Construction, validation, and application of nocturnal pollen transport networks in an agro‐ecosystem: a comparison using light microscopy and DNA metabarcoding

1. Moths are globally relevant as pollinators but nocturnal pollination remains poorly understood. Plant–pollinator interaction networks are traditionally constructed using either flower‐visitor observations or pollen‐transport detection using microscopy. Recent studies have shown the potential of DNA metabarcoding for detecting and identifying pollen‐transport interactions. However, no study has directly compared the realised observations of pollen‐transport networks between DNA metabarcoding and conventional light microscopy.


Introduction
Species interaction networks, which describe the presence and strength of interspecific interactions within ecosystems (Montoya et al., 2006), are an important tool for understanding and conserving ecosystem processes and functioning (Tylianakis et al., 2010). Currently, there is considerable interest in pollination networks as a result of ongoing global declines in pollinating insects (Potts et al., 2010) and their role in reproduction of both wild plants and crops (Klein et al., 2007;Ollerton et al., 2011).
Many flower-visiting animals are not effective pollinators and so confirming the existence of an effective pollination interaction is labour intensive (King et al., 2013). Consequently, proxies for pollination are often used to construct plant-pollinator interaction networks, which cannot strictly be referred to as pollination networks. A commonly-used proxy is flower-visitation, recorded by directly observing animals visiting flowers. This is effective for daytime sampling, although it is challenging to apply to nocturnal pollinators, such as moths (Lepidoptera; Macgregor et al., 2015) because observations are difficult and may be biased if assisted by artificial light. This may explain why plant-pollinator network studies frequently omit nocturnal moths, even though moths are globally relevant pollinators (Macgregor et al., 2015).
An alternative to direct observation is detecting pollen transport, by sampling and identifying pollen on the bodies of flower-visiting animals; this approach has been used in several previous studies of nocturnal pollination by moths (Devoto et al., 2011;Banza et al., 2015;Knop et al., 2017;Macgregor et al., 2017a). By analysing pollen transport, flower-visits where no pollen is received from the anthers are excluded (Pornon et al., 2016). This approach can detect more plant-pollinator interactions with lower sampling effort than flower-visitor observations (Bosch et al., 2009). Studies of pollen transport also permit unbiased community-level sampling of interactions without requiring decisions about distribution of sampling effort among flower species because each pollinator carries a record of its flower-visiting activities in the pollen on its body (Bosch et al., 2009). Traditionally, pollen identification is undertaken using light microscopy with a reference collection of known species (Devoto et al., 2011). However, identifications made by microscopy can be ambiguous, especially when distinguishing related species (Galimberti et al., 2014). Accurate, reproducible identification of pollen sampled from pollinators is necessary to ensure plant-pollinator networks are free from observer bias.
A recent alternative to microscopy is DNA metabarcoding: high-throughput sequencing of standard reference loci from communities of pooled individuals (Cristescu, 2014). It offers possibilities to detect interspecific interactions, including plant-pollinator interactions (Evans et al., 2016), and methods are rapidly improving, permitting greater accuracy in species identification (Bell et al., 2016a) with respect to reducing costs (Kamenova et al., 2017). Studies using metabarcoding have identified pollen sampled from honey (Hawkins et al., 2015;de Vere et al., 2017), as well as directly from bees (Galimberti et al., 2014) and flies (Galliot et al., 2017), and constructed plant-pollinator networks (Bell et al., 2017;Pornon et al., 2017). DNA sequences have confirmed identities of single pollen grains sampled from moths (Chang et al., 2018), although no study has applied metabarcoding to nocturnal pollen-transport by moths, where pollen-transport approaches may be most valuable, given the paucity of existing knowledge about moth-plant pollination interactions. Metabarcoding reveals more plant-pollinator interactions than direct flower-visitor observations (Pornon et al., 2016(Pornon et al., , 2017, although it is unclear whether this is purely because pollen-transport approaches detect interactions more efficiently than flower-visitation approaches (Bosch et al., 2009) or whether metabarcoding offers specific additional benefits. The use of a metabarcoding approach is often justified by the labour-intensive nature of microscopy-based approaches and the level of expertise required to identify pollen morphologically (de Vere et al., 2017). It is frequently suggested that metabarcoding increases the level of species discrimination compared with traditional approaches (Bell et al., 2017). Crucially, despite this assertion, no study has directly compared metabarcoding to traditional microscopy for assessing pollen transport. It is therefore unknown whether, in studies using a pollen-transport approach, the choice of detection method (light microscopy or DNA metabarcoding) can alter the realised observations of plant-pollinator interactions.
In the present study, we used matched samples of moths to construct nocturnal pollination networks using two methods: DNA metabarcoding, and the traditional light microscopy approach. The observed networks were compared considering the quantity and nature of the interactions detected and the properties of the networks themselves. We sampled moths in a U.K. agro-ecosystem because our previous study suggests that moths may have greater importance as pollinators in such systems than generally thought (Macgregor et al., 2017a). Accordingly, we developed existing pollen-metabarcoding protocols to enable detection of pollen transported by moths, and integrated molecular advances with ecological network analysis to provide a reproducible methodology for the improved study of species interactions. We present a framework for future studies of pollination networks using metabarcoding, by providing detailed descriptions of our methods and archiving all bioinformatic and statistical code. We discuss the advantages and disadvantages of each method for assessment of pollen transport by moths and other pollinator taxa, current limitations, and future research directions.

Field sampling
We sampled moths, using light traps, from four locations in a single farmland site in the East Riding of Yorkshire, U.K. (53 ∘ 51 ′ 44" N 0 ∘ 25'14" W), over eight nights between 30 June and 19 September 2015 (see Supporting information, Table S1; full details in Appendix S1). Moths were euthanised and retained individually. Because both pollen-sampling methods are destructive, it was impossible to directly compare sensitivity by sampling pollen from the same individual moth with both methods. Instead, we created two matched sub-samples of 1.

Fig. 1.
Visual summary of the two methods applied to detect and identify pollen transport by moths. Full methods are provided in the Supporting information (Appendix S1). For metabarcoding, the steps shown are: 1, field sampling of moths; 2, excision of proboscis; 3, removal of pollen by shaking; 4, extraction of DNA by the HotSHOT method; 5, amplification of DNA by a three-step polymerase chain reaction protocol; 6, sequencing of DNA; 7, assignment of DNA sequence identities; 8, analysis of interactions and construction of networks. For microscopy, the steps shown are: A, field sampling of moths; B, swabbing of proboscis with fuchsin-stained gel; C, mounting gel on microscope slide; D, identification and counting of pollen under microscope; E, analysis of interactions and construction of networks.
[Colour figure can be viewed at wileyonlinelibrary.com]. moths, each containing the same set of species, as well as the same number of individuals of each. Pollen-transport by each sub-sample was analysed using one method (Fig. 1). With both methods, we restricted pollen sampling to the proboscis because most moth species coil their proboscides unless actively feeding (Krenn, 1990). Therefore, the proboscis is unlikely to experience cross-contamination of pollen through contact with other moths (e.g. when in the moth-trap), and pollen held on the proboscis is probably the result of a flower-visitation interaction.

Method 1: Light microscopy
A standard approach for pollen sampling was applied (Beattie, 1972) in which cubes (1 mm 3 ) of fuchsin jelly were used to swab pollen from the proboscides of moths, and the pollen was examined under a light microscope at 400× magnification. Pollen morphotypes were identified using a combination of keys (Moore et al., 1994;Kapp et al., 2000) and knowledge of likely insect-pollinated plant taxa. Morphotypes (equivalent to operational taxonomic units, OTUs) represented groupings that could not be unambiguously separated to a lower taxonomic level, and might have contained pollen from multiple species.

Method 2: DNA metabarcoding
Protocols for DNA extraction, amplification and sequencing are fully described in the Supporting information (Appendix S1) and archived online (dx.doi.org/10.17504/protocols.io .mygc7tw). In brief, moth proboscides were excised using a sterile scalpel. Pollen was then removed from each proboscis by shaking for 10 min in HotSHOT lysis reagent (Truett et al., 2000) at 2000 r.p.m. on a Variomag Teleshake plate shaker (Thermo Scientific, Waltham, MA, U.S.A.). The proboscis was removed using sterile forceps, and the DNA extraction procedure completed on the remaining solution as described by Truett et al. (2000). Extracted DNA was amplified using a three-step polymerase chain reaction nested tagging protocol (modifed from Kitson et al., 2018) (see Supporting information, Appendix S1). We amplified a custom fragment of the rbcL region of chloroplast DNA, which has been previously used for metabarcoding pollen (Hawkins et al., 2015;Bell et al., 2017) and has a comprehensive reference library for the Welsh flora, representing 76% of the U.K. flora (de Vere et al., 2012), available via the International Nucleotide Sequence Database Collaboration (http://www.insdc.org; GenBank). We used two known binding sites for reverse primers, rbcL-19bR (Hofreiter et al., 2000) and rbcLr506 (de Vere et al., 2012), to produce a working forward and reverse universal primer pair, rbcL-3C (rbcL-3CF: 5 ′ -CTGGAGTTCCGCCTGAAGAAG-3 ′ ; rbcL-3CR: 5 ′ -AGG GGACGACCATACTTGTTCA-3 ′ ). Primers were validated by successful amplification of DNA extracts from 23/25 plant species (see Supporting information, Table S2). Sequence length varied widely (median: 326 bp, range: 96-389 bp); fragments shorter than 256 bp generally had no match on GenBank. Six control samples were used to monitor cross-contamination between wells (see Supporting information, Table S3).
Amplified DNA was sequenced on an Illumina MiSeq (Illumina, Inc., San Diego, CA, USA), using V2 chemistry. Taxonomic assignment of MiSeq output was conducted using the metabeat pipeline, version 0.97.7 (https://github.com/ HullUni-bioinformatics/metaBEAT). For reproducibility, all steps were conducted in Jupyter notebooks (http://jupyter.org); all bioinformatic and statistical code used in this study is archived online (dx.doi.org/10.5281/zenodo.1322712) and procedures are explained in full in the see Supporting information (Appendix S1). Taxonomic assignment of sequences was conducted within metabeat based on a blast lowest common ancestor approach similar to the one implemented in megan (Huson et al., 2007). We chose to conduct taxonomic assignment with blast because it is among the most widely-used taxonomic assignment tools and blastn specifically has a proven capacity to discriminate between U.K. plant species using the rbcL locus (de Vere et al., 2012). We used a curated database of reference sequences from plausibly-present plant species previously recorded in the vice-county of South-east Yorkshire (reference list of species archived at dx.doi.org/10 .5281/zenodo.1322712). Verbena officinalis Verbena officinalis 1 -0 †An assignment that was initially identified by metabarcoding but failed to meet the minimum read depth threshold in any sample (see Supporting information, Table S7).
‡An assignment that was re-identified by comparison with pollen of species identified by metabarcoding.
To eliminate the risk of cross-well contamination, we established a threshold for minimum read depth of 50 reads, per assignment, per well. The maximum read depth in any negative control well was 47, and the maximum read depth in any positive control well of sample assignments was 33 (see Supporting information, Table S3). Therefore, this threshold was adequate to remove sample reads from positive and negative controls. Within each well, any assignment with a read depth below 50 was reset to 0 prior to statistical analysis; this resulted in some plant OTUs being removed entirely from the dataset (however, these OTUs are indicated in Table 1).

Curation of data
We harmonised the plant identifications from each method (OTUs from metabarcoding and morphotypes from microscopy) to produce a single list of plants consistent across both methods (Table 1). Specifically, for metabarcoding, we revised family-level assignments made by blast, inspecting the range of species-level matches to identify clear taxonomic clusters within the families. For microscopy, we attempted to re-identify pollen morphotypes using images of pollen from species identified by metabarcoding for additional reference (see Supporting information, Appendix S1). Microscopic photographs of pollen were sourced from two online repositories of pollen images: Pollen-Wiki (http://pollen.tstebler.ch/MediaWiki/index .php?title=Pollenatlas) and the Pollen Image Library (http:// www-saps.plantsci.cam.ac.uk/pollen/index.htm).

Comparison of methods and statistical analysis
We tested for differences between the two identification methods, examining whether the sampling method affected the  (155) 52 (88) The samples were duplicate subsets of the total sample, and each comprised 311 individuals of 41 species. Plant types for metabarcoding were operational taxonomic units (OTUs; identified by a blast search against a curated reference database) and, for microscopy, were morphotypes (identified using identification keys). Percentages in brackets are of the relevant sub-sample.
likelihood of detecting (i) pollen on individual moths; (ii) more than one pollen species on individuals; (iii) pollen on moth species (individuals combined); and whether sampling method affected the number of pollen types detected (iv) per individual moth; and per moth species, using (v) observed richness and (vi) true richness estimated with the chao2 estimator (Chao, 1987). We used generalised linear mixed-effects models, with sampling method as a fixed effect. In individual-level analyses, we used date/light trap combination ('trap ID') and species as crossed random effects, whereas, in species-level analyses, we used moth species as a random effect to treat the data as pairs of observations (one observation, per method, per moth species). We tested significance of fixed effects using either likelihood ratio tests, for models with a binomial or Poisson error distribution, or type III analysis of variance, for models with a quasi-Poisson error distribution (error distributions used in each model are detailed in Table S4). Analysis was carried out using r, version 3.3.2 (R Core Team, 2016); all code is archived at: dx.doi.org/10.5281/zenodo.1322712.

Sampling completeness and networks
For both methods, we estimated sampling completeness of interactions, as described by Macgregor et al. (2017b). For each method, we estimated the total number of pollen types (interaction richness) for each insect species with the chao2 estimator (Chao, 1987), using the r package vegan (Oksanen et al., 2015). We calculated interaction sampling completeness for each species as 100*(observed interactions)/(estimated interactions) for each species. Finally, we calculated the mean interaction sampling completeness of all species, weighted by estimated interaction richness of each species.
We constructed pollen-transport networks from the interaction data. We used the presence of interactions between individual moths and plant taxa, rather than strength of individual interactions, because read depth (metabarcoding) and pollen count (microscopy) do not correlate between plant species (Pornon et al., 2016). We measured interaction frequency by counting interactions across all individuals in each moth species; interaction frequency correlates positively with true interaction strength in mutualistic networks (Vázquez et al., 2005). We calculated several quantitative metrics to describe the diversity and specialisation of interactions forming each network. Improved detection of interactions could increase the complexity of the network, and so we calculated two measures of network complexity: linkage density (average number of links per species) and connectance (proportion of possible interactions in the network that are realised). Likewise, improved detection of plant species with the same set of pollinator species could alter consumer-resource asymmetry and perceived specialisation of species in the network, and so we calculated H2 ′ (a frequency-based index that increases with greater specialisation), as well as generality of pollinators and plants (average number of links to plant species per pollinator species, and vice versa). Finally, the resilience of the network to cascading species loss may be influenced by its complexity (Dunne et al., 2002) and so we measured the robustness of each network (mean robustness across 1000 bootstrapped simulations of pollinator species loss). For comparison, we repeated all network analyses both (i) with plant identities aggregated at family level because the methods might differ in their ability to distinguish closely-related species and (ii) excluding all species of moth for which only one individual was sampled with each method because the influence of such singletons on network metrics could potentially be sufficiently large to bias our findings. Networks were analysed using the package bipartite (Dormann et al., 2009) and plotted using food web designer, version 3.0 (Sint & Traugott, 2016). Because we could only construct one network for each method, we recorded obvious differences between the metrics for each network but could not statistically assess the significance of those differences.

Summary
In total, we caught 683 moths of 81 species, generating two matched sub-samples, each containing 311 moths of 41 species (see Supporting information, Table S5). We detected pollen on 107 individual moths by metabarcoding (34% of the sub-sample) and 70 (23%) by microscopy (Table 2). We initially identified 20 plant morphotypes in the microscopy sample and 25 OTUs in the metabarcoding sample (Table 1). After harmonising these we recorded 33 plant identities (at varying taxonomic resolution), of which 18 were detected by both methods, 11 by metabarcoding only (including three which failed to meet the minimum read depth threshold in any sample), and four by microscopy only (Fig. 2).

Statistical comparisons between methods
Metabarcoding was significantly more likely than microscopy to detect pollen (Fig. 3) on individual moths ( 2 = 10.95, P < 0.001), as well as to detect more than one pollen type on individual moths ( 2 = 12.00, P < 0.001). However, with non-pollen-carrying moths excluded, the methods did not differ in the number of pollen types detected per individual moth ( 2 = 1.12, P = 0.290). With data aggregated per moth species, the methods did not differ in the likelihood of detecting pollen ( 2 = 0.37, P = 0.545), although metabarcoding detected significantly more pollen types per moth species ( 2 = 18.09, P < 0.001); this difference was non-significant when the estimate of true interaction richness was used ( 2 = 3.62, P = 0.057) (see Supporting information, Table S4).

Construction and analysis of networks
For each method, we constructed a quantitative pollen-transport network (Fig. 4). The estimated sampling completeness of interactions was higher for the microscopy network (75.7%) than the metabarcoding network (43.2%). Some network metrics differed markedly between the two methods (Fig. 5), although no statistical comparison was appropriate. Specifically, linkage density and generality of pollinators were higher in the metabarcoding network than the microscopy network, although all other metrics were similar. With plant assignments aggregated at family level, the metabarcoding network had higher generality of pollinators and lower generality of plants than the microscopy network (see Supporting information, Table S6). The difference between network metrics calculated with and without species of moth for which only one individual had been sampled was negligible in all cases (see Supporting information, Table S6), indicating that these singletons did not bias our results.

Methodological comparison
Our realised observations of the plant-pollinator system were generally similar between the DNA-based (metabarcoding) and microscopy-based methods for detecting and identifying pollen-transport by moths, but, nonetheless showed some key differences. Metabarcoding detected more pollen OTUs in total than microscopy, detected pollen on a greater proportion of individual moths, and was more likely to detect multiple pollen OTUs on a moth. When moths were aggregated to species level, metabarcoding detected more pollen types in total per moth species. These differences were most likely because metabarcoding had a greater ability to separately closely-related or morphologically similar pollen into multiple identities, and possibly also because the pollen capture technique for metabarcoding (shaking the whole proboscis in extraction buffer) is likely to be more efficient than the equivalent for microscopy (swabbing the proboscis with sticky gel), allowing a greater proportion of each moth's pollen load to be removed and analysed with the metabarcoding approach. Pollen capture by shaking, as used for Significance indicates likelihood ratio test for detection method in generalised linear mixed-effects models (*P < 0.05; **P < 0.01; ***P < 0.001). Error bars indicate 95% confidence intervals.
the metabarcoding approach (Fig. 1), cannot be readily adapted for a microscopy approach because collecting pollen grains from a liquid rinse for subsequent mounting on a microscope slide would not be practical. We also observed differences between the networks detected by each method. There was higher linkage density in the fully-resolved metabarcoding network than its equivalent microscopy network, although no difference in linkage density between the two networks when plant identities were aggregated at the family level (Fig. 5). This provides further evidence for the greater ability of metabarcoding to separate closely-related plant identities within families, resulting in the detection of more interactions using this approach than using microscopy. Additionally, there was higher generality of pollinators in the fully-resolved metabarcoding network than its equivalent microscopy network, whereas when plant identities were aggregated at family level, generality of pollinators was higher to a lesser degree in the metabarcoding network, but generality of plants was lower in the metabarcoding network than in the microscopy network (Fig. 5). This indicates that the metabarcoding approach detected interactions with more plant families per pollinator species, which may have been because metabarcoding had a greater ability to separate morphologically similar pollen from different families, or simply because metabarcoding detected more plant OTUs per pollinator species (Fig. 3).
Estimated sampling completeness of interactions differed conspicuously between networks (see Supporting information, Table S6). Despite containing more interactions, the metabarcoding network was estimated to be less completely sampled than the microscopy network. This is probably because metabarcoding detected more 'rare' interactions ('singletons', detected only once), being more effective at distinguishing morphologically similar pollen. This would result in a higher ratio of singletons to doubletons (interactions detected twice) and therefore a proportionally greater estimated value of interaction richness. This demonstrates that sampling method can substantially affect estimation of sampling completeness of interactions in network studies.

Pollen transported by moths
We identified several plants using metabarcoding that were not initially identified as the same species by microscopy. Because many plants have morphologically similar pollen, we conservatively chose not to identify novel moth-flower associations by microscopy unless the identification was unambiguous. Among the plants initially identified only by metabarcoding were species for which moths were not previously recorded in the literature as pollinators or flower-visitors (Macgregor et al., 2015), highlighting that much still remains unknown about pollination by moths. Some of these fitted the moth-pollination 'syndrome' (Grant, 1983), being white and fragrant: Sambucus nigra (Adoxaceae), Philadelphus coronarius (Hydrangeaceae), Filipendula ulmaria (Rosaceae), and Ligustrum vulgare (Oleaceae; although not Syringa vulgaris; not separable in the present study). However, others did not and are typically associated with other pollinators: for example, Polemonium caerulum (Polemoniaceae) and Trifolium spp. (Fabaceae) are visited by bees (Palmer-Jones et al., 1966;Zych et al., 2013), Verbena officinalis (Verbenaceae) is most likely visited by bees and butterflies (Perkins et al., 1975), whereas species of Epipactis (Orchidaceae) are generalist, with previously-known visitors including diurnal Lepidoptera (Jakubska-Busse & Kadej, 2011).
We found pollen from plants that, in this region, are chiefly associated with domestic gardens, including two species of Hydrangeaceae, species from the tribe Mentheae (Lamiaceae; includes many species grown as culinary herbs, although wild species might also have occurred), Buddleja davidii  Table S6). Solid lines connect metrics for fully-resolved data; dashed lines connect metrics when plant species were aggregated at the family level.
(Scrophulariaceae; although a railway ran adjacent to the farm and B. davidii is widely naturalised along railways in the U.K.), and Verbena officinalis (Verbenaceae). Individual moths may have carried pollen several hundred metres from the closest gardens to the field site. This provides new evidence to support previous suggestions that moths could play an important role in providing gene flow among plant populations at the landscape-scale (Miyake & Yahara, 1998;Young, 2002;Barthelmess et al., 2006), and even at continental scales for species of moths that undergo long-distance migrations (Chang et al., 2018). Such gene flow could provide benefits from nocturnal pollination even to plant species that are primarily diurnally-pollinated and not pollination-limited.
Finally, we detected several insect-pollinated crop species (only some of which require pollination for crop production): specifically, soybean Glycine max and pea Pisum sativum (Fabaceae), potato Solanum tuberosum (Solanaceae), and Brassica/Raphanus sp. (includes oil-seed rape; Brassicaceae). Floral phenology suggests that Prunus sp. (Rosaceae) was likely to be cherry (Prunus avium, Prunus cerasus or a hybrid) rather than wild Prunus spinosa. Similarly, Rubus sp. (Rosaceae) could have been wild blackberry (matching to Rubus caesius, Rubus plicatus, and Rubus ulmifolius), although it also matched raspberry Rubus idaeus. There is currently an extreme paucity of evidence in the existing global literature to support a role of moths in providing pollination services by fertilising economically-valuable crops (Klein et al., 2007;Macgregor et al., 2015). Although our findings do not confirm that any of the crops recorded receive significant levels of nocturnal pollination by moths, they do highlight a vital and urgent need for further research into the potential role of moths as pollinators of agricultural crop species.

Current methodological limitations
We identified limitations with both methods, relating to the accuracy and taxonomic resolution of pollen identification and the non-quantitative interaction data that they generated.
First, there was little initial overlap between identifications made by each method (of 20 initial assignments from microscopy and 25 from metabarcoding, only three plant identifications were shared between methods at genus or species level). Because we applied the methods to separate samples of moths, some differences were expected between the pollen species transported. In two cases (Silene and Tilia), species identified by microscopy were discarded from the metabarcoding assignments by application of the 50-reads threshold. Both species had very low abundance in microscopy samples (< 20 pollen grains per sample), suggesting precautions against cross-sample contamination with metabarcoding might mask the detection of low-abundance pollen. The remaining mismatches were most probably misidentifications by one or other method. Using images of pollen from species identified by metabarcoding as a reference for microscopy, we re-identified several pollen morphotypes, increasing agreement between the methods (19 identifications matched across methods, of which 10 were at the genus or species level) (Table 1). This indicates that creation of a reliable pollen reference collection for the field site might have improved our initial identifications made by microscopy; however, because moths can disperse (and transport pollen) over considerable distances (Jones et al., 2016), this could also have increased the risk of misidentifying pollen of a species absent from the field site (but regionally present) as morphologically similar pollen of an alternative species that was present at the field site. Misidentifications were arguably more likely under microscopy than metabarcoding, due to the conservative approach used when applying blast and the difficulty of unambiguously identifying pollen by microscopy.
Second, several assignments made with metabarcoding were not resolved beyond the family level. Although rbcL is a popular marker region for plant barcoding (Hawkins et al., 2015) and has been shown to identify over 90% of Welsh plants to at least genus level using blastn (de Vere et al., 2012), interspecific sequence diversity within rbcL is nonetheless extremely low within some families (e.g. Apiaceae; Liu et al., 2014). In some cases, reference sequences from multiple genera did not differ across our entire fragment, leading blast to match query sequences to species from several genera with equal confidence. Such instances could not have been further resolved using our fragment, even by alternative assignment methods. Sequencing a longer fragment might increase interspecific sequence variation; improvements in sequencing technology may facilitate accurate sequencing of such longer amplicons (Hebert et al., 2018). Using a locus other than rbcL might improve taxonomic resolution; loci including ITS2 and matK are also used to metabarcode pollen (Bell et al., 2016b). Sequencing two or more of these loci simultaneously might also improve assignment resolution (de Vere et al., 2012), although at a greater cost.
Third, some studies have weighted interactions in networks using the number of pollen grains transported as a proxy for interaction strength (Banza et al., 2015). This approach is impossible with metabarcoding because the number of pollen grains in a sample does not correlate with read depth (Pornon et al., 2016) and metabarcoding cannot definitively distinguish pollen from other sources of plant DNA (e.g. residual nectar on mouthparts). However, an insect's pollen load also may not be a true indicator of its efficacy as a pollinator (Ballantyne et al., 2015); pollinator effectiveness differs between pairwise interactions through variation in floral morphology, pollinator morphology and behaviour, location of pollen on the pollinator's body, and other temporal and spatial factors besides the quantity of pollen transported. Instead, interaction frequency (counting occurrences of an interaction, but disregarding individual interaction strength) predicts the relative strength of pollination interactions well (Vázquez et al., 2005), and was successfully generated with both microscopy and metabarcoding in the present study.

Merging metabarcoding and pollination network analysis
Subsequent to several recent studies that have constructed diurnal plant-pollinator networks using DNA metabarcoding (Bell et al., 2017;Pornon et al., 2017), we have further demonstrated the potential of metabarcoding by using it to construct nocturnal pollen-transport networks for the first time (Fig. 4). We provide a detailed and reproducible methodology to integrate molecular advances and ecological network analysis. Our results clearly demonstrate that the capacity of metabarcoding to generate pollen-transport interaction data is comparable with that of previously-used methods, such as microscopy. Additionally, metabarcoding may carry several practical advantages over flower-visitor observations or microscopy for studies analysing pollination networks.
One such advantage is that metabarcoding is reproducible across studies, pollinator guilds, and ecosystems. It is freed from observer biases inherent both in morphological identification of pollen, as well as in other means of detecting pollination interactions such as flower-visitor observations, where the distribution of sampling effort among flower species can affect network structure (Gibson et al., 2011) and sampling often focuses on a subset of the floral assemblage (Tiusanen et al., 2016). Metabarcoding can be conducted without system-specific expertise in morphological pollen identification, or prior knowledge about locally present plants or likely interactions (although such information can be used, if available and robust, to increase the taxonomic resolution of species identifications). Metabarcoding may reveal previously unsuspected detail in networks (Pornon et al., 2017), especially those involving moths or other under-studied pollinator taxa.
Metabarcoding may also allow more a efficient processing of samples, and therefore the analysis of larger numbers of samples, than microscopy (Fig. 6). Most pollination-network studies have focused on evaluating a single network, or a small number of networks under variant conditions (Burkle et al., 2013). Constructing multiple replicated networks across a range of treatments, sites or time points, and testing for structural differences (Lopezaraiza-Mikel et al., 2007), comprises a powerful alternative, although it can be hampered by the difficulty of generating sufficient data for multiple, well-sampled networks. For metabarcoding, investment mainly scales per plate (≤96 samples) rather than per sample (Derocles et al., 2018), whereas, for microscopy, investment of materials and especially time increases linearly for every sample, although sample processing speed might increase slightly after an initial period of learning (Fig. 6). Importantly, this increased efficiency is coupled with increased reproducibility because molecular tools treat all samples identically regardless of their complexity.
Finally, DNA metabarcoding can streamline the generation of suitable data for incorporating phylogenetic information into ecological networks (Evans et al., 2016). Recent studies have found significant relationships between phylogenetic and resource overlap in mutualistic and antagonistic networks (Rezende et al., 2007;Elias et al., 2013;Peralta et al., 2015); metabarcoding permits simultaneous generation of both interaction and relatedness data.

Conclusions
In the present study, we constructed pollen-transport networks using matched samples of moths to compare two methods used for detecting and identifying pollen: DNA metabarcoding and traditional light microscopy. We showed that the state-of-the-art DNA metabarcoding approach is capable of generating pollen-transport interaction networks that are similar to those detected using microscopy. Indeed, with metabarcoding, we detected pollen on more individual moths and detected more pollen types per moth species. These differences indicate that direct comparisons between networks constructed using metabarcoding and those constructed using traditional methods such as microscopy should be treated with the appropriate caution, although a combination of both metabarcoding and traditional methods may provide the most detailed information (Wirta et al., 2014). Metabarcoding additionally revealed a range of previously undocumented moth-plant interactions, and provided new evidence for two possible benefits of nocturnal pollination: landscape-scale provision of plant gene flow and potential provision of the pollination ecosystem service. The metabarcoding approach has considerable potential for studying pollen-transport networks and species interactions more generally. ing labwork, as well as J. Downs for assistance with fieldwork. E. Moss created the moth image used in Fig. 1.
The experiment was conceived by CJM under supervision by DME, MJOP and RF, and designed in conjunction with DHL and JJNK. Field and laboratory work was conducted by CJM with advice from JJNK. The metabeat pipeline was created by CH and metabarcoding data was processed and analysed by CJM, with advice from CH. The statistical analysis was conducted by CJM. All authors contributed to preparing the manuscript and gave final approval to the version submitted for publication.

Supporting Information
Additional supporting information may be found online in the Supporting Information section at the end of the article.
Appendix S1. Detailed description of methods Table S1. Details of moth-trapping sessions carried out to obtain samples of moths to analyse. Minimum overnight temperature and average wind speed were collected from the Met Office website, rather than measured in the field. Cloud cover indicates whether there was any cloud cover at the time of trap set-up or collection; therefore, nights for which there was cloud cover may still have experienced clear spells. Moon phase was estimated to the nearest quarter. Table S2. Details of plants used to validate novel primer combination, identified to the genus level. Table S3. Summary of negative and positive controls with a description of any contamination issues identified by each Table S4 Summary of statistical comparisons between methods (Fig. 2). 2 is from the likelihood ratio test for all error families except quasi-Poisson, when it is from the type III analysis of variance. Table S5. All moth species sampled and the number of individuals of each species assigned to each of the two subsamples were analysed with the two methods. Table S6. Summary of network descriptors calculated for the accumulated interaction data from each detection method, with assignments made both (i) to the lowest-possible taxonomic level and (ii) to family-level only, and (iii) with moth species excluded if only one individual was sampled with each method. Robustness represents the area under an extinction curve of pollinator species as plant species are removed at random, and the mean and bootstrapped 95% confidence interval (1000 simulations) is given.  Table 1. Some OTU assignments were manually revised after inspection of the log files; details are given in column 5. Table S8. Well assignments of 16 forward and 12 reverse tagged primers to create unique combinations of primers for each well, with an equal number of legal and illegal combinations.  Table S10. Species removed from the reference database by manual curation. During reference database curation, assignments made were examined on a case-by-case basis to determine whether any matches were not plausible. Species were excluded as a result of a lack of records on the BSBI Hectad Map (https://bsbi.org/maps-and-data) for the species within approximately 50 km of the field site (note that only non-cultivated species were excluded on this basis), unless indicated by: 1, shared match with related positive control; 2, no phenological overlap between flowering period and the sampling period (with a tolerance of approximately one month); 3, strict habitat requirements (e.g. sand dunes) not met in the vicinity of the field site.