Comparing environmental DNA metabarcoding and underwater visual census to monitor tropical reef fishes

Environmental DNA (eDNA) analysis is a revolutionary method to monitor marine biodiversity from animal DNA traces. Examining the capacity of eDNA to provide accurate biodiversity measures in species-rich ecosystems such as coral reefs is a prerequisite for their application in long-term monitoring. Here, we surveyed two Colombian tropical marine reefs, the island of Providencia and Gayraca Bay near Santa Marta, using eDNA and underwater visual census (UVC) methods. We collected a large quantity of surface water (30 L per filter) above the reefs and applied a metabarcoding protocol using three different primer sets targeting the 12S mitochondrial DNA, which are specific to the vertebrates Actinopterygii and Elasmobranchii. By assigning eDNA sequences to species using a public reference database, we detected the presence of 107 and 85 fish species, 106 and 92 genera, and 73 and 57 families in Providencia and Gayraca Bay, respectively. Of the species identified using eDNA, 32.7% (Providencia) and 18.8% (Gayraca) were also found in the UVCs. We further found congruence in genus and species richness and abundance between eDNA and UVC approaches in Providencia but not in Gayraca Bay. Mismatches between eDNA and UVC had a phylogenetic and ecological signal, with eDNA detecting a broader phylogenetic diversity and more effectively detecting smaller species, pelagic species and those in deeper habitats. Altogether, eDNA can be used for fast and broad biodiversity surveys and


| INTRODUC TI ON
Coral reefs represent the most diverse marine ecosystems on the planet (Fisher et al., 2015) and are also the most threatened (Williams et al., 2019). Due to their structural complexity, they host a large diversity of fish species, from tiny cryptic species to large migratory species (Collins et al., 2019;Darling et al., 2017). Because of this high species diversity, coral reefs have generally been difficult to inventory using traditional survey methods (Plaisance et al., 2011).
Moreover, global changes, including exploitation, pollution, or climate change, are degrading biodiversity on reefs (Cinner et al., 2016;Descombes et al., 2015), but it is difficult to quantify and monitor these impacts because describing species diversity and composition is generally demanding (Costello et al., 2015;Mora et al., 2008).
The monitoring of the biodiversity of coral reefs under global changes could benefit from novel solutions with lower costs and broader applicability complementing traditional methods (Thomsen et al., 2012;West et al., 2020).
Traditionally, monitoring fishes on coral reefs has been performed using underwater visual censuses (UVC) or video surveys , which offer a partial view of the dynamics of reef biodiversity, from their degradation under global changes to their recovery (Bozec et al., 2011;Cinner et al., 2016). These methods are limited in both spatial and temporal coverage and are biased toward certain categories of species (Boussarie et al., 2018). UVC is traditionally used to monitor fish diversity on coral reefs (Samoilys & Carlos, 2000). However, besides logistical difficulties to organize underwater sampling in remote locations, UVC can suffer from several observer biases, such as overlooking cryptobenthic (Bozec et al., 2011) or wideranged species such as sharks . One of the most effective approaches to circumvent the limitations of traditional survey methods in highly diverse ecosystems is environmental DNA (eDNA) metabarcoding (Cilleros et al., 2019;Gomes et al., 2017). eDNA is a noninvasive method demonstrating higher detection capabilities and cost-effectiveness compared to traditional methods, especially when deployed in remote locations (Dejean et al., 2011;Kelly et al., 2014;Thomsen & Willerslev, 2015). Before it can effectively complement traditional sampling methods, the ability of eDNA to recover signals of diversity and composition of marine systems should be evaluated.
Animals leave DNA traces in the environment (Deiner et al., 2017), which may persist from hours to days and can be detected in water samples (Collins et al., 2019;Thomsen et al., 2012). Water filtering followed by a molecular protocol to amplify and sequence target DNA can be used to recover animal DNA present in a given site. Sequences are then taxonomically assigned using a genetic reference database, which provides an integrative inventory of species and composition in aquatic systems (Deiner et al., 2017;Harrison et al., 2019). A recent synthesis counted 54 papers on tropical eDNA, whereas only 15 focused on marine systems Huerlimann et al., 2020;Sigsgaard et al., 2019;West et al., 2020). Compared to freshwater systems, the marine environment has a larger water volume to fish biomass ratio, the movement of molecules in suspension is influenced by various currents, and reef systems can contain up to hundreds of species, which might challenge the detection of individual species (Collins et al., 2019;Hansen et al., 2018;Harrison et al., 2019). Several applications demonstrated that eDNA can recover multiple components of marine ecosystems, including species richness (Jerde et al., 2019), seasonal composition variation (Djurhuus et al., 2020), rare species (Weltz et al., 2017), abundance or biomass (Knudsen et al., 2019;Thomsen et al., 2016), and the occurrence of invasive species (Nevers et al., 2018). Nevertheless, a range of methodological challenges still hampers the broad use of eDNA for the reliable monitoring of marine ecosystems, linked to the choice of markers (Collins et al., 2019;Freeland, 2017), primers sets , laboratory and sequencing protocols (Deiner et al., 2017;Goldberg et al., 2016), and bioinformatic analyses (Calderón-Sanou et al., 2020;Juhel et al., 2020), which implies further testing of the eDNA methodology in situ.
Tropical ecosystems have historically been underrepresented in research (Collen et al., 2008), and increased monitoring efforts in these regions are urgently needed, particularly under ongoing global change (Barlow et al., 2018). Different abiotic conditions and high species richness might challenge the application of eDNA in the tropics (Huerlimann et al., 2020;Jerde et al., 2019). Studies of eDNA on coral reefs have shown a strong potential for biodiversity detection (Nguyen et al., 2020;Sigsgaard et al., 2019;West et al., 2020), but the scope of methodological testing remains narrow. Dibattista et al. (2017) used fish-specific 16S mitochondrial DNA to monitor fish diversity in the Red Sea, but captured only a fraction of the local fish species pool. Stat et al. (2019) compared the signal of eDNA with observations from baited videos and detected >30% more generic richness using the combination of approaches than when either method was used alone. Sigsgaard et al. (2019) used eDNA with fish-specific 12S mitochondrial DNA across a network of sites in the Gulf of Oman and recovered sequences from a diverse assemblage of marine vertebrates, which covered approximately onethird of the bony fish genera previously recorded in this area. Using a combination of markers, West et al. (2020) detected a wide range of organisms and showed that their composition varied significantly between habitats across an entire island in the Coral Sea. Hence, attempts to survey tropical marine fish assemblages using eDNA are yielding increasingly informative results, supporting the use of seawater to trace the molecular signatures of biodiversity for monitoring purposes. Here, we compared the compositional patterns of the fish community using eDNA metabarcoding and UVCs in two different reef ecosystems in the Colombian Caribbean, the oceanic island of Providencia and Gayraca Bay in the Tayrona National Natural Park near Santa Marta.
We investigated (a) whether the species recovered with three different sets of 12S primers are complementary and consistent with species recovered with UVC; (b) whether there is a correspondence between species richness within each genus and family recovered using both eDNA and UVC, as well as a correspondence between the number of reads within each genus and family and the number of individuals; and (c) whether the divergence between biodiversity recovered with eDNA and UVC has a phylogenetic or ecological component. Additionally, we explored (d) the signal of β diversity across eDNA samples by analyzing the compositional species dissimilarity between geographic locations.

| Study areas
The study focuses on two regions of Colombia, the island of Providencia and the Tayrona National Natural Park, with extensive coral reef habitats ( Figure 1, Table S1). Providencia is located in the southwestern Caribbean Sea and is included in the UNESCO Seaflower Biosphere Reserve of Colombia. This island, which is part of the San Andres, Providencia, and Santa Catalina Archipelago, comprises a complex barrier reef on a calcareous platform surrounding an extinct Miocene volcano (Sánchez et al., 1998). The high habitat diversity provides a wide range of substratum types and coral reefs (Geister, 1992;Márquez, 1987), which shape the diversity, abundance, and distribution of coral reef fishes (Mejía & Garzón-Ferreira, 2000). The Tayrona National Natural Park is located along the continental Colombian Caribbean coast bordering the Sierra Nevada de Santa Marta.
Tayrona Park has a heterogeneous coastal topography composed of metamorphic rocks, with numerous rocky headlands, islets, and bays (Garzón-Ferreira & Díaz, 2003). Coral and other hardbottom communities are distributed along the coast, mainly as fringing reefs, while seagrass beds, mangroves, and coral reefs have developed to some extent in sheltered conditions within the bays (Garzón-Ferreira & Cano, 1991). The study was carried out in Gayraca Bay, where corals on the exposed side exhibit mainly massive to encrusting growth forms with colonies and a reef-like structure.

| Underwater visual censuses
Divers conducted underwater visual censuses, using scuba equipment to survey the composition and abundance of fishes in Providencia and in Gayraca Bay. The surveys were performed during multiple years: -2003, 2006-2007in Providencia and 1999-2011, 2013 in Tayrona National Natural Park. Data were collected using the 30-min timed roving diver fish survey method for the established depths, 4-10 m in Providencia, and 8-14 m in Gayraca, inventorying all the observed species and estimating abundances in categories following the Coral Reef Monitoring System (SIMAC) methodology (CARICOMP, 1994(CARICOMP, , 1997(CARICOMP, , 2001Garzón-Ferreira et al., 2002).
In cases of fish schools abundance was estimated in tens. Four censuses per station were implemented, resulting in a total of 120 min of sampling in each monitoring event. In Providencia, the survey was performed in eight different habitats within the reef complex ( Figure 1) and included a total of 4,200 min of sampling.
Furthermore, seagrass habitats were also sampled in four 30-min roving diver visual surveys within a predefined area of 2,500 m 2 .
In Tayrona  We followed a strict contamination control protocol in both field and laboratory stages (Goldberg et al., 2016;Valentini et al., 2016). Each water sample was processed using disposable gloves and single-use filtration equipment.

| OBITools filtering analyses for taxonomic assignments
Following sequencing, reads were processed to remove errors and analyzed using programs implemented in the OBITools package (http://metab arcod ing.org/obitools, Boyer et al., 2016) based on a previous protocol . The forward and reverse reads were assembled with the ILLUMINAPAIREDEND program, using a minimum score of 40 and retrieving only joined sequences.
The reads were then assigned to each sample using the NGSFILTER software. A separate data set was created for each sample by splitting the original data set into several files using OBISPLIT. After this step, each sample was analyzed individually before merging the taxon list for the final ecological analysis. Strictly identical sequences were clustered together using OBIUNIQ. Sequences shorter than 20 bp, or with fewer than 10 occurrences were excluded using the OBIGREP program. The OBICLEAN program was then run within a PCR product. All sequences labeled "internal," which most likely correspond to PCR substitutions and indel errors, were discarded. Taxonomic assignment of the remaining sequences was performed using the ECOTAG program with the NCBI reference sequence (www.ncbi.nlm.nih.gov, release 233, downloaded on 11 October 2019). Considering the assignment of a few sequences to the wrong samples due to tag jumps (Schnell et al., 2015) and index hopping (MacConaill et al., 2018), all sequences with a frequency of occurrence <0.001 per taxon and per library and all sequences with an occurrence of <0.0006 per taxon in the RapidRun were discarded. Sequences with <100 reads in each sample were also discarded. These thresholds were empirically determined to clear all reads from blanks and controls and were included in our global data production procedure as suggested in De Barba et al. (2014).
After the filtering pipeline, the extraction and PCR negative controls were completely clean, and no sequence reads remained in those samples.

| SWARM clustering analyses for MOTU identification
For the teleo primer set only, we used a second bioinformatics workflow based on sequence clustering using SWARM, an algorithm that groups multiple variants of sequences into MOTUs (Molecular Operational Taxonomic Units; Mahé et al., 2014;Rognes et al., 2016).
Reads were assembled using VSEARCH (Rognes et al., 2016), then trimmed using CUTADAPT (Martin, 2013) and clustered using SWARM (Mahé et al., 2014). The clustering algorithms use sequence similarity and co-occurrence patterns to delineate meaningful entities, by grouping together sequence variants generated due to PCR and sequencing errors. Sequences were first merged using VSEARCH. CUTADAPT was then used for demultiplexing and primer trimming, and sequences containing ambiguities were removed with VSEARCH. SWARM was run with a minimum distance of one mismatch to make clusters. Once MOTUs were generated, the most abundant sequence within each cluster was used as a representative sequence for taxonomic assignment. <1/1,000 reads per PCR run (i.e., tag jumps; Schnell et al., 2015) and occurring in only one PCR run from a single sample (Ficetola et al., 2015). We corrected for tag jumps following the same procedure as for the OBITools workflow.

| Taxonomic comparison of eDNA and underwater visual censuses
For both pipelines, taxonomic assignments were corrected to avoid over-confident assignment outputs from ECOTAG: We only validated identification for 100% (species level), 90%-99% (genus level), or 85%-99% (family level) identity matches, when possible. Using the outputs of the OBITools pipeline, we compared the species, genera, and families recovered by eDNA to those recorded by UVC in Providencia and Gayraca Bay. We first compared the overlap in the list of species, genera, and families recovered with each of the three 12S primers targeting vertebrates, Actinopterygii and Elasmobranchii.
Second, we evaluated whether the species, genera, and families recov- We analyzed whether detection differences between eDNA and UVC represented a phylogenetic signal and were associated with ecological traits. We performed this analysis at the genus level because the coverage of the reference database at the species level was sparser. We excluded all genera not represented in the reference database (10 genera were not detected with eDNA, were not in the reference database, but were detected in UVCs). We classified the remaining genera into ( A value around zero means that the trait is distributed on the tree as if it had evolved following a Brownian model (Fritz & Purvis, 2010).
We used the distribution of 100 super-trees (Rabosky et al., 2018) pruned at the genus level. Next, we related detection classes to a set of ecological traits assembled for each species and aggregated at the genus level. Ecological traits were gathered from FishBase (Froese & Pauly, 2018) and included body size (small <15 cm, medium and large >40 cm), trophic guild (carnivore, herbivore, piscivore, planktivore), position in the water column (benthopelagic, demersal, pelagic, reef-associated, pelagic), home range mobility (sedentary, mobile, highly mobile), and schooling behavior (of a single of two individual, schools of 3-20 individuals, schools of >20 individuals). Based on these traits, we calculated a gower distance matrix between genera and constructed a trait space using a Principal Coordinates Analysis (PCoA). We mapped and estimated the trait volume recovered by each method to identify the differences between eDNA and UVCs.
We plotted trait modalities as ellipses encompassing 90% of the genera of each modality.

| Diversity, abundance, and spatial variation in eDNA samples
We used the MOTU outputs from the SWARM protocol to perform diversity and composition analyses that did not strictly depend on We also investigated the differences in eDNA composition between the sampling stations in Providencia and Gayraca Bay together and within Providencia separately. From the MOTU presence-absence matrix, we calculated a Jaccard distance matrix. To ordinate the compositional differences between the eDNA samples collected in both sampling sites, we performed a PCoA on this distance matrix.
Using the same method, we performed a second PCoA analysis to investigate the compositional difference between the eDNA samples collected in the Providencia sampling stations. Sampling around this island covered multiple sites, following a gradient from sheltered locations to very exposed areas to marine currents. For each PCoA, we reported the explained deviance of each axis and mapped the ordination values in the geographic space.
We further calculated the pairwise Jaccard's dissimilarity index (Anderson et al., 2011; β jac ) of the compositional difference in MOTUs between (a) Providencia and Gayraca Bay and (b) between the west and east coast of Providencia. This index is expressed as:

| Comparison between eDNA primers using OBITools
For Providencia, we detected a total of 107 different species when all three primer sets were used. We detected 53 species using the teleo primers, 74 species using the Vert01 primers, and five species exclusively of Elasmobranchii using the Chon01 primers. Using the teleo and Vert01 primers together we detected all 107 species, whereas we detected 53 species when the teleo and Chon01 primers were used together and 80 when the Vert01 and Chon01 primers were used together. We detected 19 species in common between the teleo and Vert01 primers, five between the teleo and Chon01 primers, and none between the Vert01 and Chon01 prim-  Figure 2d).

| Comparison of species detection between eDNA and UVC
A total of 113 species were recorded in the UVCs around Providencia.
Using all three primers together, with eDNA we detected 35 (31%) of the 113 species that were observed in the UVCs. Out of these species, we detected 20 with the teleo primers, 25 with the Vert01 primers and 2 with the Chon01 primers. On the other hand, we detected 72 species with eDNA that were not observed during the UVCs. Overall, 41 out of 106 genera detected with eDNA were also recorded by UVC. We recorded some reef-associated species, such as the yellowhead wrasse  (Tables S5-S7).

| Comparison of species richness and abundances between eDNA MOTUs and underwater visual surveys
We performed the aggregation into MOTUs using the teleo primers, as the bioinformatics clustering pipeline using SWARM has only been developed and fully tested with this primer (Juhel et al., 2020;Marques et al., 2020). In Providencia, the eDNA clustering pipe- We tested the correlation between species richness and numbers of MOTUs in Providencia and Gayraca Bay ( Figure S1). In Providencia, we found a significant correlation between the number of species per genus and the number of MOTUs per genus (Spearman correlation test, n = 30, ρ = .37, p = .04). The genera Urobatis, Scarus, and Hypoplectrus were identified as outliers in these correlations.
We found a weaker correlation between the number of species per family and the number of MOTUs per family (n = 23, ρ = .33, p = .13).
The number of individuals was also correlated with the number of MOTU reads per genus (n = 30, ρ = .4, p = .03, Figure S2). The gen- The trait space obtained by performing a PCoA (percentage of inertia, axis 1: 24.4% and axis 2: 15.4%) on a set of ecological traits assembled for each species and aggregated at the genus level. The dark grey polygon represents the trait space covered by genera sampled by eDNA, whereas the light grey polygon represents the trait space covered by genera sampled by UVC. On the trait space, we drew an ellipse representing 90% of the points belonging to a trait category for the following traits: body size (small <15 cm, medium and large >40 cm), trophic guilds (carnivore, herbivore, piscivore, planktivore), position in the water column (benthopelagic, demersal, pelagic, reef-associated, pelagic), schooling behaviour (small groups of 1 or 2 individuals, medium groups of species gathering in schools of 3-20 individuals, schooling species of >20 individuals). In all plots, orange circles represent genera detected by the eDNA sampling method only, red by UVC only, and blue by both methods [Colour figure can be viewed at wileyonlinelibrary.com] p = .9). Finally, there was not a significant correlation between the number of individuals and the number of MOTU reads per family in this region (n = 12, ρ = .28, p = .38).

| Ecological and phylogenetic distribution of species detection
We investigated the ecological and phylogenetic distributions of detection considering all genera recorded by either eDNA or UVC and also included in the reference database. We examined the phylogenetic signal of the detection in either eDNA, UVC, or both.
For the genera detected by UVC, we found an average D-statistic of 0.18 ± 0.1 across the 100 trees, indicating that the clustering of genera identified by this monitoring technique is not different than expected under a Brownian model (p = .28 ± .12; Figure 3a).
In contrast, for the genera detected by eDNA, we found an average D-statistic of 1.16 ± 0.15, indicating that these taxonomic units detected by eDNA are widely distributed across the phylogenetic tree, as expected under a model of random phylogenetic signal (p = .66 ± .18; Figure 3a).
We related the detection classes to ecological traits using PCoA.
The percentage of inertia of the first axis of the PCoA was 24.4%, while the percentage of inertia of the second axis was 15.4%. We found that a large proportion of ecological traits was covered by the two sampling methods, even if UVC detected a smaller number of genera than eDNA. eDNA was better at detecting large piscivore and pelagic species belonging to genera such as Istiophorus, Euthynnus, Decapterus, Acanthocybium, and Strongylura, but also smaller planktivorous species of Sardinella, Cetengraulis, Lycengraulis, and Engraulis (Figure 3b). eDNA further detected more small and bottom-associated species represented by the genera Liopropoma, Hypsoblennius and Arcos.

| Spatial variation in eDNA MOTUs
We investigated MOTU composition dissimilarity among samples and found marked differences between the eDNA samples col- with 25.6% for the first axis and 17% for the second axis. We found marked differences in eDNA composition between the eastern and western sides of the island ( Figure S3). When exploring the difference between the west and east coast of Providencia, we found that the MOTU composition differed moderately (β jac = 0.27) and 97.6% of the β jac was turnover (β jtu = 0.267; β jne = 0.006). The two sides of the island had 165 MOTUs in common out of the total of 227 identified. With more taxa, the western side included some species typically associated with complex habitats of seagrasses and reef patches, such as the hogfish (Lachnolaimus maximus) and Sygnathus sp.

| D ISCUSS I ON
We showed that eDNA metabarcoding can provide a comprehensive overview of fish composition in two highly diverse tropical marine reefs of Colombia. UVC is traditionally used to monitor fish diversity on coral reefs (Samoilys & Carlos, 2000). However, besides logistical difficulties to organize underwater sampling in remote locations, UVC can suffer from several observer biases, such as overlooking cryptobenthic (Bozec et al., 2011) or wideranged species such as sharks . Compared with UVCs performed over two decades (1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017), the eDNA surveys from one year detected a large fraction of the fish species diversity, including many species that were not recorded during UVCs, and covered a wider fraction of the phylogeny and ecological space of the ichthyofauna. Moreover, we showed that eDNA has a marked spatial signal, both between the two investigated regions and within the Providencia region, supporting future local habitat monitoring of reefs using eDNA (West et al., 2020). Together, our analyses support the use of eDNA as an approach for the fast monitoring of highly diverse tropical marine ecosystems. In an eDNA study using a different marker (CO1) to detect fish, Nguyen et al. (2020) (Brandl et al., 2018). Further, eDNA sampling delivered potential new records of species for the stud- Because some taxa were detected by eDNA but not by UVC, and vice versa, we further analyzed the difference in detection between the two approaches. As the most obvious cause of discrepancy, species and genera found in the UVCs but not detected in the eDNA were missing from the reference database. We found that 60% of the genera that were recorded during UVCs but not detected by eDNA were not in the reference database extracted from NCBI, highlighting that the reference database is central to effective eDNA monitoring . Overall, eDNA analysis led to the recovery of a larger number of genera, covering a larger fraction of the phylogenetic tree and of the ecological space of fishes ( Figure 3). The fish on coral reefs tend to be phylogenetically diverse, with representatives of multiple families (Leprieur et al., 2016). We found that the genera detected using eDNA had a wide spread across the fish phylogenetic tree, while the genera observed during UVCs were phylogenetically clumped. Our results suggest that eDNA surveys are more representative than UVCs of the entire phylogenetic diversity of fishes on coral reefs. We found a positive correlation in diversity and abundance between the two sampling approaches in Providencia but not in Gayraca Bay. While the UVC sampling effort was high in Providencia, with eight UVCs targeting different habitats, the effort was lower in Gayraca Bay, where only two sites were sampled, which could explain the difference in signal between regions.
Together, this indicates a general limitation of the comparison proposed in this study, that we do not know the true compositions and abundances, as both sampling approaches involve some level of bias. Longer term, synchronous eDNA sampling and video recording could provide further validation of eDNA .
Besides species diversity, eDNA is also expected to provide information on the spatial distribution of species assemblages across different habitats (Nguyen et al., 2020;West et al., 2020). In agreement with findings from previous studies (Closek et al., 2019;Nguyen et al., 2020)  composition. Indeed, our approach captured marked differences between Gayraca Bay and Providencia, but also more locally between the east and west coasts of Providencia, corresponding to variation in habitat. The island of Providencia is composed of various habitats, and the eastern side is more exposed than the western one (Coralina-Invemar, 2012). Geomorphological diversity of the coral reef system, added to the combination of oceanic influences and terrigenous contributions from the island, lead to high variety in underwater environments and coastlines (Díaz et al., 2000). We found that the eastern side of the island has a species composition dominated by species associated with reef habitats, such as the blackear wrasse (Halichoeres poeyi) and the redtail parrotfish (Sparisoma chrysopterum); the western side is characterized by species asso-  (Deiner et al., 2017), but some uncertainties remain as regard to sampling design  and the choice of markers (Collins et al., 2019;Stat et al., 2017) and bioinformatics pipeline (Calderón-Sanou et al., 2020;Juhel et al., 2020). We tested three different primer sets for the 12S region looking for fish taxa, but we did not find a universal marker able to detect all taxa. The teleo primer generally performed best, as it was able to retrieve many teleost species, as well as five of the six species of Elasmobranchii also detected with the Chon01 primer in Providencia and one taxa of the same group at the family level in Gayraca. Nevertheless, the teleo primer did not recover some of the species that were recovered by the Vert01 primer (54 vs. 74 in Providencia and 39 vs. 64 in Gayraca), while the Vert01 primer did not recover a few species only found with the teleo primer (33 and 21 for Providencia and Gayraca, respectively). Hence, as this stage of primer development and testing, it appears that a multiprimer approach is required to capture of the entire diversity of a site (West et al., 2020). Moreover, because we found many Elasmobranchii with the teleo primer, a specialized primer for Elasmobranchii might not be needed and could be replaced by the more ubiquitous teleo primer.
In that regard, teleo is an exception among eDNA primers because other sets, such as the MiFish primers, do not amplify Elasmobranchii (Bylemans et al., 2018;Miya et al., 2015).
A mayor limitation of eDNA is the lack of completeness of the reference database. Yet, a high coverage of the reference database is crucial to allow future accurate identification of species assemblages. In fact, many species recorded by UVC were not recovered with eDNA simply because they were not represented in the reference database. In order to fully exploit the potential detection power of eDNA metabarcoding, a vast effort is needed to improve taxonomic coverage of reference databases (Schenekar et al., 2020;Weigand et al., 2019). Addressing these important database gaps requires analyses that are not based solely on species assignment.
We generated MOTUs using SWARM to get an indication of the expected overall biodiversity. However, while some MOTUs perfectly delineate true biological species without the need of a reference sequence, a fraction of these MOTUs also represent errors stemming from PCR and sequencing, overestimating true diversity (Morgan et al., 2013;Reeder & Knight, 2009), while clustering might also bind together distinct closely related species, underestimating true diversity (Huse et al., 2010). Thus, procuring a taxonomically comprehensive database with high-quality sequences and accurate data curation steps is crucial for producing robust and reproducible ecological conclusions from eDNA metabarcoding methods (Collins et al., 2019;Weigand et al., 2019).
Alternative ways to survey marine biodiversity beyond UVCs and unbiased evaluations of the ecosystem components are needed, as these provide a baseline for the management of marine protected areas . eDNA metabarcoding is becoming a more accessible method that generates reliable information for ecosystem surveillance and could prove valuable in marine monitoring programs (Lacoursière-Roussel et al., 2016). Here, we show that eDNA quickly provides a detailed picture of fish diversity and composition in two marine protected areas of Colombia, which can be used for future monitoring and management of these sites (Bálint et al., 2018).
Despite water exchange in coastal marine systems, eDNA signals are localized on coral reefs, which is promising for monitoring the health status of these ecosystems. Repeated observations of eDNA measurements at multiple stations in these areas will facilitate assessment of the status and ultimately trends in biodiversity, particularly in response to disturbance events associated with climate change (Berry et al., 2019) or pollution (Bagley et al., 2019). Our results further highlight the importance of establishing a complete reference database for eDNA analyses, as many of the sequences could not be attributed to a particular genus or species. As shown for lake ecosystems (Hänfling et al., 2016), eDNA could become an important complement to traditional UVCs for monitoring coral reef biodiversity.

CO N FLI C T O F I NTE R E S T
All authors declare that there is no conflict of interest regarding the publication of this article.