Pollinator specialization increases with a decrease in a mass‐flowering plant in networks inferred from DNA metabarcoding

Abstract How native mass‐flowering plants affect the specialization of insects at individual and species levels and the consequences for pollination networks have received much less attention than for mass‐flowering crops or alien species and basically remain unexplored. Using existing DNA metabarcoding data on the pollen loads of 402 flower‐visiting insects, we assessed the effects of a native mass‐flowering plant of high reward quality, the shrub Rhododendron ferrugineum, on pollination networks by investigating: (a) the food niches of individual pollinators and pollinator species and (b) the structure of individual and species networks in subalpine heathland patches with extremely contrasted densities of R. ferrugineum. Relative to its high abundance in high‐density patches, the shrub was greatly underrepresented and did not dominate individual's or species' generalized networks, rather individual and species specialization increased with a decrease in R. ferrugineum density. Furthermore, individuals of the more generalist dipteran Empididae species tended to extend exclusive interactions with rare plant species in low‐density networks. The same trend was observed in the more specialist Apidea but toward rare species in high‐density networks. Our results reveal a quite paradoxical view of pollination and a functional complementarity within networks. Niche and network indices mostly based on the occurrence of links showed that individual pollinators and pollinator species and networks were highly generalized, whereas indices of link strength revealed that species and above all individuals behave as quite strict specialists. Synthesis. Our study provides insights into the status of a native mass‐flowering plant in individual's and insect species' food niches and pollination networks. It revealed that a generalist pollinator species can be highly specialized at the individual level and how rare plant species coexisting with mass‐flowering plants may nevertheless be visited.

Since pollinators tend to preferentially visit the most abundant and/or rewarding resources in a given area (Dauber et al., 2010;Ghazoul, 2005), native mass-flowering plants could, like exotic plants, attract most pollinators from the surroundings and, to a certain extent, dominate the networks with possible negative effects on coflowering species. However, because pollinator behavior may change with the abundance and the quality of resources even in small areas (Dupont et al., 2014;Osborne & Williams, 2001), the reduction in mass-flowering plant cover in the community could lead to a new pollination network structure, particularly if the reduction results in a dramatic decrease in locally available resources. On the one hand, lower floral resource abundance may favor generalist species that can compensate for the decrease in mass-flowering species by using alternative resources and/or keep species specialized on the mass-flowering plants away from the vegetation patch. What is more, individuals may broaden their diet (individual generalization) in response to competition for food (Fontaine, Collin, & Dajoz, 2008;Kunin & Iwasa, 1996). Together, these processes would lead individual and species networks to be more generalized with high connectance (the proportion of links that actually occur in networks, which increases with generalization) and possibly high nestedness (the degree to which specialist plants and insects interact with generalists plants and insects). On the other hand, in response to competition for resources, a subset of individuals may shift to other resources (individual specialization) even though other members of the population continue to forage on the original resources (Bolnick et al., 2003;Svanbäck & Bolnick, 2007). Individual specialization on different floral resources would lead to generalized species networks.
Whether species networks mainly comprise either specialist or generalist individuals would have very different ecological consequences. Indeed, besides affecting the population dynamics of the pollinators themselves (Araújo, Bolnick, & Layman, 2011), the magnitude of individual specialization may strongly determine the quantity and the quality (conspecific vs. heterospecific) of pollen transferred and have considerable consequences for both plant fitness and population and community dynamics. Furthermore, generalized species networks with high connectance and nestedness are considered to be less vulnerable to species losses (Memmott, Waser, & Price, 2004). Thus, understanding the network architecture and whether the species mainly comprises specialized or generalized individuals is essential since it can affect the structure and the robustness of the network, community stability, and pollination services in the face of disturbances (Thébault & Fontaine, 2010).
In this study, we focused on the effects of the high reward native mass-flowering shrub Rhododendron ferrugineum on the structure of pollination networks by investigating: (a) individual pollinator's (a total of 402) and pollinator species' food niches and (b) the structure of individual and species networks in subalpine heathland patches with highly contrasted floral densities of R. ferrugineum.
We investigated pollinator food niches and networks by using DNA metabarcoding data that were obtained in a previous study at the site (Pornon, Andalo, Burrus, & Escaravage, 2017). Metabarcoding has already been shown in several studies to successfully identify plant taxa in pollen mixtures (Cornman, Otto, Iwanowicz, & Pettis, 2015;Galimberti et al., 2014;Galliot et al., 2017;Lucas et al., 2018;Macgregor et al., 2019;Pornon et al., 2016;Richardson et al., 2019) at a higher taxonomic resolution than microscopy (Vamosi, Gong, Adamowicz, & Packer, 2017). A recent meta-analysis (Lamb et al., 2018) concluded on a significant quantitative positive linear relationship between the proportion of sequence reads and the proportion of species in original mixtures from very different origins (stomach contents, feces, airborne pollen, etc.). Some studies provided evidence for the quantitative potential of pollen metabarcoding both with nuclear (ITS) and plastid markers (rbcL, trnL, trnH;Keller et al., 2015;Kraaijeveld et al., 2015;Pornon et al., 2016;Richardson et al., 2019;Smart et al., 2017). However, Bell et al. (2018) andRichardson et al. (2019) obtained less reliable results with ITS2. Furthermore, depending on the species, higher or lower proportions of sequence reads may be found than expected (Bell et al., 2018). Despite these limitations, in a simulation study, Deagle et al. (2018) showed that even with methodological biases are included, sequence counts can provide a more reliable description of species diet in many scenarios than link occurrence alone. Moreover, since pollen grains may K E Y W O R D S DNA metabarcoding, food niche breadth, individual pollination network, mass-flowering species, pollinator generalization, pollinator specialization, species pollination networks accumulate on insect bodies during successive visits (Delmas, Fort, Escaravage, & Pornon, 2016;Jacobs et al., 2010), the proportion of either pollen grains or sequences of a given plant species in an insect pollen load may reflect the proportion of visits to the species Stanley & Stout, 2014).
In heathland patches with a high density of R. ferrugineum and limited alternative resources, we predicted (H1) that individual insects would forage mostly on the shrub. Thus, individuals, species, and networks would mainly be specialized in R. ferrugineum with low between-individual diversity in pollen use, large food niche overlap between individuals or species, and that the individual's diet would largely resemble the species' diet. However, this overall tendency could, to some extent, be counteracted if R. ferrugineum is not their favorite source of food, if it provides resources that are difficult for some insect species to access, or if individual and/or colony development require diversified resources (Girard, Chagnon, & Fournier, 2012;Tasei & Aupinel, 2008). In patches with low R. ferrugineum density and limited alternative resources, pollinators could compete for less available flower resources (Table 1). Delmas et al. (2016) observed more frequent visits to R. ferrugineum and smaller quantities and proportions of its pollen in insect loads in low-density patches than in high-density patches, suggesting that pollinators had to share fewer resources among more individuals in low-density patches. We hypothesized that in response to global resource depletion, individuals would either diversify their diet (H2: higher individual generalization) or specialize in different floral resources (H3: higher individual specialization) with different consequences for individual and species networks.

| Study site and data collection
We sampled plants and their visiting insects in four subalpine R. ferrugineum heathland patches within a 3-km 2 area in the French Central Pyrenees (southern France) near the village of Camurac (42°46′31″N 01°55′45″E; 1,660 m a.s.l.) during the shrub's blooming period (June 2012). Heathlands were spatially delimited in visually distinct patches (i.e., aggregations of R. ferrugineum embedded in their associated flowering community; hereafter referred as "surrounding communities") and separated from other patches by a meadow. R. ferrugineum is an 0.7-0.8-m high evergreen shrub that grows as isolated individuals in vast subalpine meadows and in many other locations, covering hundreds of hectares . It bears up to 3,000 red, weakly zygomorphic tubular flowers per m 2 ) that produce nectar (0.7-1.4 μl; ) with a high sugar concentration (mean 63.4° Brix in our study site; Delmas, Escaravage, & Pornon, 2014). It is visited by many pollinators among which hymenoptera are the most typical (Delmas et al., 2014). The main characteristics of the vegetation in the patches are listed in Table 2. The four patches (low-density patches, LDP1 and LDP2; high-density patches, HDP1 and HDP2) had different floral densities (and different floral resources), mostly determined by R. ferrugineum cover. The distance between the patches LDP1 and LDP2 was 800 m, 1,500 m between the patches HDP1 and HDP2 and varied between 1,250 and 1,750 m between LDP and HDP. Since the foraging distance of most pollinators (including large bumblebees) is usually shorter than 500 m (Gathmann & Tscharntke, 2002;Zurbuchen et al., 2010), data obtained in the four patches can be considered independent.
Rhododendron ferrugineum floral density (mean no. flowers per m 2 ) was estimated from the shrub cover and measured in a 400 m 2 plot randomly chosen in the center of the patch, and the number of flowers was counted in 0.25 × 0.25 m plots placed on 20 randomly chosen individuals (Delmas et al., 2014). In each patch, the floral density of each species in the surrounding community was estimated three times during the R. ferrugineum's blooming period (early, mean, and late) in twenty 50 × 50 cm plots haphazardly chosen in   the largest number of either trnL or ITS1 sequences was to keep in matrices the results of the marker whose amplification was the most successful, in aim to prevent the underrepresentation of some species. Indeed, it is often observed during PCR that a species may be properly amplified by one marker but not by another one . For both i-sp M seq and i-sp M link , we took a count of more than 1,000 sequences (i.e., n ij > 1,000 seq.) from a given plant species as a proof of a link with the aim of removing, as far as possible, contaminations that could for instance have occurred due to airborne pollen or nonpollen plant tissues deposited on insect bodies. Therefore, all values lower than or equal to 1,000 sequences were set to zero in

| Network-level analysis
For each patch, we built ( Table 2, and their underlying statistics and mathematics are detailed in Appendix S1 and in Blüthgen, Fründ, Vázquez, and Menzel, (2008). In order to determine whether the network structure differed beyond what would be expected due to random effects, 1,000 null networks (Dormann et al., 2009) were generated using Patefield's algorithm implemented in the nullmodel function or, for i-sp N link specifically, the mgen function (with the options rep.cell set to false and autotransform set to equiprobable) of the bipartite package (Dormann, Gruber, & Fründ, 2008). Null networks were constructed through random sampling of our empirical matrices constraining the marginal totals. Thus, the relative frequencies of species in empirical matrices were conserved in null matrices.
However, with the mgen function interaction assignment in the null network still depended on marginal probabilities but the binary nature of the original i-sp N link matrices was kept. All the above-mentioned indices were assessed for the 1,000 null networks, and the mean and 95% confidence intervals were calculated.

| Food niche analysis
We calculated the mean linkage degree of individual (L i ) and species (L sp ) pollinators, that is, the mean number of plant species with which an individual insect or a species interacts. To assess whether and to what extent individual or species pollinators preferred some plant species over others, we calculated the minimum number of plant species (and therefore links) needed to account for 80% of the sequence reads for a given individual, hereafter noted L i (80%) or species, hereafter L sp (80%) . The greater the difference between L sp (80%) and L sp and between L i (80%) and L i , the smaller the number of plant species preferred by species and individuals, respectively (higher specialization). The specific attractiveness of R. ferrugineum was evaluated through the proportion of individual insects carrying its sequences.

| Effects of resource density on food niches and networks
We analyzed the effects of R. ferrugineum floral density, taxonomic insect families, and their interactions on network and on food niche indices using mixed linear models fitted with the HLfit function of the spaMM package (Rousset & Ferdy, 2014) and the prior.weights option that accounts for the different sizes of samples. We examined individual indices at family (Apidae, Empididae, and Syrphidae) instead of at species level since some species had either too few individuals or were present in only one patch. To control for spatial autocorrelations, a random patch effect was added to the models.
Residuals normality was always checked, and Box-Cox (Box & Cox, 1964) power transformation was applied if required. For the proportion of individuals carrying R. ferrugineum sequences, a binomial error was applied.

| Characteristics of the interacting communities
We captured, respectively, 97 and 123 insects in the low-density LDP1 and LDP2 patches and, respectively, 82 and 100 in the high-density HDP1 and HDP2 patches, belonging to, respectively, 29, 25, 25, and 29 species (

| Structure of species networks
Because of their large size, species networks (sp-sp N link and spsp N seq ) had low connectance (ranging from 0.12 to 0.17) and high nestedness (0.83-0.88). Species networks based on link occurrence (sp-sp N link ) were highly generalized (Table S1, S2), both globally (H ′ 2 ≤ 0.23) and at the species level (d′ insects and d′ plants generally ≤0.31) with relatively high interaction evenness (0.67-0.72).

| Characteristics of individual and species food niches
Pollinator species had large food niches, with mean species linkage degree (L sp ) ranging from 3 to 35 (Table S2), but very low mean niche overlap (SPO = 14% ± 0.012). Furthermore, they had a highly uneven diet with preferential use of a small number of plant species (L sp (80%) : 1-7 species; Figure 1). Individual pollinators were also relatively generalized (L i : 1.25-7.6), although to a lesser extent than the species. They had very few preferred plant species (L i (80%) :1-3.2) and very low diet overlap (IO = 13% ± 0.025). On average, within-individual variation in pollen type use accounted for 53% of the total species niches (WIC/TNW = 0.53 ± 0.031 SE; TNW = 1.3 ± 0.10; WIC = 0.68 ± 0.058). The proportional niche similarity between individuals and their populations was relatively low (PS i = 0.49 ± 0.016).
In summary, these results revealed that, despite visiting a relatively broad range of plant species and thus having large diet breadth (and can thus be considered as generalists), each species, or even more strikingly each individual, was strongly specialized on a few and specific floral resources, most of which differed from those used by other species or conspecifics.
Within this general trend, linear models highlighted differences between taxa with respect to (a) individual (L i , p = .0003; Figure 1) and species (L sp , p < .0038) diet breath, (b) the number of plant species preferentially foraged by individuals (L i (80%) : p < .0001) or species (L sp (80%) p = .0142), (c) the indices of within-individual variation in pollen type use (WIC; p = .0067), of total niche width (TNW, p = .0027; Figure 2) and of the WIC/TNW ratio (p = .0345), and (d) species niche overlap (SPO; p = .0086; Figure 2). Empididae and Apidae were the two taxa mostly responsible for these differences, the former being the most generalist and the latter the most specialist of the pollinator assemblage often at both individual and species levels. Thus, relative to other taxa, Empididae had higher L i , L sp , TNW, WIC, and SPO. In contrast, Apidae had lower L sp (80%) , L i (80%) , TNW, WIC, and WIC/TNW ratio than the other taxa.

| Impact of the mass-flowering plant on food niches and networks
We observed overall higher specialization of pollinators in lowdensity patches than in high-density patches. Indeed, individual (L i , p = .0086) and species (L sp , p = .0186; Figure 3)  However, we also detected taxa-floral density interactions, that is, higher niche individual-population niche similarity (PS i ) in low-density patches for Apidae, whereas the opposite was observed for Empididae (interaction between patch density and insect family, p = .0198; Figure 3). Consistently, the d' of Apidae individuals tended to decrease and the d' of Empididae individuals tended to increase in low-density networks compared to those in high-density networks (interaction between patch density and insect family, p = .0473; Figure 3). Therefore, Empididae individuals tended to extend exclusive interactions with rare plant species in low-density networks, whereas individuals Apidae did the same in high-density networks.

| D ISCUSS I ON
In this study, we investigated the impact of the native mass-flowering R. ferrugineum on individual and species networks and on the food niches of the most abundant pollinators by identifying and quantifying insect pollen loads by metabarcoding. We predicted that individual insects and insect species as a whole would mainly forage on R. ferrugineum in high-density patches (Dauber et al., 2010;Ghazoul, 2005;H1) and would be either more generalized (Fontaine et al., 2008;Kunin & Iwasa, 1996;H2) or more specialized (Bolnick et al., 2003;Svanbäck & Bolnick, 2007;H3) in response to competition for limited resources in low-density patches. Our results failed to confirm either H1 or H2, whereas, in agreement with H3, pollinators were clearly more specialized in low-density patches than in high-density ones.

| The absence of insect specialization on the mass-flowering plant
There are several possible nonexclusive explanations for the absence of pollinator specialization or constancy on R. ferrugineum in high-density patches: First, based on the proportion of insects captured while visiting R. ferrugineum (41.8% in HDP compared to 8.5% of links inferred from metabarcoding), one could have concluded that pollinator specialization was greater. Thus, specialization was F I G U R E 1 Change in linkage degrees from community to individual pollinators. L sp and L i : species and individual linkage degrees; L sp (80%) and L i (80%) : minimum number of links required to gather 80% of the pollen load of a pollinator species and of each individual, respectively. Boxand-whisker plots represent values for all combinations of patches × species that Diptera usually prefer simpler shaped actinomorphic flowers with exposed rewards Stanley & Stout, 2014) whereas there could be a trait mismatch between the tubular weakly zygomorphic R. ferrugineum corolla and the fly's buccal apparatus, (b) a decrease in pollinator density (mostly Diptera) on R. ferrugineum and an increase in the surrounding community with an increase in shrub density (Delmas et al., 2014). Our results highlight the fact that a native mass-flowering plant may play a less central role and be less dominant in networks than crops or supergeneralist alien plants (Lopezaraiza-Mikel et al., 2007;Magrach et al., 2017;Stanley & Stout, 2014;Tiedeken & Stout, 2015;Vilà et al., 2011). However, more studies are required to assess the extent to which our findings can be generalized to other mass-flowering species. Long-term evolutionary and filtering mechanisms assembling complementary plant species and/or species developing positive interactions (Ghazoul, 2006)

| The increase in insect specialization in lowdensity patches
Individual and species pollinators generally tended to reduce generally their niche breadth and have higher species diet segregation in response to the reduced floral displays (Bolnick et al., 2003). In response to the decrease in local resource availability, individual Empididae tended to extend exclusive interactions to rare plant species (higher d' insect) that differed from those used by their con- suggest that rare plant species will always be visited and will sustainably coexist with widespread mass-flowering plants. Spatiotemporal niche complementarity of pollinators is believed to be one of the main drivers of the relationship between insect diversity and pollination efficiency (Albrecht, Schmid, Hautier, & Müller, 2012;Fontaine et al., 2008). Taxonomic differences in response to floral changes may also reinforce this relationship and hence contribute to network resilience in the face of environmental changes. Although we were unable to provide direct evidence for competition, its involvement in the observed pattern is highly probable. Indeed, the decrease in pollinator abundance from high-density to low-density patches was much lower than that of floral availability. Consequently, at the same site, Delmas et al. (2014) found an increase in pollinator density on isolated R. ferrugineum shrubs and smaller insect pollen loads, suggesting stronger competition for pollen and possibly for nectar rewards. Furthermore, the fact that Polygala calcarea was not visited and therefore could not compensate for the decrease in the massflowering plant likely increased the exploitative competition in the low-density patch LDP1. The lower degree of linkage and the higher dietary segregation of insect species in low-density patches are consistent with the widely shared view that coexisting species diverge in F I G U R E 3 Effects (mixed linear model) of Rhododendron ferrugineum floral density on species and individual food niches and networks. Black lines show the independent effects of floral density, and the colored lines show family-floral density interaction effects. The size of the symbols is proportional to the size of the sample resource use to mitigate the consequences of interspecific competition (Araújo et al., 2011;Inouye, 1978;Van Valen, 1965). Therefore, the observed pattern could represent the net ultimate balance between the diversifying effect (individual specialization in various plant species) of intraspecific competition and the constraints in interspecific competition on species generalization (Bolnick et al., 2003;Roughgarden, 1972).

| Paradox and complementarity in pollination networks
The different approaches we used in this study reveal a quite para-

| Potential limits of metabarcoding
There is now enough evidence that metabarcoding is a reliable way to characterize species composition in environmental DNA samples (Ji et al., 2013) including pollen samples (Bell et al., 2018).
Further, several studies have highlighted the potential of metabarcoding for quantification of the abundance of plant species in pollen mixtures especially with plastid markers (Kraaijeveld et al., 2015;Richardson et al., 2019), whereas other markers (most often ITS2, not used in our study) were apparently less efficient (Bell et al., 2018;Richardson et al., 2019). Nonetheless, some potential biases (Deiner et al., 2017) may alter quantification accuracy.
First, a plant species-molecular marker interaction is often observed during PCR, so that a species may be properly amplified by one marker but not by another one. Such amplification failures may lead to underrepresentation of the species concerned in sequencing products and to subsequent proportional over-representation of the other species (Bell et al., 2018), thus increasing specialization level in networks. In our study, the risk of species underrepresentation (and hence the risk of proportional over-representation of other species) was limited since we combined two markers and, for every plant species, only kept the results of the marker whose amplification was the most successful . Second, the composition and relative abundance of species in multispecies pollen mixtures may also hinder quantification (Bell et al., 2018) but, at least with our protocol, only partially . Third, authors who tested metabarcoding quantification used natural (bee corbicula) or artificial mixtures with huge pollen quantities (Bell et al., 2018;Cornman et al., 2015;Galimberti et al., 2014;Keller et al., 2015;Richardson et al., 2019;Smart et al., 2017), which we did not. High pollen quantities may increase concentrations of endogenous inhibitors during PCR due, for instance, to high CG contents in ITS sequences (Mammedov et al., 2008). Fourth, the abundance of plant species in insect pollen loads could to some extent, depend on the rate of plant-to-insect pollen transfers per visit. Therefore, despite the precautions we took, we cannot totally rule out the impact of biological/manipulation biases in our study. However, using the same raw data set as in the present study, we previously found highly significant positive correlations between the number of insect visits to a plant species and the number of its sequences in sequencing products . This was the case despite the fact that visits to flowers does not translate always into pollen transport by insects (Popic, Wardle, & Davila, 2013). Finally, as long as the above biases remained consistent across floral patches (we cannot see why this should not be the case) they could not be responsible for the higher pollinator specialization in low-density patches than in high-density patches.

CO N FLI C T O F I NTE R E S T
None declared.

AUTH O R ' S CO NTR I B UTI O N S
AP supervised the study. AP, NE, MB, and CA collected the data; NE and SB performed laboratory works. CA, SB, and AP analyzed the data.
AP led the writing of the manuscript. All authors contributed critically to the drafts and gave their final approval for publication.

DATA AVA I L A B I L I T Y S TAT E M E N T
Nucleotide sequences have been published on GenBank (Accession numbers KU974005-KU974022, KU974024-KU974083).