Designing environmental DNA surveys in complex aquatic systems: Backpack sampling for rare amphibians in Sierra Nevada meadows

Funding information USDA National Institute of Food and Agriculture, Grant/Award Number: 1005908; USDA Forest Service, Pacific Southwest Research Station Abstract 1. Surveys for environmental DNA (eDNA) can provide an efficient and effective means of detecting aquatic organisms in various types of aquatic systems. 2. In the summer of 2017, the efficacy of a new, integrated eDNA backpack sampler to detect two native amphibians (Rana sierrae and R. cascadae) at risk was tested in complex mountain meadows in California. Samples were collected at 65 locations in 15 meadows where the target species were known to be present or were historically present. 3. Collection and preservation of individual samples took less than 10 min on average. Environmental DNA analysis methods detected each species at all meadows with visual detections (N = 11) except one with one frog seen away from sampling sites. Bayesian multi-scale occupancy modelling indicated that conditional detection probabilities at the sample level ranged from 0.30 (CL 0.07–0.65) at meadow heads where no frogs were observed during visual surveys to 0.93 (CL 0.77–1.00) at the meadow foot with at least one frog observed in the vicinity. 4. Compared with visual surveys, eDNA methods more frequently detected amphibians at the sampling-location scale. The improvement in detection using eDNA methods was most pronounced for samples collected at the downstream ends of meadows where water converges, where eDNA methods detected target species at 10 of 11 occupied meadows. 5. These results suggest that the addition of eDNA sampling to visual surveys in mountain meadows will improve survey accuracy and increase the probability of detecting rare frogs.


| INTRODUCTION
Accurate knowledge of where imperilled or invasive organisms occur allows the strategic allocation of limited management resources, yet determining the occupancy and distribution of these species can be difficult, especially in complex habitats with few individuals (Tyre et al., 2003;Mackenzie & Royle, 2005;Chades et al., 2008). Researchers have developed a molecular tool to address this problem in aquatic environments by surveying for the presence of trace genetic evidence (such as shed skin, faeces, urine and mucus) of species of interest in the water (Belle, Stoeckle, & Geist, 2019;Dejean et al., 2012;Ficetola, Miaud, Pompanon, & Taberlet, 2008;Goldberg, Pilliod, Arkle, & Waits, 2011;Rees, Maddison, Middleditch, Patmore, & Gough, 2014). Instead of timeconsuming survey techniques that often involve disturbing animals through netting or electroshocking, surveyors obtain water samples at strategic locations within the aquatic habitat, reducing field time, costs and risk to target organisms (Thomsen & Willerslev, 2015).
Hesitancy to apply eDNA sampling stems in part from uncertainties about how its effectiveness varies with environmental conditions and across taxa (Barnes et al., 2014;Goldberg, Strickler, & Fremier, 2018;Strickler, Fremier, & Goldberg, 2015). For example, DNA breaks down more quickly in warm or acidic water compared with cool or neutral water , and does not travel far from its source in still water conditions (Dunker et al., 2016;Goldberg et al., 2018). In addition, the production of eDNA varies among individuals and within individuals over time (Stewart, 2019). However, a better understanding of the drivers of uncertainty in eDNA detectability allows adjustments in sampling and analysis to account for this variability. For example, Goldberg et al. (2018) investigated eDNA detection for a suite of amphibian species across a gradient of environmental conditions to determine the different physiological, ecological and hydrological processes limiting detection. The findings were then used to adjust sampling protocols for the different species and environments in a way that increased detection probability to 0.95.
Advances in sampling methods should also improve the effectiveness of eDNA sampling as a monitoring tool. Recently, a backpack sampler was developed with a negative pressure inline filtration system to collect eDNA samples efficiently (Thomas, Howard, Nguyen, Seimon, & Goldberg, 2018). Using this system, water samples are pumped through a filter mounted at the end of a pole extension. Flow rate, sample volume and filtration pressure can be programmed and recorded for each sample collected (Thomas et al., 2018). In addition, Thomas, Nguyen, and Goldberg (2019) developed self-desiccating filter packs that may further increase collection efficiency and decrease contamination risk for field samples because field crews would no longer have to remove, fold and place filters in ethanol-filled vials in the field. Although promising, the effectiveness of these new tools has yet to be demonstrated for many sampling applications.
Advances in eDNA sampling methodology could facilitate the management of complex montane wet-meadow areas that provide key habitat for native amphibians at risk. For example, amphibians have experienced severe declines in California's mountains over the last century owing to habitat alteration, disease and invasive species (Adams et al., 2017;Knapp & Matthews, 2000;Piovia-Scott et al., 2015;Pope, Brown, Hayes, Green, & Macfarlane, 2014;Rachowicz et al., 2006;Wake & Vredenburg, 2008). Declines in the Sierra Nevada yellow-legged frog (Rana sierrae) and Cascades frog (Rana cascadae) have been particularly pronounced. Rana sierrae is a federally listed endangered species that inhabits highelevation lakes, streams and meadows in the Sierra Nevada. Rana cascadae is under review for listing on California's endangered species list and occurs in similar habitats in the southern Cascade and Klamath ranges of California. Severe declines and extinctions have been documented for both species in lake habitats at higher elevations resulting from a deadly fungal pathogen and the pervasive introduction of non-native sport fishes (De Leon, Vredenburg, & Piovia-Scott, 2017;Fellers, Pope, Stead, Koo, & Welsh, 2008;Knapp & Matthews, 2000;Rachowicz et al., 2006;Vredenburg, Knapp, Tunstall, & Briggs, 2010). Mountain meadows may play an increasing role in supporting these native amphibians because they provide refuge from fish and possibly disease (Pope et al., 2014).
Owing to the complexity of meadows, native frogs can be difficult to detect using conventional survey techniques, especially when populations are small. If eDNA surveys can provide an efficient and effective alternative detection method, they have the potential to greatly facilitate the conservation of rare amphibians and management of their meadow habitats.
This article describes tests carried out in the summer of 2017 on whether eDNA monitoring with the backpack sampler could be used to detect R. sierrae and R. cascadae at 15 meadows in the Sierra Nevada and southern Cascade ranges. Meadows were selected with known or expected occupancy by either species, including occupied sites with a range of abundances, and visual surveys were conducted concurrently with eDNA sampling. In addition to evaluating occupancy at the whole-meadow scale, associations between the results from eDNA sampling and specific habitat categories within meadows were also evaluated. The study had two further objectives: a beta version of the Smith-Root eDNA Sampler Backpack (Thomas et al., 2018) was tested, and newly developed self-desiccating filter packs were evaluated that may increase collection efficiency and decrease con-   Table 1). Specific meadows were chosen based on determinations of frog presence or absence from previous visual surveys. Sites were included with high densities of animals, with few animals remaining, and where target species were historically present but presumably extirpated.

| eDNA sample collection and visual surveys
Two replicate eDNA samples were collected from each of two to six localities within each meadow (Table 1, Appendix A). These included the head and foot of most meadows to determine if these were effective sampling locations to determine presence and extent of occupancy (e.g. present in meadow, but not above meadow). Additional sampling localities were selected within meadows based on the known or suspected habitat use of the focal species ( Figure 2). Sampling localities were assigned to one of four categories: (i) head of the meadow; (ii) foot of the meadow; (iii) stream channel; or (iv) offchannel pool. In conjunction with each eDNA sample collected, a visual count of amphibians within the expected area of influence of the sample locality was also conducted. The area of influence was defined as the pool where the sample was collected or, for stream samples, the reach extending 30 m upstream of the sample locality.
When field crew members did not know the current or historical location of target species within a meadow, surveys were guided by local biologists with knowledge of frog distribution in that meadow. To minimize the risk of sample contamination, eDNA samples were collected by crew members who had not handled target species.
At each sampling locality, the Smith-Root eDNA Sampler Backpack (Thomas et al., 2018) was used with a split in the sampling line for replicate sampling. An extension pole was used to collect F I G U R E 1 Meadows sampled for eDNA of Cascades frogs (Rana cascadae) and Sierra Nevada yellow-legged frogs (R. sierrae) in the Sierra Nevada and southern Cascades ranges of California (a). Images include target amphibians, R. cascadae (b) and R. sierrae (c), and an example of channel habitat in Childs Meadow. Photo credits: S. Riffle (b and c) and A. Bearer (d) [Correction added on 11 September 2020, after first online publication: The photo credits on Figure 1 has been corrected in this version.] T A B L E 1 Meadows sampled for Sierra Nevada yellow-legged frogs (Rana sierrae) and Cascades frogs (Rana cascadae), including the number of sampling locations per meadow (n), whether or not the species was detected by eDNA or visual surveys, mean sample volume filtered (±SE) and mean time to collect a sample (±SE)  A field negative sample (distilled water) was collected at every site. In addition, recommended best practices for preventing crosscontamination between samples were followed (Goldberg et al., 2016). New gloves were used whenever filters were handled, filter housings and forceps were single-use, and all other field gear was decontaminated with bleach between meadows to prevent the transfer of DNA and pathogens.

| eDNA assay development
An existing assay was available for R. sierrae (Bedwell & Goldberg, 2020), but not for R. cascadae, so a range-wide assay was developed for R. cascadae. Sequence data were provided by K. Monsen (Monsen & Blouin, 2003) and an inclusive consensus sequence for the D-loop sequence was created using Sequencher ver-

| eDNA filter processing
Environmental DNA was extracted from the field-collected filters using the Qiashredder/DNeasy method (Goldberg et al., 2011) in a laboratory dedicated to low-quantity DNA samples, using best practices for preventing and detecting contamination (Goldberg et al., 2016). A negative extraction control was included with each set of extractions and an additional negative qPCR control was run with each plate of field samples. Reactions were as described above. A well was considered as testing positive if exponential growth was produced. Each sample was analysed in triplicate and any sample that Samples from three of the 65 meadow localities were found to be strongly inhibited on both filters and were excluded from further analyses.

| Statistical analyses
To evaluate the effectiveness and consistency of the eDNA sampling method, eDNA occupancy and detection probabilities were estimated for R. sierrae and R. cascadae at the meadow, sampling locality and replicate filter scales with Bayesian, multi-scale occupancy models using the R package eDNAoccupancy v0.2.0 (Dorazio & Erickson, 2018). These models accommodate three levels of sampling so it was possible to model the probability of species occurrence among meadows (Ψ), the conditional probability of species occurrence at a sampling locality within a meadow given that the species is present in the meadow (θ) and the conditional probability of species detection on replicate filters collected at a sampling locality given that the species is present at that sampling locality ( Because the occupancy models could only be fitted with a limited number of covariates and only used presence/absence data as a response variable, linear mixed models (LMMs) were also fitted using the R package glmmTMB (Brooks et al., 2017 Burnham & Anderson, 2002). Differences between levels of categorical covariates in all LMMs were evaluated with estimated marginal means using the R package emmeans (Lenth, 2018).
Based on results from both modelling exercises, additional posthoc analyses were run to explore further the influence of certain predictors. Specifically, an assessment was made of whether the volume of water filtered differed among habitat categories within meadows using an LMM and estimated marginal means.

| Assay validation
The R. cascadae assay was successfully validated in silico, with no matches in the search. However, many species do not have sequences of the D-loop region in the database. The assay was also successfully validated against tissue samples from all co-occurring anurans.

| eDNA sample collection
Sample collection usually took <10 min (mean for R. cascadae samples = 6.1 min, mean for R. sierrae samples = 8.7 min). Filtered water volumes ranged from 0.2 to 2.2 L per filter and averaged 1.4 L. The two-sample preservation methods yielded no differences in the quantity of eDNA (mean copy number for ethanol-preserved, 1.5 per ml; for self-desiccating filters, 1.6 per ml; t = 0.7, d.f. = 11, P = 0.5).
3.3 | Detection of R. sierrae and R. cascadae using eDNA sampling Rana sierrae was detected from at least one eDNA sample at six of eight meadows sampled for the species, and occurrence was sample at five of seven meadows, and occurrence was also confirmed using visual surveys at the same five meadows (Table 1).   (Fremier, Strickler, Parzych, Powers, & Goldberg, 2019;Wilcox et al., 2016). Regardless, even with some loss of material, sampling at the foot of the meadow provides a more comprehensive sample than habitats further upstream in the meadow. It is also important to sample off-channel pools using eDNA methods because these habitats can be difficult to survey visually owing to the silty substrates where frogs and tadpoles can easily hide. In addition, sampling in pools is important for a thorough survey because transport of DNA out of the pools to downstream channel habitats is unlikely.

Results
Our occupancy model found an inverse relationship between the probability of detecting eDNA and the volume of water sampled, contrary to the findings of others Schultz & Lance, 2015). We suspect that this pattern resulted primarily from the fact that sample volume was confounded with sampling habitat type, as suggested by our LMM analysis of eDNA quantity. Samples collected at the head of the meadow tended to filter the maximum water volume (2 L per filter; Figure 5), but also tended to test positive for eDNA less often than samples collected at the foot of the meadow, in channels, or in off-channel pools (Figure 4). Samples collected at the foot of meadows and in off-channel pools often clogged early owing to fine silt in the water column, but these samples more frequently collected eDNA ( Figure 4) and tended to contain larger quantities ( Figure 6). Clearly, when sampling for rare species, it is more important to select the habitats where the eDNA of the species is likely to be found than sites where sample volume is maximized.

APP E NDIX B : MODEL SELECTION TABLE
Candidate model set ordered by AICc for linear mixed models testing the relationship between ln qPCR copy number and several relevant predictors including habitat category (HabCat), species (Sp), nearby visual counts (VC), volume filtered (Vol), and the interaction between habitat category and visual counts. R 2 m indicates the marginal R-squared value (proportion of the variance explained by the fixed effects within the model) for each model