Background Direct interrogation of microbial genomes based upon comparisons of orthologous gene sequences or metagenomic surveys provides a means to assess the diversity of microbial communities without requiring the cultivation of microbes in the laboratory. Since the cost of cloning DNA templates and capillary-based DNA sequencing constrains the number of sequences included in most of these investigations, the detection of low abundance taxa demands surveys that are many orders of magnitude larger than those reported in the literature. Massively parallel pyrosequencing on the Roche GS20 system developed by 454 Life Sciences offers a means to more extensively sample molecular diversity in microbial populations. It is now possible to generate hundreds of thousands of short (100-200 nucleotide) DNA sequence reads in a few hours without requiring the preparation of sequence templates by conventional cloning. In the near future, technical advances will likely increase the number and length of sequence reads. Pyrosequencing technology relies upon enzyme cascades and CCD luminescence detection capabilities to measure the release of inorganic pyrophosphate with every nucleotide incorporation [1]. The GS20 system takes advantage of DNA capture beads that contain, on average, one single-stranded template, which is amplified to millions of copies in an oil emulsion PCR (emPCR). The beads are then distributed on a solid-phase sequencing substrate (a PicoTiterPlate™) with 1.6 million wells that can each hold a bead and additional reagents, including polymerase, luciferase, and ATP sulfurylase. Microfluidics cycle each of the four nucleotide triphosphates over the PicoTiterPlate™, and incorporation of a nucleotide releases pyrophosphate, the substrate for a luminescence reaction, which is recorded with a cooled CCD camera. The record of intensity of each flow of a nucleotide is a flowgram, analogous to a chromatogram that reports the order of A, C, G and T residues from a DNA sequencing template. Flowgram values correspond to the homopolymer length for that base. The average number of wells with detectable sequencing templates is about 450,000, which produces about 200,000 usable reads. This new methodology brings with it different sources of error to traditional dideoxy capillary sequencing. Since the nucleotide triphosphates are flowed one at a time, substitutions are less likely than with traditional methods. However, it is sometimes difficult to resolve the intensity of luminescence produced when a homopolymer is encountered. The result can be ambiguity of homopolymer length, particularly for longer homopolymers. In addition, insufficient flushing between flows can cause single base insertions (carry forward events) usually near but not adjacent to homopolymers. Insufficient nucleotides within a flow can cause incomplete extension within homopolymers. Generally, an excess of intermediate flowgram values indicates a poor quality read [2]. The GS20 software makes corrections for carry forward and incomplete extensions (CAFIE); it shortens reads from the 3' end until fewer than 3% of the remaining flowgram values are of intermediate value, and it removes reads if the trimming falls below a threshold length. The software identifies as ambiguous flow cycles in which no flowgram value was greater than 0.5. If 5% or more of the flow cycles for a read are ambiguous, the read is removed. The assembly of many overlapping pyrosequencing reads can produce highly accurate consensus sequences [3,4]. Wicker et al. [5] compared assemblies of the barley genome produced by reads from GS20 pyrosequencing and from ABI dideoxy sequencing. Both methods produced consensus sequences with error rates of approximately 0.07% at each consensus position. Gharizadeh et al. [6] compared pyrosequences with Sanger dideoxy methods for 4,747 templates. Comparisons of the traditional capillary sequences with the 25-30 nucleotide pyrosequence reads demonstrated similar levels of read accuracy. Assemblies of massively parallel pyrosequencing reads of plastid genomes from Nandina and Platanus exhibited overall error rates of 0.043% and 0.031%, respectively, in the consensus sequence [4]. The generation of consensus sequences to improve accuracy, however, is generally not appropriate for studies that seek information about natural variation from every read. For example, in metagenomic [7] or PCR amplicon [8] libraries from environmental DNA samples, each sequence read can theoretically represent DNA from a distinct gene from a complex mixture of microbial genes. A viable but imperfect alternative to building consensus sequences for metagenomic and diversity investigations is to identify and remove pyrosequencing reads that are likely to be incorrect. For example, Gilbert et al. [9], in a study of ancient woolly mammoth mitochondrial DNA, removed pyrosequencing reads that were not 98% identical to previously sequenced mammoth mitochondrial DNA sequences, assuming that they must be poor quality. Dostie et al. [10] sequenced an amplicon library and discarded reads in which the PCR primer was not recognized by BLAST. These studies removed 15% and 7% of their reads, respectively, but it is not clear that these statistics improved the quality of the remaining data. To explore error modalities, we used the GS20 system to generate more than 340,000 reads from a PCR amplicon library that was prepared from a collection of 43 reference templates of known sequence. Each reference template contains a distinct ribosomal RNA gene (rDNA), including the V6 hypervariable region from a collection of 43 divergent bacteria [11]. Differences between pyrosequences and their cognate reference sequences identified signatures of low quality data. Results Read accuracy We obtained 340,150 reads that passed the GS20 quality filters, that is, flowgrams for each read: contained the correct base key at the start (a portion of the 454 primer used to differentiate reads from internal quality control sequences); included at least 84 flows; had fewer than 5% of flow cycles resulting in an ambiguous base call (N); and had fewer than 3% of flowgram values between 0.5 and 0.7 [12]. We aligned each read to its reference sequence using an optimized Needleman-Wunsch algorithm. Our data included 159,981 total errors over 32,801,429 bases. The error rate, defined as the number of errors (miscalled bases plus inserted and deleted bases) divided by the total number of expected bases, was 0.49%. As shown in Table 1, 39% of these errors correspond to homopolymer effects, including extension (insertions), incomplete extensions (deletions) and carry forward errors (insertions and substitutions). Carry forward occurs when an incomplete flush of base flow results in a premature incorporation of a base. The presence of homopolymers tends to increase the likelihood of both carry forward and incomplete extension with the GS20 sequencer [12]. Insertions were the most common type of error (36% of errors) followed by deletions (27%), ambiguous bases, Ns (21%), and substitutions (16%). It should be noted that the V6 region does not contain long or frequent homopolymers. The errors did not correlate significantly with distance along the sequence (R2 60%, also contain Ns. Conclusion Our analysis of the GS20 sequencing error rate of the V6 region of bacterial rRNA genes shows a marked improvement over the original error rates published by Margulies et al [2]. The largest source of errors may be due to multi-templated beads, and enhancements to both the chemistry protocol for the GS20 and the built-in bioinformatics software may account for the change in error rates. Our results highlight that a small proportion of low quality reads, presumably from multi-templated beads, are responsible for the majority of sequencing errors. The ability to identify and remove these reads is the best way to improve the accuracy of the entire dataset. It is not a replacement for assigning quality scores to detect the position of miscalled bases. The interpretation of chromatograms by programs such as PHRED [13] employs quality scores that reflect the probability of any type of base call error. Although it uses the same scale, the GS20 software generates quality values based on the probability of homopolymer extension rather than probability of a correct base call. Regardless of the ultimate cause of poor reads, the presence of even a single ambiguous base (N) was an effective indicator of low-quality sequence. The removal of all reads containing one or more Ns can drastically improve the overall quality of the remaining dataset, reducing the error rate from about 0.5% to about 0.25% (Table 2). For our data, this strategy eliminated only 6% of the total reads. By excluding approximately 1% of all reads whose lengths lie outside of the main distribution, as well as those with inexact matches to the primer, the error rate for the V6-tag data dropped to less than 0.2%. The pyrosequencing technology provides such a large number of reads that the elimination of even 10% or more of the reads in a data set should be more than offset by the increase in quality of the remaining reads. Table 2 Identifying low-quality reads and their contribution to the error rate Data selection Percent of reads Error rate All reads 100.0% 0.49% Reads with no Ns 94.4% 0.24% Reads with one or more Ns 5.6% 4.7% Reads with length ≥81 and ≤108 98.8% 0.33% Reads with length 108 1.2% 18.9% Reads with no Ns and length ≥81 and ≤108 93.3% 0.20% Reads with no proximal errors 97.0% 0.45% Reads with fewer than three proximal errors >99.99% 0.48% Reads with more than three proximal errors <0.01% 12.2% Reads with no Ns and length ≥81 and ≤108 and no proximal errors 90.6% 0.16% Removing reads with Ns is the most effective means we found of removing low-quality data and improving the error rates. Read lengths that are either longer or shorter than expected, and are outside the peak of common reads, also correlate strongly with incorrect reads. Our strategy for detecting low quality reads circumvents the need to generate consensus sequences for improving data quality in massively parallel pyrosequencing experiments of environmental DNA. Our criteria for detecting reads with errors allows for the acquisition of pyrosequencing data in the context of molecular ecology that can surpass the accuracy of traditional capillary methods. Materials and methods Generation of 1 kb clone library and selection of clones for pyrosequencing DNA was extracted according to Huber et al. [14] from diffuse flow hydrothermal vent samples as described in Sogin et al. [8]. PCR primers were designed using ARB software [15] to target the bacterial 16S rDNA. The primers used were 337 F (5' CAN CCT ACG GGN GGC NGC) and 1391R (5' GAC GGG CGG TGW GTN CA). The amplification mix contained 5 units Pfu Turbo polymerase (Stratagene, La Jolla, CA, USA), 1× Pfu reaction buffer, 200 μM dNTPs (Pierce Nucleic Acid Technologies, Milwaukee, WI, USA), and 0.2 μM each primer in a volume of 100 μl. Environmental DNA (3-10 ng) was added to 3 separate 30 μl amplification mixes. A positive control (Marinobacter aquaeolei genomic DNA) and two negative controls (no DNA and genomic DNA from the archaeon Methanococcus jannaschii) were also run. An initial denaturation step of 3 min at 94°C was followed by 30 cycles of 94°C for 30 s, 55°C for 45 s, and 72°C for 2 minutes. The final extension step was 72°C for 2 minutes. Following PCR, three reactions for each sample were combined, purified, and concentrated using the MinElute PCR Purification Kit (Qiagen, Valencia, CA, USA) according to the manufacturer's instructions. PCR product quality was assessed on a 0.8% agarose gel stained with ethidium bromide and ligated with pCR4-TOPO vector for 20 minutes at room temperature and transformed with TOP10 electrocompetent cells according to the manufacturer's instructions (Invitrogen, Carlsbad, CA, USA). Colonies for each library were randomly selected and grown in SuperBroth with 50 mg/ml kanamycin in 96 deep-well blocks overnight. Alkaline lysis template preparation was carried out on cell pellets using the RevPrep Orbit II (Genomic Solutions, Ann Arbor, MI, USA) or the Biotech RoboPrep2500 (MWG Biotech, Ebersberg, Germany). The 1,000 base-pair amplicons were sequenced bidirectionly using primers T3 (5'-ATT AAC CCT CAC TAA AGG GA) and T7 (5'-TAA TAC GAC TCA CTA TAG GG), and on an ABI 3730 × l genetic analyzer. Sequences were aligned with MUSCLE (with parameters -diags and -maxiters 10) [16] and manually manipulated in the BioEdit 7.0.1 program [17]. Distance matrixes were calculated using quickdist [8], and taxonomic identities determined using RDP-II Classifier [18]. Sequences were trimmed to include only the V6 region of the gene, the distance matrix re-calculated, and from this analysis, 43 divergent sequences were chosen for further experimentation. The average length of the V6 region for these clones was 101 bases, ranging from 95 to 109, with one longer reference of 183 bases. The 16S rDNA sequences are deposited at GenBank under accession numbers DQ909092, DQ909128 DQ909132, DQ909133, DQ909142, DQ909144, DQ909158, DQ909184, DQ909202, DQ909204, DQ909218, DQ909223, DQ909224, DQ909248, DQ909251, DQ909253, DQ909266, DQ909274, DQ909337, DQ909368. DQ909392, DQ909396, DQ909400, DQ909414, DQ909423, DQ909438, DQ909440, DQ909465, DQ909474, DQ909498, DQ909513, DQ909519, DQ909538, DQ909603, DQ909618, DQ909631, DQ909662, DQ909688, DQ909702, DQ909706, DQ909719. DQ909727, DQ909753. Generation of known V6 amplicon library We treated each plasmid with plasmid-safe DNAase (Epicentre, Madison, WI, USA) to remove Escherichia coli genomic DNA and confirmed that each plasmid produced an amplification product of the expected size with primers targeting the V6 region of the bacterial rDNA according to Sogin et al. [8]. We then pooled the individual plasmids and amplified with the primers that flank the V6 region of rRNA genes according to Sogin et al. [8]. We assessed the product quality using a BioAnalyzer Agilent DNA 1000 LabChip following the manufacturer's instructions. Three reactions were combined, purified, and concentrated using the MinElute PCR Purification Kit (Qiagen). The final amplicon library was sequenced independently by 454 Life Sciences and our own lab. Both labs used the Roche Genome Sequencer 20 (GS20) according to the manufacturer's specifications [2]. The original GS20 output files as text are available in Additional data files 3-5. Error rate calculations We combined the data from both sequencing runs for a total of 340,150 reads (226,150 and 114,000), with an average read length of 94.5 nucleotides and a total of 32,816,656 bases. These sequences are available in fasta format in Additional data file 2. To determine the reference sequence source of each pyrosequencing read, we ran a separate multiple sequence alignment of each individual read against the 43 reference sequences using MUSCLE [16] (default options plus maxiters 2, diags). We calculated the number of sequence differences between each read and the reference sequences to determine the reference sequence to which each read mapped most closely. All subsequent error calculations are based on comparing reads to their assigned reference sequence. The overall error rate is the number of errors in a read divided by the length of sequence. Specifically, we calculated errors in several ways. In all methods, each base mismatch or N in the test sequence counts as an error, and a terminal gap caused by a GS20 read terminating before the end of the reference does not count as an error. In the first and second methods each base of an insertion or deletion counts as one error. In the third method, insertions or deletions are counted by homopolymer runs. If a TAAA is inserted, it is counted as two insertions, one single-T and one multi-A insertion. The denominator for the first method was the read length. The denominator for the second and third methods was the length of reference sequence minus any discounted terminal gaps. All error rate calculations produced essentially the same results. We report error rates using all base errors (not by homopolymer run) divided by the expected length (reference sequence length minus terminal gaps). The error rates were calculated for each sequence in a pairwise comparison of the pyrosequencing read and the reference sequence to which it was assigned. We used the Needleman-Wunsch algorithm [19] for these pair-wise alignments because it selects for the best possible alignment given its run parameters. Using a set of 100 sequences and a matrix of Needleman-Wunsch run parameter combinations, we found that a gap opening penalty of 5.75 and a gap extension penalty of 2.75 minimized the calculated error rate. Error rates, reference sequences and read sequences were imported into a MySQL database for storage and analysis. Additional data files The following additional data are available with the online version of this paper. Additional data file 1 is a fasta file of the 43 known sequences used. Additional data file 2 is a gzip-compressed fasta file of the sequences output by the GS20. These sequences correspond to those included in Additional data files 3, 4, 5 but include only the final sequence information. Additional data files 3, 4, 5 are three compressed text files representing the text translations of the original GS20 binary output (sff) files for all of the sequencing used in the analysis, including sequence, flowgram and other run information. GS20 data are reported by region of the PicoTiterPlate™; we sequenced three plate regions. Supplementary Material Additional data file 1 The 43 known sequences used Click here for file Additional data file 2 These sequences correspond to those included in Additional data files 3-5 but include only the final sequence information in fasta format. Click here for file Additional data file 3 Text translation of the original GS20 binary output (sff) file for the first of three PicoTiterPlate™ regions used in the analysis. Click here for file Additional data file 4 Text translation of the original GS20 binary output (sff) file for the second of three PicoTiterPlate™ regions used in the analysis. Click here for file Additional data file 5 Text translation of the original GS20 binary output (sff) file for the third of three PicoTiterPlate™ regions used in the analysis. Click here for file
