Skip to main content
NCBI home page
Kidney360 logoLink to Kidney360
. 2025 Sep 23;7(2):353–361. doi: 10.34067/KID.0000000838

A Metabolomics Approach To Identify Metabolites Associated with Uremic Symptoms in Patients Receiving Maintenance Hemodialysis

Solaf Al Awadhi 1, Leslie Myint 2, Eliseo Guallar 3, Clary B Clish 4, Kendra E Wulczyn 5, Sahir Kalim 5, Ravi Thandhani 6, Dorry L Segev 7, Mara McAdams-DeMarco 7, Sharon M Moe 8, Ranjani N Moorthi 8, Jonathan Himmelfarb 9, Neil R Powe 10, Marcello Tonelli 11, Eugene P Rhee 5, Tariq Shafi 12,✉
PMCID: PMC12935353  PMID: 40986394

Abstract

Key Points

  • Uremic symptoms are common in patients with kidney failure, but the chemicals causing them are unknown.

  • We used state-of-the-art untargeted metabolomics and rigorous analyses to search for metabolites associated with uremic symptoms.

  • We identified metabolite-symptom associations but could not replicate findings and offer recommendations for future research.

Background

The specific toxins causing uremic symptoms (nausea, vomiting, pruritus, fatigue, difficulty concentrating, and pain) in kidney failure remain unknown. We used untargeted metabolomics to identify plasma metabolites associated with uremic symptoms in patients receiving hemodialysis.

Methods

We measured metabolites in plasma samples from Longitudinal US/Canada Incident Dialysis study participants at baseline (discovery; n=636) and year 1 (internal validation; n=260) and from Frailty Assessment in Renal Disease study participants (external validation; n=355). We used metabolite-wise linear models with empirical Bayesian inference to evaluate the association between metabolites and uremic symptom severity, adjusting for key covariates. We accounted for multiple testing using a false discovery rate (P values adjusted for false discovery rate [pFDR]) for linear models and used two machine learning models to evaluate the association consistency. We defined association as significant if pFDR < 0.1 and consistent if it had medium or high importance in both machine learning models.

Results

Participants had a mean age of 63 years, with uremic symptom prevalence ranging from 44% to 83%. We identified 627 previously characterized (known) and 35,558 unknown metabolite peaks. No known metabolites were significantly and consistently associated with uremic symptom severity across all cohorts. Within cohorts, retinol was negatively associated with nausea/vomiting in Longitudinal US/Canada Incident Dialysis at year 1, and indole-3-propionic acid was negatively associated with anorexia in Frailty Assessment in Renal Disease. Several unknown metabolites were associated with symptoms (lowest pFDR, 0.0004), but none were consistent across cohorts.

Conclusions

We identified metabolites associated with uremic symptom severity, although findings were inconsistent across cohorts. This study highlights the need for additional research on uremic toxins and clinical outcomes.

Keywords: dialysis, ESKD, hemodialysis, kidney failure, metabolism, outcomes, patient-centered care, quality of life, metabolomics

Introduction

Symptoms such as anorexia, nausea, vomiting, pruritus, fatigue, and bodily pain are common in patients with kidney failure and contribute considerably to poor quality of life.1–5 The well-recognized clinical observation that these symptoms worsen with kidney failure, reduce in intensity after dialysis initiation but persist despite achieving prescribed dialysis urea clearance targets, and improve markedly or resolve completely after kidney transplantation supports the long-standing view that retained uremic toxins that are not adequately cleared by conventional dialysis are the most likely cause of these symptoms.6–10 However, the specific uremic toxins causing these symptoms are unknown, limiting the ability to develop treatments targeting uremic symptoms.

The search for uremic toxins has followed the traditional bench-to-bedside paradigm, which involves demonstrating solute toxicity in vitro or in animal models, followed by developing targeted assays and clinical validation. This approach has identified approximately 150 uremic solutes over the past 50 years,11 but none have been directly linked to uremic symptoms. Metabolomics, the systematic analysis of metabolites in biologic specimens, can identify many metabolites in plasma and help advance our understanding of metabolic pathways. However, few studies have used metabolomics to investigate potential mechanisms of uremic symptoms.12–16

The goal of our study was to determine the association between plasma metabolite levels and the severity of several uremic symptoms in patients treated with maintenance hemodialysis for kidney failure. We used stored plasma samples from the Longitudinal US/Canada Incident Dialysis (LUCID) and the Frailty Assessment in Renal Disease (FAIR) studies. We measured plasma metabolites using a liquid chromatography-mass spectrometry (MS)–based metabolomics platform and used state-of-the-art high-dimensional data methods to evaluate the association between metabolites and patient-reported uremic symptoms.

Methods

Study Population

We included participants from the prospective, multicenter LUCID and FAIR studies1,17 because they involved patients with kidney failure on maintenance hemodialysis and symptom assessments via the Kidney Disease Quality of Life Instrument (KDQOL-36) and stored blood samples for metabolomics. LUCID enrolled 823 participants in Canada (2011–2017) who were 18 years or older and initiated hemodialysis within the past 6 months. FAIR included 3255 participants undergoing kidney transplant evaluation at Johns Hopkins (2009–2018) and the University of Michigan (2015–2016).17 Both studies were approved by the human subjects review boards of the participating institutions, and all participants provided informed consent.

For this study, we selected participants with available blood samples and data on symptoms and key covariates. On the basis of these criteria, we included 636 LUCID participants at baseline (discovery cohort), 260 of the 636 LUCID participants at year 1 (internal validation cohort), and 355 FAIR participants (external validation cohort).

Uremic Symptoms

We evaluated the following patient-reported uremic symptoms: fatigue, pruritus, anorexia, nausea/vomiting, difficulty concentrating, and bodily pain. We selected these symptoms because they are generally considered to be due to the retention of uremic toxins in individuals with advanced kidney failure, improve partially with dialysis, and often persist despite dialysis treatments.1,6 In both LUCID and FAIR, these symptoms were assessed by KDQOL-36, administered by study coordinators. For each symptom, the participants reported to what extent they were bothered by it in the preceding 4 weeks, with responses ranging from not at all bothered to extremely bothered. For this analysis, we recategorized the symptoms from the original five- or six-point scale to a three-point scale (none/mild, moderate, and severe; Supplemental Table 1) to achieve a balance between interpretability and adequate granularity.

Samples

For LUCID, the blood samples were obtained at the dialysis clinic before the start of hemodialysis and were collected in serum separator tubes and EDTA tubes and centrifuged at 3000 revolutions per minute for 10 minutes. They were aliquoted on the same day, stored at −80°C at each site, and then sent to the central storage sites in the United States and Canada. For FAIR, samples were obtained on the day of transplant evaluation, which occurs on a nondialysis day, and were processed similarly and stored at −80°C at Johns Hopkins University. For both studies, the stored aliquoted plasma samples were retrieved from the freezers and shipped to the Broad Institute on dry ice for analysis. There were no freeze-thaw cycles before the samples were shipped.

Metabolomics

The plasma samples were profiled using three liquid chromatography-tandem MS injections, which consist of Nexera X2 U-HPLC systems (Shimadzu Scientific Instruments) and Q Exactive/Exactive Plus orbitrap mass spectrometers (Thermo Fisher Scientific). Positively charged polar metabolites were extracted with acetonitrile/methanol/formic acid and separated by hydrophilic interaction liquid chromatography, and MS analyses were performed in full scan mode using positive ion mode electrospray ionization (ESI). Positively charged lipids were extracted with isopropanol and separated on a C8 column, and MS analyses were performed in full scan mode using positive ion mode ESI. Negatively charged polar metabolites were extracted from samples with methanol and separated by an NH2 column, and MS analyses were performed in full scan mode using negative ion mode ESI. Known metabolites were analyzed using TraceFinder, whereas unknown metabolites were processed with Progenesis CoMet software. See Supplemental Methods for more details on analytical methods and quality control.

A total of 630 known metabolites were detected in all the specimens in LUCID and FAIR. We removed metabolites missing in 95% of samples, resulting in 627 known metabolites across all samples. We considered the unknown metabolite peaks to be the same if the mass-to-charge ratio (m/z) of the two peaks was within 0.05 and the retention time (r/t) was within 15 seconds. On the basis of these criteria, 38,308 unknown metabolite peaks were detected in all the specimens of LUCID and FAIR. We also removed unknown metabolite peaks missing in 95% of samples, resulting in 35,558 unknown metabolite peaks across all samples.

Covariates

The covariates used in adjusted models included age, sex, height, weight, history of cardiovascular disease and diabetes, and country (the United States/Canada). We did not include participants with missing covariate information (64 LUCID participants at baseline, 40 LUCID participants at year 1, and 17 FAIR participants).

Statistical Analysis

We described participant characteristics at baseline using means, SDs, medians, and interquartile ranges (IQRs) for continuous variables and frequencies and proportions for categorical variables. We used STrengthening the Reporting of OBservational studies in Epidemiology guidelines for reporting (Supplemental Table 2).18

Metabolite Data Processing

Processing high-dimensional metabolomics data requires addressing variability and missingness. Potential sources of variability can be preanalytical (premeasurement), because of variations in sample collection, processing, and storage, or analytical (instrument drift) because of the technique used and reflected by signal intensity drifts and r/t.19,20 We evaluated instrumental drift using correlations between the quality control samples and performed principal components analysis (PCA) on the abundance of metabolites (see Supplemental Figure 1, Supplemental Methods, and Supplemental Table 3). We adjusted for instrument drift using the removal of unwanted variation (RUV) method21 (see Supplemental Figure 2 and Supplemental Methods). The RUV method used all included samples without excluding those with missing covariates. Our primary analysis uses RUV correction, and we conducted sensitivity analyses without RUV correction.

Metabolite data may be missing because of the complete absence of a metabolite in plasma or a level below the lowest detection limit on the metabolomics platform. We addressed missing metabolite data by evaluating the performance of several imputation methods, including half-minimum (uniformly distributed random values from 0 to half of the minimum observed abundance), k-nearest neighbors methods, and quantile regression imputation of left-censored (QRILC) data method22,23 (see Supplemental Figure 3 and Supplemental Methods). QRILC data showed the lowest imputation error of all methods, and QRILC-imputed metabolite data were used for all subsequent analyses.

Nonmetabolite Data Processing

We used a random forest (RF) imputation approach as implemented in the missForest R package to account for missing symptoms data.24 We excluded patients with missing covariates because of high imputation error. Because of lower imputation error for symptoms, we conducted sensitivity analyses using imputed symptom data (see Supplemental Methods and Supplemental Table 4).

Differential Analysis

We used a differential analysis framework, which is best practice when working with high-throughput assays,25–27 by a priori assuming that metabolites identified by more than one statistical model would be more likely associated with symptom severity. We applied three statistical models, including linear models with empirical Bayesian inference (as implemented in the limma software package), and the two machine learning (ML) models, least absolute shrinkage and selection operator (LASSO) and RF, to identify log2-transformed metabolite concentrations (outcome) associated with symptom severity categories (exposure). This approach, which uses metabolites as an outcome, is preferred because it allows for the accounting of instrument drift in metabolite levels using the RUV method.

Linear models with empirical Bayesian inference are a modeling strategy that shrinks sample variance toward a pooled estimate.28 It is advantageous over standard linear regression models because shrinking variance estimates leads to more stable and powerful inference (see Supplemental Methods).25,29 We adjusted P values using a false discovery rate threshold of 10% (P values adjusted for false discovery rate [pFDR] < 0.10)30,31 to identify metabolites significantly associated with symptom severity. We used the coefficients of symptom severity indicators as our effect measures, which give the adjusted log fold change in metabolite concentration between individuals with low versus medium and low versus high symptom severity.

We used the ML LASSO and RF models to crosscheck metabolites predictive of symptom severity in the linear models. We categorized variable importance from these models as high (≥90th percentile), medium (70–90th percentile), or low (<70th percentile; see Supplemental Methods).

We a priori considered a metabolite to be significantly associated with symptom severity in each cohort if it met the following criteria: (1) the pFDR was <0.1 in linear models and (2) the variable importance by both LASSO and RF was medium or high. We considered LUCID baseline as our discovery cohort, repeated measurements for LUCID participants at year 1 as our internal validation cohort, and FAIR as the external validation cohort. We defined a metabolite associated with symptoms to be externally validated if it was associated with symptom severity in the discovery, internal validation, and external validation cohorts.

Additional Analyses

Our primary analysis investigated known metabolites and included RUV correction for instrumental drift and no imputation (RUV=yes, imputation=no). Sensitivity analysis 1 did not account for instrumental drift and did not include covariate and symptom imputation (RUV=no, imputation=no), sensitivity analysis 2 accounted for instrumental drift and included imputed covariates and symptoms (RUV=yes, imputation=yes), and sensitivity analysis 3 did not account for instrumental drift and included imputed covariates and symptoms (RUV=no, imputation=yes). In addition, we explored whether any unknown metabolite peaks were associated with uremic symptoms in linear models (adjusted pFDR < 0.1 in either primary or sensitivity analyses) in any of the cohorts. All analyses used R (R Core Team, 2023).

Results

Participants' Characteristics and Symptom Prevalence

At baseline, LUCID participants had a median age of 63 years (interquartile range [IQR], 51–72); 81% were male, and 45% and 53% of patients had a history of cardiovascular disease and diabetes, respectively (Table 1). The FAIR cohort participants were younger (53 years; IQR, 43–63) and had a lower prevalence of cardiovascular disease (34%) and diabetes (39%). The prevalence of symptoms was generally higher in the LUCID participants.

Table 1.

Participants' characteristics and uremic symptom distribution

Characteristic Discovery (LUCID Baseline), n=636 Internal Validation (LUCID Year 1), n=260 External Validation (FAIR), n=355
Age, yr, median (IQR) 63 (51–72) 64 (51–72) 53 (43–63)
Male, No. (%) 514 (81) 209 (80) 238 (67)
Race/ethnicity, No. (%)
 Non-Hispanic Black 106 (17) 59 (23) 187 (53)
 Non-Hispanic White 442 (69) 171 (66) 132 (37)
 Other 85 (13) 30 (12) 36 (10)
BMI, kg/m2, median (IQR) 27 (23–31) 27 (23–31) 28 (24–32)
Diabetes, No. (%) 338 (53) 148 (57) 137 (39)
Cardiovascular disease, No. (%) 286 (45) 130 (50) 122 (34)
Symptom prevalence, No. symptomatic/No. respondent (% symptomatic)
 Fatigue 333/427 (78) 133/167 (80) 237/355 (67)
 Pruritus 279/426 (65) 120/168 (71) 195/355 (55)
 Anorexia 196/431 (45) 76/169 (45) 143/355 (40)
 Nausea/vomiting 189/429 (44) 71/169 (42) 146/355 (41)
 Difficulty concentrating 248/422 (59) 98/165 (59) 167/355 (47)
 Bodily pain 380/460 (83) 153/183 (84) 245/355 (69)
Open in a new tab

BMI, body mass index; FAIR, Frailty Assessment in Renal Disease; IQR, interquartile range; LUCID, Longitudinal US/Canada Incident Dialysis.

Overall Primary Analysis Results

In our primary analyses, which focused on known metabolites and accounted for instrument drift and did not include imputation, no metabolites were significantly (linear models with adjusted pFDR < 0.1 and ML models with medium or high importance) and consistently associated with symptom severity across the discovery and internal and external validation cohorts. Per cohort, no metabolites were associated with any of the symptoms in the LUCID baseline; one metabolite was associated with nausea/vomiting in the LUCID year 1, and one metabolite was associated with anorexia in the FAIR (Table 2).

Table 2.

Metabolites significantly associated with uremic symptom severity

Symptom HMDB ID Metabolite Molecular Weight (g/mol) Discovery Cohort (LUCID Baseline), n=636 Internal Validation Cohort (LUCID Year 1), n=230 External Validation Cohort (FAIR), n=355
Adjusted FC pFDR ML Importance Measure Adjusted FC pFDR ML Importance Measure Adjusted FC pFDR ML Importance Measure
Medium versus Low High versus Low LASSO RF Medium versus Low High versus Low LASSO RF Medium versus Low High versus Low LASSO RF
Primary analysis (RUV correction=yes; imputation=no)
 Anorexia HMDB0002302 Indole-3-propionic acid 189 1.11 0.89 0.967 Low Low 0.70 0.53 0.615 Low Medium 0.65 0.52 0.024a Higha Higha
 Nausea/vomiting HMDB0000305 Retinol (vitamin A) 287 1.02 1.03 0.999 Low Low 0.82 0.66 0.092a Higha Higha 1.00 1.08 0.969 Low Low
Sensitivity analysis 1 (RUV correction=no; imputation=no)
 Anorexia HMDB0002302 Indole-3-propionic acid 189 1.07 0.91 0.985 Low Low 0.71 0.55 0.723 Low Medium 0.65 0.51 0.028a Higha Higha
HMDB0000317 2-Hydroxy-3-methylpentanoic acid 132 1.04 1.41 0.028a Higha Higha 1.03 1.57 0.198 Medium High 1.03 1.11 0.991 Low Medium
 Bodily pain HMDB0001859 Acetaminophen 151 1.30 1.11 0.969 Medium Medium 1.49 2.63 0.517 Medium Medium 1.74 2.92 0.040a Higha Higha
Sensitivity analysis 2 (RUV correction=yes; imputation=yes)
 Anorexia HMDB0002302 Indole-3-propionic acid 189 1.03 0.88 0.963 Low Low 0.91 0.56 0.886 Low Medium 0.65 0.52 0.024a Higha Higha
HMDB0010393 C20:3 LPC_A 545 1.04 1.04 0.952 Low High 0.85 0.62 0.023a Mediuma Higha 0.97 0.96 0.995 Low Low
 Nausea/vomiting HMDB0000317 2-Hydroxy-3-methylpentanoic acid 132 1.07 1.32 0.994 High High 1.13 1.59 0.096a Higha Higha 1.03 0.97 0.949 Low Low
Sensitivity analysis 3 (RUV correction=no; imputation=yes)
 Anorexia HMDB0002302 Indole-3-propionic acid 189 0.99 0.91 0.980 Low Medium 0.92 0.58 0.886 Low Medium 0.65 0.51 0.028a Higha Higha
HMDB0000317 2-Hydroxy-3-methylpentanoic acid 132 1.04 1.38 0.051a Mediuma Higha 1.07 1.55 0.140 Medium High 1.03 1.11 0.991 Low Medium
HMDB0000300 Uracil 112 0.88 1.09 0.051a Mediuma Higha 0.93 0.90 0.886 Low Low 0.95 0.96 0.991 Low Low
NA Adenosine monophosphate 347 0.87 1.57 0.062a Mediuma Higha 0.92 0.79 0.912 Low Medium NA NA NA NA NA
HMDB0000122 Fructose/glucose/galactose 180 1.19 1.00 0.062a Mediuma Higha 1.07 1.21 0.896 Low Medium 1.00 1.02 0.991 High Low
HMDB0000289 Uric acid 168 1.00 0.85 0.081a Higha Mediuma 0.98 1.02 0.966 Low Low 0.98 1.00 0.991 Low Medium
 Bodily pain HMDB0001859 Acetaminophen 151 1.28 1.16 0.982 Medium High 1.43 2.45 0.616 Medium Medium 1.74 2.92 0.040a Higha Higha
Open in a new tab

This table includes the metabolites significantly associated (P values adjusted for false discovery rate < 0.1) with uremic symptom severity in any of the three cohorts. No metabolites were significant across all cohorts in either primary or secondary analyses.

In the primary analysis, no metabolites were associated with bodily pain, difficulty concentrating, or pruritus. In sensitivity analysis 1, no metabolites were associated with difficulty concentrating, nausea/vomiting, or pruritus. In sensitivity analysis 2, no metabolites were associated with bodily pain or pruritus. In sensitivity analysis 3, no metabolites were associated with difficulty concentrating, nausea/vomiting, or pruritus.

See Methods for details of imputation procedures.

Interpretation of the coefficients: The adjusted fold change describes the association between symptom intensity and the metabolite level. For example, in the Frailty Assessment in Renal Disease Study cohort primary analysis, the indole-3-propionic acid levels are 0.65-fold lower in those with moderate anorexia compared to those with mild/none anorexia and 0.52-fold lower in those with severe anorexia compared to those with mild/none anorexia. FAIR, Frailty Assessment in Renal Disease; FC, fold change; HMDB, Human Metabolome Database; LASSO, least absolute shrinkage and selection operator; LUCID, Longitudinal US/Canada Incident Dialysis; ML, machine learning; NA, not available; pFDR, P values adjusted for false discovery rate; RF, random forest; RUV, removal of unwanted variation.

a

Metabolites were considered to be significantly associated with symptom severity in a cohort on the basis of the following criteria: (1) The P values adjusted for false discovery rate was <0.1 in linear models, adjusted for age, sex, height, weight, cardiovascular disease, diabetes, and country (the United States/Canada) and (2) the variable importance by both least absolute shrinkage and selection operator and random forest was medium or high.

Specific Metabolites and Symptom Association Results

Indole-3-propionic acid was significantly associated with anorexia in the FAIR cohort for the primary and all of the sensitivity analyses (Table 2 and Supplemental Results). The association was negative, i.e., those with greater severity of anorexia had lower indole-3-propionic acid concentrations compared with those with no/minimal anorexia. Indole-3-propionic acid was not significantly associated with anorexia in LUCID baseline or year 1 cohorts, but there was a consistent trend toward a negative association in those with more severe anorexia symptoms.

Retinol was significantly and negatively associated with the severity of nausea/vomiting in the LUCID year 1 cohort primary analysis; however, it was NS across sensitivity analyses in the same or across other cohorts. 2-Hydroxy-3-methylpentanoic acid was significantly and positively associated in some sensitivity analyses with anorexia (LUCID baseline) and nausea/vomiting (LUCID year 1). Lysophospholipid C20:3 LPC_A was significantly and negatively associated with anorexia in one of the sensitivity analyses (LUCID year 1).

Acetaminophen was significantly and positively associated with bodily pain in two of the LUCID cohort sensitivity analyses but not in the primary analysis. In the LUCID baseline and year 1 cohorts, these associations were NS, but the directionality of the association was consistently positive.

Uracil, uric acid, AMP, and fructose/glucose/galactose were all significantly associated with anorexia in some sensitivity analyses (LUCID baseline); however, the direction of association was not definitive, and the results did not replicate in other cohorts or analyses.

Unknown Metabolite Peaks and Symptom Association Results

None of the unknown metabolite peak-symptom associations were consistent across all cohorts using a combination of primary and sensitivity analyses. Thus, no unknown metabolite peaks met our rigorous a priori criteria for discovery and internal and external validation. Unknown metabolite peaks were associated with anorexia (n=21, pFDR ≥ 0.026), nausea/vomiting (n=54, pFDR ≥ 0.0004), bodily pain (n=87, pFDR ≥ 0.007), fatigue (n=1, pFDR = 0.028), and pruritus (n=3, pFDR ≥ 0.018). None of the unknown metabolite peaks were associated with difficulty concentrating (see Supplemental Results for degree of association of all detected known and unknown metabolites with uremic symptom severity).

Discussion

Conventional, thrice-weekly hemodialysis treatments reduce the intensity of many uremic symptoms but do not fully ameliorate them. We designed this study to systematically discover and externally validate metabolites associated with uremic symptoms, using rigorous analytic techniques and state-of-the-art high-dimensional data methods. We identified several metabolites associated with specific uremic symptoms. However, none of the metabolites met our rigorous validation criteria. Our findings highlight the potential applications and limitations of an epidemiologic approach to discovering the metabolites associated with uremic symptoms.

Uremic symptoms improve with the initiation of dialysis in patients with advanced CKD and worsen with missed dialysis treatments, suggesting that dialyzable uremic toxins are responsible for these symptoms. We identified some metabolite-symptom intensity associations, although the findings did not replicate across cohorts or sensitivity analyses. We found seven metabolites associated with anorexia in individual cohorts. Of those, indole-3-propionic acid and 2-hydroxy-3-methylpentanoic acid merit closer evaluation. We found that increasing anorexia severity was associated with lower levels of indole-3-propionic acid. This association was statistically significant only in the FAIR cohort, but the direction of association was consistent in the LUCID baseline and year 1 cohorts. Indole-3-propionic acid (HMDB0002302) is a 189-Da metabolite. It is an indolyl carboxylic acid and a reductive product of tryptophan formed by bacteria in the gastrointestinal tract.32 Indole-3-propionic acid has been shown to have general beneficial effects by inhibiting the synthesis of inflammatory factors, suppressing intestinal immune responses, and maintaining the intestinal barrier by stimulating the expression of tight junction proteins that enhance barrier function.33,34 Indole-3-propionic acid and microbial byproducts of tryptophan metabolism are being investigated in a variety of disorders, including the gut-brain axis and the gut-heart axis.35 Indole-3-propionic acid has been considered a beneficial tryptophan uremic solute,36 although the associations noted in our study may simply reflect poor food intake because of anorexia. We also found that greater intensity of anorexia and nausea/vomiting was associated with higher levels of 2-hydroxy-3-methylpentanoic acid (HMDB0000317), a 132-Da organic acid generated by L-isoleucine metabolism.37 Its levels are known to be increased in patients with inborn metabolic disorders such as maple syrup urine disease.37 It crosses the blood-brain barrier and has been detected in cerebrospinal fluid.38 No previous studies have found an association between 2-hydroxy-3-methylpentanoic acid and appetite or nausea/vomiting. We also noted that greater intensity of bodily pain was associated with higher levels of acetaminophen. This expected association suggests a positive control for this study.

A limitation of all metabolomics methodologies is incomplete coverage of the metabolome, which spans tens of thousands of distinct compounds.39 Of the thousands of peaks generated by the liquid chromatography-tandem MS methods used here, only a subset have an unambiguous metabolite identification (known metabolites). To more broadly test the hypothesis that blood levels of a metabolite, even if its identity is unknown, are associated with uremic symptoms, we performed an exploratory analysis of >30,000 unknown metabolite peaks captured by our methods. As with our primary analysis, we did not find any associations that met our validation criteria.

Five previous studies have used metabolomics to search for metabolites associated with uremic symptoms; two evaluated uremic symptoms in patients with CKD not on dialysis,12,13 two evaluated metabolites associated with pruritus in patients receiving hemodialysis,14,16 and one evaluated metabolites associated with cognitive impairment in patients receiving hemodialysis15 (Supplemental Table 5). In patients with CKD not yet receiving hemodialysis, Hu et al. assessed the association of metabolites with symptom severity in the Modification of Diet in Renal Disease study cohort (n=695).12 Symptoms were assessed using the Modification of Diet in Renal Disease Symptom Questionnaire, and metabolites were measured by Metabolon. None of the associations were replicated in our study. Wulczyn et al. assessed the association of metabolites with symptom severity in the Chronic Renal Insufficiency Cohort study participants not on dialysis (n=1761).13 Symptoms were assessed using KDQOL-36, and metabolites were measured on the Broad Institute platform, both similar to our study. However, none of the metabolite-symptom associations were replicated in our study. Bolanos et al. evaluated the association of uremic pruritus with metabolites in patients with kidney failure on hemodialysis (n=57) but found no significant associations.16 Wu et al. evaluated the pruritus-metabolite associations in a study of 153 patients.14 They reported several associations but did not account for multiple comparisons. Tamura et al. evaluated the association between cognitive function and metabolites in the participants of the Frequent Hemodialysis Network trials (n=321).15 Cognition was assessed using the Trails Making Test and the Digit Symbol Substitution Test, and metabolites were measured on the Metabolon platform. They found that phenylacetylglutamine levels were associated with cognition, and phenylacetylglutamine, hippurate, and prolyl-hydroxyproline were associated with executive function. Our study used the KDQOL-36 forgetfulness question and found no associations. Notably, none of these prior studies assessed a large number of unknown metabolite peaks in addition to their primary assessment of known metabolites.

The strengths of our study include our relatively large sample size, a multicenter cohort, measurement of metabolites using a state-of-the-art metabolomics platform, and a rigorous statistical approach. Our inability to find an association despite these strengths will inform future work and could be due to several reasons. First, there may be measurement errors in the uremic symptom assessment used in our study. We used patient-reported uremic symptom intensity from the KDQOL-36, a validated quality of life tool for advanced CKD, although its validity for individual symptoms is unknown. Although the presence of some symptoms, such as vomiting, might be unambiguous, its perceived intensity could be subjective and driven by other factors, such as mood and depression. Development of a uremic symptom-specific instrument is needed, and future research could incorporate such instruments to improve accuracy and reliability in assessing uremic symptom severity. Second, we could not completely account for short-term metabolite variability and its contributing factors. Metabolite levels can vary considerably within individuals and are affected by factors such as diet and microbiome. We did not have short-term repeat measurements, dietary recall, or microbiome assessments. Our sampling time was also not standardized on the basis of the prior interdialytic interval. Future studies should incorporate more frequent or time-specific symptom assessments alongside serial metabolite measurements, including predialysis and postdialysis sampling, to better capture dynamic fluctuations in metabolites. Third, current metabolomics platforms assess total plasma metabolite levels, but the free (biologically active) fraction may be more relevant, and overall metabolome coverage remains incomplete, even when unknown peaks are included. Fourth, preanalytic variability due to variations in sample collection, processing, and storage techniques can also lead to measurement errors. Fifth, we did not have an assessment of residual kidney function, which may affect both clearance and metabolic function of the kidney; future research should incorporate this variable to improve the accuracy of metabolite-symptom associations. Sixth, our study did not account for potential interactions between metabolites, which may jointly influence symptom severity; future studies using network-based or pathway enrichment approaches could help uncover such synergistic effects. Finally, although we tested for associations, the directionality and causality of the relationship between metabolites and uremic symptoms require further investigation through longitudinal or interventional studies. Analyses of symptom-paired samples collected prospectively from patients in whom uremic symptoms were the main indication for dialysis initiation, followed by a significant improvement in symptoms after dialysis, could also provide additional information.

In conclusion, this large multicenter study identified several metabolites that were significantly associated with anorexia and nausea/vomiting in individual cohorts, but the associations were not externally validated. Our study provides important recommendations for future investigations of the elusive mechanisms of uremic symptoms in kidney failure.

Supplementary Material

kidney360-7-353-s001.pdf (3MB, pdf)
Open in a new tab
kidney360-7-353-s002.pdf (1.7MB, pdf)
Open in a new tab
kidney360-7-353-s003.xlsx (2.2MB, xlsx)
Open in a new tab

Footnotes

S.A.A. and L.M. are co-first authors.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/KN9/B281.

Author Contributions

Conceptualization: Eliseo Guallar, Eugene P. Rhee, Tariq Shafi.

Data curation: Leslie Myint, Tariq Shafi.

Formal analysis: Clary B. Clish.

Funding acquisition: Eliseo Guallar, Eugene P. Rhee, Tariq Shafi.

Investigation: Solaf Al Awadhi, Clary B. Clish, Leslie Myint, Eugene P. Rhee, Tariq Shafi.

Methodology: Clary B. Clish, Leslie Myint, Eugene P. Rhee, Tariq Shafi.

Project administration: Eugene P. Rhee, Tariq Shafi.

Resources: Jonathan Himmelfarb, Sharon M. Moe, Eugene P. Rhee, Tariq Shafi, Marcello Tonelli.

Software: Leslie Myint.

Supervision: Eugene P. Rhee, Tariq Shafi.

Validation: Eugene P. Rhee.

Visualization: Solaf Al Awadhi, Leslie Myint.

Writing – original draft: Solaf Al Awadhi, Tariq Shafi.

Writing – review & editing: Solaf Al Awadhi, Clary B. Clish, Eliseo Guallar, Jonathan Himmelfarb, Sahir Kalim, Mara McAdams-DeMarco, Sharon M. Moe, Ranjani N. Moorthi, Leslie Myint, Neil R. Powe, Eugene P. Rhee, Dorry L. Segev, Tariq Shafi, Ravi Thandhani, Marcello Tonelli, Kendra E. Wulczyn.

Funding

T. Shafi: National Institute of Nursing Research (R01NR017399).

Data Availability Statements

Data related to transcriptomic, proteomic, or metabolomic data. Data Type: Analyzable Data; Clinical Trial Data. Repository Name: LUCID study and FAIR study.

Supplemental Material

This article contains the following supplemental material online at http://links.lww.com/KN9/B282, http://links.lww.com/KN9/B283.

Supplemental Table 1. KDQOL-36 uremic symptom description.

Supplemental Table 2. STrengthening the Reporting of OBservational studies in Epidemiology statement checklist for cohort studies.

Supplemental Table 3. Coefficients of variation for metabolomics data.

Supplemental Figure 1. PCA on the metabolite abundances before adjustment for instrument drift bias.

Supplemental Figure 2. PCA on the metabolite abundances after adjustment for instrument drift bias.

Supplemental Figure 3. Evaluation of imputation methods for missing metabolite data.

Supplemental Table 4. Imputation error for symptoms and covariates.

Supplemental Table 5. Literature review summary of studies using metabolomics to assess uremic symptoms.

Supplemental Methods

Supplemental References

Supplemental Results. Degree of association of all known and unknown metabolites with uremic symptom severity.

References

  • 1.Rhee EP Guallar E Hwang S, et al. Prevalence and persistence of uremic symptoms in incident dialysis patients. Kidney360. 2020;1(2):86–92. doi: 10.34067/KID.0000072019 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.You AS Kalantar SS Norris KC, et al. Dialysis symptom index burden and symptom clusters in a prospective cohort of dialysis patients. J Nephrol. 2022;35(5):1427–1436. doi: 10.1007/s40620-022-01313-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Satti MZ Arshad D Javed H, et al. Uremic pruritus: prevalence and impact on quality of life and depressive symptoms in hemodialysis patients. Cureus. 2019;11(7):e5178. doi: 10.7759/cureus.5178 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Shetty D Nayak AM Datta D, et al. Uremic pruritus: prevalence, determinants, and its impact on health-related quality of life and sleep in Indian patients undergoing hemodialysis. Ir J Med Sci. 2023;192(6):3109–3115. doi: 10.1007/s11845-023-03393-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Hosseini M Nasrabadi M Mollanoroozy E, et al. Relationship of sleep duration and sleep quality with health-related quality of life in patients on hemodialysis in Neyshabur. Sleep Med X. 2023;5:100064. doi: 10.1016/j.sleepx.2023.100064 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Depner TA. Uremic toxicity: urea and beyond. Semin Dial. 2001;14(4):246–251. doi: 10.1046/j.1525-139X.2001.00072.x [DOI] [PubMed] [Google Scholar]
  • 7.Ozcan H Yucel A Avşar UZ, et al. Kidney transplantation is superior to hemodialysis and peritoneal dialysis in terms of cognitive function, anxiety, and depression symptoms in chronic kidney disease. Transplant Proc. 2015;47(5):1348–1351. doi: 10.1016/j.transproceed.2015.04.032 [DOI] [PubMed] [Google Scholar]
  • 8.Azar SA, Hatefi R, Talebi M. Evaluation of effect of renal transplantation in treatment of restless legs syndrome. Transplant Proc. 2007;39(4):1132–1133. doi: 10.1016/j.transproceed.2007.03.097 [DOI] [PubMed] [Google Scholar]
  • 9.Molnar MZ Novak M Ambrus C, et al. Restless Legs Syndrome in patients after renal transplantation. Am J Kidney Dis. 2005;45(2):388–396. doi: 10.1053/j.ajkd.2004.10.007 [DOI] [PubMed] [Google Scholar]
  • 10.Kovacs AZ Molnar MZ Szeifert L, et al. Sleep disorders, depressive symptoms and health-related quality of life—a cross-sectional comparison between kidney transplant recipients and waitlisted patients on maintenance dialysis. Nephrol Dial Transplant. 2011;26(3):1058–1065. doi: 10.1093/ndt/gfq476 [DOI] [PubMed] [Google Scholar]
  • 11.Duranton F Cohen G De Smet R, et al. Normal and pathologic concentrations of uremic toxins. J Am Soc Nephrol. 2012;23(7):1258–1270. doi: 10.1681/ASN.2011121175 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Hu JR Myint L Levey AS, et al. A metabolomics approach identified toxins associated with uremic symptoms in advanced chronic kidney disease. Kidney Int. 2022;101(2):369–378. doi: 10.1016/j.kint.2021.10.035 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Wulczyn KE Shafi T Anderson A, et al. Metabolites associated with uremic symptoms in patients with CKD: findings from the chronic renal insufficiency cohort (CRIC) study. Am J Kidney Dis. 2024;84(1):49–61.e1. doi: 10.1053/j.ajkd.2023.11.013 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Wu Q Zhang H Ding JR, et al. UPLC-QTOF MS-based serum metabolomic profiling analysis reveals the molecular perturbations underlying uremic pruritus. Biomed Res Int. 2018;2018(1):4351674. doi: 10.1155/2018/4351674 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Kurella Tamura M Chertow GM Depner TA, et al. Metabolic profiling of impaired cognitive function in patients receiving dialysis. J Am Soc Nephrol. 2016;27(12):3780–3787. doi: 10.1681/ASN.2016010039 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Bolanos CG, Pham NM, Mair RD, Meyer TW, Sirich TL. Metabolomic analysis of uremic pruritus in patients on hemodialysis. PLoS One. 2021;16(2):e0246765. doi: 10.1371/journal.pone.0246765 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Taylor K Chu NM Chen X, et al. Kidney disease symptoms before and after kidney transplantation. Clin J Am Soc Nephrol. 2021;16(7):1083–1093. doi: 10.2215/CJN.19031220 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.von Elm E Altman DG Egger M, et al. The strengthening the reporting of observational studies in epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Med. 2007;4(10):e296. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Fernández-Albert F, Llorach R, Garcia-Aloy M, Ziyatdinov A, Andres-Lacueva C, Perera A. Intensity drift removal in LC/MS metabolomics by common variance compensation. Bioinformatics. 2014;30(20):2899–2905. doi: 10.1093/bioinformatics/btu423 [DOI] [PubMed] [Google Scholar]
  • 20.Märtens A, Holle J, Mollenhauer B, Wegner A, Kirwan J, Hiller K. Instrumental drift in untargeted metabolomics: optimizing data quality with intrastudy QC samples. Metabolites. 2023;13(5):665. doi: 10.3390/metabo13050665 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.De Livera AM Sysi-Aho M Jacob L, et al. Statistical methods for handling unwanted variation in metabolomics data. Anal Chem. 2015;87(7):3606–3615. doi: 10.1021/ac502439y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Cao W, Wang D, Li J, Zhou H, Li L, Li Y. BRITS: bidirectional recurrent imputation for time series. arXiv. 2018. doi: 10.48550/arXiv.1805.10572 [DOI] [Google Scholar]
  • 23.Wei R Wang J Su M, et al. Missing value imputation approach for mass spectrometry-based metabolomics data. Sci Rep. 2018;8(1):663. doi: 10.1038/s41598-017-19120-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Stekhoven DJ, Bühlmann P. MissForest—non-parametric missing value imputation for mixed-type data. Bioinformatics. 2012;28(1):112–118. doi: 10.1093/bioinformatics/btr597 [DOI] [PubMed] [Google Scholar]
  • 25.Kammers K, Cole RN, Tiengwe C, Ruczinski I. Detecting significant changes in protein abundance. EuPA Open Proteom. 2015;7:11–19. doi: 10.1016/j.euprot.2015.02.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Antonelli J Claggett BL Henglin M, et al. Statistical workflow for feature selection in human metabolomics data. Metabolites. 2019;9(7):143. doi: 10.3390/metabo9070143 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Ogutu JO, Schulz-Streeck T, Piepho HP. Genomic selection using regularized linear regression models: ridge regression, lasso, elastic net and their extensions. BMC Proc. 2012;6(suppl 2):S10. doi: 10.1186/1753-6561-6-S2-S10 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Smyth GK. Linear models and empirical bayes methods for assessing differential expression in microarray experiments. Stat Appl Genet Mol Biol. 2004;3:Article3. doi: 10.2202/1544-6115.1027 [DOI] [PubMed] [Google Scholar]
  • 29.Koehrsen W. Introduction to Bayesian Linear Regression. Medium; 2018. Accessed July 19, 2023. https://towardsdatascience.com/introduction-to-bayesian-linear-regression-e66e60791ea7 [Google Scholar]
  • 30.Haynes W. Benjamini–hochberg method. In: Dubitzky W, Wolkenhauer O, Cho KH, Yokota H, eds. Encyclopedia of Systems Biology. Springer; 2013:78. doi: 10.1007/978-1-4419-9863-7_1215 [DOI] [Google Scholar]
  • 31.Chen SY, Feng Z, Yi X. A general introduction to adjustment for multiple comparisons. J Thorac Dis. 2017;9(6):1725–1729. doi: 10.21037/jtd.2017.05.34 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.International Agency for Research on Cancer. Exposome-Explorer - Indolepropionic Acid (Compound). Accessed September 3, 2024. http://exposome-explorer.iarc.fr/compounds/1690 [Google Scholar]
  • 33.Sun CY Lin CJ Pan HC, et al. Clinical association between the metabolite of healthy gut microbiota, 3-indolepropionic acid and chronic kidney disease. Clin Nutr. 2019;38(6):2945–2948. doi: 10.1016/j.clnu.2018.11.029 [DOI] [PubMed] [Google Scholar]
  • 34.Jiang H, Chen C, Gao J. Extensive summary of the important roles of indole propionic acid, a gut microbial metabolite in host health and disease. Nutrients. 2022;15(1):151. doi: 10.3390/nu15010151 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Wang YC Koay YC Pan C, et al. Indole-3-Propionic acid protects against heart failure with preserved ejection fraction. Circ Res. 2024;134(4):371–389. doi: 10.1161/CIRCRESAHA.123.322381 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Vanholder R, Nigam SK, Burtey S, Glorieux G. What if not all metabolites from the uremic toxin generating pathways are toxic? A hypothesis. Toxins. 2022;14(3):221. doi: 10.3390/toxins14030221 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Mamer OA, Reimer ML. On the mechanisms of the formation of L-alloisoleucine and the 2-hydroxy-3-methylvaleric acid stereoisomers from L-isoleucine in maple syrup urine disease patients and in normal humans. J Biol Chem. 1992;267(31):22141–22147. doi: 10.1016/s0021-9258(18)41646-8 [DOI] [PubMed] [Google Scholar]
  • 38.Fontella FU Gassen E Pulrolnik V, et al. Stimulation of lipid peroxidation in vitro in rat brain by the metabolites accumulating in maple syrup urine disease. Metab Brain Dis. 2002;17(1):47–54. doi: 10.1023/a:1014004414733 [DOI] [PubMed] [Google Scholar]
  • 39.Wishart DS Guo A Oler E, et al. Hmdb 5.0: the human metabolome Database for 2022. Nucleic Acids Res. 2022;50(D1):D622–D631. doi: 10.1093/nar/gkab1062 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

Data related to transcriptomic, proteomic, or metabolomic data. Data Type: Analyzable Data; Clinical Trial Data. Repository Name: LUCID study and FAIR study.

Articles from Kidney360 are provided here courtesy of American Society of Nephrology

RESOURCES