AI-assisted metabolomic profiling identifies candidate metabolites associated with diabetic kidney disease staging in a Korean cohort

Summary

Diabetic kidney disease (DKD) is commonly staged using albumin-to-creatinine ratio (ACR) and estimated glomerular filtration rate (eGFR), yet complementary molecular markers are needed. We profiled urine and serum metabolites from 92 Korean participants (72 with type 2 diabetes and 20 healthy controls) using gas chromatography-tandem mass spectrometry (GC-MS/MS) and liquid chromatography-tandem mass spectrometry (LC-MS/MS). An exploratory restricted Boltzmann machine framework compared five DKD staging criteria and prioritized candidate metabolites; robustness was assessed against LASSO, linear support vector machine (SVM), and random forest using nested cross-validation. ACR-based staging showed the highest metabolomic discrimination. Urinary adenosine and 5′-methylthioadenosine decreased, whereas serum N2,N2-dimethylguanosine and cis-aconitic acid increased across ACR stages. Integration with a public renal tubular microarray dataset suggested NT5E as a cross-study network hub. These cross-sectional findings define candidate metabolite signatures associated with DKD severity and require external longitudinal validation before clinical translation.

Graphical abstract

Highlights

• •ACR stages show the highest metabolomic discrimination in this cohort
• •Urinary adenosine and 5′-MTA decrease across higher ACR stages
• •Serum M22G and cis-aconitic acid increase across higher ACR stages
• •Cross-study network analysis identifies NT5E as a hypothesis-generating hub

Computational bioinformatics; Metabolomics

Introduction

Diabetic kidney disease (DKD) is the leading cause of end-stage renal disease (ESRD), and South Korea has the highest average annual increase in the incidence of treated ESRD attributable to diabetes.^1,2 The progression of DKD to ESRD contributes to increased mortality and substantial healthcare costs.³ Currently, DKD is diagnosed and staged based on the albumin-to-creatinine ratio (ACR) and estimated glomerular filtration rate (eGFR), according to the Kidney Disease: Improving Global Outcomes (KDIGO) chronic kidney disease (CKD) heatmap classification.⁴ However, ACR and eGFR have limitations in capturing early pathological changes and risk heterogeneity across individuals. Omics-based metabolomics provides complementary molecular signatures associated with DKD severity. In small to moderate-sized cohorts, machine-learning approaches should be interpreted as exploratory tools for pattern discovery and candidate prioritization rather than as deployable diagnostic or prognostic models.

Hence, there is an urgent need to identify biomarkers that can complement ACR and eGFR for DKD risk stratification. Omics-based approaches, including metabolomics, may capture disease-associated metabolic alterations. However, in cross-sectional studies, such signatures should be interpreted as associations with disease severity rather than validated predictors of progression. Therefore, there is growing interest in applying computational approaches to metabolomic data to characterize metabolic patterns associated with DKD.⁵

This study aimed to analyze the serum and urine metabolomic profiles of healthy individuals and patients with type 2 diabetes mellitus (T2DM) to determine which of the currently used combinations of ACR and eGFR best discriminates metabolomic alterations. This study also aimed to characterize serum and urine metabolomic profiles across commonly used DKD staging criteria and to identify stage-associated metabolite signatures in an exploratory manner. We further performed network-based integration with publicly available transcriptomic data to provide a hypothesis-generating molecular context for the observed metabolite patterns.

Results

Baseline clinical characteristics

The clinical characteristics of the selected participants are presented in Table S1. Compared to the healthy controls (HCs), participants with T2DM exhibited significantly higher BMI (25.2 vs. 23.2 kg/m², p = 0.019), systolic blood pressure (137.7 vs. 124.7 mmHg, p = 0.002), HbA1c, and fasting glucose levels (both p < 0.0001). Participants with T2DM also showed significantly lower eGFR (68.1 vs. 92.2 mL/min/1.73 m², p < 0.0001) and markedly elevated urine ACR levels (102.2 vs. 3.9 mg/g, p < 0.0001), with a greater proportion classified in advanced CKD stages and higher albuminuria categories (both p < 0.001). The mean duration of diabetes in the T2DM group was 15.7 ± 8.3 years. Use of antihypertensive and lipid-lowering medications was significantly more common among participants with T2DM (both p < 0.0001), reflecting the presence of comorbid cardiometabolic risk factors. These systematic differences (including medication use and comorbidities) may confound metabolite levels and were therefore considered when interpreting group separation, particularly in early-stage comparisons.

AI-based evaluation of DKD staging criteria via metabolomic profiling

We performed an exploratory comparison of five commonly used DKD staging criteria to assess which criterion best aligns with metabolic discrimination in this cohort (Figures 1A and 1B). We evaluated five distinct clinical diagnostic criteria: criterion 1 used ACR-based staging alone (S1: <30, S2: 30–300, S3: ≥300 mg/g); criterion 2 applied the complete KDIGO 2012 CKD classification⁶; criterion 3 grouped participants into HCs, CKD G1+G2, and G3a+G3b; criterion 4 further simplified into HCs, G1+G2, and G3a+G3b groups; and criterion 5 categorized participants by T2DM status and albuminuria levels (normo-, micro-, and macro-albuminuria). Detailed descriptions of each criterion are provided in the STAR Methods. Across 100 repeated training runs per criterion, we summarized performance distributions for descriptive stability assessment (Figure S1). Among the five criteria, Criterion 1, based only on the ACR, demonstrated the highest median performance for both the urine and serum metabolite-based models. Then, we performed model benchmarking to compare ACR-based DKD staging using a restricted Boltzmann machine (RBM) alongside baseline models (least absolute shrinkage and selection operator [LASSO], linear support vector machine [SVM], and random forest). Model performance was evaluated using nested cross-validation to minimize optimistic bias and information leakage, and summarized using macro-average AUROC (primary) (Figure 1C). Accuracy, F1 (macro), precision (macro), recall (macro), AUROC (macro), and learning curves of RBM and baseline models are presented in Figure S2. Across both urine and serum data, RBM demonstrated performance comparable to baseline models under the same nested cross-validation framework. Therefore, we retained RBM primarily as an exploratory representation-learning approach coupled with explainable AI (integrated gradients) to prioritize biologically interpretable metabolite candidates, rather than as a definitive best-performing classifier. Using this grouping, in which patients were categorized into three stages (S1, S2, and S3, corresponding to the KDIGO albuminuria stages A1, A2, and A3, respectively), we applied explainable AI techniques to identify the top 20 metabolite signatures that best explained each stage. These stage-specific metabolites were then compared to identify unique and shared metabolites across different stages (Figure 1D). Venn diagrams revealed the unique and shared metabolites across these stages. To better distinguish between disease stages, we focused on identifying common metabolites that showed consistent expression patterns across multiple stages. Feature selection using explainable AI identified the top biomarker candidates for both urine and serum samples. In urine, 4-hydroxyphenylpyruvic acid (4-HPPA) showed the highest attribute consistency in S3, followed by glutaric acid, and 5′-methylthioadenosine (5′-MTA) and adenosine were also identified as important features (Figure 1E). In the serum, indoleacetic acid exhibited the strongest signal, and N2, N2-dimethylguanosine (M22G) and cis-aconitate, along with other metabolites, were identified as being associated with the S3 stage (Figure 1F). This analysis yielded 9 key urine metabolites and 13 key serum metabolites as potential markers for DKD severity.

Figure 1.

AI-based metabolomic profiling approach for candidate marker discovery for DKD severity

(A) Study workflow showing the three major steps: (1) sample collection and profiling using urine and serum samples from three groups: healthy controls (HCs), patients with T2DM with mild/no DKD, and patients with T2DM with severe DKD; (2) AI-based approach developing multiple AI models based on different clinical diagnostic criteria with iterative training; and (3) model evaluation and benchmarking using nested cross-validation to compare staging criteria and to assess robustness against baseline models.

(B) Schematic representation of the restricted Boltzmann machine (RBM) neural network architecture used for metabolite feature selection. The algorithm analyzed metabolite profiles (v) through hidden layers (h) to predict patient diagnoses (y) based on five different clinical diagnostic criteria.

(C) Nested cross-validation macro-averaged ROC curves comparing RBM with baseline models (LASSO, linear SVM, and random forest) for urine (left) and serum (right) metabolite data. Macro-AUROC values are shown in parentheses for each model.

(D) Venn diagrams illustrating the overlap of metabolic signatures identified in different DKD stages (S1, S2, and S3) grouped by albumin-to-creatinine ratio for urine metabolites (left) and serum metabolites (right). Numbers indicate the count of metabolites unique to each stage or shared between stages.

(E and F) Bar charts showing biomarker candidates in (E) urine metabolites and (F) serum metabolites ranked by attribute consistency in the severe DKD stage (S3). The x axis represents convergence delta (–log10) values, indicating the strength of association with the disease state.

AI, artificial intelligence; DKD, diabetic kidney disease; T2DM, type 2 diabetes mellitus. See also Figures S1 and S2.

Differential levels of metabolite marker candidates across DKD stages

Hierarchical clustering of the selected metabolites revealed distinct expression patterns across patient samples (Figures 2A and 2B). In urine samples (Figure 2A), we observed a clear separation of metabolite expression profiles between different DKD stages, with decreases in adenosine and 5′-MTA levels in individuals with macro-albuminuria (S3) compared to those with micro-albuminuria (S2) or normo-albuminuria/HC (S1). By contrast, serum samples (Figure 2B) showed an increased expression of M22G and cis-aconitic acid in individuals with macro-albuminuria (S3).

Figure 2.

Heatmap visualization of differential metabolite levels and pathway enrichment

(A) Hierarchical clustering heatmap of top urine metabolite biomarker candidates across patient samples. Rows represent metabolites (5′-MTA, adenosine, pyruvic acid, acetoacetate, l-tryptophan, glutaric acid, maleic acid, 4-HPPA, and pipecolic acid), and columns represent individual patient samples color-coded by DKD stage (S1, S2, and S3). Red indicates a higher level, whereas blue indicates a lower level.

(B) Hierarchical clustering heatmap of top serum metabolite biomarker candidates across patient samples. Rows represent metabolites (indoleacetic acid, indolelactic acid, oxalacetic acid, M22G, cis-aconitate, phenylpyruvic acid, p-hydroxymandelic acid, hydroxyphenyllactic acid, capric acid, dodecanoic acid, uridine, l-asparagine, and suberic acid).

(C and D) Pathway enrichment analysis tables showing (C) top significantly altered pathways in urine metabolites and (D) serum metabolites. Tables display pathway names, statistical significance (–log(P)), and pathway impact scores. Tyrosine metabolism was the most significantly altered pathway in urine samples, whereas the TCA cycle was most significant in serum.

DKD, diabetic kidney disease; TCA, tricarboxylic acid cycle; 5′-MTA, 5′-methylthioadenosine; M22G, N2,N2-dimethylguanosine; 4-HPPA, 4-hydroxyphenylpyruvic acid.

We performed pathway analysis on the 9 urine metabolites and 13 serum metabolites identified as potential markers for DKD severity (Figures 1E and 1F) to determine the metabolic pathways most significantly associated with the DKD stage (Figures 2C and 2D). Among urine metabolites, tyrosine metabolism was the most significantly altered pathway (–log(P) = 3.22), primarily driven by the alteration of urine 4-HPPA (Figure 1E). This was followed by the cysteine and methionine metabolic pathway (–log(P) = 2.06) (Figure 2C). For serum metabolites, the tricarboxylic acid cycle was the most significantly affected pathway (–log(P) = 2.37), followed by alanine, aspartate, and glutamate metabolism (–log(P) = 2.08) (Figure 2D). These findings suggest distinct patterns of metabolic dysregulation in the urine and serum during DKD progression.

Metabolite signature selection based on ACR staging and their clinical correlations

We report metabolite differences across ACR stages as a clinically interpretable reference framework. Statistical validation of the most promising biomarker candidates was performed using one-way analysis of variance (ANOVA) with post hoc comparisons between DKD stages. Both urine markers, adenosine and 5′-MTA, showed significant decreases with DKD progression (p < 0.0001), with the most pronounced differences between S1 and S3 (Figure 3A). Conversely, the serum markers M22G and cis-aconitic acid significantly increased with disease progression (p < 0.0001) (Figure 3B).

Figure 3.

Expression patterns of candidate metabolite markers and their clinical correlations

(A and B) Boxplots showing expression levels of (A) urine metabolites (adenosine and 5′-MTA) and (B) serum metabolites (M22G and cis-aconitic acid) across DKD stages (S1, S2, and S3). In the one-way analysis of variance (ANOVA), p values indicate significant differences between groups, with asterisks denoting significance levels between specific groups (∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001).

(C and D) Correlation plots showing relationships between (C) urine marker candidates (adenosine and 5′-MTA) and (D) serum marker candidates (M22G and cis-aconitic acid) with key clinical parameters: age, eGFR, HbA1c, HDL, hemoglobin, and white blood cell count. R values indicate correlation strength, with asterisks denoting significance levels. Gray shading represents 95% confidence intervals.

DKD, diabetic kidney disease; 5′-MTA, 5′-methylthioadenosine; M22G, N2,N2-dimethylguanosine; 4-HPPA, 4-hydroxyphenylpyruvic acid; HDL, high-density lipoprotein; eGFR, estimated glomerular filtration rate; ns, not significant. See also Figure S3 and Table S2.

We further performed a correlation analysis to examine the relationships between these metabolite marker candidates for DKD severity and various clinical parameters (Figures 3C and 3D). In urine, adenosine levels were negatively correlated with HbA1c (R = –0.23, p < 0.05) but positively correlated with eGFR (R = 0.48, p < 0.0001). Similarly, 5′-MTA negatively correlated with HbA1c (R = –0.36, p < 0.001) and other variables, while showing a strong positive correlation with eGFR (R = 0.48, p < 0.0001). However, serum M22G and aconitic acid showed correlations with these clinical variables in the opposite direction to those observed for urinary adenosine and 5′-MTA.

Adenosine levels within the S1 group, which comprised individuals with normo-albuminuria, including both HCs and T2DM patients, appeared to segregate into two distinct clusters. This prompted us to further investigate whether the distribution of diabetes diagnoses differed between the high- and low-adenosine-level subgroups. Among individuals with high-adenosine levels, 15.4% (2/13) had T2DM, whereas 65.4% (17/26) of those in the low-adenosine level group had T2DM. This difference in T2DM prevalence between adenosine level groups was statistically significant (χ² test, p < 0.01). This observation indicates that the S1 group may contain metabolically heterogeneous states driven by diabetes status, and thus early-stage comparisons should be interpreted cautiously.

In sex-stratified sensitivity analyses, the direction of stage-associated changes was consistent across sexes. In males (S1/S2/S3 = 26/22/22), all four metabolites remained significantly different across stages (ANOVA p < 0.0001), including urinary adenosine (p < 0.0001) and 5′-MTA (p < 0.0001), as well as serum M22G (p < 0.0001) and cis-aconitic acid (p < 0.0001). In females (S1/S2/S3 = 13/5/4), urinary adenosine and 5′-MTA remained significant across stages (p < 0.001 and p < 0.05, respectively), whereas serum markers did not reach significance (p < 0.1 and p < 0.2), likely reflecting limited sample size in female S2/S3 groups (Figure S3 and Table S2).

Integration of metabolite marker candidates with publicly available transcriptomic data

To further understand the biological significance of the identified metabolite marker candidates for DKD severity, we constructed metabolite-gene interaction networks for both urine and serum marker candidates for DKD severity (Figures 4A and 4B). The networks revealed intricate relationships between the metabolites and various genes, with several hub genes (blue) mediating the interactions between multiple metabolites. Some genes in these networks (pink) were identified as marker candidates for DKD severity in our previous study,⁷ validating our approach.

Figure 4.

Metabolite-protein interaction networks and transcriptomic expression patterns

(A and B) Network visualization of (A) urine metabolite-protein interactions and (B) serum metabolite-protein interactions. Red nodes represent key metabolites, blue nodes represent hub proteins, pink nodes indicate biomarkers identified in previous studies, and gray nodes represent other interacting proteins. Black-bordered nodes indicate proteins with differentially expressed mRNAs in DKD patient samples.

(C and D) Boxplots showing differential expression of candidate genes in (C) urine metabolite-associated network and (D) serum metabolite-associated network between healthy controls (HCs) and patients with DKD. Expression levels are presented as log2 RMA16 values, with asterisks indicating statistical significance (∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001).

(E) Integrated protein-protein interaction network of differentially expressed genes identified from both urine and serum metabolite networks. Green nodes indicate downregulated genes in DKD, and pink nodes indicate upregulated genes. Node size reflects importance in the network, with larger nodes and black borders indicating hub genes that connect the urine and serum subnetworks.

DKD, diabetic kidney disease.

We then integrated our metabolomic findings with publicly available transcriptomic data (GEO: GSE30529, the renal tubular dataset of mRNA, DKD vs. HCs)⁸ to identify genes that were differentially expressed between patients compared to HCs. In the urine metabolite-associated network, several genes showed significant differential expression between the HC and DKD groups, including NT5E, nerve growth factor (NGF), DRD2, methylthioadenosine phosphorylase (MTAP), HTR1A, HTR1B, CRH, IL4I1, PRL, KLRB1, GRIK1, FOLH1, and GCDH (Figure 4C). Similarly, in the serum metabolite-associated network, we identified differentially expressed genes including NT5E, ALB, PPIG, GCG, P2RY2, TAT, IL4I1, and CYP1B1 (Figure 4D).

Finally, to identify important hub genes among those associated with the discovered key metabolites, we constructed an integrated network using the differentially expressed genes from both urine and serum networks (Figure 4E). The integrated network revealed that most nodes were downregulated in patients with DKD compared with those in HCs, whereas a smaller gene (pink) was upregulated. In this network, NT5E emerged as the central key hub gene, which was identified in our earlier study.⁷ Genes derived from urine metabolites showed multiple connections, primarily with NGF, which is linked to NT5E. Similarly, genes associated with serum metabolites are connected through ALB, which in turn is linked to NT5E.

Discussion

Our AI-based framework demonstrated that among the five clinical criteria for DKD classification based on the combination of ACR and eGFR stages, the ACR-only criterion provided the most robust framework for metabolite-based stratification in this cohort. Using explainable AI, we discovered ACR stage-specific metabolite signatures, with urinary adenosine and 5′-MTA significantly decreasing, whereas serum M22G and cis-aconitic acid increasing with disease progression. Furthermore, integration with transcriptomic data through metabolite-protein and protein-protein interaction networks identified NT5E (CD73) as a key hub gene connecting the urine and serum metabolite networks. These findings support previous evidence implicating adenosine metabolism as a central pathway in DKD and provide a systems-level perspective on pathways associated with DKD severity.

In the present study, among the ACR stage-specific metabolic markers identified, urinary adenosine levels showed a clear and stepwise decline with advancing DKD. In our combined analysis of individuals with T2DM and HCs, individuals in the high-adenosine subgroup were less likely to have a diagnosis of T2DM. This finding is consistent with our previous study,⁷ further reinforcing the robustness and reproducibility of urinary adenosine as a candidate marker associated with DKD presence and severity in cross-sectional analyses.

Notably, although urinary adenosine levels were decreased in our dataset, integration with transcriptomic data revealed increased renal expression of NT5E (CD73), the enzyme responsible for generating extracellular adenosine. Adenosine functions as an endogenous signaling molecule in the kidney, regulating renal blood flow, glomerular filtration rate, tubuloglomerular feedback, and renin secretion.^9,10 However, chronic adenosine overproduction has been associated with inflammatory and fibrotic signaling in experimental and clinical studies, although causality cannot be inferred from our cross-sectional cohort data.

Consistent with these complex and context-dependent actions, previous studies have reported conflicting findings regarding the expression and role of adenosine and NT5E in DKD.^7,11 The apparent discrepancy between reduced urinary adenosine and increased tissue NT5E expression in the present study may reflect a stage-specific physiological adaptation. In early DKD, upregulation of CD73 in the proximal tubules may enhance local adenosine production and exert reno-protective effects.¹² By contrast, during later stages, sustained adenosine signaling, particularly via A2B receptors, has been implicated in promoting pro-inflammatory and pro-fibrotic pathways.^13,14 Thus, the observed reduction in urinary adenosine in advanced DKD may represent a protective counter-regulatory mechanism that limits excessive purinergic activation.

Additional mechanisms, including elevated adenosine deaminase activity or impaired tubular secretion, may also contribute to reduced urinary adenosine levels.¹⁵ The upregulation of NT5E was derived from publicly available transcriptomic datasets, not from direct renal tissue analysis in our cohort. Therefore, future studies using paired kidney tissues and urine samples from the same patients are essential to validate and elucidate these mechanisms.

We also observed reduced urinary levels of 5′-MTA in patients with advanced DKD, which is noteworthy considering its potential role in methylation regulation.^16,17,18 5′-MTA is metabolized by MTAP, an enzyme essential to the methionine salvage pathways, with adenine produced as a byproduct.^17,19 Dysregulation of this pathway may lead to both methylation imbalance and excess adenine accumulation.²⁰ Sharma et al. have recently demonstrated that adenine accumulation in the kidneys of diabetic patients activates mTOR signaling and promotes tubulointerstitial fibrosis, thereby contributing to DKD progression.²¹ Although 5′-MTA and its analogs have been investigated as potential chemoprotective agents in other disease conditions,²² the role of 5′-MTA itself and the connection between adenine production in DKD remain largely unexplored. Reduced urinary excretion of 5′-MTA, along with adenosine and adenine, may constitute a metabolic axis in the pathogenesis of DKD.

In contrast to the decreased levels of urinary metabolites, we observed elevated levels of M22G and cis-aconitic acid in serum samples from patients with DKD. M22G is a specific nucleoside with potential as a biomarker for DKD progression,²³ and its predictive value is linked to early tubular dysfunction during nucleoside handling. The observed elevation of cis-aconitic acid in serum is consistent with, yet also contrasts with, previous findings, such as a meta-analysis²⁴ showing that aconitic acid is significantly downregulated in the urine but upregulated in the blood of patients with DKD. This bidirectional pattern across different biofluid compartments is not unique to aconitic acid. Several metabolites, including cytidine and glycine, demonstrate opposing concentration patterns in urine and blood. These compartment-specific changes underscore the complex nature of metabolic dysregulation in DKD and suggest alterations in the renal handling of both nucleosides and tricarboxylic acid cycle intermediates, reflecting the multifaceted pathophysiology of DKD.

The significant correlations observed between key metabolites and clinical parameters, such as HbA1c and eGFR, further support their clinical relevance. Associations with comorbidities, including cardiovascular disease, hypertension, and dyslipidemia, indicate that these metabolites may reflect systemic metabolic disturbances, thereby extending their utility beyond renal-specific marker candidates for DKD severity.

To further elucidate the molecular underpinnings of DKD, we integrated our cohort-derived metabolomic data with publicly available transcriptomic data through unified network analysis. This approach enables a more comprehensive evaluation of disease pathophysiology by linking metabolomic alterations to underlying gene expression patterns. Importantly, this multilayered analysis reaffirmed the central role of NT5E (CD73), the key hub gene previously identified in our integrative framework, whose network connectivity extended across both urine- and serum-related molecular pathways. Beyond NT5E, network analysis revealed that genes associated with urine metabolites converged around NGF, suggesting a potential neuroinflammatory component of DKD that warrants further exploration. Experimental models have shown that NGF upregulation in diabetic kidneys parallels NF-κB activation, and that suppressing both via mycophenolate mofetil and insulin reduces renal injury, implicating NGF in inflammatory signaling pathways relevant to DKD.²⁵ Serum metabolite-associated genes were centered around albumin, suggesting the importance of protein binding and transport in the metabolic dysregulation characteristics of DKD. Altogether, these multi-omics findings provide a hypothesis-generating context linking metabolite signatures to gene expression patterns from a public dataset. Cohort-matched transcriptomic or functional studies will be required to evaluate mechanistic relevance and therapeutic implications.

ACR-based staging showed the strongest alignment with metabolomic discrimination in this cohort. Accordingly, the identified metabolites—including urinary adenosine and 5′-MTA, as well as serum M22G and cis-aconitic acid—should be considered candidate metabolic markers associated with DKD stage and severity. Although several of these metabolites have been reported previously in DKD, our study extends prior work by applying rigorously benchmarked nested cross-validation-based model evaluation in a well-characterized Korean cohort and by integrating metabolomic findings with a public renal transcriptomic dataset to provide a hypothesis-generating molecular context.

In conclusion, our integrative analysis highlights purine-related metabolic pathways associated with ACR-based DKD staging and identifies a set of candidate metabolites that warrant further validation.

Limitations of the study

This study has some limitations. First, although our cohort included HCs and patients with T2DM who were randomly selected to ensure balanced representation across eGFR stages, it did not include a separately defined subgroup of individuals at a high risk of developing DKD. In addition, comparisons involving HCs may have been influenced by cohort or protocol differences and by differences in medication use and comorbidities, which could confound metabolite levels, particularly in early-stage groups. This may have limited our ability to identify the early markers of DKD. Second, the cross-sectional design precludes establishing causal relationships between metabolite alterations and disease development; longitudinal studies with serial sampling are necessary to determine whether the observed changes precede clinical manifestations or represent consequences of established disease. Third, serum metabolites did not differ across ACR groups in females, likely due to the small number of female participants in S2 and S3; larger cohorts will be required to evaluate sex-by-stage interactions. Finally, despite our robust AI-based approach, comprehensive metabolite identification and pathway analysis were constrained by the limited number of metabolites included and the current state of metabolic pathway databases.

Future studies should validate our findings in larger and more diverse cohorts, ideally with longitudinal follow-up, to assess their generalizability and potential prognostic utility. In addition, cohort-matched transcriptomic and experimental studies will be required to clarify whether adenosine-related pathways represent actionable targets in DKD.

By integrating metabolomic profiles with transcriptomic data and clinical parameters, we provided a more comprehensive understanding of the complex metabolic dysregulation underlying DKD progression, potentially opening new avenues for targeted interventions that address both kidney-specific and systemic aspects of the increasingly prevalent complications of diabetes.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Nan Hee Kim (nhkendo@gmail.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The public renal tubular microarray dataset analyzed in this study is available in the NCBI Gene Expression Omnibus under accession GEO: GSE30529 and is listed in the key resources table. The de-identified clinical and metabolomics data are not publicly available because the applicable informed-consent and institutional review board approvals do not permit unrestricted public deposition. Qualified researchers may request access from the lead contact; access is subject to approval by the relevant institutional review board(s) and execution of the data-use agreement.
• •The code used for the analyses reported in this paper is available from the lead contact upon reasonable request.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.

Acknowledgments

This work was supported by the Bio & Medical Technology Development Program of the National Research Foundation of Korea (NRF), funded by the Korean Government (MSIT) (NRF-2019M3E5D3073102 and NRF-2023R1A2C2003479, to N.H.K.), and by a grant from the National IT Industry Promotion Agency, also funded by MSIT (no. S0252-21-1001, Development of AI Precision Medical Solution (Doctor Answer 2.0), to N.H.K.). Additional support was provided to N.H.K. by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health and Welfare, Republic of Korea (H123C0679). This study was also supported by a grant from the National Research Foundation of Korea (NRF), funded by the Korean Government (MSIT) (no. NRF-2021R1A5A2030333, to D.H.L.). The study sponsors were not involved in the design of the study; the collection, analysis, and interpretation of data; writing the report; or imposing any restrictions regarding the publication of the report.

Author contributions

Conceptualization, I.J., S.P., N.H.K., and D.H.L. Data curation, analysis, and interpretation, I.J., S.P., D.Y.L., S.Y.P., J.H.Y., J.A.S., H.K.P., S.H.K., N.H.K., and D.H.L. Metabolomics sample preparation and data acquisition, M.J. and M.J.P. Drafting and revision, I.J., S.P., and N.H.K. Supervision, reviewing, and editing, N.H.K. and D.H.L. All the authors have read and approved the final version of this manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Study metabolomics/clinical data	This paper	Available upon request
Renal tubule microarray dataset	NCBI GEO	GEO: GSE30529
Software and algorithms
MetaboAnalyst	Pang et al.26	https://www.metaboanalyst.ca/
ECMarker	Jin et al.27	https://github.com/daifengwanglab/ECMarker
Original analysis code	This paper	Available upon request
Python 3.11.3 (original analyses); Python 3.13.9 (revision analyses)	Python Software Foundation	https://www.python.org; RRID:SCR_008394
scikit-learn 1.7.2	scikit-learn developers	https://scikit-learn.org/; RRID:SCR_002577
Python scientific and visualization packages	Package developers	NumPy 2.3.5; pandas 2.3.3; SciPy 1.16.3; statsmodels 0.14.5; matplotlib 3.10.6; seaborn 0.13.2
PyTorch 2.2.0 (original analyses); PyTorch 2.10.0 and Captum 0.8.0 (revision analyses)	PyTorch and Captum developers	https://pytorch.org/; https://captum.ai/

Experimental model and study participant details

Human participants

The participants were recruited from three hospitals in South Korea: Korea University Ansan Hospital, Gachon University Gil Medical Center, and Soonchunhyang University Seoul Hospital. Healthy controls and patients with T2DM were enrolled after obtaining informed consent. The inclusion criteria for the T2DM group were as follows: age of 30–80 years, followed up for >2 years with at least two prior measurements of both ACR and eGFR before providing informed consent, and baseline eGFR ≥30 mL/min/1.73 m², calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation. The exclusion criteria were T1DM, ESRD, non-diabetic CKD, uncontrolled hypertension, pregnancy, and severe systemic or psychiatric illness. Meanwhile, the HC group comprised participants aged 30–80 years without a history of diabetes, hypertension, or CKD, followed up for >2 years in the Korean Genome and Epidemiology Study (KoGES), and with eGFR ≥60 mL/min/1.73 m². Healthy controls were required to have stable kidney function, defined as an annual decline in eGFR of <3.3% and a total decline of <5 mL/min/1.73 m² during follow-up. The exclusion criteria for the HC group were prior renal transplantation, solitary kidney, hereditary kidney disease, history of acute kidney injury, body mass index <18.5 kg/m², malignancy within the past 5 years, chronic use of nephrotoxic drugs or systemic steroids, and any condition considered inappropriate for study participation. Healthy controls were derived from a distinct cohort framework (KoGES) with different follow-up and medication profiles. Although samples were processed using harmonized metabolomics pipelines, residual protocol-related differences cannot be fully excluded and should be considered when interpreting comparisons involving healthy controls.

Blood and urine samples were collected from healthy individuals (HC group, n = 20) and 80 patients with T2DM. Participants with T2DM were selected using a stratified sampling approach to achieve approximately balanced representation across predefined eGFR stages (G1, n = 20; G2, n = 20; G3a, n = 22; G3b, n = 18) and ACR categories (normo-albuminuria, n = 20; microalbuminuria, n = 30; and macro-albuminuria, n = 30). eGFR and ACR stages were classified based on the Kidney Disease Outcomes Quality Initiative practice guidelines.² After excluding those with insufficient sample volume, withdrawal of consent, or dropout from the study, 72 samples from individuals with T2DM were included in the final metabolomic analysis, which was conducted at Sunchon National University.

This study was approved by the institutional review boards of all participating hospitals, including Korea University Ansan Hospital (2019AS0226). All participants provided written informed consent prior to sample collection. This study was conducted following the principles outlined in the Declaration of Helsinki.

Method details

Metabolite quantitation

Analytical quality control and data preprocessing

Solvent blank injections were included throughout the analytical sequence to monitor background contamination and carryover. A blank-based low-quality filter was applied to remove metabolites with substantial background interference, excluding features with blank signals exceeding 20% of the mean signal in biological samples (blank/sample >0.2). Because a fully targeted quantitative platform was used (triple quadrupole MS operated in MRM mode), quantitative reproducibility was ensured by calibration-based quantification and internal-standard normalization, and instrument stability was monitored via calibration performance and internal standard responses across the run. Data preprocessing was performed in MetaboAnalyst (v6.0): metabolites with >30% missingness were removed, and remaining missing values were treated as left-censored (below LoD) and imputed using the LoD-based approach (replacing with one-fifth of the minimum positive value for each metabolite). All samples were analyzed in a single analytical batch per platform under identical instrument conditions; therefore, batch correction was not applied.

Gas chromatography–tandem mass spectrometry (GC-MS/MS)

Organic acids (OAs) and fatty acids (FAs) in serum and urine were analyzed by Shimadzu TQ 8040 system (Shimadzu, Kyoto, Japan) equipped with a cross-linked Ultra-2 capillary column (25 m × 0.20 mm I.D., 0.11 μm film thickness; 5% phenyl–95% methylpolysiloxane, Agilent Technologies, Atlanta, GA, USA). Each sample (1.0 μL) was injected in split mode (10:1). The oven temperature was initially held at 100 °C for 2 min, ramped to 300 °C at 10 °C/min, and maintained for 8 min. Helium was used as the carrier gas (0.5 mL/min), and argon was used as the collision gas. Electron impact ionization (70 eV) was applied.

Liquid chromatography–tandem mass spectrometry (LC-MS/MS)

Amino acids (AAs), kynurenine pathway metabolites, and nucleosides were analyzed using a Shimadzu LCMS-8050 system (Shimadzu, Kyoto, Japan) with Intrada Amino Acid (50 × 3.0 mm, 3 μm) and ACQUITY UPLC HSS T3 (100 × 2.1 mm, 1.8 μm) columns. Electrospray ionization was used under the following conditions: nebulizing gas, 3.0 L/min; heating gas, 10.0 L/min; interface temperature, 300 °C; desolvation line, 250 °C. For AAs, mobile phases were ACN/tetrahydrofuran/25 mM ammonium formate/formic acid (9:75:16:0.3, v/v/v/v) and ACN/100 mM ammonium formate in water (20:80, v/v). For kynurenine metabolites and nucleosides, 0.1% formic acid in water and ACN was used.

Sample preparation for OA and FA profiling analyses in serum and urine by GC-MS/MS

The OA and FA profiles in serum and urine were analyzed through methoxime (MO)/tert-butyldimethylsilyl (TBDMS) derivatization, as previously described.^28,29 Serum and urine samples (100 μL each) were deproteinized with 200 μL of ACN containing 0.1 μg of 3,4-dimethoxybenzoic acid and pentadecanoic acid as internal standards (ISs). After centrifugation, 1 mL water and 1 mg methoxyamine hydrochloride were added, the pH adjusted to ≥12 with 5.0 M NaOH, and incubated at 60 °C for 1 h. The aqueous phase was acidified (pH ≤ 2) with 10% sulfuric acid, saturated with NaCl, and extracted with diethyl ether (3.0 mL) and ethyl acetate (2.0 mL). After evaporation under nitrogen at 40 °C, the residue was derivatized with 20 μL MTBSTFA and 10 μL toluene at 60 °C for 1 h, and analyzed by GC-MS/MS.

Sample preparation for profiling analyses of AAs, kynurenine pathway metabolites, and nucleosides in serum and urine by LC-MS/MS

Serum (30 μL) and urine (10 μL) samples were deproteinized using ACN containing ISs (25 ng, 13C1-phenylalanine; 1.25 μg, 3,4-dimethoxybenzoic acid; 2.5 ng, 3-deazauridine), centrifuged, filtered, and analyzed by LC-MS/MS.

Restricted Boltzmann machine for DKD stage prediction

To determine the optimal combination of ACR and eGFR criteria for discriminating metabolomic alterations in DKD, we employed an RBM based on the ECMarker framework.²⁷ The RBM architecture consisted of a visible layer for metabolite data input, a hidden feature extraction layer, and an output layer for DKD stage classification (Figures 1A and 1B). Five clinical diagnostic criteria were evaluated using 100 independent learning cycles per criterion. Criterion 1 classified all participants based solely on ACR levels: S1 (<30 mg/g), S2 (30–300 mg/g), and S3 (≥300 mg/g), which aligned with the KDIGO albuminuria stages A1, A2, and A3. Criterion 2 applied the complete KDIGO 2012 CKD classification system,⁶ categorizing participants into healthy controls and CKD G1, G2, G3a, and G3b groups based on eGFR. Criterion 3 used a modified grouping as follows: healthy controls, patients with CKD G1 and G2, and a combined G3a+G3b group. Criterion 4 further simplifies this into three groups: healthy controls, the combined G1+G2 group, and the combined G3a+G3b group. Criterion 5 divided the participants into four groups: healthy controls, T2DM participants with normo-albuminuria (S1), T2DM participants with microalbuminuria (S2), and T2DM participants with macroalbuminuria (S3). Across all criteria, healthy controls were assigned to the lowest-risk categories.

In response to concerns regarding model robustness in a modest sample size, we additionally performed a formal benchmarking analysis restricted to Criterion 1 (ACR-based staging; S1/S2/S3), comparing the RBM model with three baseline classifiers: LASSO (L1-regularized logistic regression), linear support vector machine (SVM), and random forest (RF). Prior to model training, missing values were imputed using the mean value of each feature. For LASSO, linear SVM, and RF, features were standardized using Z score normalization. For the RBM pipeline, metabolite variables were scaled to a [0,1] range prior to the RBM transformation layer. To address class imbalance, class weights were set to balanced for all applicable models.

Model evaluation and benchmarking were performed using nested cross-validation (nested CV) to minimize optimistic bias and prevent data leakage. In the outer loop, 5-fold stratified cross-validation was used for performance evaluation. Within each outer training set, an inner 3-fold stratified cross-validation with grid search was used for hyperparameter tuning, optimizing one-vs-rest macro-averaged AUROC. The best estimator identified in the inner loop was refit on the entire outer training set and evaluated on the held-out outer fold.

Hyperparameter grids were defined as follows: for LASSO and linear SVM, the regularization parameter C was tuned over [0.1, 1.0, 10.0, 100.0]; for RF, n_estimators was tuned over [50, 100, 200] and max_depth over [None, 5, 10]. For the RBM benchmarking pipeline, we used a Bernoulli RBM layer followed by logistic regression, tuning n_components over [128, 256] and classifier C over [10.0, 100.0], with RBM training performed using a learning rate of 0.01, batch size of 32, and 100 epochs.

We report macro-averaged AUROC as the primary metric, along with accuracy, macro-F1, macro-precision, and macro-recall. Multi-class ROC curves were constructed using a one-vs-rest strategy based on out-of-fold predicted probabilities aggregated across the outer folds.

To identify discriminative metabolites, feature importance was quantified for each model. For LASSO and linear SVM, feature importance was computed as the mean absolute coefficient across one-vs-rest classifiers. For RF, importance was derived from normalized Gini impurity reduction.

The original ECMarker analyses were performed using Python 3.11.3 and PyTorch 2.2.0. The revision benchmarking, statistical visualization, and feature-extraction analyses were performed using Python 3.13.9, scikit-learn 1.7.2, NumPy 2.3.5, pandas 2.3.3, SciPy 1.16.3, PyTorch 2.10.0, and Captum 0.8.0.

Feature selection

To identify key metabolites contributing to the classification of each DKD stage, we implemented an explainable AI approach using the Integrated Gradients algorithm from the Captum library.³⁰ The top 20 metabolites with the highest attribution scores were selected for each disease state. Furthermore, among the commonly detected metabolites across DKD stages, we selected those that showed statistically significant differential expression across stages S1, S2, and S3 in our metabolite expression dataset as marker candidates for DKD severity for further downstream analyses.

Metabolic pathway analysis

The metabolites selected for this analysis were identified as significant contributors to the classification of different DKD stages in our AI model. The selection was based on importance scores calculated using the integrated gradients method during model evaluation. Pathway analysis was conducted using MetaboAnalyst version 6.0 (https://www.metaboanalyst.ca/),²⁶ a comprehensive web-based platform for metabolomic data analysis. Analyses were performed using the Homo sapiens pathway library as reference.

Network construction

To explore molecular context associated with DKD severity, we employed a two-stage network analysis approach. First, we constructed metabolite–gene interaction networks using MetaboAnalyst separately for urine and serum metabolites identified by our AI model. These networks mapped the relationships between the key metabolites and their interacting genes.

Second, the genes identified in these networks were used to build protein–protein interaction (PPI) via the STRING database,³¹ while preserving the original metabolite–gene connections. We identified hub genes within these networks and analyzed their differential expression in patients with DKD and healthy controls using a publicly available renal tubular tissue gene expression dataset (GSE30529).⁸

All networks were visualized using Cytoscape 3.8.0,³² which enabled us to merge the urine and serum PPI networks to identify the common molecular pathways across both biofluids.

Quantification and statistical analysis

Statistical analyses were performed using individual participants as independent biological samples; thus, n represents the number of participants. The final metabolomics cohort included 92 participants (20 healthy control and 72 participants with type 2 diabetes mellitus). For ACR-based analyses, the S1, S2, and S3 included 39, 27, and 26 participants, respectively. Sex-stratified sample sizes were 26/22/22 for males and 13/5/4 for females (S1/S2/S3). Continuous variables are presented as mean ± SD or median (IQR), as appropriate, and categorical variables as n (%). In boxplots, the center line indicates the median, boxes represent the IQR, and whiskers extend to 1.5 × IQR; all individual observations are overlaid as jittered points.

Differences among DKD stages were assessed by one-way ANOVA, followed by two-sided Welch’s t-tests for pairwise comparisons. Pairwise p values in Figure S3 and Table S2 are unadjusted, whereas analyses across all measured metabolites used Bonferroni and Holm corrections. Other two-group comparisons used Student’s t test or the Mann-Whitney U test, as appropriate; categorical variables were compared using the chi-square test, and correlations were assessed using Spearman’s rank correlation. Model performance was evaluated using nested stratified cross-validation (outer 5-fold and inner 3-fold), and permutation testing used 100 label permutations. Unless otherwise indicated, p < 0.05 was considered statistically significant. Exact n values, statistical tests, and p values are provided in the Results, figures, figure legends, and Tables S1 and S2. Analyses were performed using R 4.2.1 and Python; software and version details are provided in method details and the key resources table.