Integrated next-generation sequencing of 16S rDNA and metaproteomics differentiate the healthy urine microbiome from asymptomatic bacteriuria in neuropathic bladder associated with spinal cord injury

Patients and healthy controls were recruited into this IRB-approved study (NRH IRB# 2011–019) from the outpatient clinic and inpatient ward at National Rehabilitation Hospital (Washington, DC). Following written consent, urine samples were obtained from 26 healthy, non-SCI controls and 27 with neuropathic bladder (NB) due to spinal cord injury (SCI). Patients provided urine samples by sterile collection using the means by which they customarily empty their bladder (i.e. midstream collection during voiding, or sterile catheterization if unable to void). (Table1) The samples were coded with an anonymous research identification number and separated into two aliquots: one, for standard urine analysis and culture (Quest Diagnostics) and another for microbiome analysis. Quest Diagnostics performed analysis of urine samples for nitrite formation, leukocyte esterase, and microscopic examination for the presence and quantity of leukocytes and erythrocytes in each sample. Bacterial cultures were performed by inoculation of blood agar plates and incubation at 37°C for 48 hours.

Thawed urine samples were clarified by low-speed centrifugation and bacterial genomic DNA was extracted from urine pellets by enzymatic digestion using a final concentration of 20 mg/ml Lysozyme (Invitrogen) followed by physical lysis using Lysing Matrix B tubes (QBiogene). A previous study comparing mechanical and enzymatic methods for extracting microbial genomic DNA showed that mechanical cell disruption by bead beating produced the highest bacterial diversity[13]. The samples were vortexed at maximum speed for 45 seconds using a Fastprep fp120 (MP Biomedicals) then cooled on ice. DNA was extracted from the lysate using phenol chloroform isoamyl alcohol extraction and ethanol precipitation. 16S rDNA sequences were generated by amplifying the V1-V3 region of the bacterial 16S rRNA gene using primers 27F and 534R fused with 454 adaptors and barcodes for multiplexing. Primers targeting V2 and V3 were shown to perform as well as full-length 16S rDNA sequence for community clustering and taxonomic assignments[14].

The amplicons were normalized and pooled prior to emulsion PCR and 454 sequencing (Roche, Inc.) using titanium chemistry.

The 16S rDNA sequence-processing pipeline used for this study is composed of a selection of bioinformatics tools proven to be accurate, robust and fast. A supplementary archive contains the dot language representation of a graph depicting the entire workflow executed, outlining the specific parameters used for each command.

Initially, the SFF file, output from the sequencer, was converted into fasta and qual files using the sffinfo program included as a part of 454/Roche software package. Subsequently, the trim.seqs function in mothur[15] (version v.1.22.2) was used to de-multiplex sequencer reads. No barcode mismatches, and up to one primer mismatch were allowed past this step. The de-multiplexed reads were processed using LUCY[16‐18] to filter out reads with low quality segments. At this point, the sub.sample function of mothur was used to select an equal number of reads per biological sample (n=3671 based on the biological sample with the fewest number of reads). Subsequent, the screen.seqs function of mothur was used to remove sequences shorter than 220 bases[17]. Furthermore CD-HIT-454[19, 20] was used to collapse duplicate reads, while retaining their count for subsequent enrichment statistics, (analogous to the functionality of the unique.seqs function of mothur, but orders of magnitude faster and less demanding on the computer hardware). The sequences were aligned against the SILVA database of 16S rDNA sequences[15, 21] to verify 1) the orientation of noise-filtered sequences; and 2) the correct positioning of the reads with respect to the expectation of which variable regions should have been amplified and sequenced. Thereafter, the remaining sequences were subjected to mothur’s implementation of chimera slayer[15, 22] to filter out chimeric reads. The processed 16S rDNA data from this study can be obtained at NCBI under BioProject ID 97505.

Taxonomical classification of the final set of 82160 operational taxonomic unit (OTU)-representative reads down to the genus level was performed using mothur’s version of the RDP Bayesian classifier using a normalized RDP training dataset[23]. The final step of the pipeline clustered the sequences based on their similarity to produce OTUs. Customarily, a similarity threshold of 97% has been used to define OTUs at approximately the species level[24]. A module of CD-HIT suite[19] called CD-HIT-EST was employed to perform species-level read-clustering for subsequent analyses.

The orchestration and automation of steps has been achieved using a custom set of in-house utilities written in python and R[25] programming languages. These utilities are available online as a part of the YAP package on github[26]. JCVI grid infrastructure based on the Oracle Grid Engine (OGE) was used for all steps described. Relative abundance and diversity statistics were calculated within mothur[15]. Further statistical analyses were accomplished using R. Heat maps were generated using the heatmap.2 function of the gplots package available on CRAN[25]. OTU counts have been normalized to 100% per individual, to facilitate comparability. Only taxa (order or genus) with a standard deviation greater than 5% across all 52 individuals were used to generate the heat maps. Differences between subject OTU communities were assessed using the Bray-Curtis beta diversity statistic implemented in the vegdist function, a part of vegan R package available on CRAN. Clustering was accomplished using average-neighbor-joining method implemented in hclust function in the default installation of R. PCA analysis was performed using ade4[27]. P-values used to determine statistical significance in relative OTU difference plots were established using the default installation of R and kruskal.test functions implementing the Kruskal-Wallis rank-sum statistic test[28].

OTU-representative sequences classified as either Lactobacillales or Enterobacteriales were aligned using tools available from Release 10 of the RDP web site[29]. Specifically, sequences from OTUs composed of reads from more than one individual were aligned to the RDP reference 16S rRNA sequence, taking into account secondary structure. At most one nearest neighbor sequence from RDP was recruited into the alignment per input sequence. The alignment was downloaded and trimmed to remove columns whose gap fractions were greater than 50%, using Belvu[30]. Based on the alignment, a bootstrapped Neighbor-Joining (NJ) tree was subsequently inferred using paupFasta, an in-house wrapper script around the PAUP* program as described[31], and edited using Fig Tree[32]. In combination with nearest neighbor taxonomy, BLASTN was used against the NCBI reference RNA database, which lacks uncultured organisms, to identify certain Lactobacillales OTU-representative branches that lacked classified RDP top matches, to the species level, using a cut-off of 97% identity.

A urinary pellet specimen equivalent to 5–10 ml voided urine washed two times with ~10 ml ice-cold PBS was re-suspended in 1 ml of 10 mM ammonium bicarbonate containing 0.1% Triton-X100, 0.5% octylglucoside, 5 μg/ml leupeptin, 10 mM EDTA and 2 mM benzamidine. The suspension was heated to 85°C for 5 min followed by sonication (amplitude 4, Misonex 3000 sonicator) in 30s on/15s off cycles 10 times on ice. The suspension was centrifuged for 15 min at 16,100 x g and the supernatant recovered. Following protein quantity estimates based on SDS-PAGE analysis, an aliquot with ~10 μg protein was digested at a 1:50 ratio (trypsin/protein) using Filter-Aided Sample Preparation (FASP)[33]. The digestion mixture was reconstituted in 50 μl 0.1% formic acid. For shotgun proteomic analysis, peptides in a 20 μl sample aliquot were separated on a capillary C₁₈ LC column in 122 min binary gradient runs from 97% solvent A (0.1% formic acid) to 80% solvent B (0.1% formic acid, 90% acetonitrile) at a flow rate of 350 nl/min. Nano-electrospray into the source was followed by mass analysis and spectral acquisition in automated MS/MS mode, with the top five parent ions selected for fragmentation in scans of the m/z range 350–2,000 and a dynamic exclusion setting of 90 sec. The LC-MS/MS workflow using the LTQ-XL ion trap system (Thermo Fisher Scientific) was previously described in more detail[34]. The instrument was calibrated prior to performance of LC-MS/MS experiments with 200 nmol human [Glu¹-fibrinopeptide B (M.W. 1570.57), verifying that peaks representing ion counts had widths at half-height of <0.25 min, signal/noise ratios >200 and peak heights >10⁷. The LTQ search parameters (+1 to +3 ions) included mass error tolerances of ± 1.4 Da for peptide precursor ions and ± 0.5 Da for peptide fragment ions, selecting monoisotopic m/z values. The search engine used to select these parameters and identify peptides and proteins was Mascot v.2.3 (Matrix Science). The protein sequence database was comprised of 19 bacterial genomes (details in Additional file1) and Uniref90’s human database subset[35]. We limited this database to the human Uniref90 subset and 19 bacterial genomes frequently associated with ABU and UTI, because significant computational challenges for peptide-spectral match (PSM) assignments are encountered when very large protein databases are used[36]. As described in Additional file1, we used stringent criteria for PSMs (q-value <=0.01; PEP-value <=10^-4) using the Mascot Percolator algorithm that improves discrimination between correct and incorrect PSMs, particularly when the searched sequence space in the database is large[37].

To determine whether distinct microbial signatures were associated with gender (male versus female) and/or bladder function (healthy control versus NB), samples were grouped by these variables. Taxonomic counts were then normalized by total number of OTUs per sample, and visualized by a heatmap, clustering the distribution of OTUs at the level of bacterial genus using the Bray-Curtis index (Figure1B). The samples were colored by relative abundance (red/warm most abundant to blue/cool least abundant). Six main taxonomic profile clusters emerged with distinct patterns when grouped by gender and bladder function. Cluster 1 was composed of almost an equal proportion of healthy and NB samples, clusters 2 and 3 were dominated by healthy controls, whereas clusters 4–6 were entirely composed of patients with NB (Figure1B). The 2 “healthy” clusters, (2 and 3) were distinguished by gender, with females in cluster 2 and males in cluster 3, and by bacterial genus, Lactobacillus grouping with females in cluster 2 and completely absent in the male-dominated cluster 3. Cluster 3 was composed of different Gram-positive organisms. Cluster 1 had the most diverse bacterial genus profile, composed largely of Lactobacillus, but not as abundant as cluster 2 with elements of cluster 3 and a few potentially pathogenic genera (Enterococcus, Salmonella, and Peptoniphilus). Closer inspection of clusters 4–6 showed a very different pattern of bacteria, with known UTI pathogens dominating the profiles (Figure1B). Cluster 4 was primarily composed of Enterococcus, Escherichia and Salmonella. Cluster 5 was dominated by Klebsiella sp. and was the only cluster comprised of all males. Cluster 6 had the most Enterococcus counts of any cluster and also contained Aerococcus and Proteus sp.

By taking the difference in normalized relative abundance between controls and SCI groups, the most relative abundant bacterial taxa per group as confirmed (Figure1C). Statistical significance was established using the Kruskal-Wallis test. Significant differences (P < 0.05) between the top relative OTU counts suggest that Lactobacillus, Corynebacterium, Gardnerella, Prevotella and Enterococcus define gender differences (“+” in Figure1C). Lactobacillus, Klebsiella, Corynebacterium, Staphylococcus, Streptococcus, Gardnerella, Prevotella, Escherichia and Enterococcus defined statistically significant differences between healthy bladder controls and NB (“*” in Figure1C).

To further investigate the dynamics of the urinary microbiome in NB, the NB group was divided into bins based on the time post SCI. The bins consisted of 0–2 months (n=5), 3–12 months (n=5), 13–48 months (n=3), and 49+ months (n=14). Principal component analysis (PCA) revealed that the healthy control group and the 0–2 month duration were very similar (Figure2). Likewise, the 13–48 and 49+ month groups were similar to each other, while the 3–12 month group was too distorted by one outlier to provide meaningful information. Enterococcus and Escherichia emerged as major contributors to these profiles as illustrated in the vector diagram (inset, Figure2).

To get an overview of bacterial species variation within the NB group, differences in microbiome taxonomic profiles between males and females and among bladder management groups were plotted (Figure3). Lactobacillus species were significant (P < 0.05) contributors to female healthy controls and the void NB groups, while Corynebacterium sp. defined the healthy bladder male urinary microbiome (Figure3). The preponderance of Enterococcus sp. in male NB observed in Figure1C was primarily within the patients with NB who emptied by spontaneous voiding (Figure3).

A phylogenetic tree was constructed with OTU representatives of all OTUs with matches to Lactobacillales and nearest neighbors in the RDP database in order to provide an environmental context and determine species-level taxonomy (Figure4). Labels of the tree were colored based on NB composition. Red indicates OTUs composed only of samples from NB, light red indicates OTUs with a majority composition of NB samples, while dark blue denotes OTUs only solely of healthy controls, and light blue illustrates OTUs with a majority of reads from healthy controls. There is much greater phylogenetic diversity among the Lactobacillus and Streptococcus branches (green and gray shading, respectively, Figure4). Aerococcus and Enterococcus branches are dominated by OTUs (red leaves) from urine of NB patients, suggesting qualitatively that these two genera may be important indicators of bacteriuria, or of catheter usage. In contrast to these two branches, the Lactobacillus and Streptococcus branches are largely comprised of OTUs from healthy controls with a few exceptions (red leaves). One such exception was Lactobacillus iners (Figure4), which was previously shown to significantly contribute to UTI[39, 40]. Another exception was determined to be Streptococcus salivarius, a lactic acid-producing Gram-positive organism typically found in the oral cavity, and an opportunistic pathogen implicated in bacteremia[41‐43] and septicemia[44].

In contrast to the Lactobacillales, the Enterobacteriales had no OTUs composed solely of healthy controls; indicating this group of bacteria may be a potential indicator of future UTI (Figure5). Shaded regions of the tree noted unambiguous genus taxonomy as follows: Escherichia, Enterobacter, Klebsiella, Proteus, Morganella, and Providencia. All of these genera have been associated with UTI, and dominated the tree. Klebsiella stood out from the others, having been divided into three distinct branches. Top matches to N₂-fixing, plant-associated Klebsiella pneumoniae and K. variicola isolates rather than known K. pneumoniae clinical strains suggests that N₂-fixing, plant-associated Klebsiella may exist in urine. Further work is needed to confirm this result.

Shotgun proteomic data for a subset of nine urinary samples searching a protein sequence databases of 19 bacterial genomes were compared with the OTU analysis derived from 16S rDNA sequencing. In addition to protein identifications targeting the 19 species frequently associated with UTIs or urethra colonization, a non-redundant human protein database was searched to assess the detection of host responses towards bacterial colonization. While proteomic analysis identified more species than those diagnosed by culture methods, it did not identify the fastidious anaerobic or microaerophilic bacteria frequently assigned by the RDP classifier in the 16S rDNA analysis with the exception of Lactobacillus. We assume this to be due to species abundance issues. Proteins encoded by Escherichia, Klebsiella and Enterobacter species were identified in half of the examined samples, generally in agreement with 16S rDNA data (Table2). Metaproteomic profiles allowed preliminary insights into the production of bacterial factors interacting with the urinary tract environment. Identifications of such proteins as limited to those urine donor samples with evidence for the initiation of pro-inflammatory and microbicidal host defenses (high spectral counts for calprotectin subunits; identification of lactotransferrin, myeloperoxidase and eosinophil peroxidase). We emphasize that the sample sizes are too small to predict whether these observations will be reproducible using larger-scale proteomic surveys. Pseudomonas aeruginosa, Enterobacter hormaechei and E. coli (known opportunistic pathogens in the urinary tract) produced proteins for iron/siderophore acquisition and high mobility (flagellins) in subjects 36 and 51 (Table3). Flagellins are important for swarming and spread in the urogenital tract during infections[45]. The iron-sequestering protein lactotransferrin may be expressed and released into the urinary tract by neutrophils to sequester iron, which in turn, E. coli and E. hormaechei may counteract by addition of copious amounts of iron/siderophore receptors to the outer membrane protein repertoire (Table3). While the S100-A8 and S100-A9 calprotectins appear to have numerous physiological functions, one of them is to sequester zinc in response to sensing the presence of bacterial pathogens. This sequestration can inactivate metalloproteinases that the bacteria produce and secrete to invade the host tissue[46]. Such proteases were not identified in our datasets.

Several other proteins implicated in the innate immune response were detected (Table3). Calprotectin has acute and chronic pro-inflammatory functions and recruits immune cells to the site of inflammation[47]. Antimicrobial peroxidases, such as eosinophil and myeloperoxidase, produce reactive oxygen species during the respiratory burst of neutrophils and are directly microbicidal. Like calprotectin, annexin A1 is a calcium-binding protein that was detected in high abundance primarily in those donors where respiratory burst of neutrophils did seem to be muted (subjects 1, 16, 31, 34, 37, 39). These proteins are implicated in innate immunity, influence cell apoptosis and function as damage-associated molecular patterns (DAMPs)[48, 49]. For a complete list of bacterial and human proteins profiled in this study, see Additional file1.