Introduction

Diabetic nephropathy (DN) is one of the major complications of diabetes mellitus (DM) and has become the most prevalent cause of end-stage renal disease worldwide [1]. DN is also one of the most significant long-term diseases in terms of morbidity and mortality for individuals with diabetes [2]. Recent studies have shown that several interventions can slow the progression of DN, and their impact is greater if they are started at an early stage of the development of nephropathy [3]. Although renal biopsy is the most accurate diagnostic method for DN, routine renal biopsies are not acceptable in current clinical practice because of their invasiveness. Microalbuminuria is an alternative, non-invasive marker that can be used for DN risk assessment, and the urinary albumin-to-creatinine ratio (UACR) on first-void urine sample is recommended for DN screening. However, large prospective studies have revealed poor accuracy of this marker, even though urine samples are collected two or three times a day to normalize day-to-day variation [4]. Therefore, identifying reliable and versatile biomarkers for risk assessment of DN is important.

Mass spectrometry-based urinary proteomics is used for biomarker discoveries of DN. Dihazi et al. used surface-enhanced laser desorption/ionization time-of-flight mass spectrometry to identify three urinary proteins that differentiated patients with DN from patients with type 2 DM without nephropathy, patients with type 2 DM with micro- or macroalbuminuria, patients with proteinuria caused by non-diabetic renal disease, and healthy controls [5]. Mischak et al. profiled urinary polypeptides using capillary electrophoresis-mass spectrometry (CE-MS) and found that the MS patterns could be used to differentiate type 2 DM from healthy controls [6]. However, urinary protein markers sometimes show a wider variation than blood samples. Thus, it is necessary to discover biomarkers with small diurnal variations.

Metabolomics is the comprehensive analysis of low weight molecules in a sample, and has become a powerful tool in the biomarker discovery field. Nuclear magnetic resonance [7], gas chromatography-mass spectrometry [8], liquid chromatography-mass spectrometry (LC-MS; [9]), and CE-MS [10–12] are currently used for metabolomics. Targeted profiling, that is, detection of only a few sets of metabolites, has been used to discover biomarkers for DN. Xia et al. analyzed six intermediate metabolites of the purine degradation pathway in plasma from patients with non-DN and DN using LC with or without MS [13]. They found that adenosine, inosine, uric acid, and xanthine were useful biomarkers for monitoring DM progression. Jiang et al. used LC-tandem mass spectrometry to simultaneously quantify eight aminothiols in the homocysteine metabolic cycle in plasma and found two sulfur-containing metabolites, S-adenosylmethionine and S-adenosylhomocysteine, as potential biomarkers for DM and DN [14].

Compared to targeted profiling, comprehensive metabolome analysis of all metabolites in the given sample is a more powerful technique. Zhang et al. used non-targeted LC-MS to detect potential biomarker candidates of DN and type 2 DM, and observed significant differences in the serum levels of leucine, dihydrosphingosine, and phytosphingosine [15]. However, there are few published comprehensive metabolome profiles of DN.

Recently, we developed a non-targeted CE-MS-based metabolome profiling technique [11, 16] and applied it to biomarker discovery for acetaminophen-induced hepatotoxicity in mice [11] and several types of cancer-specific profiles in human saliva [12]. In the present study, we used CE-MS to identify serum metabolite biomarkers for DN diagnosis. Furthermore, classification models incorporating multiple biomarkers were constructed for discriminating DN from non-DN.

Materials and methods

Sample collection and metabolite extraction

All experiments were conducted in accordance with study protocol approved by the Institutional Ethics Committee of Chubu Rosai Hospital. Informed consent was obtained from all patients according to the Declaration of Helsinki as revised in 2000. Serum samples from 78 type 2 DM patients were collected and classified into the following three groups: DM group without nephropathy and albuminuria (non-DN, UACR < 30 mg/g, n = 20), early DN group with microalbuminuria (micro-DN, 30 < UACR < 300 mg/g, n = 32), and overt DN group with macroalbuminuria (macro-DN, UACR > 300 mg/g, n = 26). All serum samples were stored at −80 °C.

To extract metabolites, the frozen sera samples were thawed and 100 μl aliquots were put into 900 μl of methanol that contained internal standards (20 μmol/l each of methionine sulfone and camphor 10-sulfonic acid). The internal standards were used to normalize the extraction efficiency of metabolites during sample preparation for both cationic (methionine sulfone) and anionic (camphor 10-sulfonic acid) metabolite analysis. The solutions were mixed well and then 400 μl of Milli-Q water and 1 ml of chloroform were added, followed by centrifugation at 4,600×g for 5 min at 4 °C. The aqueous layer was transferred to a 5-kDa cutoff centrifugal filter tube (Millipore, Billerica, MA, USA) to remove large molecules. The filtrate was centrifugally concentrated at 35 °C and reconstituted with 50 μl of Milli-Q water that contained reference compounds (200 μmol/l each of 3-aminopyrrolidine and trimesic acid) immediately before CE-TOFMS analysis. These reference compounds were added to eliminate the variation in migration time of individual peaks in electropherogram among multiple datasets.

Reagents

Methionine sulfone (internal standard) was purchased from Alfa Aesar (Ward Hill, MA), and hexakis-(2,2-difluoroethoxy)-phosphazene (Hexakis) from SynQuest Laboratories (Alachua, FL). All other reagents were obtained from Sigma-Aldrich (St. Louis, MO) or Wako Pure Chemicals Industries Ltd. (Osaka, Japan). All chemicals used were of analytical or reagent grade. Water was purified with a Milli-Q purification system (Millipore, Billerica, MA).

Instruments

All CE-electrospray ionization (ESI)-TOFMS experiments were performed using an Agilent CE capillary electrophoresis system (Agilent Technologies, Waldbronn, Germany), an Agilent G3250AA LC/MSD TOF system (Agilent Technologies, Palo Alto, CA, USA), an Agilent 1100 series isocratic HPLC pump, a G1603A Agilent CE-MS adapter kit, and a G1607A Agilent CE-ESI-MS sprayer kit. The CE-MS adapter kit included a capillary cassette, which facilitated thermostating of the capillary, and the CE-ESI-MS sprayer kit, which simplified coupling of the CE system with the MS system, was equipped with an electrospray source. For system control and data acquisition, G2201AA Agilent Chemstation software was used for CE, and Agilent TOF (Analyst QS) software was used for TOFMS. The original Agilent SST316Ti stainless steel (Fe/Cr/Ni/Mo/Ti; 68:18:11:2:1) ESI needle was replaced with a platinum needle for anion analysis [17]. The resolution of the TOFMS instrument used in this study is higher than 3,000 at m/z 100 with high mass accuracy (<3 ppm).

CE-TOFMS analysis of cationic metabolites

CE-TOFMS analysis of cationic metabolites was performed as described previously [10]. Cationic metabolites were separated in a fused-silica capillary (50 μm i.d. × 100 cm total length) filled with 1 mol/l formic acid as the reference electrolyte. The sample solution was injected at 5 kPa for 3 s (approximately 3 nl), and a positive voltage of 30 kV was applied. The capillary and sample trays were maintained at 20 °C and <5 °C, respectively. The sheath liquid was methanol/water (50 % v/v) containing 0.1 μmol/l Hexakis and was delivered at 10 μl/min. ESI-TOFMS was operated in positive ion mode. The capillary voltage was set at 4 kV, and the nitrogen gas (heater temperature 300 °C) flow rate was set at 10 l/min. In TOFMS, the fragmenter voltage, skimmer voltage, and octapole radio frequency voltage were set at 75, 50, and 125 V, respectively. An automatic recalibration function was performed using the following masses of two reference standards: [13C isotopic ion of the protonated methanol dimer (2MeOH + H)]+, m/z 66.06306; and [protonated Hexakis (M + H)]+, m/z 622.02896. Mass spectra were acquired at a rate of 1.5 cycles per second from m/z 50 to 1000.

CE-TOFMS analysis of anionic metabolites

The CE-TOFMS analysis of anionic metabolites was performed as described previously [17]. Anionic metabolites were separated in a commercially available COSMO(+) capillary, which was chemically coated with a cationic polymer. Ammonium acetate solution (50 mmol/l, pH 8.5) was used as the electrolyte for CE separation. The sample solution was injected at 5 kPa for 30 s (approximately 30 nl) and a voltage of −30 kV was applied. Ammonium acetate (5 mmol/l) in methanol/water (50 % v/v) containing 0.1 μmol/l Hexakis was delivered as the sheath liquid at 10 μl/min. ESI-TOFMS was operated in negative ion mode. The capillary voltage was set at 3.5 kV. In TOFMS, the fragmenter voltage, skimmer voltage, and octapole radio frequency voltage were set at 100, 50, and 200 V, respectively. An automatic recalibration function was performed using the following masses of two reference standards: [13C isotopic ion of deprotonated acetic acid dimer (2CH3COOH-H)]–, m/z 120.03834; and [Hexakis + deprotonated acetic acid (M + CH3COOH-H)]–, m/z 680.03554. Mass spectra were acquired at a rate of 1.5 cycles per second from m/z 50 to 1,000.

Data processing

The raw data were processed using our proprietary software (MasterHands) [10, 12]. The overall data processing flow consisted of noise filtering, baseline correction, peak detection, and integration of the peak areas from 0.02 m/z-wide sections of the electropherograms. Subsequently, the accurate m/z of each peak was calculated by Gaussian curve fitting in the m/z domain, and the migration times were normalized to match the detected peaks among the multiple datasets. The peaks were identified by matching m/z values and normalized migration times of corresponding authentic standard compounds. Processed peak lists were exported for further statistical analysis.

Statistical analysis

The relative ratio of the detected peak area to that of the internal standard was used to eliminate systematic bias derived from injection volume variance and MS sensitivity. Data were analyzed with GraphPad Prism 5.0 (GraphPad Software, Inc., San Diego, CA, USA) for statistical tests. The Kruskal–Wallis test and Dunn’s post test were used to assess the statistical significance of differences among non-DN, micro-DN and macro-DN samples. The Spearman’s rank correlation test was used to calculate correlations among UACR, eGFR, and the relative ratios of peak areas of the metabolites. Multiple logistic regression (MLR) models were developed to discriminate non-DN and DN cohorts. Biomarker metabolites for these models were selected in two procedures. First, normalized data were subjected to orthogonal partial least-squares discriminant analysis (OPLS-DA) using SIMCA-P + software (Version 12.0, Umetrics, Umeå, Sweden), and a model was built and used to identify marker metabolites that accounted for differentiation of non-DN and macro-DN cohorts. Next, a stepwise variable selection method (forward and backward selection) was conducted with a threshold of p < 0.25 for adding and eliminating features using JMP 8.0 (SAS Institute Inc., Cary, NC, USA). The generalization ability of the developed MLR model was evaluated using cross-validation methods. Tenfold cross-validation was conducted 20 times with different random seeds using WEKA (ver. 3.6.1, The University of Waikato, Hamilton, New Zealand) to split the datasets into training and validation data [18]. Bootstrap analysis was also conducted to estimate the optimistic bias in the given datasets. Two hundred replicates, including the same number of patients, were computationally generated with a random selection of individuals, this permitted redundant selection, and MLR models were developed and cross-validation tests were conducted on each generated dataset.

Results

Metabolome analysis of serum samples obtained from non-DN and DN patients

Serum metabolome profiles of 78 patients in three successive stages of DN were collected using a single standard protocol [non-DN (n = 20), micro-DN (n = 32) and macro-DN (n = 26)] and analyzed. Age distribution, gender and other clinical characteristics are listed in Table 1. The ages in the micro-DN and macro-DN groups were slightly higher than in the non-DN group (p = 0.0226). The macro-DN group had significantly higher creatinine contents and lower estimated glomerular filtration rates (eGFR) than the other groups (p < 0.0001), while no significant difference was seen between the non-DN and micro-DN groups. The macro-DN group also showed significantly higher triglycerides and systolic blood pressure (SBP) compared with the non-DN group (p = 0.0172 and 0.0083, respectively). The other clinical parameters showed no significant difference among all groups (p > 0.05).

Table 1 Clinical characteristics of diabetic nephropathy patients
Full size table

On average, 4400 peaks were detected from each serum sample with CE-TOFMS. After eliminating redundant peaks, such as noise, fragments and adduct ions, 289 metabolites remained. Using this dataset, we firstly performed principal component analysis (PCA), but the resultant score plots of the PCA showed no unequivocal stage-specific clusters (data not shown). Next, OPLS-DA was performed to discriminate between DN patients (micro-DN and macro-DN) and non-DN patients based on the profiled metabolites. The OPLS-DA model demonstrated satisfactory separation between non-DN and micro-DN patients (Fig. 1a) using one predictive component and one orthogonal component (R 2 X cum = 0.21, R 2 Y cum = 0.676, Q 2cum  = 0.179), and clear separation between non-DN and macro-DN patients (Fig. 1b) using one predictive component and three orthogonal components (R 2 X cum = 0.353, R 2 Y cum = 0.946, Q 2cum  = 0.599). These results indicate that serum metabolome profile can be used to distinguish DN patients from non-DN patients.

Fig. 1
figure 1

OPLS-DA based on comprehensive metabolites data from (A) non-DN (blue dots) and micro-DN (pink triangles) samples and (B) non-DN (blue dots) and macro-DN (red triangles) samples. The ellipse in each figure indicates the Hotelling T2 (0.95) range for this model

Full size image

The resultant S plot of the developed OPLS-DA model between non-DN and macro-DN patients identified 19 metabolites (Table 2) that were highly correlated in the separation of the groups (|p(corr)| > 0.5). Of these, we were able to assign metabolite identities to eight metabolites by matching their m/z values and migration times with those of standard reagents. These metabolites were creatinine, aspartic acid, γ-butyrobetaine, citrulline, symmetric dimethylarginine (SDMA), kynurenine, azelaic acid, and galactaric acid. The composition formulae of the other metabolites were calculated based on their isotope distribution patterns as follows: C5H8N2O2 [metabolite ID (MID) 17], C9H17NO (MID 51), C9H19NO (MID 52), C2H4N2O3 (MID 158), and C6H6N4O (MID 202). Only the m/z values of the other metabolites are listed in Table 2 because of insufficient isotope peak size. The AUC values of MID 202 (0.765, 95 % CI); 0.649–0.880, p = 4.47 × 10−4) gave the best discriminating ability among these markers (Table 3).

Table 2 The 19 serum biomarker candidates that statistically differentiated the different DN stages
Full size table
Table 3 AUC values for individual markers
Full size table

Correlation between biomarker candidates and clinical parameters

Correlation analysis between these 19 serum biomarker candidates and currently available clinical parameters showed all candidate metabolites were significantly correlated with UACR (p < 0.009) (Table 4). The correlation coefficients for creatinine (r = 0.5701), aspartic acid (r = 0.4993), γ-butyrobetaine (r = 0.4942), citrulline (r = 0.4300), SDMA (r = 0.4820), kynurenine (r = 0.5351), MID 17 (r = 0.4968), MID 97 (r = 0.5223), MID 152 (r = 0.5336), MID 158 (r = 0.4980), and MID 202 (r = 0.6352) were positively correlated with UACR. Those of azelaic acid (r = –0.5210), galactaric acid (r = –0.4596), MID 51 (r = –0.4728), MID 52 (r = –0.4871), MID 96 (r = –0.3085), MID 114 (r = –0.3638), MID 127 (r = –0.2961), and MID 134 (r = –0.3669) were negatively correlated. Furthermore, 15 of 19 metabolites were significantly correlated with eGFR (p < 0.035). The correlation coefficients of creatinine (r = –0.8832), aspartic acid (r = –0.3912), γ-butyrobetaine (r = –0.6492), citrulline (r = –0.6531), SDMA (r = –0.7111), kynurenine (r = –0.5627), MID 17 (r = –0.5808), MID 97 (r = –0.7651), MID 152 (r = –0.7687), MID 158 (r = –0.6302), and MID 202 (r = –0.7455) showed negative correlation with eGFR. This indicates that they were positively associated with renal dysfunction. By contrast, those of azelaic acid (r = 0.3739), galactaric acid (r = 0.4152), MID 51 (r = 0.2204), and MID 134 (r = 0.2397) showed positive correlation with eGFR.

Table 4 Correlation analysis between the 19 biomarker candidates and clinical parameters (urinary albumin-to-creatinine ratio (UACR) or estimated glomerular filtration rate (eGFR))
Full size table

MLR model development

For the discrimination of DN (micro-DN and macro-DN) from non-DN patients, we developed a MLR model. Of the 19 biomarker candidates, γ-butyrobetaine, SDMA, azelaic acid, MID 114, and MID 127 were selected by stepwise feature selection as MLR variables. The developed model yielded high AUC values (0.927, 95 % CI, 0.870–0.983, p < 0.0001, Fig. 2a). The model also yielded high AUC values (±SD; 0.880 ± 8.62 × 10−3) in the cross-validation test. In a bootstrap test, the AUC values were 0.946 ± 0.0262 and 0.895 ± 0.0364 for training and cross-validation, respectively. To evaluate only the eight identified metabolites, we independently developed a MLR model. Stepwise feature selection selected aspartic acid, SDMA, azelaic acid, and galactaric acid as MLR variables. This MLR model also yielded high AUC values (0.844, 95 % CI, 0.754–0.934, p < 0.0001, Fig. 2b), although it performed slightly worse than the model with all metabolites, including unknown peaks. This model also yielded high AUC values (±SD; 0.792 ± 1.21 × 10−2) in the cross-validation test. In a bootstrap test, the AUCs were 0.875 ± 0.0419 and 0.820 ± 0.0543 for training and cross-validation, respectively. These results indicate that the developed model is sufficiently accurate, specific, and general.

Fig. 2
figure 2

ROC curve analyses in combination with (A) γ-butyrobetaine, SDMA, azelaic acid, MID 114, and MID 127, and (B) aspartic acid, SDMA, azelaic acid, and galactaric acid to discriminate non-DN and DN patients

Full size image

Discussion

The aim of this study was to obtain metabolic markers for early detection of DN from patient serum samples. We used CE-MS-based metabolome analysis to find differences in the serum metabolites from non-DN, micro-DN, and macro-DN samples. OPLS-DA with 289 metabolites clearly separated non-DN from macro-DN. Adequate separation of micro-DN from non-DN was also achieved. These results show that OPLS-DA is useful in this type of analysis. The resultant S-plot of the developed OPLS-DA model identified 19 metabolites that were major contributors to the separation of macro-DN from non-DN (|p(corr)| > 0.5). These metabolites showed a gradual increase or decrease with progressive development of nephropathy. Among them, eight metabolites were identified, and these markers are discussed in comparison with other published reports below.

The concentration of serum creatinine was significantly increased in the micro-DN and macro-DN groups compared with the non-DN group (p < 0.0001), and positively correlated with UACR (r = 0.5701, p < 0.0001) and negatively correlated with eGFR (r = –0.8832, p < 0.0001). Accumulation of serum creatinine was also observed in DN patients by metabolic analysis [19]. In clinical practice, creatinine is widely used as a marker of DN that reflects the renal function. Although serum creatinine had high specificity for detecting decreased GFR, the sensitivity is not sufficient because its levels do not significantly increase until the GFR is reduced to less than 50 % of normal levels [20].

The levels of amino acids, including aspartic acid (p < 0.0001), citrulline (p = 0.0005), SDMA (p = 0.0004), and kynurenine (p = 0.0002), were significantly elevated in the DN groups compared with the non-DN group. These metabolites showed high positive correlations with UACR (aspartic acid, r = 0.4993, p < 0.0001; citrulline, r = 0.4300, p < 0.0001; SDMA, r = 0.4820, p < 0.0001; kynurenine, r = 0.5351, p < 0.0001) and negative correlations with eGFR (aspartic acid, r = –0.3912, p = 0.0004; citrulline, r = –0.6531, p < 0.0001; SDMA, r = –0.7111, p < 0.0001; kynurenine, r = –0.5627, p < 0.0001). Aspartic acid and citrulline are involved in the urea cycle. Urea, is a major end product of nitrogen metabolism, and is produced by free ammonia and aspartic acid. Citrulline is normally taken up by the kidneys and converted to urea via arginine. Chuang et al. found significant accumulation of urea cycle intermediates in the patients with end-stage renal disease [21]. Because the kidneys are important in conversion of citrulline to arginine, the increase in the serum level of citrulline in DN patients could be attributed to degradation of this function.

SDMA and asymmetric dimethylarginine (ADMA), which is a structural isomer of SDMA, are formed by the enzymatic methylation of arginine residues within proteins. These metabolites have been identified as biomarkers for chronic kidney disease [22]. ADMA is metabolized by dimethylarginine dimethylaminohydrolase (EC 3.5.3.18) into citrulline and dimethylamine in the kidneys, whereas SDMA is excreted directly into the urine without further modification [23]. In this study, ADMA was under the detection limit, but SDMA was positively correlated with a decrease in function of kidney. Therefore, SDMA is a more sensitive marker than ADMA of various renal diseases, including DN.

Tryptophan is metabolized to kynurenine and further metabolized to acetyl-CoA and NAD in the tryptophan-kynurenine pathway. The rate limiting enzymes of this pathway are indoleamine 2,3-dioxygenase (EC 1.13.11.52) in the kidney and tryptophan 2,3-dioxygenase (EC 1.13.11.11) in the liver. Both these enzymes metabolize tryptophan to N-formylkynurenine, and N-formylkynurenine is subsequently catabolized to kynurenine. Saito et al. showed the peripheral kynurenine pathway accelerates in renal insufficient rats, and the reaction rate was positively correlated with the severity of the case [24]. They also found increased serum kynurenine concentrations reflected increased tryptophan 2,3-dioxygenase and decreased kynureninase (EC 3.7.1.3) activity in the liver [24]. Integration of profiling of these enzyme activities and metabolites will increase understanding of these mechanisms.

We detected a significant increase in γ-butyrobetaine in DN patients (p < 0.0001). Toyohara et al. showed a negative correlation between γ-butyrobetaine and eGFR in plasma from the patients with chronic kidney disease [25]. Because γ-butyrobetaine is converted to l-carnitine by γ-butyrobetaine dioxygenase (EC 1.14.11.1), it is assumed the increased γ-butyrobetaine arises from inhibition of this enzyme in the kidney.

The levels of azelaic acid (p < 0.0001) and galactaric acid (p < 0.0001) were significantly lower in the DN groups than the non-DN group. These metabolites also showed high negative correlations with UACR (azelaic acid, r = –0.5210, p < 0.0001; galactaric acid, r = –0.4596, p < 0.0001) and positive correlations with eGFR (azelaic acid, r = 0.3739, p = 0.0007; galactaric acid, r = 0.4152, p = 0.0002). Azelaic acid is a saturated C9 dicarboxylic acid derived from oxidation of fatty acids and inhibits the generation of reactive oxygen species on neutrophils [26]. Galactaric acid, is a natural product found in various fruits, and acts as a growth substrate for many organisms, including Escherichia coli [27]. However, biological mechanisms of decreased serum azelaic acid and galactaric acid after onset DN need to be clarified.

In this study, the obtained 19 metabolites showed relatively high separation abilities (AUC values of receiver operating characteristic curves 0.643–0.765, Table 3). To increase the separation ability, we then applied a MLR model to this dataset. The developed MLR model included five metabolites, γ-butyrobetaine, SDMA, azelaic acid, MID 114, and MID 127. This model had a higher AUC value for diagnosis of DN (0.927, p < 0.0001) than single markers, and shows the use of multiple markers is advantageous (Fig. 2a). However, this model contained two unidentified metabolites. The model using only identified metabolites was even simpler and more versatile for actual diagnosis because it could be used with quantification by another technique, such as LC, LC-MS, or an enzymatic method. Thus, we developed another MLR model using only the identified metabolites, aspartic acid, SDMA, azelaic acid and galactaric acid (Fig. 2b). This model showed high separation ability (AUC value 0.844, p < 0.0001), and could also be used to diagnose DN. However, there are several limitations to be acknowledged for this study. For example, the developed model should be further validated using larger and independent new datasets. In addition, although we evaluated the generalization ability of the developed model using cross-validation, the specificity of the model was not assessed. Especially, the specificity for DN using data obtained from study of other kidney diseases (e.g., kidney cancer) should be addressed in future study.

In conclusion, we applied CE-MS-based metabolome profiling to serum samples from diabetic patients with or without existing DN. Biomarker candidates for the early diagnosis of DN were obtained. Although a further validation study is needed, this technique has potential as a tool for biomarker discovery studies.