0 citations0 references

DNA metabarcoding—Need for robust experimental designs to draw sound ecological conclusions

Molecular Ecology2019Vol. 28(8), pp. 1857–1862

Citations Over TimeTop 1% of 2019 papers

Abstract

DNA metabarcoding, especially when coupled with high-throughput DNA sequencing, is currently revolutionizing our capacity to assess biodiversity across a full range of taxa and habitats, from soil microbes (e.g., Thompson et al., 2017) to large marine fish (e.g., Thomsen et al., 2016), and from contemporary to tens of thousands year-old biological communities (e.g., Willerslev et al., 2003). The breadth of potential applications is immense and spans surveys on the diversity or diet of species native to specific ecosystems to bioindication (Pawlowski et al., 2018). The approach is also especially cost-effective and easy to implement, which makes DNA metabarcoding one of the tools of choice of the 21st century for fundamental research and the future of large-scale biodiversity monitoring programs (reviewed in Bohan et al., 2017; Creer et al., 2016; Taberlet, Bonin, Zinger, & Coissac, 2018; Thomsen & Willerslev, 2015). However, as is often the case with any emerging technology, we feel that the rise of DNA metabarcoding is occurring at a pace and in a manner that often loses sight of the challenges in producing high-quality and reproducible data (Baker, 2016). DNA metabarcoding is by essence a multidisciplinary approach building upon many complementary expertises, including field and theoretical knowledge, taxonomic expertise, molecular biology, bioinformatics, and computational statistics. Combining all these within single studies is necessary, not so much for producing and analyzing the data per se, but rather for minimizing and controlling the possible biases that can be introduced at any step of the experimental workflow—i.e., from the sampling to data analysis—and that can lead to spurious ecological conclusions (reviewed in Bálint et al., 2016; Nilsson et al., 2019; Dickie et al., 2018; Taberlet et al., 2018). Whether the starting material consists of DNA from bulk samples (community DNA) and/or from environmental DNA (eDNA), all DNA metabarcoding studies rely on a deceptively simple succession of core experimental steps: (a) sampling and preservation of the starting material, (b), DNA extraction, (c) PCR amplification of a taxonomically-informative genomic region, (d) high-throughput DNA sequencing of the amplicons, and (e) sequence analysis using bioinformatic pipelines. Despite this apparent simplicity, each step can potentially introduce its own sources of artifacts and biases (Figure 1). For example, the sampling design might not be effective for capturing the full taxonomic diversity or the ecological processes under study, an undesired bias for studies based on species detection. The availability of DNA in the samples is governed by its production rate, transport and persistence, processes which are all largely dependent on the targeted organisms, their biomass, and the ecosystem considered. A correct assessment of an ecological phenomenon based on DNA metabarcoding require not only implementation of standardized standardized, randomized and repeatable sampling designs and procedures (Dickie et al., 2018), but also consideration of DNA dynamics in the underlying matrix (i.e., in gut, faeces, water or soil matrices from tropical or boreal organisms/ecosystems; Barnes & Turner, 2016). Likewise, the community under study can be enriched on purpose or not with specific taxa depending on how the sample is collected (e.g., filter size for water samples, removal of roots or not for soils), how it is transported/preserved, and how DNA is extracted (differential extraction efficiencies). PCR amplification is also well known to be an important source of biases, that are now fully revealed with high-throughput DNA sequencing techniques. The preferential amplification of certain taxa over other ones due to inappropriate primers provides one such example of potential bias (Clarke, Soubrier, Weyrich, & Cooper, 2014; Deagle, Jarman, Coissac, Pompanon, & Taberlet, 2014). Primer biases can both skew abundance profiles and lead to false negatives. PCR amplification can produce false negatives too through the presence of e.g., PCR inhibitors, but also many false positives through the introduction of replication errors by the DNA polymerase or the formation of chimeric fragments (reviewed in Taberlet et al., 2018). False positives can also be introduced at any step of the experimental workflow through the presence of reagent contaminants (Salter et al., 2014), or through samples, extractions or PCR cross-contaminations. An even more insidious source of false positives pertains to the occurrence of “tag jumps”, sometimes referred to as “mistagging”, “tag-switching”, or “cross-talks” (Carlsen et al., 2012; Edgar, 2018; Esling, Lejzerowicz, & Pawlowski, 2015; Schnell, Bohmann, & Gilbert, 2015). PCR amplicons are indeed often tagged with unique short nucleotide sequences added on the 5’-end of the primers (i.e., “tags”), which allow pooling all PCRs within a single sequencing run and reducing sequencing costs. Each sequence obtained can then be bioinformatically assigned back to its sample of origin on the basis of its tags (Schnell et al., 2015). However, the procedures underlying the preparation of DNA libraries and/or the sequencing can introduce these “tag jumps”, when the tag assigned to one particular sample is in fact recombined to the sequences belonging to another sample (Taberlet et al., 2018). This introduces additional, non-negligible levels of sample cross-contaminations, which primarily involve the most abundant taxa and can have a disproportionate impact on samples with low DNA concentrations (Esling et al., 2015; Murray, Coghlan, & Bunce, 2015; Schnell et al., 2015). Similarly, the Illumina index located on the P5 sequencing adaptor can be subjected to “index jumps”, resulting in apparent cross-contaminations (Taberlet et al., 2018). This bias happens when several individual Illumina sequencing libraries are pooled and loaded on the same sequencing lane (Kircher, Sawyer, & Meyer, 2012). Finally, high-throughput DNA sequencing instruments have their own error rates (Schirmer et al., 2015). The above list of problems is clearly not exhaustive, and the interested reader will find more complete reviews elsewhere (e.g. Bálint et al., 2016; Nilsson et al., 2019; Taberlet et al., 2018). Still, it illustrates that any potential bias must be considered carefully when designing an experimental protocol and when interpreting the results. This is crucial to limit their impact on downstream analyses, and to ensure that the conclusion drawn from such data are authentic. There is now an increasingly diverse range of field, laboratory (e.g., Caporaso et al., 2011; Taberlet et al., 2018; Valentini et al., 2009) and bioinformatics (e.g., Boyer et al., 2016; Caporaso et al., 2010; Dumbrell, Ferguson, & Clark, 2016) procedures aiming at reducing the amount of both false negatives (i.e., due to partial sampling, extraction, amplification or sequencing bias) and false positives (i.e., due to contaminations, “tag/index jumps”, or PCR and sequencing errors) in DNA metabarcoding experiments. However, using these protocols does not necessarily guarantee that the problem of false positives or negatives is completely under control. These protocols must continuously be reconsidered, especially alongside the emergence of novel DNA sequencing technologies that provide new opportunities, but also new challenges. Additionally, each individual study and each genomic marker comes with its own specificities, and this often requires customization of the above protocols. The sequence clustering threshold to be used to form Molecular Operational Taxonomic Units relevant to the question addressed (e.g., removing intraspecific marker variability when the species level is desired) provides such an example, and will critically depend on both the marker specificities and PCR/sequencing error rates. Bioinformatics tools can further fail to exclude molecular artifacts when the filtering thresholds are relaxed, which inflates sample diversity estimates. Likewise, they can also generate false negatives, for example when a genuine metabarcode is falsely flagged as an error or chimera, or when it is assigned to an incorrect taxon due to incomplete or inappropriate reference databases (Alsos et al., 2018; Coissac, Riaz, & Puillandre, 2012). This can be especially problematic when the question investigated strongly relies on species detection. It is therefore crucial to include several types of experimental controls so as to facilitate the exclusion of spurious signal and support the reliability of the biological conclusions (Figure 1). Amongst these controls, conducting pilot experiments is particularly helpful to assess how appropriate the sampling design is (Dickie et al., 2018). We also recommend that both biological replicates (i.e., multiple independent samples) and technical replicates (i.e., multiple extractions/PCR of the same sample and/or extract) are included in the experimental workflow to disentangle the effect of both the biological and technical variances (Ficetola et al., 2015). These replications are necessary because both sampling and PCR can introduce biases in a stochastic manner, especially when the concentration of the target DNA is low. It is also essential to analyze a sufficient number of negative controls at the field sampling, DNA extraction, PCR, and sequencing steps, as well as positive controls consisting of mock communities, known DNA samples, or even synthetic sequences reflecting the attributes of the targeted products (Figure 1). All these controls must be sequenced along the biological samples, as they facilitate the detection of sporadic contaminations and tag or index jumps while helping adjusting filtering and clustering thresholds. Ultimately, they will be a token of the reliability of the whole data curation process (De Barba et al., 2014). We also encourage careful consideration of the bioinformatics workflow itself, since the filtering steps necessary to curate the data will critically depend on the experimental design and the ecological question under study. Typically, sequences of low abundance in a given sample may be genuine or artifacts deriving from PCR/sequencing errors or tag/index jumps. The retained filtering threshold for taxon presence is thus dependent on the underlying rates of artifacts, as well as on the sequencing depth. As the different experimental controls provide direct measurements of these artifacts, they will therefore allow better tuning of the filtering thresholds. All of these technical considerations should be precisely reported within publications together with relevant illustrations and statistics characterizing the workflow, as they are necessary to assess the relevance and quality of the data underpinning specific conclusions. A last, a most obvious example of control consists in assessing the plausibility of the taxonomic composition based on a priori knowledge of the system or taxa studied. Such knowledge can be derived from data obtained with complementary sensing approaches such as visual observations. In this case, building exhaustive local reference databases of the genomic marker used from local specimens will secure the taxonomic assignment step (e.g., Alsos et al., 2018). When local information is unavailable, typically when studying microorganisms, it remains possible to assess whether the community is composed of clades that are expected to occur in the system surveyed or not, as for example soils, sediments, and gut environments harbour highly different bacterial phyla (e.g., Thompson et al., 2017). As users, readers, referees or editors, we realize that the above mentioned issues remain too often overlooked. This problematic stance can lead to unsubstantiated claims and undermine scientific advances if not resolved. Inappropriate practices such as estimating species richness from fingerprint profiles (Bent et al., 2007), the absence of biological replicates (Prosser, 2010), or that of contaminant controls (Perez-Muñoz, Arrieta, Ramer-Tait, & Walter, 2017) have been repeatedly criticized in the field of microbial ecology, and in the latter case, they contribute to the rising debate about the existence or not of a womb microbiota. Ancient DNA research has also developed rigorous standards to tackle issues related to contamination, sequencing errors, and data reproducibility (Poinar & Cooper, 2000). We believe that the community of DNA metabarcoding users has now come of age and learnt from its past errors. At a time when more and more exhaustive guides of best practices on the subject are emerging (Knight et al., 2018; Pollock, Glendinning, Wisedchanwet, & Watson, 2018; Taberlet et al., 2018), and where DNA sequencing costs are rapidly decreasing, we should be always mindful of the adage “better safe than sorry”. This note does not mean to imply that the systematic use of the highest technical and analytical standards is reasonable nor the universal remedy for all the challenges associated with DNA metabarcoding. Rather, we strongly encourage researchers and end-users to adopt reflective decision-making when designing their experiment and to critically appraise their results, with the ultimate aim to prove the robustness and reproducibility of their conclusions.

Citations Over TimeTop 1% of 2019 papers

Abstract

Related Papers