Dietary partitioning promotes the coexistence of planktivorous species on coral reefs

Abstract Theories involving niche diversification to explain high levels of tropical diversity propose that species are more likely to co‐occur if they partition at least one dimension of their ecological niche space. Yet, numerous species appear to have widely overlapping niches based upon broad categorizations of resource use or functional traits. In particular, the extent to which food partitioning contributes to species coexistence in hyperdiverse tropical ecosystems remains unresolved. Here, we use a molecular approach to investigate inter‐ and intraspecific dietary partitioning between two species of damselfish (Dascyllus flavicaudus, Chromis viridis) that commonly co‐occur in branching corals. Species‐level identification of their diverse zooplankton prey revealed significant differences in diet composition between species despite their seemingly similar feeding strategies. Dascyllus exhibited a more diverse diet than Chromis, whereas Chromis tended to select larger prey items. A large calanoid copepod, Labidocera sp., found in low density and higher in the water column during the day, explained more than 19% of the variation in dietary composition between Dascyllus and Chromis. Dascyllus did not significantly shift its diet in the presence of Chromis, which suggests intrinsic differences in feeding behaviour. Finally, prey composition significantly shifted during the ontogeny of both fish species. Our findings show that levels of dietary specialization among coral reef associated species have likely been underestimated, and they underscore the importance of characterizing trophic webs in tropical ecosystems at higher levels of taxonomic resolution. They also suggest that niche redundancy may not be as common as previously thought.

along at least one dimension of their niche space, e.g., they may occupy different microhabitats (spatial niche partitioning), have asynchronous activity patterns (temporal niche partitioning) or consume different foods (dietary niche partitioning). However, many co-occurring species show little or no apparent evidence of resource segregation (Sale, 1978;Siepielski & McPeek, 2010). This raises the question of whether the apparent niche overlap observed among many species pairs is a result of overly broad characterizations of resources (Kartzinel et al., 2015).
On coral reefs, arguably the most diverse marine ecosystem, there is a wealth of evidence for the existence of interspecific competition (Bonin, Boström-Einarsson, Munday, & Jones, 2015).
A diverse array of fishes and invertebrates associated with coral reefs capture and ingest planktonic animals from the water column as a feeding strategy. Planktivorous species are found in a majority of families of coral reef fishes (e.g., Pomacentridae, Holocentridae, Apogonidae, Balistidae) and many corals rely on heterotrophy in addition to autotrophy for nutrient acquisition (Houlbrèque & Ferrier-Pagès, 2009). Numerous invertebrates living in holes and crevices created by the three-dimensional structure of reefs also filter plankton, i.e., sponges, tunicates, bivalves, brittle stars, and polychaetes.
Planktivory can be a dominant trophic mode, with up to 50% of total fish biomass comprised of planktivores (DeMartini, Friedlander, Sandin, & Sala, 2008), and it is a critical link in the transfer of organic matter from the water column to the benthic components of oligotrophic coral reef food webs (Hanson, Schnarr, & Leichter, 2016).
Despite their abundance, diversity and functional role, the degree to which planktivorous organisms partition available food resources on coral reefs currently is not well quantified (see Frédérich, Fabri, Lepoint, Vandewalle, & Parmentier, 2009).
Pocilloporid branching corals provide habitat for a range of fishes and invertebrates, some of which share a planktivorous feeding strategy and thus potentially compete for food resources as they co-occur among the branches of their coral host. The yellowtail dascyllus (D. flavicaudus) and blue green damselfish (Chromis viridis) reside among the branches of Pocillopora which afford a refuge from predation (Holbrook & Schmitt, 2002;Schmitt & Holbrook, 1999). While feeding, they swim in the water column directly above the corals to capture zooplankton. In exchange, they promote growth of their host by releasing nutrients in the form of nitrogenous waste (Holbrook, Brooks, Schmitt, & Stewart, 2008;Holbrook, Schmitt, & Brooks, 2011;Shantz, Ladd, Schrack, & Burkepile, 2015). Notably, they are commonly found together in branching corals feeding simultaneously during the day (Hanson et al., 2016;Holbrook et al., 2008Holbrook et al., , 2011. In this study, we hypothesize that given the lack of significant temporal and spatial niche partitioning, these two planktonic-feeding fish are able to coexist, in part, because they partition their diet. Chromis viridis and D. flavicaudus have been considered to have broadly overlapping diets based on their highly similar feeding strategies and apparatus. Both rely on vision to detect and suction very small prey items in the water column and have only minor differences in the morphology of the jaw (Frédérich, Parmentier, & Vandewalle, 2006).
Dascyllus flavicaudus has larger caniniform teeth and a stronger coronoid process than C. viridis.
A high-throughput sequencing approach targeting the hypervariable mitochondrial Cytochrome c. Oxidase subunit I gene (COI) allowed us to characterize intra-and interspecific dietary differences between the two coral-dwelling fishes with an unprecedented level of taxonomic resolution. We used the same sequencing approach on polyp tissues of the coral hosts to gain insights into the poorly known diet of a sessile and more passive consumer. To characterize the pool of prey that the three planktivorous species had access to and gain further insights into their feeding behaviour (i.e., feeding selectivity), we also conducted counts of zooplankton in surrounding waters.
The study was conducted in Moorea, French Polynesia, where an extensive library of COI DNA barcodes has been built by the BIOCODE project for >3,500 marine species (Meyer, 2017). This curated library of reference sequences was used to classify many sequences recovered from fish guts and coral tissue to the species level. Our results highlight the potential of the metabarcoding approach to decipher complex mechanisms of resource use in diverse ecosystems and across a wide range of consumers.

| Sample collection
In the lagoon of Moorea, French Polynesia, Dascyllus flavicaudus and Chromis viridis (hereafter referred to as Dascyllus and Chromis) are the two most abundant fish species occupying the branching structure of Pocillopora corals. Individuals of these two species are found residing in and hovering over coral colonies in large mono-or multispecific aggregations, and both species are visual planktivores that actively feed in the water column above their host corals during the day (Hanson et al., 2016;Holbrook et al., 2008Holbrook et al., , 2011. All fishes (including Dascyllus and Chromis individuals) resident on three experimental Pocillopora eydouxi colonies that had been outplanted to a sand flat in the Maharepa lagoon (17.4751°S, 149.8088°W) approximately two months prior to initiation of the study, were collected using small hand nets and the anaesthetic quinaldine (2-methylquinoline) sulphate. Experimental corals, hereafter Pocillopora A, B and C, were representative of naturally occurring P. eydouxi colonies found within the Maharepa lagoon on the northern side of the island. Each of the three experimental colonies used in the study were approximately 1.0 m in diameter and 0.4 m in height, and were located at a depth of 2 m. The corals were spaced 10 m apart on a sand plain, which prevented target fish from moving between colonies (Schmitt & Holbrook, 1996. Fishes were collected after peak feeding as Zooplankton samples were collected during 60 min periods at night between 0200 hr and 0330 hr and again during the day between 1400 and 1530 hr over three days, August 18-20 2008, using submersible plankton pumps outfitted with precalibrated, internal flow metres, 200 μm mesh cod ends, and 2.5 cm inside diameter intake pipe openings (Alldredge & King, 2009). Pumps were deployed in pairs adjacent to each of the three outplanted Pocillopora colonies with one pump sampling the water column 0.3 m (mean ± SD = 0.32 ± 0.04 m) above the bottom and the other pump sampling 1.4 m (mean ± SD = 1.42 ± 0.09 m) above the bottom. Sampling heights corresponded to the maximum height of the outplanted corals above the bottom and the estimated maximum distance above the bottom where Dascyllus and Chromis had been observed to feed (Brooks, personal communication). Daytime plankton abundances in Moorea remain stable between noon when the fish were sampled and 1400-1530 hr when the plankton were sampled (Alldredge & King, 2009). Plankton samples were returned to the laboratory and preserved in 2% buffered formaldehyde for sorting, identification and enumeration using a dissecting microscope and plankton wheel. Flow meter data were utilized to standardize raw count data to numbers observed per cubic meter of water filtered. In four of the 36 samples, one to five Dascyllus or Chromis had entered the intake pipe and were captured in the cod end. These four samples were removed from the analyses because low plankton counts indicated substantial feeding in the cod end by the captured fish. Night samples were all collected while a bright three quarter moon was located mid-sky. Height and day/night differences in abundance were tested for statistical significance using Student's t-tests with the data log transformed where necessary to meet assumptions of normality.

| Sample processing
Fishes were thawed on 2 September 2008, measured (total length in mm) and weighed (g wet weight). The entire digestive tract of each fish was then removed and individually stored at -20°C in numbered Eppendorf tubes containing 95% ethanol. Visual examination of the stomach contents of Dascyllus (n = 34) and Chromis (n = 27) revealed a range of undigested soft-(e.g., eggs) and hard-bodied (e.g., copepods) prey from the morning's food intake. By contrast, the lower digestive tracts contained only undigested hard parts, most likely ingested the previous day. To obtain a snapshot of the diet for comparative analysis and avoid the misrepresentation of soft-bodied taxa, we focused our molecular analysis on prey removed from the stomach only.
Using sterile tools, the contents of each fish stomach and coral tissues were transferred to individual 2 ml tubes for DNA extraction using the QIAGEN DNeasy Blood & Tissue kit following manufacturer's instructions. A small tissue subsample (1 cm × 1 cm) per branch tip was excised from the Pocillopora skeleton and thoroughly rinsed with DNA-free water to remove excess DMSO buffer that could interfere with downstream analysis. The initial volume of lysis buffer was adjusted based on the amount of starting material if necessary, and samples were incubated with proteinase K overnight or until tissues were completely lysed. Genomic DNA was subsequently purified using the PowerClean DNA clean-up kit (MO-BIO) to minimize potential inhibition during Polymerase Chain Amplification (PCR). A negative control extraction and negative DNA purification were performed to check for potential contaminants.

| High throughput sequencing
To reduce the per sample cost of sequencing on the Roche FLX platform, a hierarchical tagging approach, by which each sample is tagged with a unique combination of two indices, was implemented for sample multiplexing (Table S1). The first index was incorporated during PCR amplification using indexed PCR primers. The second tag was incorporated using a ligation of indexed adaptors (Leray, Haenel, & Bourlat, 2016;Leray, Yang, et al., 2013).
First, cleaned DNA extracts (n = 64) were used to amplify a hypervariable fragment (~313 bp) of the Cytochrome c. Oxidase subunit I (COI) region with versatile PCR primers mlCOIintF and jgHCO2198 (Geller, Meyer, Parker, & Hawk, 2013;Leray, Yang, et al., 2013) known to perform well across the diversity of marine invertebrates (Leray, Yang, et al., 2013). Despite some level of amplification bias, this primer set provides useful estimates of relative abundance as shown recently for benthic samples (Leray & Knowlton, 2015).
A 6 bp index sequence was included at the 5' end of each PCR primer (Tables S1 and S2), with each index sequence differing by at least 3 bp. These index sequences were shown to induce no significant bias in operational taxonomic units (OTU) detection . To avoid the erroneous assignment of reads due to tag jumping (Schnell, Bohmann, & Gilbert, 2015), we used identical indices on the forward and reverse primer for each individual sample (e.g., Sample 1: Index1-mlCOIF/jgHCO-index1; Sample 2: Index2-mlCOIF/jgHCO-index2). Because the co-amplification of consumer DNA is known to prevent the recovery of some prey (Leray, Yang, et al., 2013;Vestheim & Jarman, 2008), consumer-specific annealing blocking primers (Table S2) were included in each PCR reaction at 10 times the concentration of versatile primers. The PCR cocktail and touchdown temperature profile can be found in Leray, Yang, et al. (2013). Three PCR replicates were conducted per sample, pooled, gel excised to ensure complete removal of primer dimers, purified using QIAGEN MinElute columns and the product eluted in 12 μl of elution buffer. A PCR reaction performed with negative control extractions confirmed the absence of contaminants (no band on 1.5% agarose gel). PCR product concentration was normalized after quantification using the dsDNA Qubit Fluorometer (Invitrogen) and equimolar amounts of each sample were pooled, with each pool containing amplicons generated with each of the eight indexed primer pairs.
Second, a total of 500 ng of PCR product was used per pool for end-repair and dA-tailed using the NEBNext Quick DNA Sample Prep Reagent Set 2 chemistry (New England BioLabs) followed by the ligation of unique 454Multiplex Identifiers (Table S1) using the FLX Titanium Rapid Library MID Adaptors Kit (Roche). The ligated PCR products were purified using Agencourt AMPure beads (Beckman Coulter Genomics), eluted in 40 μl of TE buffer, and pooled prior to emulsion PCR and sequencing in two 454 runs. Details of the experimental design of each run are provided in Tables S1 and S2.
Samples of both species were randomly assigned to each run. Note that additional samples unrelated to this study were also included in these runs.
Reads were then discarded if they (a) did not include forward and reverse primers and indices, (b) had more than one mismatch in primer index sequences, (c) had more than two mismatches in primer sequences, (d) had any ambiguous base calls (e.g., "N"), or (e) had any homopolymer regions longer than 8 bp. The fasta files were then demultiplexed based on primer indices in Mothur (Schloss et al., 2009), the sequences of both 454 runs were pooled and the data set was dereplicated using the trie function in QIIME (Caporaso et al., 2010). Following this initial quality filtering, the option "en-richAlignment" implemented in MACSE (Ranwez, Harispe, Delsuc, & Douzery, 2011) was used to align unique reads to the reference database of COI barcodes built by the Moorea BIOCODE (Leray et al., 2012;Meyer, 2017), an all-taxa biodiversity inventory of the island ecosystem. MACSE performs alignments at the amino acid level and detects interruptions in the open reading frame due to nucleotide substitution or nucleotide insertion/deletion. We selected the invertebrate mitochondrial translation code and only retained sequences without any stop codons or frameshifts for subsequent analysis. To further reduce the variability in the data set and speedup the downstream clustering, we used an initial preclustering approach implemented in Mothur to merge reads differing by three or fewer bases. This algorithm first ranks sequences in order of their abundance and then merges rare sequences with more abundant sequences within the threshold specified (Huse, Welch, Morrison, & Sogin, 2010). Reads were then screened for chimeras using UCHIME (Edgar, Haas, Clemente, Quince, & Knight, 2011)

implemented in
Mothur before discarding all remaining preclusters represented by a single sequence (singletons).
The resulting quality filtered data set was used as an input for CROP (Hao, Jiang, & Chen, 2011), a Bayesian model that delineates OTUs based on the natural distribution of sequence dissimilarity.
Rather than using a hard cut-off (e.g., 5%), CROP generates clusters within user-defined lower (-l) and upper (-u) bound levels of similarity to account for differences in rates of sequence evolution among taxonomic groups. We defined -l 3 and -u 4 because it was previously shown to delineate OTUs that closely reflect species grouping among marine invertebrates by providing the lowest frequency of false positives (splitting of single taxa) and false negatives (lumping of multiple taxa) (Leray, Yang, et al., 2013).
One representative sequence per OTU was used for taxonomic assignments following an iterative strategy. We ran similarity searches of each representative sequence against the Moorea BIOCODE reference database (BLASTn: word size = 11, e value = 1e-20), GENBANK (BLASTn: word size = 11, e value = 1e-20) and the Barcode of Life Data (search engine with default settings [Ratnasingham & Hebert, 2007]) to classify OTUs in three categories based on their level of similarity to a reference: >97%, 97%-85%, <85%. We considered that there was a species-level "match" when the similarity to a reference sequence in one of the three databases was at least 97% (Machida, Hashiguchi, Nishida, & Nishida, 2009). In the case where an OTU matched multiple species at >97% similarity, it was assigned to the lowest common taxonomic rank. OTUs with a sequence similarity to a reference barcode between 97% and 85% were assigned to the phylum of the closest match as recommended by Ransome et al. (2017). Finally, OTUs with <85% similarity to a reference COI barcode were assigned to phyla using a phylogenetic approach implemented in the Statistical Assignment Package (SAP) (Munch, Boomsma, Huelsenbeck, Willerslev, & Nielsen, 2008). We allowed SAP to build 10,000 unrooted phylogenetic trees with 50 homologues retrieved from GENBANK (>70% sequence similarity) for each query sequence (i.e., each OTU representative sequence) and accepted taxonomic assignments at an 80% posterior probability cut-off . OTUs that could not be confidently assigned using any of the methods above were labelled "unidentified".
A sample by observation contingency

| Diversity analysis
To illustrate the extent of the sequencing effort, individual-and sample-based rarefaction curves were built. The curves were computed by randomly resampling sequences and samples respectively at increasing levels of accumulation using EstimateS (Colwell, 2006).
A curve that plateaus indicates a sufficient sampling effort as only rare OTUs remain to be detected.
Unequal numbers of reads can affect estimates of alpha and beta diversity because of the positive relationship between number of sequences and number of OTUs. Hence, a subsampling procedure (rarefaction) was used to create an alternative OTU table in which the number of reads of all samples was scaled down to the smallest number of sequences that a sample contained in the data set (259).
The rarefied OTU table was used to compute distance matrices of community dissimilarity based on the Jaccard and the Bray Curtis metrics within the R package Vegan (Oksanen et al., 2009). The input table was converted to a presence/absence matrix prior to calculating Jaccard (binary = true, in function metaMDS). Jaccard ranges from 0 to 1. A value of 0 indicates that samples have exactly the same OTU composition whereas a value of 1 indicates that samples do not have any OTU in common. Bray-Curtis takes into account differences in abundance of reads between samples; a value of 0 indicates that samples are exactly identical in terms of OTU composition and abundance of reads whereas a value of 1 indicates that samples do not have any OTU in common. Bray-Curtis gives less weight to rare OTUs than Jaccard, so that samples will have lower Bray-Curtis values if they share abundant OTUs.
Patterns of species composition were visualized in two-dimensional space using nonmetric multidimensional scaling (NMDS) plots.
Differences in mean diet composition (position of groups of samples in multivariate space) were tested between species and between individual Dascyllus collected on different corals using PERMANOVA (Anderson, 2001). All statistical analyses were repeated with OTU tables rarefied down to 900 and 1,200 reads (which led to the removal of nine and 12 samples, respectively) to test for the robustness of ecological patterns to the loss of sequence data.
We further examined ontogenetic changes in fish diet by plotting pairwise Jaccard and Bray-Curtis dissimilarities in relation to differences in size (total length in mm). Linear models with 95% confidence intervals were fitted to each plot. We used the statistical programming environment R Studio v. 0.98.1056 for the computation of all statistics, the "Vegan" package (Oksanen et al., 2009) for community analysis and the "ggplot2" package (Wickham, 2009) for graphics. A total of 42 Dascyllus and 32 Chromis were collected from the three experimental coral colonies. These abundances are characteristic of the numbers of these two species found on naturally occurring colonies of P. eydouxi (Holbrook et al., 2015) in the lagoons of Moorea.
Chromis ranged in size from 31 to 65 mm TL (mean ± SD = 51 ± 9) and from 0.49 to 4.34 g wet weight (mean ± SD = 2.48 ± 1.08). Both the mean length and weight of Dascyllus differed among the three coral colonies with individuals being significantly longer (F 2,39 = 11.33, F I G U R E 1 Co-occurrence of planktivorous damselfishes surveyed on 10 Pocillopora eydouxi corals in the lagoon of Moorea p = 0.0001) on Pocillopora B than on Pocillopora A and significantly heavier (F 2,39 = 8.67, p = 0.0008) on Pocillopora B than on either Pocillopora A or C. Of these individuals, 34 Dascyllus (13 from Pocillopora A, 18 from Pocillopora B and three from Pocillopora C) and 27 Chromis (all from Pocillopora C) were used in the sequencing analysis.

| Diversity and abundance of dietary items in fish stomachs
There were a total of 354,753 reads after denoising the data set but only 251,933 (71%) met our requirements as described in the Methods section. An additional 53,932 reads were discarded because they had interruptions in the open reading frame.
The Bayesian clustering tool CROP delineated a total of 716 OTUs. Among them, seven OTUs (189 reads) were identified as prokaryotes, seven OTUs (1,135 reads) as contaminants (e.g., OTU diversity (c) and abundance (d) are also presented per individual fish and coral. An OTU was considered to match a reference sequence in the Moorea BIOCODE database, GenBank or BOLD if the level of similarity was higher than 97%. OTUs with < 97% similarity to reference COI barcodes were assigned to phyla using a Bayesian phylogenetic approach implemented in the Statistical Assignment Package (SAP) and closest Blast hits (if > 85% similarity, see Methods section). OTUs that could not be confidently assigned using any of the methods above were labelled "unidentified". The category "Other animals" comprises Bryozoa, Chaetognatha, Echinodermata, Hemichordata, Nemertea, Platyhelminthes, Sipuncula and Xenacoelomorpha. The category "Other nonanimals" comprises members of Phaeophyceae, Rhodophyta, Dinophyceae, Bacillariophyta, Heterokonta, Amoebozoa and Fungi [Colour figure can be viewed at wileyonlinelibrary.com] barcode (<97% similarity), 316 could be assigned to a higher taxonomic level (>85% similarity in GenBank or SAP assignment, see Figure 2b), while 110 (15.8%) remained unidentified (labelled "unidentified" in Figure 2a).
OTUs spanned a total of 14 animal phyla, among which Few OTUs were represented by a large number of sequences while many OTUs were rare ( Figure S1). In total, 12.7% of OTUs were represented by two sequences only and 38.7% of OTUs con-

| Patterns of dissimilarity in diet composition
Using the data set rarefied to 259 reads to control for differences in sequencing effort, there was evidence for interspecific dietary partitioning. Mean pairwise β diversity values were higher between species than within species for both the Jaccard and Bray-Curtis indices (Table 1). Samples of the three species were well separated on NMDS plots (Jaccard, Figure 4a; Bray-Curtis, Figure 4b), and these differences in mean diet composition were statistically supported by PERMANOVA tests (Jaccard: F π 2,61 = 3.5, R 2 = 0.1, p < 0.001; Bray-Curtis: F π 2,61 = 8.2, R 2 = 0.21, p < 0.001) ( TA B L E 1 Intra-and interspecific dietary overlap measured using the Jaccard (below diagonal) and the Bray-Curtis (above diagonal) dissimilarity metrics F I G U R E 4 Ordination plots comparing the composition of prey items. Nonmetric multidimensional scaling plots were computed using Jaccard (stress = 0.22), an incidence-based metric (a) and Bray-Curtis (stress = 0.22), an abundance-based metric (b). Note that the scale bar for size does not apply to Pocillopora samples. Beta diversity was calculated from the rarefied OTU table (n = 259) to control for differences in sequencing depth. Photos credit: Jeffrey Williams (fish), Gustav Paulay (  Note: OTUs that together contribute to 50% of the total difference in diet between each pairs of species are highlighted in bold (cumulated contribution of ranked OTUs). An OTU was considered to match a reference sequence in the Moorea BIOCODE database, GenBank or BOLD if the level of similarity was higher than 97%. OTUs with <97% similarity to reference COI barcodes were assigned to phyla using a Bayesian phylogenetic approach implemented in the Statistical Assignment Package (SAP) and closest Blast hits (if >85% similarity, see method section). OTUs that could not be confidently assigned using any of the methods above were labelled "unidentified". The member of each pair that consumes more of the indicated prey item is marked with a symbol. #, Chromis; *, Dascyllus; x, Pocillopora.
The number of reads reflects the individual size as well as the abundance of each OTU consumed, with larger or more numerous prey items generating more reads. Thus, the presence (or absence) of food items in the gut that could be clearly identified and assigned to a major taxonomic category was assessed to further elucidate differences in diet among the three species (Table 4)

TA B L E 4 Summary of metazoan OTUs by taxonomic group
Dascyllus had a greater diversity of fish OTUs, most certainly consumed as eggs. Chromis stomach samples yielded more reads of the OTUs identified as Labidocera sp. and larvae of shrimps and crabs compared to Dascyllus stomach samples. Only a few OTUs of cyclopoid and harpacticoid copepods, shrimp, polychaetes, appendicularians, and gastropods were present in Pocillopora tissues.
Mysids, and the cephalochordate, Branchiostoma sp., were absent from the guts analysed.

| Ontogenetic dietary shifts
There was a significant linear negative relationship between fish size (total length) and the number of OTUs in the stomach contents of Chromis (adjusted r 2 = 0.39, p < 0.001) but not for Dascyllus (adjusted r 2 = 0.01, p = 0.5) ( Figure 5). Prey composition appeared to shift during fish ontogeny based on NMDS plots (Figure 4a,b). The gradual change in diet was particularly marked for Chromis when prey abundance was taken into account in calculations of β diversity. A simple linear model explained 37% of the variance in the relationship between the Bray-Curtis index and differences in fish size (total length) (Figure 6b). The relationship was also significant for Dascyllus (p < 0.001) but the model explained a much lower proportion of the variance (adjusted r 2 = 0.02). When taking into account presence-absence to calculate β diversity (Jaccard), the relationships also were significant for both species (Chromis: p < 0.001; Dascyllus: p < 0.001), but the model explained relatively low amounts of the variance (Chromis: adjusted r 2 = 0.09; Dascyllus: adjusted r 2 = 0.02) ( Figure 6a).

| Diversity and abundance of dietary items collected by plankton pumps
Zooplankton diversity in this hyperdiverse system was expected to be very high and dominated by crustacean larvae and copepods.
Given the impossibility of identifying almost all larval forms to species morphologically and the difficulty of identifying most of the copepods to species, identification of the ambient zooplankton was generally at a higher taxonomic level than provided by the sequencing approach. Available monetary resources precluded sequencing the ambient zooplankton samples. However, these samples still provided important insights into the food availability for all three planktivores studied. Both the abundance and diversity of planktonic prey items were significantly higher at night at both depths (t = 6.573, p = <0.001) when Chromis and Dascyllus do not feed (but the Pocillopora does) compared to during the day when the fish and coral both feed. Zooplankton was relatively homogeneously distributed at the two depths during the day. Daytime abundances were low, averaging <20 items m −3 at both depths (  (Table 5). Large-bodied taxa such as amphipods, isopods, crab and shrimp larvae, other crustacean larvae, the predatory calanoid copepod Labidocera sp. and polychaetes were rare in the daytime samples but relatively abundant at night. Eggs were about half as abundant at night as during the day (Table 5).

| D ISCUSS I ON
Resolving whether species and individuals partition their food resources is key to understanding mechanisms potentially allowing a diverse community of species to coexist on coral reefs. The lack of taxonomic resolution in prey identification, the diversity of potential food resources, and the structural complexity of the reef habitat have hindered efforts to confidently identify and quantify dietary differences. It has been particularly challenging for planktivores that feed on small particulate prey, although they represent one of the dominant feeding guilds on coral reefs. In this study, we identify for the first time, pronounced interspecific dietary differences among co-occurring planktivorous fish species, which suggest the role of dietary selectivity in promoting coexistence on coral reefs.
Our investigation focused on the molecular dietary analysis of a coral host and two of its associated planktivorous fish species using high throughput sequencing of a hypervariable mitochondrial gene. Although this approach has been used extensively in microbial ecology, it has seldom been applied to decipher food webs (Leal & Ferrier-Pagès, 2016). It is now a powerful complement to traditional analyses of diets (i.e., morphological analysis of gut contents) owing to methodological improvements (i.e., primer sets for short amplicons) combined with the development of large F I G U R E 5 Diversity of OTUs in relation to fish size (total length in mm). A linear model is fitted with 95% confidence intervals. Chromis: Y = 53-0.5*X, adjusted r 2 = 0.39, p < 0.001; Dascyllus: Y = 37.8-0.1*X, adjusted r 2 = 0.01, p = 0.5. The number of OTUs was calculated from the rarefied data set (n = 259) to control for differences in sampling effort [Colour figure can be viewed at wileyonlinelibrary.com] curated databases of metazoan sequences (Machida, Leray, Ho, & Knowlton, 2017).
We found a diverse range of prey with 699 OTUs belonging to 14 phyla, of which 273 (39%) matched a reference barcode. This is relatively high for a hyperdiverse tropical ecosystem and probably reflects the extensive efforts of the BIOCODE project (Meyer, 2017) at cataloguing and barcoding the marine life of the island of Moorea.
Species that remain uncollected likely belong to morphologically cryptic species, benthic taxa living at greater depths consumed as Our findings reveal that despite having highly similar feeding behaviours and morphologies, the two planktivorous fish species that co-occur in close proximity in branching corals consumed very different portions of the available resources. Levels of interspecific dietary dissimilarity were higher than levels of intraspecific dietary dissimilarity (Table 1), a pattern consistent with predictions of niche theory (Chesson, 2000b). Interestingly, Dascyllus did not significantly shift its diet in the presence of Chromis. This suggests that the two species have intrinsic differences in their feeding behavior and hence, might not strongly compete for food (i.e., no competitive displacement). Chromis and Dascyllus are both visual particulate feeders that use suction feeding to capture individual evasive prey items. However, the presence of stronger caniniform teeth and an enlarged coronoid process in the mandible give Dascyllus an ability to seize prey in addition to sucking prey (Frédérich et al., 2006), which could explain why they are able to feed on a wider array of resources. Although it has never been reported in the literature and we never observed this behaviour, the range of benthic prey in stomachs of Dascyllus could suggest that they also pick prey off the substrate as an alternative feeding strategy. Zooplanktivorous fish rely F I G U R E 6 Correlation between prey composition and differences in fish size (total length in mm). Pairwise differences in prey composition were computed using (a) Jaccard, an incidence based metric and (b) Bray-Curtis, an abundance based metric. Linear models with 95% confidence intervals were fitted to each plot. Jaccard, Chromis: Y = 0.80 + 0.002*X, adjusted r 2 = 0.09, p < 0.001; Jaccard, Dascyllus: Y = 0.88 + 0.001*X, adjusted r 2 = 0.02, p = 0.001; Bray-Curtis, Chromis: Y = 0.50 + 0.014*X, adjusted r 2 = 0.37, p < 0.001; Bray-Curtis, Dascyllus: Y = 0.77 + 0.001*X, adjusted r 2 = 0.02, p < 0.001. Beta diversity was calculated from the rarefied OTU table (n = 259) to control for differences in sequencing depth [Colour figure can be viewed at wileyonlinelibrary.com] (a) (b) on their vision for feeding, and the dietary specialization of Chromis and Dascyllus may be driven largely by their ability to detect prey of different pigmentation, shape or behaviour. For example, the visibility of prey to a fish has been shown to depend upon how well it is able to discriminate various body structures from the background (i.e., light contrast) (Lazzaro, 1987), which is a function of the physical properties of the eyes. The successful capture of prey detected is also contingent on whether the fish is able to discriminate among individual moving prey with various motion patterns and escape strategies. For example, the predatory calanoid copepod Labidocera sp., that explained more than 19% of the variation in dietary composition between the two fish species (Table 3)   Although we did not quantify patterns of foraging by the two species of fish in the water column during this study, larger individuals of both species tend to feed in the upper part of the water column while smaller individuals feed closer to their coral host (Brooks, personal communication). In addition, Chromis do tend to feed higher in the water column than Dascyllus, which indicates that vertical segregation is likely responsible for both inter-and intraspecific dietary differences in both species.
As expected, Pocillopora has a very distinct diet owing to its feeding mode and its position in the water column. Zooplankton are a significant source of nutrients for scleractinian corals, and are essential for the maintenance of metabolic processes and skeletal growth (Ferrier-Pagès, Witting, Tambutté, & Sebens, 2003;Houlbrèque, Tambutté, & Ferrier-Pagès, 2003;Palardy, Rodrigues, & Grottoli, 2008). Pocillopora uses two main feeding strategies to capture zooplankton during both day and night (Séré, Massé, Perissinotto, & Schleyer, 2010). First, polyps immobilize individual prey using nematocyst stings and pull them towards the mouth. Second, polyps extrude mucus webs that they then pull back into the mouth once several prey have been trapped. This passive feeding strategy likely broadens the range of prey ingested. Our study revealed that Pocillopora tissues contained a diversity of fish, present in the mucus as eggs, small larvae and possibly in the form of faeces. Coral tissue also contained numerous microalgae, confirming the importance of herbivory in coral nutrition (Leal et al., 2014).
The molecular approach used here provided dietary information with an unprecedented level of resolution. Yet, it also has shortcomings (Creer et al., 2016;Leray & Knowlton, 2016;Pompanon et al., 2012). For example, primer mismatches likely create biases in the relative number of sequences per taxa (Deagle, Thomas, Shaffer, Trites, & Jarman, 2013). Polymerase chain reaction may also co-amplify secondary prey (i.e., the prey of a prey), thus artificially increasing prey richness (Sheppard et al., 2005) and inflating dietary partitioning. The detection of secondary prey is unavoidable with this sequencing approach. However, they likely make up negligible amounts of DNA, are likely highly digested in comparison to primary prey, and therefore likely account for very few sequences in the data set (Sheppard et al., 2005). For example, OTUs assigned to nonmetazoan groups (e.g., protists, fungi and algae) that are possible secondary prey for the planktivores account for only 1.8% of the sequences in this data set. Our analyses showed clear patterns of dietary partitioning regardless of the beta diversity metric used (i.e., equal or little weight given to rare OTUs) and sequencing effort (i.e., level of rarefaction), pointing to the value of these tools for understanding the complexity of trophic interactions and the role of dietary specialization for the maintenance of biodiversity on coral reefs.

ACK N OWLED G EM ENTS
We