Metabolomic studies reveal and validate potential biomarkers of diabetic retinopathy in two Chinese datasets with type 2 diabetes: a cross-sectional study Authors: Xingchen Zhou, Guixue Hou, Xin Wang, Zhaoyi Peng, Xiaoming Yin, Jing Yang, Shan Wang, Yayi He, Yue Wang, Jing Sui, Wei Qiang, Hui Guo, Yanan Wang, Liang Lin, Bingyin Shi, Mingqian He Published: 2024 ABSTRACT BackgroundDiabetic retinopathy (DR) is a major microvascular complication of diabetes mellitus and causes vision impairment and blindness. The presence of major risk factors for DR, such as high levels of HbA1c, does not predict all DR pathogenesis in the clinic, which suggests that uncovering the underlying mechanisms and identifying novel markers are needed. Previous evidence has shown that the serum metabolic signature of DR is unique and detectable compared with that of diabetes mellitus (DM). Here, we aimed to identify serum metabolites as reliable biomarkers for the presence of DR in type 2 DM (T2DM) patients.MethodsWe performed untargeted and targeted metabolomic studies using liquid chromatography‒mass spectrometry (LC‒MS) and multiple reaction monitoring (MRM) methods on the serum samples of T2DM patients. For the discovery dataset, 39 DR patients and 39 non-DR (NDR) patients were included. For the validation dataset, 95 DR patients and 95 non-DR (NDR) patients were included. Receiver operating characteristic curve analysis was performed to evaluate the discriminating power of the metabolites. Binary logistic regression models were fit to evaluate the associations of metabolite peak areas or neurotransmitter concentrations with the presence of DR and adjusted for known risk factors.ResultsA total of 7123 metabolites were tested. The 39 DR patients had a mean age of 56 years with an average diabetes duration of 12 years, and the 39 NDR patients had a mean age of 57 years with an average diabetes duration of 11 years. Nine serum candidate markers were further identified. Six out of nine markers were associated with DR after we adjusted for covariates, including blood pressure, HbA1c, diabetes duration, fasting blood glucose, triglyceride, eGFR etc. Among them, eicosapentaenoic acid (EPA) and L-tyrosine were validated in an independent, risk factor-matched sample set. The serum L-tyrosine concentration was decreased in DR group by 47% (-0.22 ± 0.87 vs. 0.48 ± 1.05, P < 0.001), of which the cutoff value was 0.10 mg/ml, with 86% sensitivity and 40% specificity (AUC = 0.62, 95% CI = 0.54–0.70, P = 0.005).ConclusionsLow levels of circulating L-Tyrosine indicate retinopathy occurrence in T2DM population.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12933-024-02535-1. Background Diabetic retinopathy (DR) is a major microvascular complication of diabetes mellitus and is the leading cause of vision impairment and blindness in working-age adults [1, 2]. The global diabetes incidence in the population aged 20–79 years in 2021 was estimated to be 10.5% (536.6 million people) and is projected to reach 12.2% (783.2 million) by 2045 [3]. Global diabetes-related health expenditures, including DRs, would thus increase correspondingly. Extensive research has shown that high fasting blood glucose (FBG) and HbA1c levels, hypertension, and a long duration of diabetes are major risk factors for DR [4, 5], but disagreements among studies remain regarding their relative importance [2, 6, 7]. Moreover, despite strict control of systemic risk factors, many patients with diabetes continue to develop DR, suggesting that other potential pathogenic factors of DR are involved [8, 9]. Thus, efficient novel DR biomarkers are urgently needed for clinical use. One possible explanation for these variations is that early exposure to hyperglycemia predisposes individuals to the presence of diabetic complications, which are described as metabolic memory [10]. Even transient exposure to hyperglycemia results in long-lasting epigenetic modifications and persistent metabolic characteristics [11]. Although the underlying mechanisms responsible for these observations in DR occurrence are still unclear [12], it is possible that qualified novel biomarkers of DR can be discovered by detecting specific metabolites. Metabolomics has been widely used for biomarker discovery in many diseases and is able to detect changes in small molecules, such as nucleic acids, proteins and lipids, during pathophysiological processes [13]. There are two analytical approaches for metabolomics studies: untargeted and targeted. Untargeted metabolomics covers complete profiling of the metabolome, aiming to uncover novel biomarkers and pathophysiological pathways of disease. Here, we chose the liquid chromatography‒mass spectrometry (LC‒MS) platform, which is highly sensitive. In contrast, targeted metabolomics aims to measure a preset list of compounds. Although only capable of providing limited information, targeted metabolomics in general has higher sensitivity and selectivity [14]. Previous metabolomic studies have revealed new pathophysiologies and mechanisms of diabetes and associated diseases [10, 15, 16]. For example, a case‒control study used an untargeted method and discovered 6 urine metabolites as potential markers for predicting progressive renal function decline; these findings were further confirmed by targeted methods in an additional dataset [17]. Neurotransmitters, such as glutamate, glycine and γ-aminobutyric acid, are among the main kinds of metabolites that perform important biofunctions while processing visual information in the retina [18]. Several studies have shown that several of these metabolites play different roles in DR pathogenesis. For example, angiotensin II contributes to neuronal dysfunction by promoting synapse degradation [19]. Some studies have indicated that an accumulation of glutamic acid in the retina causes neurotoxicity and DR in patients with diabetes [20–22]. Somatostatin is a putative neuroprotective agent that is downregulated in the diabetic retina [23]. Therefore, we paid extra attention to these neurotransmitters. Here, we hypothesize that the serum metabolic signature of patients with DR is unique and resolvable from that of patients with diabetes alone. Thus, we aimed to identify serum metabolites as reliable biomarkers for the presence of DR in individuals with type 2 diabetes. Two datasets were used for discovery and validation. The participants were divided into two groups: type 2 diabetes patients with DR and type 2 diabetes patients without DR (NDR). The clinical characteristics of the patients were compared in this study. Serum samples were collected from the participants and subjected to LC‒MS analysis. After processing the data of the discovery set to identify potential markers, another set was used for validation. Methods Sample selection This cross-sectional study screened type 2 diabetes mellitus (T2DM) patients hospitalized in the Endocrinology Department of the First Affiliated Hospital of Xi’an Jiaotong University from 2018 to 2020. DR diagnosis and staging were based on the direct ophthalmoscopy examination [24]. The sample size was determined via metabolomics data from a previous study with a 95% CI, power > 80%, and a 1:1 ratio of the sample sizes of the two groups via the normal approximation method [25–27]. Thirty-nine T2DM patients with DR were included for discovery dataset. The control subjects were 39 T2DM patients who had no DR (NDR) and were matched to the patients by age (in 5-year bands), diabetes duration (in 5-year bands), HbA1c levels (in 0.5% bands), and hypertension history. Participants with an estimated glomerular filtration rate (eGFR) of less than 90 mL/min/1.73 m² (eGFR category: G2 to G5) were excluded from selection [28]. We first conducted a population-based case-control study on 39 patients in the DR group and 39 patients in the NDR group. For the validation dataset, the sample size was determined on the basis of the results from the discovery dataset to ensure a 95% CI, power > 90%, and a 1:1 ratio of the sample sizes of the two groups via the same method as previously described [25, 26]. Ninety-five type 2 diabetes patients with DR were included. The control subjects were 95 T2DM patients who had no DR. Same criteria were applied as stated previously for discovery dataset. Propensity score matching (PSM) was performed according to age, diabetes duration, HbA1c levels, hypertension history, sex, body mass index (BMI), diastolic blood pressure (diastolic BP), systolic blood pressure (systolic BP), and eGFR. Participants with an eGFR < 90 mL/min/1.73 m2 (eGFR category: G2 to G5) were excluded from the selection. Metabolite markers of DR identified from the discovery dataset were separately quantified in a separate sample set with 95 patients in the DR group and 95 patients in the NDR group. Serum collection protocol The tubes were kept on ice and transported to the laboratory within the same working day. The tubes were centrifuged at 1,000×g for 10 min (4 °C) to separate the serum from the whole blood. Serum aliquots were stored at -80 °C. Untargeted metabolomic quantification and data processing All serum samples were processed as described in a previous report [3]. In brief, 50 µl of serum was added to 150 µl of prechilled methanol, incubated at -20 °C overnight and centrifuged for 10 min. Then, the supernatant containing polar metabolites was transferred to a new Eppendorf tube, dried in a Savant vacuum concentrator and redissolved in acetonitrile: methanol: water (1:1:1, v/v/v) for LC‒MS/MS analysis. A UPLC-QE (Thermo Fisher Scientific, San Jose, CA, USA) instrument equipped with a BEH C18 column (100 × 2.1 mm, 1.7 μm; Waters, Milford, MA, USA) was used for untargeted metabolomics detection. The mobile phase was 0.1% formic acid in water (A) and 0.1% formic acid in acetonitrile (B) at a flow rate of 0.4 ml/min, and the gradient conditions were as follows: mobile phase B (2%) was maintained for 1 min initially, increased to 20% at 3 min and 90% at 8 min, and finally maintained for 3 min. The temperatures of the column and autosampler were 45 °C and 8 °C, respectively, and the injection volume was 10 µL. Electrospray ionization was performed in both positive ion and negative ion modes for data-dependent acquisition. The resolution of the full MS scan was 30,000, and that of the MS/MS scan was 17,500. For metabolite identification and quantification, Compound Discovery 2.1 (Thermo Fisher Scientific, San Jose, CA, USA) was used for data processing and searching against an online database (mzCloud) and an in-house metabolite database annotated with MS/MS spectra and retention time information. We imputed missing values for MS features with fewer than 50% missing observations in the QC samples. Missing observations were imputed via multivariate imputation via a k-nearest neighbors (kNN) algorithm via metaX [29, 30]. In total, 7123 features were generated via LC‒MS, defined by m/z, retention time, and ion intensity. Targeted metabolite quantification by multiple reaction monitoring (MRM) For neurotransmitters, a targeted MRM method was developed on a UPLC-QTRAP6500 Plus system (SCIEX, USA). Briefly, PITC derivatization buffer was prepared immediately before serum extraction and incubated at room temperature for 10 min. The PITC derivatization solution was then removed by vacuum evaporation, and the derivatized samples were extracted with 500 µL of 1:1 methanol: acetonitrile containing 5 mM ammonium acetate using an Eppendorf ThermoMixer C at room temperature with shaking at 2000 rpm for 20 min, followed by sonication for 10 min in an ultrasonic bath. The samples were then spun at 20,000×g for 5 min and injected for LC-MRM analysis. After quantification, all the raw data were analyzed using a skyline for MRM data processing. Statistics For LC‒MS data, features were filtered if their relative standard deviations (RSDs) were > 20% in quality control (QC) samples. Samples were normalized by the median. The peak data were log-transformed (base 10) and autoscaled to reach a normal distribution. For metabolites that formed multiple derivatization peaks, the peaks with a lower coefficient of variation (CV) in the QC samples were reported. Peaks with unresolvable identifications were reported as mixtures of compounds. QC samples were not considered for further analysis. Multivariate analysis by SIMCA 14.1, including principal component analysis (PCA) and partial least-squares discriminant analysis (PLS-DA), was subsequently used to identify differentially abundant metabolites between the 2 groups. For the PCA model, the R2 cum and Q2 cum values are reported. Permutation tests were used in the PLS-DA model. The parameter R2 was used to evaluate the fitting condition for the PLS-DA models, and Q2 was used to assess the predictive ability. The metabolites in the PLS-DA model with a variable importance in projection (VIP) score > 1.2 were considered important and included in further analysis. For neurotransmitters quantified in the discovery set, features with missing values in more than 50% of the samples were discarded. For features with missing values in less than 50% of the samples, the missing values were replaced by 1/5 of the minimum positive value of each variable. Samples were normalized by the median. The peak data were log-transformed (base 10) and autoscaled to reach a normal distribution. Wilcoxon rank-sum tests were performed to compare the DR group with the NDR group. Metabolites or neurotransmitters were considered significant if their p values were < 0.05 and if their fold changes were > 1.2 or < 0.8. Multiple testing for presented p value corrections was performed via the Benjamini-Hochberg method throughout the analysis (Q ≤ 0.2) [31]. The area under the curve (AUC) for receiver operating characteristic (ROC) curve analysis was calculated to evaluate the discriminating power of the metabolites or neurotransmitter markers. The baseline clinical characteristics of the two datasets were compared between the DR and NDR groups via the t test or ANOVA for continuous variables and the χ2 test for categorical variables. Univariate analysis was performed via binary logistic regression models to evaluate the associations of metabolite peak areas or neurotransmitter concentrations with the presence of DR. Odds ratios (ORs) with 95% confidence intervals (CIs) were calculated for selected markers. Clinical variables including known risk factors for DR (Systolic BP, HbA1c, diabetes duration, age, sex, diastolic BP, antihypertensive drugs, BMI, serum creatinine, fasting blood glucose, triglyceride (TG), total cholesterol (TC), low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol, the urine albumin–creatinine ratio (UACR), and eGFR), were added as covariates to logistic regression models to calculate the adjusted odds ratios. An ordinal logistic regression model was used to evaluate the associations of markers with DR stage. These statistical tests were performed in SPSS 23.0 or MetaboAnalyst. Pathway analysis Pathway analysis was performed with MetaboAnalyst. The lists of identified metabolites (Variable importance in projection (VIP) > 1.2, raw P < 0.05 and fold change > 1.2 or < 0.8) and neurotransmitters (P < 0.05 and fold change > 1.2 or < 0.8) were used as inputs. Canonical pathways were considered significantly enriched in the dataset if their false discovery rates (FDRs) were < 0.2. Correlation analysis Correlation analysis was performed with MetaboAnalyst in the validation set. Distances between features were measured by the Pearson r test and Spearman rank correlation test. The results with feature clustering are displayed as heatmaps. Results Sample characteristics The discovery metabonomic profiling dataset consisted of 78 patients with type 2 diabetes, including 39 DR patients and 39 diabetic controls. As shown in Table 1, the patient groups did not significantly differ in most clinical characteristics. There were 23 (59%) males in the NDR group and 31 males (79.5%) in the DR group. Twenty-one (53.8%) NDR participants and 26 (66.7%) DR participants took antihypertensive drugs. In both groups, 36 (92.3%) patients took lipid-lowering drugs, and 39 (100%) patients underwent hypoglycemic therapy. The two sample groups were comparable for most traditional risk factors for DR, such as the duration of diabetes, hypertension, HbA1c and BMI. DR patients had a higher urine albumin‒creatinine ratio (UACR), urine microalbumin, and urea level and a lower eGFR than diabetic controls. Although the prevalence of microalbuminuria did not differ between the two groups, compared with diabetic controls, DR patients had a greater incidence of macroalbuminuria. In the DR group, 5 samples were diagnosed as PDR (proliferative diabetic retinopathy), 34 samples were diagnosed as NPDR (nonproliferative diabetic retinopathy). Table 1Clinical characteristics of the samples selected for metabonomic profilingClinical characteristicsNDRDRP valueN3939Sex, n (%)Male23 (59.0)31 (79.5)0.051†Age (years)57 ± 1056 ± 110.633‡Hypertension (n)26260.999†Diastolic BP (mmHg)79 ± 1179 ± 100.991‡Systolic BP (mmHg)136 ± 15141 ± 240.293‡BMI (kg/m2)25.4 ± 3.525.2 ± 3.20.804‡Serum creatinine (µmol/L)58 ± 1765 ± 190.081‡Fasting blood glucose (mmol/L)7.63 ± 2.767.49 ± 2.750.824‡Total cholesterol (mmol/L)3.90 ± 1.004.17 ± 1.220.295‡LDL cholesterol (mmol/L)2.22 ± 0.892.62 ± 1.100.079‡HDL cholesterol (mmol/L)0.95 ± 0.220.96 ± 0.230.749‡Triglycerides (mmol/L)1.50 (1.03–2.97)1.61 (1.03–2.37)0.719‡HbA1c (%)8.1 ± 1.68.0 ± 1.40.777‡Duration of diabetes (years)11 ± 612 ± 60.367‡PDR (n)5Macular edema (n)4History of using drugsAntihypertensive drugs, n (%)21 (53.8)26 (66.7)0.247†Lipid-lowering drugs, n (%)36 (92.3)36 (92.3)0.875†Statins (n)3335Fibrates (n)21Others (n)10Hypoglycemic therapy, n (%)39 (100)39 (100)0.077†Oral hypoglycemic drugs (n)1710Insulin (n)63Combined (n)1626NSAIDs, n (%)28 (71.8)29 (74.4)0.799†UACR (µg/mg creatinine)14 (7–47)36 (13–110) 0.009‡ eGFR(mL/min/1.73m2)130 ± 35120 ± 35 0.027‡ Microalbuminuria (30–300 mg/24 h)10120.615†NDR, without diabetic retinopathy; DR, diabetic retinopathy; N, number; diastolic BP, diastolic blood pressure; systolic BP, systolic blood pressure; BMI, body mass index; LDL, low-density lipoprotein; HDL, high-density lipoprotein; PDR, proliferative diabetic retinopathy; NSAIDs, nonsteroidal anti-inflammatory drugs; UACR, urine albumin–creatinine ratio; eGFR, estimated glomerular filtration rate. Continuous variables are presented as medians (interquartile ranges) for abnormally distributed data and as the means with standard deviations (SDs) for normally distributed data. Differences between two groups were tested by Pearson’s chi-squared test (†) or the Mann‒Whitney U test (‡). Table 1Clinical characteristics of the samples selected for metabonomic profilingClinical characteristicsNDRDRP valueN3939Sex, n (%)Male23 (59.0)31 (79.5)0.051†Age (years)57 ± 1056 ± 110.633‡Hypertension (n)26260.999†Diastolic BP (mmHg)79 ± 1179 ± 100.991‡Systolic BP (mmHg)136 ± 15141 ± 240.293‡BMI (kg/m2)25.4 ± 3.525.2 ± 3.20.804‡Serum creatinine (µmol/L)58 ± 1765 ± 190.081‡Fasting blood glucose (mmol/L)7.63 ± 2.767.49 ± 2.750.824‡Total cholesterol (mmol/L)3.90 ± 1.004.17 ± 1.220.295‡LDL cholesterol (mmol/L)2.22 ± 0.892.62 ± 1.100.079‡HDL cholesterol (mmol/L)0.95 ± 0.220.96 ± 0.230.749‡Triglycerides (mmol/L)1.50 (1.03–2.97)1.61 (1.03–2.37)0.719‡HbA1c (%)8.1 ± 1.68.0 ± 1.40.777‡Duration of diabetes (years)11 ± 612 ± 60.367‡PDR (n)5Macular edema (n)4History of using drugsAntihypertensive drugs, n (%)21 (53.8)26 (66.7)0.247†Lipid-lowering drugs, n (%)36 (92.3)36 (92.3)0.875†Statins (n)3335Fibrates (n)21Others (n)10Hypoglycemic therapy, n (%)39 (100)39 (100)0.077†Oral hypoglycemic drugs (n)1710Insulin (n)63Combined (n)1626NSAIDs, n (%)28 (71.8)29 (74.4)0.799†UACR (µg/mg creatinine)14 (7–47)36 (13–110) 0.009‡ eGFR(mL/min/1.73m2)130 ± 35120 ± 35 0.027‡ Microalbuminuria (30–300 mg/24 h)10120.615†NDR, without diabetic retinopathy; DR, diabetic retinopathy; N, number; diastolic BP, diastolic blood pressure; systolic BP, systolic blood pressure; BMI, body mass index; LDL, low-density lipoprotein; HDL, high-density lipoprotein; PDR, proliferative diabetic retinopathy; NSAIDs, nonsteroidal anti-inflammatory drugs; UACR, urine albumin–creatinine ratio; eGFR, estimated glomerular filtration rate. Continuous variables are presented as medians (interquartile ranges) for abnormally distributed data and as the means with standard deviations (SDs) for normally distributed data. Differences between two groups were tested by Pearson’s chi-squared test (†) or the Mann‒Whitney U test (‡). Clinical characteristics of the samples selected for metabonomic profiling Sex, n (%) Male NDR, without diabetic retinopathy; DR, diabetic retinopathy; N, number; diastolic BP, diastolic blood pressure; systolic BP, systolic blood pressure; BMI, body mass index; LDL, low-density lipoprotein; HDL, high-density lipoprotein; PDR, proliferative diabetic retinopathy; NSAIDs, nonsteroidal anti-inflammatory drugs; UACR, urine albumin–creatinine ratio; eGFR, estimated glomerular filtration rate. Continuous variables are presented as medians (interquartile ranges) for abnormally distributed data and as the means with standard deviations (SDs) for normally distributed data. Differences between two groups were tested by Pearson’s chi-squared test (†) or the Mann‒Whitney U test (‡). Chemometric analysis There were 7123 features generated via LC‒MS, defined by m/z, retention time, and ion intensity. After data processing and identification, 1026 features remained for further statistical analysis. The diabetes mellitus (DM) and DR patients were not clearly separated from each other by the PCA model, which was optimized at 5 PCs, with R2 cum and Q2 cum values of 0.23 and 0.04, respectively. The first and second PCs explained 7% and 12%, respectively, of the model variation. However, when the PLS–DA model was applied, these two groups were clearly separated from each other, with model values of R2X cum = 0.17, R2Y cum = 0.99, and Q2 cum = 0.42, which indicated the fitness and prediction accuracy of the model (Supplementary Fig. 1A). The quality of the model was evaluated by a permutation test for 200 iterations (Supplementary Fig. 1A and 1B). Metabolite markers and pathways associated with DR Sixty candidate metabolites were identified via discovery metabonomic profiling (P < 0.05, Wilcoxon rank-sum test, Q ≤ 0.2, multiple testing correction, fold change > 1.2 or < 0.8, and VIP > 1.2 according to the PLS-DA model) (Supplementary Table S1). Considering the biological function of these metabolites, 5 of them were further selected as potential markers for analysis. As shown in Table 2, the DR samples exhibited lower levels of all-trans retinoic acid (ATRA), lauroylcarnitine, and decanoylcarnitine; phenethylamine; and higher levels of eicosapentaenoic acid (EPA). For the quantification of the neurotransmitter dataset, 8 neurotransmitters were selected as potential markers (P < 0.05, Wilcoxon rank-sum test, fold change > 1.2 or < 0.8) (Table 2 and Supplementary Table S2). Four of these neurotransmitters were selected for further analysis, as shown in Table 2. In total, DR patients had lower levels of L-tyrosine, L-tryptophan, L-phenylalanine and succinic acid. Figure 1 shows the normalized peak intensities and concentrations of these 9 markers. Fig. 1Peak intensities of metabolite markers identified via discovery metabonomic profiling. A All-trans retinoic acid (ATRA), B lauroylcarnitine, C decanoylcarnitine, D eicosapentaenoic acid (EPA), E phenethylamine, F L-phenylalanine, G L-tyrosine, H succinic acid and I L-tryptophan. The data are presented as the mean ± standard deviation (SD). The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. *P < 0.05; **P < 0.01; ***P < 0.001 Peak intensities of metabolite markers identified via discovery metabonomic profiling. A All-trans retinoic acid (ATRA), B lauroylcarnitine, C decanoylcarnitine, D eicosapentaenoic acid (EPA), E phenethylamine, F L-phenylalanine, G L-tyrosine, H succinic acid and I L-tryptophan. The data are presented as the mean ± standard deviation (SD). The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. *P < 0.05; **P < 0.01; ***P < 0.001 Table 2Metabolite and neuron transmitter markers identified via discovery metabonomic profilingCompound nameFold changeP valueQ valueAUC (95% CI)ATRA0.410.0080.0330.67 (0.55–0.78)Lauroylcarnitine0.660.0180.0360.65 (0.54–0.75)Decanoylcarnitine0.670.0200.0360.65 (0.53–0.77)EPA1.380.0290.0380.64 (0.52–0.76)Phenethylamine0.440.0300.0360.65 (0.52–0.77)L-phenylalanine0.750.0070.0880.68 (0.55–0.80)L-tyrosine0.760.0090.0880.67 (0.55–0.79)Succinic acid0.710.0140.1030.67 (0.55–0.78)L-tryptophan0.80.0380.1790.64 (0.51–0.76)AUC, area under the curve; CI, confidence interval; ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid. A fold change above 1 indicated higher expression in the diabetic retinopathy (DR) group than in the without diabetic retinopathy (NDR) group. The differences between the DR and NDR groups were calculated by the Wilcoxon rank-sum test. Multiple testing for presented p value corrections was performed via the Benjamini-Hochberg method (Q ≤ 0.2). The area under the curve (AUC) (95% CI) was obtained by univariate ROC analysis. Table 2Metabolite and neuron transmitter markers identified via discovery metabonomic profilingCompound nameFold changeP valueQ valueAUC (95% CI)ATRA0.410.0080.0330.67 (0.55–0.78)Lauroylcarnitine0.660.0180.0360.65 (0.54–0.75)Decanoylcarnitine0.670.0200.0360.65 (0.53–0.77)EPA1.380.0290.0380.64 (0.52–0.76)Phenethylamine0.440.0300.0360.65 (0.52–0.77)L-phenylalanine0.750.0070.0880.68 (0.55–0.80)L-tyrosine0.760.0090.0880.67 (0.55–0.79)Succinic acid0.710.0140.1030.67 (0.55–0.78)L-tryptophan0.80.0380.1790.64 (0.51–0.76)AUC, area under the curve; CI, confidence interval; ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid. A fold change above 1 indicated higher expression in the diabetic retinopathy (DR) group than in the without diabetic retinopathy (NDR) group. The differences between the DR and NDR groups were calculated by the Wilcoxon rank-sum test. Multiple testing for presented p value corrections was performed via the Benjamini-Hochberg method (Q ≤ 0.2). The area under the curve (AUC) (95% CI) was obtained by univariate ROC analysis. Metabolite and neuron transmitter markers identified via discovery metabonomic profiling AUC, area under the curve; CI, confidence interval; ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid. A fold change above 1 indicated higher expression in the diabetic retinopathy (DR) group than in the without diabetic retinopathy (NDR) group. The differences between the DR and NDR groups were calculated by the Wilcoxon rank-sum test. Multiple testing for presented p value corrections was performed via the Benjamini-Hochberg method (Q ≤ 0.2). The area under the curve (AUC) (95% CI) was obtained by univariate ROC analysis. The odds ratios of the 9 metabolite biomarkers in the basic logistic regression models and those adjusted for systolic BP, HbA1c levels, and diabetes duration are shown in Table 3. Except for lauroylcarnitine, decanoylcarnitine and L-tryptophan, the other 6 markers were significant according to the basic logistic models. Similar results were shown in ordinal logistic regression after we further divided the DR group into a NPDR group and a PDR group (Supplementary Table S3). Although HbA1c is a known risk factor for DR, it was not a significant covariate for any candidate marker, while the adjusted odds ratios for ATRA, EPA, phenethylamine, L-phenylalanine, L-tyrosine and succinic acid remained significant. Other known risk factors, such as increased systolic BP and duration of diabetes, were not significant covariates for any of the markers in our study. ATRA, EPA, phenethylamine, L-phenylalanine, L-tyrosine and succinic acid retained significant odds ratios adjusted for systolic BP. Except for lauroylcarnitine, decanoylcarnitine and L-tryptophan, the other 7 markers retained significant odds ratios adjusted for diabetes duration. These 7 markers, namely, ATRA, lauroylcarnitine, EPA, phenethylamine, L-tyrosine, L-phenylalanine and succinic acid, also retained significant odds ratios when adjusted for systolic BP, HbA1c levels, and diabetes duration (Table 3). Adjusted odds ratios for age, sex, BMI, diastolic BP, serum creatinine, glucose, triglyceride (TG), total cholesterol (TC), low-density lipoprotein (LDL), high-density lipoprotein (HDL), the UACR, and the eGFR are shown in Supplementary Tables S4–S6. None of them were significant covariates for any marker (Supplementary Tables S4–S6). Table 3Odds ratios (95% CI) of metabolite markers adjusted for systolic BP, HbA1c, and diabetes durationOdds ratios (95% CI)Odds ratios (95% CI) in adjusted models with covariates+ Systolic BP+ HbA1c+ Diabetes duration+ Systolic BP, HbA1c, and diabetes durationATRA0.39 (0.20–0.74)*0.38 (0.19–0.73)*0.38 (0.19–0.72)*0.37 (0.19–0.72)*0.35 (0.18–0.70)*Lauroylcarnitine0.57 (0.31–1.04)†0.60 (0.33–1.09)†0.58 (0.32–1.06)†0.58 (0.32–1.06)†0.62 (0.34–1.13)Decanoylcarnitine0.67 (0.39–1.15)0.70 (0.40–1.20)0.68 (0.40–1.16)0.68 (0.40–1.17)0.71 (0.41–1.23)EPA1.73 (1.01–2.98)*1.81 (1.04–3.16)*1.95 (1.09–3.48)*1.81 (1.04–3.13)*2.18 (1.18–4.02)*Phenethylamine0.62 (0.39–0.97)*0.63 (0.40–0.98)*0.62 (0.39–0.97)*0.58 (0.36–0.93)*0.59 (0.37–0.96)*L-Phenylalanine0.53 (0.31–0.92)*0.53 (0.31–0.91)*0.54 (0.31–0.93)*0.54 (0.31–0.94)*0.54 (0.31–0.93)*L-Tyrosine0.59 (0.35–0.99)*0.57 (0.33–0.96)*0.59 (0.35–0.99)*0.60 (0.35–1.01)*0.57 (0.34–0.98)*Succinic acid0.53 (0.30–0.92)*0.53 (0.31–0.92)*0.53 (0.30–0.93)*0.53 (0.30–0.94)*0.53 (0.30–0.94)*L-Tryptophan0.59 (0.35-1.00)†0.58 (0.34–0.99)*0.59 (0.35-1.00)†0.60 (0.35–1.03)†0.59 (0.34–1.01)†CI, confidence interval; systolic BP, systolic blood pressure; ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid. Odds ratios (95% CI) were obtained by performing binary logistic regression models with or without adding covariates to logistic regression models. Odds ratios above 1 of markers indicate that the marker is a risk factor for diabetic retinopathy (DR) compared with the nondiabetic retinopathy (NDR) group. *P (model) < 0.05; †0.05 ≤ P (model) < 0.1; ‡P (likelihood ratio) < 0.05; § 0.05 ≤ P (likelihood ratio) < 0.1. Table 3Odds ratios (95% CI) of metabolite markers adjusted for systolic BP, HbA1c, and diabetes durationOdds ratios (95% CI)Odds ratios (95% CI) in adjusted models with covariates+ Systolic BP+ HbA1c+ Diabetes duration+ Systolic BP, HbA1c, and diabetes durationATRA0.39 (0.20–0.74)*0.38 (0.19–0.73)*0.38 (0.19–0.72)*0.37 (0.19–0.72)*0.35 (0.18–0.70)*Lauroylcarnitine0.57 (0.31–1.04)†0.60 (0.33–1.09)†0.58 (0.32–1.06)†0.58 (0.32–1.06)†0.62 (0.34–1.13)Decanoylcarnitine0.67 (0.39–1.15)0.70 (0.40–1.20)0.68 (0.40–1.16)0.68 (0.40–1.17)0.71 (0.41–1.23)EPA1.73 (1.01–2.98)*1.81 (1.04–3.16)*1.95 (1.09–3.48)*1.81 (1.04–3.13)*2.18 (1.18–4.02)*Phenethylamine0.62 (0.39–0.97)*0.63 (0.40–0.98)*0.62 (0.39–0.97)*0.58 (0.36–0.93)*0.59 (0.37–0.96)*L-Phenylalanine0.53 (0.31–0.92)*0.53 (0.31–0.91)*0.54 (0.31–0.93)*0.54 (0.31–0.94)*0.54 (0.31–0.93)*L-Tyrosine0.59 (0.35–0.99)*0.57 (0.33–0.96)*0.59 (0.35–0.99)*0.60 (0.35–1.01)*0.57 (0.34–0.98)*Succinic acid0.53 (0.30–0.92)*0.53 (0.31–0.92)*0.53 (0.30–0.93)*0.53 (0.30–0.94)*0.53 (0.30–0.94)*L-Tryptophan0.59 (0.35-1.00)†0.58 (0.34–0.99)*0.59 (0.35-1.00)†0.60 (0.35–1.03)†0.59 (0.34–1.01)†CI, confidence interval; systolic BP, systolic blood pressure; ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid. Odds ratios (95% CI) were obtained by performing binary logistic regression models with or without adding covariates to logistic regression models. Odds ratios above 1 of markers indicate that the marker is a risk factor for diabetic retinopathy (DR) compared with the nondiabetic retinopathy (NDR) group. *P (model) < 0.05; †0.05 ≤ P (model) < 0.1; ‡P (likelihood ratio) < 0.05; § 0.05 ≤ P (likelihood ratio) < 0.1. Odds ratios (95% CI) of metabolite markers adjusted for systolic BP, HbA1c, and diabetes duration CI, confidence interval; systolic BP, systolic blood pressure; ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid. Odds ratios (95% CI) were obtained by performing binary logistic regression models with or without adding covariates to logistic regression models. Odds ratios above 1 of markers indicate that the marker is a risk factor for diabetic retinopathy (DR) compared with the nondiabetic retinopathy (NDR) group. *P (model) < 0.05; †0.05 ≤ P (model) < 0.1; ‡P (likelihood ratio) < 0.05; § 0.05 ≤ P (likelihood ratio) < 0.1. Moreover, as shown in Supplementary Table S7, the DR-associated pathways were analyzed by MetaboAnalyst (raw P < 0.05). The phenylalanine and tyrosine metabolism pathways were significantly enriched (FDR < 0.2) among the metabolite markers identified from the discovery metabonomic profiling. The phenylalanine and tyrosine metabolism pathway was identified based on the decreased levels of L-glutamic acid, L-tyrosine, L-phenylalanine, and oxoglutaric acid (Supplementary Table S7). Validation of metabolite markers The reproducibility of the 9 candidate markers identified from discovery metabonomic profiling was investigated in another set of samples. The clinical characteristics of the participants selected for the validation set are summarized in Supplementary Table S8. The validation set included 95 NDRs and 95 DRs. Eight of 95 DR patients were diagnosed as PDR. The two groups were comparable in terms of most clinical characteristics, including the history of using drugs and known risk factors for DR, such as hypertension, BMI and HbA1c. DR patients and NDR patients did not significantly differ in terms of the variables related to lipid metabolism, such as TG, LDL, HDL and TC. Consistent with the discovery dataset, DR patients had a significantly greater UACR and urine microalbumin level than did NDR patients. The urea levels did not differ between the DR patients and the NDR patients or between the proportion of patients with microalbuminuria or macroalbuminuria in the validation dataset. In the validation set, five of the nine candidate markers exhibited significant changes between the DR and NDR patients; these five markers included EPA, phenethylamine, L-tryptophan, succinic acid and L-tyrosine. As shown in Table 4, Figs. 2 and 3, the serum EPA concentration increased by 30% significantly in DR patients (0.21 ± 0.92 vs. -0.41 ± 1.24, P < 0.001), of which a cutoff value of 0.10 mg/ml, with 78% sensitivity and 60% specificity (AUC = 0.72, 95% CI = 0.64–0.79, P < 0.001). Furthermore, the serum L-tyrosine concentration decreased by 47% (-0.22 ± 0.87 vs. 0.48 ± 1.05, P < 0.001), of which a cutoff value of 0.10 mg/ml, with 86% sensitivity and 40% specificity (AUC = 0.62, 95% CI = 0.54–0.70, P = 0.005). in DR patients. As shown in Table 5, EPA and L-tyrosine were significant according to both the basic and adjusted logistic models, which was similar in ordinal logistic regression after we further divided the DR group into NPDR group and PDR group (EPA: OR (95% CI) = 91962.90 (864.91-9778077.04), P < 0.001; L-tyrosine: OR (95% CI) = 0.62 (0.43–0.91), P = 0.014) (Supplementary Table S9). These results were consistent with those of the discovery dataset. In particular, unlike those in the discovery set, the DR samples in the validation set had higher phenethylamine, L-tryptophan and succinic acid levels. The other four markers, ATRA, lauroylcarnitine, decanoylcarnitine and L-phenylalanine, did not significantly change in the validation set. Table 4Performance of 9 metabolite markers in the validation setBiomarkersP valueFold changeATRA0.3521.17Lauroylcarnitine0.9460.83Decanoylcarnitine0.4360.94EPA < 0.001 1.30Phenethylamine < 0.001 1.26L-phenylalanine0.4010.86L-tyrosine < 0.001 0.53Succinic acid 0.003 1.12L-tryptophan < 0.001 1.44ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid; The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. Fold change values above 1 indicate higher levels in the DR group. Table 4Performance of 9 metabolite markers in the validation setBiomarkersP valueFold changeATRA0.3521.17Lauroylcarnitine0.9460.83Decanoylcarnitine0.4360.94EPA < 0.001 1.30Phenethylamine < 0.001 1.26L-phenylalanine0.4010.86L-tyrosine < 0.001 0.53Succinic acid 0.003 1.12L-tryptophan < 0.001 1.44ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid; The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. Fold change values above 1 indicate higher levels in the DR group. Performance of 9 metabolite markers in the validation set ATRA, all-trans retinoic acid; EPA, eicosapentaenoic acid; The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. Fold change values above 1 indicate higher levels in the DR group. Fig. 2Original serum concentrations of the five markers in the validation dataset. A Original serum concentration of eicosapentaenoic acid (EPA). B Original serum concentration of phenethylamine. C Original serum concentration of L-tyrosine. D Original serum concentration of succinic acid. E Original serum concentration of L-tryptophan. The data are presented as the mean ± standard deviation (SD). The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. *P < 0.05; **P < 0.01; ***P < 0.001 Original serum concentrations of the five markers in the validation dataset. A Original serum concentration of eicosapentaenoic acid (EPA). B Original serum concentration of phenethylamine. C Original serum concentration of L-tyrosine. D Original serum concentration of succinic acid. E Original serum concentration of L-tryptophan. The data are presented as the mean ± standard deviation (SD). The differences between the diabetic retinopathy (DR) group and the nondiabetic retinopathy (NDR) group were calculated by the Wilcoxon rank-sum test. *P < 0.05; **P < 0.01; ***P < 0.001 Fig. 3ROC analysis of metabolites from validation metabonomic profiling. A ROC curve of EPA and B ROC curve of L-tyrosine. EPA, eicosapentaenoic acid; ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval. The area under the curve (AUC) (95% CI) was obtained by univariate ROC analysis. *cutoff values ROC analysis of metabolites from validation metabonomic profiling. A ROC curve of EPA and B ROC curve of L-tyrosine. EPA, eicosapentaenoic acid; ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval. The area under the curve (AUC) (95% CI) was obtained by univariate ROC analysis. *cutoff values Table 5Odds ratios (95% CI) of EPA and L-tyrosine determined via metabonomic profilingOdds ratios (95% CI)Odds ratios (95% CI) in adjusted models with covariates+ Systolic BP+ HbA1c+ Diabetes duration+ Systolic BP, HbA1c, and diabetes durationEPA1.87 (1.35–2.60)*1.86 (1.34–2.59)*§1.89 (1.35–2.63)*‡1.93 (1.38–2.71)*§1.94 (1.38–2.73)*‡L-tyrosine0.48 (0.35–0.66)*0.48 (0.35–0.66)*0.48 (0.34–0.66)*0.46 (0.33–0.64)*0.45 (0.32–0.64)*EPA, eicosapentaenoic acid; CI, confidence interval; systolic BP, systolic blood pressure. Odds ratios (95% CI) were obtained by performing binary logistic regression models with or without adding covariates to logistic regression models. An odds ratio above 1 indicates that a marker is a risk factor for diabetic retinopathy (DR) compared with nondiabetic retinopathy (NDR). *P (model) < 0.05; †0.05 ≤ P (model) < 0.1; ‡P (likelihood ratio) < 0.05; §0.05 ≤ P (likelihood ratio) < 0.1. Table 5Odds ratios (95% CI) of EPA and L-tyrosine determined via metabonomic profilingOdds ratios (95% CI)Odds ratios (95% CI) in adjusted models with covariates+ Systolic BP+ HbA1c+ Diabetes duration+ Systolic BP, HbA1c, and diabetes durationEPA1.87 (1.35–2.60)*1.86 (1.34–2.59)*§1.89 (1.35–2.63)*‡1.93 (1.38–2.71)*§1.94 (1.38–2.73)*‡L-tyrosine0.48 (0.35–0.66)*0.48 (0.35–0.66)*0.48 (0.34–0.66)*0.46 (0.33–0.64)*0.45 (0.32–0.64)*EPA, eicosapentaenoic acid; CI, confidence interval; systolic BP, systolic blood pressure. Odds ratios (95% CI) were obtained by performing binary logistic regression models with or without adding covariates to logistic regression models. An odds ratio above 1 indicates that a marker is a risk factor for diabetic retinopathy (DR) compared with nondiabetic retinopathy (NDR). *P (model) < 0.05; †0.05 ≤ P (model) < 0.1; ‡P (likelihood ratio) < 0.05; §0.05 ≤ P (likelihood ratio) < 0.1. Odds ratios (95% CI) of EPA and L-tyrosine determined via metabonomic profiling EPA, eicosapentaenoic acid; CI, confidence interval; systolic BP, systolic blood pressure. Odds ratios (95% CI) were obtained by performing binary logistic regression models with or without adding covariates to logistic regression models. An odds ratio above 1 indicates that a marker is a risk factor for diabetic retinopathy (DR) compared with nondiabetic retinopathy (NDR). *P (model) < 0.05; †0.05 ≤ P (model) < 0.1; ‡P (likelihood ratio) < 0.05; §0.05 ≤ P (likelihood ratio) < 0.1. Discussion This study aimed to identify serum metabolites that could serve as reliable biomarkers for the presence of DR in individuals with type 2 diabetes by metabolomics methods. A panel of nine metabolite biomarkers was identified via discovery metabonomic profiling (Fig. 1; Table 2). To identify novel markers that complement known risk factors and improve the existing clinical stratification of DR patients, adjusted odds ratios (ORs) were calculated [4, 5]. Metabolite markers which are independent of risk factors traditionally used for progression of DR were identified via omics-based methods [32]. Our results showed that many markers were independent of known risk factors for DR, which indicates their potential use for early DR diagnosis. We found that higher EPA levels or lower L-tyrosine levels were correlated with DR incidence after we adjusted for confirmed risk factors such as HbA1c. These two markers were validated in an independent dataset with a significant cutoff value. In our study, EPA and L-tyrosine were qualified and validated as potential DR biomarkers. As shown in Supplementary Fig. 2, EPA and L-tyrosine were clustered in the validation dataset, while the distance was measured by Pearson’s r test. EPA is an n-3 polyunsaturated fatty acid (PUFA), the intake of which is thought to aid in diabetes and DR prevention [33–35]. Some researches found that EPA is inversely associated with DR in Chinese population [36, 37]. Our results also confirmed EPA as a DR predictive biomarker. In addition to EPA, L-tyrosine was also a validated biomarker with a cutoff value of 0.10 mg/ml in this study. Previous studies reported similar results; a meta-analysis of three population-based cross-sectional studies in Asia revealed that lower levels of tyrosine were positively associated with DR [38]. Luo, H. H., et al. reported that low serum phenylalanine and tyrosine concentrations work independently and synergistically to increase the risk of DR in patients with T2DM [39]. Another case–dataset study involving more than 3000 people concluded that low levels of tyrosine appear to be a marker of microvascular risk in individuals with type 2 diabetes, independent of fundamental markers of kidney function [40]. Besides the predictive value, we also found that the phenylalanine and tyrosine metabolism pathway was identified as a key metabolic dysregulation pathway associated with DR. L-Tyrosine and L-phenylalanine are aromatic amino acids that have been proven to be positively associated with insulin resistance and the risk of type 2 diabetes [41]. As a precursor of dopamine, low levels of tyrosine or L-phenylalanine may lead to dopamine deficiency and thus promote neurodegeneration and early retinal dysfunction [42–44]. Both of these metabolites met the criteria for DR-related serum biomarkers in the discovery dataset, but only L-tyrosine was validated as a predictive biomarker of DR in this study. Moreover, several clinical studies confirmed the association between low tyrosine levels and DM or DR [40, 45]. Here, our pathway mapping results further support the hypothesis that a low tyrosine level contributes to the pathogenesis of DR by inducing dopamine deficiency, which can serve as a therapeutic target in the clinic for DR treatment and provide novel insights for drug discovery to combat this disease. Despite these promising results, limitations remain. As a cross-sectional study, the serum metabotypes of the samples may only represent short-term metabolic perturbations instead of chronic risk factors associated with the presence of DR. Furthermore, the medication status of the patients was not considered in this study, which may influence the results as a confounding factor during analysis. Only the few metabolites that met the criteria were further evaluated in the validation set, in which DR patients and control subjects with diabetes were matched. The generalizability of these findings to other populations needs to be validated in the future. Conclusions In conclusion, this study identified serum metabolite signatures that distinguish patients with diabetes and retinopathy from those without retinopathy by metabolomics methods. These metabolite markers remained significant and robust while controlling for established risk factors. The metabolite markers EPA and L-tyrosine were validated in an independent, risk factor-matched sample set with a significant cutoff value in ROC curve. The phenylalanine and tyrosine metabolism pathway was identified as a key metabolic dysregulation pathway associated with DR, demonstrating its involvement in disease pathogenesis. These findings provide L-tyrosine for correlation and predictive value of metabolite profiles with the rate of DR occurrence in patients with diabetes. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1