Ethics statement

Our specimen collection and protocol were compliant with all local regulations and under the marine collection permit SC‐191140006, issued to Steven H.D. Haddock by the California Department of Fish and Wildlife. Since no vertebrates or cephalopods were involved, we did not need oversight from an animal care board.

Siphonophore collection

In order to sample a representative set of taxa across the siphonophore phylogeny, we targeted a set of 41 species (aiming for 10 specimens per species) including cystonects, apolemiids, pyrostephids, euphysonects, and calycophorans from shallow and deep waters. Most species were sampled from the Offshore California Current Ecosystem (OCCE) except for the Portuguese man-o-war P. physalis, which was collected off Bermuda in the Sargasso Sea; Sulculeolaria chuni and some Nanomia spp. (labeled as “Atlantic”) which were collected off Rhode Island in the Block Island sound; Forskalia sp. M123-SS8 and shallow Nanomia sp. KiloMoana2018-BW7-4 which were collected off the coast of Hawaii. While all the Nanomia populations sampled in this study have been referred to as Nanomia bijuga, we suspect that there may be undescribed cryptic Nanomia species among the specimens sampled based on the disparate tentillum morphologies we observed. Therefore, we decided to have them labeled at the genus level. One Nanomia specimen (KiloMoana2018−BW7−4) was collected off the coast of Kona, HI. The pleustonic (surface floating) Physalia physalis samples were collected manually using a bucket from a small boat. Species found between the 0-20m deep were collected using blue water diving techniques following the guidelines in Haddock & Heine [28]. Species from 200-4000m were collected using ROVs. All animals were collected live and brought back to the ship (or field station in Bermuda for P. Physalis) for dissection (Fig 1). Live colonies were photographed (sometimes recorded on video), and zooids of diagnostic value (nectophores, bracts, tentacles) were dissected, when possible, fixed in 4% formalin, and stored as vouchers at the Yale Peabody Museum of Natural History (voucher catalog numbers provided in specimen metadata S15 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Gut content metabarcoding workflow used in this study.

Siphonophore colony illustrated by Freya Goetz. Silhouettes in the plankton net downloaded from phylopic.org. Solid arrows indicate physical material transfer and processing, dashed lines indicate information transfer and processing. Yellow islands indicate elements processed in the laboratory bench, green islands represent bioinformatic datasets processed in the high-performance computing cluster, and red islands represent curated data products.

https://doi.org/10.1371/journal.pone.0267761.g001

Gut content metabarcoding

Shortly after collection of the live specimens, we dissected and pooled several gastrozooids from each colony, making sure that those with visible gut contents are included in addition to several other without conspicuous prey, and also including visible egested food pellets at the bottom of the sampling container. This non-random approach was aimed at increasing our prey detection rate, but may have introduced sampling bias against inconspicuous prey items, thus obscuring any meaningful quantitative analyses of the data. Nonetheless, we still gave inconspicuous prey a broad chance of being represented, since the majority of the gastrozooids we sampled lacked conspicuous prey content. We believe that pooling multiple gastrozooids as a single sample is reasonable, since all gastrozooids in a colony share an actively-flowing, interconnected gastrovascular cavity. Thus, we expected DNA from one prey capture in one gastrozooid to be present within multiple gastrozooids. For small species (such as Sphaeronectes spp.), we sampled the whole siphosome (whole colony excluding nectophores) as a pooled collection of all gastrozooids. Samples were frozen at -80°C until DNA extraction.

To extract DNA, we digested the samples with proteinase K at 56°C for 1-2h, and used the DNeasy Blood & Tissue kit (Qiagen, Hilden, Germany) eluting twice at 56°C for 10min into a final volume of 100μl. For barcode amplification, we used a set of six primer pairs that amplify six barcode regions within the 18S gene (‘V3’, ‘V5-V7S’, ‘V5-V7L’, ‘V7’, ‘V7p+V8’, and ‘V9’). The primers were designed using Geneious 11.1.5 [29], constraining the search to short (>300 bp) amplicon products with a high chance of remaining uncleaved after digestion in the gastrozooid, flanked by priming sites conserved (to a maximum mismatch of 3bp) across metazoans. The search for conserved priming sites was conducted on an alignment of 18S genes from 975 species across all metazoan phyla downloaded from GenBank (available in github.com/dunnlab/siphweb_metabarcoding/Primer_design). The primer search was optimized to only retrieve non-degenerate primer pairs with compatible annealing temperatures and without problematic dimerization and hairpin temperatures. Primer sequences are shown in Table 1, and their properties can be found in Table T1 in the protocol (dx.doi.org/10.17504/protocols.io.5qpvo57o7l4o/v2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Barcodes used in this study.

https://doi.org/10.1371/journal.pone.0267761.t001

Using these primer pairs, we ran six parallel PCR reactions for each successful extraction, selecting only those which had yielded a DNA concentration above 10ng/l. For each 25μl reaction volume, we used 2 μl of extraction template, 0.5 μl of each primer (at a 10μM concentration), 1μl of BSA, 0.2 μl of GoTaq (Promega, Madison, WI, USA) polymerase and the standard reagents and proportions of the GoTaq kit (Promega, Madison, WI, USA). The thermal cycles included an initial denaturation at 94°C for 2 min, followed by 30 cycles of denaturation at 94°C for 30 s, annealing (variable), and elongation at 72°C for 1 min, followed by final elongation at 72°C for 5 min. For barcode V9, we used an annealing temperature of 48°C for 45s per cycle. For all other barcodes, we used an annealing temperature of 54°C for 60s per cycle. Each batch of reactions for each barcode included a positive and a negative control (the elution buffer used in extraction), and the products were visualized using gel electrophoresis (2% agarose gel dyed with SYBR Safe DNA Stain) to check for amplicon size and monitor the controls. The PCR products were purified using ExcelaPure UF PCR Purification Plates (EdgeBiosystems, Gaithersburg, Maryland, USA). The DNA yield of each purified product was assessed using a Qubit 2.0 fluorometer and the dsDNA High Sensitivity assay (Thermo Fisher Scientific, USA). Purified PCR products from each barcode region for each sample were combined in equimolar pools based on their DNA yield in order to have equal representation in the sequencing lane. Further details on the quality control, PCR mix, and amplicon pooling are fully described in the online protocol (dx.doi.org/10.17504/protocols.io.5qpvo57o7l4o/v2). All molecular bench work was carried out at the Yale DNA Analysis Facility. Amplicon pools were sequenced using Illumina MiSeq (Illumina, San Diego, CA, USA) 250bp paired-end technology (except samples from specimens KiloMoana2018-BW7-4 Nanomia sp., D1019-D5 undescribed physonect, D856-SS8 Stephanomia amphytridis, D861-D12 Bargmannia amoena, D858-D6 Apolemia lanosa, and D860-D6 Erenna sirena which were sequenced using Illumina MiSeq 150bp) at the Yale Center for Genomic Analysis.

Prey reference database

In order to enhance the accuracy of the taxonomic assignments of reads, we also built an 18S gene barcoding database of potential prey items to expand on the available reference sequences in public databases. To do this, we collected 60 specimens of 30 species of zooplankton and micronekton from the OCCE using a Tucker trawl. We targeted plausible prey species from motile open-ocean taxa that cohabitate with siphonophores and are underrepresented in SILVA databases, including fishes, crustaceans, jellyfishes, urochordates, chaetognaths, polychaetes, and mollusks. Specimens were photographed alive, then tissue was sampled and frozen, and finally the rest of the animal was fixed in formalin as a voucher to be identified and preserved at the Yale Peabody Museum of Natural History. DNA extraction, quality control, PCR, and amplicon cleanup was carried out in a similar fashion as the metabarcoding protocol described above (and detailed in dx.doi.org/10.17504/protocols.io.5qpvo57o7l4o/v2), using the PCR program with an annealing temperature of 54°C, and a single pair of primers (166F and 134R), spanning the full extent of the sequence containing all barcode regions used in the gut content metabarcoding (from V3 to V9). Purified amplicons were sent in plates with the forward and reverse primer separately for Sanger sequencing from both ends at the Yale DNA Analysis Facility. A total of 89 newly-submitted sequences were then assembled and trimmed at a 95% quality cutoff in Geneious and concatenated with the latest SILVA database (SILVA_138_SSURef_NR99 downloaded on February 23, 2021) pruned to remove non-eukaryotic sequences.

Bioinformatic pipeline

Amplicon libraries were demultiplexed by primer sequence using custom bash code. Primer sequences were removed using cutadapt [31]. The forward and reverse reads were matched and repaired using bbtools [32], then denoised and de-replicated using the DADA2 [33] plugin in QIIME2 [34] with a truncation quality threshold of 28. We de novo clustered the unique features into operational taxonomic units (OTUs) using the VSEARCH [35] plugin in QIIME2 with a similarity threshold of 95%. To reduce computational load, only the top 100 most abundant features among the clustered OTUs were selected for taxonomic assignment. Taxonomic identities were assigned using the assignment software METAXA2 [36] with a 70% reliability cutoff, comparing the sequences against the SILVA123.1 reference library [37], and against our custom-built library built using SILVA138 as a foundation. The SILVA123.1 database contains 61383 eukaryotic reference sequences, while our custom database (built off SILVA138.1) contains 79044. Animals in the SILVA123.1 taxonomy are annotated to the ranks of superphylum, phylum, subphylum, class, subclass, order, family, genus, and species. However, the SILVA138.1 animal taxonomy was annotated at the levels of clade (e.g. Bilateria, Protostomia, Deuterostomia, Ecdysozoa, Lophotrochozoa), phylum, class, subclass, order, suborder, and species. All bioinformatics analyses were carried out in the Yale High Performance Computing Cluster. The taxonomic assignments and read count data were merged, then parsed to match the sample of origin and the DNA sequence they derived from. Sequence post-processing scripts can be found in the GitHub repository (https://github.com/dunnlab/siphweb_metabarcoding/Scripts).

Assignment interpretation

Different barcode regions and reference databases displayed different assignment sensitivities for different taxa. Moreover, the two reference databases were annotated at different taxonomic levels, thus revealing unequal assignment reliabilities at different phylogenetic depths. Therefore, the assignment information from different barcodes and reference databases was integrated to interpret the source and taxon of the detected reads. When the assignments from the two databases disagreed or reported suspicious (e.g. non-marine) taxa, we manually checked the sequences in NCBI BLAST. In summary, a combination of annotation database consensus, barcode region consensus, number of reads, manual BLAST checks, and natural history informed priors were used to assign these interpretations.

Taxonomic assignments were manually inspected and annotated with the interpreted consensus taxon and interpreted source (predator, prey, secondary predation, parasite, environmental, unrecognizable sequence, contamination, or cross contamination). Predator sources correspond to the siphonophore DNA from the gastrozooid. These annotations were given to OTUs with typically high read abundances, often taxonomically-assigned as siphonophores, hydrozoans, or unculturable eukaryotes. Prey sources were annotated when plausible prey taxa were assigned. We know that siphonophores can only capture prey that actively swims to trigger tentilla discharge [9, 27]. Therefore, we interpreted that DNA from non-pelagic and/or non-motile organisms cannot be sourced from dietary contributions from microorganisms, marine snow, eggs, or microscopic ciliated larvae. Secondary predation sources correspond to OTUs assigned to animals that were more likely consumed by the co-detected prey than by the siphonophore. For example, crustacean, gastropod, and larvacean sequences in P. physalis samples were interpreted as secondary predation (prey of their fish prey) given our knowledge on the prey-capture limitations of these animals and the feeding habits of their fish prey.

Parasite interpretations were annotated onto OTUs assigned as trematodes, cestodes, ichthyophonids, and myxozoans, since the most likely explanation for their presence in the samples is due to parasitism in the siphonophores or their prey. We used the environmental category to annotate OTUs likely originated from the microbial community (such as diatoms, dinoflagellates, uncultured eukaryotes) or eDNA (such as rotifers, sharks, ascidians, sponges, bivalves, anemones, echiurids, gastrotrichs, echinoderms, or bryozoans), based on their taxonomic assignment and read abundance. OTUs assigned as ‘uncultured eukaryotes’ were BLAST-checked to differentiate between environmental microbes and failed assignments of siphonophore sequences. We used the contamination category to interpret OTUs assigned to tetrapods (likely from humans), pollen, branchiopods, nematodes, mites, and insects. Amplification experiments on negative controls indicated that these contaminants originated from specimen manipulation in the field and not from the lab bench. The cross-contamination interpretation was used to annotate some suspicious OTUs with low reads in some samples that matched other taxa that were being extracted and amplified at the lab bench. Reads suspected of cross-contamination from other samples amplified in close proximity (assigned to taxa present in the potential sources of contamination, present across multiple samples in the same run with very low read abundances) were conservatively annotated as such.

When the taxonomic assignments from different barcodes disagreed, we annotated the OTUs based on the barcode majority consensus and BLAST checks. For example, Wwen all barcode regions except ‘V5-V7S’ indicate mysid prey but ‘V5-V7S’ identifies a similar number of reads as stomatopod prey, we interpreted those reads as mysid prey. Assignments of shark identities by barcode region ‘V5-V7S’ in one of the P. physalis samples (specimen BIOS19-D1-P5) were identified as ray-finned fish prey using BLAST searches and interpreted as such, in agreement with the other barcode regions. Assignments of decapod crustacean identities by barcode region ‘V5-V7S’ (in samples D1137-D7 Forskalia sp., D1243-BW25 Diphyes dispar, and D1244-SS8 Nanomia sp.) were interpreted as euphausiid prey in agreement with the assignments on the rest of the barcode regions. The taxonomic composition of the samples was analyzed and visualized in the R programming environment. Scripts and data available in the GitHub repository.

Prey field characterization

In order to compare the observed diet to the environmental abundances of potential prey taxa, we collected zooplankton and micronekton samples on the same day and station location as the relevant siphonophore gut content samples. The plankton samples paired with epipelagic siphonophore specimens were collected using a weighted hand-held plankton net (ring diameter of 1m for the Bermuda samples, 0.5m for the OCCE and Block Island sound samples) with a mesh size of 250 μm towed for ~10min between 0–20 meters depth at a maximum speed of 1 kt. In order to quantify the deep pelagic community paired with the ROV-collected siphonophore specimens, we collected zooplankton and micronekton samples using a Tucker trawl (frame area: 2 m², mesh size: 500 μm) towed for ~2h between 900 m and the surface at night. Environmental community samples were visually examined live to collect specimens to sequence for the 18S reference library and other purposes, which were annotated as removed. Samples were concentrated using metal sieves and fixed in 4% formalin. Back at the Yale Peabody Museum of Natural History, these samples were visually identified and quantified from a splitter aliquot. Identifications were carried out to the lowest taxonomic level as well as to a broad group level (e.g., copepods, decapods, krill, fish, hydromedusae, chaetognaths, polychaetes etc.). A few individual unaccounted specimens were removed from the haul before preservation to serve other scientific goals during fieldwork, and therefore these samples may be imperfect representations of the community. In order to estimate how selective siphonophore species are for different prey types in the environment, we calculated Strauss Linear Index (LI) [38] at the broad taxonomic group level.

(1)

We used this index to capture the difference between the fraction of each prey type in the environment (p_i) and the observed frequencies of prey types in the gut contents (r_i).

Comparisons to published sources

We aimed to compare and expand previous predation results from submersible observations and visual gut content inspections with the new results of DNA metabarcoding of gut contents. Therefore, we used the dietary data compiled in Damian-Serrano et al. [26] from 11 published sources divided into those that used gut content inspections and those that used human- and remotely-operated submersible observations. Many of the submersible observations correspond to ROV observations carried out in the Offshore California Current Ecosystem, spatially overlapping with the location where the majority of our metabarcoding samples were collected. Salps, ctenophores, and medusae were merged into a gelatinous prey type for comparative purposes. Published records for Apolemia uvaria were considered equivalent to Apolemia sp. for genus level comparisons. Records of all Forskalia species were considered equivalent to Forskalia sp. To evaluate the morphology-based dietary predictions generated in Damian-Serrano et al. [27], we compared the Bayesian posterior probabilities for each predicted feeding guild for each species to the metabarcoding findings. Small-crustacean guild predictions were mapped to copepod, ostracod, and cladoceran prey. Large-crustacean guild predictions were mapped to decapod, euphausiid, mysid, lophogastrid, stomatopod, and amphipod prey. Generalist guild predictions were mapped to all prey types except gelatinous prey (following the intended distinction with gelatinous specialists used in Damian-Serrano et al. [26]).

Dietary findings by taxon

Physalia physalis.

The Portuguese man-o-war is the only pleustonic (surface floating) member of the siphonophores, and the most encountered by beachgoers. Man-o-wars are well-known to feed exclusively on relatively large and motile soft-bodied prey such as fish, chaetognaths, or pelagic gastropods [40]. In our gut content samples of the Portuguese man-o-war from Bermuda, we found three specimens with ray-finned fish sequences (S1 Fig), some of which had visually recognizable fish in the gastrozooids when collected. Fish prey is congruent with published visual inspections of their gut contents [40, 41]. In all three specimens with fish prey we also found benthic and hard-bodied taxa (mysid, alpheid shrimp, spider crab, copepod, benthic gastropod, and a sipunculid worm), as well as larvacean prey sequences (S12 Table). Their nematocysts are not able to subdue crustacean prey, and their feeding reflex would not be triggered by a prey as small as a larvacean or cilia-propelled larva [42]. Therefore, we interpreted the presence of these taxa in the gut contents as secondary predation in the gut contents of the fish prey (S2 and S6 Tables). In addition, we also detected ctenophore prey in one specimen. This could be also a case of secondary predation, but we suspect a ctenophore could be large enough to be prey of the man-o-war. If that is the case, this would be the first record of P. physalis consuming gelatinous zooplankton, which would place the man-o-war as a central species in the epipelagic ‘jelly-web’ [43]. Comparisons with their surrounding prey field show these specimens were strongly selective for fish and strongly exclusive of copepods (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Species-wise grid with the frequency of the major prey types identified from the metabarcoding data and the average prey-type selectivity.

Gut content cells in white indicate absence, and cells in grey indicate presence in one specimen, or more than one specimen if labeled with a number. Selectivity colors mapped to Strauss’ L.I. values. The siphonophore cladogram (left) is a simplified version of the phylogenetic tree published in Damian-Serrano et al. [26].

https://doi.org/10.1371/journal.pone.0267761.g004

Bargmannia spp.

The three Bargmannia species considered here are frequently observed in the midwaters off Monterey Bay, and have relatively simple tentilla with large stenotele nematocysts and an undifferentiated terminal filament [27]. ROVs have recorded Bargmannia elongata consuming crustaceans and cephalopods. One B. elongata specimen had euphausiid and ostracod prey, in agreement with the DAPC prediction Bayesian posteriors for B. elongata to feed mainly on large crustaceans, but also marginally on small crustaceans (Fig 5). During specimen collection we observed a mysid prey in a specimen B. amoena. Our DNA metabarcoding identified this mysid as Boreomysis (METAXA assignment score 54.68% for V7 on SILVA123.1), and found a copepod in another specimen. Nothing was previously known, however, about the diet of Bargmannnia lata. The two B. lata specimens we sequenced consumed a ctenophore and a copepod, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Feeding interactions between siphonophores and their prey from different data sources.

Including prey identified by our metabarcoding results (red), observations published submersible observations (blue), observations published visual gut content analyses (green), and prey types predicted by the morphology-based DAPC model in Damian-Serrano et al. [27]. Gelatinous prey refers to ctenophores, medusae, and salps. Larvaceans were excluded as their own category since they are not gelatinous when swimming freely outside their mucous ‘houses’, which would be the only times they would be able to trigger a prey-capture response in siphonophore tentacles.

https://doi.org/10.1371/journal.pone.0267761.g005

The diets of these three closely-related, coexisting species appear to be non-overlapping, which could be a consequence of competitive trophic niche partitioning. The findings for B. lata consuming copepod and ctenophore prey are not congruent with the morphology-based prediction to be a large-crustacean specialist (Fig 5). We suspect that the lack of taxon sampling among the pyrostephids in [26] could have led to overfitting in the DAPC for this group. Finding ctenophore prey in this species further supports the involvement of deep-sea siphonophores in the midwater ‘jelly web’ hypothesized in Choy et al. [4].

Comparisons with visual methods

We report the first insights into the diets of nine siphonophore species and reveal 29 novel predator-prey interactions (Figs 2 and 5). When comparing our metabarcoding findings with the published visual observations from gut content inspections and submersible dives, we found five interactions congruent with ROV observations, and eight interactions (six of them involving copepods) congruent with visual gut content inspections of SCUBA-collected colonies (Fig 5).

The published records on the diets of siphonophores appear to differ in prey-type composition between epi- and deep-pelagic habitats [8]. However, the different methodological limitations inherent to each visual method (small prey underestimated by submersibles, soft-bodied prey underestimated by gut content inspections) are hypothesized to be responsible for such differences [8]. Our approach has detected prey types, such as larvaceans, ctenophores, bivalves, and ostracods previously missed by visual methods. The gelatinous animals (i.e. ctenophores, medusae, salps) identified by submersibles as prey of deep-pelagic siphonophores were found present in the gut contents of several deep species (Apolemia sp., B. lata, undescribed physonect L, and S. christiansonae), supporting the validity of these observations. However, the gelatinous prey recorded by submersibles in prayids such as Praya dubia and D. annectens [4] were not recovered in our D. annectens samples (Fig 5), suggesting that either our sample sizes were not large enough, or that ROVs had observed accidental entanglement of jellies on their tentacle nets which did not end in ingestion. In addition, we found several small crustaceans in the gut contents of epipelagic species (Forskalia sp., Nanomia sp., S. koellikeri, S. chuni, and D. dispar) in agreement with visual gut contents observations in shallow habitats. On the other hand, we also found gelatinous and soft-bodied invertebrate prey in shallow-dwelling species (P. physalis, D. dispar, and M. atlantica); as well as small-bodied animals among the prey of deep-pelagic siphonophores (Apolemia spp., Bargmannia spp., R. dunni, V. serrata, D. annectens, C. multidentata, and L. conoidea) (Fig 2). Copepods and ctenophores were the most frequent prey among bathypelagic siphonophores, while other crustaceans (such as ostracods, decapods, and krill) appeared as prey more frequently among the mesopelagic taxa. While these findings are consistent with the hypothesis that small prey is underestimated in submersible observations and rapidly-digested, soft-bodied prey is underestimated by gut content inspections, our sample sizes are insufficient to determine whether the relative contribution of these prey differs between habitats.

DNA metabarcoding was able to detect prey both small and large, gelatinous and hard-bodied, for both deep and shallow-dwelling siphonophore species. These results show that the trophic roles of siphonophores in epi- and deep-pelagic food webs could be more similar than previously-published records may indicate, due to the biases brought by the different diet-assessment methods applied in each habitat. Vertical migration is an important driver of pelagic food web structure [45, 46]. We found copepods, decapods, and euphausiids in the gut contents of both meso- and epipelagic siphonophores. These prey taxa are well-known vertical migrators [47–49], suggesting that there might be some vertical trophic connectivity between these habitats as prey migrates between them. In addition, a few siphonophore species (including V. serrata and L. conoidea in this study) are also known diel vertical migrators [50], but their patterns of feeding with depth remain unknown. Finally, our selectivity estimates (for four epipelagic and two mesopelagic species) indicate that siphonophores may play a similar role as selective predators across all depths in the water column.

Comparisons with prey field

We examined 8 prey-field samples that corresponded to the colocalized ambient prey of 15 out of 47 specimens (some trawls correspond to more than one sampled specimen). The epipelagic plankton samples from Bermuda (colocalized with the P. physalis specimens) were dominated by copepods, followed by decapod larvae and chaetognaths. While fish larvae were scarce in these samples, they were still far more abundant than in any other sampled location. The Atlantic epipelagic plankton samples (colocalized with the S. chuni and Atlantic shallow Nanomia specimens) were dominated by cladocerans, followed by copepods, larvaceans and salps. The Pacific epipelagic plankton sample from California (colocalized with the D. dispar specimens) was also dominated by copepods, followed by cladocerans and larvaceans. The quantified midwater tucker trawl from California (colocalized with V. serrata specimen D1137-D8 and Forskalia sp. specimen D1137-D9) was also dominated by copepods (albeit larger species), followed by euphausiids (both adult and larval), chaetognaths, and ostracods.

We found both positive and negative selectivity when comparing identified siphonophore prey to quantified co-localized prey fields. We found strong negative (<-0.5) selectivity for copepods in P. physalis specimens and in one specimen of V. serrata. However, in 11 siphonophore specimens from 4 species (out of the 6 species that were quantitatively assessed), we found strong positive selectivity (>0.5) for a specific prey type (S1 Fig). These cases include: selectivity for fish in P. physalis; selectivity for copepods in S. chuni, and Atlantic Nanomia sp., selectivity for bivalve larvae in V. serrata, and selectivity for salps in D. dispar (Fig 4). These selectivity values suggest a strong influence of predator-specific differences in prey capture capabilities for different prey types. However, more replication is necessary in order to test for prey-type specialization.

Epipelagic siphonophores are known to be highly selective and specialized carnivores [7, 26, 51, 52]. ROV observations have revealed that some deep-sea siphonophores are also highly specialized [4]. However, the lack of paired diet and planktonic community samples has limited an assessment of their feeding selectivity. For both the shallow- and deep-dwelling siphonophore species assessed here, we found their prey belonged to the less-abundant components of the co-localized planktonic community, demonstrating high prey-type selectivity. However, the selectivity index values presented in this study should be interpreted with care, since the prey field data is quantitative (abundance-based) but the gut content values are only binary at the specimen level, and frequency-based at the predator species level. Overall, crustaceans (especially copepods) were identified as the most frequent prey type among siphonophore diets. Copepods are typically the most abundant prey type in planktonic communities, thus being able to feed on them is likely an advantageous strategy for any planktivorous predator [53]. Fish prey were detected only in the Portuguese man-o-war samples, in agreement with published observations of man-o-war feeding [40, 41].

Our findings are congruent with the idea that siphonophores span multiple trophic positions, consuming prey across low (salps, larvaceans, copepods, ostracods) and high (fish, ctenophores, medusae) trophic levels. We found larvaceans and salps as prey of shallow- and deep-dwelling siphonophores. These urochordates have an important role in the biological carbon pump, sequestering carbon from phytoplanktonic producers into the deep sea by means of fecal matter production, mucus filter shedding, and carcass depositions [54, 55]. The role of predation on these gelatinous herbivores is often underestimated in oceanic food-web models, or primarily attributed to vertebrate predators [56]. Our results show that some siphonophores like Apolemia sp., A. lanosa, M. atlantica, and D. dispar may play an important mid-trophic role incorporating this soft-bodied herbivore productivity into the food web, and providing an alternative avenue to transfer carbon into the deep sea.

Comparisons with morphology predictions

Comparing our metabarcoding findings with morphology-based predictions [27], we found support for 10 of the 16 predicted interactions between siphonophores and prey. Among the physonects, our results supported the predictions of B. elongata eating krill and ostracods, R. dunni eating copepods, Forskalia sp. eating decapods, and H. rubrum eating lophogastrids. Among the calycophorans, we found support for the predictions of V. serrata eating decapods, ostracods, and molluscs; also C. multidentata and L. conoidea eating copepods. Among the siphonophore species studied there were 70 predicted interactions that were not found among the metabarcoding results (Fig 5). Out of the 10 taxa with both morphology-based predictions and metabarcoding results, six had all prey congruent with the predictions, three had all prey incongruent with the predictions, and Forskalia sp. presented both cases.

Food-web structure is determined largely by community composition and its patterns in time and space, as the organismal assemblages determine what predators are present and what prey is available to them [57–59]. However, organismal traits constrain which predators can eat which prey [60, 61]. The most commonly-studied trait to predict oceanic food web structure has been size [62, 63]. This is due to the importance of gape size in most predators (i.e. fish, squids, crustaceans etc.) with singular and rigid buccal openings [64, 65]. Siphonophores differ from most predators by having many gastrozooid mouths along their length, all capable of stretching out significantly to ingest prey [66], sometimes utilizing multiple zooids to wrap around large prey [67]. While prey size is still an important constraint for siphonophore-prey interactions [42], siphonophore size is far less relevant. Moreover, some studies have found that phylogenetically-conserved predator traits other than size may also be important predictors of food web structure [68, 69]. Diet is a strong predictor of both extant and ancestral siphonophore tentilla morphology, as well as of its evolutionary dynamics [26]. These relationships were utilized predict the diets of understudied siphonophore species based on the morphology of their tentilla and nematocysts [27]. Here, we were able to test these predictions for ten species and found that most of the prey items found were congruent with these predictions, indicating that tentilla morphology is a strong predictor of siphonophore diets. This finding suggests that some components of the open-ocean food web may be structured by variation in complex morphological traits exclusive to specific predator groups.

Siphonophores are hypothesized to easily evolve between feeding specializations and into a generalist diet due to their modular body plan and their functionally-specialized tentilla [26]. Our results show that closely-related species, such as those within the genera Bargmannia, Apolemia, and Nanomia, appear to feed on different prey. We hypothesize that these species could be further subspecializing to avoid competition or adapt to different prey fields at different depth habitats. This hypothesis is congruent with the conclusions in Damian-Serrano et al. [26] stating that siphonophore dietary evolution can drive rapid morphological shifts. Moreover, we find that Apolemia sp., as well as V. serrata, could be generalists feeding on a variety of crustacean and soft-bodied prey. If a more extensive and quantitative sampling of these taxa was to validate this trophic reclassification, that would suggest that a generalist diet had evolved not just three (as proposed in Damian-Serrano et al. [26]), but up to five times independently, further reinforcing the idea that siphonophore generalists were able to evolve from specialist ancestors.

Methodological considerations

While DNA-based tools can detect prey unrecognized by visual methods, they are not free of shortcomings. Since all life stages of an animal have the same genetic signature, metabarcoding tools are unable to distinguish between larval, juvenile, or adult prey. These ontogenetic stages can have vastly different ecological implications and pose different challenges during prey capture. In addition, the application of DNA metabarcoding to predator diets is usually not quantitative, since too many sources of variation may lead to differences in read abundance. For example, different animal clades have different sizes, cell densities (due to variable acellular mesoglea content), digestion rates, number of copies of the target gene, or primer affinities during the PCR [70–72]. Due to the difficulties inherent to locating and sampling the species examined in this study, frequency-based quantitative comparisons were not possible for most species either. In addition, the sample size limitations of this study may have biased the results towards higher apparent specialization, and may have missed some important components of the diets of some target species. This caveat is also common in submersible observation data and limits the reliability of comparisons across these methods.

Siphonophores differ from other consumers in several ways which impose further limitations to the value of gut content metabarcoding. The most important aspect is their feeding mode and feeding rate, especially as deep-sea ambush predators, which typically consume one prey at a time and do not get a chance to capture another until far after the former has been digested [9]. Therefore, most siphonophores are found with empty guts or digesting one or few prey items at a time. Thus, the sample size required for frequency-based analyses is much higher than for other consumers which feed more frequently. Moreover, with the exception of a couple species such as Rhizophysa eysenhardti and Rosacea cymbiformis which are diurnal feeders [7], most species also feed during the night. In the open ocean, diel vertical migration drastically changes the prey field composition for siphonophores at night [45]. Given the fieldwork limitations in this study, we were only able to collect siphonophore gut contents during the day, thus likely biasing their diet towards their diurnal prey captures. Finally, secondary predation (the prey of the prey) cannot be empirically distinguished from direct predation using DNA metabarcoding, and thus we must rely on natural-history based assumptions.