The central role of the glutamate metabolism in long-term antiretroviral treated HIV-infected individuals with metabolic syndrome

Clinical characteristics

PLWH was stratified according to the presence of MetS syndrome (100 individuals with and 100 individuals without MetS). Clinical characteristics of PLWH are summarized in Table 1. All individuals were currently on cART with similar duration of HIV-infection (p-value = 0.329) and cART exposure (p-value = 0.455). Larger prevalence of CD4 nadir < 200 cells was found among PLWH with MetS [53 (53%) vs. 38 (38%), p-value = 0.039].

Impaired amino acid metabolism in PLWH with MetS

The untargeted metabolomics identified 877 characterized biochemicals linked mainly with lipid metabolism (n = 285, 32%), xenobiotics including food components (n = 223, 25%) and amino acid metabolism (n = 220, 25%) (Figure 1A). The relative standard deviation (RSD) for the internal standards for the process variability was 5%. We observed that a total of 76 biochemicals differed significantly between PLWH with and without MetS (Mann-Whitney U, FDR <0.05) of which the majority were amino acids (n = 33, 43%) (Figure 1A). We next used metabolite set enrichment analysis (MSEA) and network analysis to identify MetS related mechanisms in PLWH and key biochemicals. The MSEA (FDR < 0.1 KEGG and HMDB) identified several affected amino acid metabolic pathways including methionine degradation, valine degradation, tyrosine degradation along the sirtuin signaling pathway that are linked to age-related diseases (Figure 1B). The network analysis identified two key molecules that interact with other metabolites, glutamate, and α-ketoglutarate, that were linked with most of the altered pathways (Figure 1C).

Figure 1. Differing metabolites and pathways found between PLWH and PLWH with MetS. (A) Doughnut charts of metabolite proportions for each super pathway for all detected metabolites (left) and metabolites with differential abundance between PLWH and PLWH with MetS (LIMMA, FDR < 0.1, n = 69). (B) Metabolites contribution to the flow of top 13 pathways represented as Sankey Plot. (C) Cytoscape network of top 13 pathways and associated enriched metabolites.

A combination of methodologies emphasizes the role of glutamate metabolism in MetS among PLWH

By combining supervised and unsupervised statistical and machine learning approaches, we sought to identify the most conservative and consensus set of metabolites associated with MetS in PLWH. Then, random forest (RF) was run on data with two classes: PLWH with and without MetS. The top-ranking metabolites (n = 21, using the feature selection by Boruta) (Figure 2A) showed several metabolites such as 4-hydroxyglutamate, γ-glutamylglutarate, glutamate, α-ketoglutarate, γ-glutamylglycine and N-acetylglutamate that highlights the central role of glutamate metabolism in MetS. This model achieved 80% predictive accuracy, sensitivity and specificity of 0.8 (Supplementary Figure 1), and an AUROC of 0.88 (Figure 2B). PLS-DA was also used to classify the whole data set (n = 877) that was submitted after scaling. A model using the two first components showed a significant coefficient of determination (R2 = 0.58) and coefficient of prediction (Q2 = 0.2). Using the predictive Variable Importance in Projection (VIP) vector, the top 30 crucial metabolites for this model were extracted (Supplementary Figure 2). LIMMA identified 69 metabolites that were significantly different between the groups (FDR < 0.05). A consensus identification among the four methods identified 11 biomarkers that were robustly identified by all approaches (Figure 2C and Supplementary Figure 3). A UMAP projection of samples based on these 11 biomarkers showed a clear separation of the uninfected controls and PLWH with MetS (Figure 2D). Seven of our potential biomarkers were upregulated in PLWH with MetS: 1-carboxyethylleucine, 4-cholesten-3-one, 4-hydroxyglutamate, α-ketoglutamate, γ-glutamylglutamate, glutamate, and isoleucine, while four were downregulated: carotene diol(2), glycerate, PSP and PC/3-MAPC (Figure 2E). γ-glutamylglycine and orotate were found only in RF while N2,N5-diacetylornithine, γ-glutamyl-alpha-lysine and γ-glutamyltyrosine were identified only by PLS-DA. Though our study population was highly matched, majority of the PLWH were male and MetS categorizations varied between male and female. We also perform sensitivity analysis restricted only to the males. Among the metabolites, 13 were identified as consensus among all the analysis (Supplementary Figure 4A). The RF model showed slightly lower performance than the full model (AUROC = 0·85, Supplementary Figure 4B). Interestingly, among the 13 potential biomarkers identified by the same process in males, eight were shared with biomarkers identified in the entire sample population and are mainly linked with glutamate metabolism (Supplementary Figure 4C). The five other biomarkers of MetS in males were γ-glutamylvaline, N-acetylglutamate, N-acetylvaline, γ-glutamylisoleucine and valine (Supplementary Figure 4D) with an increase in PLWH with MetS indicating the altered glutamate metabolism as a common feature in PLWH with MetS. Despite metabolic changes associated with MetS, confirmation that our potential biomarkers were specific for PLWH with MetS and not biomarkers of obesity or treatment was necessary.

Figure 2. Biomarkers (n = 11) with differential abundance between PLWH and PLWH with MetS identified by four methodologies. (A) Bubble plot representing Random forest variable importance based on mean decrease accuracy (a measure of the model’s performance without each metabolite). Values are scaled by the standard error of the measure. Metabolites represented at the top of the figure are the most important for prediction. (B) Receiver Operating Characteristic (ROC) curve of random forest classifier. (C) Venn diagram summarizing biomarkers identified by Mann-Whitney U test, LIMMA, Random Forest (RF), and PLS-DA. (D) UMAP visualization of the 11 biomarkers. Controls (green) and PLWH (yellow) are segregating from PLWH with MetS (red). (E) Heatmap showing log2 intensities of the 11 biomarkers in HC, PLWH without MetS and PLWH with MetS.

We thus tested whether there was an association between the biomarkers and MetS status corrected for BMI. All biomarkers were significant (FDR < 0.05) after BMI and treatment correction (Supplementary Tables 1 and 2). Pharmaceuticals can also affect metabolic profiles. Five xenobiotics were found to be significantly different between PLWH with and without MetS: hydrocinnamate, abacavir, losartan, tartronate, and thioproline. Correction for xenobiotics level was performed for each metabolite using linear regression (Supplementary Table 3). All metabolites passed the correction except glycerate. After correlation analysis, glycerate was shown to have a significant correlation with tartronate (R = 0.8, FDR < 0·00001) (Supplementary Figure 5).

Impaired central carbon metabolism with higher efflux of key carbohydrates

Higher levels of plasma glutamate and altered glutamate metabolism may indicate dysregulation of central carbon metabolism (CCM). Therefore, we investigated key metabolites from other pathways of central carbon metabolism: glycolysis, pyruvate metabolism, and the TCA cycle (Figure 3A). No changes were observed in plasma glucose levels, but we found a significant increase in pyruvate, lactate, and α-ketoglutarate levels in PLWH with MetS compared to PLWH without MetS (Figure 3A). This indicates an increased efflux of glycolytic and TCA cycle metabolites from cells into the bloodstream in PLWH with MetS. Furthermore, to characterize the uptake and secretion profile of key metabolic transporters within CCM, we measured the expression of Glut1 (glucose transporter), MCT-1 (pyruvate and lactate transporter), and xCT (transporter exchanging glutamate for cystine) in a subset of samples using flow cytometry (Figure 3B). We observed a significant decrease in classical monocytes in PLWH with MetS compared to PLWH without MetS while no difference was observed in other PBMC subpopulations (Supplementary Figure 6A). Most differences seen between the groups were on xCT and MCT-1 expression (Figure 3C). Percentage of cells expressing the different metabolite transporters were altered in PLWH with MetS compared to PLWH without MetS (Figure 3D). For Glut1, the median fluorescence intensity (MFI) was significantly higher in CD8⁺ T-cells and non-classical monocytes in PLWH with MetS compared to without MetS (Supplementary Figure 6B). PLWH with MetS had significantly increased expression of MCT-1 on all sub-populations of PBMCs compared to without MetS (Figure 3E). For xCT, expression levels were significantly increased in all monocytic cells (classical, intermediate, and non-classical) and decreased in CD4⁺ T-cells in PLWH with MetS compared to without MetS. These results indicate that PLWH with MetS had higher metabolic uptake and secretion in monocytic cell subsets compared to PLWH without MetS, specifically for glutamate transport across the cell membrane. The correlation between the transporter expression (represented as MFI) and significantly different metabolites in PLWH with MetS indicates negative co-relation of monocytes subpopulations of Glut1 with pyruvate level and CD8⁺ T-cells and classical monocytes with lactate level (Figure 3F). These further support altered metabolic state of the cells in the PLWH with MetS.

Figure 3. Central carbon metabolism with higher efflux of key metabolites. (A) Heatmap showing the level of metabolites of glutamate metabolism, glycolysis/gluconeogenesis/pyruvate metabolism, and TCA cycle. The statistically significant differentially abundant metabolites are marked with single asterisk at the level of p < 0.001 and FDR < 0.1 and double asterisk p < 0.001 and FDR < 0.05 using LIMMA. Single asterisks indicate statistically significant differences p < 0.001 and FDR < 0.1 and double asterisk p < 0.001 and FDR < 0.05. (B) Gating strategy for Glut1, MCT-1, and xCT in T cells (CD4 and CD8) and monocytes (CM, IM, and NCM). (C) Bubble plot of Glut1, MCT-1, and xCT in subpopulations. Size of the bubble represents proportion of positive cells (%). Color of the bubbles represent MFI. (D) Contour plots showing sample with median percentage of cells for each population. (E) MFI for MCT-1 and xCT in lymphocytes (CD4 and CD8) and monocytes (CM, IM, and NCM). (F) Co-relation analysis between the transporter expression and the differentially altered metabolites between PLWH with MetS and without MetS as shown in Figure 3A. Asterisk indicates p < 0.05.

Community analyses confirm identified biomarkers and display possible mechanisms of metabolic syndrome in PLWH

To establish the metabolic shifts (i.e., the metabolic alterations due to disease conditions) associated with MetS in PLWH, a weighted co-expression network analysis was performed using all detected metabolites and all patients. In this network, the nodes were individual metabolites, and edges were modeled by the positive coefficients of correlation between two metabolites (Spearman, FDR < 0·05). The resulting network consisted of 867 nodes and 45379 edges. Leiden partitioning highlights seven communities of positively intercorrelated metabolites (Figure 4 top). Analysis at the pathway level revealed that the most central community, represented as the center of the network, was enriched for the KEGG terms 2-oxocarboxylic acid metabolism, organic acids, and lipids (Metabolon™ category). This community displayed a very heterogeneous pattern, where only six differentially expressed metabolites (out of 73) were present and half had higher abundance in PLWH with MetS. Interestingly, we observed that a single community contained six potential biomarkers previously identified (Figure 4 bottom). This community also displayed 38 differentially expressed metabolites, of which only one was lower abundant in PLWH with MetS and contained six of the 11 potential biomarkers identified above. These biomarkers showed many interconnections between each other, though other biomarkers including 4-cholesten-3-one were segregated in other communities. This community was enriched in common amino acids (KEGG and Metabolon) and peptides (Metabolon).

Figure 4. The metabolome-wide weighted co-expression network. A weighted metabolite co-expression network was generated (positive Spearman rank correlations, FDR < 0·05). Significant metabolites based on LIMMA have represented opaques and non-significant transparent. Biomarkers found in the first community are represented (bottom-left).

Study population

PLWH were recruited from the Copenhagen Comorbidity in HIV-infection (COCOMO) study, an ongoing longitudinal, observational study to assess the burden of non-AIDS comorbidities in HIV infection. Of the 1099 participants in the COCOMO study, 100 PLWH with MetS were randomly selected and matched to 100 PLWH without MetS. Individuals were matched according to age, sex, duration of cART, smoking status, and current CD4⁺ T-cells count. All plasma samples were collected at the study baseline concomitant with clinical assessment. Given that untargeted metabolomics is a relative measurement of the metabolites we included HIV-negative controls (n = 20). They were also included in this study to define the level of metabolites considered as normal range in the assay run. Procedures for recruitment and data collection for COCOMO have been described elsewhere [27]. Ethical approval was obtained by the Regional Ethics Committee of Copenhagen (COCOMO: H-15017350). Written informed consent was obtained from all participants.

Clinical and biochemical assessments

Structured questionnaires were used in COCOMO to collect information about demographics, physical activity, smoking, lipid-lowering, and antihypertensive therapy [27]. Data regarding HIV-infection were obtained from a review of medical charts [27]. All physical examinations were performed by trained clinic staff, as previously described [27]. Height, weight, hip, and waist measurements and body mass index (BMI) calculations were performed according to WHO guidelines. Blood pressure (BP) was measured on the left arm after 5 minutes rest with the subject in sitting position, using an automatic Digital Blood Pressure Monitor. Non-fasting venous blood was collected and analyzed for LDL-C, total cholesterol, HbA1c, and glucose. Blood samples were analyzed at Herlev Hospital, Copenhagen. Metabolic syndrome (MetS) was defined according to the International Diabetes Federation (IDF) consensus worldwide definition of the MetS [28] as ≥3 of the following: (1) waist circumference ≥94 cm in men and ≥80 cm in women, (2) systolic blood pressure ≥130 mm Hg and/or diastolic blood pressure ≥85 mm Hg and/or antihypertensive treatment, (3) non-fasting plasma triglyceride level ≥1.693 mmol/L, (4) HDL level ≤1.036 mmol/L in men or ≤0.295 mmol/L in women, and (5) self-reported diabetes and/or antidiabetic treatment and/or plasma glucose level ≥11.1 mmol/L.

Sample preparation and untargeted metabolomics

Untargeted metabolite profiling was carried out by Metabolon Inc. (Durham, NC, USA) using ultra-high-performance liquid chromatography/mass spectrometry/mass spectrometry (UHPLC/MS/MS) as described earlier [15, 29]. Data was normalized to sample volume, log-normalized, and minimum-imputed as given by the proprietary pipeline of the provider. The metabolomics method is ISO 9001:2015 certified and the lab is accredited by the College of American Pathologists (CAP), USA.

Statistics and bioinformatics analysis

For clinical data, Welch’s T-test and Mann-Whitney U test were used to compare normally distributed and non-normally distributed continuous variables, respectively. Chi-Square Test was used to compare discrete variables if expected values of the contingency table were five or more. Otherwise, Fisher’s Exact Test was used. For metabolomics data, the median-centered data was used for all the analyses. To define metabolic alterations of the MetS in PLWH, we used a consensus of four different algorithms as the predictive signatures of MetS in PLWH. We used two conventional approaches LIMMA and Mann-Whitney U test., and one machine learning algorithm [random forest (RF)], and one linear classification model (PLS-DA) to identify the most important metabolites associated with MetS in PLWH using R packages ropls and RF, respectively [30]. Both the models separate the groups, based on the importance of metabolites. The top 30 most important variables from the PLS-DA were extracted using VipPlot [31]. For RF, feature selection was performed using the R package Boruta. A RF model using 10-cross fold validation was built using selected metabolites. Performance comparison of models was done using confusion matrix, ROC curves and area under the curve (AUC). As a sensitivity analysis, a similar analysis was performed using only male patients. To identify the consensus biomarkers, overlap of methodologies was represented as a Venn diagram using the R package eulerr ing ggplot2. Adjustment for multiple testing was performed by considering false discovery rate (FDR) <0.05.

Identification of mechanistic pathways

To highlight the most enriched pathways in PLHW with MetS group compared to PLHW without MetS, metabolites with differential abundance (LIMMA, FDR<0.05) were submitted to Ingenuity Pathway Analysis software (IPA) (Qiagen, US). The top 13 enriched pathways and associated metabolites were extracted, represented as a Sankey plot using R package ggalluvial and network using Cytoscape version 3.6.1 [32]. Metabolite-metabolite interactions with high confidence (confidence >0.7) were retrieved from databases and experiments uploaded to STITCH(v5.0) (http://stitch.embl.de/). Association analysis was performed using R and python 3. First, metabolites with low variance were removed. Then, pairwise Spearman correlations between all metabolites were performed. Only correlations with FDR<0.05 were conserved. The distribution of correlation coefficients was plotted to evaluate the size and connection strength of the networks. Accordingly, a weighted co-expression network for significant positive associations for all samples was constructed using the python module igraph (https://igraph.org/python/). Random networks with the same dimensions were also constructed and network properties analyzed. Leiden algorithm was used for partitioning the network using the python module leidenalg [33]. The average degree and clustering coefficient were computed for all communities and functional enrichment was performed for more than 30 metabolites. Metabolite set enrichment analysis (MSEA) was performed on these communities using gseapy first with an in-house script containing Metabolon terms then KEGG terms. Communities partition and network were exported from python and imported to Cytoscape. Metabolites with significant differential abundance between PLWH without MetS and PLWH with MetS based on LIMMA, as well as biomarkers identified by the four methodologies were highlighted in the global network.

Flow cytometry

In a subset of samples PLWH with MetS (n = 31) and without MetS (n = 27) we performed immunophenotyping of peripheral blood mononuclear cells (PBMCs) for transporter expression of glucose transporter (Glut1), monocarboxylic acid transporters (MCT-1), and glutamate/cystine transporter xCT on T-cells (CD4⁺ and CD8⁺) and monocytes (classical, intermediate and non-classical) as described by us recently [29]. All cells were stained with Live/Dead fixable near IR dye (Invitrogen), and cell surface markers were detected by incubating cells with relevant antibodies for 20 min on ice in flow cytometry buffer (antibodies listed in Supplementary Table 4). Cells were acquired a BD FACS Symphony flow cytometer (BD Bioscience), and data were analyzed and compensated with FlowJo 10.7 (TreeStar Inc), and Prism 8 (GraphPad Software Inc).