Multi-Omics Integration With Machine Learning Identified Early Diabetic Retinopathy, Diabetic Macula Edema and Anti-VEGF Treatment Response

Abstract

Purpose: Identify optimal metabolic features and pathways across diabetic retinopathy (DR) stages, develop risk models to differentiate diabetic macular edema (DME), and predict anti-vascular endothelial growth factor (anti-VEGF) therapy response.

Methods: We analyzed 108 aqueous humor samples from 78 type 2 diabetes mellitus patients and 30 healthy controls. Ultra-high-performance liquid chromatography–high-resolution-mass-spectrometry detected lipidomics and metabolomics profiles. DME patients received ≥3 anti-VEGF treatments, categorized into strong and weak response groups. Machine learning (ML) screened prospective metabolic features, developing prediction models.

Results: Key metabolic features identified in the metabolomics and lipidomics datasets included n-acetyl isoleucine (odds ratio [OR] = 1.635), cis-aconitic acid (OR = 3.296), and ophthalmic acid (OR = 0.836) for DR. For early-DR, n-acetyl isoleucine (OR = 1.791) and decaethylene glycol (PEG-10) (OR = 0.170) were identified as key markers. L-kynurenine (OR = 0.875), niacinamide (OR = 0.843), and linoleoyl ethanolamine (OR = 0.941) were identified as significant indicators for DME. Trigonelline (OR = 1.441) and 4-methylcatechol-2-sulfate (OR = 1.121) emerged as predictors for strong response to anti-VEGF. Predictive models achieved R² values of 99.9%, 97.7%, 93.9%, and 98.4% for DR, early-DR, DME, and strong response groups in the calibration set, respectively, and validated well with R² values of 96.3%, 96.8%, 79.9%, and 96.3%.

Conclusions: This research used ML to identify differential metabolic features from metabolomics and lipidomics datasets in DR patients. It implies that metabolic indicators can effectively predict early disease progression and potential weak responders to anti-VEGF therapy in DME eyes.

Translational Relevance: The identified metabolic indicators may aid in predicting the early progression of DR and optimizing therapeutic strategies for DME.

Introduction

Diabetic retinopathy (DR) is considered one of the main causes of blindness in the working-age population.¹ Diabetic macular edema (DME), affects about 20% of DR patients, which causes fluid accumulation in the macular region, is the primary cause of visual impairment in diabetics with a still poorly understood etiology.² Therefore quantitative mapping of the complex intraocular mechanisms behind DR onset and progression is crucial for advancing diagnosis and treatment. Recently, advanced imaging technologies like optical coherence tomography (OCT) enhance the precise localization of retinal and choroidal pathologies, aiding specialists in assessing layer-specific changes and informing treatment decisions for DR.^3-6 A key biochemical pathway involved in the onset and progression of DR is associated with vascular endothelial growth factor (VEGF).⁷ In clinical settings, anti-vascular endothelial growth factor (anti-VEGF) therapy is the primary treatment for DME; however, about 30% of patients exhibit suboptimal responses.^8-13 Thus, despite advanced imaging techniques, distinguishing DR phenotypes remains difficult, prompting researchers to investigate molecular changes in intraocular fluids. This approach could enhance retinal imaging by allowing analysis of aqueous humors (AH), potentially uncovering biomarkers linked to DR progression and providing valuable insights for targeted treatment strategies. 

Metabolomics, an emerging technology, investigates micromolecular metabolites, representing downstream products of the genome, transcriptome, and proteome, associated with disease phenotypes.¹⁴ This approach has the potential to reveal novel biomarkers and elucidate metabolic mechanisms, enhancing disease diagnosis and monitoring. Studies have revealed significant metabolic alterations associated with DR, highlighting their role in disease progression.^15,16 Nevertheless, comprehensive research into the different stages of DR, particularly DME, is still in its early stages, with limited studies on nonproliferative DR (NPDR) patients and intraocular fluid analysis. 

Machine learning (ML) stands out in omics research with its potent high-throughput data processing capabilities. ML extensively supports computer-aided diagnosis and disease target identification in the medical industry.¹⁷ Prediction models based on ML have been established to assist in the decision-making and identification of biomarkers for various eye diseases.¹⁸ By using ML to analyze high-throughput metabolomics data, we can identify the metabolites most strongly linked to diseases, thereby enhancing our understanding of the underlying mechanisms. 

This study aims to identify early DR (NPDR) and DME by establishing a risk prediction model and supplementing the metabolic spectrum of AH in DR patients using high-throughput nontargeted metabolomics. Additionally, the enrichment pathways for differential metabolites associated with DME and DR were examined. Furthermore, the treatment strategy for DME was optimized by analyzing multi-omics data that respond to anti-VEGF therapy. 

Material and Methods

Study Participants

In this prospective study, conducted between January 1, 2023, and March 31, 2024, treatment-naïve patients with DR were recruited from the Zhongshan Ophthalmic Center. The inclusion criteria of study groups were patients with a history of type 2 diabetes mellitus (T2DM) aged 50 years and above. Age- and sex-matched healthy individuals without ocular conditions other than cataract served as controls. The results from fundus color photography, spectral-domain OCT (SD-OCT) (Spectralis; Heidelberg Engineering, Heidelberg, Germany), or fluorescein fundus angiography (FFA450 fundus camera; Carl Zeiss, Jena, Germany) were used to diagnose DR. The American Diabetes Association's 2002 guidelines were followed for all diagnoses.¹⁹ T2DM patients were classified into three groups based on the Diabetic Retinopathy Disease Severity Scale: non-DR (NDR), NPDR, and proliferative DR (PDR). DME was defined as individuals with clinically diagnosed center-involving DME (CI-DME) and confirmed by central macular subfield thickness (CST) measured by OCT (305/320 µm for female/male).²⁰ The following were the exclusion criteria: (1) patients who had received anti-VEGF medication intravitreally or systemically or had undergone other interventional therapies (e.g., photodynamic therapy, retinal photocoagulation, pars plana vitrectomy, etc.) within the previous three months; (2) history of ocular trauma, inflammatory eye diseases, uveitis, retinal detachment, recent steroid treatment, or immunosuppressive drug use within the previous three months; (3) infectious diseases, diabetes-associated nephropathy (which includes patients with stage 3 chronic kidney disease, proteinuria, and macroalbuminuria, as well as hemodialysis patients), or inability to provide informed consent. Chronic kidney disorders were classified following the National Kidney Foundation Disease Outcomes Quality Initiative's clinical standards. The Zhongshan Ophthalmic Center's ethical committee Sun Yat-Sen University (IRB-ZOC-SYSU, 2023KYOJ292) authorized the research procedure, which was carried out in compliance with the Declaration of Helsinki. Written consent forms were provided by each participant. 

Each patient had a thorough ophthalmologic examination at the first visit, which included slit-lamp biomicroscopy, best-corrected visual acuity (BCVA) using the standard ETDRS letter, and CST evaluated with OCT. After retinal specialists confirmed the diagnosis of CI-DME through fundus and morphological examinations, eligible patients received intravitreal injections of conbercept (0.5 mg; Chengdu kanghong Biotech Co., Chengdu, China) following a 3+PRN regimen as per anti-VEGF treatment guidelines. Standard follow-up visits were conducted for 24 weeks after therapy, and CST and BCVA results were recorded. 

Grading of the Response to Anti-VEGF Therapy

The eyes under study were categorized into strong and weak response groups according to the BCVA and CST after anti-VEGF injections. A strong response was defined as achieving a BCVA gain of at least five, 10, or 15 letters when the baseline BCVA (Snellen equivalent) was 20/25 to 20/32, 20/40 to 20/63, or less than 20/80, respectively; and CST reduction of at least 50, 100, or 200 µm when the baseline CST was less than 75, ≥75 to <175, or ≥175 µm above standard thresholds. The weak response is defined as patients who, after 3+PRN anti-VEGF treatment, exhibit either no improvement or a decline in BCVA, or a reduction in CST that does not meet the criteria for the strong response group.²¹ 

The assessments were conducted with precision by two experienced independent readers (Y.P. and H.C.). In cases where their measurements differed by more than 10%, discrepancies were adjudicated by author Feng Wen. Notably, the interreader agreement was excellent (kappa > 0.80), which justified the use of their averaged measurements for further analysis. 

AH Sample Collection

The collection of AH samples from 48 DR patients (18 with NPDR, 30 with PDR) (25 with DME) was performed while undergoing anti-VEGF injection, as well as from 60 patients (30 NDR and 30 healthy control) while undergoing phacoemulsification surgery. AH samples were collected using a 25-gauge needle inserted through a 1 mm paracentesis. Between 50 and 100 µL of AH was collected, transferred to a 500 µL centrifugal vial, and promptly frozen in dry ice. Samples were then stored at −80 °C until analysis. 

AH Samples Preparation and Metabolomics Analysis

All experimental samples were prepared and analyzed using ultra-high-performance liquid chromatography–high resolution mass spectrometry (UHPLC/MS). Detailed descriptions of the metabolomics analysis can be found in the Supplementary Materials. 

Statistical Analysis and Pathway Analysis for Metabolomic Study

All values for the patient's clinical data were presented as means ± standard deviations (SD). Raw metabolomics data from AH were filtered (interquartile range), normalized (systematic error removal random forest), log transformed, and scaled across samples. Principal component analysis (PCA) and orthogonal partial least square discriminant analysis (OPLS-DA) were performed by SIMCA software (Sartorius AG Umetrics, Goettingen, Germany) for data analysis. A 200-iteration permutation test was run to check the model's overfitting risk. Univariate analysis, as well as independent sample t-test, and data visualization, were carried out using a proprietary cloud computing platform based on SPSS (version 26), R (version 4.1.3), and IPOS (http://82.157.20.231:3838/ipos/). Variable metabolites were uploaded into MetaboAnalyst 5.0 to conduct the analysis of metabolomic pathways. 

Model Development

Metabolomics data facilitated predictive modeling for DR, early DR, DME, and anti-VEGF responders. Least absolute shrinkage and selection operator (LASSO) regression addressed multicollinearity and variable selection. This approach generated a normalized model for application with one or ten markers. Variables identified in univariate analysis were included, and those with non-zero coefficients were retained. Logistic regression assessed selected variables. 

The prediction models were established with five classical ML algorithms (i.e., linear discriminant analysis, support vector machine, gradient boosting machine, random forest [RF], and neural networks). An evaluation of these algorithms with varying levels of complexity was conducted to gauge the equilibrium between performance and the potential for overfitting. A stratified sampling approach was used to reselect the entire group of 108 participants for calibration and validation, maintaining a 7:3 sampling ratio. Model prediction efficacy was assessed using metrics including accuracy, sensitivity, specificity, F1 score, receiver operating characteristic curve, and area under the curve. 

Results

Clinical Characteristics of the Study Population

A total of 78 patients with T2DM and 30 control subjects were included in the study, as indicated in Table 1. Of these, 30 had PDR (male: 50%), 18 had NPDR (male: 55.56%), and 30 had NDR (male: 36.67%). There were notable variations in the T2DM duration year (12.14 ± 3.18 year in PDR, 8.22 ± 3.2 year in NPDR and 7.5 ± 2.92 year in NDR, P < 0.001) between the T2DM groups. The T2DM group exhibited substantially higher levels of fasting blood glucose (FBG) and urea levels than the control group, with 7.9 ± 2.9 mmol/L versus 5.5 ± 0.71 mmol/L (P < 0.001) and 8.33 ± 4.56 mmol/L versus 4.72 ± 1.2 mmol/L (P < 0.001), respectively. Triglycerides (TG) and total cholesterol (TC) did not significantly differ across the groups. 

Table 1.

 

Baseline Characteristics of Study Participants

 

Supplementary Table S1 presents clinical and demographic data for the DME (25 patients) and non-DME (NDME, 22 patients). The baseline characteristics (gender, age, T2DM duration year, FBG, urea, TC, and TG) did not change significantly (all P > 0.05) between the DME and NDME groups. 

Notably, follow-up data for 23 of the 25 DME patients who received intraocular anti-VEGF treatment were available. Out of the 23 patients, 10 were deemed as strong response, whereas the remaining patients were classed weak response. There was no significant difference in baseline gender, age, T2DM duration year, FBG, BCVA or CST (all P > 0.05) (Supplementary Table S2). 

Metabolic and Lipid Subclasses Associated With DR Severity

To identify the metabolic characteristics correlated with the severity of DR, we compared NDR and DR (including NPDR and PDR). AH metabolomic profiling, assessed through PCA and OPLS-DA models (Supplementary Fig. S1), revealed significant separations among three groups. Univariate analysis was performed on the 314 metabolites, categorizing those with fold change > 1.2 or <0.8 and false discovery rate < 0.05 as differential metabolites. Ultimately, a comparison of the DR group with the NDR group identified 145 metabolites as differential metabolites (Supplementary Table S3). Differential metabolites were visualized via the volcano plot (Fig. 1A). A heatmap was used to display the top 30 differential metabolites between the NDR and DR groups (Fig. 1B). Additionally, pathways like the pentose phosphate pathway, arginine biosynthesis, and citrate cycle were discovered to be considerably enriched for differential metabolites in DR (Fig. 1C, Supplementary Table S4). 

Figure 1.

 

Differential metabolites and metabolic pathways between NDR versus DR (NPDR/PDR), and NDR versus NPDR groups from aqueous humor. (A) Differential metabolites between NDR and DR illustrated in volcano plot. Upregulated and downregulated metabolites are depicted in red and yellow, respectively. Nonsignificant metabolites are represented by blue dots. (B) Heatmap visualizing the group mean values of the top 30 differential metabolites between the NDR group and the DR group. Rows: differential metabolites; columns: samples. Color key indicates metabolite values (dark blue: lowest; dark red: highest). (C) Pathway enrichment analysis showed differential pathways differed between the NDR and DR groups. (D) Differential metabolites between NDR and NPDR illustrated in volcano plot. Upregulated metabolites were depicted in red and yellow, respectively. Non-significant metabolites were represented by blue dots. (E) Heatmap visualizing the group mean values of the top 30 differential metabolites between the NDR group and the NPDR group. Rows: differential metabolites; columns: samples. Color key indicates metabolite values (dark blue: lowest; dark red: highest). (F) Pathway enrichment analysis showed differential pathways differed between the NDR and NPDR groups.

Univariate analysis unveiled metabolic differences between NPDR and NDR. A total of 124 metabolites were significantly associated with NPDR (Supplementary Table S5). Differential metabolites were visualized in the volcano plot (Fig. 1D). A heatmap was utilized to display the top 30 differential metabolites between the NDR and NPDR groups (Fig. 1E). Pathway enrichment analysis highlighted significant enrichment in arginine biosynthesis, citrate cycle, and alanine, aspartate, and glutamate metabolism (Fig. 1F, Supplementary Table S6). 

Metabolic Profile and Lipid Subclasses Between DME and NDME

Differential metabolites in AH samples were explored for DME diagnosis in DR patients. Supplementary Figure S2A showed unclear group separation in the PCA model. OPLS-DA model revealed markedly different metabolic profiles between DME and NDME patients (Supplementary Fig. S2B). Differential metabolites were extracted via univariate analysis (criteria: fold change >1.2 or <0.8 and P < 0.1), visualized in the volcano plot and heatmap (Figs. 2A, 2B, Supplementary Table S7). Metabolite set enrichment analysis revealed enriched pathways in DME, including taurine and hypotaurine metabolism, glycine, serine, and threonine metabolism, and cysteine and methionine metabolism (Fig. 2C, Supplementary Table S8). 

Figure 2.

 

Differential metabolites and metabolic pathways between NDME versus DME, strong response versus weak response to anti-VEGF from aqueous humor. (A) Differential metabolites between NDME and DME illustrated in volcano plot. Upregulated metabolites are depicted in red and yellow, respectively. Nonsignificant metabolites are represented by blue dots. (B) Heatmap visualizing the group mean values of the differential metabolites between the NDME group and the DME group. Rows: differential metabolites; columns: samples. Color key indicates metabolite values (dark blue: lowest; dark red: highest). (C) Pathway enrichment analysis showed differential pathways differed between the NDME and DME groups. (D) Differential metabolites between strong response and weak response groups illustrated in volcano plot. Upregulated metabolites are depicted in red and yellow, respectively. Nonsignificant metabolites are represented by blue dots. (E) Heatmap visualizing the group mean values of the differential metabolites between the NDME group and the DME group. Rows: differential metabolites; columns: samples. Color key indicates metabolite values (dark blue: lowest; dark red: highest). (F) Pathway enrichment analysis showed differential pathways differed between strong response and weak response groups.

Metabolic Profile Between Strong and Weak Response to Anti-VEGF Treatment

The study aims to accurately distinguish between DME patients with strong response to anti-VEGF therapy and those with weak response. Metabolites that satisfied the significance threshold of P < 0.1 were chosen, plotted using volcano plots and heatmap, and examined for pathway enrichment (Figs. 2D–F, Supplementary Table S9). Differential metabolites are predominantly enriched in valine, leucine, and isoleucine biosynthesis, sulfur metabolism, and histidine metabolism (Supplementary Table S10). 

Features for Distinguishing DR, NPDR, DME, and Anti-VEGF Response

LASSO-penalized regression was used to screen high-throughput data via penalized maximum likelihood, defining the metabolite signature. A total of 10, 10, 14, 10 metabolites were screened in DR, early DR (NPDR), DME, and strong responders, respectively. Relative intensity between groups were shown in Figures 3A–D. After adjusting for age, sex, and T2DM duration year, multivariable logistic regression identified differential metabolic features for DR, including cis-aconitic acid (odds ratio [OR] = 3.296; 95% confidence interval [CI], 1.033-10.512), n-acetyl isoleucine (OR = 1.635; 95% CI, 1.084-2.467), and ophthalmic acid (OR = 0.836; 95% CI, 0.727-0.960). Moreover, differential metabolic features for NPDR were determined to be decaethylene glycol (PEG-10) (OR = 0.170; 95% CI, 0.130-0.612) and n-acetyl isoleucine (OR = 1.791; 95% CI, 1.014-3.182). Differential metabolic features for DME were found to be l-kynurenine (OR = 0.875; 95% CI, 0.814-0.970), niacinamide (OR = 0.843; 95% CI, 0.761-0.941), and linoleoyl ethanolamide (OR = 0.941, 95% CI, 0.910-0.983). Trigonelline (TRG) (OR = 1.441; 95% CI, 1.024-2.031) and 4-methylcatechol-2-sulfate (OR = 1.121, 95%CI: 1.046-1.120) were identified as differential metabolic features for strong response to anti-VEGF (Table 2). 

Figure 3.

 

Relative intensity of differential variables from AH and predictive model based on ML algorithm. (A) Relative intensity of potential biomarkers between NDR and DR groups. (B) Relative intensity of potential biomarkers between NDR and NPDR groups. (C) Relative intensity of potential biomarkers between DME and NDME groups. (D) Relative intensity of potential biomarkers between functional responders and nonresponders after anti-VEGF treatment; W-R, weak response; S-R, strong response. (E) Risk predictive model for DR was conducted using RF based on the calibration and validation set. (F) Risk predictive model for NPDR was conducted using RF based on the calibration and validation set. (G) Risk predictive model for DME was conducted using RF based on the calibration and validation set. (H) Risk predictive model for strong response to anti-VEGF in DME patients was conducted using RF based on the calibration and validation set. High Mean DecreaseGini represents metabolites that have a prominent contribution to the model; The predictive rate is represented by R². Gamma-Glu-Ile, gamma-glutamyl isoleucine; gamma-Glu-Leu, gamma-glutamyl leucine; LysoPC, lysophosphatidylcholine.

Table 2.

 

Identifying Classification Variables Through Logistic Regression

 

Prediction Models Based on ML Algorithm

Five specific ML algorithms were applied to construct prediction models based on the differential variables (Supplementary Table S11). RF exhibited significant performance on both the calibration and validation sets, demonstrating excellent accuracy, sensitivity, specificity, F1 score, and area under the curve. Therefore RF was used to establish the predictive model. Metabolites were ranked according to their impact on the model and visualized in bar plots. R² was used to measure the variation in accuracy explained by the variables in the model. The overall R² values for the calibration set in the DR, NPDR, DME, and strong response groups were 99.9%, 97.7%, 93.9%, and 98.4%, respectively. Similarly, the R² values for the validation set in the DR, NPDR, DME, and strong response groups were 96.3%, 96.8%, 79.9%, and 96.3%, respectively (Figs. 3E–H). These findings underscored accurate representation of disease status across DR, early DR, DME, strong response groups. 

Discussion

Metabolomics is effective in analyzing micromolecular substrates and diagnosing complex disorders like metabolic syndrome, T2DM, and obesity.²² In this study, we thoroughly clarified the metabolic and lipidomic profiles of DR and DME by applying the UHPLC/MS platform, and proposed a multi-omics data integration models based on ML algorithms for the identification of DR and its complications. This study deepens our understanding of the intraocular molecular changes associated with DME during anti-VEGF therapy. It provides new insights into the mechanisms of treatment action and resistance and aids in the identification of potential novel therapeutic targets. 

Through high-throughput multi-omics analysis, three metabolites were identified as the most closely associated with DR. Ophthalmic acid belongs to the class of oligopeptide organic compounds. It is a tripeptide, closely resembling the structure of glutathione (GSH). This structural similarity enables it to competitively bind to GSH as a substrate, inhibiting GSH's antioxidant function, thereby leading to the occurrence of DR.²³ Cis-aconitic acid, a tricarboxylic acid cycle (TCA) cycle intermediate, crucially regulates energy production in DR. Remarkably, renal biopsies from diabetic kidney disease patients exhibit reduced cis-aconitic acid gene expression.²⁴ Our findings suggest the TCA cycle as the primary origin of cis-aconitic acid in diabetic patients. Inhibiting TCA cycle enzymes effectively reversed renal injury in mice, offering a promising treatment strategy for diabetic microvascular complications.²⁵ 

Early detection and treatment of DR greatly improve visual prognosis. N-acetyl-isoleucine is composed of acylated α-amino acids linked to nitrogen atoms. Erroneous acetylation destabilizes protein structure, diminishing bioactivity and causing metabolic disorders. Accumulated n-acylated α-amino acids could potentially contribute to kidney damage, cardiac injury, and neurological impairment.²⁶ Notably, Park et al.²⁷ revealed that targeted therapy with n-acylated α-amino acids could prevent blood-retinal barrier breakdown and vascular leakage via VEGF receptor-2 antagonism, promising for diabetic retinopathy treatment. According to our findings, n-acetyl-isoleucine levels is elevated in the AH of NPDR patients. Disruption of protein n-terminal acetylation is proposed as a pivotal factor in early DR neurodegeneration. 

Amino acid metabolism, particularly pathways linked to arginine, is recognized as essential in DM and plays a crucial role in DR.²⁸ Our findings indicate significant enrichment of the arginine-related metabolic pathway in DR metabolism, suggesting metabolic disorders contribute to disease progression. Evidence suggests that ornithine, arginine, and citrulline were increased significantly in DR,²⁹ which is in line with our finding. Studies suggest arginine's involvement in inducible nitric oxide synthase (iNOS) uncoupling reduces reactive oxygen species, superoxide anions, and inflammatory cytokines in retinal macrophages and microglia, reversible by arginase 1.³⁰ Lee et al.³¹ suggests that arginine can inhibit iNOS mRNA translation, suppressing iNOS expression. More studies are needed to determine how arginine and its related enzymes regulate the development of DR. 

In our study, three metabolites were identified as indicators for DME. Niacinamide, or nicotinamide (NAM), a form of vitamin B3 found in food, belongs to the nicotinamide class of organic compounds.³² NAM serves as a precursor for nicotinamide adenine dinucleotide (NAD+) and NAD phosphate (NADP+) synthesis in various metabolic pathways. This redox pair significantly contributes to DR metabolism, with dysfunction being a major factor in DME.³³ In another study, the decline in NAD+ levels and NAD+/NADH ratio ultimately damaging the retinal vascular barrier.³⁴ Our results indicate reduced NAM expression in AH of DME patients, aligning with NAD+ and NADP+ depletion in DR. L-kynurenine and linoleoyl ethanolamide have been confirmed to be closely associated with reactive oxygen species generation and inflammation, which may be one of the important factors leading to fluid accumulation in DME.^35,36 

We analyzed the metabolic profiles of strong and weak response to anti-VEGF in DME patients. For the first time, our data revealed that strong response group had higher TRG levels than weak response group. TRG, a primary alkaloid from fenugreek seeds and found in sources like coffee, clover, and alfalfa, exhibits hypoglycemic effects by reducing oxidative stress and enhancing beta-cell activity.³⁷ Improving the intake of TRG is beneficial for the recovery of DME. As a component of phenyl sulfates, the role of 4-methylcatechol-2-sulfate in diabetes is still unknown. Kikuchi et al.³⁸ have described the role of phenyl sulfates as targets in diabetic nephropathy, with high levels of phenyl sulfates being associated with the progression of proteinuria. However, we found that high levels of phenyl sulfates are associated with a stronger anti-VEGF response in DME. Further investigation is warranted to understand its role in DME pathogenesis. 

Advancements in retinal imaging techniques, particularly OCT, have significantly enhanced our ability to identify structural biomarkers.^39–42 Imaging biomarkers such as sub-retinal fluid detected by SD-OCT, subtle retinal layer changes like hyperreflective intraretinal foci in swept-source OCT, and central choroidal thickness measured by enhanced depth imaging OCT have been closely associated with visual acuity, retinal sensitivity, and therapeutic responses in DME.^43–49 However, these biomarkers are limited by their indirect nature, relying on statistical interpretation rather than direct measurement.⁵⁰ Correlating these imaging findings with biochemical biomarkers could establish more direct links to the underlying mechanisms of these diseases. Future efforts to integrate liquid biopsy techniques with retinal analysis may enhance both diagnostic and prognostic capabilities, facilitating a more personalized approach to patient care. 

Despite our best efforts to demonstrate the metabolic spectrum of the DR population, our study has limitations: (1) the relatively small sample size may affect result reproducibility, necessitating larger samples in future research; (2) potential clinical biases, such as collection bias related to geographic location, ethnicity, age, and sex distribution, may arise from the single-center study design. Global multicenter research is necessary to validate and address these concerns. 

In conclusion, our study delves into the metabolic mechanisms of DR, establishing models for refined DME patient diagnosis and treatment stratification. The integration of these findings into clinical practice has the potential to advance personalized treatment strategies, ultimately benefiting patients and enhancing outcomes in the management of DR and its complications. Cis-aconitic acid and ophthalmic acid emerged as promising DR indicators, along with the enriched arginine biosynthesis pathway. Decaethylene glycol (PEG-10) and n-acetyl isoleucine showed specificity for early DR, whereas niacinamide distinguished between DME and NDME. Taurine and hypotaurine metabolism alterations were observed in DME. TRG and 4-methylcatechol-2-sulfate hold promise for predicting improvements in stronger anti-VEGF treatment response. Weak response to anti-VEGF existence underscores the need for combination or novel pharmacotherapies. This research provides a strong foundation for future investigations; however, further studies are needed to validate these findings and thoroughly assess their clinical implications. 

Acknowledgments

Supported by the Natural Science Foundation of Guangdong Province (No. 2023A1515011198), the Guangzhou Municipal Science and Technology Program (No. SL2022A03J00553), and the National Natural Science Foundation of China (No. 81970813) all provided funding for this work. 

Disclosure: Y. Pang, None; C. Luo, None; Q. Zhang, None; X. Zhang, None; N. Liao, None; Y. Ji, None; L. Mi, None; Y. Gan, None; Y. Su, None; F. Wen, None; H. Chen, None