Metabolite summary.

We detected 279 known metabolites in vaginal fluid from women with and without BV. There were significant differences in levels of 173 metabolites (62% of the 279 detected) on comparison of women with and without BV (q value = 0.02). Women with BV had higher levels of 55 biochemicals (20%) and lower levels of 118 biochemicals (42%). A summary of metabolites altered in BV based on pathway is provided in Table 2.

Hierarchical clustering depicting this set of the most variable metabolites measured in CVL samples from all women resulted in the creation of four groups, of which 2 clusters were populated by women with BV. Cluster I had 12 women without BV, and cluster II comprised 15 women with BV. Cluster III included 7 women without BV. Cluster IV included 25 women with BV and 1 woman without BV (subject 58) (Fig. 1). Vaginal fluid from subject 58, the outlier in cluster IV, had an elevated pH of 5.6 and a Nugent score of 6, defined as intermediate on this scale.

Open in a separate window

FIG 1 

Hierarchical clustering of metabolites. A dendrogram shows associations of 30% of most variable metabolites with BV status, determined using the Amsel and Nugent criteria. Clustering of the metabolites resulted in four groups (clusters I to IV). The clustering algorithm was not informed by BV status. Study participant identification numbers (IDs) are provided on the x axis, and metabolites are listed on the y axis. The heat map depicts log-transformed concentrations of the most variable metabolites between the clusters. Values ranged from −6.956 (dark green) to 4.818 (dark red).

Below, we summarize alterations in metabolism in BV for each of the areas outlined.

Amino acids.

We detected 19 of 20 important amino acids (95%) that are incorporated into proteins using the Metabolon platform (see Table S1 in the supplemental material). Levels of 18 of 19 amino acids detected (94.7%) were lower among cases; this was statistically significant for 16 of 18 amino acids (84.9%). We also noted higher concentrations of amino acid catabolites in women with BV (see Table S1). Examples of such catabolites include cadaverine (P < 0.001) and pipecolate (P < 0.001) in the lysine degradation pathway, tyramine (P < 0.001), 4-hydroxyphenylacetate (P < 0.001), and 3-(4-hydroxyphenyl)propionate (P < 0.001) in the tyrosine pathway, tryptamine (P < 0.001) in the tryptophan pathway, and citrulline (P < 0.001) in the arginine pathway. There was marked elevation of the polyamine putrescine (P < 0.001), along with lower levels of putrescine precursors arginine (P < 0.001) and ornithine (P < 0.001), in women with BV. The polyamine spermine (P < 0.001), a typical product of putrescine degradation, was lower in BV. Enhanced protein/amino acid catabolism in BV was further supported by detection of lower levels of dipeptides (n = 28); this was statistically significant for 24 dipeptides (85.7%). Levels of oxidized (P < 0.001) and reduced (P = 0.049) glutathione were lower in BV.

Carbohydrates.

Of the four amino sugars detected, N-acetylneuraminate, commonly known as sialic acid, was higher in BV (P < 0.001), while glucosamine levels were lower (P = 0.0059). Glucose oligosaccharides of various lengths, including maltotriose (P = 0.0095), maltotetraose (P = 0.0016), maltopentaose (P < 0.001), and maltohexaose (P < 0.001) were all lower in BV. Likewise, lower levels of simple sugars and sugar alcohols, such as lactate (P < 0.001), fructose (P = 0.0012), and mannitol (P = 0.001), were noted; galactose (P < 0.001) and threitol (P < 0.001) were significantly higher in BV. Succinate levels were also higher in BV (P < 0.001).

NAD.

NAD, an essential cofactor for energy metabolism, was below the limit of detection in 82.5% of BV cases and in only 10% of controls (P < 0.001). Levels of nicotinamide, a precursor to NAD biosynthesis, were also lower in BV (P = 0.06), while nicotinate (P = 0.002) levels were higher.

Lipids.

Multiple components of lipid metabolism were affected in BV. The arachidonic acid catabolite 12-hydroxyeicosatetraenoic acid (12-HETE) was higher in BV (P < 0.001), while arachidonate was lower (P = 0.0075). Deoxycarnitine, a precursor to carnitine, was higher in BV (P < 0.001), while carnitine was lower (P < 0.001). Ascorbic acid, which is essential to the synthesis of carnitine, was lower in BV (P < 0.001). Acyl-carnitines such as acetylcarnitine (P < 0.001), propionylcarnitine (P < 0.001), and butyrylcarnitine (P < 0.001), were also lower in BV. Of the four monohydroxy fatty acids detected, levels of 4-hydroxybutyrate (P < 0.001) and 13-hydroxyoctadecadienoic acid (13-HODE) (P = 0.0042) were higher in BV. Levels of biochemicals involved in glycerol metabolism, including glycerol (P = 0.046) and glycerol-3-phosphate (P < 0.001), were lower in BV.

Association between metabolites and bacterial abundance.

Pearson correlation coefficients were used to investigate associations between bacterial abundance and metabolite abundance in vaginal samples. Specifically, we analyzed the abundance of 30% of the most variable metabolites and the relative abundance of bacteria detected by broad-range PCR with pyrosequencing in a subset of 20 women with BV and 10 women without BV. A group of 10 BV-associated bacteria were highly correlated with metabolites typically associated with BV, including succinate, cadaverine, putrescine, tyramine, and deoxycarnitine (Fig. 2). Megasphaera sp. type 1 had the highest correlation coefficient with the fatty acid 12-HETE (0.48). A second cluster of BV-associated bacteria, including BVAB1, BVAB3, Megasphaera sp. type 2, and several Prevotella species, such as P. amnii, P. disiens, P. buccalis, and P. bivia, showed similar positive associations with the above-mentioned metabolites, but the correlations were less pronounced. The three lactobacilli L. crispatus, L. jensenii, and L. iners clustered together and exhibited metabolite correlation patterns that overall were in contrast to the patterns found with BV-associated bacteria.

Open in a separate window

FIG 2 

Association between metabolites and bacterial abundance. Hierarchically clustered Pearson correlation coefficients are displayed in a heat map to demonstrate associations of vaginal bacteria (relative abundance) detected using broad-range PCR and pyrosequencing (x axis) with the abundance of 30% most variable metabolites (y axis). Correlation values ranged from −0.7 (dark green) to 0.84 (dark red).

Association between metabolites and bacterial concentrations.

Bacteria were selected for qPCR based on relative abundance measured by broad-range PCR and pyrosequencing in a subset of samples (Fig. 2) and on associations made in previous studies that have demonstrated the sensitivity, specificity (31), and significance of these bacteria in BV (11, 20, 32). Associations of concentrations of key vaginal bacteria with the 30% most variable metabolites were investigated using Pearson correlation coefficients (Fig. 3). L. crispatus, L. jensenii, and L. iners clustered together, while the BV-associated bacteria clustered as two groups. Group 1 comprised Leptotrichia/Sneathia spp., BVAB2, Megasphaera spp., Prevotella timonensis, Atopobium vaginae, Gardnerella vaginalis, Eggerthella-like bacterium, and Prevotella buccalis. Group 2 comprised BVAB1, BVAB3, and Prevotella amnii. The clustering patterns noted here also reflected patterns observed in our association analysis of metabolites with bacteria detected by broad-range PCR and pyrosequencing. L. crispatus and L. jensenii, lactobacilli typically associated with vaginal health, exhibited metabolite correlation patterns that were similar; these were in striking contrast to those exhibited by BV-associated bacteria (Fig. 3). L. iners exhibited correlation patterns that were intermediate between those of BV-associated bacteria and those of L. crispatus and L. jensenii. Concentrations of L. crispatus and L. jensenii had strong positive correlations with several amino acids and dipeptides but were negatively correlated with amino acid catabolites (tyramine, pipecolate, and cadaverine) and polyamines (putrescine and agmatine). L. iners was negatively correlated with some amino acids and dipeptides (glutamate and glycylleucine) but positively correlated with others (proline, threonine, aspartate, serine, and valinylglutamate). L. crispatus and L. jensenii were positively correlated with sugars, such as maltose, maltotriose, and maltohexose, lipid metabolism biochemicals, such as arachidonate and carnitine, as well as lactate, urea, and reduced glutathione. These two lactobacilli were negatively correlated with N-acetylneuraminate, succinate, the carnitine precursor deoxycarnitine, the eicosanoid 12-HETE, the fatty acid 13-HODE, and the nucleobase uracil. BV-associated bacteria typically exhibited correlation patterns with metabolites that were opposite to those exhibited by L. crispatus and L. jensenii. Among BV-associated bacteria, BVAB1 displayed specific differences in metabolite correlation patterns compared with BV-associated bacteria from group 1. BVAB1 was negatively correlated with N-acetylaspartate, 12-HETE, and fatty acids, such as eicosenoate and dihomo-linoleate; these metabolites had positive associations with group 1 BV-associated bacteria. BVAB3 and P. amnii, members of the group 2 BV-associated bacteria, exhibited metabolite correlation patterns that were similar to each other.

Open in a separate window

FIG 3 

Association of specific vaginal bacteria with metabolites. Hierarchically clustered Pearson correlation coefficients are displayed in a heat map to demonstrate associations of key vaginal bacterial concentrations (x axis) measured using taxon-directed qPCR with the 30% most variable metabolites (y axis). Correlation values range from −0.71 (dark green) to 0.85 (dark red). Two subgroups of BV-associated bacteria were observed, and clustering patterns were similar to those noted in Fig. 2 using an untargeted approach for bacterial community analysis. L. crispatus and L. jensenii exhibited correlation patterns that were similar and opposite to those by BV-associated bacteria.

Associations between metabolites and individual clinical characteristics.

We examined associations of metabolites with each Amsel criterion for BV using penalized linear regression models (Fig. 4 and Table 3). No metabolite was associated with all four individual criteria. Lactate was negatively associated with elevated pH and an amine odor. Cadaverine was associated with elevated pH and thin homogeneous discharge, while N-acetylputrescine was associated with elevated pH and an amine odor. Deoxycarnitine and pipecolate were associated with the presence of clue cells, while the reduced form of glutathione (GSH) and glycylproline were negatively associated with the presence of clue cells. Deoxycarnitine, N-acetylputrescine, pipecolate, and cadaverine were positively associated with BV overall, while phenylalanine was negatively associated with BV.

Open in a separate window

FIG 4 

Association of metabolites with Amsel clinical criteria. Shown is a model depicting metabolites associated with individual clinical criteria used in BV diagnosis (21). Stars denote metabolites that were positively or negatively associated with BV status. BV status is indicated on the x axis of box plots. The y axis of box plots represents scaled concentrations of metabolites. The lines in box plots represent the mean, and whiskers denote 95% confidence intervals.

TABLE 3 

Association of metabolites with individual clinical criteria

Clinical criterion	Regression coefficienta
BV status
Deoxycarnitine	0.5
Phenylalanine	−0.299
N-Acetylputrescine	0.254
Pipecolate	0.148
Cadaverine	0.025
Elevated pH
Cadaverine	0.0657
N-Acetylputrescine	0.0577
Lactate	−0.0505
Glutamate	−0.0249
Sphingosine	−0.0137
Tyrosine	−0.0083
Tyramine	0.0033
Presence of clue cells
Deoxycarnitine	0.2519
Glycylproline	−0.203
GSHb	−0.0311
Pipecolate	0.0024
Presence of amine odor
N-Acetylputrescine	0.302
Lactate	−0.072
Phenylalanine	−0.049
Vaginal discharge
Agmatine	0.14
Cadaverine	0.11

Open in a separate window

^aA negative value indicates a negative association with the clinical criterion, and a positive value indicates a positive association with the clinical criterion.

^bReduced glutathione.

Alterations in key vaginal bacteria and metabolite concentration with changes in BV status.

Limited longitudinal data are presented for 8 women who returned for follow-up visits 4 weeks after baseline sample collection per the study protocol (Fig. 5; see Fig. S1 in the supplemental material). Participants diagnosed with BV were treated with metronidazole. Study participants 19, 21, 23, and 25 were cured of BV at the follow-up visit (Fig. 5). All four participants had high concentrations of lactobacilli at their follow-up visits and increased concentrations of metabolites negatively associated with BV. Women with high concentrations of L. crispatus at follow-up had high concentrations of metabolites negatively associated with BV. Women with high concentrations of L. iners also showed increased concentrations of metabolites negatively associated with BV, but these shifts were not as dramatic as increases seen in women with L. crispatus dominant communities (Fig. 5). Study participants 4 and 7 did not have BV at either time point and had high concentrations of lactobacilli at both time points (see Fig. S1). High concentrations of metabolites negatively associated with BV were also noted at baseline and follow-up. Study participants 12 and 28 had recurrent BV after antibiotic treatment. In both participants, concentrations of metabolites associated with BV declined at the follow-up visit, but there was no associated increase in metabolites negatively associated with BV (see Fig. S1). In this longitudinal analysis, changes in the microbiota were associated with shifts in metabolites in the same subject, whereas stability of the microbiota was associated with fewer changes, even after patients received antibiotic therapy.

Open in a separate window

FIG 5 

Shifts in concentrations of key vaginal bacteria and metabolites with treatment for BV. Longitudinal data are presented for four study participants (19, 21, 23, 25) who were cured of BV posttreatment with metronidazole at the 1-month follow-up visit. Bacterial concentrations are displayed as 16S rRNA copies per swab on the y axis of the x-y plots. Scaled metabolite concentrations (y axis) of 14 metabolites associated with individual clinical criteria from Fig. 4 (Table 3) have been classified as positively associated with BV and are higher in women with BV (red) and negatively associated with BV and are lower in women with BV (green). y axis scales are different for each subject reflecting differences in metabolite concentrations. All four participants had high concentrations of lactobacilli at their follow-up visits and increased concentrations of metabolites negatively associated with BV (green). Women with high concentrations of L. crispatus at follow-up had high concentrations of metabolites negatively associated with BV (green). Women with high concentrations of L. iners also showed increased concentrations of metabolites negatively associated with BV (green), but these shifts were not as dramatic as increases seen in women with L. crispatus dominant communities.

(ii) DNA extraction and qPCR.

DNA was extracted using the Ultra Clean soil DNA kit or the Bacteremia kit (Mobio, Carlsbad, CA), which gave similar results. DNA was eluted in 150 µl buffer and diluted 1:1 in 1 mM Tris–0.1 mM EDTA (TE) buffer. Sham swabs without human contact were processed as controls to assess contamination from reaction buffers or sample collection swabs. Samples were evaluated for presence of PCR inhibitors using a qPCR assay targeting a segment of exogenously added jellyfish DNA; inhibition was defined as a delay in the threshold cycle of >2 cycles compared with no-template controls (68). Control assays targeting the human 18S rRNA gene were performed to ensure that swabs contained human tissue (68). DNA from each sample was subjected to a panel of 14 qPCR assays to measure concentrations of bacteria. The bacteria targeted included three members of the order Clostridiales which are highly specific for BV, designated BV-associated bacterium 1 (BVAB1), BVAB2, and BVAB3 (8). Other vaginal bacteria targeted included Gardnerella vaginalis, Leptotrichia/Sneathia spp., Megasphaera-like bacterium (type 1 and type 2), Atopobium vaginae, Lactobacillus crispatus, Lactobacillus jensenii, and Lactobacillus iners (69, 70). We developed four additional qPCR assays targeting an Eggerthella-like bacterium, Prevotella amnii, Prevotella timonensis, and Prevotella buccalis using 16S rRNA gene-specific primers and taxon-directed hydrolysis probes. The assay conditions and primer and probe sequences for qPCR assays developed in this study are presented in Table 4. Core PCR reagents were obtained from Applied Biosystems (Carlsbad, CA), and master mixes contained buffer A (1 mM), deoxynucleotide triphosphates (1 mM), magnesium (3 mM), AmpErase uracil-N-glycosylase (0.05 U), and AmpliTaq gold polymerase (1 to 1.5 U) per reaction. Primers were added at 0.8 µM per reaction, and the final probe concentration was 150 µM. Assays underwent 45 cycles of amplification on the StepOnePlus real-time PCR system (Life Technologies, Grand Island, NY). Plasmid standards were run in duplicate from 10⁶ to 2.5 copies. Specificity and sensitivity testing was conducted as previously described (69). Two microliters of DNA (diluted 1:1 with TE buffer) was added to each qPCR mixture, and values are reported as 16S rRNA gene copies per swab.

(iii) Broad-range PCR and pyrosequencing of 16S rRNA gene amplicons.

We performed broad-range 16S rRNA gene PCR with pyrosequencing using 454 Life Sciences FLX technology (Roche, Branford, CT) targeting the V3-V4 region of the 16S rRNA gene for a subset of samples from 20 women with BV and 10 women without BV, a subset of the cohort described previously (11). Sequence reads were classified using the phylogenetic placement tool pplacer (71) and a curated reference set of vaginal bacteria (11).

(iv) Measurement of levels of biochemicals.

Chromatographic separation and full-scan mass spectroscopy (MS) were performed using the Metabolon (Durham, NC) platform (72,–74). Briefly, CVL samples were extracted to remove the protein fraction while allowing maximal recovery of small molecules using the MicroLab Star system (Hamilton, Reno, NV). The resulting extract was split into equal parts for analysis by gas chromatography-mass spectrometry (GC-MS) and liquid chromatography-mass spectrometry (LC-MS). Samples were dried by placing them on a Zymark TurboVap (American Laboratory Trading, East Lyme, CT) and stored at −80°C. Aliquots of water and solvents used in the extraction were included as controls to assess the contribution of the separation process and extraction methods to the compound signals. Additional samples with known compositions were also included with every run to ensure that there were no process deviations in each run. A reference library of 1,000 purified standards with known chemical structures was used to identify the small molecules from our samples by matching chromatographic properties and mass spectral signatures.

(v) Data analysis.

Metabolite data were median centered, and missing values were imputed as the minimum observed for that metabolite; median values were then log transformed. Pyrosequencing taxonomic counts and qPCR measures were log transformed. Welch’s two sample t tests were used to compare metabolites between BV-positive and BV-negative groups. An estimate of false discovery rate (q value) was calculated to account for multiple comparisons. A low q value (<0.1) indicates high confidence in a result. Eighty-three metabolites (30%) with the largest variability (interquartile range) were used to hierarchically cluster samples using Euclidean distance (Fig. 1); identification of four sample groups maximizes silhouette width. To explore the relationship between relative abundances of metabolite and pyrosequencing reads (Fig. 2) or qPCR measures (Fig. 3), Pearson correlations were calculated between metabolite and pyrosequencing reads or qPCR levels; these were used in hierarchical clustering using Euclidean distance. Data transformation was performed in R (http://cran.r-project.org) (75).

Lasso regression with binomial response was used to identify metabolites strongly associated with Amsel clinical criteria (Table 3). The number of metabolites retained was set to the largest value within 1 standard deviation of the cross-validation error. Analysis was performed with the R package glmnet (76).

Cohort B. (i) Study population and sample collection.

Cohort B in the independent validation study also comprised 40 women with BV and 20 women without BV enrolled in the STD clinic between April 2011 and August 2013. The BV definitions and sample collection procedures were the same as those in cohort A. We also included an additional 10 women with intermediate Nugent scores of 4 to 6 to determine if their metabolite profiles differed from those who had BV (Nugent scores of 7 to 10) or those who had Nugent scores of 0 to 3. Of these 10 women, 6 were BV negative and 4 were BV positive by Amsel clinical criteria. All women provided written informed consent, and the study was approved by Institutional Review Board at the Fred Hutchinson Cancer Research Center.

(ii) DNA extraction and molecular methods.

DNA was extracted using the Bacteremia kit (Mobio, Carlsbad, CA), eluted in 150 µl buffer, and diluted 1:1 in 1 mM Tris–0.1 mM EDTA (TE) buffer. We targeted the V3-V4 region of the 16S rRNA gene for broad-range 16S rRNA gene PCR with pyrosequencing using 454 Life Sciences titanium technology (Roche, Branford, CT). Reads were processed using the same methods as cohort A.

(iii) Measurement of levels of biochemicals. (a) Sample preparation.

Commercially available pooled serum (Innovative Research, Novi, MI) was used as the quality control (QC) sample, and aqueous metabolites were extracted using methanol (250 µl). The supernatant (200 µl) obtained postcentrifugation (20,800 × g for 10 min) was dried for 90 min using a Vacufuge Plus evaporator (Eppendorf, Hauppauge, NY) and reconstituted in 600 µl 40% solvent A (40 mM ammonium acetate in H₂O plus 0.3% acetic acid) and 60% solvent B (acetonitrile plus 0.3% acetic acid) containing 5.13 µM l-tyrosine-¹³C₂ and 22.54 µM sodium-l-lactate-¹³C₃ (Cambridge Isotope Laboratory). The isotope-labeled internal standards helped monitor LC-MS assay performance. Samples were filtered through 0.45-µm-pore polyvinylidene difluoride (PVDF) filters (Phenomenex, Torrance, CA) prior to LC-MS analysis. Acetonitrile, ammonium acetate, methanol, and acetic acid (LC-MS grade) were obtained from Fisher Scientific (Pittsburgh, PA), and stable isotope-labeled tyrosine and lactate internal standards (>99% pure) were obtained from Cambridge Isotope Laboratories, Inc. (Tewksbury, MA). The QC sample was injected every 10 vaginal samples. Vaginal swab samples and CVL fluid (100 µl) were processed using the same procedure as the QC sample but extracted using 850 µl of methanol. Sham samples (saline solution and swabs without human contact) were included to assess background signals (chemical impurities present in materials used to collect vaginal fluid). Blank saline solution and swab samples were injected once each for every 10 vaginal fluid samples.

(b) Chromatography conditions.

The LC system was composed of two Agilent 1260 binary pumps, an Agilent 1260 auto-sampler and Agilent 1290 column compartment containing a column-switching valve (Agilent Technologies, Santa Clara, CA). Each sample was injected twice, at 10 µl for analysis using the negative-ionization mode and 2 µl for analysis using the positive-ionization mode. Both chromatographic separations were performed in the HILIC mode on two parallel SeQuant ZIC-cHILIC columns (150 by 2.1 mm, 3.0-µm particle size) (Merck KGaA, Darmstadt, Germany). While one column was performing the separation, the other column was reconditioning in preparation for the next injection. The flow rate was 0.300 ml/min, the autosampler temperature was kept at 4°C, the column compartment was set at 40°C, and the total separation time for each ionization mode was 18 min. The mobile phase was composed of solvents A and B. The gradient conditions for both separations were identical and are as follows: 0 to 2 min, 25% solvent A and 75% solvent B; 2 to 5 min, 25 to 70% solvent A, linear gradient; 5 to 9 min, 70% solvent A; 9 to 11 min, 70 to 25% solvent A, linear gradient; 11 to 18 min, 25% solvent A.

(c) MS and data processing.

After the chromatographic separation, MS ionization and data acquisition were performed using an AB Sciex QTrap 5500 mass spectrometer (AB Sciex, Toronto, Ontario, Canada) equipped with an electrospray ionization (ESI) source. The instrument was controlled by analyst 1.5 software (AB Sciex, Toronto, Ontario, Canada). Targeted data acquisition was performed in multiple-reaction-monitoring (MRM) mode. We specifically developed additional assays for 19 metabolites of interest based on our results from cohort A, wherein we measured biochemicals using the Metabolon platform (see Table S2 in the supplemental material). We monitored 104 and 76 MRM transitions in the negative and positive modes, respectively (180 transitions total) (see Table S2). The source and collision gas was N₂ (99.999% purity). The ion source conditions in the negative mode were curtain gas (CUR) =25 lb/in², collision gas (CAD) = high, ion spray voltage (IS) = −3.8 kV, temperature (TEM) =500°C, ion source gas 1 (GS1) =50 lb/in², and ion source gas 2 (GS2) =40 lb/in². The ion source conditions in the positive mode were CUR =25 lb/in², CAD = high, IS = 3.8 kV, TEM =500°C, GS1 =50 lb/in², and GS2 =40 lb/in². The extracted MRM peaks were integrated using MultiQuant 2.1 software (AB Sciex, Toronto, Ontario, Canada). Standard compounds corresponding to measured metabolites were purchased from Sigma-Aldrich (St. Louis, MO) and Fisher Scientific (95 to 99% pure).

(iv) Data analysis.

Metabolite data and pyrosequencing reads were processed and analyzed as described for cohort A.

The project was funded with grants R01 AI061628 and HG005816 to D. N. Fredricks from the National Institutes of Health.

We thank Dwyn Dithmer and Kathy Ringwood for clinical contributions and Kathy Agnew and Christina Kohler for performing Gram stains of vaginal fluid. Kathy Thomas and Linda Drolette provided support with clinical data management. We are grateful to Xuezhou Hou for assistance in the development of new qPCR assays.

S.S., J.M.M., D.R., and D.N.F. conceived and designed the experiments. T.L.F. performed the biological experiments, including DNA extractions and PCRs. D.D. performed the experiments evaluating the metabolites from samples in cohort B. S.S., M.T.M., and N.G.H. analyzed the data. S.S. and D.N.F. wrote the paper. All authors edited the manuscript and approved the final draft.