Species detection from aquatic eDNA: Assessing the importance of capture methods

Environmental DNA (eDNA) is increasingly used for biodiversity monitoring, particularly in aquatic systems. However, each step, from sample collection to bioinformatic analysis, can introduce biases and influence the reliability of results. While much effort has been put into the optimization of laboratory methods,


| INTRODUC TI ON
Environmental DNA (eDNA) is increasingly used in biodiversity monitoring, both for the targeted detection of particular species (e.g., invasive species or species of conservation concern) and for characterizing the composition of whole biological communities (e.g., Thomsen et al., 2012). This DNA-based monitoring approach can be applied to a range of environments, but most studies in the fields of animal ecology and bioassessment focus on aquatic ecosystems. However, despite the advantages of eDNA to identify aquatic species, its application is still largely under evaluation due to the inherent complexity of the approach. Each step can introduce errors and biases, including sample collection, DNA extraction, DNA amplification, high-throughput sequencing, and bioinformatic pipelines (Zinger et al., 2019). A number of considerations should be taken into account when applying eDNA approaches in aquatic ecosystems and robust experimental designs are needed in order to increase the confidence on the conclusions obtained from eDNA surveys.
While many studies have focused on the optimization of laboratory methods to deal with the challenges associated with biodiversity monitoring using eDNA, comparatively less attention has been given to the evaluation of capture methods. Multiple methods are available for eDNA capture from water bodies, to the extent that few studies share the same methods (Dickie et al., 2018).
Environmental DNA capture either by centrifugation (e.g., Caldwell, Raley, & Levine, 2007), precipitation with sodium acetate and ethanol (e.g., Ficetola, Miaud, Pompanon, & Taberlet, 2008), or filtration (e.g., Jerde, Mahon, Chadderton, & Lodge, 2011) has previously been applied in aquatic systems. Centrifugation and precipitation methods are restricted to a low sample volume (usually 15 ml) which can hamper species detection, especially for low-density species (Herder et al., 2014). In contrast, filtration methods allow the capture of eDNA from larger volumes of water with previous studies reporting volumes ranging from 250 ml (Barnes et al., 2014) up to 100 L .
The most common filters used in the field are 47 mm disc filters, also called open filters. These are usually associated with small volumes due to their small surface area (c. 17 cm 2 ). An alternative, more recent, approach has been the utilization of enclosed filters (hereinafter referred to as capsules) (e.g., Lopes et al., 2017;Valentini et al., 2016). Capsules thus far used by the eDNA community have surface areas ranging from 4.5 cm 2 (Vences et al., 2016;Millex ref SLGV033RS) to 1,300 cm 2 (e.g., Lopes et al., 2017;Valentini et al., 2016;Envirochek HV ref 12099). The larger surface areas allow the filtration of much greater volumes of water, but to the best of our knowledge, a comparison of the performance of high-capacity capsules (defined here as having a surface area of more than 100 cm 2 ) to more common eDNA methods is still missing. Vences et al. (2016) did a small test with two high-capacity capsules (1,300 cm 2 ; Envirochek HV ref 12099), which the authors noted was not sufficient for statistical comparison. Spens et al. (2017) compared the performance of capsules with ethanol precipitation and disc filters; however, the surface area of the capsules used was rather small (10 cm 2 ; Sterivex ref SVGPL10RC).
Like the capture method, the species detection method applied to a given set of eDNA samples will also play an important role in any eDNA study, as different methods will have different sensitivities. Although PCR-free methods have been used to analyze eDNA samples, the most common approach is to use PCR to facilitate species detection. Currently, two main PCR-based methods are used: quantitative PCR (qPCR) and metabarcoding (PCR followed by high-throughput sequencing, HTS). qPCR is generally utilized as a species-specific assay, while metabarcoding is utilized for the simultaneous detection of multiple species and thus to assess community composition (e.g., Bálint, Nowak, Márton, & Pauls, 2017). The major difference between these two approaches is thus related to the range of organisms the eDNA survey needs to cover and the specificity of the primers used. Metabarcoding studies target a large group of species and often a single primer set is not enough to cover the biodiversity intended, whereas qPCR studies only require one set of primers to detect their target species. Metabarcoding studies are also susceptible to taxon bias, where DNA from some species is amplified more efficiently than others, potentially leading to rare species not being detected for example. The use of multiple markers (nuclear and/or mitochondrial), with different lengths, or the use of multiple primers (both group and species-specific) has previously been suggested to overcome such biases in metabarcoding studies (Harper et al., 2018). Nevertheless, metabarcoding becomes more beneficial in more diverse systems (Thomsen & Willerslev, 2015), allowing the detection of multiple species while being less time-consuming and more cost-efficient than qPCR. A common challenge to both approaches is the existence of errors and incomplete reference databases. Missing sequence information for a particular species will have an impact not only when trying to obtain a species identification but also at an earlier stage when designing the assay. These problems can be minimized using reliable genetic databases such as GenBank, that has been shown to have a very small percentage (<1%) of taxonomic errors and mislabeled sequences (Leray, Knowlton, Ho, Nguyen, & Machida, 2019), despite previous concerns raised about its accuracy (e.g., Harris, 2003). A comparison between qPCR and metabarcoding is thus essential to determine if they provide comparable results. Notwithstanding, it remains unclear which method is the best for species-specific studies, with previous research reporting different outcomes, ranging from a higher performance of qPCR (Lacoursière-Roussel, Dubois, Normandeau, & Bernatchez, 2016) to similar performance between both methods (Murray et al., 2011).
Another source of variation in eDNA research lies on the number of positive samples or replicates needed to consider species presence, to which there are no standard guidelines at the moment (Goldberg et al., 2016). This will have an impact on the final results given that different thresholds will result in different species lists (e.g., Alberdi, Aizpurua, Gilbert, & Bohmann, 2018;Allali et al., 2017;Deagle, Thomas, Shaffer, Trites, & Jarman, 2013;Mata et al., 2019). Less stringent thresholds can overestimate the presence of a species (false positives), while strict thresholds might fail to detect it despite its presence in the site (false negatives), with consequences for downstream conservation effort (Thomsen & Willerslev, 2015).
In the present study, we focused on amphibians, which are currently considered the most threatened group of vertebrates worldwide (Wake & Vredenburg, 2008), with an estimated 40% of species in danger of extinction (Bishop et al., 2012). Given their declines, the need for powerful and cost-effective methods for amphibian surveys is becoming increasingly important. The use of molecular eDNA techniques has been shown to be more efficient than traditional field surveys for amphibian detection in several cases (Dejean et al., 2012;Smart, Tingley, Weeks, Van Rooyen, & McCarthy, 2015;Valentini et al., 2016). However, amphibians often inhabit turbid environments (Lobos, Cattan, Estades, & Jaksic, 2013;Schmutzer, Gray, Burton, & Miller, 2008), such as agricultural ponds (Ferreira & Beja, 2013;Knutson et al., 2004) or shallow lakes (Jackson & Moquin, 2011), where sampling of eDNA is challenging due to reduced performance of filtration methods associated with high-sediment loads clogging filters (Hinlo, Gleeson, Lintermans, & Furlan, 2017). The efficiency of eDNA studies in turbid waters remains poorly known, and few studies have addressed the difficulties of biodiversity assessment in these environments (Egeter et al., 2018). Due to their large surface area, capsules allow the filtration of large volumes  and could help overcome the clogging problem.
While much effort has been put into the optimization of eDNA laboratory methods, less attention has been devoted to estimate the impacts of capture methods. To address this issue and better understand what influences species detection in aquatic systems, water samples were collected from turbid environments using three eDNA capture methods (precipitation, disc filters, and capsules) and their efficiency was compared in terms of volume filtered, eDNA recovered, and species detection. To cover the usual range of applications in eDNA monitoring, the study considered both the targeted detection of a ubiquitous species (the fire salamander, Salamandra salamandra Linnaeus, 1758) and the characterization of the overall amphibian community composition, using two species detection methods (qPCR and HTS) and two PCR replication thresholds (stringent and relaxed).

| Target species and pond selection
The study was conducted at the Ornithological Reserve of Mindelo and two nearby localities at Porto, Portugal, where the target species has previously been studied (e.g., Alarcón-Ríos, Nicieza, Kaliontzopoulou, Buckley, & Velo-Antón, 2020). The species selected for targeted detection, the fire salamander (Salamandra salamandra), is a urodele species widespread across Europe. The populations occurring in our study area are larviparous. Pregnant female salamanders deliver up to 90 larvae into water bodies (i.e., ponds, puddles, and streams) during the reproductive periods (Autumn and Spring), where the larvae stay until they complete metamorphosis (Velo-Antón, Santos, Sanmartín-Villar, Cordero-Rivera, & Buckley, 2015). This species was chosen due to its abundance in the study area and the relative ease with which it is detected using traditional pond sampling methods. The study system includes small ponds and temporary puddles of similar dimensions, providing suitable biological replicates, as the salamander population densities at the time of sampling were relatively homogeneous across sampling points (Table 1). Diurnal surveys were conducted in a range of ponds and puddles throughout the Porto region in late March 2018, and a total of nine ponds/puddles were sampled (Table 1).
At each sampling site, the physical characteristics of ponds (length, width, and depth) were measured. Larvae were detected using visual surveys, and larvae abundance was recorded using a transect sampling approach (Heyer, Donnelly, McDiarmid, Hayek, & Foster, 1994) (i.e., number of individuals per meter; Table 1). Several other amphibian species were detected and recorded during fieldwork, but abundance was only measured for fire salamanders, as this was the focal species for the comparison of capture and detection methods. Turbidity was measured using a Secchi disc housed in a TA B L E 1 Summary of sampling sites, including the salamander abundance observed in the field and the dimensions of the pond turbidity tube (Anderson & Davic, 2004;Myre & Shaw, 2006). The level on the turbidity tube at which the Secchi disc was no longer visible was recorded.

| Water sampling
Water collection was performed over a 10-day period using three capture methods: precipitation, disc filters, and capsules. For each site, all sampling was completed within a single sampling event (1-2 h). Precipitation samples were taken by collecting 15 ml surface water in a sterile 50-ml falcon tube. Immediately after collection, 1.5 ml of sodium acetate 3 M and 33.5 ml of absolute ethanol were added to the 15 ml water aliquots (Ficetola et al., 2008 (Hinlo et al., 2017). Equipment was sterilized between ponds with a 10% dilution of household bleach for at least 30 min and later rinsed with distilled water to remove any bleach residues. Two negative controls were collected at each pond. For the first, 15 ml of distilled water brought from the laboratory was added to a 50-ml falcon tube, along with 1.5 ml of sodium acetate 3 M and 33.5 ml of absolute ethanol. For the second, to ensure that all tubing and other reusable filtering apparatus was clean, 100 ml of distilled water was pumped through a filtering unit with a disc filter.

| DNA extraction and quantification
All DNA extractions were performed in a low-copy DNA laboratory (at CIBIO, Portugal) equipped with UV radiation. Strict protocols were followed to prevent contamination, including disposable laboratory wear, UV sterilization of all equipment before entering the laboratory and having workbenches and all the equipment needed for extraction cleaned with a 60% dilution of household bleach between extraction batches. Handling and cutting of the filters was performed on disposable aluminum sheets, changed between each filter, using forceps and scissors, which were cleaned with ethanol and flame-sterilized between samples. Additionally, a negative control was included in each batch of extractions (n = 6 batches), containing an average of 12 samples per batch.
Capsules were filled with 100 ml of resuspension buffer (50 mM Tris, 10 mM EDTA), both ends were covered with parafilm, and they were agitated manually for five minutes (e.g., Civade et al., 2016;Lopes et al., 2017). To concentrate the material in the buffer to a volume suitable for downstream extraction, the buffer was then For capsules, the extraction controls consisted of adding resuspension buffer to a clean unused capsule. Extraction controls for disc filters consisted of a sterile 2 ml Eppendorf with only n-lauroylsarcosine based buffer and no filter, while for precipitation samples, controls consisted of 50-ml falcon tubes with distilled water.
Double-stranded DNA was quantified by fluorometry (Quant-iT™ PicoGreen ® dsDNA Assay Kit, Molecular Probes), following the manufacturer's instructions. Readings were performed three times, and an average was obtained for each sample. Appendix S1). One primer set was found to be more specific and thus was chosen for further optimization: primer forward (Peixoto_ To evaluate assay sensitivity and generate a standard curve, a PCR product produced by the Peixoto-Sal-2019 primer pair was gel-extracted and cleaned with the QIAquick ® gel extraction kit (Qiagen), following the manufacturer's instructions. The purified DNA was quantified on Qubit™ following the manufacturer's instructions, the number of copies was calculated using the software DNA CALCULATOR (Sint, Raso, & Traugott, 2012), and six 10:1 serial dilutions ranging from 3.00E + 06 to 3.00E + 01 copies/µl were generated.

| qPCR
Standards were performed in triplicate and eDNA samples and negative controls in duplicate. A total of three plates were run, giving a total of six negative qPCR controls. Additionally, strict protocols were followed to prevent contamination, such as the use of separate rooms for pre-and post-PCR work, as well as the cleaning of workbenches and pipettes with bleach and ethanol before and after each use.
qPCR reactions were considered positive if a sample's fluorescence intersected the threshold line, and negative otherwise.
Exceptions to this were (a) as we did not expect overlaps above 100 bp in paired-end reads, we specified --max-overlap = 100 using flash2 and (b) to filter low-quality reads we applied a maximum expected error of 1 using the vsearch --fastq_filter command.
The exact sequence variants (ESVs) were mapped against a 12S amphibian database containing all species potentially occurring in the study area (Table S2), using the MEGABLAST algorithm (Zhang et al., 2000), and 100 results per query were kept. Hits with <70%

| Statistical data analysis
For both qPCR and HTS data, two thresholds were used for salamander detection, when S. salamandra was detected in at least one (relaxed) or both (stringent) PCR replicates of the respective sample.
Statistical analyses were performed with R (R Development Core Team, 2008) using linear modeling approaches implemented in the lme4 (v. 1.1-17) package (Bates, Mächler, Bolker, & Walker, 2015) and further assessed using the car (v. 3.0-0) package ANOVA function (Fox & Weisberg, 2011) and emmeans (v. 1.2.3) package emmeans function (Lenth, 2018). Linear mixed-effects models were used to assess the effects of capture method on the volume of water filtered and the mass of eDNA captured, as well as the effects of turbidity and salamander abundance on qPCR copies per liter of water filtered. Generalized linear mixed-effects models were used to assess the effects of capture methods on the detec- were transformed to common logarithm to meet assumptions. To ensure that variation in salamander abundance, pond area or pond depth was not confounding results, they were initially included as predictors in all models that had salamander detection and amphibian detection (only pond area and depth) as a response variable. As these factors did not contribute significantly to any models, they were subsequently excluded from final models. To compare the contributions of capture and laboratory methods to variation in S. salamandra detection, PCR replicate was nested within capture method and the mean sum of squares was used to estimate the contribution of each component (Mata et al., 2019). Although other amphibian species were incidentally detected by the HTS approach, their abundances were not measured as they were not the target study species. However, as all three capture methods were conducted at each sampling point, it was possible to assess the effect of capture method on the amphibian community composition. Moreover, overall pond effects associated with variation in nontarget species abundances were controlled by specifying sites as random factors in the final mixed models. To assess the effect of capture method on the amphibian community composition, presence/absence data (binary) were used to construct a distance matrix of Jaccard dissimilarities between eDNA samples using the vegdist function from the vegan (v. 2.5-6) package (Oksanen et al., 2019). The matrix was used as the response variable for a permutational multivariate analysis of variance using distance matrices (PERMANOVA) model with capture method as the main factor, implemented using vegan´s adonis function (10,000 permutations, with Site as strata to account for nonindependence between samples from the same Site).

| RE SULTS
A total of 54 water samples, 18 field negatives, six extraction negatives, and three PCR negatives were processed. From all controls, one field negative, from site 3, contained reads assigned to D. galganoi (n = 1,137), and so detections of this taxon were removed from all samples from this site. Approximately, 2.2 million reads were obtained for the sample set after demultiplexing, of which c. 355,000 were assigned to amphibian species and passed all bioinformatic filters. Information on volume of water filtered and DNA amount is provided in the Supplementary Material (Table S1), as well as the final taxa table with the number of reads assigned to each species (Table S3) and a summary of the read counts at each bioinformatic step ( Figure S2).

| Water sampling and eDNA capture
The volume of water filtered was significantly different (p < .0001) between capsules (x = 7.89 L, SE = 6.79) and disc filters (x = 1.10 L, SE = 1.03; Figure 1). The volume processed for precipitation was always 0.015 L. Turbidity was negatively correlated with the volume of water filtered for disc filters (R 2 = .64; p < .0001) but not for capsules There was a significant difference (p < .0001; Figure
Information on Cq-values and copies/µl for each sample is provided in the Supplementary Material ( The HTS approach generally resulted in more salamander detections than the qPCR assay, but not significantly so (Figure 3).
Precipitation consistently provided the lowest numbers of S. salamandra detections, while capsules and disc filters showed similar results ( Figure 3). PCR replicability was similar using either the qPCR or the HTS methods, as indicated by the similar values for the mean sum of squares for the nested fixed effect of Method:PCR_replicate (Table 2). The variation in detection success associated with the choice of capture method was over 10 times higher than that associated with PCR replication, regardless of the detection method used (Table 2). As expected, applying the more stringent PCR replication F I G U R E 1 Volume of water filtered using either capsules or disc filters. The three hinges starting from below correspond to the 25th percentile, median, and 75th percentile, respectively, extending to the smallest and highest value recorded. Significance values are represented by stars: *p < .05, **p < .01, ***p < .001 and ****p < .0001 although not always to a statistically significant degree (Figure 3).
For capsules, the number of reads assigned to S. salamandra was positively correlated with the volume filtered: for each increase of 1 L filtered, reads increases 358 (R 2 = .47, p < .01); while for disc filters no such trend was evident. Additionally, we did not observe a significant relationship between turbidity and salamander abundance with the total qPCR copies obtained per liter of water filtered.

| Amphibian community composition
Overall, there was a total of 49 detections (the sum of the number of species detected in all field samples) using the stringent replication threshold and a total of 104 detections using the relaxed replication threshold. Using the relaxed replication threshold, disc filters resulted in significantly higher amphibian detections than the other two capture methods (p < .05; Figure 4). However, there was no significant difference between capture methods when using the stringent replication threshold, although disc filters still resulted in highest detection levels ( Figure 4; Table 3). Using the relaxed replication threshold, the total number of amphibian detections was significantly higher than for the stringent replication threshold for both disc filters (p < .01) and precipitation (p < .01) (Figure 4), but not for capsules. There were no significant differences in amphibian community composition estimated by the three capture methods (MS = 0.35, F.Model = 1.37, R 2 = .04, p = .33; Figure 5). This was true regardless of modeling the results of single or combined PCR replicates (data only shown for combined).

| D ISCUSS I ON
In the present study, filtration techniques outperformed precipitation, generating a higher number of detections of S. salamandra and captured eDNA, while species detection was identical between disc filters and capsules. However, amphibian community characterization (i.e., species richness and composition) was not significantly affected by the choice of capture method. Overall, S. salamandra detection was similar with both qPCR and HTS. Relaxed PCR replication threshold consistently generated higher detection levels than the stringent replication threshold, although differences were not always statistically significant. It is unlikely that these key results were affected by methodological biases or artifacts, namely eventual problems associated with temporal or spatial variations in sampling conditions. For instance, while seasonal variation in eDNA F I G U R E 3 Number of samples in which S. salamandra was detected by each capture method and with each detection method and PCR replication threshold (n = 18 in all cases). Significance values are represented by stars: *p < .05, **p < .01, ***p < .001 and ****p < .0001 TA B L E 2 Variation estimated from a linear mixed-effects model in S. salamandra detection success associated with capture method and PCR replicate using either the qPCR or HTS approaches. PCR replicate was nested within capture method to have affected our results because our study was performed over a 10-day sampling period. Also, it is unlikely that spatial variations affected the results given that the three capture methods were employed at each sampling point on the same sampling occasion, and differences among ponds in environmental conditions and eDNA concentrations were controlled statistically through our mixedmodel approach.

| Filtration and precipitation capture methods
Filtration and precipitation are currently the two main approaches to capture eDNA in aquatic ecosystems (Herder et al., 2014;Hinlo et al., 2017;Li, Handley, Read, & Hänfling, 2018). Filtration is more common with disc filters, while capsules have only recently been applied in eDNA studies (e.g., Civade et al., 2016;Lopes et al., 2017).
In the present study, the choice of capture method influenced eDNA recovery and species detection, with filtration methods capturing more eDNA and detecting the target species in a higher number of samples than precipitation. Previous studies in aquatic environments have reported similar results, where precipitation resulted in lower detection rates than filtration (Eichmiller, Miller, & Sorensen, 2016;Hinlo et al., 2017;Piggott, 2016;Spens et al., 2017). The higher amounts of eDNA captured and species detection observed for filtration methods were likely associated with their higher sample volumes (Raemy & Ursenbacher, 2018).
Previous research has demonstrated that filter attributes such as pore size and membrane material can influence eDNA recoveries and detection rates (e.g., Deiner et al., 2018;Djurhuus et al., 2017;Jeunen et al., 2019). The present study did not include a comparison of pore sizes or membrane types, but as both disc filters and capsules had a polyethersulfone hydrophilic membrane and a pore size  (Vences et al., 2016), 45 L  and even 100 L . As more water is sampled, the chances of eDNA fragments being captured increases, thus explaining the higher amounts of eDNA recovered for capsules.
Nevertheless, S. salamandra detection and amphibian community composition was similar between both filtration methods and did not reflect the higher performance of capsules regarding volume and eDNA recovered. Even though disc filters provided a higher number of amphibian detection events, amphibian community composition was identical for both filtration methods. The most likely explanation is that, given the relatively small pond sizes utilized in this study, amphibian eDNA was sufficiently abundant to be detected despite filtering lower volumes with disc filters. Further studies comparing disc filters and capsules across other taxonomic groups and study systems would provide a better understanding of the efficiency of each filtration method.
Filtering higher water volumes may not always be advantageous as it might increase the concentration of inhibitors in the sample (Herder et al., 2014), usually abundant in turbid waters, constraining downstream laboratory procedures. However, we did not observe any obvious PCR inhibition in the capsule-derived eDNA samples.
The percentage of PCRs that successfully amplified was identical between capsules and precipitation samples, and both were higher than for disc filters. This indicated that inhibition was not occurring to a greater degree in capsule-derived PCRs, compared to the other capture methods. Also, we did not normalize DNA concentrations prior to the first PCRs. It might be expected that, by capturing more eDNA, the concentration of nonamphibian eDNA increases in the extracted DNA elution, which could lead to a higher proportion of nonamphibian eDNA being amplified. However, the percentage of nonamphibian reads was lower for capsules (median = 36%) than for disc filters (median = 50%), so this does not help to explain the results. Overall, capsules may be more appropriate for running waters or larger water bodies, where eDNA is more diluted (Herder et al., 2014) and filtering larger volumes can increase species detection (Lopes et al., 2017), whereas disc filters might be more suitable for smaller stagnant water bodies, where eDNA is less diluted (Herder et al., 2014). Disc filters may also prove to be more cost-effective in many situations, particularly when funds are limited, as the capsules used in our study cost c. €25 at the time of writing, while disc filters cost less than €1. One benefit of using capsules over disc filters is that they require less handling in the field, which may decrease the risk of contamination. A further alternative not tested in this study is the use of Sterivex capsules (e.g., Raemy & Ursenbacher, 2018;Sigsgaard et al., 2017). Even though their surface area is smaller than the ones compared here, they have shown to outperform disc filters (Spens et al., 2017) and are usually cheaper than high-capacity capsules.
Previous protocols using capsules have also used 5 min agitation, but followed by a centrifugation at 15,000 g of 50-ml tubes (e.g., Civade et al., 2016;Lopes et al., 2017). As not all laboratories have centrifuge equipment for these larger volumes at high speeds, we concentrated the material captured in each capsule by filtering it through a standard 47 mm disc filter. While some eDNA may have F I G U R E 5 Principal coordinate analysis plot of amphibian community dissimilarity of eDNA samples in the present study, using Jaccard distances, based on presence/absence data. Ellipses are drawn with a confidence level of 0.9.
been lost at this step, far more was still captured using this method than by filtering in the field directly through a standard disc filter.

| qPCR versus. HTS
Species detection with eDNA methods can be accomplished with either a single species or a multi-species approach. Single-species detection is generally used for endangered (e.g., Piggott, 2016) or invasive species (e.g., Hunter et al., 2015),  (Kelly, Port, Yamahara, & Crowder, 2014;Port et al., 2016;Rodgers et al., 2017). Previous studies have also successfully applied qPCR methods for detecting S. salamandra (Preißler, Watzal, Vences, & Steinfartz, 2018). While HTS is often more advantageous and cost-efficient to detect multiple species (Thomsen & Willerslev, 2015), single-species detection with qPCR is generally cheaper (Harper et al., 2018) and less timeconsuming. Additionally, HTS approaches add a level of complexity to data analyses due to the bioinformatic filtering steps required to remove sequence reads that might originate from sequencing errors or contamination (Thomsen & Willerslev, 2015).

| PCR replication threshold
An additional source of variation in eDNA research lies on the thresholds applied to the data. The use of strict filtering to reduce false positives likely reduces detection rates and thus inflate false negatives. The opposite is true for relaxed thresholds, reducing false negatives but generally at the expenses of increasing false positives and overestimating the presence of a species. The results of the present study indicate that there is much higher variability introduced from the choice of capture method than from PCR replication threshold. There were, however, differences in detection events between the thresholds used, particularly for disc filters, which suggests a higher stochasticity in PCR replicates for this capture method ( Figure 4). PCRs from capsules were more replicable, displaying overall less disparity between stringent (species detected in both replicates) and relaxed (species detected in either replicate) PCR replication thresholds for HTS, most likely due to having a greater starting mass of eDNA. Similar to our results, a recent study using water samples and both qPCR and HTS approaches for the detection of the great crested newt has demonstrated that stringent thresholds reduce detection levels of a target species (Harper et al., 2018). To enhance the reliability of a study, a balance between false positives and negatives is required, highlighting the importance of careful consideration of the most suitable threshold to apply when inferring species presence-absence, especially for endangered or protected species. According to Ficetola et al. (2015), site occupancy models can be a useful tool to estimate error rates regarding species detection and determine taxon-specific thresholds by adjusting the minimum number of replicates required to consider the presence of a species, therefore increasing confidence in the results.

| CON CLUS IONS
To the best of our knowledge, this is the first study to compare highcapacity capsules with common eDNA methods, such as precipitation and filtration with standard disc filters, highlighting the importance of choosing a suitable capture method for eDNA studies. The results indicate that the use of either disc filters or high-capacity capsules outperforms precipitation, but no major differences were found between filtration methods. Based on species detections, we cannot recommend the use of capsules over standard filters. However, as capsules filter more water and capture more eDNA, their application may be beneficial in other field situations, such as detecting low abundance species in larger or fast-flowing water bodies. The results suggest that, if eDNA assays are well-designed, the choice of capture method outweighs the choice of laboratory detection method used. However, PCR replication thresholds applied also affect the reliability of results. Identifying the best capture method is essential for accurate biodiversity surveys using eDNA techniques, and further research with larger sample sizes and a multi-taxon approach would provide a better understanding of the efficiency of each capture method, particularly relevant for capsules and disc filters.

CO N FLI C T O F I NTE R E S T
None declared.

AUTH O R CO NTR I B UTI O N S
S.P., G.V.A., B.E. and P.B. designed the study, S.P. collected all samples with help from G.V.A. and B.E. and S.P. performed all laboratory work with help from C.C. to prepare the sequencing run. Data analysis were conducted by S.P., with help from B.E., and the manuscript was written by S.P. with inputs from all authors.

DATA AVA I L A B I L I T Y S TAT E M E N T
All demultiplexed fastq files generated by the Illumina Miseq sequencer, and all final fasta sequences for the sanger-sequenced amphibian tissue samples, are available on ENA, Project Accession PRJEB35424, along with all relevant sample metadata. The final taxa table used for statistical analysis is available on BioStudies (Accession No. S-BSST308) as is a summary of the number of reads at each bioinformatic step (see also Figure S2).