effects (Cordone et al., 2018; Strona & Bradshaw, 2018) and may affect species of particular ecological and/or commercial value. In worst-case scenarios, these domino effects might ultimately drive whole regime shifts in coastal ecosystems (Wernberg et al., 2016).

The primary role of macrophytes in maintaining littoral biodiversity and coastal ecosystem functioning, and their vulnerability to human activities and climate change, call for a deeper understanding of the complex mechanisms through which macroalgae and seagrasses sustain their associated biological communities. In this context, ecological network analysis might provide a powerful tool to investigate the structure and dynamics of macrophyte–animal associations.

(2) Macrophyte–animal networks: reinventing the wheel?

Studies of aquatic macrophytes (and their associated fauna) have largely focused on coastal food webs. These are typically represented as a set of nodes, symbolising different taxa or functional groups, and links, illustrating the possible trophic interactions between each given pair of nodes (Fig. 2). Why are we then claiming novelty in the application of network theory to macrophyte–animal associations? To answer this question, it is important to clarify that when discussing macrophyte–animal networks, we are not referring to a subset of nodes and links extracted from a complete coastal food web but, instead, to an organic entity that stands on its own.

Considerable work has been done to clarify the trophic relationships between aquatic macrophytes and their consumers, especially through stable isotope analysis (Crawley et al., 2009; Porri, Hill & McQuaid, 2011; Bessa, Baeta & Marques, 2014; Kahma et al., 2021; Liénart et al., 2021). However, when considered as parts of ‘complete’ coastal food webs, the macrophyte–animal component is often regarded as a generic source of energy for the upper trophic levels, while the dynamics of the links directly connecting macrophytes to animal species are often neglected. With few exceptions (Coll et al., 2011; Kéfi et al., 2015; Cordone et al., 2018; Momo et al., 2020), macrophyte and invertebrate links are frequently pooled into single nodes in the food-web representation, while only links connecting nodes at higher trophic levels are represented in full detail. This imbalance in detail is mostly due to the fact that information on trophic interactions in coastal food webs frequently stems from the analysis of predators' diets (Olivier et al., 2019), while identifying large numbers of small-body-sized organisms – and the trophic interactions they participate in – is more challenging.

However, this over-simplification is likely to affect our understanding of the processes taking place at the macrophyte–animal interface. In particular, combining multiple species into single food-web nodes fuels the assumption that pooled species are ecologically equivalent, and hence that changes in their diversity or relative abundances will have no effects on food-web functioning and other ecosystem processes (Fig. 2A, B). This perception stands in sharp contrast with various empirical and observational studies showing how changes in macrophyte community composition can have strong effects at higher trophic levels (Edgar et al., 2004; Wernberg et al., 2016). Conversely, identifying the potentially complex structure of macrophyte–animal associations might reveal novel patterns, and enhance our ability to model processes at the macrophyte–animal level and their potential effects on the rest of the food web (Fig. 2C, D).

(1) The complexity of ecological interactions in macrophyte–animal systems

The ecological relevance of macrophytes in coastal communities is not limited to their trophic role, but encompasses a complex range of non-trophic antagonistic, mutualistic, and commensal relationships (York et al., 2018; Momo et al., 2020; see also Section II). Sometimes the boundaries between interaction types are blurred and assessing whether an animal species is beneficial or detrimental to the associated macrophytes can be complicated (see Section V.2). For example, a common form of interaction in macrophyte systems is epiphytism. In general, macroalgae and seagrasses facilitate sessile invertebrates by providing surfaces for attachment, and hence decreasing competition for space (York et al., 2018). By growing on kelp fronds, colonial hydrozoans reduce competition for space in the benthos and increase their accessibility to food resources in the water column. At the same time, they provide their macrophyte hosts with ammonium, while defending them from grazers (Hepburn & Hurd, 2005; González-Duarte, Megina & Subida, 2020). However, epiphytic organisms, such as bryozoans, often have negative effects on macrophytes, as they reduce access to light and diffusion of nutrients and gases (Muñoz, Cancino & Molina, 1991; Hurd, Durante & Harrison, 2000).

Other mutualistic interactions can be observed between mesograzers, such as amphipods, and macroalgae. Amphipods facilitate macroalgae by feeding on smaller epiphytic algae that compete for light and space, while macroalgae provide amphipods with a refuge from predators (Amsler, McClintock & Baker, 2014).

Another aspect that can complicate interpreting the nature of ecological interactions in macrophyte–animal systems is that this nature might shift when abiotic conditions change. For example, sponges and seagrasses can establish commensalistic interactions, where the growth of the former has a neutral effect on the latter (Archer, Hensel & Layman, 2018). However, under high nutrient loading, the interaction can become detrimental for the seagrass and shift to antagonistic (Archer et al., 2018).

In specific settings, macrophytes too might have a negative impact on the associated fauna, as in the case of seaweeds with large fronds, which can impede faunal settlement and grazing through the so-called ‘whiplash effect’ (Iken, 2012). Periodic shedding of outer cell walls by some algae also removes epiphytic layers and associated grazers (Russell & Veltkamp, 1984). However, in general, canopy-forming macroalgae maintain diversified assemblages of understorey algae and invertebrates (Benedetti-Cecchi et al., 2001; Strain et al., 2014).

The general picture emerging from these examples of interactions is composite and suggests that, to disentangle the complexity of macrophyte–animal associations and to explore their structure effectively, we need an appropriate analytical approach.

(2) A multi-layer framework

How to synthesise the complexity of species interactions within a single theoretical and analytical framework is a significant ecological challenge (Jordano, 2016b). For instance, in terrestrial plant–animal systems, many different trophic (e.g. plant–herbivores) and non-trophic interactions (e.g. plant–pollinators and plant–seed dispersers) have been shown to play simultaneous, intertwined roles in ecological processes (Pocock, Evans & Memmott, 2012). Traditionally, different kinds of interactions have been addressed as separate entities in specific branches of network ecology (Bascompte & Jordano, 2007; Donatti et al., 2011; Araújo, 2016). Recently, attempts have been made to move beyond the strict categorisation of networks and to investigate multiple types of ecological interactions within a unifying, multi-layer network framework (see Melián et al., 2009 and Pocock et al., 2012 for plant–animal systems, and Lafferty et al., 2008 and Kéfi et al., 2015 for food webs). Networks composed of multiple layers of interactions might offer an increasingly realistic representation of ecological complexity (García-Callejas, Molowny-Horas & Araújo, 2018) by capturing better the multifaceted nature of species interactions in natural systems.

Given the composite suite of relationships linking macrophytes and their associated fauna, a multi-layer network approach appears an optimal choice to explore these systems. However, while the approach has solid theoretical foundations from physics (see Pilosof et al., 2017), integrating multiple interaction types into single networks poses many practical challenges. As discussed above, identifying which interactions lie behind observed associations is, in general, extremely challenging. This is true also for well-investigated systems: for example, ‘cheating’ strategies in plant–pollinator (Thakar et al., 2003) or client–cleaner interactions (Guimarães Jr et al., 2007) greatly complicate the distinction between mutualistic and antagonistic interactions. In aquatic macrophyte–animal systems, the simultaneous sampling (and identification) of multiple kinds of interactions is an ambitious task, due to the variability of the associations in space and time (Christie et al., 2009; Machado et al., 2019). Moreover, there are practical challenges in identifying and recording species interactions in the marine environment compared to terrestrial habitats. However, the promising results obtained by the first studies on multi-layer networks (Lafferty et al., 2008; Melián et al., 2009; Pocock et al., 2012; Kéfi et al., 2015) should encourage ecologists to embrace the challenge.

One practical way to assemble multiple-interactions networks is to take advantage of pre-existing knowledge on species interactions within a target area or system (Kéfi et al., 2015; Puche et al., 2020). As an example, Kéfi et al. (2015) assembled a comprehensive network of trophic and non-trophic interactions in the Chilean rocky shore intertidal community from a large data set encompassing studies conducted over 40 years. However, such an approach is an exception rather than a rule. In the specific case of macrophyte–animal associations, information on the nature of ecological interactions is sparse. Although this has been thoroughly investigated in some systems, as is the case for Antarctic seaweeds and their associated fauna (Amsler et al., 2014; Heiser et al., 2020; Momo et al., 2020), sufficient pre-existing information is typically lacking. This knowledge gap highlights the need for de-novo observational and experimental studies. Considering the motivation behind this review, we are certainly at too early a stage to set such an ambitious goal, and it therefore seems appropriate to start by looking at single layers of associations. Nevertheless, it is important to identify a multi-layer representation of macrophyte–animal networks as the long-term target, to ensure that the steps taken now will continue to move us in the right direction.

In the following section, we attempt to start that process by reconstructing some preliminary macrophyte–animal networks based on data sets available in the literature. In line with the considerations above, our source data do not include details about the type of interactions (e.g. mutualism, commensalism, antagonism) behind the observed associations. However, they allow us to illustrate how a network perspective can be used to extrapolate novel valuable information not attainable with other approaches. They also allow us to demonstrate how a network approach can significantly change our understanding of the macrophyte–animal systems under study. Our hope is that these simple examples will stimulate the interest of researchers to either reconsider already-collected data from a new perspective and/or to start collecting new data with network analysis in mind.

(1) Building networks from literature data

We identify two categories of data sets that can be used to build tentative ecological networks of macrophytes and animals. The first consists of potential interaction data, which are most often represented by checklists of the fauna associated with a given and often small set of macrophyte species. Usually, such data sets report quantitative information (e.g. the number of individuals or the biomass of a given animal species found on a given macrophyte individual/species). The second category consists of co-occurrence data, for example of lists of macrophyte and animal species found within the same sampling unit. With appropriate caveats (Freilich et al., 2018; Blanchet, Cazelles & Gravel, 2020), this information can be used to identify pairwise associations, by establishing which species tend to be found together more often than predicted by chance and/or by a set of environmental/ecological factors. Pairwise associations can then be combined into association networks (Araújo et al., 2011; Berry & Widder, 2014; Harris, 2016). It should be highlighted that the idea of using co-occurrence to infer ecological interactions is currently debated (Freilich et al., 2018; Blanchet et al., 2020). A key aim of this review is to encourage studies specifically designed to collect reliable data on species interactions. Thus, we present here the approach that is best able to obtain additional information from data originally collected for purposes different from network analysis. The distinction between the two categories can be subtle, as data of fauna associated with macrophytes are somehow a (stricter) form of co-occurrence. However, the following two examples may act to clarify the distinction.

As an example of the first category (data set of pairwise associations/checklists of animals associated with a given macrophyte individual/species), we use the data from Table 1 in Huang et al. (2007). This table reports the number of gammaridean amphipods found on eight different dominant macroalgal species from the West Antarctic Peninsula. Ideally, this measure could be used to build a weighted network. The most straightforward way to do this would be to generate a link between an algal and an amphipod species, with the weight of the link equalling the number of individuals reported in the table. A further step would be to standardise the link weights by dividing them by the total number of amphipod individuals (of all species). Such a choice would emphasise the absolute strength of the interactions, by giving higher weights to dominant species compared to rare ones (Fig. 3A). An alternative approach to standardise weights would be to divide the number of amphipods found on given algae by the total number of amphipods of that species. Such an approach would emphasise patterns of amphipod specialisation, without penalising rare species occurring at low abundances (Fig. 3B).

Regardless of the criterion adopted to attribute and standardise weights, a weighted network can be investigated without transformation, or first be transformed into an unweighted network. Such transformation requires identifying a threshold for the interaction strength below which links are disregarded. For example, referring again to the data set from Huang et al. (2007) one could build an unweighted network (Fig. 3C) by arbitrarily considering an amphipod as associated with a target alga if at least 1% of the individuals of a given amphipod species was found on the target alga. This is equivalent to imposing a threshold for interaction strength of 0.01 on the network reported in Fig. 3C. While the strongest potential associations can be derived directly from the original table, exploring how the density of ecological interactions declines as we raise the threshold for interaction strength could yield insights into the degree of specialisation of the system.

Although setting a threshold might be arbitrary in itself, the exploration of how different thresholds affect the outcome can be used to guide the choice. For example, in the data set from Huang et al. (2007), consider the weighted network obtained by dividing the number of individual amphipods found on one alga by the total number of individuals of the same amphipod species found on all algae (Fig. 3B). In describing relevant amphipod/algae patterns of specialisation regardless of species abundance, it might make sense to use the minimum value ensuring that all the amphipod species are included in the network. This threshold turned out to be 0.295 (Fig. 3D). However, in the case of the network with weights standardised by the total number of individuals (Fig. 3A), it might make sense to choose a threshold regardless of its effect on the number of amphipod species left in the network. Since one interesting aspect of that network would be identifying strong associations possibly having important ecosystem effects, choosing a threshold which simultaneously eliminates links involving few individuals and rare amphipod species could be a good choice. A possible option could be that of selecting as a threshold the minimum interaction strength ensuring that all algae are left in the data set (i.e. 0.0014) (Fig. 3E). This threshold leaves us with 11 amphipod species (from the initial 32), and a quite simplified version of the network possibly representative of the ‘core’ interaction set. However, there is no right or wrong criterion in this regard, nor a ‘golden rule’ applicable to all situations.

As an example of the second possible category of data sets that might be used to obtain a network (that is data sets reporting the independent occurrence of macrophytes and animals across a set of samples), we compiled two macrophyte–animal networks using a data set from Nakamoto et al. (2018). This data set was generated by assessing the diversity and biomass of macrophytes in a subtropical seagrass–seaweed mixed bed in Okinawa. While the original study aimed to assess the effect of macrophyte phylogenetic diversity on the species diversity of associated epibenthic invertebrates, the same data can be used to identify, with a network approach, relevant associational patterns. As examples, one may score which macrophyte–invertebrate associations are maintained over time and the degree of specialisation of the system.

In the original study, researchers set twelve 50 cm × 50 cm quadrats and then counted mobile epibenthic invertebrates in each quadrat in two consecutive years. Due to the sampling procedure, the data set provides information only on the number and identity of invertebrates and the respective biomass of all macrophytes found within the same quadrat. This does not permit a direct network reconstruction. We therefore inferred potential associations based on the co-occurrence between animal and macrophytes across quadrats. For this, we first quantified the observed co-occurrence for any possible macrophyte–animal pair as the total number of quadrats where both candidate interacting partners were found. Then, to determine which species associations occur more frequently than expected by chance, we compared the observed co-occurrence with that measured in 1000 randomised assemblages obtained by randomly reallocating animals across quadrats. These assemblages were constructed by keeping the number of occurrences per macrophyte and animal operational taxonomic unit (OTU) fixed, while permitting the number of occurrences per quadrat to vary freely. In particular, we computed, for each candidate pair, a P-value as the fraction of randomised assemblages where the target pair had a co-occurrence equal or higher than that observed in the real assemblage. Finally, we generated a network by collating all the pairs where P < 0.05. We replicated the procedure individually for samples collected either in the first (2014) or second (2015) year of the survey, hence obtaining two distinct networks (Fig. 4A, B). These two networks can optionally be combined into a multitemporal network, thus allowing separation between species associations present during both years of sampling, and those present in an individual year only (Fig. 4C).

(2) Exploring ecological network structure

To illustrate the wider utility of network analysis for distinguishing differences versus similarities in the constellation of biotic associations over time, consider the two networks in Fig. 4A, B. Such networks will ideally represent the same system at two different times. We can look at those networks from a qualitative perspective, focusing on the taxonomic identity of the nodes involved in the links. In this way, the two networks can be compared in a straightforward manner, highlighting common elements and differences in the association set. The core structure of the networks appears stable between years, with more than 30% of the links being shared between the two networks regardless of changes in species composition (Fig. 4C), and all the links preserved when focusing only on species present in both surveys. These results provide support for the idea that the associations between macrophytes and animals indeed might be structured and hence be particularly suitable for network analysis. But what if, instead, we had found profound differences between the link identities? Would that have meant that the two networks were completely different?

This is where the power of network analysis resides. In Fig. 4A, B, we have plotted species so that their horizontal position is kept consistent in the different networks in order to facilitate comparison. Yet, this is the only aspect that hints at species' identity. Other than that, all the species are represented in the network with the same symbol. On one hand, this is a graphical choice: we might have used different silhouettes for the different species, or different node colours, or spelt out species' names. On the other hand, the main value of using networks is that of having an abstract, simplified and yet extremely informative representation of a complex system. In such a representation, the taxonomic identity of nodes becomes irrelevant, while what really matters is how those nodes are arranged and how that arrangement generates global and local structural patterns. In this way, networks offer a convenient framework to obtain objective measures useful to describe and compare natural systems, even when the respective networks are inhabited by taxonomically fully distinct assemblages.

Several metrics have been developed for characterising network structure. The main distinction is between local measures (i.e. properties of a node or link) and global measures (i.e. properties of the whole network). Additionally, there are ‘meso-structure’ indicators, such as core–periphery metrics. It should be noted that there are differences between the structural metrics typical of unipartite networks, such as food webs (Dunne, 2006), and those typical of bipartite networks (Ings et al., 2009) such as plant–pollinator networks and macrophyte–animal networks as we are presenting them here. In the following examples, we will refer mostly to metrics for bipartite networks. To illustrate how they may be used to resolve key features of macrophyte–animal networks, we will next consider different descriptors of specialisation and of the importance of different nodes.

To characterise network structure, both local and global measures can be used in ecological analysis. The most obvious example of a local measure is node degree, which is the number of in- and out-coming links of a target node. This local measure can be used to obtain a global ‘fingerprint’ of the network, which is the probability distribution of node degrees (a network's degree distribution). A typical procedure in network science is plotting the degree distribution with both axes (x = node degree; y = frequency) on a log scale (Fig. 5B). Such analysis can provide an immediate visual representation of specialisation patterns.

The network in Fig. 5A, which we reconstructed from the data obtained in Table 2 from Chemello & Milazzo (2002), is one of the largest networks we were able to assemble. It includes six macrophyte species, 57 invertebrate (mollusc) species and 159 links. As the information reported in the study refers to three replicate samplings, each link in our network represents a potential association identified in all three replicates. As shown in Fig. 5B, the degree distribution of the invertebrate species follows a power law, where most invertebrate nodes tend to have few connections (1 link), while only a few nodes show a relatively high generalism (>5 links).

Less-obvious measures attribute importance to a node based on different criteria which consider broader portions of the network. Ideally one could obtain some measure of node importance and then compare it with other species features to test various hypotheses. For example, one could test whether the position of a node in a network is affected by specific ecological traits and, in turn, how such network position could affect the vulnerability of the target species to indirect effects of environmental change propagating through network links (Strona & Lafferty, 2016; Escribano-Avila et al., 2018).

As an example, we compared the specialisation of associations versus macrophyte biomass in the networks obtained from co-occurrence in the data from Nakamoto et al. (2018). As a general expectation, specialised consumers should rely on ‘stable’ resources (e.g. resources that are broadly distributed, locally abundant and available through time), providing a key to the robustness and stability of ecological networks (Strona & Lafferty, 2016). For this, we find support in the network observed by Nakamoto et al. (2018). In Fig. 6 we show that the more ‘specialised’ invertebrates (that is, those connected to fewer macrophytes in the network) are preferentially associated with macrophytes having the largest overall biomass. In theory, this could make networks more robust to oscillation in resource populations compared to cases where no relationship exists between consumers' specialisation and resource availability (Strona & Lafferty, 2016). Of course, an obvious question is whether the observed pattern reflects the actual specialisation in the animal community or, instead, the observed ‘specialisation’ is an artefact of macrophyte abundances. This is where network analysis becomes important. Although we expect to observe more animal species on the more common/abundant macrophyte species simply due to chance, the simultaneous consideration of the whole network allows us to assess whether the observed degree of specialisation and network structure can be explained by the sampling process alone or whether it results from non-random associations. The consistency between network links across the two sampling seasons, and the consistency of the trajectories observed in Fig. 6, seem to suggest that the pattern is indeed a product of network structure and specialisation of associations, and not a mere result of chance.

The structure of an ecological network can reveal different processes. For example, the tendency for sharing partners in mutualistic networks (nestedness) can promote diversity by attenuating competition for resources (Bastolla et al., 2009). Similarly, the organisation of an ecological network into fairly isolated clusters of highly connected interacting species can have important implications for network persistence/stability (Thébault & Fontaine, 2010). We explored nestedness in the network from Chemello & Milazzo (2002) reported in Fig. 5A (Fig. 7). In particular, we used the popular NODF metric (Almeida-Neto et al., 2008). The metric ranges from 0 to 100, with 100 being the score of a perfectly nested matrix. We compared the NODF value observed in the original network with the NODF values of 1000 null networks. We generated the null networks using an algorithm that randomises the position of animal–macrophyte links in the network without altering the total number of animals associated with a given plant. We then obtained a P-value as the fraction of null networks with a NODF value higher than that of the original matrix, and an effect size (Z) as [NODF_obs–mean(NODF_null)]/std(NODF_null), where NODF_obs is the observed NODF, and NODF_null is the set of NODF values measured in the null networks. We performed the analyses using the freely available NeD software (Strona et al., 2014).

In addition to this analysis, for comparative purposes, we generated the ideally most and least nested networks. These can be obtained by randomly reallocating links in the network with the only constraints being maintaining the initial number of links, and at least one link per species in the network (i.e. avoiding the generation of empty rows and columns in the matrix representing the network). Alternatively, one can generate a totally random matrix (with the same constraints). In Fig. 7 we provide visual representations of the matrices corresponding to the original network (Fig. 7A), the most nested realisation (Fig. 7B), one random realisation (Fig. 7C), and the least nested realisation (Fig. 7D), together with their NODF, Z and P values. The original matrix (Fig. 7A) proved significantly nested (as commonly observed in ecological networks), and much closer to the ideal perfectly nested configuration (Fig. 7B) than to the random one (Fig. 7C), and very far from the least nested possible configuration (Fig. 7D). The presence of a nested structure is consistent with the expectation for ecological networks, because nestedness is an expected by-product of the same mechanisms illustrated above about consumers specialising on dependable resources (Strona et al., 2013). Furthermore, nestedness might be key to the emergence of complexity in multi-species communities, as the tendency to share interacting partners or resources can counteract the negative effects of interspecific competition and promote the coexistence of species (Bascompte & Jordano, 2006). Thus, the observed patterns provide circumstantial evidence confirming that the observed associations (and hence the higher-level structure emerging from those) are not a product of chance. Instead, they might indeed reflect ecological and co-evolutionary processes, although the purely illustrative nature of the example prevents us from hypothesising too far.

(3) Exploring the effect of disturbance on macrophyte–animal network structure

Networks offer a convenient framework to explore how different types of disturbance (e.g. environmental change) can propagate through (and possibly be amplified by) ecological interaction links and, more broadly, affect the structure and stability of the biological systems under study. In this section we provide two examples, one theoretical and one based on empirical data from a perturbation experiment. We then discuss the ecological information attainable by analysing macrophyte–animal network structure under different scenarios of disturbance.

(a) Mapping hypothetical co-extinction scenarios of macrophytes and animals

An informative exercise in ecological network analysis is exploring the robustness of networks to incremental species loss, that is to the subsequent removal of nodes (Memmott, Waser & Price, 2004; Dunne & Williams, 2009). The number of nodes to be removed, as well as the sequence of removal, depends on the specific objectives of the simulation. Similarly, the rules determining the effect of the loss of nodes and links (i.e. species and interactions) on the network can vary on a per case basis, depending on the starting hypotheses, the objectives of the analysis, the kind of network under examination, and the informative content of the network. For example, the rules that can be set for a weighted network differ from those for an unweighted network. The same applies to directed versus undirected networks and unipartite versus bipartite networks. In one of the simplest possible models, one can reiterate the procedure of removing one node at random from a network, and then remove the nodes left with no resources after the first removal, until no nodes are left in the network. In a food web, one can remove species in random sequence and prune the network from the portions left isolated from basal resources after each removal, hence simulating ‘extinction cascades’ (Dunne & Williams, 2009). In a bipartite network, such as a plant–pollinator network, it is appropriate to focus on the response of one of the two groups to the loss of interacting partners from the other group. For instance, a typical simulation consists of progressively removing plants, keeping track of the resulting pollinator diversity (or vice versa) (Memmott et al., 2004). The results can be plotted on a graph showing the decline in pollinator diversity following the removal of plant species (i.e. in a plot showing ‘primary extinctions’ versus ‘secondary extinctions’ or ‘co-extinctions’).

The removal of species could be based on some informed criteria, for example by combining data on species thermal tolerance with projections of future climate change (Strona & Bradshaw, 2018), or could be defined according to arbitrary criteria, such as random node removal. In the latter, the true value of the analysis is that of offering a way to compare the ‘intrinsic’ robustness of a network to disturbances. To demonstrate such an approach, we ran a set of co-extinction simulations (Fig. 8) in the network reconstructed from the data of Chemello & Milazzo (2002) (Fig. 5A), which maps 159 associations between six macroalgae and 57 animal species. In particular, we performed three sets of node-removal simulations (with 100 replicates per set) aimed at approximating worst, random and best scenarios of network collapse following macroalgal loss. For this purpose, we progressively removed macroalgae from the network in random order, or by choosing, at each step, either the most (in the worst-case scenario) or the least (in the best-case scenario) connected macroalgal species. The worst and best-case scenarios were simulated using a ‘greedy’ algorithm, that recalculates the number of connections per node at each step (as in Allesina & Pascual, 2009). After each macroalgal removal, we computed the relative fraction of animal species which were still connected to algae. We finally averaged the results, for each scenario of node removal, across the 100 replicates. For comparison, we replicated the same exercise in one of the most complete (and most famous) plant–pollinator (or better, plant–visitor) networks, mapping 15255 interactions between 456 plant species and 1044 animal species in a prairie–forest transition in Western Illinois, USA (Robertson, 1929) (note that an equivalent co-extinction simulation was conducted in Memmott et al., 2004). Rather than simulating realistic extinctions, the experiments serve to compare the trajectories of different networks, and the scenarios used here (random, best and worst) are also purely hypothetical. Nonetheless, using informed scenarios based on actual species vulnerabilities provides insights into how the primary effects of global change can propagate through network links and eventually cause the loss of species not directly affected by change. For example, we might simulate future changes by removing macrophyte species based on predicted climatic conditions. As an example of the insights gained through such approaches, Strona & Bradshaw (2018) showed how secondary effects can actually be key to mass extinctions.

Although there are clear similarities between the co-extinction trajectories observed in the macrophyte–animal network compared to those observed in the plant–pollinator network, the former appears more fragile (Fig. 8). In other words, the macrophyte–animal network shows a faster decline in animal diversity in response to plant loss than does a plant–pollinator network. As these co-exctinction simulations are based on only two networks and on hypothetical rather than informed scenarios of species loss, the results presented cannot be generalised. However, the consistency in the observed trajectories is encouraging, and urges the exploration of the structure and robustness of macrophyte–animal networks through approaches more commonly used in terrestrial plant–animal networks. Furthermore, the results suggest that interesting differences might exist between the systems. Thus, further investigation could yield insights into the mechanisms possibly controlling differential responses to global change-induced diversity loss among aquatic and terrestrial systems (Antão et al., 2020).

(b) Effects of experimental perturbation on macrophyte–animal networks

To illustrate another application of network analysis in the study of macrophyte–animal systems under disturbance, we use data generated from experimental perturbations of rocky intertidal macroalgal canopies [Ericaria amentacea (C.Agardh) Molinari & Guiry (formerly Cystoseira amentacea Bory var. stricta Montagne)]. These experiments were originally undertaken to investigate early warning indicators of regime shifts in understorey assemblages following colonisation by algal turfs (Benedetti-Cecchi et al., 2015).

The experimental design consisted of two factors: (i) a press perturbation, i.e. the annual clipping of algal canopies (“canopy”), simulating the expected effects of rising temperatures, with four levels (a control and three increasing levels of perturbation); crossed with (ii) a pulse disturbance, i.e. a one-time-removal of invertebrates and algae (“clearing”) in 10 cm × 10 cm patches within the experimental area, simulating the effect of strong storms, with two levels (uncleared and cleared). Each combination of treatment (‘canopy’ × ‘clearing’) levels was replicated five times, and the percentage cover of understorey invertebrates and algae was sampled eight times (approximately annually) in 20 cm × 20 cm quadrats over a 7-year experimental period (2007–2013) [see Benedetti-Cecchi et al., 2015 for further details].

Starting from these experimental data, we generated networks of co-occurrence of understorey invertebrates and turf-forming algae, and then assessed the effect of the two disturbance factors (“canopy” and “clearing”) on network structure and properties. We first determined which species occurred in the same sampling unit more frequently than expected by chance using the package cooccur, based on the probabilistic model of species co-occurrence from Veech (2013). We hence collated all the species pairs showing significant positive co-occurrence (P < 0.05) and generated invertebrate–algae co-occurrence networks under each combination of experimental treatments and sampling dates. We then computed the nestedness of each co-occurrence network as the effect-size (Z) obtained by standardising the observed NODF value (Almeida-Neto et al., 2008) by the NODF values of 1000 null networks (using the procedure described in Section VI.2).

A linear mixed-effect model was fitted using the package nlme (Pinheiro et al., 2007) to test the correlation between network nestedness and the combination of ‘canopy’ and ‘clearing’ treatments (fixed effects), considering the sampling date as a random effect. The two most perturbed groups showed a significant correlation between NODF and treatments (‘turf margin + clearing’, t(23) = 2.42, P < 0.05; ‘turf full + clearing’, t(23) = 2.48, P < 0.05), indicating an interactive effect of canopy clipping and clearing on network nestedness (Fig. 9). The invasion of turf-forming algae in cleared and canopy-depleted plots hence appears to drive significant structural changes in the co-occurrence networks of understorey invertebrates and algae, with respect to unperturbed plots. Disturbance can have significant impacts on the architecture of ecological networks, and, in general, perturbed natural communities appear less structured than in stable environmental conditions (Vanbergen et al., 2014; de Assis Bomfim et al., 2018). Our results reflect those reported by Benedetti-Cecchi et al. (2015), which showed dramatic shifts in understorey assemblages when press perturbation and pulse disturbance are combined.

Hence, on the one hand co-occurrence networks can provide an alternative framework to assess the effect of disturbance on macrophyte–animal assemblages in experimental studies, providing complementary, ecologically relevant information to that obtainable with other analytical approaches. On the other hand, adopting ecological networks in the context of controlled perturbation experiments allows us to establish direct causal links between shifts in species co-occurrences and network properties in response to disturbance.