GRAPHICAL ABSTRACT

Highlights

• Multi-center cohorts revealed dynamic MMP-7/SAA1/TNC changes across DKD stages.

• Machine learning improved DKD stage prediction with protein biomarkers.

• Integrating metabolites further enhanced DKD prediction accuracy.

INTRODUCTION

By 2021, an estimated 537 million adults were living with diabetes worldwide [1], with the vast majority affected by type 2 diabetes mellitus (T2DM). This figure is projected to rise, with estimates indicating 643 million cases by 2030 and 783 million by 2045 if current trends continue [1]. Among the numerous complications that accompany T2DM, diabetic kidney disease (DKD) is one of the most severe and prevalent [2]. DKD is not only the current leading cause of chronic kidney disease (CKD) and end-stage renal disease (ESRD), but also significantly elevates the risk of cardiovascular events and mortality among patients with diabetes [3,4]. However, a major challenge in DKD management is the early and precise identification of kidney damage.

The progression of DKD is particularly aggressive, driven by persistent hyperglycemia and the early involvement of harmful pathways such as hyperfiltration, inflammation, and oxidative stress [4,5]. These factors accelerate the decline in kidney function, often outpacing the progression seen in CKD caused by primary glomerular diseases. This rapid decline underscores the critical need for early DKD diagnosis to detect subtle kidney damage before it becomes irreversible. In clinical practice, DKD is typically screened and monitored using the urinary albumin-to-creatinine ratio (UACR), estimated glomerular filtration rate (eGFR), and, in some cases, renal biopsy. Although these methods are valuable, they often miss early-stage disease because initial symptoms are subtle, leading to diagnosis only after significant progression. Furthermore, not all patients with T2DM who have abnormal albuminuria or impaired renal function progress to ESRD, highlighting the complexity and variability of DKD progression [6]. This variability has driven a decades-long search for novel blood and urine biomarkers closely linked to DKD pathogenesis [7–9]. Notably, biomarkers associated with extracellular matrix (ECM) turnover, inflammation, and oxidative stress in DKD progression—such as metalloproteinase-7 (MMP-7) [10], serum amyloid A1 (SAA1) [11], and tenascin C (TNC) [12]—have emerged as promising tools for improving diagnostic and prognostic accuracy. Accumulating evidence, including our own findings, has shown that these proteins are consistently upregulated and have played prominent roles in the pathogenesis and progression of various non-DKD conditions over the past decades [13–17]. However, the ability of these biomarkers, alone or in combination, to predict early DKD remains unproven in robust clinical trials.

In addition, serum metabolomics has emerged as a powerful complementary approach for DKD diagnosis [18–20]. This approach involves the comprehensive analysis of small molecules within biological samples, offering the ability to detect subtle metabolic changes that precede overt clinical symptoms. Using metabolomics, our previous study profiled the landscape of serum metabolites at different stages of DKD [18], but the capacity of these metabolites to predict the occurrence of kidney damage after DKD was unsatisfactory. In this study, by integrating serum metabolomics with established biomarkers, we aim to develop robust diagnostic machine learning models that can detect DKD at its earliest stages. This integrated approach not only enhances the predictive power of conventional biomarkers but also paves the way for earlier intervention, ultimately leading to better outcomes for patients with DKD.

METHODS

Cohort recruitment

Five cross-sectional cohorts of patients with DKD were established, including one discovery cohort at the primary center (the Affiliated Hospital of Nanjing University of Chinese Medicine) and four external testing cohorts at the 3rd Xiangya Hospital of Central South University (C.S.), Tongji Hospital of Tongji University (T.J.), Nanfang Hospital (N.F.), and Zhongda Hospital (Z.D.). The discovery cohort (173 healthy controls [HCs], 277 patients with T2DM, 98 patients with diabetic kidney disease at an early stage [DKD-E], and 69 patients with diabetic kidney disease at an advanced stage [DKD-A]) from the primary center and two external cross-sectional cohorts, C.S. (29 HCs, 29 patients with T2DM, 20 patients with DKD-E, and 21 patients with DKD-A) and T.J. (34 HCs, 39 patients with T2DM, 25 patients with DKD-E, and 24 patients with DKD-A), were established in our previous study [18]. The N.F. cohort (30 HCs, 17 patients with T2DM, 4 patients with DKD-E, and 33 patients with DKD-A) and Z.D. cohort (79 patients with T2DM, 36 patients with DKD-E, and 15 patients with DKD-A) were newly recruited for the current study. The same inclusion and exclusion criteria were applied as previously reported [18].

Ethics compliance statement

This investigator-initiated clinical trial was conducted in accordance with the principles of the 1975 Declaration of Helsinki. Approval was obtained from the Ethics Committee of the First Affiliated Hospital of Nanjing University of Traditional Chinese Medicine (2019NL-109-02). The clinical trial was registered with the Chinese Clinical Trial Registry (ChiCTR 2000028949). In this cross-sectional study, additional serum samples were collected after routine clinical biochemical tests. All participants provided informed consent for the academic use of these extra samples.

Inclusion/exclusion criteria

The inclusion criteria were as follows: (1) age between 20 and 75 years; (2) diagnostic criteria for T2DM (all patients) and disease duration of more than 10 years; (3) DKD based on the diagnostic criteria for microalbuminuria and macroalbuminuria; (4) eGFR ≥90 mL/min/1.73 m² for the T2DM group, and eGFR above 30 mL/min/1.73 m² in both microalbuminuria and macroalbuminuria groups; (5) blood pressure <140/90 mm Hg; and (6) signed informed consent. The exclusion criteria were as follows: (1) the presence of primary kidney disease with a definite diagnosis; (2) other systemic diseases known to cause proteinuria; (3) acute diabetic complications or urinary tract infection in the past month; (4) coexisting serious primary diseases involving the cardiovascular, cerebrovascular, hepatic, renal, or hematopoietic systems, or malignancy; (5) mental illness and inability to cooperate; (6) pregnancy, lactation, or planned pregnancy; (7) current menstruation; and (8) participation in other clinical trials within the past month.

Public data mining

This study analyzed data from two published sources (GSE 131882 and GSE261545) [21–23]. Data from study GSE131882 were obtained from the Gene Expression Omnibus (GEO) database, and single-cell RNA sequencing data were processed using the Seurat R package (Satija Lab, New York, NY, USA) [24,25]. Based on the gene-by-cell expression matrix, dimensionality reduction and cell visualization were performed using the uniform manifold approximation and projection (UMAP) algorithm, where UMAP_1 and UMAP_2 represent the first and second reduced dimensions, respectively. Each dot on the UMAP plot represents an individual cell, colored by library or marker expression level. Data from GSE261545 were obtained from the GEO database and included spatial transcriptomics data from a kidney specimen from a patient with CKD caused by diabetic and hypertensive nephropathy [23]. Raw data were pre-processed using the Seurat R package. Spatial feature plots were used to visualize the gene expression within the tissue slide.

Human kidney biopsy specimens

Human kidney specimens were obtained from diagnostic renal biopsies performed at the Affiliated Hospital of Nanjing University of Chinese Medicine. Non-tumor kidney tissue samples from patients with renal cell carcinoma who underwent nephrectomy were used as normal controls. All studies involving human kidney sections were approved by the Institutional Review Board of the Affiliated Hospital of Nanjing University of Chinese Medicine. Immunohistochemical staining was performed using human samples as previously described [26]. The primary antibodies used were anti-MMP-7 (GTX11716, GeneTex, Irvine, CA, USA), anti-TNC (ab108930), and anti-SAA1 (ab199030, Abcam Inc., Cambridge, MA, USA).

Mouse models and Western blot analysis

Heart, kidney, and liver tissues from male, 18 and 24-week-old lean controls and BKS-Lepr^em2Cd479/Nju (db/db) mice were provided by Dr. Yuanyuan Wang at Guizhou Medical University (Guiyang, China). Originally, clean grade 8-week-old db/db mice were purchased from the Institute of Model Animals, Nanjing University. These mice are commonly used as a model of T2DM and organ complications because of their obesity, hyperglycemia, and insulin resistance. All mice were housed in a temperature-controlled room (22°C) under a 12-hour light dark cycle, with free access to water and food.

These harvested tissues were lysed with radioimmune precipitation assay (RIPA) buffer containing 1% NP-40, 0.1% sodium dodecyl sulfate (SDS), 100 μg/mL phenylmethanesulfonyl fluoride (PMSF), 1% protease inhibitor cocktail, and 1% phosphatase I and II inhibitor cocktail (Cell Signaling Technology, Danvers, MA, USA) in phosphate-buffered saline (PBS) on ice. The supernatants were collected after centrifugation at 13,000 ×g at 4°C for 15 minutes. Protein expression was analyzed by Western blotting as described previously [27]. In brief, the samples were separated by 12% sodium dodecyl sulfate-polyacrylamide gel electrophoresis and electrotransferred to polyvinylidene difluoride (PVDF) membranes (Microporous, Piney Flats, TN, USA). Subsequently, the membranes were blocked for 2 hours and probed with primary antibodies at 4°C for 16 hours. The chemiluminescence signal was detected after incubation with the secondary antibody for 1 hour. Finally, ImageJ software (National Institutes of Health, Bethesda, MD, USA) was used for analysis. The primary antibodies used were anti-MMP-7 (GTX11716, GeneTex), anti-TNC (ab108930), anti-SAA1 (ab207445, Abcam Inc.), and anti-β-actin (sc-130065, Santa Cruz Biotechnology, Dallas, TX, USA).

Serum proteomic analysis

As described in our previous study [18], proteomics data were collected across 12 libraries, with three libraries for each of the four groups: HC, T2DM, DKD-E, and DKD-A. After missing values were filtered out, 581 proteins were ultimately used for analysis.

Serum metabolomics

Serum samples were thawed on ice, and 40 μL of plasma was mixed with 225 μL of ice-cold methanol and, then vortexed. After 750 μL of cold methyl tert-butyl ether (MTBE) was added, the mixture was shaken at 4°C for 10 minutes. Next, 188 μL of liquid chromatography/mass spectrometry (LC/MS)-grade water was added, followed by centrifugation at 14,000 rcf for 2 minutes at 4°C. The lower layer was transfer, mixed with 750 μL of methanol (1:1), shaken for 10 minutes, and centrifuged again. The supernatant (475 μL) was dried in a SpeedVac at 45°C for 2 hours. Dried aliquots were combined with an internal standard and methoxyamine hydrochloride, vortexed, and shaken. Bis (trimethylsilyl)trifluoroacetamide (BSTFA) was then added, and the mixture was shaken before transferring to a vial for gas chromatography-mass spectrometry (GC-MS) analysis. Quality control (QC) samples were prepared by pooling aliquots from all samples, with a QC run every 10 injections. The analysis was performed on a Trace 1310 GC coupled to a TSQ 8000 mass spectrometer (Thermo Fisher, Waltham, MA, USA) [28]. A standard n-alkane mixture was used to correct retention time shifts. The extraction of metabolomics raw data was previously reported [18]. In total, 349 metabolites were identified in the discovery cohort. After raw data pre-processing and filtering of exogenous metabolites, 207 metabolites were ultimately measured and used for further analysis.

Enzyme-linked immunosorbent assay and machine learning prediction

Enzyme-linked immunosorbent assay (ELISA) kits for human MMP-7, SAA1, and TNC were purchased from R&D Systems (Minneapolis, MN, USA) or Immuno-Biological Laboratories (Fujioka, Japan). The assays used a quantitative sandwich enzyme immunoassay technique. Microplates were pre-coated with monoclonal antibodies specific to these proteins. Standards and samples were added to the wells, where the target proteins bound to the immobilized antibodies. After unbound substances were washed away, an enzyme-linked polyclonal antibody specific to the proteins was added. Following a final wash, a substrate solution was introduced, and color developed proportionally to protein levels. The reaction was stopped, and color intensity was measured, with serum protein levels expressed in nanograms per milliliter. For pairwise comparisons, the associations between individual ELISA markers and disease status were analyzed using receiver operating characteristic (ROC) curves, and the area under the curve (AUC) was calculated using the R package “pROC” [29]. For predictions using two or three ELISA markers, 5-fold cross-validation (CV) was first performed in the primary center using four machine learning methods: linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), and logistic regression (Logi) [30], applied through the R packages “MASS,” “e1071,” and “randomForest.” Next, to test the new samples, machine learning models optimized during CV were trained on the entire primary center dataset and then applied to the four external centers for performance evaluation.

Prediction evaluation

Prediction results were compared with patients’ true disease status. True positive (TP) was defined as a case correctly predicted as positive; true negative (TN), as a case correctly predicted as negative; false positive (FP), as a case incorrectly predicted as positive; and false negative (FN), as a case incorrectly predicted as negative. Accuracy, sensitivity, specificity, and the Youden index were defined as follows:

Accuracy=TP+TNTP+TN+FP+FNSensitivity=TPTP+FNSpecificity=TNTN+FPYouden index=Sensitivity+Specificity-1

Integration of ELISA and metabolomics data

ELISA and metabolite features were pooled for pairwise and multi-outcome prediction analyses. From the primary center, a total of 133 healthy, 276 T2DM, 49 DKD-E, and 68 DKD-A serum specimens had both ELISA and metabolomics markers measured. Similar machine learning algorithms were used as described in the ELISA data analysis. The top differentially expressed metabolites were selected as described in the previous study [18]. For all disease pairwise and multi-outcome predictions, the three ELISA markers, the differentially expressed metabolomics markers, or both feature sets were integrated by the machine learning classifiers in 5-fold CV.

Association between ELISA and clinical features

To test the association of clinical features with disease status, ROC curves and AUC values were analyzed using the R package “pROC” [29]. Additionally, linear regression models were established to check the association between ELISA markers and the clinical features that were highly correlated with DKD:

Clinical feature=α·MMP7+β·TNC+γ·SAA1+λ·MMP7·TNC·SAA1+ɛ

The clinical features included blood glucose, fasting C-peptide (FC), glycosylated hemoglobin (HbA1c), UACR, urinary microalbumin (UmALB), eGFR, total cholesterol (TC), triglycerides (TG), low-density lipoprotein (LDL), high-density lipoprotein (HDL), and albumin. α, β, γ, and λ are the coefficients for the three ELISA markers and the interaction term, and ɛ is the error term. Higher similarity between predicted and true values indicates a stronger association of the ELISA markers with these clinical features.

Statistics

All statistical analyses and data visualizations were performed in R using the corresponding available packages. For the animal studies presented, all data were expressed as mean±standard error of the mean. Statistical analysis of the data was performed using GraphPad Prism 9 (GraphPad Software, San Diego, CA, USA). Comparisons between two groups were made using a two-tailed Student’s t-test or the rank-sum test if data failed a normality test. Results are presented in dot plots, with dots denoting individual values. Wherever applicable, P≤0.05 (or a false discovery rate [FDR] of 5% for multiple hypothesis testing) was used to define significance. For the metabolomics subgroup, three comparisons were made per ELISA marker: HC vs. T2DM, T2DM vs. DKD-E, and DKD-E vs. DKD-A. The Bonferroni adjustment for each ELISA marker was performed when calculating the sample size. Per comparison, to control the FDR at 10%, the Bonferroni method adjusted the probability of type I error per pairwise test to 0.1/3=0.033. Sample size estimation indicated that a minimum of 23 samples per group was required to achieve 80% statistical power (power=0.8) with an adjusted type I error rate (α) of 0.033, using the non-parametric Mann-Whitney U test (Wilcoxon rank-sum test).

RESULTS

Multi-omics reveals the expression patterns of MMP-7, SAA1, and TNC in DKD patients

During DKD onset and progression, inflammatory and fibrotic processes, in addition to metabolic and hemodynamic factors, significantly contribute to the progressive deterioration of kidney function [4,5]. In serum, several prominent biomarkers have been identified as closely linked to DKD pathogenesis. For example, serum or urine matrix protein MMP-7 is increased in DKD and correlates with both cardiac function and renal filtration [10,31,32]; the ECM glycoprotein TNC is elevated and positively correlated with both UACR and hypertension [33]; and SAA predicts an increased risk of death and progression to ESRD in patients with DKD [34,35]. Clearly, these biomarkers offer potential clues for predicting DKD before significant renal damage occurs. However, the predictive value of a single circulating biomarker for identifying early DKD remains a concern [36,37]. To address this issue, we first performed public data mining in early human DKD kidneys at the single-cell and spatial levels. As shown in Supplementary Fig. 1A-C, single nucleus RNA sequencing (GSE131882) showed that MMP-7 and TNC are induced in kidneys from patients with early DKD [21]. MMP-7 was mainly produced by proximal tubules, whereas TNC was distributed in pericytes, endothelium, and podocytes. In a separate histopathology-based analysis of spatial transcriptomics data generated by the 10X Genomics Visium Platform from a human CKD kidney induced by diabetes and hypertension (GSE261545) [23], we also observed significantly increased MMP7 and slightly induced TNC, as shown in Supplementary Fig. 1D. Interestingly, in our serum tandem mass tag (TMT)-labeled proteomics analysis [18], SAA1 and TNC were increased at the early stage of DKD, but SAA1 unexpectedly showed a downregulated trend in patients with T2DM, as presented in Supplementary Fig. 1E. Given the differential expression patterns of these three markers in the diseased kidney and circulation of patients with DKD, we further validated their expression in biopsy specimens from patients with DKD. As shown in Supplementary Fig. 1F, immunohistochemical staining showed that MMP-7 and SAA1 were markedly induced in diseased tubules, whereas TNC was substantially induced in the interstitial compartment.

Global regulation of MMP-7, SAA1, and TNC in multiple organs of db/db mice after kidney damage occurs

Because MMP-7, SAA1, and TNC are secreted proteins, we next evaluated whether they could serve as circulating biomarkers for predicting DKD onset and have global regulatory functions along with DKD progression. Identifying robust circulating markers with multi-organ involvement could significantly enhance early detection and risk stratification of DKD in clinical practice, especially in patients who lack clear symptoms at early stages. To this end, we assessed their expression in multiple target organs in lean controls and db/db mice at early (18 weeks of age) and advanced stages (24 weeks of age) after kidney damage occurred, given the limited availability of human tissue samples from patients with diabetes. As shown in Supplementary Fig. 2A-C, Western blot assays demonstrated that all three biomarkers were upregulated in all T2DM target organs, including the heart, kidney, and liver, in 18- and 24-week-old db/db mice compared with db lean mice. The densitometry data are presented in Supplementary Fig. 2D. These findings support the feasibility of using circulating MMP-7, SAA1, and TNC to predict DKD in the clinic.

DKD patient cohort dynamics and multidimensional assessment

Using our previously established cohorts [18], we added two clinical cohorts to the current study to enhance the capacity of MMP-7, SAA1, and TNC for early DKD prediction. In total, 1,052 serum samples were collected from five independent medical centers, comprising one primary center and four sub-centers (designated as C.S., T.J., N.F., and Z.D.), as illustrated in Fig. 1. Among the sub-centers, the N.F. and Z.D. cohorts were newly recruited. Meanwhile, 34 additional serum samples from healthy adults were collected in the T.J. cohort. In each cohort, four groups of candidates were enrolled, including HCs, patients with T2DM (UACR <30 mg/g), patients with DKD at an early stage (DKD-E, 30≤ UACR ≤300 mg/g), and patients with DKD at an advanced stage (DKD-A, UACR >300 mg/g), according to the albuminuria category classified by the Kidney Disease: Improving Global Outcomes (KDIGO) Diabetes Work Group [38,39]. Of note, all recruited patients with T2DM in this study had been clinically diagnosed for at least 10 years. Participants recruited from the primary center were designated as the discovery cohort, and a total of 617 individuals were included. Meanwhile, 435 participants recruited from the four sub-centers, C.S., T.J., N.F., and Z.D., were designated as independent external testing cohorts. The baseline demographic and clinical characteristics of the participants in the discovery cohort are described in Table 1. The discovery cohort comprised 323 males and 294 females, with ages ranging from 20 to 75 years and body mass index ranging from 18.42 to 39.38 kg/m². As presented in Table 1, there were few differences in HbA1c, blood glucose, TC, TG, LDL, and HDL levels among all enrolled patients. In contrast to patients with T2DM and DKD-E, patients with DKD-A had significantly decreased levels of albumin and eGFR and elevated levels of blood urea nitrogen (BUN), serum creatinine (Scr), and UACR.

Newly developed machine learning algorithms optimize the combination of serum MMP-7, SAA1, and TNC in predicting early DKD

To evaluate the predictive role of these biomarkers in distinguishing DKD stages in the clinic, we first measured serum MMP-7, SAA1, and TNC levels in the discovery cohort using ELISA. As shown in Fig. 2A-C, compared with HCs, MMP-7 gradually increased with DKD progression. Interestingly, as an acute-phase inflammation marker, SAA1 levels were unexpectedly decreased in patients with T2DM, consistent with our proteomics findings presented in Supplementary Fig. 1E. However, its expression remained unchanged during the progression of kidney damage, regardless of whether the disease was at an early or advanced stage. In comparison, TNC exhibited initial upregulation in patients with T2DM, followed by a decrease at the early stage of DKD and a subsequent increase at the advanced stage. Partial least squares regression was performed on these three proteins at different DKD stages. These proteins were able to separate the four groups, as presented in Fig. 2D. Then, we applied these three markers to predict DKD stages. Five-fold CV was performed for all pairwise predictions, and four machine learning algorithms were employed, including LDA, SVM, RF, and logistic regression. As illustrated in the ROC curves in Fig. 2E, the four algorithms demonstrated comparable performance in distinguishing HC vs. T2DM and DKD-E vs. DKD-A. However, for the comparison between T2DM and DKD-E, LDA and logistic regression outperformed SVM and RF. Compared with individual biomarkers, combinations of two or three markers achieved higher AUC and accuracy values in disease prediction. Specifically, combining MMP-7 and SAA1 resulted in predictive accuracies ranging from 0.69 to 0.74 for HC vs. T2DM, 0.68 to 0.74 for T2DM vs. DKD-E, and 0.72 to 0.78 for DKD-E vs. DKD-A. When MMP7 and TNC were combined, accuracy ranged from 0.69 to 0.74 for HC vs. T2DM, 0.71 to 0.77 for T2DM vs. DKD-E, 0.75 to 0.77 for DKD-E vs. DKD-A. In comparison, predictive performance was relatively lower when SAA1 and TNC were combined. Of note, integrating all three biomarkers produced generally robust performance that was equal to or better than that of other combinations across all comparisons: 0.73 to 0.77 for HC vs. T2DM, 0.74 to 0.76 for T2DM vs. DKD-E, and 0.72 to 0.78 for DKD-E vs. DKD-A. Of particular interest, in the comparison between DKD-E and DKD-A, the combination of MMP-7 and SAA1 achieved an AUC of 0.79–0.80 and an accuracy of 0.72–0.78. Adding TNC produced a similar AUC range (0.78 to 0.80) and the same accuracy range (0.72 to 0.78). These findings suggest that, although the current biomarker panel performs reasonably well, the inclusion of additional serum markers may further enhance predictive performance. Pairwise comparisons for the remaining groups, including HC vs. DKD-E, HC vs. DKD-A, and T2DM vs. DKD-A, are presented in Supplementary Fig. 3.

External validations of MMP7, SAA1, and TNC in DKD early prediction

To verify the diagnostic value of these three biomarkers for predicting early-stage DKD, we performed external validation. First, we evaluated the predictive performance of two commonly used clinical markers, BUN and Scr, in distinguishing different DKD stages. ROC analyses were performed for three pairwise group comparisons: T2DM vs. DKD-E, T2DM vs. DKD-A, and DKD-E vs. DKD-A across independent validation cohorts (C.S., T.J., and Z.D.). Interestingly, the results indicated limited utility of BUN and Scr alone for detecting early-stage DKD, although they exhibited stronger performance in identifying advanced kidney dysfunction (Supplementary Fig. 4). The variability across cohorts also suggests the need for more robust, multi-marker approaches for early diagnosis. Therefore, Fig. 3A–D shows serum MMP-7, SAA1, and TNC levels in the recruited candidates at four different external medical centers. Similar to the primary center, MMP-7 showed an increasing trend with DKD progression. In comparison, TNC was induced at the early stage of DKD. Interestingly, SAA1 exhibited inconsistent expression patterns across the four validation cohorts. Then, for each pairwise prediction, classifiers trained from the discovery cohort were applied to these validation cohorts to evaluate model performance. As shown in Fig. 3E–G and Supplementary Fig. 5, the dot plots illustrate the probabilities of RF prediction results for each patient across the discovery and validation cohorts. The dashed-line cutoff indicates the binary split of the prediction. The cutoff values were then applied to predict the early stage of DKD in external validation cohorts. For T2DM vs. DKD-E, as shown in Fig. 3H and Supplementary Table 1, prediction accuracy was 69.4% in the C.S. cohort, 60.9% in the T.J. cohort, 70.4% in the Z.D. cohort. In the N.F. cohort, predictive accuracy increased to as high as 90.5%; this variation might be due to the relatively small DKD-E sample size in this center. In contrast, prediction accuracy reached 82% in the C.S. cohort, 84.1% in the T.J. cohort, 76% in the N.F. cohort, and 85.1% in the Z.D. cohort between T2DM and DKD-A. When DKD-E and DKD-A were compared, prediction accuracy reached 80.5% in the external C.S. cohort, 79.6% in the T.J. cohort, and 67.6% in the N.F. cohort. The detailed results of predictions performed in the four external validation cohorts are presented in Supplementary Table 1 and Supplementary Figs. 6–9. Collectively, the three selected protein biomarkers demonstrated a promising trend toward enhancing early DKD prediction accuracy.

Integrating metabolomics with MMP7, SAA1, and TNC enhances the predictive accuracy for DKD progression

Considering the reliability of serum biomarkers in early DKD diagnosis [40,41], we further mapped the patients whose serum samples were subjected to MMP-7, SAA1, and TNC measurements and metabolomics to enhance predictive power. Serum metabolomics was performed using a Trace 1310 gas chromatograph equipped with an AS 1310 autosampler connected to a TSQ 8000 triple quadrupole mass spectrometer, as previously reported [18]. Detailed information on the identified metabolites is presented in Supplementary Table 2. A total of 526 candidates were enrolled with both metabolomics and ELISA measurements. As illustrated in Fig. 4A, four different machine learning models were developed based on three protein biomarkers with or without serum metabolites. Their prediction results suggested that the combined model improved not only the accuracy but also the stability of DKD diagnosis and prediction. In particular, for DKD-E vs. DKD-A, accuracy increased by approximately 13% with the RF model, although improvement was limited for the comparison between T2DM and DKD-E. Taking the prediction results generated by the RF model as an example, the combination of proteins and metabolites enhanced prediction accuracy in HC vs. T2DM, HC vs. DKD-E, HC vs. DKD-A, and DKD-E vs. DKD-A, but accuracy for T2DM vs. DKD-E was not significantly improved compared with the protein markers alone (Fig. 4B). Given these results, more evidence is needed before noninvasive approaches can replace invasive methods for early identification of kidney damage in T2DM. From the perspective of multi-outcome prediction, the contingency heatmap presented in Fig. 4C indicates the percentage of patients whose true label is shown in the row and whose prediction label is shown in the column. Compared with the multi-outcome prediction model using only the three ELISA markers or only differentially expressed metabolites, the integration model combined both feature sets and achieved the best overall prediction accuracy. Specifically, the integration model successfully identified 85.3% of healthy participants and 73.8% of the patients with DKD-A. Additionally, it showed improved separation among the DM and DKD-E groups. This result indicates a promising future for diabetic status prediction by integrating multi-omics features.

The predictive values of MMP7, SAA1, and TNC in DKD analyzed by linear regression

In real-world applications, several clinical features are used clinically for DKD diagnosis, including blood glucose, FC, HbA1c, UACR, UmALB, and eGFR. Specifically for the discovery cohort in the primary center, these features can predict the status of DKD progression. To check the association of ELISA markers with these clinical indicators, linear regression analysis was performed using the integration of MMP-7, SAA1, and TNC to predict clinical features. As shown in Fig. 5A–D and Supplementary Fig. 10A–F, the values predicted by the three ELISA markers were highly associated with the true indicators, especially UACR and eGFR, indicating that the ELISA measures are consistent with clinical assessments. A well-calibrated model produces points close to the diagonal line, indicating strong agreement between predicted and observed values (Supplementary Fig. 10G). For most features, particularly HbA1c, eGFR, albumin, TG, and blood glucose, predictions aligned reasonably well with observations, supporting the reliability of the model. Larger deviations in variables such as LDL and mALB suggest areas where prediction performance may be more variable. Of note, HbA1c exhibited lower predictive performance than the other routinely used clinical parameters presented in the present study. Finally, we evaluated the predictive performance of each individual clinical feature alone and in combination with the three protein biomarkers using ROC curve analysis. As shown in Fig. 5E and F, combining clinical features including HbA1c, eGFR, BUN, and Scr significantly improved the AUC and accuracy in distinguishing between T2DM and DKD-E, with performance comparable to that of UACR. These findings further support the potential of the three selected protein biomarkers to enhance early-stage DKD prediction beyond conventional clinical parameters.

DISCUSSION

As the leading cause of ESRD, DKD advances more rapidly than kidney damage resulting from primary glomerulonephritis, presenting significant challenges for early, differential, and precise diagnosis [42]. Traditional diagnostic methods, including clinical assessments, urinary protein analysis, renal function tests, imaging, and biopsies, remain the cornerstone of DKD diagnosis [4]. However, the emergence of new technologies and analytical techniques has led to substantial progress in understanding DKD pathogenesis. Integrating molecular data across various dimensions that are closely related to DKD onset and progression, and applying advanced machine learning algorithms have promising potential to enhance diagnostic accuracy [18,43].

In light of these advances, our study focused on three key proteins MMP-7, SAA1, and TNC to evaluate their combined utility in predicting early DKD. Given the limitations of serum samples in early DKD diagnosis, we first confirmed that these proteins were substantially upregulated in the kidneys of patients with early DKD (Supplementary Fig. 1). In a db/db mouse model, we demonstrated that these indicators have global regulatory functions after kidney damage (Supplementary Fig. 2), suggesting their potential as circulating biomarkers. However, it is important to acknowledge that db/db mice do not fully capture the complex pathophysiology of human DKD, particularly in terms of systemic inflammation and immune cell involvement. The mechanistic role of biomarker expression changes in non-renal organs remains unclear and requires further investigation beyond the concept of “multi-organ crosstalk.” Nevertheless, to validate their clinical relevance, we established a discovery cohort and four validation cohorts across five medical centers (Fig. 1). By applying multiple machine learning algorithms, we found that the combined use of these serum markers significantly improved early DKD diagnosis compared with the use of a single marker (Fig. 2). Remarkably, integrating these biomarkers with differential metabolites improved diagnostic accuracy by nearly 13% for detecting DKD progression from early to advanced stages (Fig. 4), offering new targets and diagnostic strategies to prevent rapid DKD progression. Additionally, these biomarkers are consistent with existing clinical assessments (Fig. 5), indicating promising potential for the clinical diagnosis of diabetic diseases.

DKD often progresses silently, with only subtle early symptoms, until irreversible damage has occurred. The disease pathogenesis involves a complex interplay of metabolic, hemodynamic, inflammatory, and fibrotic processes, which collectively contribute to the rapid progression to ESRD in patients with diabetes [44]. MMP-7, a significant biomarker of DKD, is an enzyme crucial for ECM remodeling and tissue repair [15,16]. Its upregulation in response to hyperglycemia and other metabolic disturbances links MMP-7 directly to the underlying mechanisms of DKD [10]. Elevated serum or urinary MMP-7 levels can indicate early kidney injury before significant clinical symptoms or reductions in eGFR [31,45], making it a valuable tool for early diagnosis and risk stratification in patients with T2DM. Our study corroborates this, showing a gradual increase in MMP-7 levels with DKD progression across all cohorts, underscoring its reliability as a stable circulating marker (Figs. 2 and 3). However, the specificity of MMP-7 as a single biomarker is limited, given its role as a surrogate marker for non-diabetic CKD [16]. Of note, in addition to MMP-7, other members of the MMP family, such as MMP-10, have also been reported as promising biomarkers for early DKD detection [46].

The ECM undergoes significant remodeling in DKD, characterized by increased production and decreased degradation of components such as collagen and fibronectin, leading to ECM accumulation, glomerular basement membrane thickening, and interstitial fibrosis [47,48]. TNC, a glycoprotein involved in inflammation and fibrosis [13,14,49], is upregulated in response to chronic hyperglycemia in most cohorts, except in some cases in which it exhibited a downregulated trend in patients with DKD-E compared patients with T2DM (Figs. 2 and 3). Given that TNC is often localized at injury and fibrosis sites [14], its potential as a strong single circulating marker for early DKD prediction warrants further investigation. Additionally, serum TNC has been associated with increased cardiovascular events and mortality in individuals with T2DM [12].

SAA1, another biomarker selected for this study, is an acute-phase inflammatory protein primarily produced by the liver in response to pro-inflammatory cytokines [50], with potential local synthesis in the kidneys. The expression pattern of SAA1, showing downregulation in the sera of patients with T2DM and upregulation after kidney damage, underscores its relevance to DKD. Despite its utility in detecting early kidney dysfunction, our cohort studies revealed inconsistent SAA1 expression patterns across different medical centers. Previous studies have reported elevated SAA1 levels in patients with diabetes and kidney complications [11], but some data suggest that higher circulating SAA concentrations may be associated with a reduced risk of ESRD in American Indians with T2DM [11]. This variability highlights the challenges of using SAA1 as a single biomarker for early DKD prediction. Of note, among the three ELISA-based markers, SAA1 did not show significant differences across disease groups, particularly between DKD-E and DKD-A (P=0.649). To achieve adequate statistical power for detecting potential group differences, a larger sample size will be considered in future studies.

Although individual biomarkers such as MMP-7, SAA1, and TNC have shown promise in DKD diagnosis [10–12,31,32,35], their limitations in accuracy and specificity as standalone markers are evident. This study emphasizes the synergistic effects of combining these serum biomarkers to enhance early DKD prediction. Our analysis demonstrates that combining two or three biomarkers consistently outperforms the use of a single marker, reducing systematic errors and improving diagnostic performance. Among the combinations, integrating TNC with SAA1 yielded the least favorable results because of their variable serum expression in patients with DKD. The inflammatory status of individuals with DKD, especially with respect to SAA1, is challenging to evaluate, even under strict glycemic control.

To further reduce systematic errors, we integrated MMP-7, SAA1, and TNC with differentially expressed serum metabolites identified in our previous study [18]. Serum metabolomics, which analyzes metabolites in the blood, provides insights into early metabolic changes, tracks metabolite alterations, and reflects real-time kidney dysfunction. Specific metabolites associated with glucose metabolism, lipid metabolism, amino acid metabolism, oxidative stress, inflammation, and kidney function can be detected in serum, providing early indicators of renal injury. This allows for more personalized and timely adjustments to management strategies. By analyzing changes in metabolite levels, clinicians can gain insights into the functional impact of DKD on various biological pathways and organs. Combining metabolomic data with clinical and biochemical information provides a more comprehensive view of the patient’s condition, supporting better clinical decision-making. This integration significantly improved the diagnostic accuracy for distinguishing early and advanced DKD by nearly 13% (Fig. 4). However, predicting early kidney damage remains challenging, and the functional association between specific endogenous metabolites and DKD progression requires further exploration. In this study, we used all differentially expressed serum metabolites for integrative analysis, which have led to underestimation of their diagnostic value. Additionally, TNC, MMP-7, and SAA1 are functionally related to biological processes such as tubular injury, ECM remodeling, and inflammation, which overlap with the pathways captured by established markers such as kidney injury molecule-1 (KIM-1), neutrophil gelatinase-associated lipocalin (NGAL), tumor necrosis factor receptors (TNF-Rs), and soluble urokinase plasminogen activator receptor (suPAR). For example, MMP-7 has been implicated in tubular epithelial cell injury similar to KIM-1, while SAA1, like TNF-Rs and suPAR, reflects systemic and renal inflammation, suggesting potential complementarity in multi-marker diagnostic models for DKD.

Despite these advances, our study has several limitations: (1) cross-platform and cross-species integration assumes conserved mechanisms, but differences in tissue context and disease dynamics may limit direct translatability; (2) urine samples were not included; (3) correlation analysis between biomarkers/metabolites and DKD pathology is lacking; (4) the cohorts are regional, necessitating validation in broader populations with varying diets and lifestyles; and (5) unaddressed confounding factors such as medications, comorbidities, and ethnicity may influence biomarker levels and disease progression and could not be fully accounted for because of the retrospective design. In addition, the sample size for the metabolomics subgroup was relatively small. In total, 133 healthy, 276 T2DM, 49 DKD-E, and 68 DKD-A serum samples had both ELISA and metabolomics markers measured. For each ELISA marker, three comparisons were made: HC vs. T2DM, T2DM vs. DKD-E, and DKD-E vs. DKD-A. To control the FDR at 10%, the Bonferroni method was used to adjust the probability of type I error per pairwise test to be 0.1/3=0.033. Take MMP-7 HC vs. DM as an example, achieving power=0.8 and a probability of type I error=0.033 requires at least n=23 samples per group using the non-parametric Mann-Whitney U or Wilcoxon rank-sum test. Our 276 T2DM and 49 DKD-E individuals met this power requirement. Similar calculations were performed across all three markers for all three pairwise comparisons. The results showed that both MMP-7 and TNC (HC vs. T2DM and T2DM vs. DKD-E) met the power requirement, but involving SAA1 and TNC for DKD-E vs. DKD-A were underpowered. For instance, when comparing SAA1 between DKD-E and DKD-A, achieving power=0.8 and a probability of type I error=0.1/3=0.033 requires at least n=15,635 individuals per group. More individuals are expected to be collected in future studies.

In summary, our study demonstrates the potential of combining advanced technologies and algorithms to improve early DKD diagnosis. Although precise pathology remains critical, the synergy between selected serum biomarkers and metabolites offers substantial promise for transforming the clinical management of DKD. By focusing on early detection and personalized intervention, we may be able to alter the disease course, identify subtle kidney damage before clinical symptoms emerge, reduce the DKD burden, and improve quality of life for millions of patients with diabetes worldwide.

SUPPLEMENTARY MATERIALS

NOTES

CONFLICTS OF INTEREST

No potential conflict of interest relevant to this article was reported.

AUTHOR CONTRIBUTIONS

Conception or design: D.Z., H.F.

Acquisition, analysis, or interpretation of data: all authors.

Drafting the work or revising: S.L. (Silvia Liu), D.Z., H.F.

Final approval of the manuscript: all authors.

FUNDING

This is an investigator-initiated clinical study. Work in the FH’s lab was supported by the National Key R&D Program of China grant (2022YFC2502504, 2022YFC2502500), the National Natural Science Foundation of China grants (92268112, 82570872), and Guangdong Basic and Applied Basic Research Foundation (2023A1515012389).

ACKNOWLEDGMENTS

We are grateful to the volunteered staff who assisted us in collecting serum samples at the five included medical centers.

DATA AVAILABILITY

Fig. 1

Diabetic kidney disease (DKD) patient cohort dynamics and multidimensional assessment. A total of 1,052 participants were enrolled in five clinical cohorts from five independent medical centers, including 266 healthy controls (HCs), 441 patients with type 2 diabetes mellitus (T2DM; urinary albumin-to-creatinine ratio [UACR] <30 mg/g), 183 patients with diabetic kidney disease at an early stage (DKD-E; 30≤ UACR ≤300 mg/g), and 162 patients with diabetic kidney disease at an advanced stage (DKD-A; UACR >300 mg/g) in both discovery and testing cohorts. Integrative analysis combining serum protein biomarkers and metabolites was performed. SAA1, serum amyloid A1; TNC, tenascin C; MMP-7, matrix metallopeptidase-7; TMT, tandem mass tags; ELISA, enzyme-linked immunosorbent assay; C.S., the 3rd Xiangya Hospital of Central South University; T.J., Tongji Hospital of Tongji University; N.F., Nanfang Hospital; Z.D., Zhongda Hospital.

Fig. 2

Newly developed machine learning algorithms optimize the combination of serum matrix metallopeptidase-7 (MMP-7), serum amyloid A1 (SAA1), and tenascin C (TNC) for predicting early diabetic kidney disease (DKD). (A–C) Box plots showing the concentrations of MMP-7 (A), SAA1 (B), and TNC (C) in sera from healthy adults and patients with DKD at different stages. (D) Partial least squares discriminant analysis (PLS-DA) based on enzyme-linked immunosorbent assay (ELISA) markers. Four groups are included: healthy control (HC), type 2 diabetes mellitus (T2DM; urinary albumin-to-creatinine ratio [UACR] <30 mg/g), diabetic kidney disease at an early stage (DKD-E; 30≤ UACR ≤300 mg/g), and diabetic kidney disease at an advanced stage (DKD-A; UACR >300 mg/g). (E) Using a single maker or multiple markers, receiver operating characteristic (ROC) curves for each pairwise prediction, including HC vs. T2DM, T2DM vs. DKD-E, and DKD-E vs. DKD-A. Four different machine learning classifiers were employed: red line for linear discriminant analysis (LDA), blue line for support vector machine (SVM), orange line for random forest (RF), and green line for logistic regression. AUC, area under the curve; Acc, accuracy.

Fig. 3

Prediction of diabetic kidney disease (DKD) status using machine learning algorithms based on metallopeptidase-7 (MMP-7), serum amyloid A1 (SAA1), and tenascin C (TNC) expression. (A–D) Box plots showing the concentrations of MMP-7, SAA1, and TNC in sera from healthy adults and patients with DKD at various stages across different clinical cohorts, including the 3rd Xiangya Hospital of Central South University (C.S.) (A), Tongji Hospital of Tongji University (T.J.) (B), Nanfang Hospital (N.F.) (C), and Zhongda Hospital (Z.D.) (D). (E–G) Prediction probabilities based on the random forest algorithm. The best cutoff was trained from the discovery cohort and applied to validation cohorts to predict healthy control (HC) vs. type 2 diabetes mellitus (T2DM), T2DM vs. diabetic kidney disease at an early stage (DKD-E), and DKD-E vs. diabetic kidney disease at an advanced stage (DKD-A). (H) Receiver operating characteristic (ROC) curves for each pairwise prediction using four different machine learning methods in validation cohorts. Four different machine learning classifiers were employed: red line for linear discriminant analysis (LDA), blue line for support vector machine (SVM), orange line for random forest (RF), and green line for logistic regression. AUC, area under the curve.

Fig. 4

Integrative analysis of serum protein biomarkers and metabolites. (A) Integrating metallopeptidase-7 (MMP-7), serum amyloid A1 (SAA1), and tenascin C (TNC) with serum metabolites: receiver operating characteristic (ROC) curves for each pairwise prediction using four different machine learning methods in the discovery cohort. (B) Pairwise prediction accuracy based on protein features only (blue), metabolite features only (green), and integration of protein and metabolite features (orange). (C) Heatmap for multi-outcome prediction accuracy based on protein features, metabolite features, and both feature sets. The numbers in the heatmap cells represent the percentage of cases predicted as the labels on the y-axis among all true labels on the x-axis. AUC, area under the curve; Acc, accuracy; LDA, linear discriminant analysis; SVM, support vector machine; RF, random forest; HC, healthy control; T2DM, type 2 diabetes mellitus; DKD-E, diabetic kidney disease at an early stage; DKD-A, diabetic kidney disease at an advanced stage; ELISA, enzyme-linked immunosorbent assay.

Fig. 5

Linear regression to analyze the predictive values of metallopeptidase-7 (MMP-7), serum amyloid A1 (SAA1), and tenascin C (TNC) in diabetic kidney disease (DKD). Linear regression analysis for glycosylated hemoglobin (HbA1c) (A), estimated glomerular filtration rate (eGFR) (B), urinary albumin-to-creatinine ratio (UACR) (C), and urinary microalbumin (UmALB) (D). Each dot represents a patient, with their real clinical indicators shown on the y-axis and the predicted values by the three enzyme-linked immunosorbent assay (ELISA) markers shown on the x-axis. (E, F) Using a single clinical maker including HbA1c, eGFR, blood urea nitrogen (BUN), serum creatinine (Scr), and UACR (E) or respectively combine each of them with MMP7, SAA1, and TNC (F), receiver operating characteristic (ROC) curves for each pairwise prediction, including type 2 diabetes mellitus (T2DM) vs. diabetic kidney disease at early stage (DKD-E), T2DM vs. diabetic kidney disease at advanced stage (DKD-A), and DKD-E vs. DKD-A. Four different machine learning classifiers were employed: redline for linear discriminant analysis (LDA), blue line for support vector machine (SVM), orange line for random forest (RF), and green line for logistic regression. AUC, area under the curve; Acc, accuracy.

REFERENCES

• 1. Saeedi P, Petersohn I, Salpea P, Malanda B, Karuranga S, Unwin N, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: results from the International Diabetes Federation Diabetes Atlas, 9th edition. Diabetes Res Clin Pract 2019;157:107843.ArticlePubMedPMC
• 2. Alicic RZ, Rooney MT, Tuttle KR. Diabetic kidney disease: challenges, progress, and possibilities. Clin J Am Soc Nephrol 2017;12:2032-45.PubMedPMC
• 3. Agarwal R, Pitt B, Rossing P, Anker SD, Filippatos G, Ruilope LM, et al. Modifiability of composite cardiovascular risk associated with chronic kidney disease in type 2 diabetes with finerenone. JAMA Cardiol 2023;8:732-41.ArticlePubMedPMC
• 4. Thomas MC, Brownlee M, Susztak K, Sharma K, Jandeleit-Dahm KA, Zoungas S, et al. Diabetic kidney disease. Nat Rev Dis Primers 2015;1:15018.PubMedPMC
• 5. Sugahara M, Pak WL, Tanaka T, Tang SC, Nangaku M. Update on diagnosis, pathophysiology, and management of diabetic kidney disease. Nephrology (Carlton) 2021;26:491-500.ArticlePubMedPDF
• 6. Persson F, Rossing P. Diagnosis of diabetic kidney disease: state of the art and future perspective. Kidney Int Suppl (2011) 2018;8:2-7.ArticlePubMedPMC
• 7. Rico-Fontalvo J, Aroca-Martinez G, Daza-Arnedo R, Cabrales J, Rodriguez-Yanez T, Cardona-Blanco M, et al. Novel biomarkers of diabetic kidney disease. Biomolecules 2023;13:633.ArticlePubMedPMC
• 8. Barutta F, Bellini S, Canepa S, Durazzo M, Gruden G. Novel biomarkers of diabetic kidney disease: current status and potential clinical application. Acta Diabetol 2021;58:819-30.ArticlePubMedPDF
• 9. Colhoun HM, Marcovecchio ML. Biomarkers of diabetic kidney disease. Diabetologia 2018;61:996-1011.ArticlePubMedPMCPDF
• 10. Hirohama D, Abedini A, Moon S, Surapaneni A, Dillon ST, Vassalotti A, et al. Unbiased human kidney tissue proteomics identifies matrix metalloproteinase 7 as a kidney disease biomarker. J Am Soc Nephrol 2023;34:1279-91.ArticlePubMedPMC
• 11. Saulnier PJ, Dieter BP, Tanamas SK, McPherson SM, Wheelock KM, Knowler WC, et al. Association of serum amyloid a with kidney outcomes and all-cause mortality in American Indians with type 2 diabetes. Am J Nephrol 2017;46:276-84.ArticlePubMedPMCPDF
• 12. Gellen B, Thorin-Trescases N, Thorin E, Gand E, Sosner P, Brishoual S, et al. Serum tenascin-C is independently associated with increased major adverse cardiovascular events and death in individuals with type 2 diabetes: a French prospective cohort. Diabetologia 2020;63:915-23.ArticlePubMedPDF
• 13. Chen S, Fu H, Wu S, Zhu W, Liao J, Hong X, et al. Tenascin-C protects against acute kidney injury by recruiting Wnt ligands. Kidney Int 2019;95:62-74.ArticlePubMedPMC
• 14. Fu H, Tian Y, Zhou L, Zhou D, Tan RJ, Stolz DB, et al. Tenascin-C is a major component of the fibrogenic niche in kidney fibrosis. J Am Soc Nephrol 2017;28:785-801.ArticlePubMedPMC
• 15. Fu H, Zhou D, Zhu H, Liao J, Lin L, Hong X, et al. Matrix metalloproteinase-7 protects against acute kidney injury by priming renal tubules for survival and regeneration. Kidney Int 2019;95:1167-80.ArticlePubMedPMC
• 16. Zhou D, Tian Y, Sun L, Zhou L, Xiao L, Tan RJ, et al. Matrix metalloproteinase-7 is a urinary biomarker and pathogenic mediator of kidney fibrosis. J Am Soc Nephrol 2017;28:598-611.ArticlePubMedPMC
• 17. Zhu H, Liao J, Zhou X, Hong X, Song D, Hou FF, et al. Tenascin-C promotes acute kidney injury to chronic kidney disease progression by impairing tubular integrity via αvβ6 integrin signaling. Kidney Int 2020;97:1017-31.ArticlePubMedPMC
• 18. Liu S, Gui Y, Wang MS, Zhang L, Xu T, Pan Y, et al. Serum integrative omics reveals the landscape of human diabetic kidney disease. Mol Metab 2021;54:101367.ArticlePubMedPMC
• 19. Zhang H, Zuo JJ, Dong SS, Lan Y, Wu CW, Mao GY, et al. Identification of potential serum metabolic biomarkers of diabetic kidney disease: a widely targeted metabolomics study. J Diabetes Res 2020;2020:3049098.ArticlePubMedPMCPDF
• 20. Sharma K, Karl B, Mathew AV, Gangoiti JA, Wassel CL, Saito R, et al. Metabolomics reveals signature of mitochondrial dysfunction in diabetic kidney disease. J Am Soc Nephrol 2013;24:1901-12.ArticlePubMedPMC
• 21. Wilson PC, Wu H, Kirita Y, Uchimura K, Ledru N, Rennke HG, et al. The single-cell transcriptomic landscape of early human diabetic nephropathy. Proc Natl Acad Sci U S A 2019;116:19619-25.ArticlePubMedPMC
• 22. Muto Y, Wilson PC, Ledru N, Wu H, Dimke H, Waikar SS, et al. Single cell transcriptional and chromatin accessibility profiling redefine cellular heterogeneity in the adult human kidney. Nat Commun 2021;12:2190.ArticlePubMedPMCPDF
• 23. Isnard P, Li D, Xuanyuan Q, Wu H, Humphreys BD. Histopathologic analysis of human kidney spatial transcriptomics data: toward precision pathology. Am J Pathol 2025;195:69-88.PubMed
• 24. Hao Y, Hao S, Andersen-Nissen E, Mauck WM, Zheng S, Butler A, et al. Integrated analysis of multimodal single-cell data. Cell 2021;184:3573-87e29.ArticlePubMedPMC
• 25. Stuart T, Butler A, Hoffman P, Hafemeister C, Papalexi E, Mauck WM, et al. Comprehensive integration of single-cell data. Cell 2019;177:1888-902e21.ArticlePubMedPMC
• 26. Gui Y, Palanza Z, Gupta P, Li H, Pan Y, Wang Y, et al. Calponin 2 regulates ketogenesis to mitigate acute kidney injury. JCI Insight 2023;8:e170521.ArticlePubMedPMC
• 27. Gui Y, Fu H, Palanza Z, Tao J, Lin YH, Min W, et al. Fibroblast expression of transmembrane protein smoothened governs microenvironment characteristics after acute kidney injury. J Clin Invest 2024;134:e165836.ArticlePubMedPMC
• 28. Xie HH, Xu JY, Xie T, Meng X, Lin LL, He LL, et al. Effects of Pinellia ternata (Thunb.) Berit. on the metabolomic profiles of placenta and amniotic fluid in pregnant rats. J Ethnopharmacol 2016;183:38-45.ArticlePubMed
• 29. Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez JC, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 2011;12:77.ArticlePubMedPMCPDF
• 30. Bishop CM. Pattern recognition and machine learning. New York: Springer; 2006.
• 31. Ban CR, Twigg SM, Franjic B, Brooks BA, Celermajer D, Yue DK, et al. Serum MMP-7 is increased in diabetic renal disease and diabetic diastolic dysfunction. Diabetes Res Clin Pract 2010;87:335-41.ArticlePubMed
• 32. Afkarian M, Zelnick LR, Ruzinski J, Kestenbaum B, Himmelfarb J, de Boer IH, et al. Urine matrix metalloproteinase-7 and risk of kidney disease progression and mortality in type 2 diabetes. J Diabetes Complications 2015;29:1024-31.ArticlePubMedPMC
• 33. Zhou Y, Ma XY, Han JY, Yang M, Lv C, Shao Y, et al. Metformin regulates inflammation and fibrosis in diabetic kidney disease through TNC/TLR4/NF-κB/miR-155-5p inflammatory loop. World J Diabetes 2021;12:19-46.ArticlePubMedPMC
• 34. Dieter BP, McPherson SM, Afkarian M, de Boer IH, Mehrotra R, Short R, et al. Serum amyloid a and risk of death and end-stage renal disease in diabetic kidney disease. J Diabetes Complications 2016;30:1467-72.ArticlePubMedPMC
• 35. Anderberg RJ, Meek RL, Hudkins KL, Cooney SK, Alpers CE, Leboeuf RC, et al. Serum amyloid A and inflammation in diabetic kidney disease and podocytes. Lab Invest 2015;95:250-62.ArticlePubMedPMCPDF
• 36. Al-Rubeaan K, Siddiqui K, Al-Ghonaim MA, Youssef AM, Al-Sharqawi AH, AlNaqeb D. Assessment of the diagnostic value of different biomarkers in relation to various stages of diabetic nephropathy in type 2 diabetic patients. Sci Rep 2017;7:2684.ArticlePubMedPMCPDF
• 37. Vucic Lovrencic M, Bozicevic S, Smircic Duvnjak L. Diagnostic challenges of diabetic kidney disease. Biochem Med (Zagreb) 2023;33:030501.PubMedPMC
• 38. Kidney Disease: Improving Global Outcomes (KDIGO) Diabetes Work Group. KDIGO 2020 clinical practice guideline for diabetes management in chronic kidney disease. Kidney Int 2020;98(4S):S1-115.ArticlePubMed
• 39. Mogensen CE, Christensen CK, Vittinghus E. The stages in diabetic renal disease: with emphasis on the stage of incipient diabetic nephropathy. Diabetes 1983;32(Suppl 2):64-78.ArticlePubMedPDF
• 40. Sauriasari R, Safitri DD, Azmi NU. Current updates on protein as biomarkers for diabetic kidney disease: a systematic review. Ther Adv Endocrinol Metab 2021;12:20420188211049612.ArticlePubMedPMCPDF
• 41. Jung CY, Yoo TH. Novel biomarkers for diabetic kidney disease. Kidney Res Clin Pract 2022;41(Suppl 2):S46-62.ArticlePubMedPMCPDF
• 42. Fu H, Liu S, Bastacky SI, Wang X, Tian XJ, Zhou D. Diabetic kidney diseases revisited: a new perspective for a new era. Mol Metab 2019;30:250-63.ArticlePubMedPMC
• 43. Jiang X, Liu X, Qu X, Zhu P, Wo F, Xu X, et al. Integration of metabolomics and peptidomics reveals distinct molecular landscape of human diabetic kidney disease. Theranostics 2023;13:3188-203.ArticlePubMedPMC
• 44. Mohandes S, Doke T, Hu H, Mukhi D, Dhillon P, Susztak K. Molecular pathways that drive diabetic kidney disease. J Clin Invest 2023;133:e165654.ArticlePubMedPMC
• 45. Sarangi R, Sahu D, Rout NK, Padarabinda Tripathy K, Patra S, Bahinipati J, et al. Role of urinary matrix metalloproteinase-7 (MMP-7) as an early marker of renal dysfunction in diabetic individuals: a cross-sectional study. Cureus 2024;16:e66392.Article
• 46. Mora-Gutierrez JM, Rodriguez JA, Fernandez-Seara MA, Orbe J, Escalada FJ, Soler MJ, et al. MMP-10 is increased in early stage diabetic kidney disease and can be reduced by renin-angiotensin system blockade. Sci Rep 2020;10:26.PubMedPMC
• 47. Mason RM, Wahab NA. Extracellular matrix metabolism in diabetic nephropathy. J Am Soc Nephrol 2003;14:1358-73.ArticlePubMed
• 48. Kolset SO, Reinholt FP, Jenssen T. Diabetic nephropathy and extracellular matrix. J Histochem Cytochem 2012;60:976-86.ArticlePubMedPMCPDF
• 49. Ozanne J, Shek B, Stephen LA, Novak A, Milne E, Mclachlan G, et al. Tenascin-C is a driver of inflammation in the DSS model of colitis. Matrix Biol Plus 2022;14:100112.ArticlePubMedPMC
• 50. Jiang B, Wang D, Hu Y, Li W, Liu F, Zhu X, et al. Serum amyloid A1 exacerbates hepatic steatosis via TLR4-mediated NF-κB signaling pathway. Mol Metab 2022;59:101462.ArticlePubMedPMC

Figure & Data

References

Citations

Citations to this article as recorded by  

• Recent advances in point-of-care colorimetric biosensors for detecting small molecule metabolites
  Zhiqiang Zhu, Zhun Gu, Xinxing Cao, Shanshan Zhang, Shao Su
  Chemical Communications.2026; 62(48): 12000.     CrossRef

About this article

Zhou X, Gui Y, Liu JJ, Liu S, Liang D, Wang Y, Shaffer HW, Mallari SM, Jones C, Gupta P, Li D, Zhang K, Yu Y, Tao J, Wang Y, Liu S, Zhou D, Fu H. Optimizing Early Detection of Diabetic Kidney Disease through Synergistic Biomarkers and Serum Metabolites in Humans. Diabetes Metab J. 2026;50(4):752-769.

Received: Mar 08, 2025; Accepted: Sep 16, 2025

DOI: https://doi.org/10.4093/dmj.2025.0193.

Download citation ↓