|Home | About | Journals | Submit | Contact Us | Français|
Reference phylogenies are crucial for providing a taxonomic framework for interpretation of marker gene and metagenomic surveys, which continue to reveal novel species at a remarkable rate. Greengenes is a dedicated full-length 16S rRNA gene database that provides users with a curated taxonomy based on de novo tree inference. We developed a ‘taxonomy to tree' approach for transferring group names from an existing taxonomy to a tree topology, and used it to apply the Greengenes, National Center for Biotechnology Information (NCBI) and cyanoDB (Cyanobacteria only) taxonomies to a de novo tree comprising 408315 sequences. We also incorporated explicit rank information provided by the NCBI taxonomy to group names (by prefixing rank designations) for better user orientation and classification consistency. The resulting merged taxonomy improved the classification of 75% of the sequences by one or more ranks relative to the original NCBI taxonomy with the most pronounced improvements occurring in under-classified environmental sequences. We also assessed candidate phyla (divisions) currently defined by NCBI and present recommendations for consolidation of 34 redundantly named groups. All intermediate results from the pipeline, which includes tree inference, jackknifing and transfer of a donor taxonomy to a recipient tree (tax2tree) are available for download. The improved Greengenes taxonomy should provide important infrastructure for a wide range of megasequencing projects studying ecosystems on scales ranging from our own bodies (the Human Microbiome Project) to the entire planet (the Earth Microbiome Project). The implementation of the software can be obtained from http://sourceforge.net/projects/tax2tree/.
A robust universal reference taxonomy is a necessary aid to interpretation of high-throughput sequence data from microbial communities (Tringe and Hugenholtz, 2008). Taxonomy based on the 16S rRNA gene (16S) is the most comprehensive and widely used in microbiology today (Pruesse et al., 2007; Peplies et al., 2008), but has yet to reach its full potential because numerous microbes belong to taxa that have not yet been characterized and because numerous sequences that could be reliably classified remain unannotated. For example, two thirds of 16S sequences in GenBank are only classified to domain (kingdom), that is, Archaea or Bacteria, by the National Center for Biotechnology Information (NCBI) taxonomy: this taxonomy is likely the most widely consulted 16S-based taxonomy, despite its disclaimer that it is not an authoritative source, in part because classifications are maintained up to date through user submissions. Most of the un(der)classified sequences are from culture-independent environmental surveys; these sequences can swamp BLAST searches, leaving users baffled about the phylogenetic affiliation of their submitted sequences. This shortcoming has been addressed by several dedicated 16S databases, including the Ribosomal Database Project (Cole et al., 2009), Greengenes (DeSantis et al., 2006), SILVA (Pruesse et al., 2007) and EzTaxon (Chun et al., 2007), that classify a higher proportion of environmental sequences. However, improvements are still needed because many sequences remain unclassified and numerous classification conflicts exist between the different 16S databases (DeSantis et al., 2006). Moreover, the emergence of large-scale next-generation sequencing projects such as the Human Microbiome Project (Turnbaugh et al., 2007; Peterson et al., 2009) and TerraGenome (Vogel et al., 2009), and the availability of affordable sequencing to a wide range of users who have traditionally lacked access, mean that the need to integrate new sequences into a consistent universal taxonomic framework has never been greater.
The Greengenes taxonomy is currently based on a de novo phylogenetic tree of 408135 quality-filtered sequences calculated using FastTree (Price et al., 2010). De novo tree construction is among the most objective means for inferring sequence relationships, but requires either generation of new taxonomic classifications or transfer of existing taxonomic classifications between iterations of trees as the 16S database expands. Previously, we developed a tool that automatically assigns names to monophyletic groups in large phylogenetic trees (Dalevi et al., 2007), which is useful for naming novel (unclassified) clusters of environmental sequences. Here we describe a method to transfer group names from any existing taxonomy to any tree topology that has overlapping terminal node (tip) names. We used this ‘taxonomy to tree' approach to annotate the 408135 sequence tree with the NCBI taxonomy as downloaded in June 2010 (Sayers et al., 2011), supplemented with the Greengenes taxonomy from the previous iteration (Dalevi et al., 2007) and CyanoDB (http://www.cyanodb.cz). Explicit rank information, prefixed to group names, was incorporated into the Greengenes taxonomy to help users orient themselves and to improve the consistency of the classification. We assessed the consistency of the resulting classification with the NCBI taxonomy including currently defined candidate phyla (divisions), and present recommendations for consolidation of 34 redundantly named groups and exclusion of one on the basis that its sole representative is chimeric.
We obtained 16S sequences from the Greengenes database, which extracts these sequences from public databases using quality filters as described previously (DeSantis et al., 2006). We only used sequences that had <1% non-ACGT characters. The sequences were checked for chimeras using UCHIME (http://www.drive5.com/uchime/) and ChimeraSlayer (Haas et al., 2011). We only removed sequences from named isolates if they were classified as chimeric by both tools; we removed other sequences if they were classified as chimeric by either tool or if they were unique to one study, meaning that no similar sequence (within 3% in a preliminary tree) was reported by another study. Quality filtered 16S sequences were aligned based on both primary sequence and secondary structure to archaeal and bacterial covariance models (ssu-align-0.1) using Infernal (Nawrocki et al., 2009) with the sub option to avoid alignment errors near the ends. The models were built from structure-annotated training alignments derived from the Comparative RNA Website (Cannone et al., 2002) as described in detail previously (Nawrocki et al., 2009). The resulting alignments were adjusted to fit the fixed 7682 character Greengenes alignment through identification of corresponding positions between the model training alignments and the Greengenes alignment. Hypervariable regions were filtered using a modified version of the Lane mask (Lane, 1991). A tree of the remaining 408135 filtered sequences, (tree_16S_all_gg_2011_1) was built using FastTree v2.1.1, a fast and accurate approximately maximum-likelihood method using the CAT approximation and branch lengths were rescaled using a gamma model (Price et al., 2010). Statistical support for taxon groupings in this tree was conservatively approximated using taxon jackknifing, in which a fraction (0.1%) of the sequences (rather than alignment positions) is excluded at random and the tree reconstructed. We use these support values to help guide selection of monophyletic interior nodes for group naming during manual curation.
For evaluation of NCBI-defined candidate phyla, we added 765 mostly partial length sequences, that failed the Greengenes filtering procedure but were required for the evaluation, to the alignment using PyNAST (Caporaso et al., 2010; based on the 29 November, 2010 Greengenes OTU templates) and generated a second FastTree (tree_16S_candiv_gg_2011_1) using the parameters described above.
Having constructed de novo trees, the next key step was to link the internal tree nodes to known, named taxa. We used the NCBI taxonomy (Sayers et al., 2011) as the primary taxonomic source to annotate the trees, supplemented by the previous iteration of the Greengenes taxonomy (Dalevi et al., 2007) and cyanoDB (http://www.cyanodb.cz). This taxonomic annotation used a new algorithm called tax2tree. Briefly, tax2tree consists of the following steps:
Secondary taxonomies were then applied manually to the annotated recipient tree, tree_16S_candiv_gg_2011_1, in ARB (Ludwig et al., 2004) using the group tool. For the Greengenes taxonomy, this was achieved by displaying the Greengenes taxonomy field at the tips of the tree and assigning group names missed in the automated taxonomy transfer (mostly higher level ranks associated with candidate phyla). For the cyanoDB taxonomy, manual assignments were based on type species (http://www.cyanodb.cz/valid_genera; 405 instances in tree_16S_candiv_gg_2011_1; Supplementary Table S1). The manually supplemented taxonomic assignments were then exported from the curated tree as a flat file (Supplementary Figure S1) using functionality in the tax2tree pipeline and reapplied to tree_16S_all_gg_2011_1 ensuring manual updates were efficiently propagated to both trees. The tax2tree software is implemented in the Python programming language using the PyCogent toolkit (Knight et al., 2007), and is available under the open-source General Public license at http://sourceforge.net/projects/tax2tree/.
The NCBI and Greengenes taxonomies were compared for each of the 408135 sequences in tree_16S_all_gg_2011_1 making use of the explicit rank designations. The lowest classified rank for each sequence was determined and compared (Figure 2a) to estimate overall improvements in classification. Note that only contiguous classifications were used in this estimate, that is, all ranks leading to the lowest named rank also had to have names. Taxonomic similarities and differences for each sequence at each rank were also assessed by dividing sequences into five categories, (i) the two taxonomies had equal values (same name) at a given taxonomic rank, (ii) the two taxonomies had unequal values at a given rank, (iii) and (iv) one of the taxonomies lacked a value at a given rank and (v) both taxonomies lacked a value at a given rank. This provided an indication of the type of changes that had occurred between the NCBI and Greengenes taxonomies (Figure 2b).
The rapid accumulation of sequence data from across the tree of life is a boon for molecular taxonomy, but also presents a major barrier to sequence-based taxonomy curators as it is essentially impossible to manually curate trees comprising hundreds of thousands of sequences from scratch. We overcame this difficulty by developing an automated procedure based on F-measures (van Rijsbergen, 1979) for transferring any (donor) taxonomy (in a standard flat text format, Supplementary Figure S1) to any unannotated (recipient) tree (in Newick format) given common sequence identifiers. The F-measure is most often used to measure the classification performance (precision and recall) of information retrieval processes such as database searches (van Rijsbergen, 1979). This approach also has the potential to provide an assessment of the quality of fit between a taxonomy and a tree, which could be used to screen multiple taxonomies and/or trees before manual curation efforts.
Using quality-filtered sequences from Greengenes (DeSantis et al., 2006) aligned with the secondary structure-aware infernal package (Nawrocki et al., 2009), we constructed a phylogenetic tree using FastTree2 (Price et al., 2010) containing 408135 sequences (tree_16S_all_gg_2011_1). We inferred confidence estimates using taxon jackknife resampling (in which sequences, rather than positions in the alignment, are resampled) as it provides a conservative guide to group monophyly, which we found greatly assists manual group name curation between tree iterations (see below). We then applied NCBI classifications to this topology using the tax2tree algorithm (Figure 1 and methods) also taking advantage of the explicit rank designations provided by NCBI to include rank prefixes to group names (for example, p(hylum)__, c(lass)__, o(rder)__). Most sequences (69% 280488 of 408135) had uninformative NCBI classifications, that is, no rank information below domain (kingdom; Figure 2a); of these, most were environmental clones designated as ‘unclassified Bacteria'. The remaining 127647 sequences with informative classifications were then applied to the tree. This ‘unamended' approach alone resulted in an improved classification, to at least phylum-level, of nearly all of the taxonomically uninformative sequences (280452 of 280488) because most belong to known phyla but were simply deposited without classifications.
We then overlaid additional taxonomic information onto the NCBI-annotated tree by manual group name curation in ARB. This information consisted primarily of candidate phyla and other rank designations for environmental clusters imported from previous iterations of Greengenes (either assigned manually or by GRUNT (Dalevi et al., 2007)), and taxonomic information for the Cyanobacteria obtained from cyanoDB (http://www.cyanodb.cz/). This resulted in more informative classifications for 75% of sequences by at least one rank up to six ranks. These increases in classification depth are graphically shown in Figure 2a by rank.
Changes in sequence classifications between NCBI and Greengenes at each rank are summarized in Figure 2b. Most changes were from uninformative (domain/kingdom name only) in NCBI to informative in Greengenes again reflecting the classification of the large fraction of unclassified environmental sequence in NCBI. The percentage of changes to informative NCBI classifications were relatively low (<7% for all ranks), indicating the degree of congruence between NCBI and Greengenes classifications achieved in part by accommodating polyphyletic groups (see below). Of these type of changes, many occurred at the higher ranks particularly in the candidate phyla where Greengenes is manually curated most intensively (see below). A similar comparison with SILVA or RDP was not possible because of a lack of explicit ranks in these taxonomies. However, by comparing the group name immediately following the domain (kingdom) name (either Bacteria or Archaea) in the SILVA taxonomy, we estimated that only 28% of the 408135 sequences lacked phylum-level classifications in SILVA, as opposed to 68% in NCBI.
Further, the updated Greengenes taxonomy performed well in a test of reference taxonomies using a naïve Bayesian classifier (Werner et al., 2011). In this paper, we found that retraining the RDP classifier (Wang et al., 2007) using taxa from the new Greengenes taxonomy resulted in increased classification resolution relative to SILVA or RDP for a range of environments (human body habitats, snake and mouse gut and soils).
In principle, every taxon should correspond to a single monophyletic group in the 16S rRNA-based taxonomy, but practical considerations make relaxing this constraint very useful. In our approach, the back-filling step in the tax2tree (see methods) allows multiple groups to be given the same rank name. This feature is important for taxonomic groups that are well-established in the literature but polyphyletic in the reference 16S rRNA tree. A prominent example is the class Deltaproteobacteria, which rarely forms a monophyletic group in large 16S rRNA topologies and comprises five groups in the current tree_16S_all_gg_2011_1. This result may indicate that the Deltaproteobacteria do not form an evolutionary coherent grouping and will need to subdivided and reclassified. Alternatively, the Deltaproteobacteria may be a monophyletic group not resolved in 16S rRNA trees due to tree inference artifacts, chimeric sequences and/or to limits in the phylogenetic resolution of trees constructed from the 16S rRNA molecule alone. This issue can best be addressed using ‘whole genome' tree approaches that have greater phylogenetic resolution than single-gene topologies. Trees based on a concatenation of 31 conserved near ubiquitous single-copy gene families indicate that the small subset of Deltaproteobacteria with genome sequences are monophyletic (Ciccarelli et al., 2006; Wu et al., 2009). A second example, also based on concatenated conserved marker genes, indicate that the first genome sequence representative of the family Halanaerobiales (Halothermothrix orenii) is a member of the Firmicutes phylum (Mavromatis et al., 2009); which is its (contested) classification based on 16S rRNA trees (Ludwig and Klenk, 2001). Indeed, the Halanaerobiales is separate from the Firmicutes in the current Greengenes topology and is only classified as such because of the back-filling procedure.
Similarly, a number of phylum-level associations have been suggested based on concatenated gene topologies including a relationship between the class Deltaproteobacteria and phylum Acidobacteria (Ciccarelli et al., 2006). We predict that at least a subset of the currently defined candidate phyla will coalesce with other phyla once they are adequately represented by genome sequences and whole-genome trees can be constructed. Therefore, current estimates of the number of 16S-based candidate phyla (>50) should only be used as an approximation and may well drop as candidatus genome sequences accrue. However, regardless of absolute number of phyla, there is a strong need for consistent delineation of taxonomic groups between public databases, particularly candidate phyla. Thus, by retaining polyphyletic groups as sets of monophyletic taxa with the same name, we can accommodate the uncertainty in our present knowledge about both the tree and taxonomy and easily propagate whole-genome-based classification improvements in subsequent iterations of de novo 16S rRNA-based trees.
At the time of writing, the NCBI taxonomy lists 71 candidate phyla (divisions) of which 30 are represented only by partial (<1200nt) sequences. Therefore, in order to address the classification of these groups we amended tree_16S_all_gg_2011_1 with 765 mostly partial length sequences from GenBank and generated a new de novo tree using FastTree; tree_16S_candiv_gg_2011_1. We found some NCBI candidate phyla to be polyphyletic in tree_16S_candiv_gg_2011_1 because of a small number of submitter misclassifications or chimeric artifacts. In these instances, we reconciled NCBI and Greengenes designations using the majority classification for a given NCBI group. On the basis of tree_16S_candiv_gg_2011_1, we resolved the 71 candidate phyla into 45 monophyletic groups and one chimeric artifact (Table 1). Many proposed phyla appear to belong to well-established lineages including the Proteobacteria, Firmicutes, Bacteroidetes, Chloroflexi and Spirochaetes. In cases where two or more NCBI candidate phyla were combined and did not cluster with more established groups, we gave priority to either the oldest and/or the largest group. For example, we reclassified candidate phylum kpj58rc (Kelly and Chistoserdov, 2001) as OP3 (Hugenholtz et al., 1998) because of the priority of OP3 in the literature and larger number of representative OP3 sequences in the public databases. We also compared our classifications to SILVA and RDP and in many instances saw consistencies. For example, Greengenes and SILVA both classify candidate phylum CAB-I as Cyanobacteria and Greengenes and RDP both classify KSA1 in the Bacteroidetes. Conversely, in some instances we saw disagreements, such as candidate phylum GN02 (Ley et al., 2006) being classified as BD1-5, and WS5 (Dojka et al., 1998) as WCHB1-60 by SILVA (Table 1). This points to the need to consolidate classifications and also to give priority to published group names where possible.
The new NCBI-reconciled Greengenes taxonomy rescues over 200000 environmental sequences from unclassified oblivion. Moreover, the tax2tree pipeline will assist in reconciling information among the various 16S rRNA resources (Greengenes, SILVA, RDP and EZ-Taxon) with phylogenetic trees built using different methods, and will, we hope, make it easier for users to reconcile taxonomic classifications of large data sets obtained using different taxonomic schemes. This is especially important because which taxonomy is used can have a larger effect on the results than which assignment method is used (Liu et al., 2008). The new Greengenes taxonomy, along with all intermediate data products including the tree, can be downloaded from the Greengenes web site at http://greengenes.lbl.gov/.
This work, by automating the process of improving the tree and allowing import of taxonomic knowledge from elsewhere, provides the first step toward an automated pipeline that will immensely improve our ability to link organisms to environment and to understand the evolutionary change associated with phenotypic changes such as adaptation to a new host, switching to a new habitat or adapting to use a new substrate. By providing the foundation for organizing microbial knowledge, these expanded taxonomies will greatly expand our ability to understand the microbes that pervade all aspects of life on the Earth.
We thank Adam Arkin and Sean Eddy for supporting this study. The work conducted by ENIGMA was supported by the Office of Science, the Office of Biological and Environmental Research, of the US Department of Energy under Contract no. DE-AC02-05CH11231. This work was supported in part by the Sloan Indoor Environment program, the National Institutes of Health (grants UH2/UH3CA140233 and U01-HG004866), the Bill and Melinda Gates Foundation and the Howard Hughes Medical Institute.
Supplementary Information accompanies the paper on The ISME Journal website (http://www.nature.com/ismej)