4. Discussion 
    We explored the impact of the number of PCR replicates and read sampling depth, two common parameters in eDNA experimental design, on estimates of taxonomic diversity. Using eDNA extracts from six sites at three ecologically and geographically distinct locations, we performed 24 PCR replicates for each of metabarcodes: plant ITS (PITS) and fungal ITS (FITS). We then analyzed these replicates by compiling data sets that included different read sampling depths and minimum read cutoffs. We find that PCR replicates are consistent in the composition (Figure 7) and relative abundance (RA) (Figures 3 and S1) of high abundance taxa, but inconsistent in recovery of low abundance taxa, and that even large numbers of PCR replicates are insufficient to fully characterize diversity at any site.
When considering only high abundance taxa, our PCR replicates produced community profiles that distinguished sites from each other, even sites that are geographically proximate and presumably similar in community composition (Figure 4). The majority of high abundance taxa were detected in all 24 PCR replicates, excluding outlier PCRs. This result provides empirical support for the modeling-based prediction by Ficetola et al (2015) that PCR replicates will consistently detect taxa that have high “detection probability”, which they define as taxa present in high abundance relative to other taxa at a site. We also observed that the number of reads assigned to a taxon within a PCR was positively correlated to the frequency with which that taxon was observed across PCR replicates (Figures 3, 7), which was also reported by Smith and Peay (2014). Together, these results confirm that community profiles based on high abundance taxa are replicable among PCRs and capable of distinguishing sites. Minimal PCR replication is therefore necessary to characterize sites using β diversity statistics that derive from high abundance taxa.
While high abundance taxa were recovered consistently among our PCR replicates, low abundance taxa were not (Figures 1, 2, 6). Low abundance taxa rarely occurred in PCR replicates; we observed some low abundance taxa in several PCR replicates but most in only a single replicate. Unsurprisingly, this stochasticity in recovery of low abundance affected biodiversity statistics that rely on raw taxon counts, such as α diversity. While this observation has been reported previously (e.g. Beentjes et al., 2019; Ficetola et al., 2008) our results highlight how the problem can be exacerbated by shallow sequencing read depths and low minimum read cutoffs. Specifically, we find several fold differences in maximum richness and rarefied richness among replicates depending on what values we selected for read depth and minimum read cutoff. While the stochasticity in recovery of low abundance taxa poses challenges in interpretation of some biodiversity statistics, it tends not to influence β diversity between sites measured as either presence-absence (Figure 4) or RA (Figure S1), or on position with a PCoA.
At our six sites, 24 PCR replicates were not sufficient to detect all rare taxa and therefore stabilize the species accumulation curves (Figure 5). This result supports previous observations that using different numbers of PCR replicates will alter taxonomic profiles (Alberdi et al. 2017; Murray, Coghlan, and Bunce 2015). Intriguingly, Smith and Peay (2014) reported the opposite conclusion: that increasing the number of PCR replicates does not influence α diversity. As is common in eDNA studies, Smith and Peay amplify each of their PCRs over 30 cycles, whereas we estimated the optimal number of cycles for each reaction separately using qPCR, following Murray, Coghlan, and Bunce (2015). Overamplification of PCR amplicon pools can reduce the complexity of the amplicon pool as read “species” that replicate more efficiently outcompete others that replicate less efficiently (Nichols et al., 2017). Consequently, taxa that are least efficiently amplified will become increasingly rare and may not be observed, in particular at low read sampling depths.
We observed most singleton taxa in only one PCR replicate (Figure 6). This finding supports the conclusion by Leray and Knowlton (2017) that random sampling of rare taxa across PCR replicates accounts for most of the variation between PCR replicates. Increasing read sampling depth did not reduce the number of replicates that were required to stabilize the taxon accumulation curve (Table 2). However, increasing the minimum read cutoff did reduce the number of PCR replicates necessary to stabilize the curve (Table 2), presumably by removing many of the low abundance taxa from each data set such that only the high abundance taxa, most of which were common to each PCR, remained.
We found that increasing the read sampling depth significantly increased the number of taxa detected at each of our sites (Figures 1, 2; Table S4). As many eDNA studies and consortia sequence amplicon pools to the shallowest of our depths (1,000 reads), this result has implications for how biodiversity estimates based on these published data sets can be interpreted and compared. The impact of this parameter choice depends on how the data are analyzed. For example, we estimated significantly higher observed α diversity at a depth of 10,000 reads than at a depth of 1,000 reads across all sites, but found no difference between read depths when α diversity was calculated using the Shannon or Simpson metrics, which underweight low abundance taxa compared to common taxa (Hsieh et al., 2016). The significant increase in α diversity that we observed is in contrast to Murray, Coghlan, and Bunce (2015), who found that sampling depth per PCR replicate did not necessarily increase detection of low abundance taxa. This difference may be due to the use by Murray, Coghlan, and Bunce of Ion Torrent rather than Illumina sequencing technology, as the higher error profiles generated by the Ion Torrent platform require more stringent removal of rare taxa (Salipante et al., 2014). Because sites will vary in the amount of total diversity present, taxon accumulation curves such as those in Figure 1 may be useful in determining the appropriate read sampling depth for a given site.
Our results also reiterate the need to consider the physical and ecological setting during eDNA experimental design (Anderson et al., 2012; Ficetola et al. 2015). We observed the most variation in observed α diversity among PCR replicates in the PITS dataset at YL.1 (Figure 1e and 2), a lagoon basin into which water and wind carries and deposits DNA-containing materials from the surrounding environment. The constant influx of DNA from the surrounding habitats may explain why amplicon pools from this site include many low abundance taxa. Although these low abundance taxa have little effect on β diversity estimates, they are contributing members of local communities. Metabarcoding may therefore be particularly inefficient tool for estimating and comparing α diversity at sites with high biological turnover or input.
Because we sequenced each PCR replicate individually, we were also able to explore the rate of occurrence and potential impact of PCR outliers, which we define as PCR amplicon pools that differ significantly in either composition or relative abundance of taxa compared to other replicates from the same eDNA extract. We found PCR outliers to be more common at sites with high diversity, like YL.1. Increasing read sampling depth also increased the frequency of PCR outliers, but only for the FITS data sets (Figure 4), possibly because of the higher taxonomic diversity among low abundance taxa recovered by this metabarcode. While we are unable to determine the precise cause of outlier PCRs, we note that they are only observable as outliers if more than one PCR replicate is performed. This rationale is often used by groups that perform three PCR replicates per sample (Taberlet et al., 2018), which allows disambiguation between an outlier and a “normal” PCR without additional laboratory work.
Given our results, we present the following conclusions, which can serve as recommendations for experimental design in eDNA metabarcoding experiments:
  1. PCR Replication: A single PCR often captures the diversity of common taxa at a site and allows sites to be differentiated based on these common taxa. However, because outlier PCRs are a possibility, a minimum of two PCR replicates is recommended. When multiple PCR replicates are performed, the LCBD statistic can be used to identify PCR outliers by quantifying replicate uniqueness.
  2. Read sampling depth: Increasing sequencing read depth increases the chance that low abundance taxa are recovered from within the amplicon pool. However, because PCR replicates vary in taxonomic composition, exhausting the sequence complexity of an amplicon pool through deep sequencing is not the same as exhausting the sequence complexity of a DNA extract. Variation between PCR replicates in taxonomic composition or relative abundance does not diminish with increased sequencing read depth.
  3. Minimum read cutoff: Higher minimum read cutoffs remove low abundance taxa from a PCR amplicon pool. Removed taxa will include both low abundance contaminants and low abundance authentic taxa. As such, the minimum read threshold may influence α diversity but is less likely to influence β diversity.