Optimizing 16S rRNA gene profile analysis from low biomass nasopharyngeal and induced sputum specimens.

Background Careful consideration of experimental artefacts is required in order to successfully apply high-throughput 16S ribosomal ribonucleic acid (rRNA) gene sequencing technology. Here we introduce experimental design, quality control and “denoising” approaches for sequencing low biomass specimens. Results We found that bacterial biomass is a key driver of 16S rRNA gene sequencing profiles generated from bacterial mock communities and that the use of different deoxyribonucleic acid (DNA) extraction methods [DSP Virus/Pathogen Mini Kit® (Kit-QS) and ZymoBIOMICS DNA Miniprep Kit (Kit-ZB)] and storage buffers [PrimeStore® Molecular Transport medium (Primestore) and Skim-milk, Tryptone, Glucose and Glycerol (STGG)] further influence these profiles. Kit-QS better represented hard-to-lyse bacteria from bacterial mock communities compared to Kit-ZB. Primestore storage buffer yielded lower levels of background operational taxonomic units (OTUs) from low biomass bacterial mock community controls compared to STGG. In addition to bacterial mock community controls, we used technical repeats (nasopharyngeal and induced sputum processed in duplicate, triplicate or quadruplicate) to further evaluate the effect of specimen biomass and participant age at specimen collection on resultant sequencing profiles. We observed a positive correlation (r = 0.16) between specimen biomass and participant age at specimen collection: low biomass technical repeats (represented by < 500 16S rRNA gene copies/μl) were primarily collected at < 14 days of age. We found that low biomass technical repeats also produced higher alpha diversities (r = − 0.28); 16S rRNA gene profiles similar to no template controls (Primestore); and reduced sequencing reproducibility. Finally, we show that the use of statistical tools for in silico contaminant identification, as implemented through the decontam package in R, provides better representations of indigenous bacteria following decontamination. Conclusions We provide insight into experimental design, quality control steps and “denoising” approaches for 16S rRNA gene high-throughput sequencing of low biomass specimens. We highlight the need for careful assessment of DNA extraction methods and storage buffers; sequence quality and reproducibility; and in silico identification of contaminant profiles in order to avoid spurious results.


Background
High-throughput 16S ribosomal ribonucleic acid (rRNA) gene sequencing has the potential to provide detailed characterization of microbial communities from a range of ecological niches in humans [1][2][3]. These range from the densely colonized gastro-intestinal tract [4,5] to low biomass sites including the lower respiratory tract [6][7][8] and womb [9,10] which have previously been considered sterile. However, there is much controversy surrounding evidence from 16S rRNA sequencing studies supporting opposing hypotheses on womb sterility -in particular the placental microbiome [11,12] -and on the accuracy of studies of low biomass samples more broadly [13,14].
A primary concern when using 16S rRNA gene sequencing to analyse specimens with low levels of endogenous deoxyribonucleic acid (DNA) is preferential amplification and sequencing of contaminant DNA originating from reagents or the laboratory environment [15][16][17][18]. No template controls (NTCs) such as storage buffers, elution buffers, or water may serve as good approximates for "contaminants" introduced during extraction and library preparation steps [16,18,19] (Fig. 1a). However, DNA and amplicon "spill-over" from high biomass to low biomass specimens (also referred to as wellto-well contamination) adds to exogenous biological profiles in neighbouring low biomass specimens [20] (Fig. 1b). This physical exchange of DNA/amplicon between biological specimens, and between biological specimens and NTCs (Fig. 1b), calls for a rigorous in silico approach to identify and remove contamination. For example, simply subtracting NTC contaminant profiles from biological specimens introduces the risk of removing true biological profiles alongside contaminant profiles [18] (Fig. 1b). Some studies have used in silico approaches to remove previously reported potential contaminant profiles [21,22], however, contaminant profiles are likely to vary between laboratories and between experiments within laboratories. Hence, the only way to efficiently deal with contaminants is to include controls tailored to each experiment and to use optimal decontamination approaches.
In addition to addressing contamination, 16S rRNA gene sequencing studies also need to validate the process of DNA extraction and polymerase chain reaction (PCR) amplification, which is typically done via the inclusion of bacterial mock communities as extraction and sequencing controls [23]. These mixtures of known bacterial composition can be generated to mimic biological specimens and used to identify optimal DNA extraction methods for the specimen type of interest. Due to their known composition, they are also used to evaluate sequencing reproducibility and identify contaminants. However, mock community controls also need to mimic biological profiles in their biomass, as specimen biomass contributes to the level of contaminants and sequencing reproducibility observed in any given sequencing run. Therefore, biological specimens -with inherent variation in biomass -randomly selected for repeat extraction and/or sequencing within and between different sequencing runs may provide an added benefit when measuring reproducibility and identifying potential contaminants.
Standardization of bacterial profiling methods provides opportunity for multicentre comparative analyses [24], however, this may not rule out technical biases. Protocols for analysing bacterial profiles from low biomass specimens, which are at particular risk of PCR bias and contamination, need to be optimized prior to standardization. Nasopharyngeal (NP) specimens are one such example as they have bacterial densities of < 10 6 16S copies/mL [15,17,25]. In order to contribute to the development of robust 16S rRNA gene sequencing protocols for low biomass specimens, we outline important quality control steps using NP and induced sputum (IS) specimens collected from infants. Our first objective was to determine how different DNA extraction methods, bacterial biomass and specimen storage buffers could influence 16S rRNA gene sequencing profiles from bacterial mock community controls. Our second objective was to investigate whether specific characteristics from low biomass biological specimens (NP and IS) correlates with sequencing quality. Finally, our third objective was to evaluate the use of two methods for in silico identification of potential contaminants from 16S rRNA gene sequencing data generated from low biomass biological specimens (NP and IS). Overall, we aimed to provide additional insight into experimental design, quality control steps and "denoising" approaches for sequencing low biomass specimens.

DNA extraction method, specimen biomass and storage buffer influence 16S rRNA gene sequencing profiles
This section describes sequencing profiles generated using DNA extracts from high and low biomass bacterial mock community controls (Zymobiomics-Primestore-high, Zymobiomics-STGG-high, Zymobiomics-Primestore-low and Zymobiomics-STGG-low) generated using a commercial bacterial mock community (Zymobiomics-Cells) and two storage buffers (Primestore and STGG) ( Table 1, Additional file 1). For each of the four bacterial mock communities, we present sequencing profiles generated from triplicate DNA extracts using two DNA extraction kits (Kit-QS and Kit-ZB) ( Table 2). In addition, we report on sequencing profiles from two pre-extracted commercially available bacterial mock community DNA controls (Zymobiomics-DNA and BEI-DNA), included as sequencing controls ( Table 1, Table 2).
Bacterial profiles sequenced from BEI-DNA controls (n = 3) showed good reproducibility and comparability to manufacturer's specifications [27] (Additional file 2). Three genera, reported at low abundances by the manufacturer (Actinomyces 0.02%, Deinococcus 0.02% and Propionibacterium 0.2%) [27], were not detected from the DNA amplified and sequenced in our laboratory. We detected 11 additional genera (6 of which were unclassifiable at genus-level) at low abundances from BEI-DNA controls processed in our laboratory. The sequenced Fig. 1 Representation of a expected and b actual sequencing profiles from no template controls, low biomass and high biomass specimens following 16S rRNA gene sequencing. a Expected 16S rRNA gene sequencing profiles from i) no template controls (NTCs), ii) low biomass and iii) high biomass biological specimens which corresponds with their endogenous bacterial composition. b Actual 16S rRNA gene sequencing profiles generated from i) NTCs may comprise of reagent and laboratory contaminants as well as exogenous sequences from low and high biomass specimens (well-to-well contamination); ii) low biomass biological specimen sequencing profiles may be overrepresented by exogenous profiles from both NTCs (reagent and laboratory contaminants) and high biomass specimens (well-to-well contamination); whilst iii) high biomass sequencing profiles are expected to be least affected by reagent and laboratory contaminants present in NTCs and cross-contamination from low biomass specimens profiles from Zymobiomics-DNA controls (n = 8) also matched the manufacturer's specifications with the exception of a few low abundant OTUs (Additional file 2). We detected 35 additional genera (16 of which were unclassifiable at genus-level) at low abundances from the eight Zymobiomics-DNA controls processed in our laboratory (Additional file 2).
Data generated from the four bacterial mock communities (Zymobiomics-Primestore-high, Zymobiomics-Primestorelow, Zymobiomics-STGG-high and Zymobiomics-STGG-low) showed that Kit-QS extracted purer DNA compared to Kit-ZB based on the ratio of absorbance (260 nm and 280 nm) measured by the NanoDrop® ND-1000 (Table 3). We only compared ratio of absorbance measures for the two kits using high biomass bacterial mock communities, as low biomass bacterial mock community dsDNA concentrations were outside the NanoDrop® ND-1000 Spectrophotometer's lower limit of detection. As much as 100-fold more 16S rRNA gene copies per millilitre of specimen input volume was extracted from low biomass bacterial mock  Principal coordinate analysis at OTU-level showed that bacterial biomass is a key driver of 16S rRNA gene profiles (analysis of beta diversities by PERMANOVA: P = 0.001) (Fig. 2a). High biomass bacterial mock communities (Zymobiomics-Primestore-high and Zymobiomics-STGG-high) grouped together in a compact cluster alongside Zymobiomics-DNA, whilst low biomass bacterial mock communities (Zymobiomics-Primestore-low and Zymobiomics-STGG-low) clustered midway between their undiluted counterparts and NTCs (Primestore and STGG) (Fig. 2a). Beta diversities were also significantly different when comparing the two DNA extraction methods (P = 0.001) and storage buffers (P = 0.001) (Fig. 2a). When stratifying our analyses based on the two storage buffers [Primestore (Fig. 2b) and STGG ( Fig. 2c)], we observed clear differentiation in beta diversities in relation to specimen biomass (P = 0.001) and the DNA extraction method (P = 0.001) used.
In order to assess the efficiency of the two extraction methods in extracting DNA from hard-and easy-to-lyse bacteria, we compared sequencing profiles at OTU-level from Zymobiomics-Primestore-high (n = 6) and Zymobiomics-STGG-high (n = 6) to Zymobiomics-DNA (n = 8). Compared to Zymobiomics-DNA, both Kit-QS and Kit-ZB yielded an over-representation of easy-tolyse gram-negative bacteria (Pseudomonas, Salmonella and Escherichia-Shigella) and an under-representation of hard-to-lyse gram-positive bacteria (Enterococcus, Staphylococcus, Listeria and Bacillus) (Additional file 3). Differences between Zymobiomics-DNA and the two DNA extraction methods were more marked for Kit-ZB (Additional file 4 a). The hard-to-lyse gram-positive bacteria Enterococcus, Staphylococcus and Listeria were also significantly less represented by Kit-ZB when compared to Kit-QS (Additional file 4 b). We detected more "background OTUs" (OTUs not expected in mock communities) from low biomass bacterial mock communities (Zymobiomics-Primestore-low and Zymobiomics-STGG-low) compared to high biomass bacterial mock communities (Zymobiomics-Primestore-high and Zymobiomics-STGG-high), irrespective of the DNA extraction method and storage buffer used (Fig. 3). The latter was more pronounced from low biomass profiles generated using storage buffer STGG (average proportion of "contaminant OTUs": 9.5%) when compared to Primestore (average proportion of "contaminant OTUs": 1.5%).
In summary, we show that the bacterial biomass is an important determinant of 16S rRNA gene sequencing profiles and that storage buffers and DNA extraction methods further influence these profiles. When comparing Kit-QS to Kit-ZB, we observed better quality of DNA and a better representation of hard-to-lyse bacteria from profiles generated using Kit-QS. Finally, we show that the use of different storage buffers impacts on low biomass "background OTUs", with the commonly used STGG buffer associated with increased background.
Quality of 16S rRNA gene sequencing data from respiratory tract specimens correlates with specimen biomass We validated 16S rRNA gene sequencing data generated across 11 sequencing runs from low biomass technical repeats (NP and IS specimens) and determined criteria for excluding specimens based on the likelihood of spurious results. We compared 1) technical repeat profiles to NTC profiles, 2) examined sequencing reproducibility and 3) investigated the presence of "spurious OTUs" in relation to specimen biomass, participant age at specimen collection and read counts following bioinformatic processes. In total, we analysed 244 sequencing libraries generated from 35 Primestores (NTCs) and 209 technical repeats (NP and IS specimens) ( Table 1, Table 2). All technical repeats had a minimum of two sequencing profiles available for analysis. We generated 86 sequencing profiles from withinrun repeats (43 NP/IS specimens processed in duplicate buffers. a Differences in beta diversities (calculated at OTU-level) measured from all bacterial mock community controls and no template controls (NTCs). b Differences in beta diversities measured from bacterial mock community controls and NTCs generated using Primestore storage buffer. c Differences in beta diversities measured from bacterial mock community controls and NTCs generated using STGG storage buffer. The proportion of variance captured by coordinate analysis axes are shown in the bottom left corner of each panel. Blue and red colours represent DNA extraction methods Kit-QS and Kit-ZB, respectively. Shades of chartreuse filled circles represent bacterial mock communities generated using Primestore storage buffer (solid-filled chartreuse circles: high biomass bacterial mock communities; pattern-filled chartreuse circles: low biomass bacterial mock communities). Shades of emerald filled circles represent bacterial mock communities generated using STGG storage buffer (solidfilled emerald circles: high biomass bacterial mock communities; pattern-filled emerald circles: low biomass bacterial mock communities). Dark green filled circles represent Zymobiomics-DNA. Chartreuse and emerald pattern-filled squares represent Primestore and STGG NTCs, respectively within respective runs); and 123 sequencing profiles from between-run repeats [NP/IS specimens processed in duplicate (n = 30), triplicate (n = 9) and quadruplicate (n = 9) across different runs].
Sequencing profiles from low biomass technical repeats (NP and IS specimens), collected during the first two weeks of life, are similar to profiles from no template controls (Primestore) Using logarithm of ratio-transformed data (log-ratio) biplots, we observed two distinct clusters in relation to participant age at specimen collection ( Fig. 5a) and specimen biomass (16S rRNA gene copies/μl) (Fig. 5b). Almost all technical repeats collected at < 14 days of life clustered closely with Primestore (Fig. 5a). This was also evident for a subset of specimens collected at > 14 days of life. The majority of technical repeats with < 500 16S rRNA gene copies/μl clustered with Primestore (Fig. 5b). Although less marked, we also observed a tendency of specimens with lower read counts to cluster with Primestore (Fig. 5c).
Specimen features, such as participant age at specimen collection, 16S rRNA gene copy numbers and read counts as proxy for sequencing reproducibility We determined the reproducibility of sequenced profiles among technical repeats included in our dataset. We calculated a R 2 value for each of the duplicate sequencing profiles (n = 73) present in our dataset, three R 2 values for each of the triplicate sequencing profiles (n = 27), and six R 2 values for each of the quadruplicate sequencing profiles (n = 54). Overall, sequencing reproducibility was high across the entire dataset [median R 2 = 0.992 (IQR: 0.951-0.999)]. Sequencing profiles generated from (See figure on previous page.) Fig. 4 Participant age at specimen collection, read counts and alpha diversity relative to specimen biomass for no template controls (Primestore, n = 35) and technical repeats (n = 209). a Scatter plot of participant age at specimen collection b read counts following bioinformatic processes and c Shannon diversity indices (alpha diversity) at OTU-level in relation to specimen biomass (16S rRNA gene copies/μl) plotted on log e scale. Vertical pink shaded area highlights 16S rRNA gene copies/μl < 500 Fig. 5 Logarithm of ratio-transformed data (log-ratio) biplots in relation to participant age at specimen collection, 16S rRNA gene copies/μl and read counts following bioinformatic processing. Data points are coloured according to a participant age at specimen collection (in days), b 16S rRNA gene copies/μl and c read counts available for downstream analyses. Technical repeats (n = 209) are represented using filled circles. No template controls (Primestore, n = 35) are represented using filled triangles within-run technical repeats (n = 86) were more reproducible compared to those from between-run repeats (n = 123) [median R 2 = 0.999 (IQR: 0.991-1.000) vs 0.982 (IQR: 0.942-0.997)].
In summary, low biomass technical repeats -represented by low 16S rRNA gene copy numbers (< 500 copies/μl) -produced fewer reads and were primarily collected during the first 2 weeks of life. These specimens tended to have higher alpha diversity, produce more OTUs (including "spurious OTUs") and had less reproducible sequencing profiles. In addition, the majority of low biomass technical repeats clustered with Primestore when assessing beta-diversity.
Improved quality of 16S rRNA gene datasets is attainable via the decontam R package for identification of "potential contaminants" We aimed to validate two in silico approaches ("NTConly" and "NTC + decontam") to identify potential contaminants from bacterial profiles generated from technical repeats (NP and IS specimens processed in duplicate, triplicate or quadruplicate) ( Table 1, Table 2). Prior to analysis, we denoised the dataset consisting of 35 Primestores and 209 technical repeats by removing 249 of the 1252 OTUs identified as "spurious OTUs" across the dataset. We included technical repeats with > 500 16S rRNA gene copies/μl measured from their template (n = 148) as true biological specimens (based on Fig. 6 Bacterial composition in no template controls (Primestore, n = 35) and low biomass technical repeats (n = 209). a Dendogram representing unsupervised hierarchical clustering distances are based on Bray Curtis dissimilarity indices calculated at OTU-level. The dendogram is colourcoded based on specimen type (Primestore: darkturquoise; technical repeats: deeppink). b Differences between Primestore and technical repeats are shown at genus-level, with colour-codes representing phylum-level classification (Shades of blue: Proteobacteria, shades of red: Firmicutes). Genera with proportions < 1% in each of the specimens are grouped together as "Other" and shown in grey. c Most abundant genera within each the specimens, specimen type, participant age at specimen collection (in days) and 16S rRNA gene copy numbers (copies/μl) are summarised at the bottom of the Fig. A-N-P-R: Allorhizobium-Neorhizobium-Pararhizobium-Rhizobium Fig. 7 Associations between reproducibility and a participant age at specimen collection, b 16S rRNA gene copy numbers, and c read counts. Reproducibility is measured by coefficient of determination (R 2 ) values, calculated by comparing proportions of each OTU present between technical repeats. Horizontal blue bars highlight R 2 values > 0.90. Different shades of vertical blue bars represent a < 7, < 14, < 30, < 60 days; b < 100, < 500, < 1000 copies/μl; and c < 2000, < 4000, < 6000, < 8000 and < 10,000 reads; respectively. For b and c, each set of technical repeats had two 16S rRNA gene copy number/read count measures shown as two points connected by a horizontal line on the X-axis results reported in the previous section). The "NTConly" approach designates an OTU as a "potential contaminant" when it is detected across both Primestore (n = 35) and technical repeats (n = 148) [16,18,19,28]. The "NTC + decontam" approach designates an OTU as a "potential contaminant" via implementation of the decontam package in R [29]. We identified potential contaminants via the "NTC + decontam" approach based on 1) their frequency as a function of the specimen biomass; and 2) their prevalence in true biological specimens (n = 148) compared to Primestore (n = 35).
The "NTC + decontam" approach results in improved removal of "potential contaminants" when compared to the "NTConly" approach Figure 9 summarises the per specimen proportions observed for five genera previously reported as "potential contaminants" prior to and after decontamination (Additional file 9). Both "NTConly" and "NTC + decontam" approaches completely removed Aquabacterium, Acidovorax and Noviherbaspirillum profiles from the majority (100, 99 and 72%) of specimens positive for these genera (Fig. 9a-c; Additional file 9). We observed lower proportions for specimens positive for Acinetobacter and Stenotrophomonas following decontamination using the "NTConly" compared to the "NTC + decontam" approach ( Fig. 9d-e; Additional file 9). Both Acinetobacter and Stenotrophomonas have been reported as potential contaminants at genus-level in 16S rRNA gene sequencing datasets [15,30], but are also reported as of the upper airway bacteria [31][32][33]. When investigating both Acinetobacter and Stenotrophomonas genera at OTUlevel, we observed that the "NTConly" approach identified OTU_15 (A. baumannii), OTU_863 (A. johnsonii), OTU_358 (A. calcoaceticus-A. baumannii complex) and OTU_56 (S. maltophilia) (Additional file 10) as potential contaminants, all of which are commonly isolated from the respiratory tract.
In summary, the "NTConly" approach to decontamination resulted in larger shifts in proportions of genera commonly detected from the nasopharynx compared to using the decontam package implemented through R software ("NTC + decontam" approach). Furthermore, the "NTC + decontam" approach better differentiated likely "true nasopharyngeal bacteria" from "potential contaminants" at OTU-level compared to the "NTConly" approach.

Discussion
The complexities of 16S rRNA gene sequencing of low biomass specimens are increasingly recognised and have broad applicability [13,14,30,34]. In order to deal with such complexities, we provide, in a stepwise manner, a comprehensive overview of several key components of a quality control process for low biomass 16S rRNA gene sequencing studies. We highlight the importance of evaluating DNA extraction protocols prior to the implementation thereof, and the need to account for potential effects of different storage buffers used to preserve specimens. We further show the value of determining specimen exclusion criteria via the use of optimal sequencing Fig. 8 Shifts in profiles of bacterial genera commonly detected from the nasopharynx prior to and following decontamination via two in silico approaches. Per specimen shifts (n = 148) in bacterial proportions are shown for bacterial genera commonly detected from the nasopharynx: a Moraxella, b Corynebacterium 1, c Haemophilus, d Staphylococcus and e Streptococcus. Open circles and smoothing splines (representing a factor of 2x the standard deviation) denote bacterial proportions (Y-axis) for each of the specimens (X-axis). Red: Proportions prior to decontamination; Blue: Proportions following the removal of "potential contaminants" identified using the "NTConly" approach; Yellow: Proportions following the removal of "potential contaminants" identified using the "NTC + decontam" approach Fig. 9 Shifts in profiles of potential contaminants prior to and following decontamination via two in silico approaches. Per specimen shifts (n = 148) in bacterial proportions are shown for bacterial genera commonly described as "potential contaminants" in 16S rRNA gene sequencing datasets a Aquabacterium, b Acidovorax, c Noviherbaspirillum, d Acinetobacter and e Stenotrophomonas. Open circles and smoothing splines (representing a factor of 2x the standard deviation) denote bacterial proportions (Y-axis) for each of the specimens (X-axis). Red: Proportions prior to decontamination; Blue: Proportions following the removal of "potential contaminants" identified using the "NTConly" approach; Yellow: Proportions following the removal of "potential contaminants" identified using the "NTC + decontam" approach controls and exploratory analyses and highlight the importance of optimized approaches for contaminant identification and removal in any low biomass 16S rRNA gene sequencing study. We believe that our specific and detailed approach to dealing with low biomass respiratory tract specimens provides a rigorous method to assessing and responding to the complexities inherent to low biomass 16S rRNA gene sequencing experiments in general.
Prior to the advent of 16S rRNA gene highthroughput sequencing, bacterial DNA extraction platforms have primarily been evaluated based on DNA yield and purity, DNA shearing, bacterial diversity (using fingerprinting techniques) and how well DNA from hard-and easy-to-lyse bacterial cells are represented (using targeted PCR). Recently, mock community controls have been implemented to further assess the performance of extraction platforms [35,36]. When carefully selected, these mock controls may allow for the evaluation of extraction efficiency, serve to determine extraction reproducibility, and provide insight into contamination profiles. However, in order evaluate these important aspects of 16S rRNA gene sequencing, mock community controls need to represent bacterial communities expected from the specimen of interest and optimally reflect its biomass. In this study, we used both high and low biomass bacterial mock community controls to evaluate the performance of two DNA extraction kits. Overall, specimen biomass had a noticeable impact on sequencing profiles irrespective of the storage buffer and DNA extraction method used. These findings, together with previous reports on breast milk and faecal 16S rRNA gene sequencing profiles [37,38], emphasise the importance of careful consideration of the type of specimen and its biomass when performing these comparisons. For example, previous studies have shown that low biomass breast milk specimens are influenced by contaminants and the DNA extraction method used [37], whilst diluted faecal specimen microbiota profiles are not much affected by contamination but largely impacted by the DNA extraction method used [38]. Our study showed that specimen biomass is a primary driver of sequencing profiles, and that the effect of DNA extraction kits become more evident with reduced specimen biomass. We further evaluated the effect of different storage buffers on sequencing profiles as storage buffers from different manufacturers may contain different levels of endogenous contaminants (contaminants introduced during the manufacturing process). Again, the effect of storage buffers on bacterial mock community profiles was more apparent among low biomass controls, which further highlights the importance of careful consideration when selecting NTCs for these types of studies [30]. If selected carefully, NTCs will ameliorate the detection and removal of "background OTUs", whether inherent to storage buffers, or originating from cross-contamination between wells [20] or DNA extraction methods used [37,38].
In addition to evaluating DNA extraction kits and storage buffers, 16S high-throughput studies also need to consider potential "batch effects" between independent runs. Large-scale experiments processed over long periods of time are highly susceptible to batch effects introduced by reagents (DNA extraction kits, PCR reagents and sequencing kits), laboratory personnel and laboratory environment [18,39]. Although downstream statistical tools have been introduced to correct for batch effects [40], it is crucial to minimize the probability of confounding during wet lab processes. Sources of variation arising from different biological groups (for example, specimens collected at different ages, from cases and controls, or from intervention and control groups) should be distributed evenly across different batches to avoid batch-effect biases. In order to evaluate batch effects across different runs, researchers need to carefully consider sequencing controls most suitable for their study design. Studies have shown that bacterial mock community controls are not only useful for evaluation of DNA extraction processes, amplification and sequencing steps [41][42][43][44], but may also provide insight into reproducibility across runs [45]. Although we support the use of mock community controls for reproducibility assessment [30]; these controls may not accurately represent inherent features of the biological specimens under study and should be supplemented with replicate biological specimens. Our study provides comparisons between biological specimens repeated within and between different sequencing runs and highlight their role in improving the quality of 16S rRNA gene sequencing data generated from low microbial biomass specimens [30]. As expected, our study showed slightly higher reproducibility between specimens repeated within the same runs compared to specimens repeated across different runs. However, we found that the biological nature of NP and IS specimens under investigation had a larger effect on reproducibility. Biological specimens of low biomass (< 500 16S rRNA gene copies/μl) showed lower sequencing reproducibility [46]. This also correlated with the age at which specimens were collected -the majority of specimens with low reproducibility (and low biomass) were collected during the first 2 weeks of life. The latter is an important observation as analyses may be biased when including "early life" specimens without incorporating the types of quality checks we report here. Hence, the sole use of bacterial mock communities may not be sufficient to determine reproducibility of low biomass specimens. Of note, correlations between bacterial biomass and reproducibility measures are only reliable when using appropriate quantification tools and we highly recommend the use of qPCR approaches targeting the 16S rRNA gene [47]. In support of the latter, studies have shown a general over-estimation of DNA concentrations measured by the NanoDrop® ND-1000 Spectrophotometer in comparison to fluorescence-based quantification methods such as the Qubit [48,49], whilst fluorescencebased quantification methods are less accurate than qPCR approachesparticularly for low biomass specimens [50][51][52].
Low biomass biological specimens produce less reproducible sequencing profiles compared to high biomass specimens as a result of contaminant profiles originating from reagents or the laboratory environment [15][16][17][18], and from neighbouring high biomass specimens [20]. Cross-contamination between wells may be minimised by the use of individual tubes as opposed to plates during DNA extraction and amplification processes, whilst reagent contaminants introduced during amplification processes may be reduced by PCR master mix decontamination protocols [53,54]. However, these precautionary steps may still need to be supplemented with in silico identification and removal of contaminants when dealing with low biomass specimens. In order to optimally identify and apply in silico decontamination to sequencing datasets, NTCs need to be selected carefully [16,55]. As endogenous NTC contaminant profiles are introduced to very low biomass biological specimens during the process of specimen collection and storage, neat storage buffers should be processed alongside biological specimens. In the event that storage buffers are not used to preserve specimens, the DNA extraction kit lysis buffer could be implemented as NTC during the process of DNA extraction. If DNA extraction kit lysis buffer is used as NTCs, it is recommended that the lysis buffer is exposed to items used during specimen collection (such as collection tubes, swabs, scoops, etc.) prior to the process of DNA extraction. The use of water as NTCs during DNA extraction steps is not recommended, as it may introduce external contaminant profiles which are not representative of the specimens or the DNA extraction kit [56][57][58]. Both exogenous and endogenous contaminant profiles present in NTCs add to the complexity of in silico contaminant identification as they may differ across batches of specimens collected, batches of reagents ordered, batches of DNA extractions and batches of library preparation. This highlights the importance of including a good representative of NTCs with each sequencing project -starting with specimen collection and ending with downstream in silico analyses.
Our study further showed that the in silico approach taken to identify and remove potential contaminants is just as important as identifying and including a good representative of NTCs during specimen collection, DNA extraction, library preparation and sequencing steps. We compared decontamination results from a widely used method in which all OTUs present in NTCs are flagged as contaminants and removed from biological specimens [16,18,19,28] ("NTConly" approach) to a more sophisticated method using the decontam package [29] ("NTC + decontam" approach). To date, few studies have performed these types of comparisons, especially in relation to respiratory tract specimens. In addition, we evaluated these two in silico decontamination processes by subtracting the maximum proportion of each OTU identified as a contaminant among NTCs, as opposed to the complete removal thereof. The decontam package [29] provides a statistical classification method for identifying contaminants derived from biological profiles based on their DNA concentrations [17,18,21,35,59]. This approach appeared to perform well on our dataset which showed clear trends between specimen biomass, diversity, reproducibility and the frequency of spurious OTUs. Our results showed that the identification of potential contaminants from NTCs without the decontam package ("NTConly" approach) more frequently identified bacteria commonly isolated from the nasopharynx as potential contaminants [15,60,61]. In contrast, the decontam package allowed for better discrimination between likely "true" and "contaminant" profiles at OTU-level, e.g., OTUs representing S. equi [62] but not S. pneumoniae, were identified by the "NTC + decontam" approach. This emphasises the importance of investigating sequencing data below genus-level resolution [using OTUs or amplicon sequence variants (ASVs)] in addition to genus-level analysis when considering potential contaminants. The result obtained for OTU_56 (S. maltophilia) and others, for example OTU-358 (A. calcoaceticus-A. baumannii complex), OTU-1390 (S. anginosus group) and OTU_863 (A. johnsonii), further highlight potential risks of using a relative abundance threshold to remove low abundant reads in silico [35]. All above-mentioned OTUs had less than 100 reads sequenced per specimen, however, may represent rare features truly present in biological specimens [31][32][33]63]. Also, by randomly setting thresholds to remove only low abundant reads from the dataset will not remove abundant contaminants which poses higher risks for downstream analysis interference. Finally, our results do not support the use of a "blacklist" approach to remove reads or taxa previously reported as common contaminants [21,22]. For example, at genus-level, Streptococcus, Staphylococcus, Haemophilus, Corynebacterium, Acinetobacter and Stenotrophomonas have all been listed as potential contaminants by previous reports [15,30], but are also common indigenous bacteria of the nasopharynx.
There are several limitations to this study. In the first section of our paper, only two DNA extraction kits were evaluated which do not represent the current assortment of kits available for DNA extraction from biological specimens. In addition, these kits were not selected for evaluation based on their aptness to extract DNA from low biomass specimens, but rather based on the fact that they were readily available for use in our laboratory. Further, limited numbers of repeats were included as mock controls during the evaluation of extraction methods. In the second section of our paper, we primarily included NP specimens to validate sequencing performance with limited numbers of IS specimens included for repeat processing. Finally, in the last section of our paper, we only assessed two in silico approaches for the identification of "potential contaminants". We did not investigate the use of tubes as opposed to plates during DNA extraction and amplification steps, nor did we investigate PCR master mix decontamination protocols [53,54] prior to amplification and sequencing.

Conclusions
We have described a stepwise approach to ensuring reliable, reproducible results from 16S rRNA gene sequencing of low biomass respiratory specimens. In our approach, we 1) investigate the performance of DNA extraction kits and the use of different storage buffers on low and high biomass mock community controls; 2) include relevant NTCs representing potential background OTUs present in biological specimens that undergo the same processing steps as biological specimens; 3) account for batch effects by including adequate sequencing controls; 4) establish specimen exclusion criteria based on reproducibility measures and clustering patterns as functions of specimen biomass, demographic data (such as "participant age at specimen collection"), and/or read counts; 5) denoise sequencing data by removing spurious OTUs; and 6) explore different in silico approaches to best correct for contamination. We show the value of detailed exploratory analyses of sequencing controls to identify and reduce experimental error.

Methods
We processed NP and IS specimens included in this study over a total of eleven sequencing runs as part of a study nested within the Drakenstein Child Health Study (DCHS) [26]. The DCHS is a population-based birth cohort study in a peri-urban area of South Africa which investigates the role of low biomass NP and IS microbial communities in the development of lower respiratory tract infection or wheezing illness during infancy and childhood. Different types of analyses, extraction and sequencing controls were included to address the three objectives of this study. A reference guide to DNA extraction kits, storage buffers/no template controls, bacterial mock communities, technical repeats and decontamination approaches used in this study is provided in Table 1.

Extraction and sequencing controls
In summary, each of the eleven sequencing runs processed as part of the study nested within the DCHS [26] consisted of four 96-well plates. We included controls for DNA extraction ("extraction controls") (Tables 1 and 2) and other downstream processes including PCR amplification, library preparation and sequencing steps ("sequencing controls") ( Tables 1 and 2).

Extraction controls
We addressed the first objective of our study by including four bacterial mock communities as extraction controls across two of the eleven sequencing runs ( Table 2). We used ZymoBIOMICS™ Microbial Community Standard bacterial cells (Catalog No. D6300, Zymo Research Corp., Irvine, CA, United States) ("Zymobiomics-Cells") to generate the four bacterial mock communities ( Table  2 Zymobiomics-Cells were mixed with two storage buffers [PrimeStore® Molecular Transport medium ("Primestore") [64,65] and a medium containing skim milk, tryptone, glucose, and glycerine ("STGG")], to further evaluate the effect of different storage buffers on 16S rRNA gene sequencing profiles generated from two DNA extraction methods. Primestore is used as a storage buffer for NP and IS specimens collected for molecular testing by the parent DCHS study [65], whilst STGG is widely used to preserve NP specimens intended for culture [65,66] and has also been used for 16S rRNA gene sequencing [67]. In order to limit reagent batch effects, we respectively pooled Primestore and STGG buffers with corresponding batch numbers prior to their use.
We generated two bacterial mock communities by suspending 900 μl of Zymobiomics-Cells in 3600 μl of each of the storage buffers, respectively (Additional file 1). These bacterial mock communities represented high biomass bacterial communities ("Zymobiomics-Primestorehigh" and "Zymobiomics-STGG-high") ( Table 2, Additional file 1). We further made a 1-in-10 4 fold dilution of each of the two high biomass bacterial mock communities ( Table 2, Additional file 1), each representing low biomass bacterial communities ("Zymobiomics-Primestore-low" and "Zymobiomics-STGG-low"). High and low biomass bacterial mock communities were included to evaluate the effect of bacterial biomass on resultant bacterial profiles following the use of two DNA extraction methods ( Table 2, Additional file 1).
Using two extraction methods (automated and manual), we extracted DNA in triplicate from each of the four bacterial mock communities (Zymobiomics-Primestore-high, Zymobiomics-STGG-high, Zymobiomics-Primestore-low and Zymobiomics-STGG-low). In preparation for DNA extraction using the automated method, we vortexed the bacterial mock communities for 5 s and transferred 400 μl of homogenised bacterial mock communities to ZR BashingBead™ Lysis Tubes containing 0.5 mm bashing beads (catalogue no. ZR S6002-50, Zymo Research Corp., Irvine, CA, United States). We performed a mechanical off-board lysis step at 50 Hz for 5 min using the TissueLyser LT™ (Qiagen, FRITSCH GmbH, Idar-Oberstein, Germany). We centrifuged the lysate at 10000 rpm for 1 min and loaded 250 μl of the supernatant onto the QIAsymphony® SP instrument (Qiagen, Hombrechtikon, Switzerland) for DNA extraction. We used the DSP Virus/Pathogen Mini Kit® (catalogue no. 937036, Qiagen GmbH, Hilden, Germany) ("Kit-QS") to carry out automated extractions of DNA with the elution volume set to 60 μl. For manual DNA extractions, we followed the manufacturer's instructions as outlined by the ZymoBIOMICS DNA Miniprep Kit (catalogue no. ZR D4300, Zymo Research Corp., Irvine, CA, United States) ("Kit-ZB"). We vortexed the bacterial mock communities for 5 s and transferred 250 μl of homogenised bacterial mock communities to ZR BashingBead™ Lysis Tubes containing a mix of 0.1 and 0.5 mm bashing beads (catalogue no. ZR S6012-50, Zymo Research Corp., Irvine, CA, United States) together with 750 μl ZymoBIOMICS™ Lysis Solution as per manufacturer's instructions. We performed a mechanical lysis step at 50 Hz for 5 min using the TissueLyser LT™ (Qiagen, FRITSCH GmbH, Idar-Oberstein, Germany) before continuing with the manual extraction steps as outlined by the manufacturer's protocol. We eluted the DNA in a final volume of 100 μl as per manufacturer's recommendations. We processed neat Primestore and STGG as no template extraction controls (Table 2) alongside the four bacterial mock communities using the two methods described above. All NTC extraction aliquots were obtained from the respectively pooled Primestore and STGG buffers used to generate the bacterial mock communities. This was done to limit reagent batch effects as the primary aim was to determine the effect of different storage buffers and bacterial biomass on extraction profiles using two DNA extraction kits.

Sequencing controls
We allocated six wells per 96-well plate to sequencing controls during each of the eleven sequencing runs. Sequencing controls included "bacterial mock community DNA", "technical repeats" and "no template controls" ("NTCs") ( Table 2). Technical repeats and NTCs were used to address the second and third objectives of our study.
Bacterial mock community DNA are commercially available pre-extracted genomic DNA mixtures from bacterial communities which did not undergo DNA extraction in our laboratory. We included a minimum of one of two types of bacterial mock community DNA controls per 96-well plate across the eleven runs Deinococcus radiodurans (0.02%) [27]). ZymoBIOMICS™ Microbial Community DNA Standard represents an even mixture of genomic DNA from eight genera (theoretical 16S rRNA gene composition: previously described for Zymobiomics-Cells in the DNA extraction section above). For the purpose of this study, we only report on data from BEI-DNA (n = 3) and Zymobiomics-DNA (n = 8) processed in the same sequencing runs as the bacterial mock communities used to compare DNA extraction methods. Zymobiomics-DNA was used as a reference when evaluating two DNA extraction methods.
Technical repeats refer to DNA extracts from NP and IS specimens, randomly selected for repeat amplification and sequencing across the eleven runs (Table 2). We extracted DNA from all NP and IS specimens processed across the eleven runs using the automated extraction platform and Kit-QS, as described above. We randomly selected a single biological specimen's DNA extract from any of the 90 biological specimens processed per 96-well plate for repeat amplification and sequencing ("withinrun repeats"). We further randomly selected a single biological specimen's DNA extract from any of the 360 biological specimens processed per run for repeat amplification and sequencing on an independent sequencing run ("between-run repeats"). Technical repeats were used to validate 16S rRNA gene sequencing data from low biomass specimens and to investigate decontamination processes.
In addition to technical repeats, we selected neat Primestore extracted using Kit-QS (processed alongside NP and IS specimens) as NTCs. The latter was included as NTCs due to the fact that NP and IS specimens collected for molecular analyses in the DCHS parent study (also included as technical repeats in this study) are stored in Primestore and extracted using Kit-QS [65]. We processed a minimum of one Primestore per 96-well plate, but only included Primestores with 16S rRNA gene quantification data for downstream analysis ( Table  2). All Primestores included in this study underwent DNA extraction, PCR amplification, library preparation and sequencing steps alongside technical repeats. Primestores processed across the eleven sequencing runs were used to evaluate the quality of 16S rRNA gene sequencing data and to correct for potential contamination inherent to 16S rRNA gene sequencing. For the purpose of comparing different extraction methods, we also included Primestore and STGG extracts (each extracted in triplicate using Kit-QS and Kit-ZB) processed alongside extraction controls (Table 2).

Amplicon library preparation and Illumina sequencing
We measured DNA yield and purity from extraction and sequencing controls using the NanoDrop® ND-1000. We further determined total bacterial load present in extraction and sequencing controls using a previously described qPCR method targeting the 16S rRNA gene [47].
We amplified the V4 hypervariable region of the 16S rRNA gene using a two-step amplification approach [68]. A total of 7 μl of DNA from all extraction and sequencing controls was included as template during the first PCR. During the second PCR, we used 7 μl of the PCR product from the first PCR as template to add adapters, barcodes, 12-15 staggered nucleotides and priming regions [68]. PCR conditions and modified primers used in the two-step amplification approach have previously been published [68,69].
We purified amplicons by adding Agencourt® AMPure® XP PCR Purification solution (catalogue no. A63881, Beckman Coulter, CA, USA) to amplicons from the second PCR at a 0.65:1 (bead:amplicon) ratio [68]. We used agarose gel electrophoresis and the GloMax®-Multi Detection System (Promega Corporation, Madison, WI, USA) together with the QuantiFluor® dsDNA System (catalogue no. E2670, Promega Corporation, Madison, WI, USA) to verify and quantify amplicons. Following quantification of amplicons, we pooled amplicons from each run (384 wells per run) at 70 ng and purified the pool using Agencourt® AMPure® XP PCR Purification solution at a 1:1 ratio. We quantified the purified pool using the Qubit® Fluorometer (Invitrogen, Life Technologies, CA, USA) and Qubit™ dsDNA BR Assay Kit (catalogue no. Q32850, Invitrogen, Life Technologies, CA, USA). We loaded a total of 7000 ng of the purified pooled library on a 1.6% agarose gel and performed electrophoresis at 35 Voltz for 30 min, 40 Voltz for 45 min, 70 Volts for 180 min and 50 Voltz for 60 min. Purification of the excised 16S library using the QIAquick Gel Extraction kit (QIAgen, MA, USA) followed agarose gel electrophoresis [68].
For each of the eleven runs, we quantified libraries and determined their fragment sizes using the KAPA Library Quantification Kit (catalogue no. KK4844, KAPA Biosystems, Boston, MA, USA) and the Agilent DNA 1000 kit (Agilent Technologies, CA, USA), respectively. We diluted libraries to 4 nM using Buffer EB (Qiagen, Hilden, Germany), after which we denatured and neutralized libraries using 0.2 N NaOH and HT1 hybridization buffer. Over the span of eleven sequencing runs, we gradually increased library concentrations with each run (ranging between 4 pM and 7 pM per library) to reach optimal flow cell loading concentrations. Each sequencing run contained the PhiX internal sequencing control spiked at 15%. We loaded denatured libraries onto the MiSeq Reagent Kit v3 (600-cycle) (Illumina, San Diego, CA, USA) and performed sequencing on the Illumina® MiSeq™ platform as per manufacturer's instructions [70,71].

Bioinformatic steps
We used FastQC and MulitQC packages [72,73] to assess sequence quality of FASTQ files. We merged paired-end sequence reads and performed quality filtering using the UPARSE algorithm in USEARCH version 10.0; whereby UPARSE merge_fastq (fastq_maxdiff set to 3) and UPARSE filter_fastq (sequences truncated to 250 bp and fastq_maxee set to 0.1) commands were used, respectively [74]. We used USEARCH10 sortbysize [75] to de-replicate sequences occurring more than once; USEARCH10 cluster_otus (clustering radius set to 3) [75] to cluster sequences into operational taxonomic units (OTUs), USEARCH10 uchime2_ref tool [75] to remove chimeras; and USEARCH10 usearch-global [75] to determine OTU counts. We used the USEARCH10 uchime2_ref tool for further chimera detection and removal in addition to the USEARCH10 cluster_otus command as it provides a reference for common chimeric sequences, considering the low biomass nature of specimens included in our study. We used the Quantitative Insights Into Microbial Ecology (QIIME 1.9.0) suite of software [76] to assign taxonomy (using SILVA database [77,78] and a sequence similarity set to 97%) via the RDP classifier method [79] and the assing_taxonomy.py command [76]. The Nextflow tool [80] was used to loop the bioinformatics processing workflow.

Data analysis
We used R software version 3.5.1 and RStudio software version 1.1.456 [81] for data analysis and visualisation. Prior to analyses, we transformed count data to compositional data [82][83][84]. We conducted all analyses at OTUlevel.
In order to evaluate the effect of two DNA extraction methods (Kit-QS and Kit-ZB) on 16S rRNA gene bacterial profiles, we compared sequencing profiles from DNA extracts from high and low biomass bacterial mock communities and no template controls (NTCs). Each of these sets of controls were generated using two storage buffers (Primestore and STGG). We used Principal Coordinate Analysis of beta diversities such that the distances between points approximate the beta diversity between each pair. We conducted Permutational Multivariate Analysis of Variance (PERMANOVA) using the function adonis from the package vegan [85] on beta diversities with 1000 permutations. We further evaluated the performance of the two DNA extraction methods by comparing how efficiently they extract hard-and-easyto-lyse bacteria in relation to the reference profile (Zymobiomics-DNA). Since the data is compositional, the isometric logratio transformation (ilr) was applied with the [pivotCoord] function in the R package rob-Compositions [86] such that the bacteria of interest represented the pivot coordinate. This result in the data being represented in an equivalent Euclidean space where single factor analysis of variance (ANOVA) is performed, and Tukey Honest Significant Difference simultaneous confidence intervals computed.
We evaluated the quality of 16S rRNA gene sequencing data from low biomass specimens by comparing OTU-level profiles [including alpha diversity (Shannon diversity index [87])] generated from technical repeats and Primestore]. We investigated correlations between specimen biomass and specimen features including participant age at specimen collection, read counts and alpha diversity. Lambda scaled [88] logarithm of ratiotransformed data (log-ratio) biplots (incorporating data adjusted in a Bayesian context to remove zeros [89][90][91]), were used to compare bacterial profiles obtained from technical repeats and Primestore. The [vegdist] function offered by the vegan package [85] in R was used to calculate the Bray Curtis dissimilarity index [92][93][94][95] required for complete linkage (furthest neighbour) clustering analyses. The [hclust] function offered by die stats package [81] in R was used to perform unsupervised agglomerative clustering analysis at OTU-level. We calculated sequencing reproducibility [coefficient of determination in linear regression analysis (R 2 )] by comparing the proportions of each OTU present in a specimen to proportions present in their technical repeats [96].
We addressed two in silico approaches to correct for potential contamination inherent to 16S rRNA gene sequencing using technical repeats and Primestore in three steps. The first step entailed denoising of the dataset. We removed "spurious OTUs" (defined as OTUs with < 5 reads across all sequenced technical repeats and Primestore) from the dataset. During the second step, we removed biological specimens with 16S rRNA gene copy numbers < 500/μl as these low biomass specimens produced sequencing profiles with poor reproducibility. The 16S rRNA gene copy number cut-off of < 500/μl was based on data generated by this study (summarised in the previous section), which corresponds with findings from previous reports [46]. Following steps 1 and 2, we compared two in silico approaches ("NTConly" and "NTC + Decontam") for identifying "potential contaminants" from the dataset generated. Using the "NTConly" approach, we compared all OTUs sequenced from Primestore to those sequenced from technical repeats. If we observed a match between OTUs present in both Primestore and technical repeats, we referred to these OTUs as "potential contaminants". We subtracted the maximum proportions of each contaminant OTU present in Primestore from technical repeats. The second approach, "NTC + Decontam", used the decontam package in R [29] to identify potential contaminants. We implemented the isContaminant function [29] and a combination of the "frequencyand prevalence-based methods" offered by the decontam package [29]. The "frequency-based method" identifies contaminants based on the frequency of each OTU as a function of the concentration of specimen biomass. The "prevalence-based method" offered by the decontam package identifies contaminants based on the prevalence of each OTU in true positive (biological) specimens versus the prevalence in NTCs [29]. Following the identification of potential contaminants using the decontam package, we subtracted the maximum proportions of each contaminant OTU present in Primestore from technical repeats.
Additional file 1. Four bacterial mock communities used to evaluate the effect of DNA extraction methods, storage buffers and bacterial biomass on 16S rRNA gene sequencing profiles. Zymobiomics-Primestore-high and Zymobiomics-Primestore-low: ZymoBIOMICS™ Microbial Community Standard bacterial cells in DNA/RNA Shield™ (Zymobiomics-Cells) suspended in PrimeStore® Molecular Transport medium