Cross-sectional examination of metabolites and metabolic phenotypes in uremia

As shown in Table 1, the mean age of the study population was 69.5 years and 69 % of subjects were white. There was an equal representation of males and females and nearly half of the individuals had a history of diabetes or had diabetes listed as their cause of ESRD (49 %). The mean BMI was 26.5 kg/m² (SD ±7.6 kg/m²) and the mean SBP was 144 mmHg (±27 mmHg). A minority of individuals reported a lipid disorder (12 %) and the median total cholesterol level was 162 mg/dL (quartile1-quartile3, 127–188 mg/dL). The median serum albumin level was 3.6 g/dl (3.2-3.8 g/dL).

The PLS-DA approach allowed us to visualize and extract the metabolites that best separated individuals (Figs. 1,2,3,4 and 5; left panels) across phenotype tertile or class (Table 2). Because four of the clinical phenotypes we studied (BMI, serum albumin, total cholesterol, and SBP) are continuous measures, we recognize that tertile cut-offs do not demarcate distinct physiologic or pathophysiologic processes. Thus, plot overlap across tertiles was expected. Variable importance in projection (VIP) provided a score for each metabolite, ranking the metabolites according to their predictive power in the PLS model; the fifteen metabolites with the highest VIP scores for each plot are shown in Figs. 1, 2, 3, 4 and 5 (right panels), and the levels of these metabolites across tertiles (or class) of the phenotypes with corresponding test statistics are shown in Tables 3-7.

Values are median peak area for the metabolites (quartile 1, quartile 3)*P-value significant at the Bonferroni adjusted level of 3.0 × 10−4

For individuals with or without diabetes, glucose was one of the strongest discriminating metabolites, with higher levels among the individuals with diabetes (Fig. 1). Xanthosine-5-phosphate (XMP) and vanillylmandelic acid (VMA), however, had even higher VIP scores than glucose, with XMP levels higher and VMA levels lower among the individuals with diabetes (P = 5.0 × 10⁻⁷ and 1.5 × 10⁻⁶ for XMP and VMA, respectively). Other significant differences included higher levels of 3-phosphoglyceric acid (P = 4.9 × 10⁻⁵), methionine (P = 4.5 × 10⁻⁵), mevalonate-5-phosphate (P = 2.8 × 10⁻⁴), and ornithine (P = 3.2 × 10⁻⁴) among individuals with diabetes. Tyrosine, an aromatic amino acid, had a nominal association with diabetes status (P = 1.0 × 10⁻³). There were no significant associations between tyrosine levels, or any aromatic or branched-chain amino acid levels, with BMI (Table 4). In fact, none of these metabolites had even a nominal association with BMI in univariate comparisons (data not shown). Overall, the ability of metabolite profiles to discriminate individuals across tertiles of BMI was poor (Fig. 2). Among the top metabolites by VIP score, only 20:4 carnitine (arachidonyl carnitine), β-hydroxybutyrate, and benzoate had nominal P-values in univariate analysis (Table 4).Values are median peak area for the metabolites (quartile 1, quartile 3)*P-value significant at the Bonferroni adjusted level of 3.0 × 10−4Values are median peak area for the metabolites (quartile 1, quartile 3)*P-value significant at the Bonferroni adjusted level of 3.0 × 10−4

Thus, our data from an ESRD population do not recapitulate observations derived in the general population linking diabetes and obesity with elevations in branched chain and aromatic amino acids (except for the nominal association with tyrosine noted above), short chain acylcarnitines, the glutamate/glutamine ratio, or bile acids [2]. One potential explanation is that kidney function (and dysfunction) has a direct effect on select metabolites. The interaction between kidney function and amino acid metabolism, in particular, has been closely examined. For example, the kidney is known to make a substantial contribution to whole body tyrosine appearance via intra-organ phenylalanine hydroxylation [46]. Further, the metabolic acidosis that results from renal failure leads to increased leucine oxidation [47], and indeed, generalized muscle catabolism. The significant associations we did identify with diabetes status in ESRD, e.g. XMP (a purine breakdown product) and VMA (a catecholamine metabolite), have not been reported in the general population. To what extent these alterations reflect distinct pathophysiologic processes in uremia warrant further investigation.

For individuals in different tertiles of serum albumin, PLS-DA demonstrated moderate separation across groups (Fig. 3). Tryptophan and its downstream metabolite kynurenine were significantly higher in individuals with higher serum albumin levels (P = 1.8 × 10⁻⁴ and 2.5 × 10⁻⁴, respectively). Two additional tryptophan metabolites, indoxyl sulfate and indole lactate, had a trend for higher levels among individuals with higher serum albumin levels. Given recent observations that ascribe a functional role for tryptophan metabolism, specifically through indoleamine-2, 3-dioxygenase (IDO), in modulating the immune system and vascular tone [48, 49], it is an appealing link between metabolism, nutritional status, inflammation, and cardiovascular risk in ESRD. However, these data should be interpreted with caution. Because tryptophan and its catabolites are hydrophobic, their positive correlation with albumin may reflect their significant protein-binding [50] – our LC-MS method measures total, not free, plasma metabolite levels. A similar mechanism may underlie the trend for association between thyroxine and albumin levels. By contrast, creatine, the metabolite with the third highest VIP score, had a trend for higher levels among individuals with lower serum albumin levels. In a study that examined spent media from cultured muscle cells treated with mitochondrial respiratory chain inhibitors, as well as plasma obtained from individuals with respiratory chain diseases, creatine levels were reproducibly elevated [51]. In cell culture, extracellular creatine was inversely correlated with the intracellular phosphocreatine/creatine ratio, suggesting that elevated plasma creatine may signal a low energetic state in tissues using the creatine phosphate shuttle. Interestingly, elevated plasma creatine and decreased plasma phosphocreatine levels both have a nominal association with increased 1-year cardiovascular mortality among incident dialysis patients in ArMORR [27]. These observations raise the possibility that impaired mitochondrial respiration in muscle is linked to the pathogenesis of hypoalbuminemia in dialysis.

For individuals across tertiles of total cholesterol and SBP, PLS-DA again demonstrated moderate separation across tertiles (Figs. 4 and 5). For both analyses, long-chain acylcarnitines were among the metabolites with the highest VIP scores, with 18:2 carnitine (linoleylcarnitine, P = 9.0 × 10⁻⁷), 18:1 carnitine (oleoylcarnitine, P = 6.0 × 10⁻⁷), and 18:0 carnitine (stearoylcarnitine, P = 7.7 × 10⁻⁵) all significantly higher in individuals with lower total cholesterol levels (Table 6,) and linoleylcarnitine (P = 5.5 × 10⁻⁵) and oleoylcarnitine (P = 8.7 × 10⁻⁶) significantly higher in individuals with lower SBP (Table 7); select short-chain and medium-chain acylcarnitines were also significantly elevated among individuals with lower SBP. Like albumin, low cholesterol and blood pressure are established predictors of increased mortality in ESRD [8, 9, 11, 12]. These findings are consistent with our observation that long-chain acylcarnitines are associated with 1-year cardiovascular mortality in ESRD [27], and motivate further interest in long-chain acylcarnitines as markers, or even mediators, of altered lipid metabolism and cardiovascular function in uremia. Tryptophan, one of the strongest discriminators of albumin tertile, was also a top hit in regards to total cholesterol and SBP, with higher tryptophan significantly associated with lower levels for both. Finally, higher arginine levels were significantly associated with higher total cholesterol and higher deoxyadenosine levels were significantly associated with lower SBP.Values are median peak area for the metabolites (quartile 1, quartile 3)*P-value significant at the Bonferroni adjusted level of 3.0 × 10−4

To examine the specificity of the long-chain acylcarnitine-phenotype associations described above, we looked more broadly at the correlation between all of the metabolites captured on our platform that are uremic retention solutes and 24 phenotypes captured in ArMORR. Correlations between previously reported uremic retention solutes measured by our LC-MS platform, including long-chain acylcarnitines, and a broad array of clinical phenotypes measured in ArMORR are shown in Fig. 6 (full results are shown in Additional file 1: Tables S2 and S3). The correlations for glucose, as measured by the mass spectrometer, provide an anchor for interpreting this data – LC-MS glucose had a strong positive correlation with glucose measured by the clinical laboratory (r = 0.80, P <10⁻¹⁵), and a modest negative correlation with serum sodium (r = −0.21, P = 7.0 × 10⁻³) that did not reach statistical significance, but is well recognized in the clinic (i.e. dilutional hyponatremia due to hyperglycemia). Aside from the correlation between the two glucose measures, the strongest correlations identified by this broader examination of uremic metabolites were indeed between long-chain acylcarnitines and lipid traits (total and LDL cholesterol) and blood pressure. Among lipid phenotypes, long-chain acylcarnitines such as oleoylcarnitine were more strongly correlated with total (r = −0.45, P = 2.6 × 10⁻⁹) and LDL cholesterol (r = −0.43., P = 1.7 × 10⁻⁸) than with HDL cholesterol (r = −0.15, P = 5.7 × 10⁻²) or triglycerides (r = −0.27, P = 5.0 × 10⁻⁴). Similarly, oleoylcarnitine was more strongly correlated with SBP (r = −0.38, P = 3.4 × 10⁻⁸) than diastolic blood pressure (r = −0.25, P = 3.0 × 10⁻³). Long-chain acylcarnitines were not significantly correlated with any non-lipid or non-blood pressure phenotype.

Among the other uremic retention solutes included in our analysis, only kynurenic and anthranilic acid (both tryptophan metabolites downstream of IDO) had more than one significant correlation with a clinical variable (Fig. 6). Kynurenic acid was significantly correlated with diastolic blood pressure (r = 0.32, P = 5.0 × 10⁻⁶), bicarbonate (r = −0.31, P = 1.1 × 10⁻⁵), and phosphorous (r = 0.38, P = 5.7 × 10⁻⁸), whereas anthranilic acid was significantly correlated with alkaline phosphatase (r = 0.29, P = 6.3 × 10⁻⁵) and ferritin (r = 0.35, P = 9.7 × 10⁻⁷), a frequently used marker of inflammation in hemodialysis. Notably, the other metabolites that had a nominal association with ferritin included two other tryptophan metabolites, kynurenic acid (r = 0.16, P = 2.6 × 10⁻²) and quinolinate (r = 0.20, P = 6.3 × 10⁻³), and malondialydehyde (r = 0.24, P = 9.0 × 10⁻⁴), an oxidative stress marker. These findings reinforce a potential link between the IDO pathway and uremic inflammation. Sorbitol (r = −0.34, P = 3.4 × 10⁻⁵), cytidine (r = −0.40, P = 1.1 × 10⁻⁶), and uracil (r = −0.42, P = 1.7 × 10⁻⁷) were significantly correlated with plasma glucose; urate (r = −0.35, P = 5.2 × 10⁻⁶) and malondialdehyde (r = 0.38, P = 5.7 × 10⁻⁷) were significantly correlated with triglycerides; hippurate (r = 0.31, P = 7.0 × 10⁻⁶) and oxalate (r = 0.36, P = 1.6 × 10⁻⁷) were significantly correlated with phosphorous; and indoxyl sulfate (r = −0.28, P = 6.8 × 10⁻⁵) was significantly correlated with bicarbonate. Indoxyl sulfate, which has been found to have several proinflammatory effects in model systems [52], had no correlation with ferritin (nor transferrin saturation, nor white blood cell count), but had a trend for a negative correlation with BNP (r = −0.44, P = 4.0 × 10⁻⁴).

Our study has several limitations. First, by examining cross-sectional relationships between metabolite levels and various clinical phenotypes, our study is unable to address causation. Instead, by applying parsimonious selection procedures and conservative adjusted significance thresholds, we seek to highlight notable associations for future investigations. Second, study subjects were selected in the context of a case–control study of 1-year cardiovascular mortality, raising the possibility of confounding by case status. However, as noted in the Methods, the statistical significance of the findings are not changed when analyzed stratified by case status. Nevertheless, we acknowledge that our sample is not a random selection of incident dialysis patients, potentially limiting the generalizability of our findings. Third, residual confounding from other sources is likely to influence the associations described, as multivariate adjustments were not pursued. However, because the relationship between various clinical phenotypes and metabolite levels in uremia is unknown, we did not want to obscure novel biological associations by statistical adjustment. Thus, we present these raw associations as a framework for interpreting and appropriately adjusting select findings in future metabolomics studies in ESRD. Finally, our results require replication in an independent sample, ideally selected randomly and including both incident and prevalent dialysis patients.