Mock community and vaginal microbiome sequencing

A synthetic microbial community (the mixed vaginal panel, MVP) was generated as previously described [28] by combining 20 cloned cpn60 barcode sequences from vaginal isolates in an equimolar mixture (S1 File). Plasmid DNA was extracted from E. coli with the EZ-10 Spin Column Plasmid DNA Miniprep Kit (Bio Basic Inc., Markham ON, Canada). Purified plasmid DNA was quantified and diluted to 1 x 10⁹ plasmid copies/μL with 10 mM Tris pH 8.5. Five μL of each plasmid solution was pooled to form the synthetic community. cpn60 UT PCR was performed in duplicate with a 1:3 mixture of Illumina adapted primers M729/M280:M1612/M1613 (S1 Table) as previously described [29]. Sequence library preparation was conducted according to the Illumina MiSeq 16S Metagenomic Sequencing Library Preparation (Illumina Inc., San Diego, CA) protocol with some minor changes as optimized for the cpn60 UT: in PCR clean-up steps, 32.5 μL of magnetic beads were utilized. Libraries were sequenced on the Illumina MiSeq plaform using a 500 cycle kit with 400 cycles for Read 1 (5´ end of the cpn60 UT) and 100 cycles for Read 2. Only Read 1 sequences were used in downstream analysis. This asymmetric, directional sequencing approach simplifies subsequent bioinformatics steps since the length of the UT sequence prevents complete coverage with a paired end 2 x 250 strategy.

cpn60 UT amplicon sequence data from vaginal microbiome samples (n = 45) generated using the method described above were selected for analysis. Samples were from pregnant women and were selected based on the availability of 75,000–96,000 raw reads per sample, providing even levels of input in the different analyses. The original collection of the human microbiome samples was approved by the University of British Columbia Children’s & Women’s Research Ethics Board (certificate no. H17-02253). All sequence data used in comparison of methods is deposited in the National Center for Biotechnology Information (NCBI) Sequence Read Archive and is associated with BioProject accession PRJNA625735.

Cutadapt version 1.15 [30] and trimmomatic version 0.32 (quality score 30, sliding window of 4) [31] were used for amplification primer sequence removal and quality/length trimming, respectively.

Variant calling

Filtered sequence reads were imported to QIIME2 [32], and DADA2 [27] with a specified truncation length was executed. OTU (variant) sequences were exported from QIIME2 and aligned using watered-BLAST [33] to cpnDB_nr, a non-redundant version of the chaperonin database cpnDB [5,6] to identify the nearest neighbour (NN) for each read. cpnDB_nr consists of cpn60 UT sequences with a single record for each species, and a preference for the type strain if available (downloaded from http://cpndb.ca). Formed OTU sequences that were over 55% identical to anything in cpnDB_nr were considered to be cpn60 sequences and retained for further analysis. The 55% cutoff was established in a previous study showing that OTU with <55% identity to anything in the reference database correspond to host DNA sequences, or off-target (non-cpn60) amplicons [34]. The feature table was also exported, and sequence read counts were summed for OTU with the same NN species.

To evaluate an optimal truncation length for the cpn60 UT in terms of the quality of the sequence classification and data retention, truncation lengths of 50, 100, 150, 200, 250 and 300 were tested with DADA2. The OTU sequences formed at each truncation length were aligned to cpnDB_nr with watered-BLAST returning all significant hits [33]. An OTU sequence was considered unambiguously identified if there was only one resulting match to the database that did not tie with any other results.

De novo assembly

Trinity version 2.4.0 [35] with kmer size of 31 was used to assemble quality filtered reads into OTU. OTU sequence frequencies were determined by mapping quality filtered reads on to the assembled OTU sequences using Bowtie2 version 2.3.3.1 [36]. A feature table including the OTU frequency (read count) in each sample was created by processing the SAM file output from Bowtie2 using components of mPUMA [23]. Assembled OTU sequences were aligned to cpnDB_nr to determine nearest neighbours (NN, best sequence match). For community membership analyses, read counts for individual OTU with the same nearest neighbour species were combined.

Reference mapping

Filtered reads were mapped to cpnDB_nr or the VOGUE Reference Assembly with Bowtie2 version 2.3.3.1 [36]. The VOGUE Reference Assembly is a curated set of OTU previously generated from a de novo assembly of sequence reads from 546 vaginal microbiomes [14] and labeled according to nearest neighbour in cpnDB_nr. The mock community (MVP) was also mapped on to the 20 cpn60 UT sequences comprising the MVP. Following Bowtie2 mapping to the selected reference database, a feature table containing OTU names and read abundances was created as previously described [17].

Microbial community composition analysis

Alpha diversity metrics Shannon diversity and Chao1 estimated species richness were calculated for each sample as mean values over 100 subsamples with a sequence depth of 25000, using QIIME2 [32]. Alpha rarefaction plots were visualized to confirm that adequate sampling depth was achieved. The Kruskal-Wallis test by ranks was used to compare diversity and richness metrics resulting from the three methods, and if significant differences were found, post hoc pairwise comparisons were done using the Wilcoxon rank sum test.

For initial comparison and visualization of the relationships among microbiome profiles resulting from each OTU generation method in terms of species presence and abundance, a Jensen-Shannon distance matrix was calculated in R using a custom distance function that calculates the square root of the Jensen-Shannon divergence [37]. Hierarchical clustering was conducted using the hclust function in R with the Ward 2 linkage method. Vaginal Community State Type (CST) classification for vaginal samples was based on an assessment of the proportionally dominant species in each sample, and comparison to previously published definitions of vaginal CST (CST I (Lactobacillus crispatus dominant), CST II (Lactobacillus gasseri dominant or co-dominant), CST III (Lactobacillus iners dominant or co-dominant), CST V (Lactobacillus jensenii co-dominant), CST IVa (variable, heterogenous mix) and CST IVc (G. swidsinskii and/or G. leopoldii) [14,38,39].

Relationships among microbiome profiles generated for the vaginal samples were also examined by PERMANOVA analysis. Feature tables containing read counts were first subject to centre log-ratio (CLR) transformation to estimate the likelihood of true zero values in the dataset (recommended for sparse, compositional datasets like microbiome profiles) using the aldex.clr function of ALDEx2 [40]. Only OTU present in at least 10% of samples were retained for downstream analysis. CLR-transformed data were converted to a propr object using the aldex2propr function of ALDEx2 and principle component analysis (PCA) was performed on the resulting propr object. PCA data were plotted as compositional biplots (ggbiplot) to visualise similarity of microbiome profiles. To test for significant differences in microbiome profiles by OTU generation method and between samples, the adonis function of vegan R package was used to construct a Euclidean distance matrix from the CLR-transformed data and perform a PERMANOVA on the resultant matrix. As PERMANOVA is sensitive to differences in dispersion of data within groups, it assumes a homogenous within-group dispersion; therefore, we checked this assumption with the betadisper and permutest functions of vegan. The former determines dispersion of data within each group while the latter permutes the dispersion data to test for significant variability of within-group dispersion. Finally, post-hoc testing of permutation data from permutest was performed using the TukeyHSD function to determine where differences in dispersion lay.

Optimal length for variant calling

Variant calling using the DADA2 denoising algorithm in QIIME2 employs a truncation length parameter to which all input sequence reads are trimmed uniformly. If a read is under that length then it is not included in the process [27]. It was therefore desirable to first determine an optimal truncation length to use with the cpn60 UT sequences, given input reads of up to 400 bp and the trade-off between minimum length requirement, and proportion of reads retained in variant calling. This parameter would then be used for the analysis of the variant calling approach when comparing other OTU generation methods.

To investigate the optimal truncation length for cpn60 variant calling with DADA2 without the confounding effect of diminishing read numbers retained as the truncation length requirement increases, we generated a single set of input reads from the vaginal microbiome sequence data that were all at least 300 bp in length after quality filtering. The Trimmomatic quality filtering step was performed so that only reads above 300 bp with a quality threshold of 30 over a 4 bp sliding window would be used as a consistent input for DADA2 (n = 346526 reads). As expected, there were more variants (OTU) produced with the greater sequence truncation lengths (Table 1). This was expected given the way in which DADA2 determines variants; longer lengths mean the algorithm has more chances to detect differences. The OTU identified at each truncation length corresponded to a decreasing number of nearest neighbours as more, lower-quality matches were replaced with fewer, higher quality matches. This observation was also reflected in the increasing proportion of OTU with a single best match. The proportion of OTU unambiguously identified reached a plateau at ~90% starting at a truncation length of 150 bp (Table 1). Given the protocol with which the samples were sequenced, all reads start from the 5´ end of the cpn60 UT. Any OTU aligned in the reverse direction are therefore considered nonsense and cannot be interpreted for species identification. By 150 bp, 100% of the OTU were aligned in the correct direction (Table 1). Taken together, these results indicate that 150 bp from the 5´ end of the cpn60 barcode sequence is sufficient for microbiome profiling with species-level resolution, consistent with the known features of the cpn60 barcode sequence [3]. Using truncation lengths >150 bp could be desirable depending on the amount of data available, but the obvious trade-off between minimum length and number of reads retained for analysis should be carefully considered. The effect of loss of shorter reads is that the overall sequencing depth of the samples is reduced, and thus the chance of detecting rare organisms, which likely outweighs the benefit of the small increase in resolution that would be achieved using greater lengths. Shorter read lengths could also decrease the cost of sequencing.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Variant calling results for vaginal microbiome samples processed at different truncation lengths with a consistent input of 346526 quality filtered reads.

https://doi.org/10.1371/journal.pone.0235682.t001

Additional support for the 150 bp length was obtained when we examined the pairwise percent identities calculated from multiple sequence alignments of bacterial sequences from cpnDB_nr (n = 5974), trimmed to lengths ranging from 50 bp to >500 (full length) from the 5´ end (S1 Fig). The mean and median percent identities of the pairwise comparisons for alignments of 150 and 200 bp were significantly lower than for any other lengths of the cpn60 barcode. This finding is consistent with that of a previous investigation demonstrating relatively higher inter-species sequence differences in the 5´ end of the cpn60 barcode sequence relative to the 3´ end [3].

Considering all the trade-offs of data retention, cost of sequencing, and species resolution, a DADA2 truncation length of 150 bp was used for the subsequent comparison of methods of OTU generation.

Accuracy of OTU generation methods for a mock community

The MVP synthetic community was used to determine accuracy of the three methods of OTU generation, since unlike biological specimens the composition of the MVP is known. Following amplification primer removal and quality trimming, 54702 and 40128 reads were available for analysis from the duplicate sequencing runs, respectively. The number of OTU produced from de novo assembly, variant calling (truncation length 150), and reference mapping to either the VOGUE reference assembly or MVP sequences was determined (Fig 1). Reference mapping directly to the MVP gave the expected result of 20 OTU in both library duplicates, confirming that all constituents of the mock community were amplified and sequenced. When the MVP OTU were aligned to the MVP dataset using watered-BLAST to identify nearest neighbours, they aligned to only 19 different species. This difference was not unexpected since two of the cloned cpn60 barcode sequences in the MVP (corresponding to Lactobacillus gasseri and L. johnsonii) are 97% identical. Reference mapping to the VOGUE reference assembly was included in the comparison to determine the effect that a larger and more diverse mapping database could have on the results, and we found that far more OTU and NN were generated from the MVP sequence libraries using this approach (Fig 1). Further analysis of the profiles showed that many of the ‘extra’ OTU that were produced were the result of reads mapping to highly similar sequences in the reference database. This result illustrates a limitation of the reference mapping approach, which is dependent on the Bowtie2 algorithm to determine the best match for each read. Bowtie2 alignments close in score may be chosen over one anther unpredictably, or if the program’s search limit has been reached it may not determine a better alignment if one exists, which may lead to changes in resulting profiles [36].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Number of operational taxonomic units (OTU) and nearest neighbour species (NN) from each method applied to the MVP mock community of 20 vaginal bacteria.

The synthetic community was sequenced in duplicate in separate runs. For each pair of duplicates, the lighter bar represents the first run and the darker bar the replicate. Methods: 1 = de novo assembly, 2 = reference mapping to MVP, 3 = reference mapping to VOGUE reference assembly, 4 = variant calling.

https://doi.org/10.1371/journal.pone.0235682.g001

Variant calling and de novo assembly of MVP derived sequence reads both produced more OTU than the 20 expected, up to 29 in the first replicate analysed by variant calling (Fig 1). Additional OTU were most likely formed as a result of minor sequence difference due to PCR errors in the barcode sequence amplification, which was not done with a proof-reading polymerase. Support for this explanation is also found in the observation that the ‘extra’ OTU corresponded to the expected number of nearest neighbours in the second replicate. These observations are typical for descriptions of microbial communities based on barcode amplicon sequencing. OTU formed, especially rare ones, may be either real biological entities or artefacts due to PCR and sequencing errors or the bioinformatic processes used to designate and identify them [41]. We reported all OTU in this analysis regardless of abundance, but a variety of approaches have been applied to curation of OTU, including establishment of cut-offs based on abundance or degree of co-occurrence among samples [42].

Taken together, the results of analysis of the MVP synthetic community sequence data demonstrates that while variant calling and de novo assembly produced the expected community composition, results of reference mapping were heavily influenced by the qualities of the database used and could lead to a significant over-estimation of the number of species in the community.

Effects of OTU generation method on vaginal microbiome profiles

Microbial community profiles derived from barcode amplification and sequencing include both an inventory of the distinct OTU present as well as their relative abundance based on sequence read numbers. Various statistical approaches and clustering methods can then be applied to investigate beta-diversity within collections of samples with the goal of identifying similarities and differences among samples that may be associated with defined groups in a study, for example pre- and post-treatment groups, or subjects with different clinical outcomes or phenotypes. To determine the effects of the different OTU formation, identification, and frequency calculation methods on overall microbiome profiles, we exploited previously generated data sets derived from the vaginal microbiome (n = 45, 3.8 million raw reads). After quality filtering (q30, minimum length 150), 2 million reads were available for analysis with an average of 44600 reads per sample. For de novo assembly and variant calling, sequence read counts for OTU with identical nearest neighbours in cpnDB_nr were combined to facilitate merging of the resulting read count tables with the output from reference mapping. Results of de novo assembly, reference mapping, and variant calling (using a 150 bp truncation length) were then compared to determine if the methods produce different results in terms of community composition or different conclusions regarding the relationships among samples (alpha and beta diversity).

There was no difference in diversity (Shannon) of the vaginal microbiome profiles determined with the three OTU generation methods (Kruskal-Wallis, P > 0.05), but there was a significant difference in richness (Chao1) (Wilcoxon rank sum test, all pairwise comparisons P < 0.02) (Fig 2), which was not unexpected given the characteristics of the methods. The relative overestimation of community richness resulting from the reference mapping approach that we observed in the synthetic community sequencing experiment was also observed here. Richness of communities based on OTU generated with variant calling was lower overall than when OTU were generated by de novo assembly, which may reflect the removal of variants determined by the DADA2 denoising algorithm to be the result of sequencing error [26].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Comparison of alpha diversity metrics.

Shannon diversity (A) and Chao1 (B) values were calculated for each vaginal microbiome sample (n = 45) and each method of OTU formation. P values are indicated above the bars, * = P < 0.02, ** = P < 0.001, ns = not significant.

https://doi.org/10.1371/journal.pone.0235682.g002

The microbiome profiles of each sample determined with each method were subjected to cluster analysis to determine if the OTU formation method affected the overall microbiome profile and interpretation of the relationships among samples. In most cases, the three profiles for each sample clustered together regardless of OTU generation method (Fig 3). Clusters of samples corresponding to previously defined vaginal microbiome “community state types” (CST) were identified based on visual inspection of proportional abundance of key taxa (S2 File). CST identified were similar to those reported in previous cpn60-based studies of vaginal microbiota of Canadian women: CST I (Lactobacillus crispatus dominant), CST II (Lactobacillus gasseri dominant or co-dominant), CST III (Lactobacillus iners dominant or co-dominant), CST V (Lactobacillus jensenii co-dominant), CST IVa (variable, heterogenous mix) and CST IVc (G. swidsinskii and/or G. leopoldii) [14]. Most (43/45) samples received the same CST classification regardless of the OTU generation method used, demonstrating that one would not make a different conclusion regarding CST affiliation using any of the three methods tested in this study (Fig 3, S2 File). The exceptions were sample numbers 19 and 29. In the results of reference mapping and variant calling for these samples, the most abundant OTU (accounting for 35% and 37% of these samples, respectively) was identified as Gardnerella leopoldii, and the second most abundant OTU (29% or 27%) was identified as Lactobacillus jensenii, consistent with CST V. Although the proportional abundance of L. jensenii was the same in the de novo assembly results for either sample, the most abundant OTU that accounted for 34% and 36% of the microbiome profiles was identified as G. swidsinskii (S2 File). The cpn60 sequences of G. leopoldii and G. swidsinskii are 97% identical to each other. It is surprising that this difference alone would result in a change in the CST identification for these samples and so it is likely that differences in detection of other lower abundance species also contributed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Comparison of microbiome profiles resulting from three methods of OTU generation.

Clustering of profiles from vaginal samples based on nearest neighbour species abundance with Jensen-Shannon distance calculation and hclust in R. In cases where the sample profiles from all methods cluster together, only the sample number is shown. RM = reference mapping, DN = de novo assembly, VC = variant calling. Clusters corresponding to CST I-IV are indicated, and are described in the text.

https://doi.org/10.1371/journal.pone.0235682.g003

Additional support for the lack of effect of OTU generation method on beta diversity was observed when PERMANOVA analysis was performed. Given that this method is sensitive to group dispersion effects in unbalanced study designs, we also tested for homogeneity of variance. PERMANOVA reported no significant clustering of the vaginal samples by OTU generation method (P = 0.986) (Fig 4A); however vaginal microbiome profiles significantly clustered based on sample number (P < 0.001) (Fig 4B), meaning that for each sample, the assembly method used to generate the profile did not significantly affect the community composition. No differences in dispersion between groups were seen for either assembly method or sample number (P = 0.738 and 0.930, respectively).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Assembly method does not affect apparent microbiome composition: Centre log-ratio transformation (CLR) was performed on read count data from all 45 vaginal microbiome profiles generated using three different methods (n = 135).

PCA plots grouped by assembly method (A) or sample number (B) were generated and Euclidean distance matrices of CLR-transformed data were analysed by PERMANOVA. Significant clustering of vaginal microbiomes by sample, but not assembly method was evident from PCA plots and supported by PERMANOVA results (P < 0.001 and P = 0.986, respectively).

https://doi.org/10.1371/journal.pone.0235682.g004

Practical considerations and recommendations

All three approaches to OTU generation use open source software so there is no cost factor to consider. When the real time for all processes in each OTU formation method applied to the 45 vaginal samples (conducted on a Dell PowerEdge R910 with 250 GB RAM running CentOS Linux 7) were summed, reference mapping was the most time-consuming method (58.7 min), followed by de novo assembly (23.7 min) and variant calling (16.1 min). Times for reference-mapping are dramatically increased relative to the other approaches as the size of the reference database increases.

The reference mapping approach can be efficient if a reference database customized for the environment under study, but such reference sequence databases are not available for all bacterial communities. The de novo assembly method has been used extensively for cpn60-based microbiome profiling, but is slower, and requires more software dependencies than variant calling. Taken together, the results of the current study suggest that both of these approaches, especially reference mapping, may overestimate microbial community richness. Variant calling was the fastest method we tested, and truncation lengths of at least 150 bp resulted in species level identification of OTU. The use of denoising methods for variant calling has also been suggested for detection of otherwise unobserved 16S rRNA sequence variants [26]. This approach is particularly attractive for users with limited experience with data analysis, since it can be easily installed and accessed within the QIIME2 package [32]. Another significant benefit of variant calling is that unique identifiers are assigned to each variant making it simple to compare detection of variants across studies.

The choice of method for sequence classification from barcode sequence-based studies of microbial communities is influenced by a number of factors including pre-existing knowledge of the environment under study, numbers of samples, sequencing depth, desired level of taxonomic resolution, computational resources, and the need to pool results from independent studies. Exact sequence variant calling with DADA2 using 150 bp from the 5´ end of the cpn60 barcode sequencing is a recommended strategy for rapid, relatively simple to execute, high-resolution profiling of microbiomes that facilitates the comparison of microbiome profiles across studies.