Using individual-based trait frequency distributions to forecast plant-pollinator network responses to environmental change

Determining how and why organisms interact is fundamental to understanding ecosystem responses to future environmental change. To assess the impact on plant-pollinator interactions, recent studies have examined how the effects of environmental change on individual interactions accumulate to generate species-level responses. Here, we review recent developments in using plant-pollinator networks of interacting individuals along with their functional traits, where individuals are nested within species nodes. We highlight how these individual-level, trait-based networks connect intraspecific trait variation (as frequency distributions of multiple traits) with dynamic responses within plant-pollinator communities. This approach can better explain interaction plasticity, and changes to interaction probabilities and network structure over spatiotemporal or other environmental gradients. We argue that only through appreciating such trait-based interaction plasticity can we accurately forecast the potential vulnerability of interactions to future environmental change. We follow this with general guidance on how future studies can collect and analyse high-resolution interaction and trait data, with the hope of improving predictions of future plant-pollinator network responses for targeted and effective conservation.

pressure from a variety of factors associated with high rates of anthropogenic environmental change over the past century (Arce et al., 2023).These include agricultural intensification (Hemberger et al., 2021;Outhwaite et al., 2022), pesticides (Brittain et al., 2010;Kenna et al., 2023), climate change (Kerr et al., 2015;Settele et al., 2016) and emerging diseases (Fürst et al., 2014;Graystock et al., 2020).It is therefore a research priority to develop approaches that study responses of whole communities of interacting species in a single analytical framework.In recent decades, this area has benefited from the application of graph theory, which allows quantification of the 'roles' that taxa play and the rewiring of network architecture along environmental gradients and in response to perturbations (Box 1; Alarcón et al., 2008;Bascompte et al., 2019;Bramon Mora et al., 2020;CaraDonna et al., 2017;Carstensen et al., 2016;Dalsgaard et al., 2013;Poisot et al., 2012).Further, these network properties have been shown to impact network stability and robustness to perturbations (Bastolla et al., 2009;Jiang et al., 2018;Lever et al., 2014).In these networks, interacting taxa have traditionally been represented as species-level nodes, by aggregating individual-level interactions.However, the resolution to which interacting taxa are incorporated in networks may dramatically change how we estimate interaction probabilities (Guimarães Jr, 2020).Aggregating interactions to species level may obscure trait-based mechanisms of interactions, by overlooking the importance of trait variation among conspecific individuals for underpinning interaction preferences (Liang et al., 2021).Untangling the role of intraspecific variation within plant-pollinator networks is therefore an urgent area of investigation, especially given that accurate representations of interactions are needed to forecast whether a plant-pollinator network is structurally stable (Rohr et al., 2014) or estimate the probability that species will persist under environmental change (Saavedra et al., 2020).
In this paper, we discuss recent research developments that consider intraspecific trait variation (how

BOX 1 Glossary
• Feasibility = A range of conditions in which all modelled species 'have constant and positive abundances' (Rohr et al., 2014).• Functional traits = For this paper, any morphological, behavioural or phenological characteristic of an organism that determines fitness by affecting growth, reproduction or survival (Violle et al., 2007).• Individual-level network = A network where interactions are represented as occurring between individuals (not species), for at least one trophic level in plant-pollinator networks (Guimarães Jr, 2020).• Interaction plasticity = Context dependency by which organisms vary with whom they interact, such as under different environmental conditions (Baracchi, 2019;Mery & Burns, 2010).Here, we refer to plasticity in realised interactions, rather than newly evolved interactions.• Modularity = A network pattern consisting of densely connected groups of nodes with sparse connections to nodes in other groups (Beckett, 2016).A module refers to one of these groups.• Network architecture = Specific arrangement of plant-pollinator interactions in a given network.
• Plant-pollinator interaction network = A method of representing plant-pollinator interactions, where taxa (typically, species) are represented as nodes, and links between nodes indicate a pollinator visiting a plant (or a pollinator making contact with a plant's reproductive structures; Petanidou & Potts, 2006; but see Ballantyne et al., 2015).Links or interactions can be binary (indicated as being either one or zero; a pollinator does or does not visit a plant, respectively); weighted (where link thickness typically represents the number of observed pollinator visitations to a plant) or probabilistic (representing the likelihood of a pollinator visiting a plant; Poisot et al., 2016).In this paper, we primarily discuss weighted interaction networks.• Trait filtering = When environmental conditions restrict the occupancy of individuals based on their functional trait(s) (Diaz et al., 1998;Keddy, 1992).• Trait frequency distribution = The distribution of standardised trait values measured from multiple individuals, which can be used to show how the trait varies within (intraspecific) and between (interspecific) taxonomic groups.• Trait matching = When individuals possess congruent characteristics (such as having compatible morphologies or large phenological overlap) that increase the probability of an interaction (e.g.Stang et al., 2009).• Trait space = For the purpose of this paper, a multidimensional representation of functional traits.An individual's 'position in trait space' is defined by the specific combination of its trait values.• Unipartite network projection = When a bipartite network is reprojected such that there is only one trophic level represented as nodes.Links between nodes then depict how partners from the absent trophic level are shared.phenotypes vary between individuals within a species), how trait distributions change over environmental gradients (space and time, or in response to changes in environmental conditions) and how they combine to shape plant-pollinator interactions.We illustrate how these approaches can reveal and better account for high degrees of interaction plasticity over differing spatiotemporal scales and other environmental gradients.We posit that harnessing this approach will be important for developing predictive frameworks to focus on why interactions vary and rewire, and the underlying mechanisms (Barbour & Gibert, 2021;Bascompte & Scheffer, 2023;Ramos-Jiliberto et al., 2012), and improve predictions of how plant-pollinator networks will respond to environmental change for targeted conservation.

Why intraspecific trait variation is important for understanding interaction probabilities
When representing plant-pollinator interactions within large communities, observations of individual interactions are typically aggregated into their respective species node in the network.This is intuitive and sensible, but does not consider the variation in traits that compose the node, and so may inadvertently overlook the trait-based mechanisms that drive interactions (Arroyo-Correa et al., 2021;Rumeu et al., 2018).Indeed, how the composition of traits (here, phenology, morphology and behaviour) within populations is shaped (e.g.Bjorkman et al., 2018a for tree morphology; Burkle et al., 2019 for plant and pollinator morphology; Duchenne et al., 2020 for pollinator phenology) can drive interactions between mutualists, with loss of trait diversity potentially leading to loss of interactions (Baruah, 2022).To illustrate this, when considering a pollinator species node, there is likely variation in individual floral preferences occurring within a population (Arroyo-Correa et al., 2023;Tur et al., 2014; see Barbour et al., 2016 for a food-web example;Crestani et al., 2019 andGuerra et al., 2017 for frugivory examples; Figure 1).For example, suppose pollinator species A forages equally on two plant species: Y and Z.When all individuals within a species are represented as a single node, it is unclear whether individuals within A split their time equally between Y and Z, or whether there is specialisation within A, in which half the individuals exclusively visit Y and the other half exclusively visit Z (Figure 1).If the latter, the mechanisms driving these divergent preferences between individuals of the same species are obscured by aggregating these interactions into a species-level node.Individuals acting differently to the species-averaged trend (i.e.individual niche variation) is a widely observed phenomenon in ecology (see Bolnick et al., 2011).Models describing such individual variation have been developed from optimal foraging theory (Pulliam, 1974;Schoener, 1971;Svanbäck & Bolnick, 2005), tested empirically (Costa-Pereira et al., 2019;Fontaine et al., 2008;Tinker et al., 2012), and integrated with models of population dynamics to predict species' responses to perturbations (such as the consumer-resource model of population dynamics with adaptive foraging in Valdovinos et al., 2018).More recently, we have also seen the development of individual-individual (Rumeu et al., 2018) or individual-species networks (where one trophic level is resolved to individuals, and the other trophic level to species; Arroyo-Correa et al., 2021;Dupont et al., 2014;Tur et al., 2014; Box 2; Figure 2).These higher resolution networks have tended to reveal that individuals within species can range from 'specialist' (low degree) to 'generalist' (high degree), even within relatively generalized species (Arroyo-Correa et al., 2023;Tur et al., 2014).Moreover, individuals in close spatial proximity also tend to interact more similarly or within the same module (Arroyo-Correa et al., 2021;Dupont et al., 2014, respectively).Therefore, only by incorporating intrapopulation differences in resource use and individual specialisation into plant-pollinator networks can we begin to reveal the mechanisms underpinning a given species' set of interactions (Cantwell-Jones et al., 2023;Cope et al., 2022; Figure 1).
One benefit of individual-level networks is that they allow us to investigate the scale at which specific mechanisms-spatial distribution, phenology, morphological traits, abundance and competition-become important.For example, Dupont et al. (2014) showed that on a scale of just 50 × 50 m, bumblebee individuals partitioned resources spatially such that different network modules corresponded to different locations within the survey plot (also see Arroyo-Correa et al., 2021).Similar results have been found for the temporal partitioning of interactions (Valverde et al., 2016).Moreover, by mapping traits at the individual level across networks, we can discern how the functional outcomes of interactions are linked to specific phenotypic or even genotypic differences (Arroyo-Correa et al., 2021;Cantwell-Jones et al., 2023;Kuppler et al., 2016;and Barbour et al., 2016 for a food web example).We can even group individuals by similarity in traits rather than species identity, to see how traits underpin network architecture (Rumeu et al., 2018) and how this changes over environmental gradients (Cantwell-Jones et al., 2023).Finally, individual-level networks are starting to reveal the long-term and evolutionary consequences of interactions.For example, an individual's position or role within a network can be linked to their fitness, with more central individuals having greater fitness in flower-visitor (Gómez & Perfectti, 2012) and frugivory networks (Crestani et al., 2019).Additionally, intraspecific variation in traits and interactions has been linked to dietary diversification over evolutionary timeframes (Pfaender et al., 2011).
At a community level, it is still not well formalised how quantifying intraspecific variation can improve our understanding of emergent community-level network properties.A growing body of empirical research, however, has shown that intraspecific variation in traits and interactions impacts networks by increasing complexity (Barbour et al., 2016), robustness (Cantwell-Jones et al., 2023) and feasibility (Arroyo-Correa et al., 2023).In food webs and flower-visitor networks, the abundance and diversity of nodes for one trophic level are associated with the genetic (and phenotypic) diversity of nodes on another trophic level (Barbour et al., 2016;Crutsinger et al., 2006;Genung et al., 2010).With plantpollinator interactions, Arroyo-Correa et al. (2023) observed that intraspecific variation in a pollinator attraction trait (mean number of flowers) shaped the plant species-level degree, driving community-level network nestedness and feasibility.These empirical findings support previous theoretical research (primarily on food webs) in which processes acting on the individual level (or other levels of biological organisation) scale to shape networks at higher levels (Bolnick et al., 2011;Lemos-Costa et al., 2016;Valdovinos et al., 2010Valdovinos et al., , 2013;; see Montiglio et al., 2020).For example, modelled food webs that include adaptive changes to individual feeding (adaptive trophic behaviour) tend to have more stable dynamics, be more resilient to perturbations and have more complex structures (Valdovinos et al., 2010(Valdovinos et al., , 2013)).In contrast, when simulating the impact of intraspecific variation on empirical mutualistic networks, Baruah (2022) found that networks with high intraspecific variation were more likely to experience abrupt collapses, with tipping points occurring earlier, despite tending to have higher feasibility.Overall, disentangling how intraspecific variation in traits and interactions contributes to network functioning and eco-evolutionary dynamics has never been more important (Barbour et al., 2020;Barbour & Gibert, 2021;Melián et al., 2011), given the rapid (and likely underestimated) rate at which intraspecific variation is currently being lost (Des Roches et al., 2021).Emerging techniques, such as the use of multilayer networks and agent-based modelling (see "Future directions" section), may help bridge this research gap.
F I G U R E 1 Mapping intraspecific variation in traits can reveal potentially hidden variation within species interactions, which enables a better understanding of how functional traits underpin interaction probabilities and network structure.(a) Hypothetical individual (pollinator)-species (plant) system with one pollinator species (in orange) and two plant species (in blue and grey) with equally weighted links, as seen through a more traditional lens.Three scenarios (b-d) could be underlying the species-level network in panel (a).However, without explicitly associating an individual's trait value with that individual's interaction, we would not be able to discern which scenario is occurring.Scenario (b): Individuals within the pollinator species do not differ in their floral preferences and visit both plant species equally.Scenario (c): Individuals within a pollinator species prefer different plant species, but body size is not driving this.Scenario (d): Individuals within a pollinator species prefer different plant species, potentially driven by body size.In this figure, to illustrate the concept, we only consider variation in the pollinator node, but intraspecific variation (in traits or interactions) in the plant nodes also likely contributes to structuring interactions.

How might we incorporate intraspecific trait variation to improve mechanistic predictions of plant-pollinator interactions?
For an interaction to occur, the activities of species need to overlap in space and time (i.e.their spatial distribution and phenology; e.g.CaraDonna et al., 2017;Olesen et al., 2011;Schwarz et al., 2020Schwarz et al., , 2021;;Simanonok & Burkle, 2014), and individuals need to be attractive and morphologically compatible with each other (e.g. a plant with a long corolla (nectar tube) being primarily visited by long-tongued pollinators; Klumpers et al., 2019;Stang et al., 2009).However, spatiotemporal overlap of traits can be influenced by population size and how the morphological traits are distributed across individuals (intraspecific variation).Indeed, individuals within species are rarely all active (emerge or flower) on the same day, are rarely distributed homogeneously in space or have the exact same morphological traits.Instead, these individual-level traits (in the following examples, body size) exist along a continuum (i.e. a trait frequency distribution) that can change over seasonal or spatial scales (Cantwell-Jones et al., 2023), be filtered by the environment (Classen et al., 2017) or shaped by rearing conditions (Chole et al., 2019).Therefore, most traits will show intraspecific variation, though the degree of variation will differ between species (e.g.Rixen et al., 2022).Although the importance of intraspecific variation in mediating network interactions has been appreciated for over a decade (Bolnick et al., 2011), it has only recently started to be explicitly incorporated in studies of plantpollinator networks, for example using exponential random graph models (e.g.Arroyo-Correa et al., 2021; see 'Future directions'), structural equation modelling (e.g.Arroyo-Correa et al., 2023), regressing individual-level plant and pollinator traits (e.g.Liang et al., 2021) or placing individuals across species into 'functional groups' based on trait similarity (Cantwell-Jones et al., 2023;Rumeu et al., 2018).
Moving forward, to improve predictions of plantpollinator network responses to environmental change, future studies should consider: (1) how traits within species vary over spatiotemporal scales (Classen et al., 2017;Cope et al., 2022;Gómez et al., 2020;Pellissier et al., 2018); (2) the role that intraspecific trait variation plays in determining a species' probability of interacting with a mutualistic partner (Arroyo-Correa et al., 2021;Liang et al., 2021) and (3) how environmental variables shape intraspecific variation within and among populations across species assemblages (Burkle et al., 2019).By better mapping intraspecific trait variation, ecologists can then examine how different biotic or abiotic conditions could change trait overlap between species in the same trophic level (e.g.pollinator-pollinator) or trait compatibility between trophic levels (e.g.plant-pollinator).Such trait-distribution overlap can be estimated, for example by constructing kernel density estimators (Mouillot et al., 2005) and hypervolumes (Mammola, 2019), followed by calculating the intersection between species.This could reveal how levels of competition between species might be affected by environmental change or how interactions might rewire following species loss or population abundance change (Barbour & Gibert, 2021;Ramos-Jiliberto et al., 2012).Although we focus here on plant-pollinator networks, we posit that incorporating information on frequency distributions of traits will improve modelling of any type of inter-and intraspecies interactions (e.g.Barbour et al., 2016;Faticov et al., 2020;Wetzel et al., 2016).For example, individual genotype or phenotype preferences can generate within-species clusters (i.e.modules) of interacting parasitoid and host individuals (Lavandero & Tylianakis, 2013) such that consumer intraspecific trait variation can determine the proportion of its resource species consumed and its potential to rewire to other resource species (Barbour & Gibert, 2021).These modules may also be hypothesised to dynamically buffer the individual-based network in the same way as modules of species-level interactions (Stouffer & Bascompte, 2011).

Moving away from binary and mean trait values
Studies looking to predict interactions based on functional traits have often used binary or mean measures of a species' trait.This includes categorising a species as being seasonally active or inactive or being spatially present or absent; or assigning a mean value for a morphological trait, typically calculated from a small subset of individuals or trait dataset.Yet, these binary and mean trait values mask true intraspecific variation.Species can show relatively high intraspecific variation through genetic variation or plasticity between individuals (e.g.Dai et al., 2017;Szigeti et al., 2020), sexual dimorphism (Ashman, 2000;Okamoto et al., 2013) or various lifehistory groups or stages co-existing (e.g.castes in social insects, or ontogeny more broadly; Nakazawa, 2015).Using phenology as an example, the degree of overlap between interacting species is affected not only by the breadth of their distributions (which has been typically considered) but also by the peaks and shapes (Stemkovski et al., 2023; Figure 3).Therefore, if a plant species' peak flowering coincides with a pollinator species' peak activity, they are more likely to interact than if only the first 5% of the plant species has flowered, given the importance of abundance in determining the probability of two species encountering each other (Bartomeus, 2013;Simmons, Cirtwill, et al., 2019;Vázquez et al., 2007).Similarly, species also vary in abundance throughout their spatial ranges, meaning the likelihood of individuals encountering an interaction partner will also vary throughout the spatial range, even if they tend to co-occur (Poisot et al., 2012(Poisot et al., , 2015)).Moreover, as species with larger intraspecific trait variation often have larger ranges (Sides et al., 2014;Treurnicht et al., 2020), their likelihood of encountering more interaction partners should also increase.This more accurate prediction of interaction likelihoods is key to forecasting how a species' population size might change following a perturbation in the interaction network, such as another species becoming locally extinct or shifting their phenology earlier (Figure 3c).For instance, by combining accurate interaction probabilities derived from trait distributions with population dynamic models (such as Lotka-Volterra type models or dynamic consumer resource models; Valdovinos, 2019), it could be possible to assess the potential future stability of the network and persistence of constituent species following a perturbation (Rohr et al., 2014;Saavedra et al., 2020).
When considering morphological traits, allocating a mean value assumes that the average can accurately represent all individuals within a species, and that overlap

BOX 2 Moving from species-to individual-level networks
To represent interacting taxa, networks (or 'graphs') are often used, in which organisms are represented as nodes, and interactions are represented as links between nodes, which can be weighted to reflect interaction strength.Traditionally, nodes have represented species, in which individual interactions are collated.This form of network representation shows levels of generalism or specialism for each species but effectively assumes similar interaction preferences per individual within a given species.However, individuals within species can vary in their dietary breadth (Tur et al., 2014;Wang et al., 2021), level of 'generalism' (Arroyo-Correa et al., 2023) and interaction plasticity (Araújo et al., 2008 andTinker et al., 2012 for food web examples).As a result, in individual-level networks, individuals within the same module may not necessarily be of the same species (in Tur et al., 2015, this is due to phenology).Thus far, three main types of individual-based networks have been used (Figure 2): (1) individual-individual, where each node represents a single individual, and therefore links can reveal the precise traits or conditions under which that interaction occurred (e.g.Dupont et al., 2014); (2) individual-species or resource, where there are two sets of nodes -one set representing individuals and another representing species or resources (e.g.Arroyo-Correa et al., 2023;Isla et al., 2022;Kuppler et al., 2016;Tur et al., 2014).This allows us to track the specific individual-level trait associated with using a given resource; and (3) niche-overlap networks, where nodes are individuals (of the same or different species), and links are the relative overlap in resource use between individuals (such as the unipartite projection in Arroyo-Correa et al., 2021;Gómez et al., 2011).Although the study of individual-level networks is still in its infancy, these highly resolved networks allow the opportunity to investigate mechanistically how patterns of resource use at the individual level scale up to drive the trends observed at the level of species, or even communities (Guimarães Jr, 2020).between species is small.While for some study systems this assumption may apply, it often does not hold for plants and pollinators.For example, Gentile et al. (2021) observed that the body size range in Arctic fritillary butterflies (Boloria chariclea) overlapped with the mean body sizes of approximately 30% of butterfly species in Canada.Similarly, individuals of bumblebee species can be more similar in body size to individuals of different bumblebee species than to conspecifics (Cantwell-Jones et al., 2023).Therefore, when testing the role of functional traits in underpinning interactions, incorporating information on the shape of a species' distribution of functional traits (such as height, variance, skewness and kurtosis) is vital to model how trait overlap between plants and pollinators predict their interaction likelihood (Figure 4a).For example, Bartomeus et al. (2016) modelled the probability of an interaction between pairs of species given their traits and abundances.Their model includes parameters accounting for the probability of observing specific traits in the regional pool, and they provide an example of the framework being used on both a predator-prey dataset and individual plant-pollinator interactions.While this model performed well in the predator-prey dataset, challenges remain in improving model predictive power for plant-pollinator interactions (Bartomeus et al., 2016; but see Eklöf et al., 2013).Modelling the entire trait distributions of interacting species (i.e.beyond mean-trait values) will not only enable better quantification of trait compatibility and thus interaction probabilities between plants and pollinators (for example, by assessing the proportion of the pollinator population with tongues long enough to reach a range of nectar-tube depths), but may also improve estimates of overall network or ecosystem functioning (Gross et al., 2017).

Trait change in space and time, and under different environmental conditions, could decouple interaction partners
When considering the assumption of morphological compatibility between interaction partners, meantrait values are further limited for species in which F I G U R E 3 By considering the temporal distribution of plant and pollinator activity (phenology), we may be able to better predict plantpollinator network changes over time.(a) Activity distributions of two plant species (blue and grey) and one pollinator species (orange) are shown as either binary (absence/presence, shown by the filled timelines at the top) or as a distribution (the dynamic number of individuals observed to be active/flowering over time, at the bottom).(b) Hypothetically, if the interactions between the plants and pollinators were observed at two discrete points during the season (T1 and T2) then the predicted interaction strengths (that is, predicted number of interactions between species; link thickness) would appear consistent between T1 and T2 when considering binary values of phenology.In contrast, when frequency distributions of activity are considered, the predicted interaction strengths would vary through the season, depending on the abundance of the respective plant and pollinator species at each time point.This highlights the importance of understanding the temporal overlap of trait frequency distributions to understand interaction probabilities, and that extrapolation across a season based on a small number of collection time stamps could provide erroneous predictions about interaction strengths.(c) Hypothetical impact on pollinator population size ('pop.size') of each plant species going extinct, based on the predicted interaction strengths of panel (b).As binary measures of overlap lead to potential over-or underestimation of interaction strengths between plant and pollinator species, estimates of how species respond to perturbations (here, extinction of a plant species and the resulting impact on the pollinator population) will likely be erroneous.In the top part of panel (c), this is shown by pollinator population size decreasing equally under extinction of either plant species (where 'pre' refers to pollinator population size before flower species extinction, and 'post', to after flower species extinction).In contrast, with accurate estimates of interaction strengths from using trait frequency distributions, more realistic estimates of the pollinator species' sensitivity to plant extinction(s) are expected (lower part of panel (c)).
morphology at the population-level changes over time.An example of this is when earlier and later-emerging generations, sexes or castes of a species differ in their functional traits (Cantwell-Jones et al., 2023;Fric & Konvicka, 2002;Steward et al., 2022).The Mediterranean plant Moricandia arvensis (Brassicaceae) illustrates this well, with individuals exhibiting a seasonal change from producing large, lilac flowers in spring to small, white flowers in summer, following increases in temperature and photoperiod, leading the flowers to be visited by different suites of pollinators (Gómez et al., 2020).Pollinators can also show temporal intraspecific variation, such as butterfly generations (Wilson et al., 2019) and social insect castes differing in size through a season (Cantwell-Jones et al., 2023).Similarly, populations of the same species differently located in space can vary in mean-trait values and associated variance (phenotypic plasticity), skewness or kurtosis (Figure 4b; Classen et al., 2017;Le Bagousse-Pinguet et al., 2017;Taylor-Cox et al., 2020;Valladares et al., 2014), whether through local adaptation, variation in environmental filtering or stochastic processes (e.g.Rodríguez- Alarcón et al., 2022; but see Le Bagousse-Pinguet et al., 2017).For example, Classen et al. (2017) found that bee size increased with elevation (Mt.Kilimanjaro) within species, despite large-bodied species being filtered from the community at high elevation.If traits related to morphological compatibility and attractiveness are also filtered, then the range of potential interaction partners for a species (or the likelihood they interact) could also be impacted.Such systematic trait changes in space or time could confound the apparent influence of spatial and phenological F I G U R E 4 Using information on the shapes of species' frequency distributions of functional traits, we can assess how trait overlap between plants and pollinators predicts their interaction likelihood.(a) A series of pollinators (icons along the top), with specific frequency distributions (in orange) of functional traits, could potentially interact with a series of flowers (icons on the left-hand side), with their own frequency distributions of functional traits (in grey).If a plant-pollinator interaction requires matching of both plant and pollinator functional traits, we might predict that plants and pollinators with the greatest overlap in functional trait distributions would have the greatest likelihood of interacting (represented by blue circles, with larger circles representing increased likelihood).(b) Illustrates how differences in trait distribution shapes can be captured by the four moments (mean, variance, skewness, kurtosis; recreated from Wieczynski et al. (2019)).For skewness and kurtosis, '+ve' means positive and '-ve', negative.
overlap discussed in the previous section above.As among-population trait variation can drive changes in the resulting community-level networks (Start, 2019; in experimental food webs), investigating how the degree of overlap of trait frequency distributions changes in different environmental contexts may explain potential variation in community network structure when comparing localised populations of the same species.
In the context of environmental change, laboratory experiments and field observations have already highlighted how climate-mediated shifts in functional traits could impact interactions (e.g.Descamps et al., 2021;Gallagher & Campbell, 2017;de Manincor et al., 2023;Miller-Struttmann et al., 2015; see Barbour and Gibert (2021) for a review of how climate-driven trait change might rewire food webs).For instance, Hoover et al. ( 2012) found that pumpkins (Cucurbita maxima) altered their floral morphology, phenology and nectar chemistry under a combination of environmental change factors (elevated CO 2 , N and temperature).Moreover, the change to nectar chemistry impacted its attractiveness to visiting pollinators (bumblebees) and affected pollinator longevity.Gallagher and Campbell (2017) found that corolla length and seed set increased in Mertensia ciliata with increasing water availability, whereas pollinator visitation was highest at intermediate levels of water availability.Given that plants and pollinators can exhibit varying trait responses to global change-and may not even 'experience' the same environmental conditions (Hegland et al., 2009)-understanding how local environments influence species' trait frequency distributions will be vital for predicting plant-pollinator network responses (Schleuning et al., 2020).

Investigating vulnerability of plant-pollinator networks to environmental change using multidimensional trait space
Given the diversity of ways that environmental variables affect species traits, and in turn the suite of traits that contribute to a given interaction, it may be useful to visualise species in multidimensional trait space [e.g. using hypervolumes (Blonder et al., 2014(Blonder et al., , 2018) ) or trait probability density functions (Carmona et al., 2016(Carmona et al., , 2019))].In multidimensional trait space, each axis can represent a different trait (or combination of traits, if dimensionality reduction is performed), and an individual's position (as coordinates) corresponds to its given trait value along each axis (Blonder et al., 2018).Species can then be represented as having a volume in trait space that encompasses all its conspecific individuals, enabling calculation of species trait overlap (for example, hypervolume intersection) or alternatively how singular their traits are (i.e.unique component of trait space occupied, compared with the proportion of trait space shared with other species; Mammola, 2019).By quantifying overlap in trait space, we can then assess the degree to which morphologically similar species compete, that is share interaction partners, or whether species with larger intraspecific variation (i.e.greater volumes in trait space) interact with a greater variety of interaction partners.Further, if pollinator species that share a large area of multidimensional trait space compete for similar floral resources, this could help predict how a network may rewire following introduction of an invasive pollinator species (Valdovinos et al., 2018), by looking at where the invasive species sits in trait space (Parra-Tabla et al., 2019).Alternatively, it has been shown empirically that niche packing in plant-frugivore networks emerges through higher resource overlap (measured as resource trait probability densities and hull volume overlap), with the increasing competition being offset by less frequent use of shared resource species (Dehling et al., 2022), such that the use of a subset of possible resources by individuals may be a critical determinant of this complementarity.
By mapping changes in trait space under environmental change, we should be able to forecast the potential vulnerability or resilience of a plant-pollinator network (Theodorou et al., 2021).For example, if, under an environmental perturbation (such as drought), we see a contraction in trait space (i.e.species converge on a similar set of traits; Pradervand et al., 2014;Rodríguez-Alarcón et al., 2022), we might expect to see higher levels of competition within the plant-pollinator network, as individuals with similar traits converge on the same interaction partners, leading to increased redundancy in the network (Figure 5).Alternatively, if species disperse in trait space (e.g.species-specific responses cause species to become less similar in their traits), we might expect higher levels of specialisation in the network.The development of 'holes', or unoccupied areas of trait space (Blonder, 2016), could suggest a loss of function, such as the disappearance of interactions.Further, simultaneous mapping of both plant and pollinator trait-space changes could reveal how levels of overlap between interaction partners change.This approach would allow predictions of whether they would continue to interact under environmental change, or whether they would rewire to interact with new species (Barbour & Gibert, 2021;Burkle et al., 2013, but see Leimberger et al., 2023).Forecasting species' ability to rewire interactions is key, given that this plasticity increases network robustness (Ramos-Jiliberto et al., 2012;Vizentin-Bugoni et al., 2020).However, a key first step for such forecasting is to elucidate the role, potential explanatory power and context dependency of (individual-level) traits for predicting plant-pollinator interactions.

Species identity is also important
Focusing on functional traits, such as morphology and phenology, to understand interactions provides us with 14610248, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/ele.14368by Umea University, Wiley Online Library on [04/02/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License a common framework that is generalisable across species (Adler et al., 2013) and incorporates intraspecific variation in interactions.However, within networks, individuals of different species with similar functional traits often have different diets, even within the same consumer guild or trophic level, and species with similar traits do not always forage similarly (as shown in food webs and plant-pollinator networks; Jacob et al., 2011;Resetarits & Chalcraft, 2007;Rudolf et al., 2014;Sponsler et al., 2022).Individuals within the same species may share inherent behavioural preferences that can be difficult to capture with a small number of functional traits (Poisot & Stouffer, 2018;Segar et al., 2020; but see Eklöf et al., 2013;Pigot et al., 2020).As these differences could result from shared evolutionary history, combining phylogenetic data with functional-trait values could provide the most useful approach for modelling plant-pollinator interactions (Salguero- Gómez et al., 2018).
The relative importance of species identity (or evolutionary history) and traits for predicting interactions may also be context dependent.For instance, if two species have co-evolved together, are obligate mutualists or strongly rely on specific guilds (Vamosi et al., 2014; such as plants that rely on buzz pollination; Armbruster et al., 2013), then species identity might be more predictive than intraspecific trait variation for forecasting interaction probabilities.Conversely, for species where intraspecific variation is much larger than interspecific variation, the trait values of individuals may be more useful for understanding interactions (Szigeti et al., 2020).Aside from the evolutionary context of the interaction, the relative importance of species identity and trait values may also be determined by the environmental or spatiotemporal context (Cantwell-Jones et al., 2023;Sponsler et al., 2022).For example, in plant-hummingbird networks, the relative importance of morphological matching versus species abundance varies with latitude, where morphological matching is more important in low-latitude networks (Dalsgaard et al., 2021;Simmons, Sweering, et al., 2019;Sonne et al., 2020).Additionally, F I G U R E 5 Environment-driven changes in trait space could fundamentally reshape interactions within a dynamic plant-pollinator community.Using a hypothetical example, in (a) near the base of the mountain, three pollinator species (in different colours) fill relatively large areas in trait space (each species has large intraspecific variation).All pollinator species overlap to a degree in trait space, and therefore should share plant partners (black nodes) in the interaction network, assuming that trait matching drives interactions.As a result, the pollinator species show a degree of functional redundancy (i.e. they are morphologically and functionally "homogeneous").In (b) at the top of the mountain, the three pollinator species separately converge on their respective centroids in trait space due to environmental filtering (i.e.their intraspecific variation declines), and this leads to species becoming less functionally similar.In other words, this reduced overlap in trait space reduces functional redundancy, with fewer plant partners being shared.
14610248, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/ele.14368by Umea University, Wiley Online Library on [04/02/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License species may modify their interactions or traits when in the presence (or absence) of another species (Blanchet et al., 2020;Brosi & Briggs, 2013;Helsen et al., 2020).Therefore, when moving towards predictive frameworks of plant-pollinator interactions, we must consider that functional traits are nested within species identity and the species' evolutionary history, as well as that interaction networks are nested within a local environmental context (e.g.Cusser et al., 2019;Zhao et al., 2022).To achieve this, careful design of fieldwork experiments and annotated archival metadata are needed to collate highresolution interaction and trait data across environmental gradients.

Future directions: Methodologies to collect and analyse the data needed for predictive frameworks of plant-pollinator interactions
Thus far, we have reviewed how considering individuallevel trait data could conceptually enable a greater mechanistic understanding of plant-pollinator interactions.However, what data are needed to build such a framework, and how can they be feasibly collected?Ideally, data would be collected on (1) spatiotemporally explicit individual plant-pollinator interactions, (2) corresponding functional trait data for both individual plants and pollinators, (3) abundances of the observed plant and pollinator species (independent from the interaction data) and (4) local environmental data, including macro-and microclimate variables.These data would be collected at a temporal resolution that can well capture dynamics across the growing season (or however long plants and pollinators are active in the system) across multiple years, and over appropriate spatial scales such as along informative environmental gradients (Figure 5; Pellissier et al., 2018;Tylianakis & Morris, 2017).Collecting such high-resolution, individual-level data is, however, timeintensive, although existing datasets that meet (some of) these requirements do exist (e.g.Cantwell-Jones et al., 2023;Liang et al., 2021;Rumeu et al., 2018;Wei et al., 2021).Although financial and time constraints can hamper the physical collection of such high-resolution datasets, recent technological advances may assist data collection (Høye et al., 2021).Below, we suggest statistical methods that allow us to make better use of existing data to answer evidence gaps raised in this paper, by (A) using it to inform theoretical research, (B) explicitly linking trait and interaction data, and (C) analysing network structure in higher resolution.We additionally provide (D) methods that enable more efficient data collection.
(A) Codeveloping theoretical and empirical plantpollinator networks to better understand interaction likelihoods: • Agent-based or individual-based models: One approach to circumvent some of the constraints of data collection (especially at the individual level) is to use existing empirical networks to inform parameters in agent-based or individual-based models.These models create a virtual system, in which individual agents (such as plant or pollinator individuals) are simulated.These agents can interact with each other and their environment, and the emergent consequences of the agents' behaviours can be studied (e.g.Gegear et al., 2021;Kortsch et al., 2023;Newton et al., 2018).
Agent-based models have so far been used to investigate how factors impacting individual pollinator behaviour (for instance, land fragmentation, forest cover or impaired cognition through sublethal stressors) can scale up to impact the number of plants pollinated and the resulting plant-pollinator network (Gegear et al., 2021;Kortsch et al., 2023;Newton et al., 2018).• Empirical dynamic modelling (EDM; Sugihara & May, 1990): EDM is an equation-free, mechanistic modelling approach that can take a time series of a system and accurately forecast future species dynamics (Sugihara & May, 1990;Ye et al., 2015).EDM has also been used to forecast ecological interactions in real time, by enabling predictions of interactions based on the 'context' of the system (such as species abundances, resource availability or environmental conditions; Deyle et al., 2016).Incorporating individual-level data could therefore further enhance tracking species interactions over time and improve future predictions of species dynamics.• Estimating potential missing links and using probabilistic networks: When conducting plant-pollinator observations, it is often impossible to detect every species or interaction, even with extensive sampling (Chacoff et al., 2012), potentially leading to inaccurate estimates of network structure and levels of specialisation (Fründ et al., 2016).Two solutions include either estimating potential missing links (such as with Bayesian frameworks; Cirtwill et al., 2019;Young et al., 2021) or using probabilistic networks, which represent interactions as probabilistic events that are inherently variable in space and time (Poisot et al., 2016).By providing estimates of interaction certainty or likelihood, these methods can help identify interactions that would benefit from greater sampling effort (Young et al., 2021).
(B) Linking traits and interactions in plant-pollinator networks: • Machine learning: Although traits have shown strong predictive power in predatory-prey and seed-dispersal interaction networks (e.g.Bender et al., 2018;Dehling et al., 2016;Woodward & Hildrew, 2002), power has been more modest for plant-pollinator interactions (e.g.Bartomeus et al., 2016;CaraDonna et al., 2017; but see Eklöf et al., 2013).However, machine learning is being increasingly used to predict pairwise interactions at the species level, due to its ability to recognise the complex patterns underpinning interactions, such as those driven by multiple functional traits (Borowiec et al., 2022;Strydom et al., 2021).Some machine learning methods (e.g.random forest, boosted random forest or deep neural networks) can outperform traditional statistical techniques in their ability to correctly predict interactions or the functional traits underpinning them (Pichler et al., 2020).Due to the ability of machine learning to detect patterns in large and complex datasets (Greener et al., 2022;Strydom et al., 2021), it will likely prove a useful tool for predicting how interaction networks rewire following a local species loss or change in the distribution of functional traits.

• Exponential random graph modelling (ERGM):
Another approach to explicitly link species or individual-level trait data to interactions is to use ERGMs (e.g.Arroyo-Correa et al., 2021;Miguel et al., 2018 for an example in plant-frugivore networks), which allow properties of the entire network to be predicted by node traits or local structural properties.For example, ERGMs could be used to assess whether the odds of two species sharing interaction partners depend on their overlap in multidimensional trait space, and how important this overlap is relative to shared evolutionary history between species or local environmental conditions.
(C) Analysing the drivers of network structure at higher resolution: • Motif analysis: This breaks a network down into its constituent subnetworks ('motifs'), which have been referred to as 'building blocks' of networks (Milo et al., 2002;Simmons, Vizentin-Bugoni, et al., 2019).Motifs can be used to measure network structure sensitively (Simmons, Cirtwill, et al., 2019); quantify 'roles' of species in a network (Bramon Mora et al., 2020;Cirtwill et al., 2018) or identify distinct behaviours within networks (Pasquaretta et al., 2021).As using motifs enables analysis of indirect interactions (e.g.Dritz et al., 2023), they can likely be used to assess how competition for pollinators or floral resources changes following environmentally driven trait change.• Ecological multilayer networks: Multilayer networks allow networks across space, time, different interaction types or different levels of biological organisation to be analysed in a single framework as different 'layers' (Hutchinson et al., 2019;Pilosof et al., 2017).Links between (interlayer edges) or within layers (intralayer edges) can be weighted according to the proportion of individuals of a given species that interact with a partner species or engage in different types of interactions (Hervías-Parejo et al., 2020).Multilayer networks may enable ecologists to assess how population-level processes cascade to impact community or metacommunity network structure (Scotti et al., 2013).
For example, Hervías-Parejo et al. ( 2020) joined avian seed-dispersal and pollination networks into a multilayer network, which revealed that, while the majority of bird species (80%) engaged in both processes, it was only a minority of individuals from these species (7%) that exhibited this dual-function behaviour.This highlights how multilayer networks can be used to distinguish between the roles of species versus individuals (potentially with specific functional traits) in contributing to ecosystem functioning (Hervías-Parejo et al., 2020).
(D) Collecting more data, more efficiently: • Automated detection methodology: Recent years have seen increased recording of plant-pollinator interactions or plant phenology using cameras (e.g.Arroyo-Correa et al., 2023;Mann et al., 2022), allowing a much greater volume and higher resolution of data to be collected.Not only does using cameras reduce the likelihood that interactions go unobserved, but they can also be combined with automated detection technology (Bjerge et al., 2023;Mann et al., 2022), thus potentially reducing the time intensity typically required for fieldwork (Besson et al., 2022).• Pollen analysis: Plant-pollinator interactions can be inferred through analysis of the pollen carried by insects (e.g. de Manincor et al., 2020), using DNA metabarcoding (Bell et al., 2023;Gill et al., 2016;Richardson et al., 2021) or automated palynology using machine learning (e.g.Sevillano et al., 2020).Although pollen analysis may not always be spatially explicit, analysing pollen allows a highly resolved individual pollinatorplant species network to be constructed (Biella et al., 2019;Tur et al., 2014).• Metabarcoding of environmental DNA (eDNA): A promising method for detecting interactions is metabarcoding of environmental DNA (eDNA), such as that left by visitors on flowers (Harper et al., 2023;Newton et al., 2023;Thomsen & Sigsgaard, 2019).eDNA metabarcoding has so far proven to be an efficient, non-invasive approach to assessing species composition of large communities (Banerjee et al., 2022).However, there are still challenges around translating eDNA quantification metrics into organismal abundance, managing the detection of trace DNA, and DNA dispersing or remaining on surfaces over long periods (especially extracellular DNA, like pollen; Banerjee et al., 2022).• Estimating trait variation: Automating data collection of individual plant or pollinator traits will likely be challenging, but in the meantime certain multidimensional trait space methods (e.g.trait probability density method, implemented in the 'TPD' R package; Carmona et al., 2019) allow estimates of trait variation and covariation to be inputted alongside mean trait values (instead of individual measures), improving estimates of trait overlap between species over mean values alone.Moreover, online trait datasets (e.g. for plants: TRY [Kattge et al., 2011], FloRes [Baden-Böhm et al., 2022], and Tundra Trait Team database [Bjorkman et al., 2018b]; for pollinators: EuroBaTrait [Froidevaux et al., 2023], LepTraits [Shirey et al., 2022], GlobalAnts [Parr et al., 2017], AVONET [Tobias et al., 2022] and a dataset for Brazilian bees [Borges et al., 2020]) with georeferenced measurements may help provide localised species-specific estimates of intraspecific variation.Although the availability of trait data for many pollinators currently lags behind that of plants, the increasing use of trait data from (digitised) museum specimens may help close this gap (e.g.Arce et al., 2023;Oliveira et al., 2016).In particular, body size (length or mass) is frequently the best predictor of interactions in food webs and mutualistic networks (Eklöf et al., 2013), potentially because it correlates with many life-history traits (Woodward et al., 2005), and it is also one of the simplest to measure (even on pinned specimens).

Conclusions
Global environmental change is placing plant-pollinator networks under pressure, with uncertain consequences for the species involved and the functions they provide.This synthesis puts forward the urgent call for researchers to bring together individual-level interactions and traits across appropriate spatiotemporal scales and environmental gradients.This should become increasingly feasible with improvements in modelling methods and automation of data collection.We believe such data are key to building a more generalisable, predictive framework, and will provide the tools to forecast how species interactions can be reshaped, gained, or lost, which will have large implications for quantifying and predicting ecosystem health and productivity.

AU T HOR CON T R I BU T ION S
ACJ and RJG conceived the synthesis with inspiration and advice from JMT and KL.ACJ and RJG wrote the manuscript with significant contributions from JMT and KL.

F
I G U R E 2 Different methods of representing interactions at the individual level.(a) Five different pollinator individuals (represented as nodes of different colours) are interacting with two plant species or resources (nodes with different shades of green), with link thickness representing interaction strength.(b) Pollinator individuals are interacting with five plant individuals belonging to two different species (again represented with different shades of green).(c) Unipartite projection of panel (b), where links (and their respective thicknesses) between pollinator individuals represent the degree to which they visit the same plant individuals (i.e. share resources).

•
Maclean et al., 2019)20) The expanding use of climate loggers (e.g. the SoilTemp database;Lembrechts et al., 2020)coupled with improved methods for interpolating microclimate data (e.g.Maclean et al., 2019)allow modelling of the environmental conditions experienced by plants and pollinators in unprecedented detail.
ACJ is funded by the NERC Science and Solutions for a Changing Planet Doctoral Training Programme, Imperial College London (NE/S007415/1).The project was also supported by grants from INTERACT (funded by H2020-agreement no.730938) and Quekett Microscopical Club awarded to RJG, as well as the Genetics Society Heredity Field Grant and RGS-IBG Geographical Fieldwork Grant.JMT is funded by the Bioprotection Aotearoa Centre of Research Excellence.We would also like to thank Jonas Lembrechts, Olivia K. Bates, the Gill & Graystock research groups, and the many MRes students and field interns who have contributed to discussions.The manuscript benefitted greatly from comments from four reviewers.The peer review history for this article is available at https:// www.webof scien ce.com/ api/ gatew ay/ wos/ peer-review/ 10. 1111/ ele.14368 .