MiFish, a set of universal PCR primers for metabarcoding environmental DNA from fishes: detection of more than 230 subtropical marine species

We developed a set of universal PCR primers (MiFish-U/E) for metabarcoding environmental DNA (eDNA) from fishes. Primers were designed using aligned whole mitochondrial genome (mitogenome) sequences from 880 species, supplemented by partial mitogenome sequences from 160 elasmobranchs (sharks and rays). The primers target a hypervariable region of the 12S rRNA gene (163–185 bp), which contains sufficient information to identify fishes to taxonomic family, genus and species except for some closely related congeners. To test versatility of the primers across a diverse range of fishes, we sampled eDNA from four tanks in the Okinawa Churaumi Aquarium with known species compositions, prepared dual-indexed libraries and performed paired-end sequencing of the region using high-throughput next-generation sequencing technologies. Out of the 180 marine fish species contained in the four tanks with reference sequences in a custom database, we detected 168 species (93.3%) distributed across 59 families and 123 genera. These fishes are not only taxonomically diverse, ranging from sharks and rays to higher teleosts, but are also greatly varied in their ecology, including both pelagic and benthic species living in shallow coastal to deep waters. We also sampled natural seawaters around coral reefs near the aquarium and detected 93 fish species using this approach. Of the 93 species, 64 were not detected in the four aquarium tanks, rendering the total number of species detected to 232 (from 70 families and 152 genera). The metabarcoding approach presented here is non-invasive, more efficient, more cost-effective and more sensitive than the traditional survey methods. It has the potential to serve as an alternative (or complementary) tool for biodiversity monitoring that revolutionizes natural resource management and ecological studies of fish communities on larger spatial and temporal scales.


Introduction
Environmental DNA (eDNA) in aquatic environments refers to genetic material found in the water column. In the case of multicellular organisms, eDNA originates from various sources, such as metabolic waste, damaged tissue or sloughed skin cells [1]. Ficetola et al. [2] was the first study demonstrating the use of eDNA for detecting an aquatic vertebrate species (invasive American bullfrog) from controlled environments and natural wetland, published in 2008. Subsequently, eDNA from fishes has been detected from various aquatic environments, including ponds [3][4][5], streams [6], rivers [7][8][9][10] and seawater [11,12]. Such ubiquitous presence of eDNA from fishes in the water column has led to the increasing use of this technique as a tool for detections of invasive [3,[7][8][9], rare or threatened species [5,6], investigations of local fauna [10,13], or in a larger mesocosm [12] with known species composition. These pioneering studies have shown the use of eDNA to be appropriate as a non-invasive genetic monitoring tool in various fields of fish biology.
For monitoring the occurrence of a single or few fish species, short species-specific eDNA fragments (72-312 bp) have been used [3,[5][6][7][8][9], with earlier studies detecting those species based on the presence/absence of PCR products by visually inspecting the products on an agarose gel stained with ethidium bromide [7][8][9]. More recently, quantitative PCR (qPCR) using probe-based chemistries has been employed for the detection of target species [3][4][5][6] owing to the method's sensitivity, specificity and potential to quantify the target DNA [6]. For example, Takahara et al. [4] estimated the biomass of common carp (Cyprinus carpio) in a natural freshwater lagoon, using the qPCR approach (real-time PCR), based on the positive relationships between eDNA concentrations and biomass in aquaria and experimental ponds.
For monitoring fish assemblages with broader taxonomic scopes, Minamoto et al. [10] designed degenerate PCR primers to amplify a short fragment of the mitochondrial cyt b gene (285 bp) with reference to those sequences from the local freshwater fish fauna. Based on PCR amplification of the fragment and subsequent subcloning and sequencing of the product, they successfully detected multiple species in eDNA from the controlled aquaria (one to five spp.) and three stations in the Yura River, central Japan (two to four spp.) [10]. Thomsen et al. [11] developed two generic and four species-specific PCR primer sets for amplifying short fragments of the cyt b gene (32-51 bp), in order to detect marine fish species from three sampling sites at a coastal zone in Denmark. Using a next-generation sequencing (NGS) platform (Roche 454 GS FLX), they detected 15 species in the amplicons, including both important commercial fishes as well as some species rarely recorded by conventional monitoring methods [11]. More recently, Kelly et al. [12] attempted to estimate the fish fauna in a large tank at the Monterey Bay Aquarium with known species composition by sequencing PCR amplicons from eDNA using an NGS platform (Illumina MiSeq). They used a set of published universal PCR primers to amplify a 106 bp fragment of the mitochondrial 12S rRNA gene [14] for metabarcoding fish species in the tank. Although they detected seven of the eight species of bony fishes present, they were able to identify those species only to taxonomic family or genus owing to the limited sequence variability within the amplicons. In addition, they failed to detect all three elasmobranchs (sharks and rays) contained in the tank [12].
These earlier studies on eDNA metabarcoding (high-throughput multispecies identification using degraded DNA extracted from an environmental sample [15]) have shown both potential and limitations. They are non-invasive and are demonstrably more efficient and cost-effective than the traditional monitoring methods, such as visual surveys, trawls and seines [11,12]. The former two studies [10,11], however, required development of PCR primers specifically designed with reference to DNA sequences from the known local fish fauna and those primers are of limited uses in future studies with little prior knowledge on the faunal composition. The latter study [12] employed PCR primers that have been developed using the computer software 'ECOPRIMERS' [14] and that are supposedly universal among vertebrates. Despite the use of universal primers, the successful detection in the aquarium tank was dependent on the taxonomic groups (e.g. no detection for ocean sunfish and all elasmobranchs), and the amplified products, if any, exhibited little sequence variability to correctly assign fish species in the same family or genus [12].
The primary objective of this study was to circumvent these problems associated with PCR primers. To achieve this goal, we: (i) developed universal primers for fish eDNA that amplify a short fragment (less than 200 bp) containing sufficient sequence variation to correctly assign fish species; (ii) tested versatility of the primers across a taxonomically and ecologically diverse range of fishes using eDNA from aquarium tanks with known species compositions; and (iii) preliminarily examined the use of the primers for detecting eDNA from fishes inhabiting natural seawater environments with unknown species composition and abundances in an open ecosystem.
The development of the universal primers (MiFish-U/E) was based on the aligned whole mitochondrial genome (mitogenome) sequences from 880 fish species, which was supplemented by partial mitogenome sequences from 160 elasmobranchs. The primers are targeted to amplify a hypervariable region of the 12S rRNA gene (163-185 bp), which contains sufficient information to unambiguously identify fishes we tested to taxonomic family, genus and species, with one exception (closely related congeners of Thunnus). We tested the versatility of those PCR primers using eDNA from four tanks in the Okinawa Churaumi Aquarium and from natural seawaters near the aquarium in the subtropical western North Pacific. Using a high-throughput Illumina MiSeq platform, we detected eDNA from 232 fish species from those seawaters, which are taxonomically diverse and are distributed across 70 families and 152 genera. In addition to eDNA, this metabarcoding approach is applicable to bulk samples (total DNA), such as those from net collections containing a diverse range of fish eggs, larvae, juveniles or damaged specimens with few diagnostic characters present for species identification.

. Selection of genetic marker
Mitochondrial DNA (mtDNA) was chosen as the genetic marker because copy number of mtDNA is greater than that of nuclear DNA per cell, and detection rate therefore is expected to be higher in the former, even where DNA is present at a low concentration and/or is degraded [16]. In order to select a suitable region in the mitogenome for species identification based on eDNA, 1044 whole mitogenome sequences were batch downloaded from the database MITOFISH v. 2.80 [17] in a FASTA format as of 20 April 2013. After removing problematic sequences involving large-scale gene rearrangements [18], the remaining 880 sequences (electronic supplementary material, table S1) were subjected to multiple alignment using MAFFT v. 6.956 [19] with a default set of parameters. The aligned sequences were imported into MESQUITE v. 2.75 [20] for visual inspection of the conservative and hypervariable regions. The search for a short hypervariable region (up to 200 bp for paired-end sequencing using the Illumina MiSeq) flanked by two conservative regions (ca 20-30 bp) across 880 species was performed on the entire set of aligned mitogenomes. The conservative and hypervariable regions were highlighted by a 'Select' function in MESQUITE (a submenu 'Variable among taxa' in 'Select Characters') [20].

Primer design
To facilitate primer design based on comparisons of diverse sequences from 880 fish species, a base composition for a selected position in the conservative region was shown using a 'Show Selection Summary Strip' function in MESQUITE [20]. The base compositions in selected characters were manually recorded in a spreadsheet for the primer design. In the primer design process, we considered a number of technical tips that enhance the primer annealing to the template without the uses of degenerate bases [21]:

Primer testing with extracted DNA
In order to test whether these newly designed PCR primers were universal or not, we first tested MiFish-U-F/R (no adapter sequences) using extracted DNA from 96 species representing all the four major lineages of fishes (Agnatha, Chondrichthyes, Actinopterygii and Sarcopterygii) placed in 47 orders and 96 different families (table 1). Double-stranded DNA concentrations from those fishes were measured with a NanoDrop Lite spectrophotometer (Thermo Fisher Scientific, Wilmington, DE, USA) and the extracted DNA was diluted to 15 ng µl −1 using Milli-Q water. PCR was carried out with 30 cycles of a 15 µl reaction volume containing 8.3 µl sterile distilled H 2 O, 1.5 µl 10 × PCR buffer (Takara, Otsu, Japan), 1.2 µl dNTPs (4 mM), 1.5 µl of each primer (5 µM), 0.07 µl Taq polymerase (Z Taq; Takara) and 1.0 µl template. The thermal cycle profile after an initial 2 min denaturation at 94 • C was as follows: denaturation at 98 • C for 5 s; annealing at 50 • C for 10 s; and extension at 72 • C for 10 s with the final extension at the same temperature for 5 min.
Double-stranded PCR products were purified using Exo SAP-IT (USB, Cleveland, OH, USA) to remove redundant dNTPs and oligonucleotides from primers. Direct cycle sequencing was performed with dye-labelled terminators (BIGDYE terminator v. 1.1; Applied Biosystems, Foster City, CA, USA) following the manufacturer's protocol and the purified PCR products were sequenced for both strands on the ABI 3130xl Genetic Analyzer (Life Technologies, Carlsbad, CA, USA). The DNA sequences were edited and assembled using GENETYX-MAC v. 17 (Genetyx, Tokyo, Japan) and deposited in DDBJ/EMBL/GenBank databases.

In silico evaluation of interspecific variation
Interspecific differences within the amplified DNA sequences are required for accurate assignments of taxonomic categories. To computationally evaluate levels of interspecific variation in the target region (hereafter called 'MiFish sequence') across different taxonomic groups of fishes, 1361 whole mitogenome sequences were batch downloaded from MITOFISH v. 2.89 [17] as of 3 September 2014. After removing duplicate sequences (e.g. multiple sequences from subspecies), uncertain taxonomic status (e.g. hybrids) and possible erroneous sequences (e.g. unable to annotate using MITOANNOTATOR [17]), the MiFish sequences were extracted from the remaining 1324 sequences using custom Ruby scripts (available from: http://dx.doi.org/10.5061/dryad.54v2q) and they were subjected to calculation of pairwise edit distances. The edit distance quantifies dissimilarity of sequences in bioinformatics [23] and is defined as the minimum number of single-nucleotide substitutions, insertions or deletions that are required to transform one sequence into the other. For comparisons, metabarcode sequences amplified by 12S-V5 primers [14] (forward: 5 -ACTGGGATTAGATACCCC-3 ; and reverse: 5 -TAGAACAGGCTCCTCTAG-3 ) (hereafter called 'ecoPrimer sequences') were also extracted from the 1324 sequences and their interspecific variation was evaluated as described for MiFish sequences. The ECOPRIMER pair amplifies the same gene (mitochondrial 12S rRNA gene) as that of the MiFish-U/E primers, but the two primer pairs are designed to amplify two different regions adjacent to each other (12S-V5-F primer is located within MiFish-U-R primer). The ECOPRIMER pair was used in a metabarcoding study of fishes by Kelly et al. [12] who attempted to estimate an artificial fish fauna using eDNA in the large tank at the Monterey Bay Aquarium.  Table 1. A list of fish species for testing MiFish-U primers (without adapter sequences) using extracted DNA diluted to 15 ng µl −1 , subsequently sequenced with a Sanger method.

Water sampling and DNA extraction
All sampling and filtering equipment was exposed to a 10% bleach solution for at least 30 min before use. For water samplings in the aquarium, approximately 10 l of seawater was collected from the surface using multiple casts of an 8 l polyethylene bucket fastened to a 10 m rope. The bucket was thoroughly prewashed with tank water. The sampling was conducted between 10.00 and 13.00 before daily feeding on two consecutive days (2 and 3 June 2014). The sampled water was stored in a valve-equipped 10 l book bottle and immediately brought to the laboratory before subsequent filtering. For water samples from the coral reefs near the aquarium, 10 l of seawater was collected in a similar manner on 4 June and 7 November 2014. One to three 2 l lots of seawater from the 10 l samples were vacuum-filtered onto 47 mm diameter glass-fibre filters (nominal pore size, 0.7 µm; Whatman, Maidstone, UK). Each filter was wrapped in commercial aluminium foil and stored in −20 • C before eDNA extraction. Two litres of Milli-Q water was used as the negative control and treated identically to the eDNA samples, to monitor contamination during the filtering and subsequent DNA extraction.
DNA was extracted from the filters using the DNeasy Blood and Tissue Kit (Qiagen, Hilden, Germany) in combination with a spin column (EZ-10; Bio Basic, Markham, Ontario, Canada). After removing the attached membrane from the spin column (EZ-10), the filter was tightly folded into a small cylindrical shape and placed in the spin column. The spin column was centrifuged at 6000g for 1 min to remove redundant seawater for DNA extraction. The column was then placed in a new 2 ml tube and subjected to lysis using proteinase K. Before lysis, Milli-Q water (400 µl), proteinase K (20 µl) and buffer AL (180 µl) were mixed and the mixed solution was gently pipetted onto the folded filter in the spin column. The column was then placed on a 56 • C preheated aluminium heat block and incubated for 30 min. The spin columns were covered with commercial aluminium foil and a clean blanket for effective incubation at the specified temperature. After the incubation, the spin column was centrifuged at 6000g for 1 min to collect the DNA. In order to increase DNA yields from the filter, 300 µl of sterilized TE buffer was gently pipetted onto the folded filter and the spin column was again centrifuged at 6000g for 1 min. The collected DNA solution (ca 900 µl) was purified using the DNeasy Blood and Tissue Kit following the manufacture's protocol.

Paired-end library preparation and MiSeq sequencing
Two to five eDNA samples from each of the four aquarium tanks (total 14 samples; figure 1a-d) and four eDNA samples from the coral reefs (figure 1e,f ) were used for multiplex PCR using two universal primer pairs (MiFish-U/E). Of these 18 eDNA samples, five samples from the Kuroshio tank were additionally used for multiplex PCR using two universal plus one genus-specific primer pairs (MiFish-U/E/tuna) for correct assignments of Thunnus species. Prior to library preparation, work-space and equipment were sterilized, filtered pipet tips were used and separation of pre-and post-PCR was carried out to safeguard against contamination. We also employed controls to monitor contamination including PCR blanks for each experiment.
Massively parallel paired-end sequencing on the MiSeq platform (Illumina, San Diego, CA, USA) requires PCR amplicons to be flanked by: (i) primer-binding sites for sequencing; (ii) dual-index (i.e. barcode) sequences; and (iii) adapter sequences for binding to the flowcells of the MiSeq. We employed a two-step tailed PCR approach to construct the paired-end libraries (figure 2).
The first-round PCR (first PCR; figure 2) amplified the target region using primers 5 -ACACT CTTTCCCTACACGACGCTCTTCCGATCTNNNNNN + MiFish gene-specific sequences-3 (forward) and 5 -GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTNNNNNN + MiFish gene-specific sequences-3 (reverse). The first 33 and 34 nucleotides (nt) are partially used for primer-binding sites for sequencing and the following six random hexamers (N) are used to enhance cluster separation on the flowcells during initial base call calibrations on the MiSeq platform. Illumina); the 5 -end sequences are adapters that allow the final product to bind or hybridize to short oligos on the surface of the Illumina flowcell; and the 3 -end sequences are priming sites for the MiSeq sequencing.
The first PCR product was diluted 10 times using Milli-Q water and used as a template for the second PCR. The second PCR was carried out with 12 cycles of a 12 µl reaction volume containing 6.0 µl 2× KAPA HiFi HotStart ReadyMix, 0.7 µl each primer (5 µM), 3.6 µl sterile distilled H 2 O and 1.0 µl template. Different combinations of indices (chosen from A/D501-508 for forward primers and A/D701-712 for reverse primers) were used for different templates for a massively parallel sequencing using the MiSeq platform. The thermal cycle profile after an initial 3 min denaturation at 95 • C was as follows: denaturation at 98 • C for 20 s; annealing and extension combined at 72 • C (shuttle PCR) for 15 s with the final extension at the same temperature for 5 min.
The indexed second PCR products were pooled in equal volumes and the pooled libraries (total 100 µl) were subjected to agarose gel electrophoresis using 2% L03 (Takara). A target size of the libraries (ca 370 bp) was excised from the gel and purified using a MinElute Gel Extraction kit (Qiagen) with an elution volume of 12 µl. The library concentration was estimated using a Qubit dsDNA HS assay kit and a Qubit fluorometer (Life Technologies). Double-stranded DNA concentration of the pooled library was adjusted to 4 nM (assuming 1 bp equals 660 g mol −1 ) using Milli-Q water and 5 µl of the 4 nM library was denatured with 5 µl of fresh 0.1 N NaOH. Including HT1 buffer (provided by the Illumina MiSeq v. 2 Reagent kit for 2 × 150 bp PE), the denatured library (10 µl; 2 nM) was diluted to the final concentration of 12 pM for sequencing on the MiSeq platform. A 30 µl of PhiX DNA spike-in control (12 pM) was added to improve data quality of low diversity samples such as single PCR amplicons used in this study.

Data pre-processing
An overall quality of the MiSeq reads was evaluated by the programs FASTQC (available from http://www.bioinformatics.babraham.ac.uk/projects/fastqc/) and SUGAR [24]. After confirming a lack of technical errors in the MiSeq sequencing, low-quality tails were trimmed from each read using DynamicTrim.pl from the SOLEXAQA software package [25] with a cut-off threshold set at a Phred score of 10 (= 10 −1 error rate) [26]. The tail-trimmed paired-end reads (reads 1 and 2) were assembled using the software FLASH [27] with a minimum overlap of 10 bp. The assembled reads were further filtered by custom Perl scripts in order to remove reads with either ambiguous sites (Ns) or those showing unusual lengths with reference to the expected size of the PCR amplicons (297 ± 25 bp). Finally, the software TAGCLEANER [28] was used to remove primer sequences with a maximum of three-base mismatches and to transform the FASTQ [29] format into FASTA.

Taxonomic assignment
The pre-processed reads from the above custom pipeline were dereplicated using a 'derep_fulllength' command in UCLUST [30], with the number of identical reads added to the header line of the FASTA formatted data file. Those sequences represented by more than or equal to 10 identical reads were subjected to the downstream analyses and the remaining under-represented sequences (with less than 10 identical reads) were subjected to pairwise alignment using a 'usearch_global' command in UCLUST. If the latter sequences observed from less than 10 reads showed more than or equal to 99% identity with one of the former reads (one or two nucleotide differences), they were operationally considered as identical (owing to sequencing or PCR errors and/or actual nucleotide variations in the populations) and they were added to the more than or equal to 10 reads.
The processed reads were subjected to local BLASTN searches [31] against a custom-made database. The latter was generated by downloading all whole and partial fish mitogenome sequences deposited in MitoFish [17] and whole mitogenome sequences from tetrapods deposited in NCBI Organelle Genome Resources (http://www.ncbi.nlm.nih.gov/genomes/OrganelleResource.cgi taxid=32523) to cover those tetrapods occurring in aquatic environments. In addition, the custom database was supplemented by assembling new sequences in M.M.'s laboratory (electronic supplementary material, table S3). As of 4 October 2014, the database covers approximately 4230 fish species distributed across 457 families and 1827 genera. According to the latest edition of 'Fishes of the World' [32], fishes comprise 515 families, 1827 genera and 27 977 species with our custom-made database covering 88.7% of the families, 40.6% of the genera and 15.1% of the species.
The top BLAST hit with a sequence identity of more than or equal to 97% and E-value threshold of 10 −5 was applied to species assignments of each representative sequence. We found that this cut-off value maximally recovered the species composition from each tank, while avoiding erroneous taxonomic assignment. Reliability of the species assignments were evaluated based on a ratio of total alignment length and number of mismatch bases between the query and reference sequences. For example, if a query sequence was aligned to the top BLAST hit sequence with an alignment length of 150 bp with one mismatch present, the ratio was calculated as 150/(1 + 1). Value one is added to the denominator to avoid zero-divisors. This ratio was calculated for the top and second BLAST hit species, and a log of odds ratio (LOD) score between these ratios was used as the comparable indicator of the species assignment. Results from the BLAST searches were automatically tabulated, with scientific names, common names, total number of the reads and representative sequences noted in an HTML format. Moreover, biological information for each detected species is available from the hyperlink in the table, such as that of FishBase (http://fishbase.sinica.edu.tw), Barcode of Life (http://www.boldsystems.org), GBIF (http://data.gbif.org), MitoFish (http://mitofish.aori.u-tokyo.ac.jp) and NCBI (http://www.ncbi.nlm. nih.gov) for quick evaluation and credibility of the bioinformatic identification.

Results and discussion
We visually inspected the aligned sequences throughout the entire mitogenomes across the 880 species (electronic supplementary material, table S1) by highlighting variable and invariable sites using MESQUITE [20]. After repeated inspections, we found a short hypervariable region (ca 170 bp) within the 12S rRNA gene, which was flanked by highly conservative regions (ca 20-30 bp) across the 880 species (table 2). Note that we were unable to find such a region within the barcoding region of the aligned COI gene sequences, which have been frequently used as the marker of choice also in fishes [33]. This observation is consistent with a recent argument against the use of the COI gene as a genetic marker for metabarcoding studies [34].
The hypervariable region in the 12S rRNA gene includes multiple segments that are forming big loops in a proposed secondary structure of the molecule [35,36]. In particular, four segments of the loops were so variable in length (involving multiple insertions/deletions) that they were considered unalignable even among closely related gobioid fishes in a previous study [37]. The two highly conservative regions, on the other hand, exhibit no length variations among the 880 species and were located on the two stem regions (stem nos. 15/16 and 24/25 in [35,36]), which undergo secondary structural constraints through strong Watson-Crick base pairings [35]. Following these empirical and theoretical observations, we decided to design a new primer pair located on the two conservative regions, thereby amplifying the highly taxonomic informative hypervariable region in between.
In the initial stage of this study, we designed degenerate PCR primers to accommodate sequence variations among taxa, but found that such degenerate primers did not amplify the target eDNA when they were used with long adapter sequences in the tailed PCR (figure 2). We redesigned a new set of primers without degenerate sites (MiFish-U) using various technical methods related to construction of adequate primers (see Material and methods). The new forward (MiFish-U-F) and reverse (MiFish-U-R) primers consist of 21 and 27 bases (table 2) with G/C contents of 57% and 44% and T m of 56.6 • C and 56.5 • C, respectively.
With the redesigned MiFish-U primers (without adapter sequences), we confirmed successful amplifications of the hypervariable regions using extracted DNA from 96 species representing all of the four major lineages of fishes (Agnatha, Chondrichthyes, Actinopterygii and Sarcopterygii) distributed across 47 orders and 96 different families (table 1). With these PCR products, we successfully determined their nucleotide sequences using the conventional Sanger sequencing method. All the sequence data are available from DDBJ/EMBL/GenBank databases with accession numbers shown in table 1.

MiFish-E
During the preliminary experiments using eDNA from the aquarium tanks, we found that only a few assembled reads from the MiSeq sequencing represented elasmobranchs (sharks and rays). The lack of elasmobranch sequences was totally unexpected, because we included a number of elasmobranchs while designing the universal primers (13 spp.; see the electronic supplementary material, table S1) and more than 100 large-sized individuals of various elasmobranchs (mostly more than 1 m in total lengths; figure 1a) were present and active in the Kuroshio tank. We suspected that absence of the elasmobranch sequences resulted from PCR bias derived from primer-template mismatches. Inspection of the newly downloaded 160 elasmobranch sequences found only a few such mismatches (table 3), with significant ones being restricted to two sites near the 5 -end of the forward primer and in a single site near the 3 -end of the reverse primer. The newly designed primers for the elasmobranchs based on these mismatches were proved effective for amplification of the region, with all the species with reference sequences being detected by the MiSeq sequencing (see below). The new forward (MiFish-E-F) and reverse (MiFish-E-R) primers were designed in an identical region to that of the universal primers, consisting of 21 and 27 bases (table 3) with G/C contents of 52% and 41% and T m of 54.1 • C and 55.2 • C, respectively, and were used with MiFish-U in multiplex PCR.

MiFish-tuna
In addition to newly constructed pairs of the universal primers (MiFish-U/E), preliminary experiments showed that nucleotide differences in the MiFish sequences from tunas (seven species of Thunnus) were so small that the bioinformatic pipeline was unable to assign assembled reads to the correct species (see below). We visually inspected the entire mitogenome sequences from the seven species of tunas and found a region with sufficient interspecific variations among constituent species. The newly designed genus-specific forward (MiFish-tuna-F) and reverse (MiFish-tuna-R) primers amplify a portion of the ND5 gene (180 bp), consisting of 22 and 21 bases with G/C contents of 55% and 57% and T m of 56.9 • C and 57.8 • C, respectively (see table 3 for primer sequences with adapters).

In silico evaluation of interspecific variations
The pairwise edit distances from MiFish and ecoPrimer sequences were calculated for all combinations of 1324 fish species distributed across 59 orders, 319 families and 890 genera (total 1324 C 2 = 875 826 pairs) and the resulting distances were sorted into between-order, family, genus and species (table 4). As expected from the size difference between MiFish and ECOPRIMER sequences (average lengths 172 bp versus 106 bp), the former appears to have more variation than the latter and also outperforms the latter in unambiguously assigning each taxonomic category (table 4). In particular, MiFish sequences perform well for higher taxonomic categories; for example, all the between-order edit distances are larger than 10 in MiFish sequences, while the smallest one in ECOPRIMER sequences is three (four pairs). Also, two pairs of the between-family edit distances from ECOPRIMER sequences are zero, indicating that interfamilial discrimination is not feasible for these two pairs. For lower taxonomic categories such as genus and species, MiFish sequences also outperform ECOPRIMER sequences in terms of unambiguous taxonomic assignments. For example, the number of pairs with smaller between-genus and species edit distances (e.g. less than or equal to 3) in MiFish sequences are 4.17 and 2.48 times lower than those in ECOPRIMER sequences, respectively (table 4).
It appears that MiFish sequences still have inherent limitations to unambiguously assign lower taxonomic categories, such as genus and species. Actually, there are 32 and 98 between-genus and specific pairs with the edit distances of zero, respectively (table 4). For those taxonomic groups with no or a few nucleotide differences in MiFish sequences, we need to develop new molecular markers that contain sufficient information to discriminate constituent species. Development of the new marker for correct species assignments of tunas in this study (MiFish-tuna) represents a good example of such a case (see below).
It should also be noted that those zero distances in the intergeneric comparisons from MiFish sequences (total 32 pairs) are restricted mostly to specific groups of fishes, such as Cichlidae (cichlids; 14 pairs) and Istiophoridae (billfishes; 14 pairs), whose limited genetic divergences in mtDNA are well established (and sometimes misleading owing to gene introgression) compared with their distinct morphological divergences [38][39][40]. The remaining four pairs include that of Cyprinidae (carp and minnow), Engraulidae (anchovy), Mormyridae (freshwater elephantfish) and Mirapinnidae (hairyfish), all of which are under taxonomic revisions at various taxonomic categories [41][42][43][44]. Actually, a recent study [42] demonstrated that members of the latter family Mirapinnidae simply represent larval stages of the different whalefish families, indicating that current fish taxonomy is still in a state of flux.   Table 4. Frequency distributions of the interspecific edit distances of the MiFish (above) and ECOPRIMER (below) sequences among 1324 fish species deposited in the MitoFish database [16]. (The edit distances are sorted into between-order, family, genus and species. Only edit distances from 0 to less than or equal to 10 are shown.) We first tested MiFish-U primers (without adapter sequences) using eDNA from the aquarium tanks in preliminary experiments and observed consistent amplifications across different samples on an agarose gel stained with ethidium bromide (results not shown). The PCR bands from those amplifications, however, were often smearing, with occasional extra bands being observed outside the expected size of the products (ca 220 bp).
Following the partial success of PCR using eDNA, we constructed MiFish-U primers for the first PCR by appending adapter sequences at their 5 -ends (figure 2; for primer sequences, see table 5). Optimal experimental conditions for the first PCR with these primers were achieved through trial and error, and we found that choice of a PCR kit (KAPA HiFi HotStart ReadyMix) and associated high-annealing temperatures (65-67 • C) in the first PCR are the two most important factors contributing to successful amplifications showing distinct single PCR bands on the agarose gel.
Based on the above empirical observations, we constructed 14 dual-indexed, paired-end libraries through two-step tailed PCR (figure 2) for two to five water samples from each of the four aquarium tanks.

MiSeq sequencing and data analysis
The MiSeq paired-end sequencing (2× 150 bp) of the 14 libraries, together with another 129 libraries (total number of libraries = 143), yielded a total of 14.86 million reads, with an average of 95.0% base calls being Phred quality scores of more than or equal to 30.0 (Q30; error rate = 0.1% or base call accuracy = 99.9%). This run was highly successful considering that the quality scores specified by Illumina is more than 80% bases higher than Q30 at 2 × 150 bp (Illumina Publication no. 770-2011-001 as of 27 May 2014).
After demultiplexing and subsequent pre-processing of the raw data from MiSeq, the outputs were subjected to the BLAST searches for taxonomic assignment. In total, 4 322 882 reads were assigned to fish species with more than or equal to 97% identity to reference sequences in the custom database. Of these, 4 053 184 (93.4%) are identified as those fishes contained in one of the four tanks (hereafter called 'tank species') and the remaining 286 446 (6.6%) are derived from 'non-tank species' (table 6), discussed below.
According to the unpublished monthly report from the aquarium, the four tanks harboured a diverse range of 249 fish species distributed across 64 families and 146 genera at the time of sampling. Of these 249 species, we confirmed that 180 species have reference sequences in the custom database (tables 7 and 8) and detected eDNA from 168 species (93.3%; table 6). In the following, we describe and discuss results from the metabarcoding analyses of each tank separately.

Kuroshio tank
The  Table 5. A list of primers for the first and second PCR used in the paired-end library preparation for the MiSeq analyses; indices (=barcodes) are highlighted with an underline. (Note that those index sequences for the reversal primers (R) are read by MiSeq on the opposite strand and should be reverse/complement in the sample sheet for MiSeq runs.)      a Those reads with less than 97% sequence identity are excluded from the above table for simplicity. They are 285 172 reads in total; 57 572 reads from the Kuroshio, 222 897 reads from the tropical fish, 1093 reads from the deep-sea and 3610 reads from the mangrove tanks, respectively. than 10 m in total length). It predominantly keeps large-sized fishes characteristic to areas around the Kuroshio, one of the western boundary currents flowing northeastwards along the entire length of Japan, including the Okinawa Islands. Preliminary experiments showed that the exclusive use of an MiFish-U primer pair was unable to detect most species of the elasmobranchs (including whale sharks); subsequent development of MiFish-E primers and application of multiplex PCR (MiFish-U/E), however, enabled us to detect all species of the elasmobranchs contained in the tank (table 7). Out of the 63 fish species with reference sequences in the custom database, we detected 61 species (96.8%) including 17 and 44 species of elasmobranchs and teleosts, respectively, which are collectively distributed across 17 families and 44 genera (table 7). The two undetected species (3.2%) are carangids (Carangoides orthogrammus and Pseudocaranx dentex; table 8) and we visually confirmed their presence in the tank. There were no extra carangid sequences referable to those two species in the MiSeq outputs, suggesting that they may represent an example of false negative in our metabarcoding analyses.
Although yellowfin and Pacific bluefin are the only tuna species contained in the Kuroshio tank, our custom bioinformatic pipeline erroneously assigned assembled reads into supposedly six tuna species (table 9). This is apparently owing to small interspecific nucleotide differences among the seven species of tunas, with a mean pairwise p-distance of only 2. 22 (table 7). It should be noted that MiFish-U/E primers also amplified eDNA from a non-fish marine vertebrate (spotted dolphin, Stenella attenuata) also present in the Kuroshio tank (excluded from table 7). We actually found many reads from the dolphin across the five samples totalling 37 056. A comparison between the primer sequences of MiFish-U-F/R and priming sites of the dolphin (EU557096) indicates that there is only one mismatch in the middle of the forward primers (excluding two T/G bonds), suggesting that the primers are also useful for detecting non-fish vertebrates by accommodating their unique nucleotide variations at the priming sites.  we confirmed reference sequences for 105 species in the custom database (tables 7 and 8) and detected eDNA from the 95 species distributed across 32 families and 65 genera (tables 6 and 7). The detection rate (90.5%) is somewhat lower than those of the other tanks (96.8-100%; table 6) and the 10 undetected species are taxonomically diverse, distributed across 10 families within 10 genera (table 8). We visually recognized the presence of these 10 species in the tank and reconfirmed detection of eDNA from the same families or genera of those 10 species. This suggests that strong PCR bias derived from primertemplate mismatches seems unlikely and the lack of eDNA from these 10 fish species may represent false negatives. Note that co-occurrences of multiple species from some of the speciose genera, such as Epinephelus (five spp.), Lutjanus (six spp.) and Scarus (four spp.) (table 7), do not confuse the taxonomic assignments, because all undetected species from these genera show significant nucleotide differences from those congeners (p-distance = 2.9−16.6%). The detection rate might also be affected by uncertainty in the species identification based on morphology for the tank species and/or for voucher specimens of the reference sequences. The large species diversity in this tank (155 spp.) also highlights the importance for taxonomic coverage of the reference sequences in the custom database [45], which only attain approximately twothirds of the tank species (105 spp.). For the tropical fish tank, we subjected 1 524 620 reads to BLAST searches and were unable to assign 222 897 reads (14.6%) into any species with more than or equal to 97% sequence identity (not shown in  Figure 4. Compositions of the non-tank species (with more than or equal to 97% sequence identity to reference sequences in the custom database) for eDNA from the four tanks in the Okinawa Churaumi Aquarium. Percentages in parentheses are based on the total number of reads with sequence identity of more than or equal to 97% (table 6). For classification of the non-tank species, see text.

Tropical fish tank
than 97% sequence identity are referable to the tank species when they have congeners in the reference sequences and represent single members of those genera in the respective tanks (see footnotes in table 7). By contrast, such cases are quite rare in the tropical fish tank and presence of multiple confamilial or congeneric species with less than 97% sequence identity hinders further taxonomic assignments.

Deep-sea tank
The deep-sea tank (figure 1c) keeps 15 species of benthic and benthopelagic fishes from elasmobranchs to higher teleosts commonly found in slope waters off Okinawa. Of these 15 deep-sea fish species, we confirmed reference sequences for 13 species in the custom database (table 7) and detected all of these 13 species with eDNA (100%; tables 6 and 7).

Mangrove tank
The mangrove tank exhibits the brackish-water mangrove swamps in Okinawa (figure 1e), keeping eight species of teleosts common to those environments. We confirmed reference sequences for all of these eight teleosts in the custom database (table 7) and detected eDNA from all of them (100%; tables 6 and 7).

Detection of non-tank species
The most serious pitfall of eDNA is the risk of contamination, which remains among the greatest experimental challenges to this field [45,46]. To avoid such risk, we performed decontamination procedures for laboratory spaces and equipment and physically separated pre-and post-PCR work spaces (see Material and methods), which are known to significantly limit the contamination [47]. Despite these efforts, a total of 286 446 reads (6.6%) were considered as those from non-tank species and most of them may represent false positives from various sources. In a similar metabarcoding study using universal primers, Kelly et al. [12] reported that approximately 25.5% of the tank sequences were assigned to taxa not living in the mesocosm tank (non-tank species) at the Monterey Bay Aquarium.
Although this study is not designed to rigorously determine the extent of detection rates of such false positives, it would be useful for future eDNA research using the metabarcoding approach to list possible sources of the non-tank species as exogenous DNA with some comments. They can tentatively be classified into: (i) other tank species (62 218 reads; 23.8%); (ii) species from other libraries on the same run (8925 reads; 3.1%); (iii) fish feed (86 204 reads; 30.1%); (iv) non-fish vertebrates (68 735 reads; 2.4%) excluding a spotted dolphin contained in the Kuroshio tank; and (v) unknown (116 264 reads; 42.3%) ( figure 4).
One of the most noteworthy examples is detection of non-tank species showing abundant reads in their respective tanks. Those tank species with pooled reads of more than 100 000 were consistently found across other tanks and even from some negative controls, including four species of tunas and mackerels (Rastrelliger kanagurta, Thunnus albacares, T. orientalis, Katsuwonus pelamis) plus a fussiler (Pterocaesio marri) from the Kuroshio tank, a parrotfish (Scarus ghobban) from the tropical fish tank, a snake mackerel (Thyrsitoides marleyii) from the deep-sea tank and a moonyfish (Monodactylus argenteus) from the mangrove tank. The occasional detection of those reads in the negative controls strongly suggests cross contamination in the laboratory, which seems unavoidable in eDNA studies using PCR amplifications  [45]. Although we are unable to pinpoint the experimental step of such contamination, PCR-amplified eDNA during the library preparation, which generate billions of DNA copies in a single reaction, would be the most critical source for large amounts of exogenous DNA [45].
Detection of such non-tank species can be partly explained by re-intake of discharged seawater from the aquarium as it continuously pumps fresh seawater into the facility from the outer reef slope at a depth of 20 m (350 m offshore). Subsequently, the water is directed to various tanks after filtration and is finally led through a drain discharging on the same outer reef slope. Because of the close proximity of the influx and outflow of water (300 m separation), eDNA from non-tank species are likely to occasionally circulate in other tanks as exogenous DNA.
We also encountered putatively exogenous DNA from other libraries ( figure 4), which notably consists of subarctic pelagic and benthic fishes from the Bering Sea and adjacent waters (e.g. salmon, northern smoothtongue, sculpins; 8925 reads; 3.1%). All of these dual-indexed paired-end libraries were constructed in other laboratories and cross contamination is highly unlikely. Kircher et al. [48] demonstrated such misassignment on the Illumina sequencing platform and the Illumina document (pub. no. 770-2013-046 as of 20 November 2013) recently acknowledged that it can occur during the demultiplexing, a process by which reads are assigned to the sample of origin.
Another source of exogenous DNA includes fish feed (e.g. mackerel, herring, flying fish). They are predominant in the Kuroshio tank ( figure 4) where large amounts of those fishes are regularly fed to large-sized elasmobranchs, teleosts and dolphins. We also detected exogenous DNA from non-fish vertebrates (figure 4), mostly from that of humans and domesticated animals such as chickens and pigs, similar to that observed in the mesocosm tank at the Monterey Bay Aquarium [12]. Human eDNA is obviously present from staff diving and maintenance, whereas domesticated animal DNA have frequently been found in chemical reagents [49].
Finally, significant amounts of eDNA from non-tank species are derived from unknown sources other than fish or non-fish vertebrates listed above (116 264 reads; 40.6% among non-tank species and 2.5% among tank + non-tank species). Most of those reads comprise eDNA from non-subtropical marine and freshwater fishes from various localities. It should be noted that such dubious reads are few in eDNA from natural seawater (see below), only comprising 0.58% (5502 reads) of the total reads with more than or equal to 97% sequence identity (954 326 reads). This suggests that seawater from the aquarium tanks contain more exogenous DNA with unknown sources than those from natural environments. Further investigations are needed to more rigorously specify the identity of those dubious sequences from unknown sources.

Primer testing with eDNA from natural seawaters
In addition to the aquarium tanks, we also sampled natural seawater from a rocky coast around the coral reef nearby the aquarium (figure 1e,f ) on two separate days (4 June and 7 November 2014). Using eDNA from four 2 l samples, we prepared four dual-indexed libraries and they were subjected to the MiSeq paired-end sequencing. After demultiplexing and subsequent pre-processing of the raw data from MiSeq, the outputs were subjected to the BLAST searches for taxonomic assignments. In total, 954 326 reads were assigned to fish species with more than or equal to 97% sequence identity to reference sequences in the custom database, of which 948 824 (99.4%) were putatively considered as endogenous eDNA.
From the four water samples, we detected 93 fish species distributed across 36 families and 62 genera (table 10). We confirmed that all of these species occur in the subtropical western North Pacific, although most of them are not particularly obvious and colourful, usually small-sized and/or fossorial reefassociated fishes unsuitable for the aquarium display. Of these 93 fish species, 64 are unique in these samples not detected in the four aquarium tanks and 11 families are new to the taxonomic list (table 10). Unfortunately, there is no background faunal information on fishes in this area, and we are unable to compare the present results with those from previous studies.

Concluding remarks
With the use of newly developed universal primers (MiFish-U/E) and a high-throughput NGS platform (Illumina MiSeq) in a metabarcoding approach to fish eDNA, we confirmed the detection of 232 fish species distributed across 70 families and 152 genera from four aquarium tanks and coral reefs in the subtropical western North Pacific. Those 232 species are not only taxonomically diverse, ranging from sharks and rays to higher teleosts, but are also greatly varied in their ecology, including both pelagic and Table 10. Taxonomic composition and read numbers for the 93 species of teleost fishes detected in the MiSeq analyses of eDNA samples from a rocky coast near the aquarium. (Only those species with identity more than or equal to 97% are shown with numbers of pooled reads from two samples. Asterisks indicate those species also occur in the four aquarium tanks (    (total DNA), such as those from net collections containing multiple life stages and damaged specimens with no diagnostic characters for species identification. Furthermore, the detection of various mammals suggests the broad applicability of this approach to non-fish vertebrates with slight modifications of primer sequences to accommodate unique nucleotide variations among those organisms.

Family Serranidae
Nevertheless, there are several methodological challenges that must be addressed before this metabarcoding approach is likely to become a mainstream technology in fish biodiversity research. The first one would be to explore a method that generates a greater diversity of MiFish sequences at a lower cost to avoid PCR dropouts (=false negatives). Those taxa that are prone to the dropouts can potentially skew the relative abundance in eDNA sequences, making it difficult to assess biologically relevant differences across taxonomic groups [34]. Considering stochasticity of individual PCR reactions and PCR bias derived from primer-template mismatches, optimal number of PCR replicates and use of multiple annealing temperatures should be explored to comprehensively detect fish eDNA without the dropouts. In a fungal metabarcoding study, pooling multiple repeated PCRs and using multiple annealing temperatures were recommended to facilitate the recovery of more correct species richness [50].
The second one is false positives that are consistently observed in our metabarcoding analyses of the four aquarium tanks (figure 4). Although sources of the majority of those reads (57.7%) can be identified (e.g. exogenous DNA from other tank species, other libraries, fish feed, non-fish vertebrates), there are a significant number of reads from unknown sources other than the former (42.3%; 2.5% of the total number of reads with more than or equal to 97% sequence identity). Such dubious reads are relatively few in eDNA from the coral reefs near the aquarium (0.58%) and subsequent analyses of eDNA from oceanic waters that are remote from human activities support this observation (results not shown). This also illustrates the limits of the eDNA metabarcoding approach that cannot discriminate sources of eDNA from either exogenous or endogenous origins.
The third one is completeness of the reference sequence database, which is indispensable for correct taxonomic assignments. Reference sequences in the custom database used in the present analyses were derived from two data sources. The first one is MitoFish, from which all whole mitogenome sequences (1324 sequences) and partial mitogenome sequences containing MiFish sequences (2953 sequences) were obtained. The second one is supplementary MiFish sequences assembled in M.M.'s laboratory (648 sequences; electronic supplementary material, table S3). In total, it covers approximately 4230 fish species distributed across 457 families and 1827 genera as of 4 October 2014. Obviously, this taxonomic coverage is far from satisfactory, considering the enormous diversity of fishes with at least 27 977 species placed in 515 families and 1827 genera [32]. Nevertheless, total number of fish whole mitogenome sequences in MitoFish [17] has steadily increased since its 2006 onset and the number of original MiFish sequences has increased considerably as a result of recent massive sequencing of the two large tissue collections (figure 5), currently reaching 2364 sequences from a wide variety of fish taxa. Obviously, our custom-made database for newly designed eDNA markers is not compatible to that of other online resources. For example, the Fish Barcode of Life project (http://www.fishbol.org/index.php) currently