Introduction

Metatranscriptomic analysis using pyrosequencing is dramatically improving our understanding of gene expression in natural microbial communities (DeLong, 2009; Poretsky et al., 2009; Shi et al., 2009). In these analyses, cDNA is synthesized from total RNA extracted from an environmental sample and used directly for massively parallel shotgun sequencing. Metatranscriptomes from functionally diverse habitats, including seawater and soil, can be sequenced using pyrosequencing methodologies (GS 20 or GS FLX systems, Roche 454 Life Sciences, Branford, CT, USA), yielding tens to hundreds of thousands of sequence fragments from the RNA pool (Frias-Lopez et al., 2008; Urich et al., 2008; Poretsky et al., 2009). Such analyses provide detailed information on the taxonomic and functional diversity in the transcriptionally active community, as they simultaneously characterize both the ribosomal and messenger RNA components of the transcript pool. However, to advance the generic application and utility of pyrosequencing-based transcriptomics to microbial ecology studies, it is necessary to develop quality assurance and methodological troubleshooting techniques that both enhance current protocols and minimize bias in the interpretation of the read data.

It is not yet clear to what extent pyrosequencing methods capture the full breadth of expressed functional genes in microbial community transcriptomes. Indeed, pyrosequencing likely fails to capture many functionally important transcripts that occur at low frequencies. This is particularly true of data sets dominated by ribosomal RNA sequences (for example, prokaryotic 5S, 16S and 23S rRNA, eukaryotic 18S and 28S). In the first study of a marine microbial metatranscriptome sequenced using pyrosequencing (Roche GS 20 system), rRNA reads represented 53% of total sequences (Frias-Lopez et al., 2008). A similar study using the same technology reported rRNA contamination at 37% of total sequences (Poretsky et al., 2009), following the application of two commercial kits designed to enrich for mRNA (mRNA-ONLY Prokaryotic mRNA isolation kit, Epicentre; MICROBExpress Bacterial mRNA enrichment kit, Ambion, Austin, TX, USA). Subsequent analyses of marine transcriptomes using the Roche GS FLX platform, which yields average read lengths (200–250 bp) over twice those produced via the GS 20 technology, have shown rRNA abundances that often exceed 90% of total reads (DeLong lab, unpublished data; Hewson et al., 2009b; Hewson I, personal communication). The increase in rRNA abundance observed with FLX-based technology relative to GS 20 is likely due to differences in the pyrosequencing protocols themselves. For example, the FLX protocol includes a bead-binding purification step that selects for longer transcripts, relative to GS20, which may increase the relative rRNA representation. As read lengths increase—the Roche GS FLX Titanium reagents now facilitate average read lengths greater than 400 bp—achieving adequate sequencing depth of non-rRNA reads, primarily mRNA, requires effective subtractive procedures to minimize rRNA abundance. Ribosomal RNA subtraction becomes particularly relevant as metatranscriptomics moves from a purely descriptive phase to one in which the method is applied experimentally to track low frequency changes in gene expression, for example, in response to environmental perturbations (DeLong, 2009).

Here, we introduce a sample-specific method for the subtraction of rRNA from total RNA. The method employs subtractive hybridization using antisense rRNA probes transcribed in vitro from PCR products amplified from coupled DNA samples, thereby ensuring the specificity of the probe mix. We tested this method by sequencing rRNA-subtracted and unsubtracted transcriptomes of open ocean bacterioplankton communities at two depths in the North Pacific Subtropical Gyre (Station ALOHA, North Pacific), showing a substantial reduction in the targeted RNA fraction (bacterial 16S and 23S rRNA). Our method can be generally extended to any microbial community sample, and can be easily expanded to target other rRNA transcripts, including those of archaea or eukaryotes.

The increasing use of pyrosequencing-based metagenomics and transcriptomics would also benefit from explicit quality control tests to determine the quantitative reproducibility of the data generated and to ensure reasonable extrapolations of read abundance to in situ transcript abundance. Gomez-Alvarez et al. (2009) showed that metagenomic data sets generated using pyrosequencing contain significant numbers (11–35% of total reads) of sequencing artifacts in the form of replicate sequences. These artificial replicates, putatively generated during the emulsion PCR stage of pyrosequencing (Briggs et al., 2007; Gomez-Alvarez et al., 2009), can erroneously increase the apparent abundance of transcripts from which these sequences derive. To address these and other concerns related to the reproducibility of expression profiles, we performed a set of technically replicated pyrosequencing-based transcriptomic analyses. These analyses determined criteria for replicate removal, underscored the potential for pyrosequencing artifacts (replicate reads) to confound interpretation of transcript diversity and abundance, and highlighted the potentially limited extent to which standard sequencing depths reveal the diversity of total transcript pools.

Materials and methods

Sample collection

Planktonic microorganisms were sampled from the photic zone (25 m and 75 m) at Station ALOHA (22° 45′N, 158° 00′W) as part of the Hawaii Ocean Time-series (HOT) program (HOT-186 cruise, casts S2C27 and S2C30) in October 2006, as described in Shi et al. (2009). Replicate seawater samples (1.8–2.0 l) for RNA extraction were prefiltered through 1.6 μm GF/A filters (47 mm dia., Whatman, Kent, UK) and collected onto 0.22 μm Durapore filters (25 mm dia., Millipore, Billerica, MA, USA) using a peristaltic pump. Filters were immediately transferred to microcentrifuge tubes containing 300 μl RNAlater (Ambion) and frozen at −80 °C. Less than 20 min elapsed between sample collection (arrival on deck) and fixation in RNAlater. Samples for DNA extraction were collected from the same water sample used for RNA collection as in Frias-Lopez et al. (2008). For each sample, seawater (1.8–2.2 l) was prefiltered through a 1.6 μm GF/A filter (125 mm dia., Whatman) onto a 0.22 μm Steripak-GP20 filter (Millipore). The filter units were filled with lysis buffer (50 mM Tris-HCl, 40 mM EDTA and 0.75 M sucrose), capped, and frozen at −80 °C until extraction.

RNA and DNA isolation

Total RNA was extracted from filters using a modification of the mirVana miRNA Isolation kit (Ambion) as described previously (Shi et al., 2009). Briefly, samples were thawed on ice, and the RNAlater surrounding each filter was removed by pipetting and discarded. Filters were immersed in lysis/binding buffer (Ambion) and vortexed to lyse attached cells. Total RNA was then extracted from the lysate according to the manufacturer's protocol, incubated (37 °C for 30 min) with TURBO DNA-free to remove genomic DNA, and purified and concentrated using the RNeasy MinElute Cleanup kit (Qiagen, Valencia, CA, USA). Genomic DNA was extracted from Steripak filters as described previously (Frias-Lopez et al., 2008).

rRNA subtraction

Subtractive hybridization using sample-specific biotinylated rRNA probes was used to remove bacterial 16S and 23S rRNA from total RNA samples (Figures 1 and 2). The probe synthesis reaction was adapted from an in situ hybridization method by DeLong et al. (1999) and combined with a subtractive hybridization protocol similar to that of Su and Sordillo (1998). Ribonucleotide probes targeting bacterial 16S and 23S rRNA genes were generated from community DNA samples collected in tandem with each total RNA sample. Templates for probe generation were first prepared by PCR using universal primers flanking nearly the full length of the bacterial 16S gene and 85% of the 23S rRNA gene, with reverse primers modified to contain the T7 RNA polymerase promoter sequence (Table 1; DeLong et al., 1999; Hunt et al., 2006). PCRs (50 μl each) included 100 ng template DNA, 1 μl of Herculase II Fusion DNA Polymerase (Stratagene, La Jolla, CA, USA), 1 × Herculase reaction buffer, 10 mM dNTP and 10 μM each of forward and reverse primers. Reaction conditions were as follows: 2 min at 92 °C; 35 cycles of 20 s at 95 °C, 20 s at 39 °C (23S reactions) or 55 °C (16S reactions), 75 s (16S) or 90 s (23S) at 72 °C; 3 min at 72 °C. Resulting products were visualized via gel electrophoresis and purified via the QIAquick PCR purification kit (Qiagen). The above rRNA probe generation step can also potentially be modified by generating amplicons via reverse transcription (RT)-PCR, using RNA (rather than DNA) as starting template. We deliberately chose to amplify from the DNA pool, however, as RT-PCR is less efficient at longer sequence lengths, and we sought to maximize probe coverage along the length of all rRNA genes.

Figure 1
figure 1

Flowchart of the metatranscriptomic sample processing pipeline, illustrating steps for the sample-specific subtraction of bacterial ribosomal RNA (16 and 23S) from total RNA in an environmental sample.

Figure 2
figure 2

Size distribution of total RNA in unsubtracted and rRNA-subtracted portions of the HOT-186 25 m sample.

Table 1 Primers used for PCR to generate bacterial 16S and 23S rRNA probes and for reverse transcription of polyadenylated cDNA

Biotinylated antisense rRNA probes were generated by in vitro transcription (IVT) with T7 RNA polymerase using T7 promoter-containing 16S and 23S amplicons as templates. IVT was conducted using the MEGAscript High Yield Transcription kit (Ambion), with the following modifications. Probes for 16S and 23S rRNA were generated separately in 20 μl reactions, each containing: 1 × buffer, T7 RNA polymerase, SUPERase·In RNase inhibitor (10 U), ATP (7.5 mM), GTP (7.5 mM), CTP (5.625 mM), UTP (5.625 mM), biotin-11-CTP (1.875 mM, Roche), biotin-16-UTP (1.875 mM, Roche) and 16S/23S DNA template (250–500 ng). Reactions were run at 37 °C for 4–5 h, then DNAse digested with TURBO DNAse (Ambion) for 15 min at 37 °C. Products were purified using the MEGAclear kit (Ambion). Assuming the template amplicons used for IVT were 50% GC, the biotin labeling density in the resulting 16S/23S probes was 1 in 8 nucleotides.

Biotinylated rRNA probes were hybridized to complementary rRNA molecules in the total RNA sample. Hybridization reactions (50 μl), each containing formamide (20%), 1 × SSC buffer (0.15 M sodium chloride, 0.015 M sodium citrate), SUPERase. In RNase inhibitor (20 U), template RNA (25 m 600 ng; 75 m 200 ng), and equal amounts of 16S and 23S rRNA probes at a final template-to-probe ratio of 1:2 (mass, per probe), were denatured at 70 °C for 5 min and incubated at room temperature (RT) for 3 min. Biotinylated double-stranded rRNA was then removed from the sample by hybridization (10 min at RT) to streptavidin-coated magnetic beads (New England Biolabs, Ipswich, MA, USA; 50 μl aliquot, washed 3 × in 1 × SSC), followed by separation on a magnetic rack (2 min) and removal of the rRNA-subtracted supernatant via pipet. An additional 50 μl 1 × SSC was applied to the beads for washing, separated as above, and pooled with the original supernatant. The pooled products were purified via the RNeasy MinElute Cleanup kit (Qiagen). Subtraction efficiency was evaluated by monitoring the removal of 16S and 23S peaks from total RNA profiles using a 2100 Bioanalyzer and the RNA 6000 Pico chip kit (Agilent, Santa Clara, CA, USA; Figure 2).

Minor updates to this protocol were implemented following characterization of the Station ALOHA (HOT 186) samples. The fully optimized and updated protocol is included as a pdf file in the Supplemental Online Materials. In brief, (1) an initial wash with 0.1 N NaOH was incorporated into the streptavidin bead preparation steps to ensure complete removal of RNases, (2) the denaturation/hybridization step was changed from 5 min at 70 °C followed by 3 min at RT to 5 min at 70 °C followed by a step-down procedure with 1 min each at 5 °C intervals from 65°C to 25 °C, and (3) probes for archaeal and eukaryotic large and small subunit rRNA were incorporated into the protocol (see Supplementary text and Supplementary Table S3 for primer design and sequences), and additional streptavidin-coupled beads were used to ensure complete removal of these additional probes.

RNA amplification and cDNA synthesis

rRNA-subtracted and unsubtracted total RNA (35–80 ng) was amplified using the MessageAmp II-Bacteria kit (Ambion) as described previously (Frias-Lopez et al., 2008; Shi et al., 2009). Briefly, total RNA was polyadenylated using Escherichia coli poly(A) polymerase. Polyadenylated RNA was converted to double-stranded cDNA via reverse transcription primed with an oligo(dT) primer containing a promoter sequence for T7 RNA polymerase and a recognition site for the restriction enzyme BpmI (T7-BpmI-(dT)16VN, Table 1). cDNA was then transcribed in vitro at 37 °C (25 m for 7 h, 75 m for 14 h), yielding large quantities (10–100 ug) of single-stranded antisense RNA. Amplified RNA (5–10 ug aliquot) was then converted to double-stranded cDNA using the SuperScript III First-Strand Synthesis System (Invitrogen, Carlsbad, CA, USA) with priming via random hexamers for first-strand synthesis, and the SuperScript Double-Stranded cDNA synthesis kit (Invitrogen) for second-strand synthesis. cDNA was then purified with the QIAquick PCR purification kit (Qiagen), digested with BpmI for 2–3 h at 37 °C to remove poly(A) tails, and used directly for pyrosequencing.

Pyrosequencing and technical replicates

Before sequencing, poly(A)-removed cDNA was purified via the AMPure kit (Beckman Coulter Genomics, Danvers, MA, USA). Purified cDNA was used for the generation of single-stranded DNA libraries and emulsion PCR according to established protocols (454 Life Sciences, Roche). Clonally amplified library fragments were then sequenced on a Genome Sequencer FLX System (Roche).

To produce a technically replicated transcriptomic analysis, equal aliquots of the HOT-186 75 m total RNA sample were separated and used independently for rRNA subtraction, RNA amplification and pyrosequencing—see samples 75 m A and B, unsubtracted and rRNA-subtracted (Table 2). All sample processing parameters were kept consistent across replicates.

Table 2 Read numbers and statistics

Data analysis

Ribosomal RNA-derived reads were identified using BLASTN to compare all reads against an rRNA database composed of both prokaryotic and eukaryotic small and large subunit rRNA nucleotide sequences (5S, 16S, 18S, 23S and 28S rRNA) from available microbial genomes and sequences in the ARB SILVA LSU and SSU databases (http://www.arb-silva.de). Reads producing alignments with bit scores greater than 50 were identified as rRNA sequences and removed from pyrosequencing data sets.

Non-rRNA sequences were checked for replicate sequences using the open-source program CD-HIT (Li and Godzik, 2006) according to the protocol of Gomez-Alvarez et al. (2009). Replicates were defined as sequences sharing greater than 99% nucleotide identity, with an allowable length difference of 1 bp, and a requirement that the first 3 bp of the replicate sequences be identical. This cutoff was chosen by comparison of multiple pyrosequencing runs prepared from a single pyrosequencing library (sample 75 m A, rRNA-subtracted), as a conservative measure that significantly reduced the size of sequence clusters appearing at high frequency in one sequencing run but at low frequency in repeat sequencing runs (Supplementary Table S1 and Figure S2). Additional details on criteria for identifying replicates can be found in the Supplementary Online Material.

Non-rRNA sequence reads were compared with the National Center for Biotechnology Information non-redundant protein database (NCBI-nr, as of 1 February 2009) using BLASTX. Top BLASTX hits with e-values less than 1 × 10−5 were used for nr-protein designations. Reads with multiple, equal hits were assigned to the reference protein (hit) with the highest number of previously assigned reads. The total number of reads assigned to each reference protein was tracked through all analyses for consistent assignment between pyrosequencing runs. Statistically significant differences (two-tailed P<0.05) in the expression levels (abundances) of nr-designated transcripts were determined in pairwise comparisons between data sets using the method of Audic and Claverie (1997, AC test), which accounts for variation in database size and assumes a Poisson distribution for the number of transcripts representing a given gene. P-values were adjusted using a false discovery rate (FDR) correction to account for potential false positives because of multiple comparisons, as in Benjamini and Hochberg (1995).

Nucleotide sequences generated in this study are available from the NCBI Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra) under accession # SRA012008.

Results and discussion

Pyrosequencing read statistics

A total of 1 786 949 sequence reads representing 350 Mbp over 10 pyrosequencing runs were generated from rRNA-subtracted and unsubtracted samples from depths of 25 and 75 m in the open ocean photic zone (Table 2). Reads with significant BLASTN hits (bit score >50) to either prokaryotic or eukaryotic rRNA sequences represented 80–88% of total reads in unsubtracted samples. Of these, bacterial rRNA represented 74–83% of total reads (Table 2), with archaeal and eukaryotic rRNA representing 0.9–1.4 and 3.5–4.8%, respectively. The high percentages of rRNA reads observed here are consistent with experimental evidence suggesting that rRNA accounts for 80–90% of total RNA in a typical bacterium (Wendisch et al., 2001). Indeed, upon transitioning to the 454 GS FLX-sequencing system, the percentage of rRNA observed in unsubtracted cDNA data sets derived from marine bacterioplankton has averaged 88% in our lab (range: 74–97%; n=20 FLX cDNA data sets), underscoring the necessity for an effective rRNA subtraction approach when using our linear amplification protocol.

rRNA subtraction

Subtractive hybridization to sample-specific rRNA probes lowered bacterial rRNA abundance by 40–58% relative to unsubtracted samples, reducing bacterial rRNA to 35–46% and total rRNA to 52–61% of pyrosequencing reads (Table 2, Figure 2, Supplementary Figure S3). In response, the fraction of non-rRNA reads in each data set increased up to fourfold, raising the proportion of reads with significant BLASTX hits to NCBI-nr proteins from 3.1–4.9% in unsubtracted samples to 7.3–20.4% in subtracted samples (Table 2). Bacterial 16S rRNA showed a greater proportional decrease in abundance than 23S rRNA following rRNA subtraction (Table 2), which may be due to a combination of broader coverage across diverse bacterial phyla by the 16S primers relative to the 23S primers, and to differences in the extent to which 16S and 23S primer sets span the full length of the target molecules (95% of 16S rRNA, 85% of 23S rRNA). As our probe sets targeted only bacterial rRNA, the relative percentages of archaeal and eukaryotic rRNA reads increased approximately three to fivefold (between 3.5–4.8% and 10.7–12.2% of total reads, respectively; Table 2). However, among those reads not identified as bacterial rRNA, the percentage represented by archaeal and eukaryotic rRNA reads did not differ substantially between unsubtracted (25–29%; mean: 23.8%) and subtracted (20–29%; mean: 26.6%) samples, suggesting little non-specific probe binding but emphasizing the need to implement additional probe sets to target these rRNA fractions.

The subtraction of rRNA occurred non-uniformly along the length of rRNA transcripts. For example, among the rRNA reads remaining in the subtracted sample, the proportion mapping to the central region of the 23S rRNA (bp 1000–2500) decreased relative to the unsubtracted sample, whereas the proportion mapping to the terminal region (bp 2500–2900) increased substantially (by 50%; Supplementary Figure S1). This pattern may be caused in part by the exclusion of the terminal 400 bp of the 2900 bp 23S rRNA gene by our probe set (Table 1), although a clear, but as of yet unexplained, bias toward 3′ 23S rRNA reads was also observed in the unsubtracted sample. A shift in the relative abundances of reads representing varying regions of the 16S rRNA was also apparent following subtraction, with a noticeable proportional increase in reads mapping to the 5′ 16S rRNA region (Supplementary Figure S1). These patterns indicate differential subtraction efficiencies along the rRNA transcripts, and therefore the potential that these molecules had been fragmented in the pre-subtracted samples.

After developing and testing the rRNA subtraction protocol described above, we have applied our method to a diverse range of samples and further optimized it to include primer sets targeting both archaeal and eukaryotic rRNAs (see Supplementary Table S3). Using this revised protocol, rRNA abundance in microbial community cDNAs has averaged 36% of total reads (range: 28–55%) across 10 different bacterioplankton samples from three distinct oceanic regions (Table 3; see Supplementary Figure S3 for a representative total RNA profile following domain-specific rRNA subtraction). We have also applied the method to a pure monoculture of actively growing Dokdonia sp. using 16S and 23S rRNA probe sets that perfectly match this target species. For this pure culture, our method successfully reduced rRNA to an average of 8% of the total cDNA reads (range: 3–11%; n=3; see Supplementary Table S3 for primers).

Table 3 rRNA abundance in metatranscriptomic profiles of diverse samples following subtractive hybridization

In addition to the method presented here, several commercial rRNA reduction protocols are also now available. One commonly used commercial product, the MICROBExpress Bacterial mRNA enrichment kit (Ambion), employs a subtractive hybridization to proprietary oligonucleotide probes, followed by rRNA removal via bead-immobilized capture oligonucleotides (in contrast with sample-specific, near full-length probes and biotin-streptavidin capture in our protocol). Although the MICROBExpress kit has been shown to be compatible for rRNA subtraction from a variety of bacteria, the efficiency of rRNA removal using this method can vary widely for community RNA samples (for example, Poretsky et al., 2005; McGrath et al., 2008; Hewson et al., 2009a, 2009b), as well as for single-species analyses (for example, Yoder-Himes et al., 2009). Indeed, oligonucleotide capture probes used in this method are predicted to be sensitive to target sequence variability known to be present in microbial community rRNAs, and the manufacturers explicitly state that the commercial kit is only partially compatible, or even incompatible, with a variety of microorganisms, including all Archaea (http://www.ambion.com/techlib/misc/microbe.html). Another commercial rRNA removal strategy, the mRNA-ONLY kit (Epicentre), uses specific exonucleases to selectively digest rRNA, but not mRNA. A potential complication of the exonuclease method in the mRNA-ONLY protocol is its potential to catalyze secondary, non-5′ monophosphate RNAse activity that may degrade specific mRNA transcripts in addition to the rRNA (Epicentre website). In a recent analysis of a microbial community associated with a Trichodesmium bloom, the mRNA-ONLY kit, used in conjunction with the MICROBExpress kit, was unable to reduce rRNA abundance below 94% of total FLX-based reads (Hewson et al., 2009b; Hewson I, personal communication), again suggesting potentially high variability in commercially available protocols. Recognizing the potential limitations of these protocols is important in selecting or developing rRNA subtraction methods for microbial transcriptome analyses.

Other alternatives to commercially available subtraction techniques have also been explored. Noting mixed results with the MICROBExpress method, McGrath et al. (2008) proposed physical removal of rRNA bands from RNA samples by agarose gel electrophoresis and extraction. Although effective, this technique requires large starting concentrations of RNA and may bias functional gene expression profiles by eliminating mRNA transcripts that co-migrate with the rRNA fraction. Gilbert et al. (2008) reported low rRNA representation in a metatranscriptomic survey following amplification of the cDNA via multiple displacement amplification (MDA) using the GenomiPHI V2 kit (GE Healthcare, Piscataway, NJ, USA). However, this kit is optimized for amplification of genomic DNA, and has not been extensively tested for reproducibility and bias when used for amplification of short cDNA fragments. Indeed, numerous studies have shown that even for single template genomic DNA, MDA using φ29 polymerase amplifies unevenly across different loci, suggesting that the reliability and utility of MDA for quantitative analyses is uncertain at best (Bergen et al., 2005; Ballantyne et al., 2007). Recently, Armour et al. (2009) proposed a novel rRNA depletion method based on the use of a computationally defined subset of hexameric primers that exclude target rRNA sequences during cDNA synthesis from total RNA. Although potentially effective in low-complexity samples or single-species monocultures, this method, as for most published rRNA subtraction procedures, is not easily extrapolated to complex metatranscriptomic samples with diverse rRNA pools. A definitive comparison of rRNA removal protocols from previously published metatranscriptomic studies is not feasible, as previous studies did not include unsubtracted controls for comparison, used distinct cDNA synthesis and downstream pyrosequencing preparation protocols and analyzed microbial communities with differing compositions.

The protocol we describe here offers some potential advantages over the existing rRNA-subtraction procedures noted above. The method can be easily tailored to synthesize sample- or taxon-specific probes targeting either specific strains or a broad array of archaeal, bacterial and eukaryotic rRNAs. Although we chose to target only bacterial rRNA for the initial development of this method, the protocol can be expanded to include both archaeal and eukaryotic rRNA probe sets generated using the broad-specificity primers listed in Supplementary Table S3. The method can also be applied to less complex, non-environmental samples (for example, experimental cultures or consortia). If necessary for such samples, taxon-specific primers can be used in place of universal primers for probe generation, as shown for a Dokdonia culture analyzed in our lab (Table 3 and S3). When followed by a linear RNA amplification step, this subtractive protocol can be used with relatively small amounts of starting material; in our hands, subtraction of rRNA from as little as 20 ng total RNA has yielded amounts of mRNA-enriched template sufficient for amplification and pyrosequencing. As the protocol implemented here biases the composition of any rRNA reads remaining after subtraction (as does any rRNA subtraction procedure), our method is specifically designed for maximizing coverage of the functional RNA pool, potentially identifying unique or interesting transcripts that can inform or suggest more targeted gene-specific studies to follow.

Replicate reads

Sequencing artifacts in the form of replicated sequences are a common source of error in pyrosequencing data sets (Briggs et al., 2009; Gomez-Alvarez et al., 2009). Careful identification and removal of such sequences, which can account for more than 30% of pyrosequencing reads (Gomez-Alvarez et al., 2009), is therefore essential for accurately extrapolating observed read abundance to in situ DNA or RNA abundance. Replicates are hypothesized to originate during the emulsion PCR step of pyrosequencing. Incomplete emulsion can result in the attachment of a single PCR product to multiple beads (Briggs et al., 2009; Gomez-Alvarez et al., 2009), which upon sequencing results in an increase in the number of observed occurrences of that sequence in the data set. In an idealized case, sequences generated from identical, replicated template molecules would share start and stop sites, resulting in identical sequence along their full length. However, sequencing errors and quality trimming result in the generation of non-identical sequences of different lengths from identical template molecules. The challenge therefore is to differentiate imperfectly sequenced artificial replicates from legitimate sequences derived from multiple, similar DNA fragments.

Gomez-Alvarez et al. (2009) calculated that, for the genomic DNA of an idealized microbial community, the probability of multiple reads starting at the same position in a randomly sheared metagenome analysis is extremely low (1 × 10−10). These authors therefore identify artificial replicates as sequences (of potentially varying lengths) sharing greater than 90% nucleotide identity and having identical beginning sequences (over the first 3 bp). This definition is likely robust for microbial community DNA, but is perhaps overly conservative for microbial community transcriptomes. Unlike random genomic DNA fragments, transcripts have clearly defined start and stop sites and routinely occur in multiple copies per cell. As a result, criteria for identifying replicates from metatranscriptomes must, to the greatest extent possible, distinguish artifacts from legitimate re-sampling of multiple transcripts originating from the same gene.

We analyzed pairs of sequencing runs (independent emulsion PCR and sequencing reactions generated from the same adaptor-ligated template library) to establish criteria for replicate removal from our transcriptomic libraries, identifying replicates as sequences differing by no more than 1 bp in length, sharing 99% nucleotide identity, and having identical start sites (first 3 bp) (see Supplementary Online Text, Supplementary Table S1 and Figure S2). Using these criteria, re-sampled transcripts were rare (0.3–3.4% of sequences) compared with artificial replicate transcripts (6–39% of sequences). Though clusters of replicate reads were identified in all samples examined, the percentage of replicated reads to total reads (replicate frequency) varied greatly among the 10 different pyrosequencing runs (Table 2). Substantial variation occurred even between multiple runs generated via independent emulsion PCR and sequencing reactions from the same template library. Notably, replicate frequency varied between runs from 9% to 38% and 14–36% in the rRNA-subtracted and unsubtracted 75 m A samples, respectively, and from 11% to 39% between runs of the unsubtracted 75 m B sample (Table 2). In each of these comparisons, replicate frequency was negatively correlated with the number of total reads per run, suggesting a link between artifact generation and the efficiency of the emulsion PCR and pyrosequencing steps. This immediately suggested important criteria for quality control and assessment of individual sequencing runs.

The presence of these replicates can bias the apparent diversity of the transcript pool. For example, before removal of replicate sequences the slopes of rarefaction curves describing the number of unique non-redundant (nr) peptide reference genes identified per data set as a function of sequencing depth differed notably between repeated runs (1 and 2) of the subtracted 75 m A samples (Table 2; Figure 3). Removal of replicate sequences eliminated this discrepancy, yielding highly similar rarefaction curves. Together, these data indicate that artificial replicates are a common and potentially biasing component of pyrosequencing-based transcriptomic data sets. However, effective criteria for replicate removal may vary among samples of differing taxonomic complexity and functional state. Though not yet routine in pyrosequencing-based analyses, systematic comparisons of technically repeated pyrosequencing runs may help delineate sample-specific criteria for replicate removal.

Figure 3
figure 3

Rarefaction curve for HOT-186 75 m sample AS. The number of unique nr reference genes identified via BLASTX (at e-values 1 × 10−5) is shown as a function of sequencing depth. Runs 1 and 2 represent multiple pyrosequencing runs from the same adapter-ligated sample library, before and after removal of replicate sequences.

Statistical comparison of transcript abundances in technical replicates

Validation of rRNA subtraction and computational replicate removal methods required statistical determination of changes in transcript abundance with a high degree of taxonomic resolution. Statistical comparisons of metagenomic and transcriptomic profiles have largely focused on gene clusters and functional groupings rather than individual genes (Rodriguez-Brito et al., 2006; Frias-Lopez et al., 2008; Huson et al., 2009; Poretsky et al., 2009). However, this approach potentially lacks the resolution to detect changes in expression of specific transcripts (for example, due to the non-specific binding of rRNA probe to an mRNA transcript during subtractive hybridization). Here, to assess the impact of rRNA subtraction and replicate removal on expression profiles, non-rRNA transcripts were mapped to specific protein sequences in the NCBI-nr database, and the relative abundances of transcripts matching single reference genes were compared between data sets (Table 4, Figures 4 and 5). Dalevi et al. (2008) showed that a similar mapping approach accurately represented the functional and taxonomic characteristics of 100 bp DNA fragments; we expect an even greater degree of assignment accuracy given our longer average read length (>200 bp). A statistical test from the expressed sequence tag literature was then applied to identify differentially represented reference genes (AC test; Audic and Claverie 1997), along with a FDR minimizing test (Benjamini and Hochberg, 1995) to correct for the large number of comparisons (P<0.05). We used this test to examine the reproducibility of transcriptional profiles generated by pyrosequencing, and the effect of rRNA subtraction on the abundance on non-rRNA transcripts.

Table 4 Dataset (DS) comparisons—non-rRNA sequences mapped to non-redundant (nr) NCBI reference sequences
Figure 4
figure 4

Relative abundance of NCBI-nr reference genes in rRNA-subtracted pyrosequencing 75 m A data sets. Reference genes representing >0.1% of the 75 m A library are shown in descending order. Their abundance in a replicate library (75 m B rRNA-subtracted) is shown in red. Reference genes with significantly different abundances are labeled with a FDR-corrected P-value.

Figure 5
figure 5

Relative abundance of NCBI-nr reference genes in HOT-186 25 m pyrosequencing data sets with and without rRNA subtraction. Reference genes representing >0.1% of the unsubtracted library are shown in descending order. Their abundance in the library generated from post-subtraction RNA is shown in red. Reference genes with significantly different abundances are labeled with a FDR-corrected P-value.

Removal of replicate sequences was key to the generation of reproducible transcriptional profiles (see starred data sets in Table 4). For example, the variation between rarefaction curves generated from the raw sequencing data from the rRNA-subtracted 75 m A sample (Figure 3) was also reflected in reference gene abundances; 144 genes were identified as ‘differentially expressed’ between the two sequencing runs (Table 4). Removal of replicate sequences reduced this to just 30 references with significantly different abundances. The majority of these differences involved genes represented by very few (or zero) reads in one of the two data sets. Replicate removal may therefore have the most significant effect on apparent differences between low abundance transcripts, for which even small numbers of artificial replicates may have a disproportionately large effect on apparent expression level.

A thorough analysis of the reproducibility of metatranscriptomic profiles must also take into account experimental variation, as the generation of transcriptomic libraries from extracted total RNA requires extensive processing steps, including linear amplification in our procedure that might introduce variability into the observed transcriptional profile. To address this issue, we subdivided the total RNA sample from 75 m to generate four samples (unsubtracted and rRNA-subtracted samples A and B), which were processed in parallel to examine the reproducibility of resultant transcriptomic data sets. For the unsubtracted 75 m sample, no nr reference genes were differentially expressed between replicates A and B (Table 4). For the rRNA-subtracted 75 m sample, following removal of replicate reads, only 0.03% (7 of 25 174) of total identified references were differentially expressed (Table 4; Supplementary Table S1). Of those genes represented by greater than 0.1% of the sequence reads in the subtracted 75 m A data set (n=42), only three varied in expression between replicates (Table 5, Figure 4). As observed in comparisons of replicate runs from a single template library, the majority of the significant differences between the full technical replicates involved relatively low abundance reference genes having significantly higher representation in one of the two data sets (Table 5).

Table 5 NCBI-nr reference genes differing significantly in abundance between technical replicates A and B of the rRNA-subtracted HOT-186 75 m sample

Statistical comparison of rRNA-subtracted and unsubtracted samples

Having established the reproducibility of transcriptional profiles generated during metatranscriptomic analyses, we examined bias potentially introduced during the removal of rRNA by subtractive hybridization. Subtractive protocols have the potential to alter functional gene expression profiles if removal of non-rRNA transcripts occur because of non-specific probe binding. In the 25 m sample, only 0.03% of total nr reference genes (13 of 48 090 total, which includes 2378 shared between data sets +45 712 unique to either data set) showed significantly different abundances between subtracted and unsubtracted samples (Table 4). Of these, five were represented at abundances greater than 0.1% of total reads in the unsubtracted data set (Figure 5). In the 75 m sample, only 0.01% (3 of 32 340) were differentially represented. These low levels of variation are within the range observed between replicate pyrosequencing runs derived from independent emulsion PCRs using template from the same adapter-ligated library (0–0.2% of genes differentially expressed for comparisons 25 m S1 vs S2, 75 m AS1 vs AS2, AU1 vs AU2, BU1 vs BU2, Table 4). This suggests that these differences are due to stochastic variation introduced during pyrosequencing, and that the subtractive hybridization protocol does not significantly alter the apparent expression profile generated from transcript libraries.

Conclusions

A distinguishing characteristic of community DNA and RNA sequencing efforts is the high complexity of the resulting sequence data. For example, in this study, pyrosequencing of a single sample of seawater collected at 25 m generated 266 859 unique non-rRNA sequences (including runs from both rRNA-subtracted and unsubtracted libraries), 117 809 (44%) of which had significant hits to 48 090 unique nr reference proteins. An initial, half-plate run derived from the RNA-subtracted portion of the 25 m sample yielded 21 011 nr references, of which only 17% were recaptured in a second full-plate run, which contained nearly twice as many sequences (Table 4). Indeed, no two data sets analyzed in this study, including data sets generated from the same single-stranded library, shared more than 17% of unique nr reference genes (range: 5–17%). This suggests that the vast majority of unique transcripts in the RNA pool may be un-sampled due to stochastic variation in the pyrosequencing procedure. Sequencing depth can therefore clearly limit the analytical capability of metatranscriptomic analyses. Deeper sequencing not only increases the likelihood of sampling novel transcripts but also facilitates statistically significant comparisons of transcripts appearing across multiple data sets (for example, experimental treatments). This is particularly relevant for low-abundance transcripts that constitute the majority of diversity in the mRNA pool.

Though changes in the transcriptional profiles of highly expressed genes yield important insight into microbial communities (for example, Frias-Lopez et al., 2008; Hewson et al., 2009b), the relatively shallow extent to which pyrosequencing captures gene-level diversity among functional transcripts increases the need to optimize message recovery. Ribosomal RNA represented 80–88% of sequences recovered from the ocean samples analyzed in this study. Given the mRNA transcript diversity encountered here, the presence of rRNA at such high proportions hinders the detection of potentially tens to hundreds of thousands of unique functional transcripts via a standard 454 pyrosequencing run. Even incremental reductions in the rRNA pool can therefore substantially increase our knowledge of genes expressed at low frequency. When analyzing expression at such resolution, it becomes increasingly important to distinguish sequencing artifacts from genuine variation in transcript abundance. This analysis, along with other recent studies (for example, Gomez-Alvarez et al., 2009), highlights the need for meaningful criteria for identifying and removing artificially replicated sequences that confound statistical comparisons of expression. Comparisons between technically replicated libraries constitute an effective method for establishing such criteria and should become more commonplace in pyrosequencing-based analyses. Upon removal of sequencing artifacts, transcriptomic pyrosequencing data sets appear highly reproducible and, in conjunction with rRNA-subtraction methods that maximize message recovery, can provide new insights into the diversity and dynamics of less abundant transcripts. This is particularly relevant as microbial metatranscriptomics is increasingly used to monitor community responses to experimentally induced perturbations, some of which may elicit subtle, but important, functional changes in non-dominant community members.