Will different OTU delineation methods change interpretation of arbuscular mycorrhizal fungal community patterns?



See also the Commentary by Powell and Sikes

Progress in microbial community ecology is challenged by the fact that individuals often cannot be morphologically identified and counted, and the great majority of taxa are not phenotypically characterized because they lack recognizable traits and are not in culture. As a result, microbes are often distinguished from one another using molecular sequence data. For example, ≥ 97% sequence similarity in the 16S rRNA gene is commonly used to separate species in bacteria (Stackebrandt & Goebel, 1994; Schloss & Handelsman, 2006) although some argue that 99% is more appropriate (Acinas et al., 2004). Such differences in universal thresholds may influence our ability to understand the ecology of sequence clusters, or operational taxonomic units (OTUs). If sequence similarities are too narrow (e.g. 99%), comparisons among communities are difficult, and if they are too broad (e.g. 90%), OTUs with different ecological roles and distribution patterns may be erroneously lumped together (discussed in Knights et al., 2011a). Universal thresholds also do not consider differences in speciation and substitution rates among lineages and may therefore not capture equivalent units of diversity. Indeed, Koeppel & Wu (2013) showed that many bacterial OTUs are paraphyletic and span multiple ecological habitats, and Youngblut et al. (2013) demonstrated that ecological patterns may be optimally detected by using different OTU delineations among lineages.

Arbuscular mycorrhizal fungi (AMF) present similar challenges. AMF are asexual, obligate symbionts that exchange nutrients and other services for plant carbon (Smith & Read, 2008). Long-term asexual evolution has led to high genetic and ecological diversity within known AMF species (Munkvold et al., 2004; Koch et al., 2006; Rosendahl, 2008; Stockinger et al., 2009) that traditionally have been identified based on spore features. Molecular techniques show that, like bacteria, many environmental sequences do not cluster with cultured species (Rosendahl, 2008), but there is currently no consensus on how to best organize sequences into biologically and ecologically meaningful taxonomic units (Öpik et al., 2010). While some researchers use universal – albeit varying – OTU thresholds (Dumbrell et al., 2011) or BLAST matches against published sequences (Lekberg et al., 2007), others attempt to control for evolutionary differences by identifying OTUs as groups forming monophyletic clades (Rosendahl & Stukenbrock, 2004; Sykorova et al., 2007). How these methods affect our understanding of AMF community patterns along environmental and spatial gradients is more or less unknown (but see Powell et al., 2011). This is a pressing question given that high-throughput pipelines for post-sequencing analyses of next generation sequencing (NGS) data often utilize universal thresholds for OTU delineations (Caporaso et al., 2010), as groupings based on evolutionary relationships are computationally expensive for large data sets.

Here we compare how similar two OTU delineation methods, one that considers evolutionary origins and another based on 97% sequence similarity, distribute AMF OTUs along environmental and spatial gradients. For the evolutionary approach, we manually combined sequence groups into OTUs that formed monophyletic clades (hereafter referred to as the monophyletic clade approach, or MCA). This resulted in a sequence variability within OTUs of 3–11.5%. For the 97% universal approach, we used a standard bioinformatics workflow developed using the open source Quantitative Insights into Microbial Ecology software package (QIIME, v.1.6.0, Caporaso et al., 2010) with a 97% universal OTU threshold (i.e. a sequence variability within OTU of 3%) and the UCLUST clustering algorithm (hereafter referred to as the 97% approach). Both approaches are outlined in more detail in the Supporting Information Methods S1. Using the same data, quality filtering (QC) and removal of low abundance OTUs, we applied both methods to three published NGS (454-Titanium) AMF datasets of LSU rRNA gene sequences (all targeting the D2 region). The three datasets range in spatial scale from one Danish grassland (meters; Lekberg et al., 2011; hereafter referred to as ‘Site’; primers glo454-NDL22; 8912 sequences, post QC), 22 local plant communities in south-western Montana (tens of kilometers; Lekberg et al., 2013; hereafter termed ‘Local’; primers FLR3-FLR4; 302 527 sequences) and 19 sample sites located across regions in the western United States (hundreds of kilometers; Bunn et al., 2014); termed ‘Regional’; primers FLR3-FLR4; 28 711 sequences). The three datasets contained 37 independent variables (Table S1) that were used to assess if and how the two OTU delineation methods distributed AMF communities differently along spatial and environmental gradients. The correspondence between the two OTU delineation methods was assessed using four complimentary tests. Mantel tests were used to determine the correlation and significance of numerical variables for AMF community patterns, and Analysis of Similarities (ANOSIM) was used for categorical variables. BEST analysis constructed the optimal n-parameter model of all numerical variables at once, and Procrustes analyses were used to compare distributions of AMF communities in PCoA space (Caporaso et al., 2012). Distance-based Redundancy Analysis (db-RDA) with forward selection was used to select a set of metadata variables that significantly explained non-overlapping portions of the AMF community variance (see ‘ordistep’ in the vegan package for R). All analyses were run in QIIME (Caporaso et al., 2010) and are outlined in more detail in the Supporting Information.

We also applied a supervised learning approach (Breiman, 2001; Knights et al., 2011b) to tests if and how the choice of universal OTU threshold (90–99%) influenced the ability to assign an AMF community to the correct environmental category. This approach requires a strong categorical predictor and we therefore used the Local dataset in which AMF communities differed significantly among plant community types (Lekberg et al., 2013). Ninety percent of the sequence data was chosen at random to train the learning algorithm to recognize AMF assemblages associated with four plant community types (knapweed (Centaurea stoebe), cheatgrass (Bromus tectorum), spurge (Euphorbia esula) and mixed remnant native) and the remaining 10% of the data was classified based on the inferred function generated by the learning algorithm (Knights et al., 2011b). In this scenario, the error rate expected by chance was 75% given that there were four plant community types (i.e. three in four designations would be incorrect due to chance alone). To obtain robust estimates of generalization error, 10-fold cross-validation was employed (Knights et al., 2011b).

Compared with the MCA, the 97% approach drastically increased OTU numbers from 33 to 76 in the Site dataset, from 46 to 1083 in the Local dataset, and from 30 to 278 in the Regional dataset (Figs S1–S3). This increase may be due, in part, to sequencing errors (Dickie, 2010; Tedersoo et al., 2010) because the recommended denoising step was not conducted (Reeder & Knight, 2010). An increase in OTU numbers was expected with the 97% approach, however, because many OTUs were manually combined in our MCA method, especially within the Rhizophagus irregularis and Glomus microaggregatum groups that had a within-OTUs sequence variability of up to 11.5%. This is very similar to the 11.8% intraspecific variation in the LSU gene region of Rhizophagus intraradices (FL isolate) reported by Stockinger et al. ( 2009). Based on this, the 97% approach may resolve smaller genetic differences within species, as has been shown for all fungi using the Internal Transcribed Spacer (ITS) region (Blaalid et al., 2013). Indeed, the most abundant MCA OTU in the Site dataset (Rhizophagus P) clustered with R. irregularis (FR750199) in Krüger et al. (2012), whereas the 97% approach divided this monophyletic clade into three OTUs. Coincidentally, perhaps, the online database MaarjAM that targets the 18S gene region also separates this species into three distinct virtual taxa (Öpik et al., 2010). Another unknown abundant OTU in the Site dataset, Rhizophagus D, was split into five OTUs with the 97% approach.

The finer-scale division of OTUs in the 97% approach increased (pairwise t-test < 0.001) β-diversity (γ/α) by an average 148% compared with MCA in the Local dataset. However, the increase in richness with the 97% approach did not appear to separate local ecotypes because the amount of variation explained by the spatial distribution of samples was not higher with the 97% approach compared with MCA (Table S1). Also, in spite of the dramatic increase in OTU numbers with the 97% approach and a change in absolute OTU richness, differences in relative richness among vegetation types observed in Lekberg et al. (2013) remained largely the same. If highly variable groups were over-represented in particular vegetation types, this correspondence may not have been observed.

OTU numbers aside, the distribution of AMF communities in ordination space, and their responses to environmental and spatial variables were very similar between the two OTU delineation approaches. More specifically, Mantel-r values, which assess the relationship between individual spatial and environmental variables and AMF community compositions, were highly correlated between the MCA and the 97% approach (Fig. 1). That is, important variables in the MCA (indicated by high Mantel-r and low P-values) were also important in the 97% approach (Table S1). The analyses of categorical variables were also in agreement, and the BEST analyses identified the same (Site and Local) or similar (Regional) predictor variables (Table S2). Both methods clustered the AMF communities according to vegetation type in the Local dataset (Fig. 2), and significant (< 0.01) Procrustes analyses (Caporaso et al., 2012) were observed for all three studies (Table S3, Fig. S4).

Figure 1.

Correlation of Mantel-r values obtained from statistical analyses with explanatory variables and arbuscular mycorrhizal fungal (AMF) communities generated from a monophyletic clade approach (MCA) or a universal (97%) threshold in all the three datasets (Site, closed circles; Local, open circles; Regional, triangles). The Mantel-r values were highly correlated (= 0.96, < 0.001) overall, as were separate correlations within datasets (Site: = 0.99, < 0.001; Local: = 0.96, < 0.001; Regional: = 0.97, < 0.001). The red circles indicate variables that explained a significant proportion of variation (< 0.1) in both operational taxonomic unit (OTU) delineation approaches using Distance-based Redundancy Analysis (db-RDA) and forward selection. The db-RDA eliminates potential issues with autocorrelations among explanatory variables and thus represents a very conservative, albeit still highly significant (= 0.93, < 0.001), measure of the relationship between the two OTU delineation approaches.

Figure 2.

Procrustes analysis of arbuscular mycorrhizal fungal (AMF) communities among cheatgrass (red), native (orange), knapweed (blue) and spurge (green) communities for the Local dataset. Duplicate points (connected by black lines) represent data processed through the monophyletic clade approach (MCA) and 97% (circled) operational taxonomic unit (OTU) picking pipelines. The lines between points highlight the Euclidean distance between points that represent the two different methods. Procrustes plots for the Site and Regional datasets are in Supporting Information Fig. S4.

Is this correspondence to be expected given that the MCA often lumps OTUs generated by the 97% approach? Not necessarily, because one may have predicted that either (1) combining sequence types, and possibly ecotypes, in the MCA would have obscured environmental signals resulting in inflated P-values, or, on the contrary, (2) splitting sequence types within lineages, and possibly diluting ecologically relevant sequence types with the 97% approach would have reduced the power to identify significant environmental variables. We found neither and thus conclude that AMF community patterns are robust across different OTU delineation methods (at least within the datasets and DNA target region tested here). We found support for this in the supervised learning, because even though OTU numbers increased exponentially with increasing similarity cut-offs, and ranged from 146 (90% threshold) to 2077 (99% threshold), the error rate was similar at 90 and 99% thresholds (Fig. S5). That is, the computer was equally good at assigning an AMF community to the correct aboveground plant community using AMF communities that had been delineated using 90% and 99% thresholds, which suggests that differences in AMF community patterns among plant community types in this dataset were deeply phylogenetically rooted. This was further supported by almost identical clustering of AMF communities among plant communities using the whole dataset and universal thresholds of 90, 97 and 98% (ANOSIM = 0.60–0.62, < 0.001).

Overall, our results agree remarkably well with findings by Powell et al. (2011), which showed that an equivalent amount of variation was accounted for using a universal OTU delineation approach (97%) and one based on evolutionary processes (general mixed Yule-coalescent or GMYC model). Their comparison used Sanger sequencing datasets that targeted the more conservative 18S gene region (Stockinger et al., 2010), which suggests that our findings may extend beyond the three datasets included here. Powell et al. (2011) argued for the use of the GMYC model because the same amount of variation was explained by fewer OTUs. The known difference in lineage age and divergence among known AMF species also favors an evolutionary approach because OTUs can be more easily aligned with known species (Krüger et al., 2012) or virtual taxa (Öpik et al., 2010). Niche-based methods using variable thresholds are being developed for 16S gene sequence data (Koeppel & Wu, 2013), but they are currently only computationally feasible for smaller-sized datasets (< 10 000 sequences). Until these approaches become readily available, our results indicate that even though comparisons of richness are difficult, AMF community patterns generated using universal thresholds do not differ from those grounded in evolutionary theory. While a more unified approach in AMF community ecology is desirable, cross-validations against expert-curated databases (Öpik et al., 2010) will allow for comparisons among studies regardless of OTU delineation approach, and will forward our understanding of AMF ecology and biogeography.


The authors thank Roger Koide, Philip Ramsey, Beau Larkin, Dan Mummey and four anonymous reviewers for their constructive comments on the manuscript, Rebecca Bunn for providing comments and allowing the authors to use the Regional dataset, Yvette Ortega for inputs on statistics and MPG Operations for funding. S.M.G. is supported, in part, by an EPA STAR Graduate Fellowship.