We studied individuals of pollinator and plant species in both datasets to evaluate intraspecific variation and specialization in pollen-transport and pollen-transfer comparing these to the species-based networks in the original studies. All individuals within a given species were considered as independent nodes in these networks. Networks were defined as weighted Aij adjacency matrices, thus considering the frequency of interactions among nodes. For pollen-transport networks, i represents a pollinator species in the species-based networks and an individual pollinator in the individual-based networks. Then, the aij element corresponds to the amount of pollen loads on pollinators’ body at the species and individual levels, respectively. For pollen-transfer networks, i is a plant species in the species-based networks, and an individual plant at the individual-based networks. The aij element is the amount of pollen grains deposited on plants’ stigmas at the species and individual levels, respectively. The amount of pollen grains on stigmas is estimated by the average pollen load of three open flowers per individual plant (if that was not possible, we considered the number of pollen grains present in one or two flowers, as available in the sampling area). Thus, the aij is the sum of the average pollen load of all individuals of a given plant species in the species-based, pollen-transfer networks and the average pollen load per individual plant at the corresponding individual-based networks. In all cases, j corresponds to a pollen type, keeping the number of columns in the adjacency matrix constant in both individual- and species-based networks (Fig. 1, Table S1).

Fig. 1.

A hypothetical diagram showing how the different types and levels of animal-mediated pollination are assembled in ecological networks and the role of intraspecific variation on the arrangement of network links. Species-based networks aggregate all the individual information of each plant and pollinator species of a given community into a single unit – the species. Individual-based networks consider each individual of a given species as a single unit in the network, therefore taking into account the intraspecific variation that may affect structural patterns. Plant-flower visitor networks use flower visitation as a proxy for pollination; pollen-transport networks consider the pollen load on pollinator’s body after a flower-visitation event while pollen-transfer networks consider the pollen deposition on stigmas as a baseline in which pollination takes place.

(0.42MB).

Balsam pollen load on pollinators’ bodies was measured as the number of balsam pollen grains counted on the body of each individual of A. mellifera and B. pascuorum recorded in the pollen-transport networks. Balsam pollen deposition on plant stigmas was measured as the number of balsam pollen grains deposited on individual stigmas of C. sepium, E. hirsutum and C. lutetiana likewise recorded in the pollen-transfer networks. Both individual pollen load and pollen deposition were compared to the quantities recorded for the corresponding species level data (see section on the statistical analyses).

We estimated node specialization considering the target pollinator (A. mellifera and B. pascuorum) and plant species (C. sepium, E. hirsutum, C. lutetiana) by calculating their k degree and d’ specialization index (Blüthgen et al., 2006). Here, degree is the number of distinct pollen types found on pollinator’s bodies and/or deposited on plant stigmas for pollen-transport and transfer networks respectively, at the individual- and species-based levels. For all cases, d’ specialization estimates how selective a given plant or pollinator individual/species is regarding the potential partners available and is based on the full communities in which they were sampled.

Considering the entire dataset of both target studies, i.e., the networks containing all the plant and pollinator species sampled, we aim to understand whether changing from species- to individual-based networks, and from pollen-transport to pollen-transfer networks increases overall network structural specialization and pollinators’ niche partitioning. The results are expected to help to understand the generalism/specialism conundrum in pollination systems (Waser et al., 1996; Armbruster, 2017; Ashman et al., 2020). To do this, we used three complementary, non-correlated metrics, that give us different perspectives about the overall structural specialization, as follows:

i) H’2 specialization: this metric describes how the observed distribution of interactions deviates from an expected distribution based on the number of interactions per species in the network (Blüthgen et al., 2006). The higher the pairwise species selectivity, the higher its contribution to the overall network specialization. This index is based on the Shannon entropy, constraining the total number of interactions for each species; its final value is scaled between H’2min and H’2max such that the higher the value, the higher the network specialization.

ii) Interaction evenness (IE): this estimates the heterogeneity of links between the two interacting groups within the bipartite network. It is calculated as the sum of the proportion of the ij interaction (Shannon-diversity) divided by the total number of interactions, as follows:

iii) Pollen niche overlap: this estimates the mean similarity of interactions of the pollen type level, calculated by the Horn’s index (Horn, 1966; Dormann et al., 2009). It’s value ranges from zero to one; a niche overlap of zero means that no interactions are shared between pollen types given the use of pollinators’ bodies for transport, or the stigmas for deposition (i.e., higher specialization in the choice of transport or deposition sources); a niche overlap of one means that all pollen species are transported around by the same pollinators’ bodies, or are deposited on the same stigmas (i.e., low specialization).

To test for intraspecific variation within the target pollinator and plant species (objective 1 and 2), we used the Jarque-Bera Normality Test (Jarque and Bera, 1980) within the ‘moments’ package (Komsta and Novomestky, 2015). It measures the shape of the individual data distribution for each species, testing whether the distribution is left or right skewed (i.e., skewness), and whether it differs from a normal distribution (i.e., kurtosis); values far from zero indicates the data do not follow a normal distribution. Here, we would expect a normal distribution if most of the individuals perform similarly within the network context, i.e., most pollinators carry similar numbers of balsam pollen grains and most plants have similar numbers of stigmal balsam pollen deposition while only a few individuals would have extremely high or extremely low number of pollen grains, therefore representing the upper and lower tails of a normal distribution. In turn, we would expect a right-skewed distribution if only a few individuals of plants or pollinators have a high number of balsam pollen grains while most of the population have a low number, including many zeros.

To understand whether node specialization (objective 3) varies between network type (transfer and transport) and level (individual- and species-based) we used General Least Square (GLS) and Linear Mixed Models (LMM) in the nlme package (Pinheiro et al., 2016). We fitted individual models for each proxy of node specialization (i.e., degree and d’ specialization) used as the response variable tested against the interaction between network type and network level. Degree was log-transformed and d’ logit-transformed, as it is bounded between 0 and 1, to fit normal distributions. We then selected the best fixed and random structures based on AIC model selection and corrected for boundary effects following Zuur et al. (2009). We started by choosing the best random structure among the following random model candidates: species nested within treatment - to control for pseudoreplication effects within each treatment as the same species could be counted more than once in each plot given the natural variation in the number of individuals; treatment only – to control for differences between invaded and non-invaded areas; and species only – to control for differences in species occurrences among plots. The best random structure model was then used to proceed with the selection of the best fixed effects.

Finally, we fitted individual models for each proxy of structural specialization (i.e., H’2, interaction evenness, niche overlap) that were used as the response variable tested against the interaction between network type and network level (objective 4). Treatment (invaded, non-invaded sites) were added as a random factor to account for potential effects of balsam invasion. Data were log- and/or logit transformed to fit normality assumptions where needed (Warton and Hui, 2011). Model selection based on the Akaike Information Criteria (AICc) follows the protocols suggested in Zuur et al. (2009). We fitted the best model structure to linear regressions with “maximum likelihood estimation” while applying the “dredge” function (Barton, 2013) to select the best-fixed structure according to Akaike Information Criteria (AIC). Then, we used the Akaike weight of evidence (wAICc) to obtain the relative importance of the different models (Burnham and Anderson, 2002). Network metrics were calculated using the bipartite package in R (Dormann et al., 2009); all analyses were run in R v 3.0.1 (R Core Team, 2014).

We found large variation in the number of balsam pollen grains transported by different individuals within pollinator species: A. mellifera: 7623.14 ± 7475.71 (mean ± SD; range 807-39071); B. pascuorum: 7234.58 ± 7362.13 (mean ± SD; range 78–40333). The distribution of pollen grains on the pollinators’ bodies is right-skewed, i.e., a few individuals carry a large number of pollen grains while most individuals carry a few (Fig. 2a and b). Likewise, there was significant intraspecific variation in the number of balsam pollen grains deposited on stigmas: C. sepium: 360.98 ± 589.79 (mean ± SD; range 0–2519); E. hirsutum: 9.79 ± 35.96 (mean ± SD; range 0–378), C. lutetiana: 12.07 ± 42.94 (mean ± SD; range 0–384). Similar to the pollen load, the right-skewed distribution of data shows that a few stigmas received high numbers of pollen grains whilst the majority received very few, or none (Fig. 2c–e). In all cases, the shape of the data distribution in terms of skewness and kurtosis was significantly different from a normal distribution; thus, only a few individual plants and pollinators have high numbers of balsam pollen grains, while most of the population have just a few (Fig. 2).

Fig. 2.

The frequency distribution of the number of balsam pollen grains on the bodies of individuals of the commonest pollinators in the pollen-transport networks (Apis mellifera; Bombus pascuorum) and deposited on the stigmas of individuals of the commonest plant species in the pollen-transfer networks (Calystegia sepium; Epilobium hirsutum; Circaea lutetiana). Note the variation in the number of balsam pollen grains among individuals within each species. JB is the Jarque-Bera Normality Test showing the data is significantly right-skewed compared to a normal distribution. For visualization purposes, we present the original values without the log-transformation required for the applied normality test.

(0.38MB).

Three models were chosen as best explanations when testing differences in degree between network type and network level while one single model was selected for d’ specialization in the same context (Table 1). In both cases, species nested within treatments (i.e., invaded vs non-invaded areas) was selected as the best random structure, explaining 34.48% and 35.5.% of the variance for degree and d’ specialization models, respectively. In turn, treatment explained 0.01% and 34.13% of the variance for degree and d’ specialization, respectively. The intra-specific variation in the distribution of degree and d’ among species and between network type and level can be seen in Fig. 3.

Fig. 3.

The inter- and intra-specific variation of node specialization of individual and species-based pollen-transport and pollen-transfer networks, considering the most common pollinator and plant species in the study sites invaded by Impatiens glandulifera (balsam). Differences on degree and d’ specialization between network types and individual- and species-level networks. For simplicity, data from invaded and non-invaded sites are not shown, either because the treatment was non-significant in the model, or it explained very little of the total variance.

(0.48MB).

Degree was always higher at the species level than at the individual level as expected given that it is the aggregation of all individuals within each species (Fig. 4, Table 1). Among the selected models, the first ranked one showed a significant interaction between network type and level, showing that degree was higher for nodes within pollen-transfer networks but only at the individual level (Table 1, Fig. S1). This apparently counterintuitive result is related to a higher number of distinct pollen types recorded at the stigmas’ surface in comparison with those recorded at the pollinators' bodies when individuals are considered, a finding masked when data is pooled in species-based networks. Indeed, there were significant differences between individual-based and species-based networks in all models, including when level was selected as the only response variable. In comparison, network type did not significantly affect the degree of the target species.

Fig. 4.

Differences in node specialization (degree, d’ specialization) between network types (pollen-transport, pollen-transfer) and levels (individual-based, species-based) of invaded pollination networks. For simplicity, data from invaded and non-invaded sites are removed because treatment was non-significant in the models or explained very little of the total variance.

(0.27MB).

The full model was chosen as the explanation for differences in d’ specialization across the different levels of biological organization (Table 1, Fig. S2). Like degree, a significant interaction between network type and level was detected when analyzing d’ specialization, i.e., nodes showed higher selectiveness in the interactions they perform in the pollen-transfer networks, but only at the species level. Such a pattern was driven by differences between levels, in which individual-based networks are significantly more specialized than species-based networks (see results in the next section; Fig S3).

Overall, the three complementary metrics used here as proxies for specialization showed significant differences for network type and level (Table 2, Fig. 5, Fig. S3). Specifically, H’2 specialization index only differed between network types, in which pollen-transfer networks were significantly more specialized than pollen-transport networks while there were no differences between individual- and species-based networks. The models for interaction evenness showed a significant interaction between type and level, in which individual-based networks showed higher uniformity in the distribution of links within the pollen-transfer networks but not for pollen-transport networks (Table 2, Fig. S2). When considered independently, the difference in the magnitude of interaction evenness was greater for individual-based networks and there were also significant differences between network types, in which greater evenness was observed in the transfer of pollen to stigmas compared to the movement of pollen on pollinators’ bodies.

Fig. 5.

Empirical examples of pollen-transport networks (a, b) and pollen-transfer networks (c, d) at the species (a, c) and individual level (b, d) in areas invaded by Impatiens glandulifera (balsam). Blue bars represent the insect pollinators at both (a) species- and (b) individual-based networks; red bars represent the plant stigmas at (c) species- and (d) individual-based networks; yellow bars are the pollen types recorded on (a, b) pollinators’ bodies and (c, d) deposited on stigmas. Species are connected by weighted links whose width corresponds to the frequency of interactions. Note that the number of species in the upper side of the networks remains constant within network types because we are evaluating intraspecific variation in just one side of the networks, i.e. at the pollinator or stigma level. The target pollinators (Bombus pascuorum - Bp, Apis mellifera - Am) have their names highlighted in shades of blue (a, b) and the target stigmal plant species (Calystegia sepium - Cs, Circaea lutetiana - Cl, Epilobium hirsutum - Eh) in red (c, d). See Suplementary Material for species abbreviation names (Table S1, S2).

(1.04MB).

Finally, niche overlap of pollen types in the ‘use’ of pollinator’s bodies and stigmas was significantly lower in the individual-based networks but did not differ between network types (Table 2, Fig. 5, Fig. S3). Yet, there is a tendency of lower niche overlap of individual-based, pollen-transfer networks in comparison to individual-based, pollen-transport networks while no difference seems to occur at the species-based networks, as indicated by the marginally significant interaction between network type and level. Treatment (invaded, non-invaded areas) as a random factor was not selected as the best model structure in any of the above models. These results suggest that pollen-transfer networks are more specialized than pollen-transport networks, at least at the individual level, and that niche partitioning tends to be higher given the pollen transfer to stigmas in comparison with pollen transport data from pollinators` bodies when intraspecific variation is considered.

Although the networks studied here are from the same study system, i.e., they both study the invasion of Impatiens glandulifera in urban and semi-natural areas around the city of Bristol in the UK, the pollen-transport and pollen-transfer networks were built independently, in different years, by different ecologists, with a different design and different number of replicates. Thus, caution is needed when comparing both networks due to possible fluctuation in species composition between years and methodological differences (see Arceo-Gómez et al., 2018). That said, samples were collected from the same regional pool of species and while rare species in networks vary year to year, common species and network topology tend to remain the same (Alarcón et al., 2008; Petanidou et al., 2008; Dupont et al., 2009). Thus, Forup et al. (2008) found that, when abundance is taken into account, up to 94% of the species in networks remained the same between samples with three years intervals (2001 and 2004). Plus, the analyses estimating intraspecific variation and node specialization were run with the five most abundant plant and pollinator species in the area, therefore minimizing the possible effects of abrupt temporal fluctuations that could influence our results. Ideally, sampling the four types of networks should take place concurrently (a huge task though), but our results provide some intriguing pointers to what would be expected. Moreover, our results are largely in agreement with what would be expected given our understanding of pollinator behavior. It is also important to note that, as any other study without temporal dimensions, our findings are the results of metrics and models based on a static picture of the community and we cannot capture fluctuations over time. Thus, some of the uncertainties on the underlaying causes of the patterns observed here can only be addressed in future studies that use experiments designed to address the relationship between different types of networks and built on different levels of biological organization over time.

Darwin (1859) pointed out that intraspecific variation in traits is the basic material of natural selection, which allows individuals to co-exist and evolve. This is recognized in other fields of biology such as on studies of niche partitioning (e.g., Bolnick et al., 2003; Pauw, 2013), foraging behaviour (e.g., Jakobsson et al., 2008; Song and Feldman, 2014) and predator – prey interactions (e.g., Araújo et al., 2011; Tinker et al., 2012). Intraspecific variation has been explored rather little in pollination networks though (but see Dupont et al., 2011; Tur et al., 2014 for exceptions). Here we showed that individuals of two common pollinator species, A. mellifera and B. pascuorum, show considerable variation in the amounts and diversity of pollen they carry on their bodies. The same pattern was observed in pollen deposition on stigmas of three common plant species, C. sepium, E. hirsutum and C. lutetiana. Our findings suggest that the two patterns could be linked: the intraspecific variation of pollen types transported by individual pollinators translating into the intraspecific variation observed in pollen deposition. If this is the case, one can expect a cascading effect of intraspecific outcomes through the different levels of biological organization when considering the movement of pollen grains from plant anthers to plant stigmas.

Pollen-transfer networks were more specialized than pollen-transport networks. Both types of networks are poorly explored in pollination studies, which traditionally use plant-flower visitor interactions as a proxy of pollination. Pollen-transfer networks are less well understood and to date, as far as we know, very few studies have used this approach to understand pollen movement at a community level (see Fang and Huang, 2013; Emer et al., 2015; Ballantyne et al., 2015; Parra-Tabla et al., 2020). Because pollen deposition on a conspecific receptive stigma is a key step for pollination success, an interplay of mechanisms of competition and facilitation are expected to be in place to optimize the probability that the right pollen reaches the right stigma in multi-species communities which share pollinators (Ashman et al., 2020). Indeed, our analyses of niche overlap showed that pollen types overlap less in deposition on stigmas, either at the species or individual level, in comparison to the use of pollinators` bodies as transport vectors, which suggest the existence of pollen filtering mechanisms. It is beyond the scope of this study to identify these mechanisms, but we advocate that further studies should incorporate pollen-stigma chemical and physiological interactions (reviewed in Ashman et al., 2020). These could provide an explanation for the structure of the pollen-transfer networks as our findings suggests that individual variation in the morphological (e.g., Montgomery and Rathcke, 2012), chemical and/or physiological (e.g., Hiscock et al., 2002; Allen et al., 2011) traits of the interacting pollen and stigma are in place to avoid heterospecific pollen adherence to stigmatic surfaces (Arceo-Gómez et al., 2016, 2019). Electrostatic forces between pollinators and plants also play important roles in pollen transfer (Clarke et al., 2013 and references therein). Thus, mechanisms such as chemical reactions and electrostatic fields that occur at the individual level could explain, or be part of the explanation, concerning the species-specificity of pollen load on plant stigmas identified here. Plus, factors observed to influence plant fitness such as plant pollinators’ attractiveness (Chittka and Schurkens, 2001), abundance and morphological traits (Willmer and Finlayson, 2014; Lázaro et al., 2019) of both interacting players might influence pollen movement at the community level and could explain the observed selectiveness of interactions regarding pollen movement from anthers to stigmas.

Although many studies have investigated balsam invasion, there is still no consensus on its impact on native communities. For instance, while Chittka and Schurkens (2001) experimentally report negative effects of balsam on the seed set of a native species, the observational studies of Cawoy et al. (2012) and Bartomeus et al. (2010) found no evidence of balsam affecting the seed set of native species or outcompeting for pollinators, respectively. In turn, Vilà et al. (2009) compiled community-wide effects of alien species across Europe, including balsam, reporting that its presence did not affect network structure. Our results shed some light on the observed variation in balsam impact. Thus, when revisiting Lopezaraiza-Mikel et al.’s (2007) dataset we found considerable intraspecific variation in the quantity of balsam pollen transported on the bodies of the two commonest pollinator species, A. mellifera and B. pascuorum. A. mellifera and B. pascuorum are well known as generalist species in plant-visitor networks (Goulson, 2003; Valido et al., 2019; Arroyo-Correa et al., 2020), and they visit many plant species. However, our results show that this generalization is not constant among individuals of the same species, rather there is substantial variation. One of the reasons for this intraspecific variability could be that different individuals show different levels of floral fidelity, depending on competition, niche partitioning and the resources available (Gegear and Thomson, 2004; Brosi and Briggs, 2013; Pauw, 2013; Huang et al., 2015). Pollinators which show a high fidelity to balsam would also explain the low rates of balsam pollen transfer to stigmas of other species. If this is the case, balsam would be unlikely to negatively impact the overall native plant reproduction. Why some plant species consistently received more balsam pollen deposition than others though, as observed by Emer et al. (2015), remains an open question that requires further investigation.

Balsam pollen transfer to interspecific stigmas is likely to be affected by the patterns of pollinators’ grooming. Grooming by foraging bees is common after flower visitation, varies in frequency and timing (Rademaker et al., 1997; Holmquist et al., 2012), and it is known to influence pollen transfer from donor to recipient flowers (Harder and Aizen, 2010; Willmer and Finlayson, 2014). One of the reasons pollinators’ groom is to remove excessive pollen grains attached to their bodies that can affect foraging ability (Harder and Aizen, 2010). Thus, we can hypothesize that higher frequency and/or intensity of pollinator’s grooming after visiting high pollen-producers such as balsam (Baude et al., 2016) could explain the low transfer of alien pollen to the stigmas of native plants.

The intraspecific variation of invaded pollen-transport and pollen-transfer networks are a double-edged sword with respect not only to the impact of balsam invasion but also to solve the puzzle in understanding how invasive species integrate into ecological networks (Emer and Timóteo, 2020). On one hand, the few individuals of pollinators with very high numbers of alien pollen grains on their bodies could function as “super-spreaders” of invasive species in a network context (Fu et al., 2015). Super-spreaders are usually rare and characterized by nodes within the network population that are particularly effective in spreading information, such as an infected organism which transmits a disease to a high number of individuals or a politician that strongly influences their many voters (Zhang et al., 2019). Here, the very few pollinators that carry large amounts of alien balsam pollen grains on their bodies could potentially act as super-spreaders of invasive species, therefore strongly affecting the dynamics of biological invasions. That said, most stigmas had very low numbers of alien balsam deposition which limits the impact of the invasive species. Overall, our results suggest that, even if a highly invasive plant species produces vast quantities of pollen and shares many pollinators with native species, most alien pollen-transport is mediated by relatively few pollinator individuals and transferred to only a few plant individuals in the community. Therefore, the negative effects, if any, on native plant reproduction due to alien pollen interference (e.g., stigma clogging), are likely to only affect a few plants, buffering the impact of the plant at the population and community level. In a broader context, one can speculate that the integration of invasive plant species in ecological networks may be mediated by a few individuals that function as super-spreaders at different levels of biological organization. In the case of pollination, super-spreaders could be present from the flower-visitor to the pollen-transfer level, and their effects may be different depending on the scale in which they act.

Finally, our findings showing that only a few individuals within a population can drive the patterns observed at the species level support the idea the that intraspecific effects can be comparable to, and sometimes stronger than, species effects (Des Roches et al., 2018). This reinforces the importance of quantifying the magnitude of intraspecific variation. In the specific case of invasion biology, knowing whether individual pollinators within a species vary in their transport of heterospecific pollen and whether individual plants vary in their stigmal pollen load capacity, may help to explain why the impact of invasive species can be positive (Molina-Montenegro et al., 2008), neutral (Bartomeus et al., 2008a, b), or negative (Chittka and Schurkens, 2001, Richardson and Traveset, 2020).