Characterization of residual kidney function in chronic hemodialysis patients using plasma metabolomics

Abstract

In dialysis, preservation of residual kidney function (RKF) is associated with a favorable prognosis. We aimed to identify differences in plasma metabolites depending on the presence of RKF and to evaluate whether a subset of biomarkers could assist in classifying RKF. A prospective study was conducted among chronic hemodialysis (HD) patients in Geneva, Switzerland. Presence of significant RKF was defined as residual clearance of urea (KRU) ≥ 2 mL/min on 24 h urine collection. A total of 89 a priori selected plasma metabolites were analyzed by liquid chromatography-mass spectrometry. We included 136 patients, 54 with KRU ≥ 2 mL/min (40%) and 82 (60%) with KRU < 2 mL/min. The overall metabolic profile strongly differed between patients with and without significant RKF, despite similar total (residual + dialytic) clearances. Classification based on the metabolic profile achieved a 93% accuracy to predict significant RKF at a 2 mL/min cut-off. A subset of 3 biomarkers allowed RKF classification with 87% accuracy. RKF has a decisive influence on the overall metabolic profile of HD patients, which may partly explain its clinical benefit. This provides a physiological rationale for RKF preservation in incremental HD programs, while a limited subset of metabolites could prove useful in monitoring RKF in this setting.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-47357-1.

Introduction

The worldwide estimated number of patients on kidney replacement therapy (KRT) for end-stage kidney disease (ESKD) is close to 4 million, with the highest prevalence in high-income regions such as Europe and North America¹. With an overall increase in the incidence and prevalence of KRT over the last decade, hemodialysis (HD) remains the most common KRT modality concerning more than 80% of incident patients^2,3. In the setting of dialysis, residual kidney function (RKF) is defined as preserved production of a clinically significant urine output. It is observed that dialysis patients with preserved RKF have a more favorable prognosis than those who are anuric^4,5. The reason for this beneficial effect of RKF is not formerly understood but a significant urine output in the setting of dialysis offers several theoretical advantages. Mitigation of intra-dialytic weight gain and subsequent reduction in ultrafiltration rate is associated with improved survival⁶. Improved endogenous clearance of uremic toxins is another potential mechanism and higher RKF associates with lower serum beta2-microgobulin concentration, in turn corresponding to improved survival⁷. Moreover, RKF allows clearance of secreted solutes otherwise not readily removed by dialysis⁸. In the appropriate setting, a significant RKF could allow for the prescription of a less intensive dialysis regimen (i.e. incremental dialysis). As exposure to the dialysis procedure itself conveys sub-clinical target organ damage, RKF could indirectly improve patients’ prognosis by limiting their exposure to dialysis treatment^9–12.

Analysis of plasma metabolites using liquid chromatography coupled to mass spectrometry (LC–MS) has been extensively used in the field of chronic kidney disease (CKD)^13,14. Metabolomics is able to identify qualitative and quantitative differences in plasma metabolites profiles between non-dialysis CKD and HD patients¹⁵. Such differences could contribute to the dim prognosis of patients on dialysis as compared to those with milder stages of CKD. In the setting of ESKD however, differences in kidney function between patients with and without RKF are minimal and the extent to which preserved RKF contributes to their metabolic homeostasis is unknown. Consequently, we investigated in a prospective observational study whether patients undergoing chronic HD would present measurable differences in a priori selected plasma metabolites depending on the presence or absence of a clinically significant RKF. We secondarily wished to determine whether a limited subset of biomarkers could assist in determining the presence of significant RKF in individual patients.

Methods

Participants and study design

We conducted a prospective observational study at two distinct dialysis centers: the dialysis unit of a tertiary hospital (Geneva University Hospitals, Geneva, Switzerland) and a private dialysis center (“Quavitae Rive Gauche”, Geneva, Switzerland). We screened all incident and prevalent adult patients undergoing chronic HD with or without RKF between December 2022 to July 2024. Exclusion criteria were: (1) age < 18, (2) presence of AKI, (3) not willing to participate and (4) unable to provide written informed consent. Medical management and dialysis prescription were left to attending physician’s discretion. Both study centers have established routine incremental HD programs, where individual RKF is considered when prescribing dialysis dose and regimen. Patients were dialyzed using Fresenius 5008 (Fresenius MC, Bad Homburg, Germany) or Braun Dialog IQ (B. Braun Avitum, Melsungen, Germany) dialysis machines with polysulfone high-flux dialyzers. Patients received high-flux hemodialysis or online post-dilution hemodiafiltration with ultra-pure dialysis water. Plasma and urine samples were collected immediately prior to the mid-week dialysis session for patients on a thrice-weekly regimen or the second dialysis session of the week for patients on a twice-weekly regimen.

Variables definition

Presence of RKF was defined as a daily urine output ≥ 200 mL. In patients with RKF, residual clearance of urea (KRU) was calculated on a 24 h urine collection using the pre-HD serum urea concentration¹⁶. A KRU cut-off of 2 mL/min was used to further classify those patients. This cut-off was a priori specified and selected based on clinical relevance as well as reconnaissance in related guidelines¹⁷. Patients with daily urine output < 200 mL despite maximal diuretic therapy were considered anuric and were attributed a corresponding theoretical KRU of 0 mL/min. Total body water, approximating urea distribution volume (V), was calculated using Watson formula¹⁸. Residual KtV per week was calculated by multiplying KRU by (7 × 1440) (min) and dividing by V (mL). Dialytic single-pool KtV (spKtV) per session was calculated using the second generation equation developed by Daugirdas et al.¹⁹. Dialytic equilibrated KtV (eKtV) per session was calculated using the equation developed by Tatterall et al.²⁰. Dialytic standard KtV (stdKtV) per week was calculated using the equation developed by Leypoldt et al.²¹. Total (residual + dialytic) stdKtV per week was calculated by adding residual KtV per week to dialytic stdKtV per week.

Metabolites selection

A total of 89 metabolites were a priori selected based on two distinct criteria. First, we analyzed data from a previous pilot study, including 35 prevalent patients on chronic HD (unpublished data). Metabolites whose concentration markedly differed between patients with and without RKF were included in the present study. Second, metabolites known from prior scientific studies to be associated with RKF preservation or clinical prognosis in ESKD were also included in the present study. Data on the clinical relevance of those metabolites can be found in the related literature^22–25. Finally, those 89 metabolites were categorized in three groups of decreasing clinical relevance to prioritize data acquisition (Supplementary Table S1).

Data acquisition and analysis

Sample preparation and LC methods were adapted from Gagnebin et al.²⁶. Briefly, plasma samples were protein-precipitated using cold methanol containing internal standards, followed by centrifugation, drying, and storage at − 80 °C. Extracts were reconstituted in injection solvent, respectively, incubated at − 20 °C for 2 h, centrifugated, and diluted 20-fold before injection. A detailed protocol can be found in the supplementary material (Supplementary Methods S1). To address the wide range of analyte polarities, two complementary LC methods were employed. Hydrophobic metabolites were analyzed using reversed-phase chromatography (RP) on a Kinetex C18 column (150 × 2.1 mm, 1.7 µm; Phenomenex, California, USA), while polar compounds were assessed via hydrophilic interaction chromatography (HILIC) on a Acquity Premier BEH amide column (150 × 2.1 mm, 1.7 µm; Waters Corporation, Milford, MA, USA). Details are provided in the supplementary material (Supplementary Methods  S2). Targeted analysis was conducted on a Xevo TQ-XS triple quadrupole mass spectrometer (Waters Corporation, Milford, MA, USA) equipped with a ZSpray electrospray ion source, operated in polarity switching mode with scheduled multiple reaction monitoring (MRM). Two transitions were acquired per analyte, sourced from literature and databases such as the MassBank of North America (MoNA), and validated with standard and plasma injections. MS acquisition details are available in the supplementary material (Supplementary Table S2). Internal standards were spiked at the protein precipitation step of the protocol to control variability. Batch effect was corrected with LOESS correction and PQN normalization, using an in-house script²⁷. Principal Component Analysis (PCA) was performed using SIMCA software to assess batch correction. The sequence validity parameters can be found in the supplementary material (Supplementary Table S3 and Supplementary Fig. S1).

Statistical analysis

Between groups comparison was conducted using Student’s t-test and Chi-square test for continuous and categorical variables respectively. Missing values were not imputed. A two-sided p-value < 0.05 was considered significant. Descriptive statistical analyses were conducted using STATA software version 17 (StataCorp, 4905 Lakeway Drive, College Station, Texas 77845 USA).

Multivariate modeling included orthogonal partial least square discriminant analysis (OPLS-DA), Monte Carlo uninformative variable elimination (MC-UVE), elastic net regression, support vector machine and neural networks. Specification of those models are given in Supplementary Methods S3. Paired prediction comparison was carried out using McNemar’s test.

Ethics

This study was approved by the local ethics committee “Commission cantonale d’éthique de la recherche” (CCER) (2022-01398), Geneva, Switzerland and performed according to the declaration of Helsinki. Informed consent in written form was obtained for all patients included in this study.

Results

In total, 136 patients were included in the study. Mean age was 61.7 ± 16.5, with 82 (60%) men. Among those 136 patients, 58 (43%) were anuric with a theoretical KRU of 0 mL/min, 24 (18%) had RKF with KRU < 2 mL/min and 54 (40%) had RKF with KRU ≥ 2 mL/min. Consequently, 82 patients had KRU < 2 mL/min (Supplementary Fig. S2). Clinical characteristics of patients according to a threshold of KRU = 2 mL/min are given in Table 1. By definition, patients with KRU ≥ 2 mL/min had higher KRU and urine volume as compared to those with KRU < 2 mL/min, while other clinical parameters were similar. Regarding dialysis parameters, patients with KRU ≥ 2 mL/min were less frequently treated with HDF (as opposed to HD), had shorter dialysis vintage, were more frequently on a 2×/week dialysis regimen and had longer interdialytic interval. Regarding adequacy parameters, patients with KRU ≥ 2 mL/min had lower dialytic KtV per session and lower dialytic KtV per week but similar total (residual + dialytic) KtV per week as compared to those with KRU < 2 mL/min.

Table 1.

Characteristics of patients according to a threshold of KRU = 2 mL/min (N = 136).

	KRU < 2 mL/min (N = 82)	KRU ≥ 2 mL/min (N = 54)	P value
Clinical parameters
Age (years)	62.3 ± 16.9	60.7 ± 16.2	0.594
Gender (men)	51 (62%)	31 (57%)	0.577
Smoker	17 (32%)	14 (30%)	0.805
Diabetes	24 (29%)	14 (26%)	0.671
Hypertension	64 (79%)	42 (78%)	0.864
ESKD etiology
Diabetes or hypertension	38 (46%)	26 (48%)	0.689
Glomerulonephritis	16 (20%)	7 (13%)
ADPKD	9 (11%)	5 (9%)
Other	19 (23%)	16 (30%)
Charlson score	5.9 ± 3.0	5.8 ± 2.8	0.447
Daily urine volume (mL)	138 ± 270	1614 ± 602	< 0.001
KRUa (mL/min)	0 (0–0.39)	3.46 (2.79–4.90)	< 0.001
Dialysis parameters
HDF (vs HD)	54 (68%)	16 (30%)	< 0.001
AVF (vs KT)	41 (51%)	20 (37%)	0.105
Dialysis vintage (months)	49 (18–84)	5 (1–14)	< 0.001
Dialysis regimen
2x/week	2 (3%)	40 (74%)	< 0.001
3x/week	76 (95%)	14 (26%)
> 3x/week	2 (3%)	0
Interdialytic days	2 (2–2)	3 (2–4)	< 0.001
Adequacy parameters
Residual KtV per weeka	0 (0–0.11)	1.03 (0.79–1.35)	< 0.001
Dialytic KtV per sessiona	1.55 (1.38–1.93)	1.13 (0.85–1.47)	< 0.001
Dialytic KtV per weeka	2.36 (2.19–2.62)	1.31 ( 1.09–1.62)	< 0.001
Total (residual + dialytic) KtV per weeka	2.37 (2.23–2.68)	2.40 (2.18–2.95)	0.164

KRU, residual clearance of urea; ESKD, end-stage kidney disease; ADPKD, autosomal dominant polycystic disease; HDF, hemodiafiltration; HD, hemodialysis; AVF, arteriovenous fistula; KT, catheter.

^aSee text for methodology regarding those parameters.

Bold values indicate p < 0.05.

From 89 a priori selected metabolites, 57 (64%) were detected in all plasma samples from the 136 patients included in the study (Supplementary Table S4). The final list of 57 identified metabolites is provided in Supplementary Table S5. PCA shows grouping of the QCs samples across the three analyzed batches, and the IS RSDs were under 20% for the three batches in the two LC platforms (Supplementary Fig. S1). A graphical representation illustrating the log fold change in relative mean abundances of the 57 metabolites between KRU ≥ 2 mL/min and KRU < 2 mL/min is provided in Fig. 1. Overall, KRU ≥ 2 mL/min was associated with higher levels of serotonin, deoxycholic acid, carnitine, cortisone, p-cresyl sulfate and tryptophan (p < 0.05 for all). Conversely, KRU ≥ 2 mL/min was associated with lower relative levels of 4-pyridoxic acid, citramalic acid, hippuric acid, 5-hydroxyindolacetic acid (5-HIAA), kynurenic acid, riboflavin, indoxyl sulfate, 4-hydroxybenzoic acid and phenylacetylglutamine (p < 0.05 for all). Levels of other identified metabolites were similar between KRU ≥ 2 mL/min and KRU < 2 mL/min.

Fig. 1.

Log fold change in relative mean levels of the 57 metabolites between KRU ≥ 2 mL/min and KRU < 2 mL/min (N = 136). A log scale was used to allow direct comparison of symmetrical quantitative effect in both directions. A negative fold change indicates a lower relative abundance in patients with KRU ≥ 2 mL/min. A positive fold change indicates a higher relative abundance in patients with KRU > 2 mL/min.

Supervised learning for KRU classification

In a first set of analyses, supervised learning (OPLS-DA) was applied to characterize the contribution of metabolites to the classification of KRU status at a cut-off of 2 mL/min.

To control for the inherent inter-individual variability in metabolites concentration and normalize results across the study population, 1596 unique ratios of paired metabolites were computed by combinatorial calculation of the 57 identified metabolites. OPLS-DA score plot for the 136 patients and loading plot for the 1596 ratios of paired metabolites for KRU classification at a cut-off of 2 mL/min are presented in Supplementary Fig. S3a and b respectively. R²Y was 0.68 and leave-one-out cross-validation Q²Y was 0.45. The overall accuracy to classify KRU at a cut-off of 2 mL/min was 82% with the matrix of confusion provided in Supplementary Table S6.

To increase model robustness, uninformative variable elimination (UVE) was applied. This process reduced the number of unique ratios of paired metabolites included in the model from 1596 to 140. Among the 57 identified metabolites, 48 putative biomarkers appeared in those 140 ratios of paired metabolites while 9 metabolites did not appear in any ratio. The relative occurrence of those 48 metabolites in the 140 ratios is illustrated in Supplementary Fig. S4. A second OPLS-DA model was then computed based on the selected 140 ratios for KRU classification at a cut-off of 2 mL/min. Score plot for the 136 patients and loading plot for the 140 ratios of paired metabolites are presented in Fig. 2a and b respectively. R²Y was 0.72 and leave-one-out cross-validation Q²Y was 0.66. The overall accuracy to classify KRU at a cut-off of 2 mL/min was 93% with the matrix of confusion provided in Table 2. A volcano plot illustrating the contribution of each ratio of paired metabolites to KRU classification at a cut-off of 2 mL/min is provided in Fig. 3. The relative abundances of the three metabolites with the most occurrences in those 140 ratios according to a threshold of KRU = 2 mL/min are presented in Fig. 4. Levels of 5-HIAA, hippuric acid and citramalic acid were lower in patients with KRU > 2 mL/min as compared to those with KRU < 2 mL/min (p < 0.001 for all).

Fig. 2.

OPLS-DA model for KRU classification at a cut-off of 2 mL/min. (a) Score plot for patients (N = 136). (b) Loading plot for ratios of paired metabolites (N = 140).

Table 2.

Performances of OPLS-DA model for KRU classification at a cut-off of 2 mL/min based on 140 ratios of paired metabolites (N = 136).

	Predicted KRU < 2 mL/min	Predicted KRU ≥ 2 mL/min
True KRU < 2 mL/min	80 (92%)	2 (4%)
True KRU ≥ 2 mL/min	7 (8%)	47 (96%)

KRU, residual clearance of urea.

Percentages are given in columns.

Overall accuracy is 93%

Fig. 3.

Volcano plot illustrating the contribution of each ratio of paired metabolites to KRU classification at a cut-off of 2 mL/min (N = 136). A negative fold change indicates a lower value in patients with KRU ≥ 2 mL/min. A positive fold change indicates a higher value in patients with KRU ≥ 2 mL/min.

Fig. 4.

Relative abundance of selected metabolites in patients with KRU above or below 2 mL/min (N = 136). Metabolites were selected based on the highest occurrences in ratios of paired metabolites.

Parsimonious linear models for KRU prediction

In a second set of analyses, sparse linear modeling was applied to evaluate the performance of parsimonious subsets of metabolites to predict KRU at a cut-off of 2 mL/min.

A regression model with elastic net regularization was applied to the 140 ratios of paired metabolites. A population of models was investigated allowing alpha (α) and lambda (λ) parameters to vary from 0 to 1. In this setting, the magnitude of the regularization (λ) indirectly determines the number of ratios (and thus of biomarkers) included in the model. A visual representation of the accuracy of this model to predict KRU at a cut-off of 2 mL/min is provided in Fig. 5. The overall accuracy of the model was relatively independent of the magnitude of the regularization (λ), indicating that a limited number of metabolites may be sufficient for reliable prediction. This was further illustrated in Fig. 6 representing the accuracy of the model to predict KRU at a cut-off of 2 mL/min according to the number of metabolites included in the model. Inclusion of a larger number of metabolites only marginally improved predictive accuracy while inclusion of a minimal number of metabolites allowed to retain satisfactory classification performance. Specifically, a strongly regularized model including only 3 metabolites (2,3-dihydroxybenzoic acid, citramalic acid and tryptophan) allowed an 85% overall accuracy to predict KRU at a cut-off of 2 mL/min. A McNemar’s test comparing paired predictions between a 3-metabolite model (85% accuracy) and a 29-metabolite model (92% accuracy) yielded a p-value of 0.026, indicating that parsimony comes at the cost of a slight decrease in performance.

Fig. 5.

Accuracy of considered metabolites to predict KRU at a cut-off of 2 mL/min using a regression model with elastic net regularization (N = 136). Alpha (α) parameter indicated the balance between lasso and ridge penalties with α = 0 corresponding to pure ridge regression and α = 1 corresponding to pure lasso regression. Lambda (λ) parameter indicated the magnitude of regularization with λ = 0 corresponding to no regularization and λ = 1 corresponding to strong regularization.

Fig. 6.

Accuracy of considered metabolites to predict KRU at a cut-off of 2 mL/min according to the number of metabolites included in the model, as indirectly determined by the magnitude of regularization (λ) (N = 136). Illustrative points were empirically selected with X indicating the number of metabolites included in the model and Y the overall accuracy of the model (%).

Parsimonious non-linear models for KRU prediction

In a third set of analyses, the parsimonious subset of 3 relevant biomarkers obtained with the previous regression model (2,3-dihydroxybenzoic acid, citramalic acid and tryptophan) was further evaluated to predict KRU at a cut-off of 2 mL/min using non-linear models.

Various models, including logistic regression, random forest, naïve bayes, gradient boosting and k-nearest neighbors, offered relatively similar performance overall, while support vector machine and neural network algorithms provided the highest ROC AUC at 0.84 and 0.88 respectively (Supplementary Fig. S5). The overall accuracy of support vector machine and neural network algorithms to classify KRU at a cut-off of 2 mL/min was 87%. The matrix of confusion for those models are provided in Supplementary Tables S7 and S8.

Discussion

In this prospective study, we used plasma metabolomics to characterize RKF in chronic dialysis patients. We observed major differences in metabolites profiles between patients with KRU above or below 2 mL/min despite similar total (residual + dialytic) clearances. Furthermore, a very limited subset of biomarkers proved sufficient to accurately distinguish patients with or without significant RKF. Those findings highlight that RKF in dialysis patients plays a significant role in metabolic homeostasis and its characterization may inform a variety of clinical questions.

When considering metabolites individually, we found significant differences in relative abundance depending on RKF. We observed that hippuric acid, kynurenic acid and indoxyl sulfate, that are all recognized as protein bound uremic toxins (PBUT), were present in lower abundance in patients with KRU ≥ 2 mL/min. Those substances are derived from the intestinal metabolism of amino acids. Dysfunction of the intestinal microbiota as well as decreased active tubular secretion result in PBUT accumulation in ESKD²⁸. Clearance of those toxins with traditional HD techniques is notably limited even when using convective clearance²⁹. The gut–kidney axis has recently emerged as an important mediator of uremic toxin accumulation. Beyond the established role of the gut bacteriome, the gut mycobiome also contributes to the host metabolome in ESKD. Gut fungal dysbiosis in ESKD patients—characterized by an increased abundance of opportunistic pathogens—explained over 10% of the variance in the serum metabolome³⁰. Importantly, fungi that are enriched in ESKD patients were also positively correlated with serum levels of creatinine, homocysteine and phenylacetylglycine, a gut-derived uremic solute. These findings expand the microbial contribution to uremic toxin accumulation beyond bacteria and suggest that the influence of RKF on uremic toxin levels observed in our study may partly be mediated through modulation of the intestinal fungal microbiota. Our findings confirm that levels of most PBUT are significantly lower in patients with preserved RKF despite a lower dialysis dose as well as a lower convective clearance. A causal influence on patients’ prognosis is possible as PBUT could induce endothelial dysfunction and vascular calcification in experimental models^31,32. Surprisingly, in contrast to other PBUT, we found that p-cresyl sulfate was present in higher abundance in patients with KRU ≥ 2 mL/min. This result is in agreement with prior reports based on the Hemodialysis (HEMO) trial illustrating that p-cresyl sulfate was indeed found in higher abundance in patients with preserved RKF³³.

Tryptophan metabolism is profoundly affected by CKD. Synthesis of kynurenine from tryptophan is mediated by the indoleamine dioxygenase 1 (IDO1) enzyme and represents the rate-limiting step of the kynurenine pathway. With increasing CKD severity, tryptophan levels decrease while IDO1 activity and levels of tryptophan catabolites of the kynurenine pathway increase^15,34. Low tryptophan levels and activation of the kynurenine pathway is thought to contribute to systemic inflammation, subclinical atherosclerosis and subsequent cardiovascular diseases^35–37. A recent longitudinal study including 184 patients with CKD reported that low tryptophan levels were associated with incident cardiovascular disease³⁸. In the present study, we observed that tryptophan levels were significantly higher in patient with preserved RKF. Of note, we and others have previously showed that tryptophan abundance and IDO1 activity were not significantly affected by dialysis^15,34. Our present results therefore extend the importance of tryptophan metabolism on cardiovascular prognosis to dialysis patients with very low levels of renal function. In both membranous nephropathy and CKD, depletion of specific Lactobacillus species, particularly Lactobacillus johnsonii, has been linked to an unfavorable shift in tryptophan-derived indole metabolites and activation of intrarenal aryl hydrocarbon receptor (AhR) signaling. The higher tryptophan levels observed in HD patients with RKF in our cohort may therefore reflect a more favourable balance within this tryptophan–indole–AhR axis, which could contribute to the prognostic advantage of RKF^39,40.

Beyond individual metabolites, we analyzed the impact of RKF preservation on the overall metabolic profile of ESKD patients on dialysis. An inherently high inter-individual variability in metabolites concentrations was expected and combinatorial ratios of metabolites were used as a natural normalization method. As compared to patients without significant RKF, those with preserved RKF had a very different metabolic profile, despite virtually identical total (residual + dialytic) clearances. Consequently, the presence or absence of a significant RKF could be reliably determined by observing the overall metabolic profile alone. Classification of KRU at a cut-off of 2 mL/min was achieved with an overall accuracy of 93%. More importantly, positive and negative predictive values to classify KRU ≥ 2 mL/min and < 2 mL/min were 96% and 92% respectively. Interestingly, this performance could be relatively maintained when considering only a very restricted subset of metabolites. Specifically, an 87% overall accuracy for KRU classification at a cut-off of 2 mL/min was achieved using non-linear methods based on 3 metabolites only, with corresponding positive and negative predictive values of 88% and 86% respectively. Interestingly, several metabolites emerging as key discriminants of RKF in our study, such as hippuric acid, indoxyl sulfate and kynurenic acid, also feature prominently in recent metabolomic studies of DKD and CKD‑associated secondary hyperparathyroidism^41,42. This convergence across different CKD phenotypes supports the notion that RKF preservation modulates core uremic toxin pathways rather than disease‑specific signatures.

The general clinical implications of our findings are manifold. First, we illustrate that the presence of RKF allows for maintenance of a certain metabolic profile that cannot be replicated by a more intensive dialysis schedule in anuric patients. Consequently, the clinical benefit of RKF preservation demonstrated in observational studies is likely to extend beyond the mere handling of volume status or the clearance of small molecules such as urea and creatinine. This also tends to reinforce incremental approaches to dialysis prescription, where dialysis dose is tailored to individual patient needs accounting for RKF^12,43. In the setting of incremental dialysis, it is usually accepted that a certain amount of RKF confers a greater clinical benefit than the same amount of dialytic clearance, so that every 1 mL/min of RKF lost is replaced by a higher dose (e.g. 1.5 mL/min) of dialytic clearance^43,44. Our findings highlight the potential physiological reasons underlying this practice. Second, we illustrate that an apparently negligible endogenous clearance is in fact associated with easily measurable differences in levels of various metabolites. Prior reports have illustrated that minimal RKF was still associated with measurable variation in the concentration of certain individual uremic toxins⁴⁵. Our findings extend this concept to a broader range of uremic toxins and in patients undergoing various dialytic regimen. Finally, we show that relative abundance of a very limited subset of biomarkers was able to reliably characterize the presence or absence of significant RKF. Interestingly, urea and creatinine were not among this subset, with poor discriminative capacities. Prescription of an incremental dialysis regimen requires repeated 24 h urine collections to verify that RKF is sufficient to maintain a reduced dialysis dose. Those urine collections are burdensome and prone to errors⁴⁶. Various groups have tried to rely on endogenous serum biomarkers to approximate RKF without requiring a urine sample^47–49. Our present results indicate that a combination of less commonly used biomarkers might prove useful in dichotomous identification of individual patients with an endogenous KRU above or below 2 mL/min. Such classification could assist the clinician in the follow-up of an incremental dialysis prescription.

Readers should bear in mind limitations of our study. First, present data are cross-sectional in nature. Thus, dynamic evolution of metabolic profiles over time and formal characterization of intra-individual variability could not be described. Likewise, potential relationship with later clinical outcomes could not be established. Second, spent dialysate was not analyzed in this study. Consequently, the effect of the dialytic procedure on the clearance of considered metabolites could not be directly measured. Our results should be seen as an overview of the metabolic state of chronic dialysis patients with different levels of RKF and corresponding appropriate dialysis doses. Third, some metabolites investigated could be influenced by dietary intake, nutritional status and gut microbiota, which were not fully characterized in this study. However, the consistent differences in metabolic pattern according to RKF and the a priori clinically driven metabolites selection suggest that our findings are less likely to be explained by exogenous factors. Finally, our metabolomic platform relied on relative quantification of identified metabolites. Absolute concentrations could therefore not be measured preventing establishment of externally valid reference values. Simultaneous absolute quantification of dozens of metabolites is currently however not realistic for routine use, although novel strategies might arise in the future^50,51. Moreover, as our study was designed to test clinically driven hypotheses grounded in biological plausibility, we deliberately focused on a predefined set of metabolites, which theoretically may have introduced selection bias.

Conclusion

Using a multicentric prospective cohort, we show that preserved RKF has a decisive influence on the overall metabolic profile of HD patients treated with a similar total (residual + dialytic) clearance. These metabolic differences could partly explain the clinical benefit of RKF preservation observed in longitudinal studies. In this context, the influence of RKF on protein-bound uremic toxins and tryptophan metabolic pathways appears pivotal. Overall, our results provide a physiological rationale for the preservation of RKF in dialysis patients. The current development of incremental hemodialysis programs will certainly benefit from these results. We also show that measurement of a limited number of biomarkers could prove useful in quantitative assessment of RKF at the individual level. Conceptually, regular measurement of these metabolites might reduce reliance on repeated urine collections in incremental dialysis programs, pending external validation and assessment of feasibility in various settings.

Author contributions

DAJ designed the study, recruited the patients, analyzed the data, interpreted the results and wrote the manuscript. JB, OS, MAN and GV acquired the data and revised the manuscript. VJ, NM, PS and SDS recruited the patients, interpreted the results and revised the manuscript. SR and BP designed the study, supervised the study and revised the manuscript.

Funding

This study was founded by an academic research & development (PRD) grant (13–2022-II) from Geneva University Hospitals, Geneva, Switzerland. BP was partly funded by the Marie-Heim Votglin grant of the Swiss National Science Foundation (PMPDP3_171352 & PMPDP3_186203).

Data availability

The data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.