Antibiotic resistance potential of the healthy preterm infant gut microbiome

Background Few studies have investigated the gut microbiome of infants, fewer still preterm infants. In this study we sought to quantify and interrogate the resistome within a cohort of premature infants using shotgun metagenomic sequencing. We describe the gut microbiomes from preterm but healthy infants, characterising the taxonomic diversity identified and frequency of antibiotic resistance genes detected. Results Dominant clinically important species identified within the microbiomes included C. perfringens, K. pneumoniae and members of the Staphylococci and Enterobacter genera. Screening at the gene level we identified an average of 13 antimicrobial resistance genes per preterm infant, ranging across eight different antibiotic classes, including aminoglycosides and fluoroquinolones. Some antibiotic resistance genes were associated with clinically relevant bacteria, including the identification of mecA and high levels of Staphylococci within some infants. We were able to demonstrate that in a third of the infants the S. aureus identified was unrelated using MLST or metagenome assembly, but low abundance prevented such analysis within the remaining samples. Conclusions We found that the healthy preterm infant gut microbiomes in this study harboured a significant diversity of antibiotic resistance genes. This broad picture of resistances and the wider taxonomic diversity identified raises further caution to the use of antibiotics without consideration of the resident microbial communities.


Introduction
Over recent years the composition of the gastrointestinal (GI) microbiota has been increasingly implicated in health and disease, with bacterial populations harbouring both Manuscript to be reviewed beneficial commensals and pathogens. A diverse bacterial population results in greater genetic content, but some of this additional genetic material is less welcome. Previous studies have implicated the GI microbiota as a reservoir of antimicrobial resistance (AMR) genes (Penders et al., 2013), held by, or capable of being transferred to, potential pathogens. Whilst often benign, during bacterial infection transfer of AMR genes can occur, which -coupled with selection pressures arising through antimicrobial therapy -can make treatment difficult, increasing the time taken to cure the infection. Furthermore, antimicrobial therapies are generally (ideally) tailored towards acute infections targeting a single pathogen, with little consideration of the wider microbial communities which reside in the microbiome, leading to a situation in which the use of antibiotics may cause unintentional harm to the host.
As our understanding of the microbiome has developed, the collection of AMR genes within a bacterial population has recently been defined as the resistome (Penders et al., 2013).
Antibiotics have a role in shifting the profile of the resistome within the population (Jernberg et al., 2007), with low antibiotic-use communities harbouring lower AMR gene frequencies (Walson et al., 2001;Bartoloni et al., 2009). Heavy treatment of bacterial populations with antibiotics can lead to the long term overrepresentation of AMR genes. Such dynamics are evident in the microbiome of preterm neonates, who receive multiple antibiotic courses, and are cared for in an Intensive Care Unit environment potentially contaminated with multi-resistant bacteria.
Antibiotic treatments for both term and preterm neonates have demonstrated lasting effects on the microbiota (Tanaka et al., 2009;Arboleya et al., 2015), with the trajectory of population development diverging from untreated controls, leading to a potential scenario of prolongedeven life-long -high frequency AMR reservoirs through the selection of bacteria within the population that are most resistant. A wide range of AMR genes have been found in neonatal populations (de Vries et al., 2011;Zhang et al., 2011), some shown to be present from birth Manuscript to be reviewed (Alicea-Serrano et al., 2013;Gosalbes et al., 2015), whilst twin pairs have been shown to have GI communities with similar distributions of both organisms and resistance genes (Moore et al., 2015). These observations suggest vertical transmission as a source, with discrepancies between mothers and babies being due to the substantial shifts in the microbiota adapting to the very different environment of a newly born infant's GI tract (Gosalbes et al., 2015).
The GI tract of a premature neonate is a particularly unusual scenario for observation of AMR genes, due to greatly reduced bacterial immigration as a result of the isolated, sterile environment of incubators and very controlled enteral feeds; donor breast milk may be pasteurised and, whilst unpasteurised maternal milk (which harbours specific bacteria (Beasley & Saris, 2004;Jimenez et al., 2008;Martin et al., 2009)) is given where possible, there is a likelihood of little or no breastfeeding due to extreme prematurity.
In these circumstances, the GI community and the resistance genes present are likely in the main to be derived from the mother, and acquired during birth. Whilst limited bacterial numbers and diversity will initially be transferred, mechanisms are available for the dissemination of AMR through the expanding bacterial population (as reviewed by van Hoek et al (2011)) with transfer having been documented within the gut environment (Shoemaker et al., 2001;Karami et al., 2007;Trobos et al., 2009). Heavy use of antibiotics in the course of care of premature infants would not only then skew the bacterial population and drive resistance selection, but has been shown to increase the activity of some transposable elements due to stressing of bacterial populations (Beaber, Hochhut & Waldor, 2004).
In this study, we present a detailed investigation of the resistome from the GI microbiota of eleven premature infants, with detailed information on antibiotic receipt and maternal antibiotic use. The microbiota of premature infants has been subjected to such investigations before, but through targeted techniques such as PCR or qPCR (Gueimonde, Salminen & Isolauri, 2006; Alicea-Serrano et al., 2013;von Wintersdorff et al., 2016) or through functional metagenomics (de Vries et al., 2011;Moore et al., 2015), which has the disadvantage of not being able to quantify the antibiotic resistance potential of a community (Forslund et al., 2014). We have used shotgun metagenomic sequencing to describe the resistome in its entirety, moving from species level taxonomic profiling, to characterisation of the resistance landscape, including typing of metagenomes identified as potentially harbouring mecA, conferring resistance to methicillin and other β-lactam antibiotics.

Study population
The study was approved by West London Research Ethics Committee (REC) Two, United Kingdom, under the REC approval reference number 10/H0711/39. Parents gave written informed consent for their infant to participate in the study.
Faecal samples analysed were collected from premature infants, defined as less than 32 completed weeks of gestation. Premature infants were recruited to the study at the Imperial College Healthcare National Health Service Trust neonatal intensive care unit (NICU), at Queen Charlotte's and Chelsea Hospital, between January 2010 and December 2011.

Sample collection
Almost every faecal sample produced by each participant between recruitment and discharge was collected by nursing staff from diapers using a sterile spatula. Samples were placed in a sterile DNase-, RNase-free Eppendorf tube, stored at -20 °C within two hours of collection and stored at -80 °C within five days. A single faecal sample from each of twelve infants who had no diagnosis of necrotising enterocolitis or blood-stream infection during their admission was selected for metagenomic sequencing. DNA from one faecal sample did not complete library preparation (see below); clinical characteristics of the remaining eleven infants and faecal sample metadata are presented in Table S1.

DNA extraction and shotgun library preparation
DNA extractions were performed as described previously (Rose et al., 2015), but with the following modifications: DNA extracts were prepared from approximately 200 mg of faeces, which were re-suspended in 10x volume:weight filtered 1x phosphate-buffered saline (PBS), with addition of 1:1 (volume:volume) 2% 2-mercaptoethanol diluted in 1x filtered PBS. The MolYsis selective lysis kit (Molzym) was used for the selective lysis of eukaryotic cells, incorporating the modifications previously described (Rose et al., 2015). Bacterial lysis was performed by addition of 50 µl lysozyme (Sigma), 6 µl mutanolysin (Sigma) and 3 µl lysostaphin (Sigma) to 100 µl of re-suspended bacterial pellet, and incubated at 37 o C for 1 h. This was followed by addition of 2 µl proteinase K and 150 µl 2x Tissue and Cell lysis buffer (Epicentre) and incubated at 65 o C for 30 min. Lysates were added to 2 ml tubes containing 0.25 ml of 0.5 mm beads and beaten on a Fast Prep 24 system at 6 m/s for 20 s and repeated once after 5 min. Finally, DNA was purified using the MasterPure complete kit (Epicentre) according to the manufacturer's instructions, eluted in 50 ul 0.1 x TE buffer (Sigma) and stored at -80 o C.
Extracted DNA was fragmented using the NEBNext dsDNA fragmentase kit (NEB) according to the manufacturer's instructions. Shotgun DNA libraries were subsequently prepared using the KAPA HyperPrep kit (KAPA Biosystems) according to the manufacturer's instructions. Ligated libraries were amplified by PCR with the number of cycles being dependant on starting material biomass, varying between 2 and 8 (mean 3 cycles). A negative extraction control was included consisting of 1 ml filtered 1x PBS and processed alongside the samples. After library amplification, the negative extraction control and one preterm infant faecal sample required >8 PCR cycles owing to very low starting pre-PCR biomass (DNA concentration <0.1 ng/ul), therefore these samples were excluded from downstream analysis, leaving faecal samples from eleven premature infants.

Shotgun metagenomic sequencing
Library insert size and quantity was assessed for each sample by Bioanalyser and qPCR as described previously (Rose et al., 2015).
Following assembly, very short binned S. aureus contigs (<1 kb) and partial assemblies, in this case those with less than half of the median S. aureus genome size (<1.5 Mb), were excluded positive result based on on fragment sizes ± 5 bp of those expected, and peak concentration >= 500 pg/ul. In addition to the above controls, extraction and PCR negative controls were included, which substituted input genomic DNA for purified water.

Statistics
Species richness and the evenness of their abundance were quantified using the Shannon-Weaver index ecological measure, calculated within MEGAN. Visualisation of samples was performed by hierarchical clustering using the UPGMA method and principal coordinates analysis (PCoA), all based on a matrix of Bray-Curtis distances calculated within MEGAN.
Correlations and t-tests were performed within R (v 3.2.5) (R Developement Core Team, 2015).

The healthy preterm metagenome
Using shotgun metagenomic sequencing we have captured an early snapshot of the antimicrobial resistance landscape within the gut microbiota of eleven premature infants who did not have proven sepsis or necrotizing enterocolitis. Infants were born either vaginally (N=6) or by caesarean section (N=5), with gestational ages ranging 24-31 weeks (mean 26.9 weeks). Ages of the infants at which the samples were taken ranged from 5 -43 days (mean 25.7 days) (

Manuscript to be reviewed
The eleven sequenced samples and four replicates were analysed using a blastx type analysis with filtering by the Lowest Common Ancestor (Huson et al., 2007;Buchfink, Xie & Huson, 2014), which enabled assignment of taxonomic labels for 71.5% of the reads within the complete dataset to at least the level of Kingdom (Table S2). As an alternative method, we also profiled the dataset using a marker based approach (Truong et al., 2015), which was highly congruent to species level relative abundances, as well as higher taxonomic ranks, to the blast based method used (Pearson R = 0.9 -1.0) ( Table S3). Replicate sequencing of samples also demonstrated reproducibility of the method, either by cluster analysis (Figure 1) or pairwise correlations ( Figure S1).
Moving to taxonomic composition, each sample was marked by a few highly abundant species, such as sample Q216 with 85.1% Clostridium perfringens, Q189 with 73.1% Klebsiella pneumoniae, and Q83 with 85.9% Enterococcus faecalis ( Figure 1A). In terms of prevalence, the previous three species, as well as Enterobacter cloacae and Staphylococcus epidermidis, were found at over 50% relative abundance in one or more samples. Furthermore, S. epidermidis and S. aureus were ubiquitous, ranging from 0.06% to 57.1% abundance in all samples ( Figure 1B).
Principal coordinate analysis (PCoA) demonstrated three loose sample groups based on a high abundance of S. epidermidis, K. pneumoniae, and either B. breve, S. aureus or C. perfringens ( Figure 1C). In total we identified a non-redundant set of 172 species across all samples (see Table S4 for complete dataset).

Prevalence of antimicrobial resistance
Before focusing on individual AMR genes, we measured the α-diversity (Shannon-Weaver index) within each sample and, although a small sample set, compared this to the known Manuscript to be reviewed antibiotic exposure of the preterm infants ( Figure S2). Eight of the eleven infants had received a course of prophylactic antibiotic treatment consisting of co-amoxiclav (Table S1), whilst a second course was administered to four infants, consisting of combinations of co-amoxiclav, tazocin or vancomycin. In total, exposure ranged from 2 -8 days of antibiotics before samples were taken, excluding infants Q87 and Q89 which received no antibiotics. Antibiotics were also administered maternally to three infants (Q26, Q117 & Q189), but this did not include the two above infants with no antibiotic treatment. Sample diversity ranged from 0.9 -2.9 (SD ± 0.5), but when compared to cumulative antibiotic exposure expressed in days, no significant difference was found between the taxonomic diversity and amount of antibiotic exposure for untreated and treated infants (unpaired t test, P = 0.17) ( Figure S2). It is important to stress however that the small and heterogeneous nature of the sample set will have reduced the power to detect differences between antibiotic exposure in this study, and so prevented any meaningful stratification by other clinical variables such as mode of delivery or day of life.
A mapping based approach against a comprehensive collection of acquired antibiotic resistance genes was next used to quantify AMR within the eleven metagenomes (Inouye et al., 2014). In total 143 AMR genes were identified, consisting of a non-redundant set of 39 different AMR genes (Figure 2 and Table S5). Per infant, the average number of genes identified was 13 (ranging 5 -22 genes), and AMR genes were found across eight different antibiotic classes, including aminoglycosides and fluoroquinolones (Table S6). Mean sequence coverage across the sequence database was 99.0%, and sequence divergence ranged from no difference to 12.3% (Table S6). In total over 1,600 alleles were searched for, and notable AMR genes not detected included those involved in carbapenem and vancomycin resistance, the latter of which was administered to three preterm infants prior to sample collection (Table S1). The class most frequently detected were β-lactamases, comprising ten different genes (Table 2) Manuscript to be reviewed blaZ gene was present in every infant. Interestingly, within this set of β-lactamasae genes, mecA was found in four infants (Q87, Q117, Q175, and Q189), and at a mean depth of coverage ranging from 3.9 to 52.2 bp (Figure 2). mecA confers resistance to methicillin as well as other βlactam antibiotics, and is carried on the SCCmec mobile element found across several Staphylococci species. Identification of four infants with potential methicillin resistant S. aureus (MRSA) or S. epidermidis (MRSE) carriage, along with high abundances and prevalence of both S. aureus and S. epidermidis species across the dataset, could indicate a significant reservoir for AMR transfer between the species, as well as highlight the seeding of the infant gut microbiome from an early stage.

Focus on S. aureus species detected
We next wanted to understand the relationship of the S. aureus species within the mecA positive as well as negative samples, as the premature infants overlapped in time and so could harbour closely related strains. This was undertaken to firstly confirm in silico prediction of mecA using an established molecular based typing method, but also to push the metagenomic analysis further on what was known to be a challenging dataset owing to the range of identified S. aureus as described above, with relative abundance ranging from 0.06% -39.8% (Table S4).
We first tested the computational prediction of mecA experimentally using a multiplexed PCR typing method (Milheiriço, Oliveira & De Lencastre, 2007), which provides detection of the  (Table S7). The exception, sample Q87, generated the expected mecA amplicon size but the concentration of this fell below the threshold for detection (<500pg/ul) and so was excluded.
In an attempt to understand strain relatedness directly from the metagenomic data we undertook in silico MLST analysis using an S.aureus schema as well as metagenome assembly.
The MLST was able to classify four of the eleven samples, all with different ST types -ST8, ST1027, ST22, ST25, although the last two had some degree of uncertainty in their assignment (Table S8). This suggests that for at least these four samples, the S. aureus strains are unrelated and unlikely a result of transmission. We were interested to know if de novo assembly of the metagenome could be utilised to resolve these and any of the remaining unclassifiable samples further. Following assembly and identification of S. aureus contigs (see Methods), we found that it was not possible to capture more than a fifth of the expected genome size for the above unclassified samples, with an abundance of > 3% necessary to achieve over 90% estimated capture, which was achieved in four cases (Table S9). Phylogenetic reconstruction of these four genomes alongside a collection of published S. aureus genomes (Table S10), provided confirmation of the diversity of S. aureus identified ( Figure S3), enabling placement across a global collection of strains.

Discussion
It is recognised that one of the most important public health threats worldwide is antimicrobial resistance. Here we report on the gut composition and AMR diversity for eleven healthy but premature infants. Recent studies have shown that the initial seeding of the infant gut microbiome is influenced by the microorganisms in the immediate environment, and whilst colonisation by bacteria with AMR genes has been demonstrated (Brooks et al., 2014), 'stable' microbiome. Although we found no correlation between diversity and antibiotic exposure, with infants treated with either no antibiotics (including during pregnancy), to up to 8 days of antibiotic administration, effects such as relatively small sample size, as well as day of life of sample and normal gut development are biases to this finding, which is contrary to other studies within infants (Greenwood et al., 2014;Merker et al., 2015). It could be that at this very early stage, the microbiota is influenced to a greater extent by seeding during birth from the mother and environment than antibiotic treatment, or that not enough time has passed to detect differences from the antibiotics administered; larger sample numbers would be required, A threat to this development is the acquisition of antibiotic resistant bacteria, which can potentially seed the infant microbiome. Coupled with the high rate of horizontal gene transfer within the commensal community (Stecher et al., 2012), the preterm infant gut microbiome has the potential to be a reservoir for AMR. With dominance of the preterm gut by species known to carry clinically relevant antibiotic resistance, we next quantified the burden of antibiotic resistance genes within the infant's faecal flora, which identified an average 13 genes per infant.
Previous targeted or functional studies based on infants have found some of the AMR genes also identified here, including those for Tetracycline (tet) (Gueimonde, Salminen & Isolauri, 2006; Alicea-Serrano et al., 2013) and β-lactam (bla) (Fouhy et al., 2014). In a wider context, it is known that AMR genes are a common feature of bacterial populations, found in communities inhabiting the soil, rivers and even deep-sea sediment (Knapp et al., 2010;Qin et al., 2011;Kittinger et al., 2016). Therefore, whilst their presence in the human gut microbiome should be of little surprise (Bailey et al., 2010), identification of genes such as mecA demonstrates the prevalence of some clinically significant resistant bacteria from birth.
One of the advantages of the method used in this study is the utility of the results generated, enabling multiple avenues of questions to be addressed. However, short read sequencing remains a challenge when applied to the linkage of resistance elements, such mecA, to specific genome sequences (strains), which is made difficult by the nature of metagenomic samples containing multiple alleles from different closely related species, as well as potentially multiple strains of the same species. Secondly, the methods used here were inherently restricted to identification of known AMR genes found within the ARGannot database used in this study, which contains those genes involved in acquired resistance only, therefore chromosomal mutations, such as those conferring resistance to rifampicin as well as novel resistance genes would have been missed, leading to potential underrepresentation of resistance in this study.

Conclusions
The healthy preterm infants sampled within this study harboured multiple AMR genes, representing a potential reservoir for later disease onset. In particular, detection of clinically important AMR genes, such as mecA, highlights the need to further understand the impact that this reservoir could have on later treatment regimes.. From a methodology point, this approach was able to provide a comprehensive snapshot of the complete taxonomic diversity and resistome in one assay. Although tracking of the movement of such AMR genetic elements would be enhanced by improved handling of the dynamic ranges of abundances; different methods at the level of sample preparation, such as sample normalisation, may offer potential answers to such hurdles. Overall this study leads to questions such as how this resistance potential contributes to later clinical intervention or disease onset, and if antibiotic treatment without knowledge of prior AMR burden could lead to unintentional harm. More broadly, this and other studies show the great promise that shotgun metagenomics holds for clinical microbiology.