Original Article
JOURNAL OF COMPUTATIONAL BIOLOGY Volume 22, Number 00, 2014 # Mary Ann Liebert, Inc. Pp. 1–9 DOI: 10.1089/cmb.2014.0268
Inconsistent Denoising and Clustering Algorithms for Amplicon Sequence Data ¨ RKROTH,2 and JENNI HULTMAN 2 KAISA KOSKINEN,1 PETRI AUVINEN,1 K. JOHANNA BJO
ABSTRACT Natural microbial communities have been studied for decades using the 16S rRNA gene as a marker. In recent years, the application of second-generation sequencing technologies has revolutionized our understanding of the structure and function of microbial communities in complex environments. Using these highly parallel techniques, a detailed description of community characteristics are constructed, and even the rare biosphere can be detected. The new approaches carry numerous advantages and lack many features that skewed the results using traditional techniques, but we are still facing serious bias, and the lack of reliable comparability of produced results. Here, we contrasted publicly available amplicon sequence data analysis algorithms by using two different data sets, one with defined clone-based structure, and one with food spoilage community with well-studied communities. We aimed to assess which software and parameters produce results that resemble the benchmark community best, how large differences can be detected between methods, and whether these differences are statistically significant. The results suggest that commonly accepted denoising and clustering methods used in different combinations produce significantly different outcome: clustering method impacts greatly on the number of operational taxonomic units (OTUs) and denoising algorithm influences more on taxonomic affiliations. The magnitude of the OTU number difference was up to 40-fold and the disparity between results seemed highly dependent on the community structure and diversity. Statistically significant differences in taxonomies between methods were seen even at phylum level. However, the application of effective denoising method seemed to even out the differences produced by clustering. Key words: amplicon sequencing, clustering, denoising, operational taxonomic unit, taxonomy.
1. BACKGROUND
N
ext-generation sequencing technologies (NGS) have become an essential tool in studying diverse microbial communities. Especially platform 454 has been widely applied in microbial ecology after the machinery became easily available and the cost per sequence decreased. Typically, one to three variable regions of the 16S rRNA gene are amplified, and the resulting fragments sequenced. Analyzing these massive datasets has become the next challenge.
1 Institute of Biotechnology, and 2Department of Food Hygiene and Environmental Health, Faculty of Veterinary Medicine, University of Helsinki, Helsinki, Finland.
1
2
KOSKINEN ET AL.
There is a variety of tools available for analyzing 16S rRNA gene sequence data. The main objectives are removal of artifacts and low-quality data (homopolymer errors, single-base insertions, and quality problems at the end of sequence reads) (Huse et al., 2007), clustering reads into operational taxonomic units (OTUs), and assigning taxonomy. The data analysis typically starts with denoising, which is an important quality control stage and has a strong effect on downstream analyses (Schloss et al., 2011). In this first step, the raw data are corrected by clustering the flowgrams using frequency-based heuristics or approximate likelihood with empirically derived error distributions, or fasta-formatted files are corrected based on alignments (Reeder and Knight, 2010; Quince et al., 2011; Bragg et al., 2012). Clustering sequences into OTUs can also have a strong impact on the number of observed and estimated community richness. Whether two sequences are clustered together into the same OTU depends both on their similarity and other sequences in the studied dataset, as well as used clustering algorithm (Schloss et al., 2011). There are four classic hierarchical clustering algorithms: the nearest (single-linkage), furthest (complete linkage), weighted and average (Unweighted Pair Group Method with Arithmetic Mean), and neighbor algorithms (Legendre and Legendre, 1998). These algorithms were written long before the era of NGS, and are therefore computationally demanding. Hence, to optimize the process, heuristic algorithms have been developed to cluster sequence reads into OTUs, allowing analysis of large datasets without access to highly powerful computer resources. The clustering rate is typically much faster and maximum memory requirement considerably lower, but the heuristic algorithms do not work as accurately as the best of the classic hierarchical clustering tools (Schloss et al., 2011). Additionally, different algorithms can produce surprisingly divergent results, particularly in OTU numbers and diversity estimates. In this study, we compared publicly available sequence denoising and clustering tools to estimate which algorithms and parameters work most reliably.
2. METHODS Two published datasets were used in analyses: (1) Quince Titanium dataset (Quince et al., 2011) and (2) modified atmosphere packaged poultry microbiome dataset (Nieminen et al., 2012). We utilized both quality-filtered and raw sequence data of the latter. Quince Titanium dataset is an artificial soil microbe community, made up by amplifying and sequencing selected clones in known proportions. The modified atmosphere packaged poultry microbiome is an amplicon sequence dataset, dominated by lactic acid bacteria, which has been thoroughly characterized by traditional microbiological methods and next-generation sequencing (Nieminen et al., 2012). The samples consist of DNA isolated and analyzed from marinated and natural poultry products. Quince Titanium dataset was used as a benchmark, because its community structure and diversity are fully known, and modified atmosphere packaged poultry microbiome was used to test how consistently the algorithms work with datasets with varying diversity profiles and community compositions. The studied datasets were subjected to several analysis pipelines. We used DeNoiser (Reeder and Knight, 2010), mothur shhh.flows (Schloss et al., 2011), and Acacia (Bragg et al., 2012) for denoising the sequence reads. Figure 1 summarizes the tested datasets and workflows. DeNoiser flowgram clustering process (Reeder and Knight, 2010) and subsequent quality trimming step were run in QIIME package version 1.8.0 (Caporaso et al., 2010) with default settings. Following quality trimming parameters were applied: minimum length 200 bp, no ambiguous nucleotides, 6 nucleotides as the maximum length for homopolymers, no difference to the barcode sequence, and no difference to the primer sequence. The mothur shhh.flows flowgram clustering algorithm was applied with default settings following mothur 454 SOP (Schloss et al., 2011) protocol (May 11, 2014), trimming the data to first 450 flows. The data were also trimmed based on quality scores. The following parameters were used: no ambiguous nucleotides, 8 nucleotides as the maximum length for homopolymers, no difference to the barcode sequence, maximum of 2 nucleotides difference to the primer sequence, minimum average quality score 30 in a window of 50 nucleotides, and minimum length of 200 or 250 nucleotides. Moreover, we analyzed the datasets without quality trimming to compare the effects of denoising, quality trimming, and lack of both. Then, we only removed the sequences shorter than 200 or 250 nucleotides, containing longer homopolymers than 8 nucleotides or differing more than 2 nucleotides from the primer sequence. For barcode sequences, we did not accept any differences in the sequence.
INCONSISTENT DENOISING AND CLUSTERING ALGORITHMS
3
FIG. 1. Flowchart representing the materials and methods of the study. The two datasets used (poultry and Quince Titanium) are presented in blue, and tools utilized in green.
Acacia denoising, which produces alignment-based error-corrected fasta files (Bragg et al., 2012), was conducted to the fastq-files with Acacia version 1.52.b0 and default settings. The sequence reads were clustered to OTUs using mothur average neighbor algorithm (Schloss et al., 2009), standalone programs ESPRITTree version 11152011 (Cai and Sun, 2011) and BEBaC version 12.06.2012 (Cheng et al., 2012), as well as UCLUST version v1.2.22 (Edgar, 2010) and CD-HIT version 4.5.4 (Niu et al., 2010) integrated into QIIME package (Caporaso et al., 2010). These clustering tools were applied with denoising and quality trimming in various combinations (Fig. 1). Clustering sequences to OTUs using mothur average neighbor algorithm followed the mothur 454 SOP protocol (Schloss et al., 2011) (May 11, 2014). The sequences were aligned to SILVA reference alignment (Pruesse et al., 2007), screened, filtered, and preclustered. The distances between sequences were calculated, and the OTUs were defined based on the average distances. A representative sequence from each OTU at distance level 0.03 was picked and subjected to taxonomic assignment. ESPRITTree (Cai and Sun, 2011) clustering was conducted following the user guide. The sequence reads were clustered to OTUs with default settings and the representative sequences at distance level 0.03 were picked for taxonomic identification. BEBaC (Cheng et al., 2012) clustering was applied as recommended in the user manual. No denoising or quality score filtering was conducted, but the sequences were trimmed for minimum length of 410 bp, and sequences including homopolymers longer than 8 nucleotides and sequences differing more than 2 nucleotides from the primer sequence and at all from the barcodes were excluded from the analysis. The remaining sequences were chopped to maximum length of 510 to minimize the effect of sequence length differences on clustering. First, the initial clusters were created, then a distance matrix was calculated, and finally crude clusters were formed. BEBaC created consensus sequences for every crude cluster (OTU), and the consensus sequences were subjected to taxonomic classification. Using BEBaC, all the sequence data of a project can be analyzed as a whole or sample specifically, if the sequences are sorted to samples prior to BEBaC analysis. In this study, we employed both approaches with quality-filtered poultry product microbiome to test whether these methods give similar or dissimilar results. CD-HIT (Niu et al., 2010) and UCLUST (Edgar, 2010) clustering were conducted in QIIME package (Caporaso et al., 2010) using default settings. The OTUs were picked with 97% sequence similarity, and the first sequence in each OTU was selected as a representative sequence for taxonomic assignment. To compare the community membership and structure produced by applied algorithms, we subjected the representative sequences from each OTU from all treatments to beta diversity measurements. The representative sequences were aligned, distances between sequences calculated, and, finally, the sequences were clustered. We generated Venn diagrams to represent the overlap between communities. These beta diversity measurements were conducted using mothur software package (Schloss et al., 2009). Taxonomic affiliations were defined using BLAST algorithm (Altschul et al., 1997) and Greengenes taxonomy (February 4, 2011, version11) in QIIME package (Caporaso et al., 2010). E-value 0.0001 was selected for a threshold. Taxonomies were defined for compared datasets and each data analysis pipeline tested, and OTU tables containing taxonomic affiliations and the number of sequences in each OTU
4
KOSKINEN ET AL.
constructed. Statistically significant differences between relative abundance of bacterial taxa created by tested analysis methods were detected using STAMP (Parks and Beiko, 2010) ‘‘two samples analysis’’ with the following parameters: Fisher’s exact test for statistical hypothesis, Newcombe–Wilson for difference between proportions and confidence interval methods, and Storey’s FDR for multiple test correction method. We calculated the proportion of taxa that differed statistically significantly between analysis pipelines and drew distance matrices illustrating the magnitude of differences. We examined the results at phylum, class, order, family, and genus levels, and calculated the average value that represented all levels.
3. RESULTS AND DISCUSSION Denoising amplicon sequence data has a major influence on the error rate and the number of usable reads in sequence dataset, and thereby the number of observed OTUs in downstream analyses (Kunin et al., 2010; Quince et al., 2011; Schloss et al., 2011). However, the differences in OTU numbers and taxonomies caused by the publicly available algorithms are less reported. In this study, we compared the effect of denoising and clustering algorithms on the number of observed OTUs and the abundance of identified taxa in studied datasets (Fig. 2a). Our study suggests that the number of OTUs is highly dependent on selected denoising, quality trimming, and clustering algorithms. The results indicate that the choice of denoising method influences the number of OTUs multifold. Of the used methods, Acacia produced two to over five times more OTUs, when mothur shhh.flows and DeNoiser generated more similar results. There were differences and inconsistencies between datasets, suggesting that data quality and community structure may affect the magnitude of observed differences. According to Bragg et al. (2012) Acacia’s error correction should be comparable to AmpliconNoise (Quince et al., 2011), the new version of PyroNoise (Quince et al., 2009), but our results do not support their claim: mothur shhh.flows (translated from PyroNoise algorithm) and Acacia produced highly different numbers of OTUs with all tested datasets. The largest OTU numbers were achieved, as expected, when the datasets were not denoised or trimmed based on quality scores. The magnitude of difference in the numbers of OTUs between denoised and untrimmed data depended on the analyzed datasets: the effect was about 4 times with Quince Titanium dataset, but up to 10 times with poultry product microbiome (Fig. 2a). Additionally, the settings of denoising algorithms have also been shown to affect the results (Gaspar and Thomas, 2013); thus, we selected to use default parameters. We compared five different OTU clustering approaches (Fig. 1). The number of OTUs detected in samples with different OTU clustering was found particularly differing: 43–2162 ( – 617.5) and 47–2158 ( – 651.3) in poultry dataset and 72–794 ( – 285.5) in Quince Titanium data. The selected denoising algorithms markedly affected the number of observed OTUs, but the real distinction between OTU numbers was caused by clustering: the choice of clustering algorithm inflected the results by more than one order of magnitude. Alike varying diversity has also been observed in mock communities (May et al., 2014) although the differences were not that striking. The most dramatic differences between clustering methods were observed when modified atmosphere packaged poultry data were not denoised or quality trimmed but only clustered. The difference in OTU numbers between ESPRITTree and BEBaC clustering was over 40fold. The smallest number of OTUs was produced using BEBaC clustering. The Quince dataset constructed from 88 clones produced only 72 OTUs, indicating that BEBaC clustering did not register all the diversity in the dataset. Additionally, with poultry product microbiome samples, the difference between clustering the sequences of two samples together and separately with BEBaC returned almost twofold difference in the OTU numbers.
‰ FIG. 2. (a) Bar chart depicting the number of operational taxonomic units (OTUs) produced with applied methods. The number of OTUs was highly dependent on used combination of denoising and clustering algorithms: the difference between OTU numbers generated using Acacia, Denoiser, and shhh.flows denoising algorithms was several fold, while the choice of clustering method inflected the OTU numbers by more than one order of magnitude. ‘‘Quince clones’’ represent the 88 clones in the original clone data, and ‘‘Modified atmosphere packaged poultry’’ represents the number of OTUs when the data were first published (Nieminen et al., 2012). (b) Bar chart describing the percentage of identified taxa and corrected errors induced by amplification and sequencing compared to the original Quince clone data. 100% = all taxa present in clone library were identified, and all amplification and sequencing errors were corrected during the analysis.
5
6
KOSKINEN ET AL.
UCLUST and CD-HIT produced very similar numbers of OTUs but considerably more than using mothur average neighbor. Similarly to denoising algorithms, also clustering results were greatly dependent on the analyzed dataset and efficient denoising with shhh.flows or DeNoiser prior to clustering made the results more uniform. In addition to denoising and clustering, the selected minimum sequence length affected the OTU numbers by about twofold with poultry product microbiome when the data were not denoised but trimmed based on quality scores. This effect was smaller with Quince Titanium data. Commonly, when a diverse microbial community is studied by amplicon sequencing, there is no accurate a priori information on the diversity of the studied samples and the results gained using different methods may seem equally credible. Here, Quince Titanium dataset served as a benchmark with fully known community structure and diversity. The results show that none of the applied combinations of denoising/ trimming and clustering methods gave exactly the accurate number of OTUs and community composition. BEBaC clustering produced the closest estimation of observed OTUs, 72 in total, and was the only method underestimating the diversity in this bacterial community. The other applied methods overestimated the diversity by 1.5–9 times (Fig. 2a). DeNoiser and mothur shhh.flows denoising algorithms applied with CDHIT, UCLUST, ESPRITTree, and mothur clustering produced somewhat similar OTU numbers, but Acacia denoising inflated the diversity (Fig. 2a). Acacia results were comparable to those when no quality trimming was applied. Additionally, when the sequence data were not denoised, but trimmed based on quality scores and length in QIIME, or ignoring the quality trimming, the number of observed OTUs was rather high (Fig. 2a). Taxonomic affiliations were assigned by querying the representative sequence from each OTU against Greengenes database. The results solidly show that the taxonomic affiliations were also affected by the choice of data analysis methods. The abundant phylotypes were detected by all methods but the relative abundances varied. However, the presence and absence of rare taxa varied more between methods. With Quince Titanium dataset, the performance of the analysis pipelines was assessed from various points of views: how well the results resembled the original clone library community, how many bacterial groups at different taxonomic levels were wrongly filtered out during the analysis, and how many spurious taxa (not present in the original clone or amplified library) were created during the analysis. Creating the amplicon library was the major source of error: about half of the bacterial groups identified in raw amplicon sequence data were false, from 13% at phylum level to almost 85% at species level (Supplementary Material file 1, available online at www.liebertpub.com/cmb). Majority of these false taxa were removed during the analysis but the performance of the selected analysis tools varied: at this respect the best tools (shhh.flows/ESPRITTree, shhh.flows/mothur, shhh.flows/ mothur length ‡ 250, and BEBaC) corrected 99% of extra taxa, and the poorest (Acacia/UCLUST) only 89% (Fig. 2b). Consequently, taxa absent from the original clone data were identified from Titanium dataset by several of the selected analysis methods. At the same time, filtering out the real diversity was most efficiently achieved by the analysis pipelines that produced low number of OTUs (shhh.flows/ ESPRITTree, DeNoiser/mothur, q ‡ 30/mothur, q ‡ 30/mothur length ‡ 250, and BEBaC) (Fig. 2b). However, analysis using DeNoiser/CD-HIT, DeNoiser/Uclust, or DeNoiser/ESPRITTree also produced small numbers of OTUs but less of the real diversity was filtered out. In addition to the inability to distinguish between real diversity and amplification and sequencing errors, Acacia denoising algorithm introduced a spurious taxon that was present neither in the original clone data nor in the amplicon library. This phenomenon was not detected when observed phyla were compared, but at class level, one spurious bacterial class, Epsilonproteobacteria, was identified by Acacia/ESPRITTree, Acacia/mother, and Acacia/UCLUST algorithms. The same spurious taxon was detected at order (Campylobacterales), family (Helicobacteraceae), and genus (Sulfurimonas) levels. At family level also
‰ FIG. 3. Distance matrices illustrating the share of identified bacterial taxa differing statistically significantly between applied data analysis methods. The figure shows the average difference of five taxonomic levels: phylum, class, order, family, and genus. The separate figures of these taxonomic levels are presented in supplementary information (Supplementary Data Files 2–4). (a) Quince Titanium dataset, (b) modified atmosphere packaged poultry, natural, and (c) modified atmosphere packaged poultry, marinated. The number of statistically significantly ( p ‡ 0.05) differing taxa was derived from STAMP (Parks and Beiko, 2010) analysis and the numbers in distance matrices represent the proportion of taxa that differed between the two analysis methods in question when the number of differing taxa was compared to all identified taxa.
7
8
KOSKINEN ET AL.
Catabacteriaceae was detected by Acacia/ESPRITTree and Acacia/UCLUST. These results suggest that Acacia denoising algorithm introduces errors that may skew the community structure. Statistically significant differences between relative abundance of bacterial taxa created by analysis methods were analyzed using STAMP (Parks and Beiko, 2010). First, we compared the 88 clones with different analysis pipelines. Only three of the tested methods resulted in no statistical difference between the clone library and the analyzed Quince Titanium amplicon dataset: shhh.flows/ESPRITTree, shhh.flows/ mother, and shhh.flows/mothur length ‡ 250 (Fig. 3a). According to the analysis, the choice of denoising method affected the results most strongly even at phylum level. mothur shhh.flows and Acacia were the most divergent, as denoising using Acacia yielded similar results as when no quality trimming was applied. STAMP analysis of the modified atmosphere packaged poultry datasets revealed the evident effect of community structure and diversity on differences between denoising and clustering methods: the difference between Acacia and shhh.flows denoising was seen in more diverse, natural poultry sample but not in marinated, and the largest dissimilarity was detected between BEBaC clustering and all other methods (Fig. 3b,c). The share of identified taxa differing statistically significantly between the analysis pipelines at different taxonomic levels and datasets is presented in Supplementary Material files 2a–e, 3a–e, and 4a–e. Similar taxonomic affiliations did not ensure similar OTUs. Venn diagrams illustrating the differences between OTUs derived from different data analysis pipelines are presented in Supplementary Material files 2f, 3f, and 4f. The datasets analyzed in this study were different: the Quince Titanium dataset was dominated by three phyla, Proteobacteria, Actinobacteria, and Acidobacteria, and the distribution of taxa was more even compared to poultry microbiomes. The analyzed poultry microbiomes, natural and marinated, were quite simple communities, dominated by two bacterial phyla, Firmicutes and Proteobacteria. The natural community was, however, slightly more diverse and this small difference between the communities was apparent in taxonomic assignment results: the natural poultry product bacterial community was more affected by the analysis methods, even at phylum level when the marinated community manifested the differences clearly only at family and genus levels (Supplementary Material files 3 and 4).
4. CONCLUSIONS According to our study, the analysis methods substantially affect the amplicon sequencing results: clustering method has a great impact on the number of OTUs and denoising algorithm influences more on taxonomic affiliations. The disparity between results seems to be highly dependent on the community structure and diversity, and significant differences in taxonomies between methods are seen even at phylum level. Thus, testing these tools merely with mock communities may not result in representative results with real data, as the diversity and evenness of the natural microbial communities vary and the true performance of tested algorithms remains unclear. When comparing between different datasets, we recommend that the comparisons should be done using the same analysis pipeline in order to minimize the effects of denoising and clustering tools. The results also apply for Illumina sequencing with different sequencing-based error sources but similar challenges with preprocessing and clustering-based variance in OTU numbers. The results from this study can be utilized in selection of the best pipeline for microbial community analysis.
AUTHORS’ CONTRIBUTIONS K.K. and J.H. designed and performed the analyses and wrote the article with the help of K.J.B. and P.A. All authors read and approved the final article.
ACKNOWLEDGMENTS We thank members of Auvinen laboratory for valuable comments on the article. The financial support from the Finnish Funding Agency for Technology and Innovation projects (TEKES 40046/11) and Academy of Finland (Centre of Excellence in Microbial Food Safety Research) is gratefully acknowledged.
INCONSISTENT DENOISING AND CLUSTERING ALGORITHMS
9
AUTHOR DISCLOSURE STATEMENT No competing financial interests exist.
REFERENCES Altschul, S.F., Madden, T.L., Schaffer, A.A., et al. 1997. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res. 25, 3389–3402. Bragg, L., Stone, G., Imelfort, M., et al. 2012. Fast, accurate error-correction of amplicon pyrosequences using Acacia. Nat. Meth. 9, 425–426. Cai, Y., and Sun, Y. 2011. ESPRIT-Tree: hierarchical clustering analysis of millions of 16S rRNA pyrosequences in quasilinear computational time. Nucleic Acids Res. 39, e95. Caporaso, J.G., Kuczynski, J., Stombaugh, J., et al. 2010. QIIME allows analysis of high-throughput community sequencing data. Nat. Meth. 7, 335–336. Cheng, L., Walker, A.W., and Corander, J. 2012. Bayesian estimation of bacterial community composition from 454 sequencing data. Nucleic Acids Res. 40, 5240–5249. Edgar, R.C. 2010. Search and clustering orders of magnitude faster than BLAST. Bioinformatics 26, 2460–2461. Gaspar, J.M., and Thomas, W.K. 2013. Assessing the consequences of denoising marker-based metagenomic data. PLoS One 8, e60458. Huse, S.M., Huber, J.A., Morrison, H.G., et al. 2007. Accuracy and quality of massively parallel DNA pyrosequencing. Genome Biol. 8, R143. Kunin, V., Engelbrektson, A., Ochman, H., et al. 2010. Wrinkles in the rare biosphere: pyrosequencing errors can lead to artificial inflation of diversity estimates. Environ. Microbiol. 12, 118–123. Legendre, P., and Legendre, L. 1998. Numerical Ecology. Elsevier, Amsterdam. May, A., Abeln, S., Crielaard, W., et al. 2014. Unraveling the outcome of 16S rDNA-based taxonomy analysis through mock data and simulations. Bioinformatics 30, 1530–1538. Nieminen, T.T., Koskinen, K., Laine, P., et al. 2012. Comparison of microbial communities in marinated and unmarinated broiler meat by metagenomics. Int. J. Food Microbiol. 157, 142–149. Niu, B., Fu, L., Sun, S., et al. 2010. Artificial and natural duplicates in pyrosequencing reads of metagenomic data. BMC Bioinformatics 11, 187. Parks, D.H., and Beiko, R.G. 2010. Identifying biologically relevant differences between metagenomic communities. Bioinformatics 26, 715–721. Pruesse, E., Quast, C., Knittel, K., et al. 2007. SILVA: a comprehensive online resource for quality checked and aligned ribosomal RNA sequence data compatible with ARB. Nucleic Acids Res. 35, 7188–7196. Quince, C., Lanze´n, A., Curtis, T.P., et al. 2009. Accurate determination of microbial diversity from 454 pyrosequencing data. Nat. Meth. 6, 639–641. Quince, C., Lanzen, A., Davenport, R.J., et al. 2011. Removing noise from pyrosequenced amplicons. BMC Bioinformatics 12, 38. Reeder, J., and Knight, R. 2010. Rapidly denoising pyrosequencing amplicon reads by exploiting rank-abundance distributions. Nat. Meth. 7, 668–669. Schloss, P.D., Gevers, D., and Westcott, S.L., 2011. Reducing the effects of PCR amplification and sequencing artifacts on 16S rRNA-based studies. PLoS ONE 6, e27310. Schloss, P.D., Westcott, S.L., Ryabin, T., et al. 2009. Introducing mothur: open-source, platform-independent, community-supported software for describing and comparing microbial communities. Appl. Environ. Microbiol. 75, 7537–7541.
Address correspondence to: Dr. Kaisa Koskinen Institute of Biotechnology P.O. Box 56 (Viikinkaari 4) 00014 University of Helsinki Finland E-mail:
[email protected]
