L. Zinger and A. Gobet contributed equally to this work.
Evolution du Plancton et PaléoOcéans (EPPO), Station Biologique de Roscoff, UMR7144 Centre National de la Recherche Scientifique (CNRS) et Université Pierre et Marie Curie (UPMC), Place Georges Teissier, 29680 Roscoff, France
Genoscope (CEA), CNRS UMR 8030, Université d’Evry, 2 rue Gaston Crémieux, BP5706, 91057 Evry, France
Aquatic environments harbour large and diverse microbial populations that ensure their functioning and sustainability. In the current context of global change, characterizing microbial diversity has become crucial, and new tools have been developed to overcome the methodological challenges posed by working with microbes in nature. The advent of Sanger sequencing and now next-generation sequencing technologies has enabled the resolution of microbial communities to an unprecedented degree of precision. However, to correctly interpret microbial diversity and its patterns this revolution must also consider conceptual and methodological matters. This review presents advances, gaps and caveats of these recent approaches when considering microorganisms in aquatic ecosystems. We also discuss potentials and limitations of the available methodologies, from water sampling to sequence analysis, and suggest alternative ways to incorporate results in a conceptual and methodological framework. Together, these methods will allow us to gain an unprecedented understanding of microbial diversity in aquatic ecosystems.
Aquatic ecosystems, referred here as any water body, account for >70% of the Earth’s surface (excluding ice and groundwater ecosystems) and provide various goods and services for human populations, representing gigantic economic value (Costanza et al. 1997). Planktonic microorganisms (including Bacteria, Archaea, members of Eukarya (protists and fungi) and viruses) dominate these ecosystems in terms of both abundance and biomass. A litre of sea water contains approximately ∼106 eukaryotic cells (Brown et al. 2009), ∼108 prokaryotic cells (Whitman et al. 1998) and ∼109–1011 virus-like particles (Wilhelm & Matteson 2008). Aquatic microorganisms also represent a large and diverse pool of species (Slapeta et al. 2005; Wilhelm & Matteson 2008; Auguet et al. 2010); for instance, Bacteria within the global ocean are estimated to consist of more than ∼2.106 species (Curtis et al. 2002) and conduct a vast array of metabolic functions (Venter et al. 2004; Rusch et al. 2007). This biological pool sustains major biogeochemical processes (Cotner & Biddanda 2002; Falkowski et al. 2008): phytoplankton perform the majority of primary production in the Ocean and nearly half of the net primary production on Earth (Field et al. 1998), whereas virioplankton and heterotrophic prokaryotes and protists, together forming the ‘microbial loop’, contribute predominantly to organic matter and nutrient recycling (Azam et al. 1983; Pernthaler 2005; Pomeroy et al. 2007). As reported in ocean ecosystems (Behrenfeld 2011), primary productivity is likely to be affected by global warming, which may impact the microbial food web and diversity and thus threaten freshwater and marine resources (Dudgeon et al. 2006; Worm et al. 2006; Nogales et al. 2011). Higher biodiversity is assumed to increase ecosystem capacity to resist and recover from perturbation both by maintaining ecosystem functioning despite species loss and by diversifying the responses to this perturbation (Loreau et al. 2001). For example, diversified freshwater microalgal communities improve buffering of nutrient pollution by occupying a larger range of ecological niches (Cardinale 2011). Studying biodiversity of aquatic environments is therefore necessary for assessing, monitoring and anticipating their processes and sustainability (Duffy & Stachowicz 2006).
In its broadest meaning, measuring biodiversity consists of characterizing the number, composition and variation in taxonomic or functional units over a wide range of biological organizations (from genes to communities; Green et al. 2008; Magurran 2004). Microbial diversity has thus far been characterized extensively from a taxonomic angle at the community level, using different measures of diversity (Box 1). When describing microbial communities, numerous authors have underlined the difficulty in choosing the appropriate unit to measure diversity and this point has been extensively discussed elsewhere (Rossello-Mora & Amann 2001; Wilhelm & Matteson 2008; Caron et al. 2009). In particular, microbes’ short generation time and capacity for clonal reproduction results in a continuum of genetic diversity in microbial genomes that greatly complicates the identification of closely related microbial taxa (Acinas et al. 2004a). Such biological characteristics, together with the lack of clear morphological delineation and gene exchange among genetically unrelated taxa (i.e. horizontal transfer, Ochman et al. 2000; and multiple copies per genomes; Acinas et al. 2004b; Zhu et al. 2005), make the classification of microorganisms into discrete units difficult. This partly led microbial ecologists to adopt barcoding approaches (sensu lato; Valentini et al. 2009) where microbial species are equated to ‘operational taxonomic units’ (OTU), mostly based on small subunit ribosomal RNA gene similarities (Olsen et al. 1986). Although imperfect, this approach has offered several new insights into biogeographical patterns of aquatic microbial community, such as the effect of ecosystem types, taxa–area relationships or latitudinal gradients (Duffy & Stachowicz 2006; Pommier et al. 2007; Zinger et al. 2011), as well as their response to anthropogenic perturbations (Nogales et al. 2011).
Table Box 1. Measures of biodiversity
Studying diversity provides different pictures depending from which angle it is looked at. A first fundamental distinction arises from the consideration of species incidence, or their relative abundance. While the first aspect indicates the extent of resource partitioning between species, the second rather gives information on the way species interact for sharing these resources (Whittaker 1972). Second, diversity is a comparative science referring to spatially or temporally organized units (Magurran 2004). This organization is expressed by the following notions (Whittaker 1972; Fig. 1).
Alpha diversity refers to the diversity within one location or sample. It is often measured as species richness (i.e. number of species), seldom as species evenness (extent of species dominance). Species richness is strongly sensitive to sampling effort, and requires standardized samples, or the use of estimators that corrects undersampling biases, such as Chao1 or ACE. Evenness is less affected by undersampling biases and is usually assessed with Simpson’s or Pielou’s indices or rank abundance curves (review in Magurran 2004).
Beta diversity consists in determining the difference in diversity or community composition between two or more locations or samples (i) by considering species composition only, and use incidence data with associated metrics such as Jaccard or Sorensen similarity indices or (ii) by taking species relative abundances into account, and use Bray–Curtis or Morisita–Horn dissimilarity measures (Anderson et al. 2011). Using abundance data is, however, strongly discussed among microbiologists when dealing with rRNA gene data because of variations in gene copy number among strains (Acinas et al. 2004b; Zhu et al. 2005) as well as PCR artefacts.
Gamma diversity, or regional diversity, is similar to alpha diversity but applies for a larger area that encompasses the units under study.
Finally, the spatial scale of investigation can produce very different results and should be consistent in cross-study comparisons (Magurran 2004).
Measuring microbial diversity: from culture to pyrosequencing (Fig. 2)
Microbial diversity was initially studied through microscopy, and cultivation by using specific liquid and solid media (e.g. Zobell medium). Assigning taxonomy of phytoplankton and protists required tedious observations that relied completely on morphological traits, and bacterial diversity was assessed solely by morphotype description of the colony they would form on specific media. Rapidly, microbiologists realized that only 1% of the bacteria counted under the microscope could be cultivated on solid or in liquid media, and called this discrepancy the ‘Great plate count anomaly’ (Staley & Konopka 1985). Advances in molecular biology partly solved this problem using methods such as DNA–DNA re-association (Stackebrandt & Goebel 1994) or flow cytometry sorting of size-specific groups (Dorigo et al. 2005). Later on, the use of ribosomal RNA (Olsen et al. 1986) enabled the description of microbial taxonomic diversity, (i) by means of fingerprinting methods, which separate rDNA fragments according to their length and/or their nucleotide composition [i.e. automated rRNA intergenic spacer analysis (ARISA), terminal restriction fragment length polymorphism (T-RFLP), temperature or denaturing gradient gel electrophoresis (TGGE or DGGE) and single-strand conformation polymorphism (SSCP)], (ii) by microscopy, using FISH (fluorescence in situ hybridization) and derived methods (CARD-FISH, MAR-FISH), or (iii) by cloning 16S rRNA gene fragments and subsequently sequencing the clones following the Sanger sequencing method (Fig. 2, reviewed in Dorigo et al. 2005). Cloning/sequencing may be preceded by a flow cytometry size-sorting step, allowing an improved description of diversity, as shown for eukaryotic picophytoplankton (Shi et al. 2009). Although fingerprinting technologies enable the processing of many samples, they are inadequate for taxonomic identification and suffer from a lack of resolution. More importantly, calculating richness from DNA fingerprinting techniques remains impossible (Bent et al. 2007). Additionally, cloning/sequencing and FISH are not directly compatible with high-throughput approaches. The quest to describe microbial communities has now reached a new stage with the development of next-generation sequencing techniques (NGS), leading towards a high-throughput description of the microbial world at a higher level of detail than cloning or sequencing (MacLean et al. 2009).
Water samples sequencing: gaps and caveats
An overview of the literature focusing on microbial diversity in aquatic ecosystems through sequencing approaches is illustrated in Fig. 3. Although not exhaustive, this list still allows the identification of general trends from the last two decades. An increasing interest in microbial diversity in aquatic environments combined with the decreasing cost of molecular methods has produced a noticeable surge in publications since 2004 (Fig. 3a). Considering the methods used in the selected literature (Fig. 3a), one-third of the studies applied sequencing on isolates or selected bands of molecular fingerprints only, suggesting that a large part of the studies focused more on identifying selected microbial taxa rather than on quantifying the diversity of the overall community (Fig. 2). Cloning/sequencing remains by far the most widely used method for assessing microbial richness since 2004, although the use of NGS methods such as 454 pyrosequencing is expected to increase in the next few years (Fig. 3a). We considered four main water body classes: (i) saline and (ii) nonsaline environments, because these two water bodies harbour genetically distant organisms of Bacteria, Archaea, microbial eukaryotes and viruses (Logares et al. 2009); (iii) geothermal waters, which are characterized by high temperatures, high mineral concentrations, highly variable fluids and chemical properties, and where the food web revolves around autotrophs (e.g. hot springs lakes) and/or chemotrophs (e.g. deep-sea vents) depending on light availability (Rothschild & Mancinelli 2001; Fisher et al. 2007); and (iv) impacted water bodies, although the pollution originated from diverse sources (Fig. 3b). Considerable efforts have been made to characterize coastal waters and surface waters from the open ocean. Accordingly, patterns of diversity for Bacteria, Archaea and microbial eukaryotes have been identified in marine environments (Giovannoni & Stingl 2005; Pommier et al. 2007; Rusch et al. 2007; Wilhelm & Matteson 2008; Fuhrman 2009; Zinger et al. 2011). In contrast, and despite their higher economic value (Costanza et al. 1997), freshwater bodies have received less attention. Recent meta-analyses in both marine and freshwater environments revealed a higher bacterial and archaeal diversity in inland freshwaters, which is suspected to result from the lower connectivity and greater heterogeneity of these ecosystems (Auguet et al. 2010; Barberan & Casamayor 2010). This feature seems to also hold true for microbial eukaryotes and viruses (Logares et al. 2009). Among the nonsaline water bodies, lakes have been extensively studied (Fig. 3b), with a particular focus on the monitoring of blooms of certain microbial algae showing toxic/pathogenic properties. However, sequencing-based descriptions of microbial diversity remain scarce in the other nonsaline ecosystems (i.e. rivers and groundwaters). Considering lakes, rivers and coastal waters, only a few studies described their response to global changes and anthropogenic disturbances (see impacted sites in Fig. 3b), even though these ecosystems are experiencing such pressures (Dudgeon et al. 2006; Halpern et al. 2008). A better picture of microbial community composition and diversity in these natural ecosystems is therefore urgently needed. In this regard, NGS could rapidly provide significant insights. Finally, a fair number of studies have been carried out on Bacteria and, to a lesser extent, Archaea in geothermal ecosystems, aquifers and activated sludge waters.
Among the microbial domains studied, Bacteria undoubtedly have received most of the attention in almost all of the ecosystems that we defined, representing 77.6% of the selected literature (Fig. 3b). In contrast, and despite their abundance, diversity and significant role in aquatic ecosystems (Auguet et al. 2010), archaeal diversity was less studied, especially for nonsaline waters. Similarly, based on sequencing approaches, there are fewer descriptions of Eukarya and most viruses despite their high abundance and diversity (Finlay 2004; Slapeta et al. 2005; Fernandez-Alamo & Farber-Lorda 2006; Wilhelm & Matteson 2008; Stoeck et al. 2010). Both viral and eukaryotic diversity has mostly been characterized in nonsaline lakes and marine ecosystems (Fig. 3b), whereas Eukarya also occur in more extreme ecosystems (Rothschild & Mancinelli 2001). To conclude, the lack of reports regarding nonbacterial domains constitutes a gap in our understanding of microbial diversity. This emphasizes the need of cross-domains surveys to better understand overall patterns of diversity. So far, only 11.2% of the selected literature focused on at least two microbial domains, 2.2% on at least three of them.
Sampling issues and potential solutions
Community and sample representativeness
Our vision of the diversity and composition of any biotic assemblage strongly depends on the sampling procedure, which has to be established with careful consideration (Magurran 2004). The obvious issue of contamination will not be addressed here, but should clearly be kept in mind. The sampling strategy applied relies on the initial question being addressed: the estimation of parameters, such as richness (or alpha diversity, Box1), requires more extensive sampling effort so as to increase the accuracy of parameter estimates. In contrast, the resolution of patterns of diversity instead requires standardized sampling strategies so as to maximize the power to detect differences between sites (Kenkel et al. 1989). However, there is debate over this latter point, because obtaining a representative description of microbial communities may require different degrees of sampling effort depending on habitat heterogeneity and patchiness. If one attempts to compare two habitats displaying contrasting heterogeneity, using standardized sampling strategies might therefore render an altered picture of diversity patterns (Cao et al. 2002).
Marine, inland and groundwaters display various degrees of heterogeneity on a broad range of spatial and temporal scales in terms of particles, organic matter and nutrient distribution, physico-chemistry and physical mixing (Karl 2007; Griebler & Lueders 2009; Grossart 2010). This also holds true for microbiological properties. For instance, global estimates of prokaryote cell density have been reported to be much lower in sea water than in freshwater (Whitman et al. 1998). At a regional scale, bacterial cell density has been observed to be higher in coastal waters than in the open ocean (Acinas et al. 1997). Furthermore, the scale of bacterioplankton community turnover varies, horizontally, vertically and temporally in the ocean (reviewed in Fuhrman 2009), and similarly, the taxa–area relationship of aquatic microbes has been reported to be steeper in island-like habitats, such as lakes, than in contiguous habitats such as the open ocean (Prosser et al. 2007).
These properties emphasize the need to maximize the community representativeness even in the case of pattern recognition. Considering a given habitat or a given environmental condition, community attributes will be better characterized by increasing the community autosimilarity, i.e. the average similarity in community composition and diversity among replicate samples (Cao et al. 2002). This can be achieved by increasing the number of replicate samples and/or the sampling volume. Replication is of particular concern for microbial ecologists that are facing strong undersampling biases when collecting samples on the field, representing only a tiny portion of the ecosystem surveyed from a microbial point of view. Such undersampling is a source of great variability between samples originating from a single habitat. For instance, based on microscopic observations, a recent study showed that seawater samples collected only a few minutes and metres apart harboured planktonic protistan communities varying in their composition and structure, mainly owing to rare taxa (Dolan & Stoeck 2011). This renders the comparison of two different habitats impossible without sufficient replication. Use of the appropriate number of replicates is therefore of the utmost importance, and this should be prioritized instead of obtaining many sequences from a single sample (Magurran 2004; Prosser 2010).
Besides replication, the sample size, or volume, may enhance the community representativeness. According to the taxa–area relationship detected in various aquatic environments and for different microbial taxa (Fuhrman 2009; Prosser 2010), increasing sampling size would inherently increase the likelihood of capturing higher numbers of cells and/or to encompass a larger number of microhabitats harbouring different taxa number and abundance (Magurran 2004). This relationship is, however, not uniform among habitats (Prosser et al. 2007), complicating the definition of a unique standard sample volume. As a compromise, one could find a standard sample volume that is technically feasible while at the same time minimizes the variability of the attributes of the ecological communities (Cao et al. 2002). This question has received little attention so far in aquatic microbiology, and it is generally left to the discretion of the investigators to choose the appropriate sample volume, which typically ranges from 10 mL to more than 200 L. Nevertheless, fingerprint profiles of increasing volumes of water (from 10 to 1000 mL) showed no significant differences in richness (Dorigo et al. 2006) and low variability of community structure when sampling more than 50 mL of seawater (Ghiglione et al. 2005), but this has not been broadly confirmed using sequencing-based approaches. Gomez-Pereira et al. (2010) provided some insights into the effect of water volume for detecting rare flavobacterial clades by means of FISH, showing that cell abundance variability was reduced by two when using at least 250 mL of water. Hence, the sampled volume effect remains to be determined when using NGS approaches where the genetic diversity revealed is much higher.
The same considerations hold when deciding the sequencing effort to apply on each sample, as it constitutes a second sampling step for which microbiologists again struggle with under-representation. For instance, Quince et al. (2008) estimated that assessing 90% of the diversity based on 16S rRNA genes would require a sequencing effort five times higher in surface open ocean waters than in the case of the GOS survey (∼7000 sequences initially used; Rusch et al. 2007) and 280 times higher in Sogin’s rare biosphere study on vent fluids (∼442 000 sequences initially used; Sogin et al. 2006). Nevertheless, sequencing effort does not appear to affect the rankings of the observed and estimated richness when comparing different conditions (Shaw et al. 2008), and seems of importance only for parameter estimation. It is worth mentioning here that the sequencing effort required may also differ among taxa, as Bacteria appear to have diversified an order of magnitude more than Archaea and Eukarya (Amaral-Zettler et al. 2011).
Aquatic microbiologists often resort to filtration to concentrate diluted microbial cells from water. Usually, sufficient amounts of DNA for subsequent molecular analyses are obtained by using 0.22-μm polycarbonate filters (see Bej et al. 1991 for an overview of filter types), although this mainly depends on the filtered water volume. Water filtration is the only option when processing large volumes, as these cannot be readily handled with centrifugation. However, filtration may also select subsets of the existing microbial populations: as mentioned earlier, microbes inhabit, grossly, the free-living and the attached fractions of the water (Grossart 2010). The latter including ‘marine snows’ also recognized as microscale environments may be lumped together during filtration processes (Azam & Long 2001; Kiorboe et al. 2003). Although the attached microbes likely account for an important fraction of microbial diversity in the water, most surveys only focus on the free-living fraction, emphasizing the need to develop alternatives to water filtration (Grossart 2010).
Disentangling the extracellular DNA from the dead, the dormant and the active cells
Microbial ecologists all share the frustration of dealing with large communities mixing live, dormant and dead cells. Although the dead cell DNA is substantially involved in ecological and evolutionary processes (see Lorenz & Wackernagel 1994; Vlassov et al. 2007 for reviews), one may prefer to assess the actual, living microbial cells for diversity estimation and pattern recognition. Because dissolved DNA is generally assumed to pass through filtration, this issue is often overlooked in environmental studies, despite the high concentration of naked DNA in the aquatic environment, ranging from 0.2 to 44 μg/L, with higher values especially in estuarine and coastal ecosystems (reviewed in Lorenz & Wackernagel 1994).
To focus on the living microbial fraction for diversity assessment, one could subject the samples to treatment with propidium monoazide (PMA; Nocker et al. 2007), propidium iodide (Luna et al. 2002) or ethidium monoazide (Soejima et al. 2008) prior to DNA extraction (see Cenciarini-Borde et al. 2009 for a review). These molecules intercalate between DNA strands but cannot penetrate intact cells, therefore rendering the naked DNA nonamplifiable by PCR. The contribution of dead cells in microbial community structure in both freshwater and seawater samples was recently tested using PMA and 454 pyrosequencing (Nocker et al. 2010). Slight changes in bacterial phyla proportion were observed with PMA treatment alone, but were enhanced with the addition of an extra heating step, causing cell membrane damage. These results prove the efficiency of PMA and suggest that naked DNA hampers the assessment of bacterial community structure at broad taxonomic resolutions. However, more effort needs to be invested in accounting for the effects of dead material at finer taxonomic levels and for diversity estimation.
The mixing of DNA from both active and dormant cells in a single DNA extract may also cause frustrations. Dormancy is a common response of microbes to cope with stressful conditions and has also important ecological implications, especially regarding ecosystem resilience (Lennon & Jones 2011). The proportion of dormant cells within a sample represents up to ∼35% and ∼50% of the total cell amount in marine and freshwaters, respectively (Lennon & Jones 2011). Here again, depending on the initial question, one may assess the active microbial cells only, so as to better link microbial diversity to ongoing ecosystem processes. In this case, the integration of dormant cells constitutes a potential bias because this ‘seed bank’ may inflate alpha diversity while decreasing beta diversity at the same time (see Box 1 for definition of terms; Lennon & Jones 2011).
Sequencing the rDNA and reverse-transcribed rRNA fragments so as to obtain both total DNA pool and metabolically active microbial populations is a solution for excluding dormant cells. This approach is increasingly used for seawater ecosystems (e.g. Moeseneder et al. 2005; Frias-Lopez et al. 2008; Ghiglione et al. 2009; Rodriguez-Blanco et al. 2009), but to our knowledge, has not been yet used for freshwater or groundwater. Most of these studies reported some differences in diversity and taxa proportions between total and active prokaryotic communities, suggesting that metabolically active taxa are not necessarily the most abundant (Moeseneder et al. 2005). To our knowledge, the very few studies focusing on active aquatic Eukarya revealed similar trends (Stoeck et al. 2007; Not et al. 2009).
Another alternative to avoid both naked DNA and dormant cells is provided by incubating samples with bromodeoxyuridine (BrdU), a thymidine analogue, which is incorporated into DNA of growing microbial populations (Urbach et al. 1999). The resulting DNA extract can then be processed in combination with BrdU magnetic bead immunocapture and sequencing, giving access to the active microbial diversity. By doing so, Taniguchi & Hamasaki (2008) showed net differences in community composition and structure between the active and total microbial pool. Such differences may be not only attributed to high mortality of particular microbial taxa owing to grazing and viral lysis, but also to lower growth rates of active microbial taxa. Although a few microbial isolates have been reported to not incorporate BrdU (Urbach et al. 1999), this approach seems a good alternative to assess the active fraction of microbial diversity. Finally, DNA stable-isotope probing (DNA-SIP) relies on similar concepts but incorporates stable-isotope-labelled compounds 13C, 15N in the newly synthesized DNA, which can be further processed after isopycnic centrifugation and the identification of enriched DNA (see Chen & Murrell 2010; Morales & Holben 2011 for reviews).
Pitfalls and potential solutions in molecular approaches
Molecular-based studies intrinsically contain biases that may introduce discrepancies in measures of microbial diversity. First, in regard to DNA extraction steps, there have been several reports of cultures and samples recalcitrant to DNA and/or RNA extraction because of cell resistance to lysis and the presence of PCR inhibitors, such as humic acids or proteins (von Wintzingerode et al. 1997). Consequently, a panel of additional steps has been proposed to improve DNA extraction, such as sample freeze–thaw, additional chemical lyses or bead-beating to break down recalcitrant cells (Bej et al. 1991; Ferrera et al. 2010). Similarly, various methods and commercial kits for DNA purification have been developed to reduce the amount of PCR inhibitors in DNA extracts (Miller et al. 1999; Jiang et al. 2005).
The second source of limitation is the PCR performed to amplify targeted genes. Describing established pitfalls and limitations of PCR is out of the scope of this study (see von Wintzingerode et al. 1997 for a review), and we will instead focus on issues arising from the emergence of next-generation sequencing (NGS). First, different universal PCR primers have been reported to miss a large part of both prokaryote and protistan diversity when using 454 pyrosequencing (Jeon et al. 2008; Hong et al. 2009). Hence, the degree of universality of the chosen primers (or cocktails of primers) ought to be carefully considered when interpreting the resulting structure and diversity patterns (Huber et al. 2009). Second, PCR does not represent the real community structure as it unequally amplifies DNA fragments according to the number of PCR cycles performed and the DNA polymerase used. Further, polymerase enzymes often produce errors such as mutations, chimeras or heteroduplexes (von Wintzingerode et al. 1997). Such wrongly amplified fragments may pollute public databases and generate cascading mistakes (Hugenholtz & Huber 2003).
Nevertheless, these biases can be corrected in the ‘wet’ laboratory using proof-reading enzymes (Acinas et al. 2004a) and by adjusting PCR conditions (Acinas et al. 2005). Further, the likelihood of observing these errors is reduced by targeting shorter DNA fragments (Liu et al. 2007; Huber et al. 2009), which is particularly relevant in the context of NGS, which generates, for now, relatively short DNA sequences (∼100–400 bp). Another alternative to PCR biases may be emulsion PCR, where each fragment is amplified separately in a microdroplet, thereby avoiding the formation of heteroduplexes and chimera (Nakano et al. 2003). This method is systematically used prior to processing DNA with NGS (MacLean et al. 2009) on crude DNA extracts or PCR products. Finally, with the decrease in costs, PCR-free approaches such as whole-genome amplification (WGA; e.g. Gonzalez et al. 2005), whole-genome sequencing (WGS, Rusch et al. 2007; Venter et al. 2004) or direct sequencing (Shendure & Ji 2008; MacLean et al. 2009) may become valuable approaches.
The emergence of NGS has great potential for assessing microbial diversity (see extensive reviews on the methods in (MacLean et al. 2009; Shendure & Ji 2008), but the relative novelty of the technique precludes a clear identification of their limitations. Currently, Roche® 454 pyrosequencing technology has been the most widely used NGS method for characterizing microbial diversity (Fig. 3a). A recent comparison of this technique with Illumina® technology for microbial diversity assessment found both approaches to be similar, although the taxonomic assignment of 454 Titanium reads provided longer fragments, whereas Illumina allowed a greater coverage (Claesson et al. 2010). The read quality generated by 454 pyrosequencing generated a debate in the microbial ecology community because of inaccurate assessment of homopolymer length in two strictly identical DNA fragments, leading to an inflation of microbial diversity estimates (Quince et al. 2009). On the other hand, 454 pyrosequencing appears better for metagenomic surveys compared to fosmid libraries coupled with Sanger sequencing, the latter one tending to over-represent GC-rich fragments (Temperton et al. 2009).
Current tools/pipelines available to assess diversity in the ‘dry lab’
As mentioned previously, the DNA barcoding approach (sensu lato; Valentini et al. 2009) has become fairly popular for the rapid assessment of microbial diversity through the use of partial or complete ribosomal genes (Olsen et al. 1986). Ideally, these barcodes have (i) to be short enough and flanked by highly conserved regions for targeting a given taxa in an order, (ii) to be suitable for all taxonomic groups, (iii) to allow taxonomic assignment and (iv) should permit the definition of taxonomic levels, e.g. OTUs as surrogates of species, from sequence data sets (Valentini et al. 2009). Unfortunately, the situation is not ideal in this regard given the lack of ‘universality’ of most primers currently used and limited databases, especially for microbial eukaryotes (Stoeck et al. 2010). Most importantly, the inherent differences in evolution rates among microbial taxa (Giovannoni & Stingl 2005; Thornhill et al. 2007) may make it difficult to choose a universal sequence similarity threshold for defining taxonomically meaningful OTUs. This issue applies differently according to the rRNA hypervariable region chosen (Schloss 2010) and is further clouded by the presence of artefact sequences generated during PCR or sequencing. Cutting into the similarity tree of DNA sequences for creating OTUs is therefore a delicate task and requires making compromises between all potential sources of genetic variability in sequence data sets.
Postsequencing in silico approaches can handle some of these issues (Table 1). First of all, errors because of sequencing have to be removed to avoid artificial inflation of diversity estimates (Reeder & Knight 2009) and can be identified through sequence quality. For instance, the per nucleotide-error rate of sequence data (∼0.25% for 454 pyrosequencing; Huse et al. 2007) may be lowered by removing sequences containing one or more unresolved nucleotides (N’s), errors in the primer sequences, and sequences where length differs significantly from the expected (Huse et al. 2007). Programs to remove 454 pyrosequencing and PCR errors have been developed: PyroNoise and the DeNoiser cluster flowgrams (Quince et al. 2009; Reeder & Knight 2010), whereas the single-linkage preclustering (slp) approach (Huse et al. 2010) uses sequences. AmpliconNoise is a development of PyroNoise that first clusters flowgrams to remove 454 errors and then sequences to remove PCR errors (Quince et al. 2011). AmpliconNoise and PyroNoise are both iterative probabilistic methods, whereas the DeNoiser and slp use faster agglomerative strategies. AmpliconNoise is able to remove more noise than the other programs without overclustering and removing true variation which the agglomerative algorithms are prone to do (Quince et al. 2011). Chimeric DNA sequences constitute a second source of bias and are harder to detect. Initially, several programs, e.g. ChimeraCheck (Cole et al. 2005) or Bellerophon (Huber et al. 2004), were developed for chimera detection and have been extensively used on classical clone libraries. More recently, algorithms adapted to the length and large amount of sequences generated by NGS, such as ChimeraSlayer and Perseus, have also been proposed (Table 1).
Table 1. Examples of tools to process DNA sequences (mainly for NGS data sets)
Agglomerative approach using flowgram alignments. Flowgram clustering based on the assumption that sequences with errors tend to be rare and should be similar to true abundant sequences. Cons: Highly computer-intensive
Agglomerative approach using a flowgram alignment. Flowgram clustering based on centroid approach prior to classical clustering step. Pros: Fast to compute. Cons: Misassignment of reads resulting in loss of accurate OTU clustering
Agglomerative approach using sequences by applying a preclustering at 98% sequence similarity with a single-linkage approach before the classical clustering step. Based on pairwise alignments. Pros: Improve OTU clustering. Cons: Possible misassignment of reads resulting in inaccurate diversity estimates.
Improved version of PyroNoise (AmpliconNoise) with chimera detection (Perseus). Perseus uses sequence abundances in the classification of PCR chimeras. Pros: More accurate than DeNoiser and slp to estimate OTU numbers. Does not require a set of nonchimeric sequences. Able to find 99% chimeras (more than ChimeraSlayer)
Preprocessing algorithm that trims, filters, dereplicates sequence reads and removes chimeras. Pros: Performs equally well with any type of sequence data set (DNA libraries or pyrosequencing reads). Keeps a maximum of sequences (no overtrimming). Friendly user interface. All preprocessing steps can be manually verified.
MSA. Pairwise alignments on sequences and construction of a similarity tree for combining the alignments. Use neighbour-joining algorithms for tree construction Pros: Phylogenetic inferences Cons: Low-quality alignments when processing large and diversified sequence data sets.
MSA. Pairwise alignments on partial sequences and construction of a similarity tree for combining the alignments. Includes various MSA strategies. Pros: MSA optimized for large data sets. Choice in the alignment method. Cons: Low-quality alignments depending on the algorithm chosen.
MSA. Pairwise alignments on partial sequences and construction of a similarity tree for combining the alignments. Iterations for alignment improvement. Pros: MSA optimized for extremely large data sets. Cons: Low-quality alignments when processing large and diversified sequence data sets.
Fast divisive clustering algorithm for graphs based on simulation of the flow in the graph. The graph is composed of all sequences, more or less connected according to their similarities. Pros: Fast and sensitive. Cons: May fail to accurately cluster too dense graphs.
Commands from several tools for sequence processing (e.g. chimera removal, sequence alignment, OTU clustering), alpha diversity (e.g. rarefaction curves, diversity estimators) and beta-diversity analyses (e.g. distance indices, Unifrac). Pros: Ease-to-use commands with a well-described Wiki. Different methods available
Sequence processing and diversity assessment (e.g. low-quality reads removal, pairwise sequence alignment, clustering, species estimates). Pros: Parallel structure implemented, for large data sets. Cons: Restriction in analyses available
Integrates several tools for sequence processing (e.g. quality filtering, denoising, sequence alignment, clustering, taxonomic assignment, alpha diversity and beta diversity analyses). Pros: Ease-to-use, data visualization. Cons: Restriction in analyses available
Agglomerative approach using sequences. Quality filtering at 0.2% per-base error probability. Clusters sequences at 97% by using PyroClust or Uclust. Taxonomic assignment and chimera identification. Pros/Cons: Data upload on the web interface may be time-consuming. The pipeline can be run locally.
Sequence reads are processed in two steps: (i) sequence trimming and filtering and (ii) analysing step, including NCBI-base taxonomic assignment and diversity analysis (e.g. rarefaction estimates). Pros: Clear manual leading the user. Cons: Can only be run if eight additional programs are downloaded locally. Some basic Perl programming skills may be needed.
Many tools for sequence processing (e.g. sequence alignment, clustering, taxonomic identification) and diversity analyses (alpha and beta diversity). Pros: Ease-to-use, formats data for other applications (e.g. MOTHUR, EstimateS) Cons: No sequence trimming tool
Sequencing processing (e.g. quality filtering, sequence alignment, clustering, taxonomic identification) and diversity analyses (alpha and beta diversity). Pros: Ease-to-use Cons: Split the data set into classified and unclassified data sets before clustering
Many tools for sequence processing (e.g. low-quality reads removal, clustering, taxonomic identification) and diversity analyses (alpha diversity). Pros: Ease-to-use, highly flexible, outputs are generated at each steps Cons: No alignment step is used, which might result in less accurate clusters
After trimming, the data are then submitted to alignment algorithms in order to calculate dissimilarities between sequences. The choice of the algorithm is of prime importance and is dependent on the phylogenetic inferences made, for which multiple sequence alignment tools (MSA) are necessary. However, recent concerns have arisen regarding the alignment quality provided by classical MSA (e.g. CLUSTALW or MUSCLE; Table 1). For instance, these algorithms have been shown to result in different tree topologies even by excluding gapped sites, and perform poorly when large and diversified sequence data sets are analysed (Wong et al. 2008). Another study showed that Greengenes and MUSCLE algorithms tended to misestimate both alpha diversity and beta diversity (Schloss 2010). MSA incorporating the rRNA gene secondary structure (e.g. SINA, Pruesse et al. 2007) rely on biological concepts and should be preferred (Schloss 2010), but the use of pairwise algorithms alignments is still valid (Huse et al. 2010; Schloss 2010), especially when one focuses on poorly referenced taxa or non-rRNA genes. Distance matrices calculated from sequence alignments are then used for clustering sequences into OTUs based on their similarities (e.g. MOTHUR, Schloss et al. 2009) or patristic distances from phylogenetic trees (Pommier et al. 2009; Table 1). A large panel of clustering algorithms (a few examples are given in Table 1) is available for this purpose, of which some are tested and discussed in Schloss & Westcott (2011). These methods still require an arbitrary cut-off similarity threshold, but a promising alternative, CROP (Table 1), has been recently proposed (Hao et al. 2011) and may cope better with the differences in evolution rates among taxa. With the routine use of sequencing technologies, an increasing number of pipelines that include most or all aforementioned steps are now available, are able to handle NGS data sets and are user-friendly (Table 1).
The development of NGS provided access to the microbial ‘rare biosphere’ and its patterns of diversity (Sogin et al. 2006; Galand et al. 2009), but its existence has been highly debated in the light of potential NGS technology biases (Reeder & Knight 2009). In fact, more than 50% of the OTUs obtained are represented by a few or even one single sequence, even when sequence trimming was applied (e.g. Gilbert et al. 2009; Pommier et al. 2010; Agogue et al. 2011). These OTUs are often suspected to be artefacts and may be discarded in further analyses (Reeder & Knight 2009). By doing so, a few studies have tested the effect of rare OTUs on the resulting diversity patterns by removing an increasing proportion of them in the whole sequence data set (Galand et al. 2009; Pommier et al. 2010; Agogue et al. 2011), and a specific tool has been developed to test for such an effect in complex data sets (Gobet et al. 2010). The effect of rare OTU removal has not been described for patterns of alpha diversity yet. Our analysis shown in Fig. 4a–c suggests a negligible effect, but this might change according to community evenness and would require further work. Patterns of beta diversity appeared conserved without singletons or rare OTUs (Gobet et al. 2010; Pommier et al. 2010; Fig. 4d–f), but the noise introduced by rare OTUs in the data may reduce pattern detection in certain cases (Agogue et al. 2011). Finally, depending on the study, these rare OTUs have been more or less successfully taxonomically assigned (e.g. Galand et al. 2009; Agogue et al. 2011). A systematic removal of these rare OTUs may thus not always be justified, but could certainly enhance the detection of diversity patterns at the community scale.
Final recommendations and future challenges
As in several fields of biology, microbial ecologists have welcomed NGS technologies. Owing to their novelty, many discussions about their potential limitations and biases are still ongoing at the expense of other methodological aspects that are likely to deeply alter our perception of microbial diversity. Here, we have presented an overview of tools and concepts existing to assess microbial diversity through sequencing approaches, from sampling to the analysis of environmental sequences. We do not recommend one single methodological pipeline over the others, as all existing methods are differently adapted to each nuance of biodiversity science. The correct pipeline should therefore be chosen carefully, by considering all sample processing steps and by keeping in mind the initial question, the background knowledge of the organisms and ecosystems studied, as well as methodological limitations.
Because of their deeper coverage, one might wonder whether NGS technologies may render null and void all results previously achieved by means of cloning/sequencing or fingerprinting approaches. Studying diversity in the first place is a comparative science (Magurran 2004), and when considering community richness or evenness (Box 1), one usually attempts to identify the conditions harbouring higher or lower diversity, but not their actual values, the latter obviously varying according to the method and estimator used. Although fingerprinting methods poorly estimate alpha diversity (Bent et al. 2007), they may be useful for a ‘quick and dirty’ snapshot of the communities (Gillevet et al. 2009) and may help sample selection for deeper characterization through NGS sequencing. However, such a feature remains bound to the use of similar regions of the rRNA genes, which may provide different patterns of alpha diversity (Schloss 2010). Cloning/sequencing and NGS may give similar patterns of richness, but different pictures of species evenness, as shown by in silico sequencing depth simulations (Shaw et al. 2008). Regarding patterns of beta diversity, all these methods seem to provide similar results for salt marsh communities (Gillevet et al. 2009), but this pattern ought to be confirmed in aquatic environments.
Finally, two main lessons can be learned from the last two decades of water sequencing. First, the use of sample replicates and their individual sequencing is indispensable. Lack of replication, or pooling replicate samples prior to sequencing alter community representativeness, renders an inaccurate picture of microbial diversity and further precludes comparisons of different conditions in a statistically reliable way (Prosser 2010). We acknowledge that sample replication might have been difficult when using cloning/sequencing approaches that are both costly and time-consuming. However, NGS technologies are in essence high-throughput methods and produce relatively good-quality data when trimmed by appropriate bioinformatic tools and have become as cheap (per sample cost) as fingerprinting approaches when considering multiple samples. Furthermore, in the case of diversity pattern detection at the community scale, the main ecological signal is not affected by the removal of rare taxa (Agogue et al. 2011; Gobet et al. 2010; Pommier et al. 2010; Fig. 4). In this context, deep sequencing is not necessary and might allow processing of more replicate samples at reduced costs.
Second, our understanding of microbial diversity and its underlying processes would be fruitless without establishing correlation, or lack of correlation, between microbial diversity and environmental parameters. Through increasing awareness of this issue, robust and standardized measurements of environmental parameters have now become imperative, especially in the age of vast sampling campaigns (e.g. the Sorcerer Global Ocean Sampling Expedition http://camera.calit2.net/about/gos.shtm, the International Census of Marine Microbes, http://icomm.mbl.edu/ or the Tara Oceans Expedition, http://oceans.taraexpeditions.org/) and the increase in open access sequence data in concert with cross-study meta-analyses. The Genomic Standards Consortium initiative has consequently developed a list of minimal information that should be provided with any marker gene sequences deposited in international databases (MIMARKS, Yilmaz et al. 2011), namely the geographical xyz coordinates as well as a standardized description of the environment (Environment Ontology, http://gensc.org/gc_wiki/index.php/Habitat-Lite). With geographical coordinates, additional environmental data can be easily retrieved from online databases such as Megx (http://www.megx.net/, Kottmann et al. 2010) or Pangaea (http://www.pangaea.de/), but to our knowledge, these databases have exclusively been developed for ocean and coastal waters, emphasizing the need for similar development in freshwater ecosystems.
Challenges remain in estimating microbial diversity that cannot be reached even with NGS technologies (Quince et al. 2008). The limitations in the characterization of rare taxa or the ‘seed bank’ (Lennon & Jones 2011) reduce our ability to capture microbial community responses to environmental changes and by extension, to understand ecosystem resilience. This supposes an understanding of the functional diversity of microbes as well, which is increasingly studied through ‘omics’ methods; these have so far allowed the discovery of new functional genes, metabolic pathways and their biogeographical patterns (Venter et al. 2004; DeLong & Karl 2005; Rusch et al. 2007). Concepts and methods are still evolving and propose promising perspectives for establishing a link between microbial diversity, their functions and ecosystem processes and stability (Green et al. 2008; Hofle et al. 2008; Morales & Holben 2011).
What is the future of environmental sequencing?
As with the appearance of the Sanger method in the early 1970s, two major recent progresses in molecular biology, i.e. high-throughput sequencing and active-cell probing, have revolutionized our perspective on current aquatic microbial ecology (Hofle et al. 2008). Yet, most studies fail to use both methods conjointly or only target a single gene, with a clear preference for 16S rRNA genes (Chen & Murrell 2010). While the adoption of NGS is continuously increasing (McCarthy 2010), single-cell sequencing approaches also bring appealing perspectives to environmental biology (Woyke et al. 2010). We foresee the development of new approaches that will exploit these tools jointly, expand their targets to full genomes and associate them with quantitative measurements. This ‘Real-Time Metagenomic’ would target all active members (including abundant and rare individuals) of the studied community, measure the single-cell content of transcribed genes and its full genome capability. Behind the conceptual and methodological innovation, such an approach ought to include strict sampling replication and proper statistical analyses and would illuminate one of the most complex and important trophic guilds on the planet.
We are indebted to Fabrice Not, Alice Valentini, Bénédicte N. Poncet, Hannah Marchant, Tim Vines and two anonymous reviewers for valuable comments and suggestions that improved the manuscript. LZ thanks Antje Boetius for the opportunity to work in her group. Her work was financed by the Leibniz program of the DFG to AB. Work from AG was financed by the ANR project PROMETHEUS [ANR-09-GENM-031].
The authors have a long-standing experience in the production and analysis of microbial diversity datasets based on sequencing. They are interested in inferring the ecology, diversity and biogeography of microbes, by using concepts of classical ecology together with the development of molecular and bioinformatic tools. L.Z. is a post-doctoral researcher whose research focuses on soil and marine prokaryotic and fungal diversity. A.G. is a post-doctoral researcher specialized in the diversity of protists and bacteria in marine waters and sediments. T.P. is a microbial ecologist whose research has long focused on marine diversity and distribution, he now focuses on the functional diversity of bacteria involved in the N cycling in soil and in the turnover of organic matter in aquatic environments.