The reproducibility of pyrotag sequencing in our workflow was generally high, with phylum/class-level read abundances across biological replicates retrieved at an average standard deviation (SD) of ~0.6% (, maximum SD of 4.3%). Also at a higher, family- or genus-level resolution, selected dominating members of the aquifer microbial community were retrieved at comparable reproducibility (average SD of ~0.9%, max. SD of 2.7%; ). Over all different taxonomic levels and abundances of taxa defined by the RDP classifier, abundant ones (>10% abundance, typically phylum, class, or order-level resolution) were recovered with an average SD of ~1.7% read abundance (). Less abundant taxa were retrieved at comparable relative reproducibility, e.g. with an average SD of 0.9, 0.2 or 0.04% for taxa below either 10%, 2%, or 0.3% read abundance, respectively. Reproducibility was comparable, or even better for technical replicates, where phylum-level max. SD was 2% read abundance (supporting Figure S1
Reproducibility of pyrotag read abundance over biological replicates of aquifer DNA extracts.
Read abundance standard deviation (SD) for dominating and less abundant taxa in pyrotag libraries.
Linkage clustering-based comparison of OTU occurrence (97% sequence similarity) showed that 45±5% of OTUs in each library were present as singletons (only one read per library). Nevertheless, OTUs overlap (Sørensen similarity) for triplicated pyrotag libraries was high, and at 88, 82, and 67% for sediments of 2006, 08 and 09, respectively. This is in contrast to a much lower OTU overlap reported recently for replicated pyrotag libraries from soil 
. A number of different factors may have contributed to this distinction: the soil samples may have harboured a substantially more diverse bacterial community than our aquifer sediments, making effects of undersampling in pyrotag libraries more severe. Second, that study used post-PCR ligation of 454 sequencing adaptors and ‘classical’ 454 sequencing, while our amplification primers already contained sequencing adaptor tags and Titanium chemistry was used. The comparative effects of both on pyrotag library reproducibility have not been specifically addressed to date. Last but not least, our workflow integrates an initial quality trimming step of sequencing reads based on base calling confidence scores 
, the effects of which on the reduction of sequencing noise and overall library similarity should not be underestimated.
In our hands, total OTU overlap of pyrotag libraries from technical replicates was even higher (>96%). Thus to a certain extent, also singleton OTUs seemed reproducible over replicated libraries. We can cautiously speculate on whether such singletons already represent members of the rare biosphere, however much ‘deeper’ sequencing (i.e. >100.000 s, not ~10.000 reads per library) would have been necessary to truly address this question. We expect that while ‘deeper’ sequencing would not have altered the reproducibility of abundant OTUs, overall diversity and also the stochastic appearance of rare OTUs would have increased. Also Shannon diversity H’ was highly reproducible for biological replicates of pyrotag libraries and comparable to H’ reproducibility obtained in T-RFLP fingerprinting of the same samples (). However, total T-RF diversity was only ~50% of pyrotag OTU diversity, which was expected, since taxon-specific resolution of pyrotag sequencing is much higher than fingerprinting.
Reproducibility and comparison of Shannon diversity (H’) and selected OTU abundance in T-RFLP fingerprinting and pyrotag libraries.
As our workflow allows for the linking of read abundances within defined assembled contigs to that of specific ‘in vivo
’ T-RFs, we compared OTU abundance retrieved via both methods for representative samples ( and ). With this unique approach, it was possible to demonstrate the highly similar community structure and OTU abundance patterns retrieved by both methods. In essence, it becomes clear that both T-RFLP and pyrotag sequencing are capable of recovering the same amplicon pools from environmental samples, and yield highly comparable overall microbial community patterns. As expected, T-RFs predicted in silico
vs. those measured in vivo
mostly differed by a few bp 
, which prevented a straightforward calculation of overall community similarity e.g. via Sørensen OTU overlap or comparable indices. Nevertheless, the functional organisation (Fo
) of bacterial communities as inferred from Pareto-Lorentz curves of cumulative OTU abundances 
for both approaches supported highly similar overall community structure. Thus, Fo
was 0.76 vs. 0.71 for in vivo
T-RFs vs. pyrotag contigs in 2006, and 0.53 vs. 0.54 in 2009. In 2008 however, inferred Fo
was higher for T-RFs than for pyrotag contigs (0.74 vs. 0.59), which can be explained by the exceptionally high yield of reads recovered in that specific pyrotag library (Table S1
), resulting in more OTUs from assembled contigs passing our subjectively defined 20 reads-per-contig threshold. A library-specific cut-off for defining contig assembly read thresholds could help to alleviate this limitation of our present workflow.
Comparison of bacterial community structure as recovered in T-RFLP fingerprinting and pyrotag libraries.
The most significant difference in T-RF patterns of both methods was related to an abundant 282 bp T-RF found ‘in vivo’
in the 2008 sample, but not predicted ‘in silico
’ for any dominating pyrotag contig (). Upon second consideration, this was identified as a pseudo T-RF 
of abundant Spirochaetes
-related templates in this sample, for which a secondary Msp
I restriction site at 285 bp was predicted, succeeding their primary restriction site at 208 bp. Thus it is plausible that this population was actually represented by both the 204 and 282 bp T-RFs detected ‘in vivo
’, with the second one being a pseudo T-RF. Other, minor differences were related to a higher relative abundance of specific ‘in vivo
’ T-RFs compared to the respective pyrotag contigs. This was expected, since it is known that several phylogenetically distinct rRNA gene populations may share identical terminal restriction sites, and thus contribute to the same ‘in vivo
’ T-RFs. In fact, especially the 159 bp T-RF, representing a population of primary toluene degraders (unclassified Desulfobulbaceae
) in situ
, was up to 2-fold more abundant in fingerprinting than in pyrotag sequencing (). Thus here, additional bacterial populations may have contributed to this T-RF. For other important populations at the site such as Geobacter
spp. and the Comamonadaceae
, the abundance of the respective T-RFs (129 bp; 137 & 228 bp; 486 bp) was more similar to pyrotag read abundance. Nevertheless, also the fact that different reverse primers were used in both approaches (Ba519r for pyrotags vs. 907r for T-RFLP) may have introduced further distinctions in the recovery of community structure cannot be excluded.
Amplicon pyrosequencing seems to hold also the potential for a quantitative recovery of template input ratios. Spiked A. fisheri
reads were retrieved with linearity (R2
0.99) over three orders of magnitude reflecting qPCR-defined amendment ratios in a reliable manner (). The maximum amendment (20%) of external A. fisheri
DNA caused decreases in abundance of maximally ~2% for other important intrinsic taxa (Figure S2
). This linear representation is again in contradiction to the apparently random recovery of template amendment ratios recently reported 
. In that study, 0.1% of genomic DNA of S. oneidensis
was spiked to soil DNA. Very likely, the use of qPCR guided template spiking differentiate our results. It is clear that DNA with quantified 16S rRNA gene content can be spiked in a much more meaningful manner than amendments guided by bulk DNA quantification. qPCR-defined template mixtures have already been crucial earlier, in demonstrating the semi-quantitative robustness of rRNA gene-based T-RFLP fingerprinting 
. Nevertheless, also the possibly pronounced impact of our initial pyrotag data quality control measures (confidence trimming) on template abundances, linear amendment recovery, and overall library similarity cannot be ignored.
Semi-quantitative recovery of spiked A. fisheri 16S rRNA genes in pyrotag libraries.
Finally, we want to mention that our comparison of 1- and 2-step PCR for the generation of pyrotag libraries did not produce pronounced distinctions in read abundance of dominating lineages and overall community structure (Figure S2
). In accordance to a recent report 
, 2-step pyrotag libraries were significantly more diverse (Supplementary Table S2
). In our hands, 2-step pyrotag libraries also contained ~5 times higher ratios of shorter reads (<250 bp). We are still examining whether this phenomenon could potentially be connected to the apparently increased diversity of 2-step libraries. However, the 2-step PCR did not affect the semi-quantitative recovery of A. fisheri
16S rRNA gene sequences (Figure S2
Since each template of our study was screened using different MID adaptors and also variable pools of amplicon preparation and mixing (Supplementary Table S1
), the highly reproducible read abundances obtained across biological and technical replicates seem to suggest that none of the used MID adaptors introduced a systematic, adaptor specific bias in overall community structure, or that all were connected to a similar bias. Nevertheless, since we did not systematically compare 1-step vs. 2-step amplicon libraries for technical replicates generated with the same MID adaptors, these results do not exclude potential biases introduced by specific MID adaptors.
In conclusion, our study demonstrates that 454 pyrotag sequencing is a robust and reproducible method for the reliable recovery of the diversity and structure of complex natural microbial communities, with a reproducibility certainly comparable to that of established screening tools such as T-RFLP fingerprinting. As for every other analytical technique, biological and technical replication is essential to obtain an accurate measure of semi-quantitative results 
. Nevertheless, each of our pyrotag libraries showed consistent read abundances with subsequent replicate means, and the most important community distinctions were recovered also in non-replicate libraries. The possibility of spiking quantitative template amendments to pyrotag libraries brings new exciting applications in the study of complex microbial communities. Amendment with multiple standards not present in the sample, or even with synthetic DNA, could be an interesting future development.