Dysbiosis of the gut microbiota is associated with HIV disease progression and tryptophan catabolism

Gut bacterial microbiota composition differs between HIV- and VU subjects

To capture a wide range of stages of HIV disease progression and treatment, a cohort of 32 male subjects was assembled to include six viremic, untreated (VU) HIV-infected subject, 18 virally-suppressed subjects on HAART with varying degrees of CD4⁺ T cell recovery (HAART), one HIV-infected long-term non-progressor, and nine uninfected risk-matched controls (HIV−); see Table S1 for defining characteristics of the cohort. Rectosigmoid biopsy specimens and peripheral blood were collected for immunophenotyping of immune cell subpopulations and for profiling of microbial communities. Biopsy tissue was chosen over stool sampling based on evidence that immunomodulatory bacterial species adhere closely to the gastrointestinal mucosal epithelium in the mouse (19, 20), comprising a community that is unique from that within the gut lumen (28). Microbial profiling was performed using a standardized, high-density microarray containing oligonucleotide probes targeted against hypervariable regions of the bacterial 16S rRNA gene (G3 PhyloChip (29)). This approach is ideally suited for high-resolution microbiome comparative analyses and detection of lower-abundance species that may not be identified using sequencing approaches of relatively shallow read depth (30). This is particularly relevant given recent evidence demonstrating the presence of low-abundance “keystone” species that despite their low numbers are pivotal in modulating commensal community composition and host-microbe interactions in the oral cavity (31).

Analysis of rectosigmoid biopsies showed that total bacterial load was similar across all subjects, regardless of infection status (Fig. S1A), and that total bacterial load measures were concordant across different biopsies from a given individual as well as across different DNA extraction methods (Fig. S1B). Additionally, comparative analysis of gross microbial community metrics showed no significant differences between HIV-infected and uninfected subjects in community evenness or richness (Fig. S1C). By contrast, analysis of between-subject microbiota composition using a Canberra community dissimilarity matrix revealed distinct clustering of the subject groups representing extremes of health and disease: the HIV- and the VU group, respectively (Fig. 1). Permutational multivariate analysis of variance confirmed that the observed variability in community composition was significantly related to subjects belonging to the VU or HIV- group (Adonis; P = 0.002), indicating that untreated HIV infection is associated with a distinct mucosal microbiota composition. Interestingly, HAART-treated patients, though uniformly exhibiting viral suppression (i.e., plasma viral load < 40 copies/ml), exhibited highly dissimilar community compositions: the microbiota of some patients clustered closely with VU subjects while others were more similar to HIV- subjects.

To identify differences in the adherent bacterial communities of HIV- and VU subjects, a permutational analysis algorithm with integrated false-discovery correction (SAM (32)) was used to detect specific bacterial taxa in differential fold abundance across these two subject groups. Of the 33,951 taxa (defined as a group of bacteria having at least 99% 16S rRNA sequence identity) that were detected in at least one of the samples studied, 625 exhibited significantly different relative abundance between VU and HIV- after stringent data filtering. This list comprised 579 taxa that were enriched and 45 taxa that were depleted in VU subjects compared with HIV- subject samples (Fig. 2 and Table S2 for list of taxa). Among the taxa enriched in VU subjects, the most enriched was Erysipelotrichaceae in the class Mollicutes, a class that has been associated with obesity and heightened cardiovascular morbidity (33–35). A majority of taxa enriched in VU subjects included members of the phylum Proteobacteria, with a notably high representation of the Enterobacteriaceae family, a finding verified by quantitative polymerase chain reaction (PCR) (Fig. S2) and supported by a prior report (36). Enriched genera from within this family included known pro-inflammatory pathobionts such as Salmonella, Escherichia, Serratia, Shigella, and Klebsiella species. Additionally, Staphylococcus, Pseudomonas, and Campylobacter spp., known opportunistic pathogens and sources of bacteremia in HIV-infected subjects (37–39), were highly enriched in the mucosae of VU subjects. Significant reductions in the relative abundance of specific members of Clostridia and Bacteroidia were observed in VU subjects, with the greatest degree of depletion amongst members of the Bacteroides and Alistipes genera, which have been previously found to be depleted in inflammatory bowel disease (40). We refer to this community of bacterial taxa (identified as differing significantly in abundance between VU subjects and HIV-uninfected subjects) as the “disease-associated microbial community” (DMC).

Intestinal dysbiosis correlates with markers of disease progression among HIV-infected subjects

While levels of T cell activation (as measured by co-expression of CD38 and HLA-DR) and viral load are well-accepted markers of HIV disease progression (3), it is now also appreciated that other immune parameters are linked to morbidity and mortality, including plasma markers of inflammation, IDO1 activity, and markers of microbial translocation (7, 11, 17). Of note, specific murine gut commensal species have been shown to adhere to mucosal epithelia and induce altered inflammatory cytokine secretion patterns (19, 20), some in association with IDO1 upregulation (20). To determine whether associations exist between the mucosa-associated DMC and HIV disease progression, peripheral blood was drawn at the time of mucosal sampling and additional biopsies were obtained for immunophenotyping (see Table S3 for a complete list of immune parameters measured). Correlations between the relative abundance of taxa within the DMC and immunologic parameters among all HIV-infected subjects (HAART and VU) were examined (Fig. 3A). Strong associations were not found between taxon abundances and peripheral blood CD4⁺ T cell count or gut HIV RNA or DNA levels. However, taxa that were enriched in the DMC correlated with elevated levels of T cell activation, of catabolism of tryptophan through the kynurenine pathway (a marker of IDO1 activity measured by the ratio of plasma concentrations of the downstream product, kynurenine, to the parent compound, tryptophan [Kyn:Trp]), and of soluble plasma markers of inflammation (such as IP-10, soluble TNF receptor II, and IL-6), and diminished levels of mucosal T cells secreting IL-17 and IL-22. Though relative abundances of many taxa were found to correlate with these immune parameters (Fig. 3A), phylogenetically distinct members of the Enterobacteriaceae family were among the strongest correlates of elevated gut CD4⁺ and CD8⁺ T cell activation (unadjusted P = 0.00004, R_S = 0.82; unadjusted P = 0.00006, R_S = 0.80; Table S3; R_S, Spearman rho correlation coefficient) while members of Staphylococcaceae were among the strongest correlates of higher Kyn:Trp and of elevated plasma IP-10 levels (unadjusted P = 0.00007, R_S = 0.76; unadjusted P = 0.0014, R_S = 0.64). Taxa within the families Micrococcaceae, Rhodobacteraceae, Halomonadaceae, and Pasteurellaceae also exhibited strong correlations with numerous markers of disease progression. Conversely, those taxa depleted in the DMC were found to correlate with lower levels of T cell activation and higher levels of mucosal T cells secreting IL-17, with the strongest correlates of lower T cell activation belonging to Bacteroidaceae (unadjusted P = 0.0017, R_S = −0.68) and the strongest correlates of higher gut IL-17 production belonging to taxa within the Rikenellaceae family (unadjusted P = 0.0004, R_S = 0.72; see Fig. S3 and Table S3 for full list of all correlation coefficients). Collectively, these findings indicate higher abundance of multiple, diverse microbial taxa within the DMC is linked to immune activation and to poor gut IL-17 responses, and that this state is concomitantly characterized by an apparent depletion of taxa linked to immune homeostasis and conditions which promote gastrointestinal integrity, i.e., lower levels of T cell activation and higher levels of IL-17 secretion.

Gut-resident microbial communities function as discrete ecosystems bound together by interdependent metabolic exchange and collective protective mechanisms. Accordingly, many taxa of the DMC exhibited similar patterns of correlation with immune parameters (Fig. 3A). Therefore, a holistic approach was taken to understand such trends in gross community abundance as they may relate to inflammatory states. Principal component analysis (PCA) was performed to consolidate patterns among co-varying microbial abundances into a handful of summary variables and to compare such variables to immune parameters. The first principal component (PC1) of this analysis comprised 52.8% of DMC variance (Fig. 3B and Fig. S4A), was able to differentiate HIV- from VU (Fig. 3C), and was chiefly characterized by opposing trends in the relative abundance of Proteobacteria (including Pasteurellaceae, Mycobacteriaceae, Pseudomonadaceae, and Enterobacteriaceae members) and Bacteroidaceae (Fig. S4B). We used PC1 as a measure of the degree to which a subject’s gut microbiota was similar to that of VU or HIV− (and therefore a surrogate marker for extent of dysbiosis), and related it to markers of immunopathology and disease progression among all HIV-infected subjects. PC1 was found to associate strongly with plasma IP-10 (adjusted P = 0.044), a marker of systemic inflammation shown to correlate with accelerated HIV disease progression (18, 41), with Kyn:Trp (adjusted P = 0.007), and with T cell activation in the gut and peripheral blood (respective significance values shown in figure) (Fig. 3D, Table S4).

Heterogeneity in microbiota of subjects undergoing HAART correlates with markers of disease progression

Even though HAART is effective in diminishing plasma viremia to nearly undetectable levels, subjects receiving long-term treatment can exhibit persistently elevated levels of chronic immune activation (16), of which the etiology is unknown. We hypothesized that persistence of even a subset of the DMC could continue to sustain pathologic chronic immune activation in this population despite suppression of viral replication. To address this hypothesis, the relative abundance of DMC members was averaged across each of the VU, HAART, and HIV- subject groups and compared between these groups. The DMC in HAART subjects was found to represent an intermediate state between that observed in VU subjects and uninfected controls (Fig. 4A). Additionally, longitudinal sampling of two HIV-infected subjects prior to and following nine months of effective HAART demonstrated a relative shift in the DMC to more closely resemble that of the HIV- group (Fig. S5). However, individual subjects within the HAART group exhibited a spectrum of relative abundance for members of the DMC, such that the microbial community composition of some subjects was more similar to that of HIV- subjects while the composition of others was more similar to that of VU subjects (Fig. 4B).

We considered a number of possibilities to account for the presence of this spectrum in the DMC of subjects on HAART. Since prophylactic antibiotics are commonly used by subjects on HAART and may influence microbial community composition, relationships between PC1 and antibiotic administration history (for which data was available for 13 out of 18 subjects) were analyzed. No significant trends were observed, as was the case for comparisons to age and time on HAART (Fig. S6). Furthermore, gut inflammation itself has been found to stimulate the outgrowth of facultative anaerobes, including members of the Enterobacteriaceae family (24, 42, 43), by providing growth substrates in the form of anaerobic respiration terminal electron acceptors. Expression of genes involved in oxygen and nitrogen radical production (NADPH oxidase [CYBB subunit], arginase 1, nitric oxide synthase 2 [NOS2], type I interferons [IFN-α, IFN-β], type II interferon [IFN-γ], tumor necrosis factor-alpha [TNF-α]) was accordingly quantified in whole biopsies and compared to PC1, and no significant associations were found (Table S6). Rather, associations were found between PC1 and numerous markers of disease progression, including levels of IL-6 in plasma, Kyn:Trp, soluble TNF-α receptor II levels in plasma, blood coagulation marker D-Dimer levels in plasma, gut IDO1 transcript, blood memory CD8 T cell activation, plasma IP-10 levels, and gut CD4⁺ T cell activation (P < 0.05, |R_s| > 0.5; Table S5). After adjusting for multiple comparisons, significant relationships remained between PC1 and two disease markers: plasma Kyn:Trp and plasma IL-6 concentration (Fig. 4C). To understand whether these associations were influenced by potentially confounding variables and to incorporate a greater proportion of microbial community abundance trends, multiple linear regression analysis was performed. This approach allows for simultaneous consideration of multiple adjustment variables (antibiotics usage history, days on HAART, age, race) and the five most representative principal components of DMC variance (PC1, PC2, PC3, PC4, and PC5) toward understanding which variables are most predictive of the immunologic parameters of interest (Kyn:Trp and plasma IL-6). Out of all possible models involving these variables, the model with the best leave-one-out cross-validated (44) (LOO-CV) predictive accuracy was chosen and p-values were obtained by permutation. PC1 and ethnicity were significantly correlated with Kyn:Trp in the best scoring model (P = 0.0006 and P = 0.0036, respectively). In the case of plasma IL-6 levels, the variable lending greatest predictive power to the most predictive multivariate model was PC1 (P = 0.018), while four other variables contributed less significantly: PC4 (P = 0.032), ethnicity (P = 0.048), age (P = 0.027), and duration of HAART (P = 0.020). Comparisons of Kyn:Trp and plasma IL-6 to PC1 alone are shown in Fig. 4C.

Members of the DMC participate in immunoactive tryptophan catabolism

IDO1 is induced during inflammation and is thought to actively regulate immune activation by its numerous effects on T cells, including inhibition of IL-17 secretion and induction of Foxp3-expressing regulatory T cells (T_reg) (45). These effects are mediated by its production of kynurenine pathway compounds from its enzymatic substrate tryptophan, a reaction for which IDO1 is thought to govern the rate-limiting step (46). Consistent with previous findings (47), we found that IDO1 is strongly associated with type II interferon expression in colonic mucosal tissue and the ratio of plasma concentrations of kynurenine to tryptophan (Kyn:Trp) was indeed associated with reduced IL-17 secretion in CD4⁺ T cells and higher abundance of T_reg in the colon (Fig. S7A–B). However, though there was a significant positive correlation between IDO1 expression and plasma Kyn:Trp in VU subjects, there was no such association in the HAART HIV-infected subject group (Fig. S7C).

Given that a significant association was found between our marker of dysbiosis (PC1) and Kyn:Trp within the HAART subject group, we postulated that the DMC may directly contribute to immunomodulatory tryptophan catabolism through the kynurenine pathway. Comparing taxon abundances to plasma Kyn:Trp, we found that taxa within 140 genera positively correlated significantly with Kyn:Trp (Table S7). While bacteria are among the primary producers of the essential amino acid tryptophan, relatively few are known to encode the enzymatic machinery for tryptophan catabolism (48–51). Indeed, analysis of genomic and proteomic annotations from the UniProt Consortium (52) revealed several bacterial genera within Proteobacteria that encode enzymes homologous to those necessary for catabolism of tryptophan to the most potent immunomodulatory product of the kynurenine pathway (9, 13), 3-hydroxyanthranilic acid (3-HAA; Fig. 5A, Table S7). Comparing this list to that of genera enriched in the DMC, we found that genera that bore genetic homologs to tryptophan catabolism enzymes of the kynurenine pathway were preferentially enriched in the DMC (Fig. 5B), including Pseudomonas, Xanthomonas, Burkholderia, Stenotrophomonas, Shewanella, and Bacillus, as well as members of the families Rhodobacteraceae, Micrococcaceae, and Halomonadaceae (see Table S7 for full list). Furthermore, genera for which at least one member bore genetic homologs to 3–4 enzymes within the immunoactive kynurenine pathway correlated more strongly with Kyn:Trp ratios than genera that bore 1–2 tryptophan catabolism enzyme homologs, and even more so than genera that had no genetic evidence for kynurenine pathway metabolic machinery (Fig. 5C, Fig. S8).

To confirm the capacity for production of kynurenine pathway catabolites among members of the DMC, a representative species of the strongest bacterial correlate to plasma Kyn:Trp ratios (genus Pseudomonas) was assayed for kynurenine production in pure culture. Pseudomonas fluorescens strain A506 was found to produce significant amounts of kynurenine from tryptophan (Fig. 5D) in vitro, further supporting the hypothesis that gut mucosal-adherent bacteria associated with HIV infection might directly contribute to immunoactive tryptophan catabolism during HIV disease.

Study design

The primary aims of the present study were to understand whether differences in mucosal-adherent bacterial communities exist between HIV-infected untreated, HIV-infected on HAART, or HIV-uninfected subjects, and whether such differences are associated with immunological perturbations characteristic of HIV disease. To accomplish these aims, microarray-based microbial community profiling was employed and assessment of immunologic status of subjects was performed using clinical data, flow cytometry, ELISA, liquid chromatography-tandem mass spectrometry, and quantitative polymerase chain reaction. The investigators were not blinded to serostatus nor to viral load measurements, medication history, or peripheral blood CD4⁺ T cell counts. The control subjects were seronegative risk-matched men who had a history of having sex with men (MSM) and/or with histories of exposure to HIV from interactions with seropositive partners. The initial cohort of HIV-infected subjects on HAART comprised HIV-infected patients who had maintained undetectable viral loads on stable antiretroviral regimens for at least one year, with equal numbers of individuals with peripheral blood CD4⁺ T cell counts > 500 cells/mm³ and with CD4⁺ T cell counts < 350 cells/mm³, though CD4⁺ T cell counts at times of sample collection spanned the range between these benchmark CD4⁺ T cell counts. All subjects were included that met the criterion of having had no recorded antibiotics usage within two months of mucosal sampling, regardless of peripheral blood CD4⁺ T cell count, and the final list of subjects included nearly equivalent numbers of subjects with < 350 CD4⁺ T cells/mm³ and > 500 CD4⁺ T cells/mm³. Triplicate 16S amplicon hybridizations using separate G3 PhyloChip microarrays (Second Genome, Inc.) were performed using mucosal biopsy samples from two different study subjects. All triplicates were found to cluster closest together in a global hierarchical clustering analysis based on a Canberra community dissimilarity matrix.

Subject sample collection

All participants gave written informed consent for research sigmoidoscopy and biopsies, a protocol approved by the Committee on Human Research, UCSF. Study participants underwent a blood draw and received a Fleet enema prior to the procedure, and rectosigmoid biopsies (each ~3 mm in diameter) were obtained between 10 and 20 cm from the anus using jumbo forceps. Four biopsies were formalin-fixed and paraffin-embedded for immunohistochemistry; 12 biopsies were placed immediately in 15 ml of RPMI 1640 with 15% fetal calf serum, and processed for flow cytometric analysis within 4 hours of collection; two biopsies were placed in RNAlater (Life Technologies) and frozen at −80°C for quantitation of inflammatory gene expression; finally, another two allocations of two biopsies each were snap-frozen in liquid N₂ and stored at −80°C for HIV viral nucleic acid quantitation and bacterial community profiling by G3 PhyloChip microarray.

Nucleic acid extraction and quantitative polymerase chain reaction (qPCR)

For quantitation of inflammatory and interferon-stimulated gene expression, RNA was extracted from snap-frozen biopsies frozen in RNAlater (Life Technologies) after homogenization in Trizol (Life Technologies) and purification using the RNeasy mini-kit (Qiagen). RNA was also prepared from whole blood, via collection using the PaxGene Blood RNA Kit (Qiagen). cDNA was prepared from RNA using Omniscript RT Kit (Qiagen) using random hexamer primers. qPCR was carried out using TaqMan Gene Expression MasterMix (Applied Biosystems) in a StepOne Plus Real Time PCR system (Applied Biosystems). IDO1 transcript was measured using custom primers and probe (F: 5’- GCCAAATCCACAGGAAAATCT, R: 5’- GCTGTGACTTGTGGTCTGTGA, P: 5’- AAACATCTGCCTGATCTCATAGAGT). Interferon stimulated genes (ISG) were measured using commercially available primer-probe sets from Applied Biosystems: GBP1 Hs00977005_m1, IFI27 HSs00271467_m1, MX1 Hs00895601_m1, OAS1 Hs00973637_m1. ISG expression was consolidated using the geometric mean of relative expression values for all four aforementioned genes. Quantitation of expression of inflammatory genes (NADPH oxidase [CYBB subunit], arginase 1, nitric oxide synthase 2, type I interferons [IFN-α, IFN-β], type II interferon [IFN-γ], interferon-stimulated gene expression, tumor necrosis factor-alpha) was performed using the Fluidigm platform, with gene-specific pre-amplification using proprietary computationally designed primers (DELTAgene from Fluidigm). Relative expression was calculated using the ΔΔCT method, using HPRT as housekeeping gene.

Total bacterial burden quantification was carried out using DNA extracted from separate snap-frozen rectosigmoid biopsies that were collected at the same time as those above. Biopsies were resuspended in RNAlater ICE (Life Technologies), drained, and transferred to Lysis Matrix B (MP Biomedicals) tubes containing 600 ul RLT buffer and subjected to 30 seconds of bead beating at 5.5 m/s. Samples were briefly spun down and DNA was extracted using the AllPrep Kit (Qiagen). TaqMan qPCR was performed, as above, using the following primers: P891F, 5’-TGGAGCATGTGGTTTAATTCGA; P1033R, 5’-TGCGGGACTTAACCCAACA, and UniProbe 5’-CACGAGCTGACGACARCCATGCA. A standard curve of known 16S rRNA copy numbers was included with each run and was used to calculate total bacterial burden normalized to total 16S rRNA genes. Enterobacteriaceae were quantitated using Brilliant II SYBR Green (Agilent Technologies) and the following primers: F: 5’- ATGGCTGTCGTCAGCTCGT; R: 5’- CCTACTTCTTTTGCAACCCACTC (72). Bacteroides fragilis was also quantified using Brilliant II SYBR Green, and the following primers: F:5’- GAAAGCATTAAGTATTCCACCTG; R: 5’- CGGTGATTGCTCACTGACA (73).

T cell phenotyping analysis by flow cytometry

Mononuclear cells were prepared from peripheral blood by density gradient centrifugation and isolation using Ficoll-Paque (STEMCELL Technologies). Rectosigmoid biopsy mononuclear cells were isolated from biopsy specimens using a protocol optimized for lymphocyte viability and yield (74). Biopsy pieces were digested in three rounds of 0.5 mg/mL collagenase type II (Sigma-Aldrich), after which tissue was disrupted with a syringe bearing a 16-gauge blunt-end needle and passaged through a 70mm cell strainer.

Cells were then stained directly or after phorbol 12-myristate 13-acetate (PMA)/ionomycin stimulation in the presence of brefeldin and monensin for surface marker expression, and then permeabilized and stained intracellularly for FoxP3 or cytokine production using the FoxP3 Fixation/Permeabilization Staining Set (eBiosciences) or Cytofix/Cytoperm (BD Biosciences), respectively. Acquisition was performed on a Becton Dickinson LSRII (BD Biosciences) and analyzed using FlowJo software, version 9 (Treestar). Naive T cells were identified as CD27+ cells lacking CD45RO expression, and were excluded to yield total memory T cells. Proportions of activated CD4⁺ and CD8⁺ T cell were measured by the fraction of these cell populations staining positive for HLA-DR and CD38. Proportions of regulatory T cells (T_reg) from total CD4⁺ cells were determined by co-staining of CD25 and FoxP3.

Immunohistochemistry and viral load quantification in rectosigmoid tissues

Four rectosigmoid biopsies were collected into 4% paraformaldehyde, embedded in paraffin, cut into 5 µm sections and mounted on glass slides. Slides were deparaffinized in Clear-Rite 3 and rehydrated in an ethanol series, and antigenic epitopes were then retrieved by placing slides in a pressure cooker for 20 minutes at 121°C in pH6 citrate buffer. Slides were blocked and stained with polyclonal rabbit anti-CD3 (DAKO), and stain developed with DAB chromogen (anti-rabbit Polink-2 Plus HRP Detection Kit, GBI Labs). Whole slide composite images were acquired on a ScanScope (Aperio) and analyzed with the accompanying software to quantify CD3 stained cells per area of tissue. To determine the number of CD4⁺ cells per area, the CD3+ cells/mm² value was multiplied by the fraction of CD3+ cells that stained with CD4 as determined by flow cytometry. Two rectosigmoid biopsies were snap-frozen in liquid N₂ for quantification of HIV DNA and RNA as previously described (75).

Plasma inflammatory marker measurements

Plasma was collected from peripheral blood by centrifugation and stored at −80°C. The following soluble plasma proteins were measured by ELISA using commercially available kits and according to the manufacturer’s instructions: IP-10, sTNF-RII, IL-6, FABP2, sCD14 (all from R&D Systems), zonulin (Immundiagnostik), D-dimer (Diagnostica Stago). Concentrations of kynurenine, tryptophan, and 3-hydroxyanthranilic acid were quantitated in plasma and in bacterial growth cultures by liquid chromatography-tandem mass spectrometry, as previously described (9).

16S rRNA amplicon preparation, Phylochip hybridization, scanning, and data normalization

Nucleic acid extraction was performed on snap-frozen rectosigmoid biopsies by re-suspending in RNAlater ICE (Life Technologies), according to the manufacturer’s protocol and as previously described (76). G3 PhyloChip microarrays were obtained from Second Genome, Inc. PCR amplification of the bacterial 16S rRNA gene, Phylochip hybridization, Phylochip scanning, and fluorescence intensity normalization and initial probeset filtering of Phylochip data were performed as previously described (29), with the exception of performing 25 cycles of 16S PCR amplification rather than 30 and the usage of customized quartile ‘r score’ cutoff values. For each 16S rRNA gene probeset, an r score was independently calculated that incorporated mismatch probe fluorescence intensity as well as distribution of total chip fluorescence to account for per-chip background fluorescence. Quartile r score cutoffs were chosen based on r scores retrieved for control probes spiked-in at known concentrations, as recorded from all Phylochip microarrays of the current study, and were as follows: rQ₁ > .30, rQ₂ > .65, rQ₃ > .88.

Phylochip analysis

Following probe intensity normalization and to reduce the dataset for analyses, probesets belonging to phylogenetically related (based on species-level identity) 16S rRNA sequences and bearing similar probeset response patterns across all samples were represented by a single taxon, while others were eliminated from the dataset. This process was performed by calculating the variance in fluorescence intensity for each probeset and its two most phylogenetically related neighbors, and was repeated for each taxon across multiple different samples; the sum of all variances was used to define groups of taxa whose probe sets responded similarly. The resulting measures were ranked to select for removal the top ~15% of probesets exhibiting similar response trends across all samples, allowing reduction of redundant probesets. Next, the ‘vegan’ package within the R programming environment was used to perform non-metric multidimensional scaling (NMDS) based on community dissimilarity using the Canberra dissimilarity metric. The ‘Adonis’ package was used to estimate strength of relationship between community dissimilarity and classification of subjects into either VU or HIV-uninfected. To identify bacterial taxa in differential abundance among viremic untreated and HIV- subjects, the software package Statistical Analysis of Microarrays (SAM (32)), which incorporates the Storey (77) method of false discovery rate estimation in the form of adjusted q-values, was used. Taxa were filtered down to a final list of 547 taxa by the following criteria: q-value < 0.05, raw MFI fold change > 2, and Mann-Whitney U test P value < 0.03. Genera to which belonged taxa within the final list were visualized using the Interactive Tree of Life (78, 79) (iTOL, http://itol.embl.de/).

Comparisons of disease-associated microbial genera to immunologic parameters

The R software package ‘multtest’ was used to calculate Spearman rho correlation coefficient (R_S) values for comparisons of all 562 bacterial taxa that comprise the disease-associated microbial community (DMC) to each of the 24 immunologic parameters measured. Measures of correlations between immunologic variables and taxa were aggregated and condensed to genus-level classifications for visualization purposes in Fig. 3A, using the following methodology: for bacterial genera that were represented by more than one taxon in the DMC, R_S values representing correlations between taxa belonging to the same genus and an immunologic parameter were averaged to yield a R_Save that represents the correlation between relative abundance of that genus and each immune parameter. This was performed on the remaining immunologic parameters and all genera represented by more than one taxon, and was visualized using Cluster3 and Java TreeView (Version 1.1.6r2). To represent unadjusted significance values for each correlation shown in Fig. 3A, boxes were made to surround each correlation for which unadjusted P < 0.02. For genera that were represented by more than one taxon in the DMC, the median unadjusted P value for comparisons among immunologic parameters and taxa within that genus was used. Choice of mean for computing aggregated R_S and median for aggregated unadjusted P values for each genus was informed by assessment of global R_S and P value distributions (normal and non-normal, respectively, Fig. S9).

Principal component analysis

To quantify trends in co-variance among taxa that have similar patterns in relative abundance across all subject samples, a principal component analysis (PCA) was performed within the R programming environment (package ‘prcomp’). This allowed the condensation of trends in co-varying microbial abundances into discrete variables, thereby yielding single measures that represented broad trends in abundance of multiple taxa. To reduce the number of relationships tested, Spearman rank correlations with immunologic parameters were calculated only for the principal component that represented the greatest portion of co-variance (PC1), and the resultant P values were adjusted for multiple comparisons by the Benjamini-Hochberg technique (80).

Multiple linear regression model selection analysis

Among treated subjects, associations were confirmed using linear regression, allowing for the inclusion of a variety of adjustment variables comprising age, race, antibiotic use, and duration of HAART, as well as the first five principal components representing the greatest proportion of microbial abundance co-variance (PC1, PC2, PC3, PC4, PC5). All possible models were fit and compared according to their predictive accuracy. Predictive accuracy was defined as the mean squared error (MSE) calculated from leave-one-out cross-validation (LOO-CV). Under this approach, the model is fit repeatedly, leaving out one observation each time in order to assess a given model’s ability to predict new observations. For each outcome, the model with the lowest LOO-CV MSE was chosen. Given the small number of observations available, p-values were then obtained by permutation of residuals (81).

Analysis of kynurenine pathway enzyme annotations and associations with plasma Kyn:Trp ratios

The UniProt Consortium database (52) (www.uniprot.org) was queried for all annotated bacterial homologs of enzymes within the pathway to catabolize tryptophan to 3-hydroxyanthranilic acid (Enzyme Commission numbers: EC 1.13.11.11, EC 3.5.1.9, EC 3.7.1.3, EC 1.14.13.9). Phylum classifications for each annotation were combined, counted, and represented as pie charts with colors representing phyla as denoted in legend of Fig. 5A. ‘Totals’ represent numbers of unique genera within the domain Bacteria that bear annotated homologs in the highly curated and comprehensive UniProt database for each respective enzyme at the time of publication.

Spearman correlation P values were calculated for each taxon in relation to subject plasma Kyn:Trp ratios. For Fig. 5C, each genus was ranked based on P values of the taxon with the strongest correlation (as defined by the taxon with the lowest P value within a given genus). Annotations for tryptophan catabolism enzymes for all sequenced species and strains within each genus were considered when counting numbers of tryptophan catabolism enzymes for each genus.

Pseudomonas fluorescens kynurenine production in bacterial cultures

Pseudomonas fluorescens strain A506 (ATCC 31948) was grown in King’s B Medium (82) within U-bottomed 96 well polystyrene culture plates at an initial inoculum of 2.85 × 10⁷ cfu per well. King’s B Medium was prepared using Bacto Proteose Peptone No. 3 which, according to manufacturer-provided information, would contain 0.3 mM in the King’s B Medium preparation. Thus, bacterial cultures contained the free tryptophan already present plus 0, 0.5, 1.0, and 5.0 mM tryptophan added exogenously. Cultures were grown at room temperature and, at 20 hours post inoculation, were spun at 12,000 rpm for 10 minutes in a tabletop microcentrifuge. Supernatants were filtered through a 0.22 µm filter and kanamycin was added to a final concentration of 50 µg/ml. Liquid chromatography-tandem mass spectrometry was used to assay concentrations of kynurenine and tryptophan in growth culture supernatants as described previously (9). Three temporally independent biological replicates were assessed for each condition.