The study adhered to the Declaration of Helsinki and was approved by the Ethics Committee of the First Affiliated Hospital of Fujian Medical University. Each patient has provided written informed consent prior to enrollment in the study. Thirty-eight diabetic patients (23 female) were enrolled.

Diabetic patients (with Type 2 Diabetes Mellitus, T2DM) who requested fundus examination at First Affiliated Hospital of Fujian Medical University Eye Center from November 2017 to September 2018 were enrolled. Diabetes was diagnosed according to the diagnostic criteria by the American Diabetes Association (ADA) (20). Patients were excluded from the study if they had other diseases that cause retinal hemorrhage and proliferation, such as retinal vein occlusion, retinal detachment, or macular degeneration. In addition, patients who suffered severe or complex systemic disease were also excluded. Patients who were incapable of carrying out study-related visits were also excluded. All patients were under standard glycemic control regimens, including oral hypoglycemic agents and/or insulin. None of the included patients had received prior anti-VEGF therapy, intravitreal steroids, or pan-retinal photocoagulation (PRP) for PDR.

DR grading was based on the fundus examination carried out by the same two senior ophthalmologists. Florid PDR was defined by the presence of active retinal neovascularization with or without vitreous hemorrhage, while Gliotic PDR was characterized by predominant fibrovascular membrane formation with signs of gliosis and traction (9).

Patients were stratified into three groups: Florid PDR (n=16): Neovascularization-dominant; Gliotic PDR (n=13): Fibrovascular membrane-dominant; Control group (n=9), termed “Non-Proliferative DR (NPDR/NDR) group” for clarity, consisted of patients with No DR (NDR, n=4) or non-proliferative DR (NPDR, n=5). NDR patients were enrolled during routine annual diabetic eye screening. Including NPDR allowed us to compare methylation profiles across the continuum of DR severity, from non-proliferative to proliferative stages.

All patients in this study were subjected to a detailed medical history taking, including age, gender, height, weight, Body Mass Index (BMI), history of diabetes and hypertension, use of diabetes medication and hypertension medication, history of smoking and drinking, daily exercise, family history, fasting blood glucose level, and glycated hemoglobin (HbA1c).

Peripheral venous blood was collected for serum preparation. The samples were incubated undisturbed for 30 minutes at room temperature and then centrifuged. The serum was transferred into fresh 2 mL tubes, frozen, and stored at −80°C until analysis. DNA isolation from frozen whole-blood samples stored at −80 °C using the phenol–chloroform extraction method was performed in the biobank setting.

The DNA methylation levels of specific CpG sites were determined by MethylTarget sequencing (Genesky Biotechnologies Inc., Shanghai, China), a method using next-generation sequencing-based multiple targeted CpG methylation analysis (21, 22). Primer design and validation were performed by Methylation Primer software on bisulfate-converted DNA. Primer sets were designed to flank each targeted CpG site in 100–300 nucleotide regions. Genomic DNA was extracted from frozen samples using Genomic Tip-500 columns (Qiagen, Valencia, CA, USA), and bisulfite-converted DNA was extracted using the EZ DNA Methylation™-GOLD Kit (Zymo Research, CA, USA) according to the manufacturer’s protocols. After PCR amplification (HotStarTaq polymerase kit, TAKARA, Tokyo, Japan) and library construction, samples were sequenced (Illumina HiSeq Benchtop Sequencer, CA, USA) using the paired-end sequencing protocol according to the manufacturer’s guidelines (23).

Gene segments and CpG sites were numbered sequentially according to their genomic positions relative to the transcription start site. The format “Gene_X_Y_Z” indicates: X = amplicon number, Y = segment identifier within the amplicon, and Z = nucleotide position relative to the amplicon start. For example, “EGLN1_1_2_73” refers to amplicon 1, segment 2, CpG site at position 73.

Sample size was determined based on previous epigenetic studies in DR, with an estimated effect size of 0.8, alpha of 0.05, and power of 80%.

SPSS software version 23 (SPSS Inc, Chicago, IL, USA) was used for statistical analysis. Data normality was assessed using the Kolmogorov-Smirnov test. The descriptive data were presented as means and standard deviations (SD). Methylation levels are expressed as median (%)and extreme values (%). Binary logistic regression was performed to assess associations between methylation levels and PDR subtypes, with results expressed as odds ratios (OR) and 95% confidence intervals (CI). Statistics were performed by t-test and ANOVA. Student’s t-test was used for comparisons between two groups. One-way ANOVA was used to compare three or more groups. P-value of less than 0.05 was considered statistically significant.

The study recruited 38 patients, including 23 women. The basic characteristics of patients in each group are shown in Table 1. Groups were not fully matched for age and gender; however, these variables were adjusted for in regression analyses.

As shown in Figure 1, AKR1B1 (P = 0.036), KDR (P = 0.004), and MMP9 (P = 0.018) showed low methylation in PDR patients compared to controls.

As shown in Figure 2; Table 2, AKR1B1 (P = 0.039) and MMP9 (P = 0.023) showed low methylation in the Florid PDR group compared to the control group.

In Figure 3; Table 3, the EPO (P = 0.025), KDR (P = 0.023), MMP9 (P = 0.014), and UCP1 (P = 0.048) genes in the Gliotic group are hypomethylated compared to the control group.

For all 16 target genes, no statistically significant differences in whole-promoter methylation levels were detected between the Florid and Gliotic PDR groups (P > 0.05 for all genes).

Although no statistically significant differences were detected at the whole-promoter level for the 16 target genes between the two PDR subtypes, exploratory analysis of finer-resolution regions revealed several nominally significant differences at the gene segment and individual CpG site levels.

Figures 4–6 show box plots with superimposed dots comparing methylation levels of all analyzed gene segments between the Gliotic and Florid PDR groups. As shown in these figures, most gene segments exhibited comparable methylation levels across the two subtypes, whereas some segments showed a trend toward differential methylation.

Table 4 lists the gene segments that reached nominal statistical significance (P < 0.05, uncorrected) in the comparison between Gliotic and Florid PDR groups. Notably, two segments of the EGLN1 gene (EGLN1_1_2_ and EGLN1_2_) showed significantly lower methylation in the Gliotic PDR group compared to the Florid PDR group.

Table 5 presents the individual CpG sites that exhibited nominal statistical significance (P < 0.05, uncorrected) between the two subtypes. A total of 20 CpG sites across multiple genes showed differential methylation, including sites within EGLN1, SOD2, AKR1B1, FLT1, HIF1A, PGF, SERPINF1, and TGFB1.

Notably, within the EGLN1 gene, locus-specific patterns were observed. In the EGLN1_1_2_ and EGLN1_2_ segments, which showed overall segment-level hypomethylation in Florid PDR (Table 4), the constituent CpG sites (sites 73 and 85, respectively) also demonstrated significantly lower methylation in the Florid group, consistent with the segment-level findings. In contrast, within the EGLN1_1_1_ segment—which did not reach statistical significance at the whole-segment level—two CpG sites (sites 78 and 150) showed significantly higher methylation in the Florid PDR group, suggesting that segment-level non-significance may reflect averaging across sites with opposing or heterogeneous methylation patterns. For other genes, multiple CpG sites showed differential methylation between the two subtypes, with varying directions of effect.

Given the exploratory nature of these analyses and the lack of multiple testing correction, all findings presented in Table 4 and Table 5 should be interpreted as hypothesis-generating signals requiring validation in independent cohorts.

Although no statistically significant differences were observed at the whole-promoter level across the 16 target genes between the two PDR subtypes, exploratory analysis of gene segments and individual CpG sites revealed several nominally significant differences within the EGLN1 gene, with notable locus-specific heterogeneity.

There were some consistent findings at the segment and CpG levels. The EGLN1_1_2_ and EGLN1_2_ segments showed significantly lower methylation in Florid PDR compared to Gliotic PDR at the whole-segment level (Table 4). This pattern was corroborated at the individual CpG level, with sites 73 (EGLN1_1_2_) and 85 (EGLN1_2_) also exhibiting hypomethylation in the Florid group (Table 5). This concordance suggests that the differential methylation in these regions is robust at multiple resolution levels and may have functional implications for EGLN1 expression. In contrast, within the EGLN1_1_1_ segment—which did not reach statistical significance at the whole-segment level—two CpG sites (sites 78 and 150) showed significantly higher methylation in the Florid PDR group. This apparent discrepancy highlights an important biological consideration: gene segments may contain CpG sites with opposing methylation patterns that cancel each other out when averaged across the entire segment. Such intra-segmental heterogeneity may reflect differential regulation of individual CpG sites by distinct transcription factors or chromatin modifiers, and has been increasingly recognized in epigenetic studies of complex diseases (11).

EGLN1 (also known as PHD2) encodes a prolyl hydroxylase that serves as a key oxygen sensor, targeting hypoxia-inducible factor-1α(HIF-1α) for proteasomal degradation under normoxic conditions (52). Previous studies have shown that hypomethylation of EGLN1 regulatory regions is associated with enhanced expression in response to hypoxic stress (53), representing a negative feedback mechanism that fine-tunes HIF-1α activity by promoting its degradation when oxygen levels normalize. In contrast, the predominantly higher methylation of EGLN1 CpG sites observed in Gliotic PDR in our study would be expected to have the opposite effect: reduced EGLN1 expression, leading to impaired HIF-1α degradation and consequent stabilization of HIF-1α even under conditions where it would normally be targeted for proteolysis. This stabilization may drive a transcriptional program that promotes extracellular matrix deposition and fibrosis over angiogenesis, aligning with the clinical phenotype of Gliotic PDR characterized by fibrovascular membrane formation rather than active neovascularization.

This interpretation is consistent with the established biology of the HIF pathway: EGLN1 acts as a rheostat that adjusts HIF-1α levels in response to oxygen availability (52). Epigenetic modulation of EGLN1 expression—whether through hypomethylation (as in adaptive responses to hypoxia) or hypermethylation (as potentially in Gliotic PDR)—can therefore shift this rheostat, tipping the balance between angiogenic and fibrotic outcomes.

However, given the exploratory nature of this analysis, the lack of multiple testing correction, and the complex, locus-specific patterns observed, these findings should be interpreted as hypothesis-generating signals requiring validation in larger cohorts and functional studies in ocular tissues. Future investigations should examine whether the differential methylation of specific EGLN1 CpG sites correlates with gene expression levels in retinal cells and contributes to the fibrotic phenotype in PDR.

Beyond EGLN1, exploratory analysis identified multiple CpG sites across several other genes that showed nominal methylation differences between Florid and Gliotic PDR (Table 5). These included multiple sites within SOD2 and AKR1B1, as well as single sites in FLT1, HIF1A, PGF, SERPINF1, and TGFB1.

Superoxide dismutase 2 (SOD2) encodes a mitochondrial antioxidant enzyme that protects against oxidative stress, a key pathogenic mechanism in DR (48). Differential methylation of multiple SOD2 CpG sites between subtypes may reflect altered oxidative stress responses, though the functional consequences of these specific methylation changes remain unknown. AKR1B1, discussed earlier in the context of PDR progression, also showed differential methylation at three CpG sites between subtypes, suggesting that this gene may be subject to complex epigenetic regulation at multiple levels.

The remaining genes—FLT1, HIF1A, PGF, SERPINF1, and TGFB1—are all implicated in angiogenesis, hypoxia response, or fibrosis. The observation of single CpG sites with nominal differences in each of these genes raises the possibility of broader epigenetic dysregulation, but the lack of consistent patterns across multiple sites within each gene precludes firm conclusions.

Importantly, for all of these genes, the differences were observed only at the individual CpG site level, with no corresponding differences at the whole-gene or gene-segment level. Given the exploratory nature of these findings and the lack of multiple-testing correction, these observations should be interpreted with caution and validated in larger cohorts.

There are some limitations in the present study. First, the sample size is relatively small, which limits statistical power and increases the risk of both false-positive and false-negative findings. Second, we acknowledge a methodological limitation in our statistical approach. The use of parametric tests (t-test, ANOVA) assumes normality of data distribution, which may not be fully met by percentage-based methylation data. Furthermore, we did not perform multiple testing correction (e.g., for the 16 target genes), meaning the reported P-values should be interpreted with caution as exploratory indicators rather than confirmatory evidence. Third, the lack of a non-diabetic healthy control group prevents distinguishing epigenetic changes specific to DR progression from those associated with diabetes or chronic hyperglycemia per se. In addition, the use of peripheral blood, while minimally invasive, may not fully capture the epigenetic landscape within the ocular microenvironment (e.g., retina, vitreous, fibrovascular membranes). This is a critical limitation when studying a localized disease like PDR. Finally, all participants had T2DM, limiting generalizability to Type 1 diabetic retinopathy. Other potential confounders, such as medications and comorbidities, were also not fully controlled for.

Future studies are essential to validate and extend these preliminary observations. Priorities should include: 1) Conducting well-powered studies with larger, matched cohorts that apply appropriate non-parametric statistics and multiple testing corrections; 2) Investigating methylation patterns in direct ocular samples (e.g., vitreous humor, epiretinal membranes) to confirm local relevance; and 3) Performing functional assays to determine the biological consequences of the observed methylation changes. Only through such rigorous follow-up can the potential of these epigenetic signatures be fully evaluated. Furthermore, to directly address the limitation of using peripheral blood, methylation analysis of matched vitreous humor and epiretinal membrane samples from the same cohort is currently in progress. The results will be critical for validating the local relevance of the blood-based signatures identified in this study.

In summary, this exploratory study provides the first evidence that Florid and Gliotic PDR may be associated with differential methylation patterns at specific CpG sites and gene segments. While no statistically significant differences were observed at the whole-promoter level for any of the 16 target genes between the two subtypes, segmental and CpG-level variations—particularly within EGLN1—suggest potential epigenetic heterogeneity that may contribute to the distinct clinical phenotypes of angiogenesis-dominant (Florid) versus fibrosis-dominant (Gliotic) PDR.

Several genes showed hypomethylation in PDR patients compared to non-proliferative controls. In Florid PDR, AKR1B1 hypomethylation may enhance polyol pathway activity and oxidative stress, while MMP9 hypomethylation aligns with its established role in extracellular matrix degradation and neovascularization (49).

In Gliotic PDR, hypomethylation of EPO and KDR may promote fibrovascular proliferation through angiogenic and inflammatory pathways (28, 34), and UCP1 hypomethylation—though modest in magnitude—could reflect compensatory anti-inflammatory or metabolic adaptations (41).

These preliminary findings offer a foundation for future research into subtype-specific epigenetic signatures in PDR. However, given the exploratory nature of this study, the small sample size, and the lack of multiple testing correction, all observations require rigorous validation in larger, independent cohorts and in ocular tissues. Future studies should also investigate the functional consequences of the identified methylation changes and explore their potential as minimally invasive biomarkers for risk stratification or as guides for subtype-specific therapeutic strategies.