Evaluation of microbiome enrichment and host DNA depletion in human vaginal samples using Oxford Nanopore’s adaptive sequencing

Vaginal samples from swabs mostly yield small amounts of DNA (< 40 ng/µl) for library preparation and subsequent sequencing, making PCR amplification often mandatory, which can introduce amplification bias and alter the microbial composition.

Influence of nonspecific amplification-based library preparation for the determination of microbial communities

The nanopore amplification-based library preparation kit (RPB004) uses transposase-mediated cleaving of DNA molecules to attach the primer binding sites for PCR amplification, which should reduce PCR amplification bias.

We initially assessed this bias by determining the microbial composition of the ZymoBIOMICS Microbial Community Standard²³ (control) by sequencing using the RPB004 PCR-based library preparation kit and compared the abundance of the different species to the native PCR-free library preparation kit (LSK109). DNA of the mock community cells was isolated simultaneously in three replicates to address experimental variations. Each replicate was sequenced with the LSK109 and the RPB004 kit. Accordingly, all samples have the same “lysis and DNA isolation” bias; therefore, the library preparation kits are the only parameter that differentiates the sequenced samples.

The reads were mapped against the microbial genomes via minimap2 v.2.19²⁴, counted via samtools depth v1.11²⁵ (bases sequenced per organism), and summarized via ggplot2 (Fig. 1). All reads could be mapped to the reference genomes of the mock community.

Both sequencing kits detected all ten organisms of the ZymoBIOMICS Microbial Community Standard and the results of the sample’s replicates exhibited only negligible deviation.

Compared to the expected abundance in the control, Gram-positive bacteria and yeast were underrepresented in the sequencing data obtained by both library preparation methods (amplification free: average 0.60 fold, min 0.34 fold, max 0.79 fold; amplification-based: average 0.70 fold, min 0.35 fold, max 0.97 fold), while Gram-negative bacteria were overrepresented (amplification free: average 1.84 fold, min 1.80 fold, max 1.88 fold; amplification-based: average 1.64 fold, min: 1.01 fold, max: 1.99 fold). The amplification-based library preparation approach shows a considerable difference to the PCR-free library preparation method for Pseudomonas aeruginosa (0.55 fold of PCR-based), Lactobacillus fermentum (0.47 fold of PCR-based), Staphylococcus aureus (2.21 fold of PCR-based), and Cryptococcus neoformans (1.57 fold of PCR-based). Six organisms show minor differences to the PCR-free library preparation (Bacillus subtilis, Enterococcus faecalis, Escherichia coli, Listeria monocytogenes, Saccharomyces cerevisiae, Salmonella enterica).

We expected the PCR-free library preparation approach to represent the microbial community standard more accurately since it was previously validated by other groups²⁶. However, it showed clear variation in abundances compared to the control, which might be attributed to the different cell disruption device used in this work. The description of the ZymoBIOMICS Microbial Community Standard states that it mimics a mixed microbial community of well-defined composition, containing three easy-to-lyse Gram-negative bacteria, five tough-to-lyse Gram-positive bacteria, and two tough-to-lyse yeasts. Thus, Gram-positive bacteria and yeast were underrepresented due to differential lysis rather than differences in the library preparation protocols. We did not observe a significant advantage or disadvantage in choosing the amplification-based library preparation method over the PCR-free library preparation method to assess the microbial composition as both similarly overrepresent Gram-negative bacteria (Fig. 1), but interestingly the PCR amplification-based library preparation seemed to represent the control slightly better than the PCR-free library preparation.

Adaptive sampling for metagenomes: enrichment or depletion?

ONT’s adaptive sampling method enables it to either enrich or deplete DNA. To evaluate if microbial target enrichment or host depletion is more suitable for human vaginal metagenome sequencing (providing more microbial sequencing data), we compared both methods against a control experiment without adaptive sampling.

We sequenced a human vaginal metagenome (87.93% human host contamination) from a pregnant woman to derive species information for the enrichment process first. In a second step, we performed a depletion experiment using a human genome as reference (GCF_000001405.39). Finally, we performed an enrichment experiment using nine bacterial genomes downloaded from NCBI as reference based on the most abundant identified species from the first control sequencing experiment (see “Methods” section: “Nanopore sequencing”).

Each read passing the nanopore during adaptive sampling was mapped against a single or multiple reference genome(s) (e.g., human reference genome or multiple bacterial genomes) while sequencing. The mapping occurred in intervals of several bases, and three types of decisions were made: (1) ‘no_decision’—the read has been continued and mapped against the reference(s) after several bases again (‘no decision’), (2) ‘stop_receiving’—the read was accepted and fully sequenced (‘accepted’), (3) ‘unblock’—the sequencing was immediately stopped and the read was rejected by reversing of the voltage (‘rejected’). The base pairs required until a decision has been made were summarised in Fig. 2 B for all reads. For both methods, read rejections occurred within approx. 400–800 bp. Accepting reads started at approx. 400 bp or 4000 bp for the enrichment or depletion protocols, respectively. More generally, read lengths of at least 400 bp were required for both adaptive sampling methods to start the individual reads’ decision-making process.

The enrichment experiment yielded a higher total reads’ number (5.67 million, 1.50 fold more than depletion, of which 5.44 million were rejected reads), followed by the depletion experiment (3.79 million reads, 1.39 fold more than the control, of which 3.07 million reads were rejected). The control yielded 2.73 million reads. One should note that experimental variations affect the total sequencing performance, but the yield increase via depletion of human sequences was further validated (see “Performance of human host depletion via adaptive sampling in human vaginal metagenomic samples”).

Due to short read lengths, which result from the high rejection rate and the fast decision process (Fig. 2B.2), the enrichment experiment yielded the least amount of total bases and microbial bases while ‘human depletion’ yields the most microbial bases (Table 1). Without adaptive sampling, the proportion of sequenced human reads was unsurprisingly highest (87.93%) but could be strongly reduced by the depletion approach to 34.73% and by the bacterial enrichment down to 8.29% (Fig. 2A). The ‘human depletion’ method rejected almost 81.01% of all reads, which was lower than the total abundance of human DNA in the control experiment, suggesting that the chosen human genome might be insufficient for a complete depletion of all human reads or the adaptive sampling process itself is prone to error. The bacterial enrichment method rejected 95.93% of all reads, which indicates that some bacterial reads were also rejected. We identified 5.48% of Gardnerella reads, 2.41% of Lactobacillus reads, and 2.20% of other microbial reads in the ‘rejected’ fraction of the bacterial enrichment experiment. Simultaneously, the proportions of essential vaginal microorganisms, like Lactobacillus and Gardnerella, could be increased by both methods but higher by the enrichment protocol.

We compared the proportions of bacterial genera of the experiments identified from the reads of the ‘accepted’ and ‘no decision’ category to validate whether the overall microbial composition was retained despite adaptive sampling (Fig. 2C). The human depletion method clearly showed very similar proportions to the control for 34 of 36 microorganisms except for Escherichia and Luteimonas. The difference in the proportions of these two organisms could be attributed to experimental variations, especially since their frequency in the control experiment was only 0.05% (Escherichia) and 0.03% (Luteimonas). Conversely, the enrichment method shows significant differences in most genera, including important vaginal microorganisms like Gardnerella, Lactobacillus, and Ureaplasma.

Therefore, we assume that an enrichment approach might be unsuitable for investigating the microbial composition between metagenomic samples if not all species can be reliably provided as target sequences during the enrichment. On the other hand, the ‘human depletion’ experiments maintained a comparable microorganism composition as the control experiments and considerably (53.20%) reduced the number of human reads, making it a robust choice for clinical metagenomic samples with high amounts of human host DNA.

Performance of human host depletion via adaptive sampling in human vaginal metagenomic samples

We collected 15 vaginal samples of pregnant women (see “Methods” section: “Sample selection”) with high proportions of host DNA (> 90%).

First, each of the 15 samples were sequenced without adaptive sampling serving as a control experiment and ground truth of their metagenomic composition to track possible changes introduced via adaptive sampling. We then sequenced the same isolated DNA from the control experiments while using adaptive sampling (human DNA depletion) and compared the overall sequencing performance to the previously sequenced controls (Fig. 3). Additionally, a negative control of a swab without patient material following the same sample gathering and sequencing approach yielded 126 reads (ranging from 10 to 4000 bp) but none of the reads were classifiable and might be attributed to some PCR-primer and sequencing adapter constructs.

All reads were taxonomically classified via centrifuge v1.0.4²⁷ to investigate their taxonomic composition (centrifuge database: Human-Virus-Bacteria-Archaea 01.2021)²⁸. 99.43–99.85% of the reads generated in the control experiments could be taxonomically classified, while 97.44–99.83% of the reads in the depletion experiments could be taxonomically assigned.

On average, depletion experiments yielded 1.71 fold (± 0.27 fold) more reads (including rejected reads) than the corresponding control experiments (Fig. 3A). This corresponds to a yield of ~ 1.7 flow cells from a standard Nanopore sequencing experiment, with the only difference being that unwanted DNA molecules (human) were only partially sequenced. Reads of the categories ‘no decision’ (average: 5.38%, ± 7.24%) and ‘accepted’ (average: 0.23%, ± 0.35%) contributed a small overall proportion of all sequenced reads due to the high amount of human DNA. On average, adaptive sampling categorized 92.05% (± 7.42%) of reads as ‘rejected’. ‘Accepted’ reads were comparably long (Fig. 3C) with a median read length of ~ 4000 bp, which is in line with previous results (Fig. 2B).

Samples of the control experiments contained 97.59% (± 7.63%) human reads, while bacteria reads made up to 2.22% (± 7.49%) (Fig. 3C). Without splitting into the three read categories, the depletion runs showed similar proportions of species like the control. The ‘rejected’ fraction of the depletion experiments contained 99.80% (± 0.09%) human reads and 0.02% bacterial reads (± 0.02%), indicating a very selective depletion process. Reads of the category ‘accepted’ contained low amounts of human reads 1.23% (± 8.13%) and 97.42% (± 8.38%) bacterial reads. Three samples contained 4, 20, and 92 reads only within the ‘accepted’-fractions and were excluded in the previous calculation. The ‘no decision’-fractions contained 62.37% bacterial reads (± 34.76%), but also 25.06% human reads (± 36.19%).

In summary, most of the not ‘rejected’ reads were placed into the ‘no decision’ category as the ‘accepted’ decision was rarely made. The ‘no decision’ category also included many human reads and most of the bacterial reads. Combining the ‘accepted’ and ‘no decision’ fractions of the bacterial reads generally yielded more microbial reads compared to the control experiment underlining the capabilities of adaptive sampling to increase the sequencing depth (Fig. 3C). Furthermore, the decision made for the ‘rejected’ category was remarkably accurate as it contained almost exclusively human reads. However, those human reads were still identified in the ‘accepted’ and ‘no decision’ fractions as observed in previous experiments. Thus, to reliably remove all human reads during sequencing seems rather elusive.

Depletion does not alter the species distribution in samples

Adaptive sampling selectively depletes human reads while simultaneously enriching microbial reads due to the increased sequencing depth but may alter species representation and metagenomic composition. We, therefore, compared the taxonomically classified reads of the 15 vaginal metagenomic samples in detail, to compare the individual abundance between control and the host depletion experiments (Fig. 4, Supplementary Figures S1, S2).

The proportion of each genus was calculated in relation to the total amount of reads generated for each sample, including ‘rejected’ reads for the depletion experiment. We only included genera with at least 30 reads to avoid over-interpreting uncertain taxonomic classifications. This resulted on average in five bacterial genera (min 1, max 20) and in six bacterial species (min 1, max 26; Supplementary Figure S2), which correspond well with the expected vaginal microbiome²⁹. Across all 15 samples, the bacterial proportion varied from the corresponding control experiments on average by 1.03 fold, indicating a similar representation of genus abundance levels by the depletion experiment (Fig. 4). Low abundance genera with reads counts between > 30 and < 100 showed higher discrepancies (± 0.23 fold). Higher read counts showed higher reliability (e.g. > 500 < 1000 reads (± 0.03 fold) and > 1000 reads (± 0.04 fold). This higher discrepancy in genera with a small number of reads is expected, as the experimental variability’s influence is more prominent. We did not detect any organisms that were found solely by only one method using the applied cutoff. Overall, both experiment groups performed highly similarly in relation to the detected genus abundance levels between the control and depletion experiments.

In addition to the abundance comparison, we assessed the Bray–Curtis-Dissimilarity and the Spearman correlation for a more robust statistical analysis. Low Bray–Curtis dissimilarity values for all samples (0.005–0.077) indicate high similarity between each pair of metagenomes (control and depletion), underlining previous findings using only the abundance comparison (Table 2). The Spearman correlation test resulted in statistical significance (2.02E−05 and 1.92E−06) and positive coefficients (rho: 0.98 and 0.93) for samples 1 and 9 (Table 2). In other words: if, e.g. Lactobacillus is highly abundant in the control of sample 1, it is also highly abundant in the corresponding depletion experiment. Samples 3, 10 and 15 were not included in the Spearman correlation calculation since only one pairwise case (organism) occurs in these samples. The P value indicates no statistical significance for the remaining samples, probably due to few pairwise cases (< 10 pairwise cases).