Quality checks

Every imported SSU and LSU sequence had to pass a multi-stage quality inspection. Sequences were rejected if they were shorter than 300 unaligned nucleotides, if they were composed of more than 2% of ambiguities or more than 2% homopolymeric stretches longer than four bases, which means only bases exceeding homotetramers are counted, or if they had more than 5% identity to vector sequences. The identity was checked by querying a database of commonly used vector sequences, based on the EMVEC (http://www.ebi.ac.uk/blastall/vectors.html) and UniVec (http://www.ncbi.nlm.nih.gov/VecScreen/VecScreen.html) databases using the blastn tool. All thresholds to reject sequences were defined based on statistical analysis of the retrieved SSU and LSU sequences. Each sequence in the SILVA databases carries the percentages of ambiguities, homopolymers, and vector contamination. Additionally, a summary ‘sequence quality’ score is calculated according to the following formula, where Sq = sequence quality, A = % ambiguities, H = % homopolymers and V = % vector identity:

This score represents the mean of the three individual parameters, such that 100 is the best possible value. All sequences that passed the quality thresholds were automatically aligned against the seed alignment using the new SILVA INcremental Aligner (SINA).

Aligner

To cope with the huge amount of sequence information and to minimize the workload for manual curation, a new dynamic incremental profile sequence aligner (SINA) was developed. In the first step the aligner uses the suffix tree concept of ARB (8) to search for up to 40 closely related sequences in the seed alignment. The reference sequences from the seed are transferred into a partial order graph as used in (18), while preserving the positional identity from the reference alignment. The sequence under investigation is then aligned to this graph using a variant of the Needleman-Wunsch algorithm (19) with affine gap penalties and cost free overhang. The graph concept allows ‘jumping’ between the different references to find an optimal alignment for the different sequence regions. This technique enables the algorithm to correctly place bases that were e.g. deleted from the closest relative, but are present in the candidate sequence and other relatives. It also eliminates the need for synthetic full length sequences in the reference alignment as introduced for the NAST aligner (20). To further improve the alignment quality a variability statistic is used to give more weight to conserved positions. Results of each step of the aligner are reported to the database and shown in the corresponding fields of the exported ARB file (Tables 1–3). The ‘alignment quality’ score is a measure of the similarity with the seed sequences that are taken into account for the alignment process. The score is derived from the dynamic programming score by removing the effects of sequence length and positional weighting. High values (>90) mean that nearly identical sequences have been found within the seed alignment, resulting in a high likeliness for the alignment to be accurate. Low values indicate a high distance as perceived by the aligner, making the alignment task more difficult and lowering the average accuracy. Due to the size of the seed alignment, low values are rather rare and ask for manual inspection of the alignment. The ‘basepair’ score is calculated from the number of bases involved in helix binding according to the secondary structure model of Gutell et al. (21) as already implemented in the ARB package. Canonical and non-canonical base pairings are evaluated, weighted according to the cost model implemented in the Probe_Match (‘weighted mismatches’) tool in ARB (8). To fit our unified scoring scheme, the alignment quality and the base pair score were normalized to values between 0 and 100, such that 100 represents the maximum score. After aligning, the number of successfully aligned bases was again counted and sequences with less than 300 bases within the boundaries of the respective SSU or LSU rRNA genes were discarded.

Anomaly check

To check for sequence anomalies, a custom version of the Pintail software (22) was used. The software was specifically adapted for batch processing by the RDP II team. By design, Pintail can only detect anomalies between two sequences. To circumvent this limitation, a pairwise comparison of all sequences in the seed against a group of 20 sequences was performed. If a majority of the comparisons was deemed anomalous, the sequences were iteratively eliminated from the seed alignment until no such sequences remained. Subsequently, all aligned sequences of the SSU database were tested against their five closest relatives within this pruned seed. The number of ‘yes’, ‘likely’ and ‘no’ reported by Pintail was counted for each sequence and transferred into the ‘Pintail quality’ value. This score was normalized between 0 and 100, such that 100 indicates the best quality and a low probability that the sequence is anomalous or chimeric. Only SSU sequences were checked for anomalies because the Pintail software is currently not designed to handle LSU sequences.

Taxonomy

Every sequence in the SILVA databases carries the EMBL taxonomy assignment. Where available, the greengenes and RDP taxonomies where added for comparison. The EMBL taxonomy was retrieved simultaneously with the sequence, whereas the other taxonomies have been assigned to the sequences based on accession numbers. The greengenes taxonomic outline was acquired in June 2007 from the greengenes website (http://greengenes.lbl.gov/) and the RDP Nomenclatural Taxonomy was acquired from RDP II release 9.51. At the moment, no other up to date databases containing aligned LSU sequences are available. Therefore, the only taxonomy provided with the LSU database is the taxonomy used by EMBL. Type strain information has been added to the field ‘strain’ and is indicated by [T]. Mapping was done based on the RDP II dataset and is therefore only available for Bacteria.

Nomenclature

All organism names have been synchronized with the ‘Nomenclature up to date’ website of the Deutsche Sammlung für Mikroorganismen und Zellkulturen” DSMZ (released June 2007, http://www.dsmz.de/download/bactnom/names.txt) in order to stay consistent with the constant renaming of validly described species according to the recommendations published in the ‘International Journal of Systematic and Evolutionary Microbiology’. All former names are stored in the database and are visible on the web page, as well as in the corresponding field of the ARB databases (Tables 1–3).

SSU and LSU rRNA databases for ARB

Two types of precompiled databases for both small and large subunit ribosomal RNA sequences are available in the ARB format: the high-quality Ref databases and the comprehensive Parc databases. The Ref databases are subsets of Parc, which are exclusively comprised of nearly full length 16S/18S and 23S/28S rRNA sequences. A sequence is accepted if it is at least 1200 bases long. Additionally, sequences as short as 900 bases are included if they belong to the domain Archaea. Applying a strict cut-off at 1200 bases would result in the loss of the majority of sequences of this domain. Sequences in the LSU Ref database have to be at least 1900 bases long. For quality control, all sequences that could not be unambiguously aligned (alignment quality score <50 and <30 for SSU and LSU, respectively) were removed from the Ref databases. Both Ref databases are supplemented with a guide tree based on the full length sequence tree of the ARB Jan 04 SSU and the Ludwig LSU databases, respectively. The trees were built using the ARB parsimony tool with filters to remove highly variable positions. Common filters like the positional variability filters were recalculated for the Ref databases. Sequences with long branches in combination with low alignment qualities (<80) were removed from the Ref databases.

The rRNA Parc databases are a collection of all quality checked and automatically aligned rRNA sequences longer than 300 bases of the aligned rRNA gene (field ‘nuc_gene_slv’, Tables 1–3). The name Parc has been chosen according to the UniProt concept (23), where Parc stands for the comprehensive protein sequence archive. All sequences in the SILVA databases are associated with a rich set of sequence and process parameters. Included is information from the initial quality checks to the alignment process, as well as information taken directly from the EMBL entry (Tables 1–3). Together with the search and query functionalities on the web site and in ARB, one can quickly search for problematic sequences or generate individual high or low quality sequence subsets for further processing or curation. The ARB package can be used to export sequences in various formats like EMBL, GenBank, or aligned and unaligned FASTA.

Availability/Webpage

The SILVA databases are available via a web-based interface at http://www.arb-silva.de. The web interface is divided into six sections: the browser, search, list, download, background, and FAQs pages. Download of the complete Parc and Ref datasets in ARB format is available in the download section. It is also possible to download files that gain additional sequences from new releases. Subsets of aligned sequences from the Parc dataset can be retrieved from the website. This is currently possible via two entry points: a taxonomic browser and advanced search functions. After selecting a database and the desired taxonomy in the browser, the user can navigate through the taxonomy by clicking on the respective nodes. A cart system is used to easily select subsets of single sequences, complete groups or even phyla. The cart system keeps the selections for the SSU and LSU databases distinct. This allows the user to select sequences from both databases simultaneously without mixing the two sequence types. However, it must be noted that any misclassification or erroneous information provided by INSDC is currently propagated on the SILVA webpage.

Additionally, the advanced search functions of the SILVA website can be used to build custom subsets of sequences. In addition to simple searches e.g. for accession numbers, organism names, taxonomic entities, or publication DOI/PubMed IDs, complex queries over several database fields using constraints such as sequence length or quality values are possible. The results can be sorted according to accession numbers, organism names, sequence length, sequence and alignment quality and Pintail values. Before download, the search results must be added to the ‘List’. This can be done by either manually selecting the sequences by mouse click or by clicking on ‘Add complete result to List’ to mark and transfer all results.

The coloured bars on the search page and in the short and detailed sequence views of the browser given a fast overview of the different quality aspects assigned to every sequence. The length of the bars is a graphical representation of the respective quality value. The colours classify the information into four categories: A green bar represents a value equal to or greater than 75. Yellow bars stand for values equal to or greater than 50 but less than 75. Values less than 50 are expressed by an orange bar. Red bars are only used for scores of 0. Since ‘problematic’ sequences, sequences of inadequate quality, as well as insufficiently aligned sequences were discarded from the databases only the Pintail scores can have 0.

In the ‘List’ section of the website, the entries can be inspected, items can be deleted, and the download files can be created. By clicking on the ‘generate download’ button the user will be asked whether he would like to download the sequences as a multi-FASTA or ARB file from the download section of the web page. All generated files will be available for download on the download page for up to 24 h. The background section of the website provides additional information about the current status of the databases, and the FAQ section describes the main steps necessary to download subsets of sequences and how to merge the retrieved ARB databases with the user's personal ARB database.

Alignment and aligner

The current SILVA alignment is based on 46 000 and 150 000 alignment positions for the small and large subunit rRNA, respectively. The reasons for the large amount of alignment positions are: (i) large insertions often present in Eukarya and (ii) sequencing errors, such as additional artificial bases often found in homopolymeric sequence stretches. Such errors are common and require placement to be filtered before phylogenetic tree reconstruction, without corrupting the rest of the alignment.

In the ‘align-to-seed’ approach implemented in the SILVA system, well aligned sequences from seed datasets are used as references for new, unaligned sequences. Therefore, the quality of the final alignment strongly depends on the accuracy of the seed alignment. To further improve the quality of the SSU and LSU seed databases a manual curation process was performed on the latest officially released ARB dataset from January 2004.

The SSU seed hosts currently over 1000 unpublished sequences that primarily cover the domain Archaea. These sequences further improve the alignment in regions of the original SSU January 2004 dataset with sparse sequence coverage. In summary, the quality and consistency of all of the seed alignments is excellent. Only minor inconsistencies could not be resolved in the Eukarya. Nevertheless, the Parc datasets exceed the corresponding SSU and LSU seeds by a factor of 8 to 25. This probably indicates that not every phylum is equally represented in the seed. Hence, before using the alignments for in-depth phylogenetic analysis, the alignment of the selected sequence should be carefully checked. Problematic sequences can be easily filtered out by the quality values followed by manual curation. The SILVA team highly appreciates the return of manually inspected and corrected alignments of sequence subsets for inclusion in the SILVA seed. This will allow us to further increase the quality of future alignments.

To manage the deluge of data currently available in the public databases, a new aligner (SINA) has been developed. Similar to existing aligners, such as the Fast Aligner implemented in ARB (8) or the NAST aligner (20), the tool uses related sequences from the reference alignment as a template. For benchmarking the performance of SINA, standard tools, such as BAliBASE (27), could not be used since they are restricted to protein sequences. Benchmark results were obtained by removing and realigning each sequence from the seed. The results were internally compared to the original alignment by counting the number of matching and non-matching columns. Overall, SINA correctly placed 99.8% of all bases in the alignment. Furthermore, 33% and 80% of all sequences tested had no, or less than 1%, alignment errors, respectively. The high accuracy was gained in extensive test runs by changing parameter sets for gap insertions/extension parameters and family sizes combined with subsequent manual inspection of the results by expert curators. The design and implementation of SINA as individually running processes allows distributed aligning on cluster nodes. More than one sequence per second can be aligned per CPU.

We would like to thank Ralf Westram for expert assistance with the ARB software suite, the company Pixelmotor for designing and implementing the webpage and all colleagues and students who helped with the manual curation of the databases. We would also thank James Cole, George Garrity and the RDP II team for help with Pintail and fruitful discussions. We are grateful for funding from the Max Planck Society. Funding to pay the Open Access publication charges for this article was provided by the Max Planck Society.

Conflict of interest statement. None declared.