PCR-Induced Sequence Artifacts and Bias: Insights from Comparison of Two 16S rRNA Clone Libraries Constructed from the Same Sample

Abstract

The contribution of PCR artifacts to 16S rRNA gene sequence diversity from a complex bacterioplankton sample was estimated. Taq DNA polymerase errors were found to be the dominant sequence artifact but could be constrained by clustering the sequences into 99% sequence similarity groups. Other artifacts (chimeras and heteroduplex molecules) were significantly reduced by employing modified amplification protocols. Surprisingly, no skew in sequence types was detected in the two libraries constructed from PCR products amplified for different numbers of cycles. Recommendations for modification of amplification protocols and for reporting diversity estimates at 99% sequence similarity as a standard are given.

Estimation of the extent of PCR-induced artifacts in microbial diversity studies remains an important task in the search for patterns and extent of microbial diversity. The basic types of PCR artifacts have been shown in controlled laboratory studies and can be divided into two categories: those resulting in sequence artifacts (PCR errors), and those skewing the distribution of PCR products due to unequal amplification (PCR bias) or cloning efficiency. Sequence artifacts may arise due to (i) the formation of chimerical molecules (3, 10, 14, 15, 25, 26, 37, 38), (ii) the formation of heteroduplex molecules (25, 27, 29, 32), and (iii) Taq DNA polymerase error (4, 25). PCR bias is thought to be due to intrinsic differences in the amplification efficiency of templates (23) or to the inhibition of amplification by the self-annealing of the most abundant templates in the late stages of amplification (31). However, it remains difficult to translate these results to environmental samples in which target genes are orders of magnitude more highly concentrated than in the simple mixtures of templates generally used in controlled laboratory studies.

Here, we address the following questions. (i) To what extent do different PCR errors contribute to overestimation of microbial diversity? (ii) Do these PCR errors suggest differences in community structure? (iii) To what extent does PCR bias result in different template distributions after various cycle numbers? Finally, we derive and reiterate recommendations to minimize PCR artifacts.

We have recently generated two large 16S rRNA gene libraries (∼1,000 sequences each) from a single bacterioplankton sample (1), providing an opportunity to evaluate PCR artifacts in a realistic setting. The first (standard) library was constructed using 35-cycle amplification to mimic commonly used protocols. The second (modified) library was based on the following amplification protocol to reduce the accumulation of PCR artifacts: limitation to 15 cycles of amplification to decrease PCR bias (23) and accumulation of Taq DNA polymerase errors and chimerical sequence formation (25), followed by 3 additional cycles in a fresh reaction mixture (reconditioning PCR step) to minimize the formation of heteroduplex and Taq DNA polymerase errors (32). In addition, we identified Taq DNA polymerase errors in sequences from the modified library by manual reconstruction of 16S rRNA secondary structures (1). This accounted for the occurrence of Taq DNA polymerase errors to a high degree, with an error rate of 3.3 × 10⁻⁵ per nucleotide per duplication, which approximated the theoretically expected error rate of 2 × 10⁻⁵ per nucleotide per duplication for the Taq DNA polymerase used (1). Finally, a combination of three bioinformatics tools was used to identify putative chimeras in both libraries (1).

Table 1 demonstrates the strong effect of a decrease in cycle numbers from 35 to 15 + 3 on the accumulation of PCR artifacts (1). This is evident from several results, such as (i) a decrease in unique 16S rRNA sequences (ribotypes) from 76% to 48% in the standard and modified libraries after correction for chimeras and Taq DNA polymerase errors (1); (ii) a greater-than-twofold decrease in sequence diversity, from 3,881 to 1,633 sequences, among libraries based on the Chao-1 nonparametric richness estimator (1); and (iii) an increase in coverage index from 24% in the standard library to 64% in the modified library (Table 1), suggesting that much fewer clones would have to be sequenced to obtain a representative sample of the modified library.

TABLE 1.

Compositions of the standard and modified 16S rRNA gene clone libraries

Clone library	No. of PCR cycles	No. of sequences analyzed	% Chimeric sequences	No. of ribotypes (100%; 99%a)	Estimated total no. of sequencesb (100%; 99%a)	% Unique 16S rRNA sequences (100%)	Library coveragec (100%; 99%a)
Standard	35	909	13	692; 302	3,881; 962	76	24; 66.5
Modified (+)d	18	1,131	NDg	686; 326	3,087; 699	61	39.4; 71.2
Modified (−)e	18	1,033f	3	516; 240	1,633; 520	48	64; 89.3

^aValues calculated for 100 and 99% sequence similarity cutoffs, respectively.

^bThe total number of sequences was estimated using the Chao-1 nonparametric richness estimator (11).

^cLibrary coverage was calculated by the method of Good (9).

^dModified library (+) data set analyzed before removal of chimeras and Taq error correction.

^eModified library (−) data set analyzed after removal of chimeras and Taq error correction.

^fNumber of sequences after removal of chimeric sequences and sequences shorter than 600 bp used for rarefaction and phylogenetic analyses. Thus, the number differs slightly from those in Table 2, which includes shorter sequences.

^gND, not determined.

An important question is to what extent chimeras and heteroduplex molecules contribute to diversity estimates in clone libraries, as these represent amalgams of existing sequences that would be scored as novel lineages. While our experimental design does not allow differentiation between chimeras and heteroduplex molecules, we have shown previously that the inclusion of a “reconditioning” step reduces the occurrence of heteroduplex amplicons to a negligible level (undetectable at a detection level of 1%) (32). Moreover, the combination of three bioinformatics tools suggests that the incidence of chimeras dropped from 13% in the standard library to only 3% in the modified library (Table 1), again suggesting a strong effect of modified amplification protocols (32).

The strong effect of Taq DNA polymerase errors can be seen from a comparison of the frequencies of shared and unique sequence types among the two-16S rRNA gene clone libraries (Table 2). For example, there was an almost-twofold-higher incidence of singletons (unique sequences occurring only once) in the standard library (61.5%) compared to that in the modified library (36%) (Table 2). Indeed, if sequences are clustered into 99% consensus groups, ∼80% of lineages are shared among the libraries (Table 2).

TABLE 2.

Occurrences of shared and unique sequence types within the two 16S rRNA gene clone libraries

Clone library	No. of sequences analyzed	% (no.) of repeated sequences shared within libraries at a sequence similarity cutoff of:		% (no.) of singleton sequences at a sequence similarity cutoff ofa:		% (no.) of unique clusters at 100% similarity cutoffb
		100%	99%	100%	99%
Standard	902	32.5 (293)	80 (722)	61.5 (555)	13.6 (122)	6 (54)
Modified	1,082	48 (520)	81 (876)	36 (389)	10.7 (116)	16 (173)

^aSingleton sequence, single sequence unique to one of the libraries.

^bUnique cluster, cluster containing more than one sequence unique to either one of the libraries.

Given that the two libraries showed considerable differences in sequence diversity, we explored whether a statistical comparison would identify both libraries as being drawn from the same sample. Coverage curves generated with the LIBSHUFF program (Fig. 1) indicated that the modified library would be judged significantly different from the standard library (P = 0.001) and that therefore both libraries would be interpreted as samples from different communities (P < 0.05). However, the analysis demonstrates that the differences are due to genetic distances of less than 0.01 and that both libraries share most phylogenetic lineages at evolutionary distances (D) greater than 0.02. This result supports the likelihood that Taq DNA polymerase errors are the major contributor to artificial sequence divergence. Similar results arose from lineage-versus-time plots (data not shown). We conclude that the incidence of Taq DNA polymerase errors is sufficiently accounted for by the clustering of sequences into 99% identity groups, and we recommend that sequence diversity be reported at this similarity cutoff.

FIG. 1.

LIBSHUFF coverage curve comparison for the modified and standard libraries. Shown are the homologous coverage curve (Cx) of the modified library (squares), the heterologous coverage curve (Cxy) of the modified library versus the standard library (triangles), and the value of (Cx-Cxy)² (crosses) for the samples at each D value. D is equal to the Jukes-Cantor evolutionary distance determined using PAUP software. Filled circles indicate the P value of 0.05 for (Cx-Cxy)² for the randomized samples.

The construction of two well-sampled libraries from the same sample gave us the unique opportunity to examine the effect of PCR cycle numbers on the relative distribution of sequence types (i.e., PCR bias). We examined the relative abundances of distinct phylogenetic groups in both libraries by (i) grouping of sequences using neighbor-joining and parsimony methods implemented in the ARB sequence analysis package (17) and (ii) statistical comparison of differences in taxonomic composition of the two libraries using the Ribosomal Database Project Classifier (5). Surprisingly, neither approach showed a significant difference in distribution of major phylogenetic lineages among the libraries (Fig. 2; see also S3 in the supplemental material). The three most abundant groups (Bacteroidetes, α-Proteobacteria, and γ-Proteobacteria) were all similarly represented (32.8% versus 32.5%, 29.3% versus 28.7%, and 29.3% versus 28.7%, respectively) (Fig. 2). These three groups comprise more than 80% of the total clones in each library. However, even the less-well-represented Actinobacteria showed only very minor differences in abundance, constituting 8 and 5.5% of the total clones (Fig. 2). Well-represented clades such as SAR11 or Roseobacter, which allow comparison with finer taxonomic resolution, also display similar distributions among the libraries (see S3 in the supplemental material).

FIG. 2.

Relative frequency distribution of major phylogenetic groups detected among the environmental 16S rRNA sequences from the standard (gray bars) and modified (black bars) libraries, respectively.

In order to explore potential reasons for the unexpected similarity of the relative abundances of the ribotypes in the two libraries, we investigated whether the amplification process essentially stopped soon after the 18th cycle due to reagent limitation. Using quantitative PCR (qPCR), we determined the kinetics of product accumulation mimicking PCR conditions utilized for clone library construction (100 nM primer concentration and 5 to 10 ng DNA) (see S4 in the supplemental material). The results show that the reaction was saturated around the 30th cycle, indicating that the modified library underwent ∼12 additional cycles of amplification (see S4 in the supplemental material). Because the slopes of the product accumulation curves indicate that the amplification efficiency was between 83 and 88%, and assuming an initial target concentration of ∼10⁷ rRNA gene copies (∼4.9 × 10⁶ genomic templates × ∼2.5 rRNA operons per genome [2]), these 12 extra cycles led to an approximately 1,410- to 1,949-fold-higher product concentration. Thus, even a minor difference in amplification efficiency among different templates would have resulted in a noticeable difference in ribotype abundance in the two libraries.

PCR bias was previously attributed to intrinsic differences in amplification efficiencies as a consequence of differences in primer binding energy (12, 23, 35, 36) and to inhibition of amplification due to the reannealing of templates that occurs once they reach saturation concentration (30, 31). Indeed, a primer binding efficiency different from that of genomic templates may still play a significant role in biased amplification but may be lessened during later cycles when amplification proceeds primarily from PCR amplicons, which display a perfect match to the primers. Further, Suzuki et al. already discussed the possibility that PCR bias due to reannealing should be small in environmental DNA samples composed of highly diverse templates because no template may reach saturation concentration (30, 31). Thus, it may not be surprising that several other studies comparing amplicon distributions in community fingerprints (terminal restriction fragment length polymorphism and automated ribosomal intergenic spacer analyses) generated at different cycles similarly suggested an absence of bias (7, 18, 22).

These considerations thus focus attention on the first few cycles of the PCR as a major source of bias to libraries generated from complex environmental samples.

Although we have no direct way of estimating bias during the first few cycles, the distribution of different taxa observed in our clone libraries is strikingly similar to previous quantitative estimates of different bacterioplankton groups. For example, representatives of the Cytophaga-Flavobacterium-Bacteroides group (Bacteroidetes) accounted for 32% of the sequences in our library, while they were previously detected at ∼30% by fluorescent in situ hybridization (FISH) in other coastal samples (6, 13). Similarly, the SAR11 and Roseobacter clades comprised 12% and 8% of the total clones in the libraries, respectively. These values are within the ranges observed for 16S rRNA genes based on genomic shotgun sequencing of Sargasso Sea bacterioplankton, where the SAR11 group comprised 4.7 to 37.3% (20, 34), and in FISH enumerations (1 to 32%) (21). The Roseobacter clade accounted for 10 to 40% of the total bacterioplankton cells determined by FISH analysis (8) and between 7 to 20% of the total bacterial counts determined using a dilution culture approach in combination with other molecular methods (28). For less-abundant taxa, we confirmed the abundance of Vibrio-related sequences in our library by qPCR estimation in the same environment. These made up a total of 1.8% of the clones, which corresponds well to their abundance within total bacterioplankton (33). Nonetheless, the current data stem from a single experimental system with a single primer set, and there are cases where results from qPCR or probing did not agree with the distribution of clones in libraries in similarly complex systems (19, 24).

We summarize and reiterate the following suggestions for minimizing PCR artifacts in environmental clone library construction. (i) To minimize PCR drift, several replicate PCR amplifications should be combined (23, 36). (ii) To minimize chimeras and Taq DNA polymerase errors, the smallest possible number of PCR amplification cycles should be carried out (e.g., until a band is barely visible on agarose gels) (23, 30, 31). (iii) To reduce PCR bias, high ramp rates between the denaturation and annealing steps and low annealing temperatures should be used, while long extension times (>180 s) should be avoided (12, 16). (iv) To minimize the presence of heteroduplex molecules, a reconditioning PCR step (i.e., three to five additional PCR cycles using fresh reagent mixture) should be included (32). Finally, our analysis suggests that Taq DNA polymerase errors are sufficiently constrained when sequences are clustered into 99% similarity groups so that these groups should always be reported in estimates of sequence diversity, particularly since higher similarity cutoffs (e.g., the commonly used 97%) may mask microdiverse clusters, which we have recently suggested as representing important units of differentiation among marine bacterioplankton (1).