At the heart of modern biodiversity science are a set of concepts and theories about species richness, stability, and function (Loreau and de Mazancourt 2013). These relate to the abundance, distribution, and services that biodiversity provides, and how biodiversity (as an interconnected set of species) responds to multiple stressors. Documenting interactions between and among species is thus one of the fundamental building blocks of community ecology, providing a powerful abstraction and platform for mathematical and statistical modelling of biodiversity to make predictions, mitigate threats, and manage services (Windsor et al. 2023).Such network representations of biodiversity are increasingly argued to be an asset to understanding and predicting the abundance, distribution, dynamics, and services provided by multiple species facing multiple stressors (Simmons et al. 2021). However, there is a growing discourse around limitations to the interpretation and applied use of networks (Dormann 2023; Blüthgen 2010), primarily as the result of shortcomings regarding their conceptualisation (Blüthgen and Staab 2024).
We propose that every network embeds assumptions about the process(es) that determine interactions, and about the levels of organization at which this occurs (i.e. the biological, ecological, spatial/temporal scale). The differences in these assumptions ultimately influence the nature and scope of inference that can be made from a given network (Proulx, Promislow, and Phillips 2005). Fundamentally, we are talking about an intersection of the data used to construct the network and the underlying theory as to what drives the occurrence of interactions between species. Although there have been extensive discussions about the challenges relating to data collection and observation (e.g., Blüthgen and Staab 2024; Brimacombe et al. 2023, 2024; Moulatlet et al. 2024; Pringle and Hutchinson 2020; Polis 1991; Saberski et al. 2024) we still lack a clear framework framed by the different assumptions and scale dependent processes.
In this perspective we aim to provide an overview of the different food web representations, particularly how these relate to the terminology used to define a food web, and how this is influenced by both the processes that determine interactions Section 2, as well as how this relates to the way in which we construct the resulting networks Section 3. This allows us to deliver an overview of fundamental questions in ecology that we think can benefit from network thinking and a proposal that such thinking can accelerate our capacity to predict the impact of multiple stressors on biodiverse communities. Specifically, we finish this perspective with an overview of fundamental questions in ecology that we think can benefit from network thinking and a proposal that such thinking can accelerate our capacity to predict the impact of change on biodiverse communities.
1 Setting the Scene: The Not So Basics of Nodes and Edges
Networks often have multiple uses: an ‘object’ from which inferences are made (e.g., topological inference about biodiversity, interactions among species, and community structure, [REF]); a platform for evaluating ‘downstream’ responses to stressors [REF]; and a platform for evaluating mathematical and statistical models of ‘generative processes’ [REF]. Against this backdrop of multiple research agendas, it should come as no surprise that the definition of ‘edges’ and ‘nodes’, as well as the levels of organisation at which they are collated takes many forms (Poisot, Stouffer, and Kéfi 2016; Moulatlet et al. 2024), while also encoding a series of assumptions within a network.
1.1 How do we define a node?
Although this may seem elementary that a node should represent a (taxonomic) species, the reality is that nodes often represents non-taxonomic units such as a trophic species (e.g., Yodzis (1982); Williams and Martinez (2000)), a feeding guild (e.g., García-Callejas et al. 2023), or a segregation of species by life stages (e.g., Clegg, Ali, and Beckerman 2018). Such granularity and variation can limit the ability to make (taxonomic) species specific inferences (e.g., does species \(a\) eat species \(b\)?), affect inference made from networks, including estimates of complexity and structure (Beckerman, Petchey, and Warren 2006; Clegg, Ali, and Beckerman 2018) and make it challenging to use networks in ‘downstream analyses’, for example, of extinction or invasions. Despite these implications, there may also be value in having nodes that represent an aggregation of species, as the distribution of the links between them may be more meaningful in terms of understanding energy flow and distribution within the system.
1.2 What is captured by an edge?
Links within food webs can be thought of as a representation of either feeding links between species (be that realised or potential (Dunne 2006; Pringle 2020), or fluxes within a system e.g., energy transfer or material flow as the result of the feeding links between species (Lindeman 1942; Proulx, Promislow, and Phillips 2005). These correspond to different ‘currencies’ (the feasibility of links or the energy that is moving between nodes). There are also a myriad of ways in which the links themselves can be specified. Links between species can be treated present or absent (i.e., binary), may be defined as probabilities (Banville et al. 2024; Poisot et al. 2016) or by continuous functions which further quantify the strength of an interaction (Berlow et al. 2004). How links are specified will influence the structure of the network. For example, taking a food web that consists of links representing all potential feeding links in a collection of species will be meaningless if one is interested in understanding the flow of energy through the network as the links are not environmentally/energetically constrained.
1.3 Network representations
Against these definitions, networks fall into two major ‘types’: metawebs, traditionally defined as all the potential interactions for a specific species pool (Dunne 2006); and realised networks, which is the subset of interactions in a metaweb that are realised for a specific community at a given time and place. The fundamental differences between these two network representations are the spatial scale at which they are constructed and the associated processes that are assumed to drive pattern at these scales.
A metaweb is at its core a list of feasible interactions between pairs of species. The feasibility for a given pair is derived from the complementarity (phylogenetic relationships) of their traits (representing a global metaweb), which can be further refined by co-occurrence (representing a regional metaweb). By this definition, metawebs provide a means to identify links that are not ecologically plausible, i.e., forbidden links (Jordano 2016b), or provide an idea of the complete diet of a species (Strydom et al. 2023).
In contrast realised networks are relatively localised in space and time, and the links between species are contingent on both the co-occurrence of species, the role of the environment, and mechanisms of diet choice. Fundamentally this means that the presence/absence of a link is the result of the ‘behaviour’ of the species.
This distinction between metawebs and realised webs lead to some further definitions. Links that are absent in a metaweb can conceptually (although not always practically) be treated as being truly absent. However, links that are absent in a realised network cannot be considered as truly absent but rather as absent due to the broader environmental/community context. Furthermore, a realised network is not simply the downscaling of a metaweb to a smaller scale (e.g., moving from the country to the 1x1 km2 scale based on fine-scale species co-occurrence). Instead, realised webs capture processes that determine the realisation of an interaction. Specifically, in realised webs, the definition of an edges shifts from being determined by feasibility to that of choices and consequences that centre around energy. Meaning if one were to take the same community of species and constructed both a metaweb and realised network the two networks might have the same species but would be structurally different, owing to the differences in the ‘rules’ constraining the presence of links (Caron et al. 2024).
2 From Nodes and Edges to Process and Constraints
In the previous section we discussed how the definition of nodes and edges at representing different biological and ecological and processes associated with them lead to the concept of a metweb and a realised web. Here we expand this discussion, introducing five core constraints across these scales that further expose processes that determine the links among species: evolutionary compatibility, co-occurrence, abundance, diet choice, and non-trophic interactions Figure 1.
2.1 Processes that determine the feasibility of an interaction
Here we introduce evolutionary compatibility and co-occurrence as processes that ‘act’ at the species pair of interest, that is the possibility of an interaction being present/absent is assessed at the pairwise level.
Here we introduce evolutionary compatibility and co-occurrence as processes that ‘act’ at the species pair of interest. The scale of inference and set of processes embodied in these two constraints combine to define a ‘list’ of interactions that are viable/feasible and defined as present/absent. It is however possible to build a network from this information. However, it is important to be aware that the structure of this network is not constrained by any community context and so just because species are able to interact does not mean that they will (Poisot, Stouffer, and Gravel 2015).
Evolutionary compatibility
This constraint is defined by shared (co)evolutionary history between consumers and resources (Segar et al. 2020; Gómez, Verdú, and Perfectti 2010; Dalla Riva and Stouffer 2016) which, in the more proximal sense, is manifested as ‘trait complementarity’ between two species (Benadi et al. 2022). In this body of theory, one species (the consumer) has the ‘correct’ (multivariate) set of traits that allow it to chase, capture, kill, and consume the other species (the resource) and interactions that are not compatible are defined as forbidden links (Jordano 2016b); i.e., not physically possible and will always be absent within a network.
Networks arising from this constraint can be binary (possible vs forbidden) or probabilistic (Banville et al. 2024), e.g., the metaweb constructed by Strydom et al. (2022) uses probabilities to quantify their confidence with regards to the possibility of a specific interaction existing between two species. A network constructed on the basis of evolutionary compatibility is conceptually aligned with a ‘global metaweb’, and gives us information as to the feasibility of links between species despite the fact that they do not co-occur (as shown in Figure 1).
(Co)occurrence
The co-occurrence of species in both time and space is a fundamental requirement for an interaction between two species to occur (at least in terms of feeding links). Although co-occurrence data alone is insufficient for building an accurate and ecologically meaningful representation of feeding links (Blanchet, Cazelles, and Gravel 2020), it is still a critical process that determines the realisation of a feeding link and allows us to spatially constrain a global metaweb based on local communities (Dansereau, Barros, and Poisot 2024), in the context of Figure 1 this would be the metawebs for regions one and two.
2.2 Processes that modify the behaviour (preference) of species
Here we will showcase three processes that will ultimately influence the realisation of an interaction between species and form the conceptual basis for realised networks. As we show in Figure 1 a ‘truly realised’ network is the product of different facets of both the properties of the community (abundance and non-trophic interactions) as well as the individual (profitability). This represents a conceptual shift where the presence (realisation) of an interaction is no longer constrained to evaluating the viability between a pair of species but rather takes into consideration information about the community and the individual (Quintero et al. 2024), and as discussed in Section 1.3, links are now constrained by consumer choice.
Abundance
The most basic abundance constraint linked to foraging biology is the principle that organisms feeding randomly will consume resources in proportion to their abundance (Stephens and Krebs 1986), and interactions are not necessarily contingent on there being any compatibility between them (E. Canard et al. 2012; Momal, Robin, and Ambroise 2020; Pomeranz et al. 2019). Alternatively the abundance of different prey species will influence the distribution of links in a network (Vázquez et al. 2009), by influencing which prey are targeted or preferred by the predator, as abundance influences factors such as the likelihood of two species (individuals) meeting (Poisot, Stouffer, and Gravel 2015; Banville et al. 2024). Thus, if abundance data are combined with a derived metaweb, there is a basic ruleset that can define the distribution (e.g., structure) and potentially the strength of links.
Profitability
It is well established that consumers make more active decisions than eating items in proportion to their abundance (Stephens and Krebs 1986). Ultimately, consumer choice is underpinned by an energetic cost-benefit framework centred around profitability and defined by traits associated with finding, catching, killing, and consuming a resource (Wootton et al. 2023). Although energetic constraints can be invoked in a myriad of ways (e.g., Pawar, Dell, and Savage 2012; Portalier et al. 2019; Cherif et al. 2024) we select profitability as a term to capture rules linked to optimal foraging (Pyke 1984) and metabolic theory (Brown et al. 2004); it is a sensible ‘umbrella concept’ for capturing the energetic constraint on of the distribution and strength of interactions.
Non-trophic interactions
Perhaps not as intuitive when thinking about the previous constraints, non-trophic interactions (Ings et al. 2009) specifically modify either the realisation or strength of trophic interactions (Golubski and Abrams 2011; Pilosof et al. 2017; Staniczenko et al. 2010; Kamaru et al. 2024). Non-trophic interactions can modify interactions either ‘directly’ e.g., predator a outcompetes predator b; or ‘indirectly’ e.g., mutualistic/facilitative interactions. Altogether they can alter the fine-scale distribution and abundance of species as well as their persistence (Kéfi et al. 2012, 2015; Buche, Bartomeus, and Godoy 2024).
3 Network construction: a case for models
3.1 Why construct networks?
Broadly the desire to construct a network has arisen for two different purposes; building networks that can be used in real-world, applied contexts (have actionable consequences?), and building networks that allow us to interrogate, generate, and reflect upon different ecological theories. The act of constructing a ‘real world’ network through the empirical collection of interaction data is both costly and challenging to execute (Jordano 2016a, 2016b), which has led to the development of a suite of approaches that allow us to predict the interaction between two species, or network structure (see Strydom et al. 2021 for a broader discussion), or identify missing interactions (gap fill) within existing empirical datasets (e.g., Biton, Puzis, and Pilosof 2024; Stock 2021; Dallas, Park, and Drake 2017). However, working with ‘real-world networks’ is data-hungry and cumbersome, and has driven the development of models that construct ecologically plausible networks. These models often explicitly model one or a few of of the processes discussed in Section 2 and in doing so allow us to better understand the different constraints determining interactions (Stouffer 2019; Song and Levine 2024).
3.2 Construction through induction
Tools developed in the context of constructing networks allow a user to take a collection of species and determine wht the interactions between them could be. Being able to predict a network is useful for determining all feasible interactions for a specific community, and the tools that have been developed in this context have the potential to allow us to construct first draft networks for communities for which we have no interaction data (Strydom et al. 2022). Making them valuable for interpolation in data poor regions and predicting interactions for ‘unobservable’ communities e.g., prehistoric networks (Yeakel et al. 2014; Fricke et al. 2022; Dunhill et al. 2024) or future, novel community assemblages. Additionally, an understanding of the role of interactions between species has allowed us to better determine the distribution of a species by accounting not only for the role of the environment but also the role of species interactions (Higino et al. 2023; Pollock et al. 2014).
Owing to the intense amount of data one would need at the community level to make predictions about the realisation of networks Section 2.2 the tools that predict interactions typically only asses the feasibility of interactions and typically focus on capturing some pairwise assessment of the likelihood of an interaction being present between two species. Resting on the assumption that there are a set of ‘feeding rules’ that can be used to make this assessment (Morales-Castilla et al. 2015). The determination of these feeding rules is typically done in a few ways ways, each with their own constraints and assumptions. Rules can be defined a priori based expert knowledge opinions, typically this is done on a trait-based basis e.g., the paleo food web model (Shaw et al. 2024) specifies a series of rules for four different sets of traits and interactions are deemed feasible of all conditions are met. Alternatively rules can be elucidated by correlating real world interaction data with a suitable ecological proxy for which data is more widely available (e.g., traits). These rules can be used by a binary classifier to determine if a link is present (see Pichler et al. (2020) for an overview), including generalised linear models (e.g., Caron et al. 2022), random forest (e.g., Llewelyn et al. 2023), trait-based k-NN (e.g., Desjardins-Proulx et al. 2017), and Bayesian models (e.g., Eklöf, Tang, and Allesina 2013; Cirtwill et al. 2019). Finally, graph embedding uses the structural features of a known network to infer the position of species in the network (see Strydom et al. (2022) for a detailed review of methods).
Data implications for these approaches are that they require good datasets from which we can infer the rules, but because they contain real world species it does make it easier to validate them…
3.3 Construction through deduction
(I don’t know how to phrase this better.) As opposed to inferring interactions from known interaction, models are typically more formalised and deduced from a body of theory. SOmething about synthetic networks??
3.3.1 Species agnostic networks
These models define networks via an assumption that the interactions between species occurs irrespective of the identity of the species (i.e., species have no agency). Here there some assumption as to the expected structure of a network i.e., the links between the nodes and how they might be distributed, typically constrained by connectance. There are three broad group of models based on some assumption.
First, links are randomly distributed throughout the network (e.g., Fortuna and Bascompte 2006; Bascompte et al. 2003), these models are often used as a ‘null hypothesis’ to ask questions about network structure (e.g., Banville, Gravel, and Poisot 2023; Strydom, Dalla Riva, and Poisot 2021).
Second. Interactions that occur between species are due to the abundance of species within the community (Pomeranz et al. 2019; E. F. Canard et al. 2014; Krishna et al. 2008)
Third. Based on the idea that networks follow a trophic hierarchy and that network structure can be determined by distributing interactions along single dimension (the “niche axis”, Allesina, Alonso, and Pascual (2008)). Essentially these models can be viewed as being based on the idea of resource partitioning (niches) along a one-dimensional resource which will result in the standard ‘trophic pyramid’ to ensure that all species can ‘fit’ along this resource (which has strong ties back to the idea of intervality) e.g., Cascade model (Cohen, Briand, and Newman 1990), Niche model (Williams and Martinez 2000), Nested hierarchy model (Cattin et al. 2004).
These models are data light but there are some decisions that need to be made regarding what the expectations are on network structure.
3.3.2 Species-specific networks
In terms predicting interactions current models are rooted in feeding theory and allocate the links between species based on energy e.g., diet models (Beckerman, Petchey, and Warren 2006; Petchey et al. 2008) have been used construct networks based on both profitability (as determined by the handling time, energy content, and predator attack rate) as well as abundance (prey density). (Wootton et al. 2023).
At a ‘coarser’, functional level there are models that are based on the compartmentation and acquisition of energy for species at different trophic levels (Allesina and Pascual 2009; Krause et al. 2003). Models that determine structure are based on the idea that networks follow a trophic hierarchy and that network structure can be determined by distributing interactions along single dimension [the “niche axis”; Allesina, Alonso, and Pascual (2008)], while parametrising an aspect of the network structure (although see Allesina and Pascual 2009 for a parameter-free model).
They are ‘costly’ to construct in real world settings (requiring data about the entire community, as it is the behaviour of the system that determines the behaviour of the part) and also lack the larger diet niche context afforded by metawebs.
4 Making Progress with Networks
It is probably both this nuance as well as a lack of clear boundaries and guidelines as to the links between network form and function (although see Delmas et al. 2019) that has stifled the ‘productive use’ of networks beyond the inventorying the interactions between species. Although progress with using networks as a means to address questions within larger bodies of ecological theory e.g., invasion biology (Hui and Richardson 2019) and co-existence theory (García-Callejas et al. 2023) has been made we still lack explicit guidelines as to what the appropriate network representation for the task at hand would be, and as highlighted in Box 1, underscores the need to evaluate exactly what process a specific network representation captures as well as its suitability for the question of interest. Below we present a mapping of what we believe are some of the key questions for which interaction networks can be used to the different networks representations that are most suitable, as well as highlight some of the methodological challenges that still need to be improved upon.
4.1 Making use of the different network representations
Methodological challenges
- Tools that allow us to estimate both the feasibility as well as realisation of links: Currently most approaches to modelling realised networks fail to explicitly account for any form of evolutionary constraint Wootton et al. (2023) and we need to develop either an ensemble modelling approach (Becker et al. 2022; Terry and Lewis 2020) or tools that will allow for the downsampling of metawebs into realised networks (e.g., Roopnarine 2006).
- Is there something in generalisable models that ‘combine’ different processes/aspects (e.g., using body size as a catch all) versus limited models that allow you to unpack things bit-by-bit (i.e., process by process). So Wootton et al. (2023) may (TBD) span the gamut but it lacks the ability to unpack… Although myabe the terms do?
- Modelling interaction strength: Although realised networks are more closely aligned with explicitly capturing interaction strength we lack models that allow us to quantify this (Wells and O’Hara 2013; Strydom et al. 2021).
- How do we validate our predictions?: Progress has been made to assess how well a model recovers pairwise interactions (Strydom et al. 2021; Poisot 2023), but we still lack clear set of guidelines for benchmarking the ability of models to recover structure (Allesina, Alonso, and Pascual 2008)
- Something about making what we do with networks more tractablie in the applied space? e.g., Dansereau et al. (2024)
Theory challenges
- Core Theory Advancement: Do the decades of insights arrived at for stability-diversity-productivity relationships with tri-trophic or diamond shaped models hold for complex communities (10’s-100s) (Danet et al. 2024); How will spatial and temporal variation in climate and productivity drive change in complex ecosystems. Necessary to move to predicting changes in biodiversity per se, ecosystem functions and identifying sensitive and robust species and portions of communities.
- How will novel communities interact? How will range shifts and invasions result in new/novel community assemblages. And then also the intentional changes of species compositions through rewilding.
- Does rewiring happen and does it deliver robustness? Specific sub points to consider here is persistence, especially persistence to perturbation. Again, dynamic networks and network/community assembly and finally extinctions (Dunhill et al. 2024).
- When do invasive species enhance or decimate communities? When do reintroductions work? (Wooster et al. 2024)
- Are there temperature threshold to community collapse
- Can socioeconomic networks combined with biological networks drive understanding of externalities?
- Can paleoecological data from deep time hyperthermal events provide sufficient insight into the targets, pace and recovery times from rapid climate events?
5 Concluding remarks
Having a clear understanding of the interplay between network representations and the processes that they are capable of encoding is critical if we are to understand exactly which networks can be used to answer which questions. As we highlight in Box 1 the different network representations have different potential uses and it should be clear that there is no ‘best’ network representation but rather a network representation that is best suited to its intended purpose. In providing a formalisation regards to the assumptions and mechanisms that need to be explicitly taken into consideration when deciding to use (and construct) networks we hope to prevent the unintentional misuse or misinterpretation of networks as well as provide a starting point from which we can develop a better framework for the applied use of networks to answer questions that are not only pressing within the field but also within broader biodiversity science.