Identification and analysis of oxidative stress-related genes associated with the occurrence and development of diabetic retinopathy Authors: Liangjie Niu, Yifan Zhang, Huika Xia, Jianmin Wang Published: 2026 ABSTRACT AbstractDiabetic retinopathy (DR) develops via complex interactions of multiple risk factors and molecular pathways. However, the association between oxidative stress-related genes (OSRGs) and DR progression remains poorly understood. We analyzed two DR datasets (GSE221521, GSE189005) and 986 OSRGs. Differentially expressed genes (DEGs) from GSE221521 were intersected with OSRGs to generate candidate genes; those with protein-protein interaction (PPI) network associations were defined as candidate hub genes. Machine learning further filtered final hub genes, which underwent cross-dataset expression validation. Biomarkers were identified by receiver operating characteristic (ROC) curve analysis (AUC > 0.7) and subjected to nomogram construction, functional annotation, and immune correlation analysis. We obtained 52 candidate genes, 39 of which had PPI interactions. Machine learning yielded 13 final hub genes; CD36, DDIT3, E2F2, NFATC4, and RPL11 showed consistent expression trends across datasets. CD36, DDIT3, E2F2, and RPL11 (AUC > 0.7) were validated as biomarkers, with robust DR predictive performance in nomograms. These genes were enriched in ribosome and spliceosome pathways. Immune infiltration analysis identified three differential cell types: M0 macrophages, monocytes, and activated memory CD4 + T cells. CD36, DDIT3, E2F2, and RPL11 are potential diagnostic biomarkers for DR. The present bioinformatic findings provide preliminary clues for subsequent exploration of DR pathogenesis.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-53316-7.Keywords: Diabetic retinopathy, Oxidative stress, Biomarkers, Immune infiltrationSubject terms: Biomarkers, Computational biology and bioinformatics, Diseases, Genetics, Immunology Introduction Diabetic retinopathy (DR) is a series of retinal lesions caused by long-term hyperglycemia, leading to microvascular leakage and obstruction in the retina, such as microangiomas, hard exudates, cotton wool spots, neovascularization, macular edema, proliferative membrane formation, and even tractional retinal detachment1,2. As one of the most common and serious microvascular complications of diabetes mellitus (DM), DR has emerged as the leading cause of visual impairment and blindness worldwide3,4. At present, the main treatment methods for DR include intravitreal injection of drugs, retinal laser photocoagulation, vitrectomy, and other interventions5. However, these treatment methods cannot completely cure DR, and can only delay the progression of DR as much as possible-eprimarily because the specific molecular details underlying DR, especially those related to oxidative stress, have not been fully clarified. Once poorly controlled, it will inevitably lead to very serious consequences. Therefore, exploring the signaling pathways and related molecular associations involved in the occurrence and development of DR at the molecular level is of great significance for further understanding the pathogenesis of DR and guiding clinical prevention and treatment6. Among the multiple pathogenic factors implicated in DR, oxidative stress has emerged as a central mechanism linking hyperglycemia to retinal damage. Redox reactions are the most fundamental chemical reactions in the human body, which are in a state of balance under physiological conditions. When this balanced state is disrupted, oxidative stress occurs—it refers to a state of imbalance between oxidative and antioxidant effects in the body, caused by excessive production of reactive oxygen species (ROS), which further leads to cell damage and subsequent loss of cellular function7. DR arises and evolves through a complex interplay of multiple risk factors and molecular pathways; its specific pathogenic associations related to oxidative stress are still not fully understood. Currently, it is widely believed that high blood glucose-induced retinal oxidative damage involves four classic metabolic abnormalities: protein kinase C (PKC) pathway, polyol pathway, hexosamine pathway, and advanced glycation end products (AGEs) formation8,9. DR patients are in a long-term state of hyperglycemia, and this long-term hyperglycemia damages the retinal capillary wall by activating various abnormal metabolic pathways of oxidative stress, leading to the occlusion of the retinal capillary wall9. In addition, chronic hyperglycemia can induce low-grade inflammation, increase ROS production, disrupt the balance of redox reactions in the body, and lead to early retinal nerve cell death. Subsequently, this retinal nerve cell death directly triggers the collapse of the blood-eye barrier, an increase in vascular permeability, and the acceleration of DR progression10,11. Relevant research shows that targeting oxidative stress can be used as a strategy to prevent and treat microvascular complications of DM, such as DR12. However, so far, the biological associations of oxidative stress-related genes (OSRGs)—as the core molecular mediators of oxidative stress—in the occurrence and development of DR have not been fully explored. Although a recent large-scale proteomic study13 has established a comprehensive plasma proteome atlas for DR, transcriptomic and oxidative stress-related gene signatures remain less explored. The present study focuses on oxidative stress-related genes and immune regulatory networks at the transcriptome level, providing biological complementarity to the existing proteomic findings. Therefore, this study utilized DR related public datasets and comprehensive bioinformatics methods to identify biomarkers associated with oxidative stress during the occurrence and development of DR. From this bioinformatics study, we analyzed the biological pathways they participate in, their association with the immune microenvironment, and explored their potential correlations as well as related diseases and drugs, providing new directions for further research and treatment of DR. Results Identified and functional enrichment of candidate genes After differential expression analysis, we identified 3,008 differentially expressed genes (DEGs) from GSE221521, including 2,420 up-expressed and 588 down-expressed genes (Fig. 1A and B), which were intersected with OSRGs taken to get 52 candidate genes (Fig. 1C). In GO enrichment analysis, candidate genes were enriched to a total of 945 functional entries, including response to oxidative stress, intrinsic apoptotic signaling pathway, etc. (Fig. 1D). In addition, a total of 29 signaling pathways were shown to be associated with candidate genes in KEGG, such as MAPK signaling pathway, neutrophil extracellular trap formation and so on (Fig. 1E). Moreover, MRA analysis showed that RPL10A, RPL35A, RPS8, etc. inhibited the activation of RPL11, and OSMR, PLEKHM3, KMT2C, etc. promoted the activation of EP300, and so on (Fig. 1F). Identified and functional enrichment of candidate genes. (A) Volcano plot of differential analysis. (B) Heatmap of differential analysis(C) Venn for DEGs and OSRGs. (D) GO enrichment analysis results for 52 candidate genes. (E) KEGG enrichment analysis chart results for 52 candidate genes. The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway map was obtained from KEGG(https://www.kegg.jp/). KEGG is a publicly available resource under the terms of the academic uselicense (Kanehisa et al., 2016: Kanehisa & Goto,2000)57–59. (F) MRA analysis regulatory network between 5 candidate genes with other genes. Identification of 13 final hub genes in machine learning In the PPI network, 39 candidate hub genes interacted with each other, such as H3C14-H4C4, H3C12-EP300, etc. (Fig. 2A). In addition, 19 feature genes 1 in Boruta (Fig. 2B), 34 feature genes 2 in SVM REF (Fig. 2C), and 22 feature genes 3 in LASSO (Fig. 2D) were identified, respectively. After taking the intersection of these genes, 13 final hub genes were identified for subsequent analyses (Fig. 2E). Identification of 13 final hub genes in machine learning. (A) PPI network of 39 candidate hub genes. (B) Boxplot showing the importance scores of 19 feature genes (feature genes 1) identified by Boruta analysis. The x-axis represents genes, and the y-axis denotes feature importance calculated based on the random forest algorithm. Boxplots illustrate the distribution of feature importance for each gene across multiple iterations. The color codes are defined as follows: green (Confirmed) indicates genes identified as stable important features by Boruta; red (Rejected) denotes genes regarded as unimportant features; yellow (Tentative) represents marginal features with ambiguous importance classification; blue (Shadow) refers to shadow features generated by random permutation of original features, which serve as the baseline control for feature importance evaluation. (C) The accuracy of the Support Vector Machine model identified 34 feature genes. (D) LASSO logistic coefficient penalty chart and cross-validation error curve identified 22 feature genes 3. (E) Venn diagram showing the intersection of feature genes screened by the three machine learning algorithms, yielding 13 final hub genes. Construction of nomogram for CD36, DDIT3, E2F2, and RPL11 The expression profiles of final hub genes were analyzed in the GSE221521 and GSE189005 datasets. Among them, CD36, DDIT3, E2F2, NFATC4, and RPL11 presented consistent expression patterns and statistically significant differences between DR and control groups across the two datasets (Fig. 3A). ROC analysis revealed that the four genes including CD36, DDIT3, E2F2 and RPL11 displayed favorable discriminative ability between DR and control samples, with AUC values greater than 0.7 in both datasets. Accordingly, these four genes were screened as core biomarkers (Fig. 3B).A nomogram model integrating the above biomarkers was subsequently constructed to estimate the occurrence probability of DR according to the total scores (Fig. 3C). The overall AUC of the nomogram model was above 0.7 according to ROC analysis (Fig. 3D). The calibration curve indicated that the predicted values were relatively consistent with the actual observations, with a slope close to 1 (Fig. 3E). Furthermore, decision curve analysis demonstrated that the model could provide certain reference value for clinical assessment (Fig. 3F). In summary, the established nomogram exhibited moderate predictive performance, and its external reproducibility and further application still require subsequent validation. Construction of a nomogram for CD36, DDIT3, E2F2, and RPL11. (A) Expression boxplots of final hub genes in the train set and validation set, where CD36, DDIT3, E2F2, NFATC4, and RPL11 showed consistent expression trends in the two datasets and were significantly different between the DR and control group. (B) AUC values of the ROC curves of 4 genes were greater than 0.7 in the train set and validation set. CD36, DDIT3, E2F2, and RPL11were considered as biomarkers. (C) Nomogram of the biomarkers. (D) ROC curve of the nomogram. (E) Calibration curve of the nomogram. (F) DCA curve of the biomarker column chart. In the GSE60436 dataset, CD36, E2F2 and RPL11 were significantly upregulated, while DDIT3 was remarkably downregulated in the DR group relative to the control group (Supplementary Figure S1A).Meanwhile, in the GSE160306 dataset, only DDIT3 and RPL11 were detected with available expression profiles, and both genes exhibited significantly lower expression in the DR group compared with controls (Supplementary Figure S1B). The expression levels of the biomarkers were further verified by quantitative real-time polymerase chain reaction (qRT-PCR) using a small clinical cohort (5 pairs). Among these, DDIT3 and RPL11 were significantly downregulated in DR samples (p < 0.05), whereas CD36 and E2F2 did not reach statistical significance, which may be limited by the small sample size (Table 1). Table 1Clinical sample test results.ControlDRp-valueDDIT31 ± 0.66090.1463 ± 0.07740.0233CD361 ± 0.59281.0606 ± 0.46320.8615E2F21 ± 1.16610.1859 ± 0.15930.1603RPL111 ± 0.43240.1686 ± 0.06130.0112Conduct statistical analysis on the differences in biomarker expression levels between DR group samples and control group samples to obtain corresponding p-values. A p-value less than 0.05 was considered statistically significant. Clinical sample test results. Conduct statistical analysis on the differences in biomarker expression levels between DR group samples and control group samples to obtain corresponding p-values. A p-value less than 0.05 was considered statistically significant. Biomarkers were all enriched to ribosome and spliceosome In this study, the biological functions of the biomarkers were further explored by GSEA. The results showed that the biomarkers were all significantly involved in the ribosome and spliceosome pathways. Meanwhile, NFATC4 and RPL11 were both significantly enriched in oxidative phosphorylation (Fig. 4A and D). In addition, RPL11 showed the strongest positive correlation with oxidative phosphorylation, with a correlation coefficient of 0.75 in biological state or biological process correlation analysis (Fig. 4E). Biomarkers were all significantly enriched in the ribosome and the spliceosome. (A-D) GSEA analysis results of biomarkers. (E) Correlation analysis between biomarkers and biological states or biological processes. Differential immune cells identified in the training set In the immune infiltration analysis, the proportions of 22 immune cells were calculated for each sample in the training set (Fig. 5A), among which CD8 T cells and neutrophils showed the strongest negative correlation with a correlation coefficient of -0.62. Plasma cells and M0 macrophages showed the strongest positive correlation with a correlation coefficient of 0.37 (Fig. 5B). In addition, the abundance of M0 macrophages, monocytes in DR samples was higher, while the opposite was true for activated memory CD4 T cells (Fig. 5C). Subsequently, activated memory CD4 T cells showed negative and positive correlations with E2F2 (cor=-0.356) and RPL11 (cor = 0.362), respectively. Meanwhile, there was also a significant positive correlation between CD36 and monocytes (cor = 0.438) (Fig. 5D). Differential immune cells identified in the training set. (A) In the immune infiltration analysis, the proportions of 22 immune cells for each sample in the training set are shown. (B) Heatmap of immune cell correlation. (C) Comparison of differences in the abundance of immune cell infiltration between disease and normal samples. (D) Bar chart of the correlation between biomarkers and differential immune cells. Regulatory networks of biomarkers were generated In this study, the 184 miRNAs predicted by ENCORI and 136 miRNAs predicted by NetworkAnalyst were intersected to obtain 14 targeted miRNAs. Based on this result, a total of 33 lncRNAs were also predicted. The regulatory relationships they formed in the lncRNA-miRNA-mRNA network included MALAT1-hsa-miR-96-5p-DDIT3, XIST-hsa-let-7d-5p-E2F2, and so on (Fig. 6A). Afterward, we also predicted 25 RBP in the ENCORI database, and they formed 38 regulatory relationships with biomarkers, including E2F2-CHTOP, DDIT3-SCAF4, and so on (Fig. 6B). Regulatory networks and drug prediction network of biomarkers were generated. (A) mRNA-miRNA-lncRNA regulatory relationship network. (B) Biomarker-RBP regulatory relationship network. (C) Drug prediction network. Diseases and drugs corresponding to biomarkers The diseases associated with biomarkers were also predicted in this study, where platelet glycoprotein IV deficiency, steatohepatitis, white blood cell count procedure, and ventricular septal defects were associated with CD36, DDIT3, E2F2, and RPL11, respectively (Supplementary Figure S2-5). Subsequently, the drugs of biomarkers were predicted, among which DDIT3, RPL11, CD36, and E2F2 corresponding drugs included Curcumin, Exaluren, ABT-510, and Edifoligide Sodium, respectively (Fig. 6C). Discussion DR is a severe vision-threatening ocular disease worldwide. Although the exact pathogenesis has not been fully clarified in DR patients, oxidative stress—driven by the accumulation of ROS—appears to play a pivotal role in DR onset and progression [8–11]. Therefore, exploring the biological pathways and molecular associations underlying the link between oxidative stress and DR pathogenesis and progression at the molecular level is of critical importance for further understanding the pathogenesis of DR and guiding clinical prevention and treatment. Notably, a recent large-scale proteomic study [13] has systematically identified risk biomarkers and predictive models for DR at the plasma protein level. In contrast, the present study was performed at the transcriptome level. In this study, based on the analysis of DR-related public datasets (GSE221521 and GSE189005), we identified 52 oxidative stress-related DEGs that were distinct in the DR group relative to the control group. Among these DEGs, CD36, DDIT3, E2F2, and RPL11 were further identified as potential biomarkers associated with oxidative stress in DR, a diagnostic potential strongly supported by receiver operating characteristic (ROC) curve analysis. The biomarkers and ribosome and spliceosome pathways identified in this study were not highlighted in the aforementioned proteomic research, thereby providing valuable and complementary bioinformatic evidence for a more comprehensive understanding of DR. The CD36 gene encodes a member of the class B2 scavenger receptor family. This gene is expressed in retinal myeloid cells [14], retinal Müller glial cells [15], retinal neovascularization [16], and peripheral blood [17]. Under hyperglycemic conditions, CD36 expression is upregulated in endothelial cells, which may lead to increased oxidized low-density lipoprotein (OxLDL) uptake [18], which in turn enhances microvascular damage [19]. One of the main manifestations of DR is dysfunction of the blood-retinal barrier (BRB), which can lead to lipoprotein leakage into retinal tissue. The accumulation of LDL in the retina only occurs when the BRB is abnormal, whereas individuals with a normal BRB can prevent the leakage of LDL [20]. Early researchers [21] have demonstrated that OxLDL exists in the retina of diabetic patients, even before the occurrence of clinical retinal lesions, but does not exist in non-diabetic retinopathy (NDR). In the diabetic context, OxLDL in human plasma only accounts for a small part of the total LDL, but most of it is modified into glycosylated OxLDL [22,23]. CD36 is one of the receptors for oxidized lipoproteins and can promote the uptake of such oxidized lipoproteins [24,25]. This also suggests that most of the lipoproteins infiltrating and accumulating in the retina have undergone pathological modifications. Related studies indicate that CD36 may participate in the uptake of oxidized low-density lipoproteins [25] and is involved in the regulation of microvascular endothelial function and insulin signaling [19]. Additionally, glycosylated oxidized low-density lipoproteins may exert certain cytotoxic effects on retinal cells, exacerbate local oxidative stress, and consequently affect the survival and physiological functions of retinal vascular cells and retinal pigment epithelial cells. DDIT3 (CHOP, c/ebp homologous protein), an important transcription factor, critically regulates the cellular stress response, especially endoplasmic reticulum stress [26]. Previous studies have shown that CHOP is a pro-apoptotic transcription factor that can induce ROS production [27]. However, the overexpression of hydrogen peroxide (H2O2), one of ROS, will trigger the interruption of protein folding, and the accumulation of misfolded or unfolded proteins in the endoplasmic reticulum cavity [28], which will activate endoplasmic reticulum stress [29] and upregulates CHOP expression. It contributes significantly to the accelerated apoptosis of retinal ganglion cells (RGCs) [30,31]. Apoptosis of RGCs apoptosis is the main pathological manifestation of DR neuro-degeneration [32]. Further supporting this mechanism, there is a negative correlation between RGCs cell lines (mRGCs) and CHOP when inducing ROS overexpression [33]. These findings confirm that DDIT3 is indeed involved in DR progression, which is consistent with our research results. In addition, H2O2 produced under hyperglycemia not only induce apoptosis in retinal nerve cells (including RGCs), but also trigger the occurrence of the degenerative cascade of the microvascular system, including the destruction of BRB [34,35], in other words, H2O2 involved in DR progression through another aspect. To our knowledge, this study provides the first preliminary evidence linking E2F2 and RPL11 to DR. Notably, these two genes have been predominantly studied in the context of cancer due to their close association with tumor progression, and their potential associations with DR-related oxidative stress remain unreported. Thus, these observations may provide testable hypotheses for future investigations into their potential functions in DR, rather than confirming established mechanisms. The transcription factor E2 (E2F) family plays a key regulatory role in transcriptional control of cancer gene networks, and is crucial for the progression of the cell cycle, proliferation, apoptosis, and differentiation [36,37]. E2F2 is an important member of the E2F family [38–41], and its abnormal expression is strongly associated with unfavorable clinical outcomes in various cancers. However, whether and how E2F2 participates in DR remains largely unknown and requires further validation. The ribosomal protein L11 (RPL11) encoded by the RPL11 gene is a component of the ribosomal 60 S subunit. Recent studies have confirmed that RPL11 can stabilize and activate p53 by binding to and inhibiting the E3 ubiquitin ligase activity of murine double minute 2 (MDM2), forming the RP-MDM2-p53 pathway [42,43]. This pathway is involved in regulating ribosomal homeostasis, nutrient sensing, and metabolic balance, and is particularly important under conditions of high glucose and oxidative stress [44,45]. RPL11 is the first ribosomal protein confirmed to interact with MDM2 and plays a critical role in maintaining p53 stability and cellular metabolic homeostasis. This study found that RPL11 expression was decreased in DR patients. We speculate that downregulation of RPL11 may attenuate the activity of the RP-MDM2-p53 pathway, which could potentially contribute to metabolic disturbances and retinal dysfunction. These inferences remain hypothetical and require further experimental validation. This study also performed GSEA analysis, with results showing significant enrichment of biomarkers in two signaling pathways: ribosomes and spliceosomes. While these pathways have not been extensively studied in DR, emerging evidence suggests their potential association with oxidative stress processes linked to disease progression. Ribosomes are organelles mainly composed of proteins and RNA (one large and one small subunit), with the primary function of protein translation—critical for cell survival and growth. Previous studies demonstrate that oxidative stress downregulates intracellular translation [46,47]. For example, in yeast (exposed to H₂O₂ or chronic oxidative stress) or Drosophila embryos under oxidative stress, cysteine residues in ribosomal proteins are preferentially oxidized [48]. Among these ribosomal proteins, Rps26 and Rpl10 exhibit the highest oxidation levels, indicating they are most sensitive to oxidative stress [47–49]. Yang et al. [50] further confirmed this, supporting ribosomes as key targets of oxidative stress—suggesting oxidative modification of ribosomal proteins (e.g., Rps26, Rpl10) may impair protein translation and thus be associated with oxidative stress-related DR pathological progression. In most eukaryotes, genes are transcribed into precursor mRNA (pre-mRNA), which must undergo splicing to become mRNA capable of being translated into proteins; this splicing step is critical for gene expression. Splicing is catalyzed by spliceosomes—dynamic ribonucleoprotein complexes with variable conformations and compositions—that ensure the accuracy and flexibility of the splicing process. Abnormal pre-mRNA splicing underlies increased severity of many human diseases—e.g., urothelial carcinoma [51]. However, research on the spliceosome pathway’s relationship with DR processes remains limited. Given the spliceosome pathway’s enrichment in our biomarker GSEA analysis and its core function in mRNA maturation, we reasonably speculate that under DR-related pathological conditions (e.g., hyperglycemia, oxidative stress), spliceosomes may indirectly be associated with DR progression by modulating the quantity and quality of biomarker production. However, the causal relationship between these pathways and DR pathogenesis requires further experimental confirmation. This study performed immune infiltration analysis on the training set and found significant differences between the control and DR groups: M0 macrophages and monocytes were upregulated, while activated CD4 + T cells were markedly elevated in the DR group. Previous studies have also observed CD4 + T cell accumulation in DR patients’ vitreous [52]. Additionally, recent bioinformatics studies reveal the DR biomarker OSER1 drives DR progression by affecting activated CD4 + T cells to promote inflammation and vascular damage [53]; CD4 + T cells also synthesize VEGF under pathological conditions, and when aggregated at inflammatory sites to directly contact endothelial cells, even low VEGF secretion suffices to promote local neovascularization [54]. It should be emphasized that these immune infiltration results are derived from in silico prediction and lack experimental verification; therefore, they can only be regarded as exploratory hypotheses rather than definitive biological evidence. This study identifies the critical associations of oxidative stress-related genes—including the previously reported CD36 and DDIT3, as well as the newly identified E2F2 and RPL11—in DR pathological progression. Additionally, our findings highlight the potential involvement of the ribosomal pathway, spliceosome pathway, and CD4 + T cell-mediated immune infiltration in disease development, providing new insights for understanding DR pathogenesis. It is worth noting that both the training set and validation set in this study were derived from peripheral blood samples rather than retinal tissue, which is an important limitation of this research. In clinical practice, retinal tissue from living patients cannot be obtained non-invasively, and most publicly available retinal transcriptomic data come from cadaveric donor samples. To address this issue, we further validated the expression levels of the four core biomarkers in an independent retinal tissue dataset. Encouragingly, the expression trends of CD36, DDIT3, E2F2, and RPL11 in peripheral blood were completely consistent with those in retinal tissue, demonstrating that peripheral blood-based transcriptomic signatures can reflect the pathological changes of diabetic retina. Although these molecular features partially reflect the systemic diabetic state of the body, the consistent aberrant expression patterns in both tissues suggest that they may serve as systemic indicators to map the retinal pathological processes specific to DR. Furthermore, qRT-PCR validation was performed using only 5 pairs of clinical samples with a limited sample size. Only DDIT3 and RPL11 achieved preliminary verification with expression trends consistent with the training dataset, whereas validation of CD36 and E2F2 requires further confirmation in larger cohorts. Some core molecular mechanisms remain incompletely elucidated. For instance, the putative molecular interactions by which RPL11 may mediate metabolic abnormalities in DR through the RP-MDM2-p53 pathway are based on computational inference and remain hypothetical, requiring further experimental validation. Future research could proceed in two directions: on one hand, expanding the sample size and incorporating multi-center cohorts to validate the reliability of the core findings; on the other hand, combining DR animal models with in vitro retinal cell experiments to deeply analyze the associations between genes and pathways, to provide a more solid theoretical foundation for the screening of potential therapeutic targets and the development of immune intervention strategies for DR. Materials and methods Data source In this study, DR-related datasets were all obtained from the GEO databases (https://www.ncbi.nlm.nih.gov/geo/). Among them, GSE221521 (GPL24676) was used as a training set, including 69 DR and 50 control blood leukocyte samples. While GSE189005 (GPL23126) contained a total of 45 samples, the blood samples of 10 DR patients with type 2 diabetes and 9 controls were used as the validation set. In addition, we also obtained 986 OSRGs from the GSEA database (http://www.gsea-msigdb.org/gsea/index.jsp). In particular, gene expression validation was performed using the GSE160306 (GPL20301) and GSE60436 (GPL6884) datasets. The GSE160306 dataset consisted of 20 normal retinal samples and 39 retinal samples from patients with DR. The GSE60436 dataset contained 3 control fibrovascular membrane samples and 6 DR-associated fibrovascular membrane samples. All datasets were downloaded from the GEO database using the GEOquery package (v 2.68.0). The raw count expression matrix, FPKM expression matrix of the training set, and RMA-normalized expression matrix of the validation set were obtained accordingly. No batch effect was observed within the intra-dataset analyses. Identification and functional analysis of candidate genes In GSE221521, the DEGs in control and DR groups were screened based on the DESeq2 package (v 1.40.2) [55] using |log2FC|>0.5 and p.adjust < 0.05 as the screening criteria. This threshold enables the identification of early and subtle yet pathologically meaningful molecular alterations in genes [56]. Moreover, the candidate genes were obtained by taking their intersection with OSRGs. Subsequently, the candidate genes were applied for GO and KEGG [57–59] functional enrichment to explore their biological functions by the clusterProfiler package (v 4.8.2) based on the background gene set in org. Hs.eg.db package (p.value < 0.05) [60,61]. Afterward, the candidate genes were subjected to master regulator analysis (MRA) with all the genes in the training set, and the top 5 candidate genes with the smallest p-values were selected for presentation. Protein-protein interaction (PPI) network construction and machine learning A PPI network of candidate genes was constructed based on the STRING database, in which genes that had interactions at the protein level were taken as candidate hub genes for machine learning analyses (interaction score ≧ 0.4). To avoid bias caused by feature selection using only a single method, three independent algorithms were adopted for gene screening. Specifically, in the training set, Boruta analysis was performed on candidate hub genes using the Boruta package (v 8.0.0) [62]. Based on random forest iterative denoising, genes identified as confirmed were screened as feature genes. Subsequently, SVM-RFE analysis was conducted via the e1071 package (v 1.7.13) [63], with the method set as svmLinear. 10-fold cross-validation was performed, and genes corresponding to the minimum error rate were selected as feature genes. Furthermore, LASSO regression analysis was implemented using the glmnet package (v 4.1.8) [64], with parameters set as family = binomial and type.measure = deviance. 10-fold cross-validation was applied to calculate error rates under different feature combinations, and genes with non-zero coefficients at the optimal minimum lambda value corresponding to strongly correlated features were identified as feature genes. Finally, the intersection of feature genes obtained from the three machine learning algorithms was taken to identify the final hub genes.Thereafter, the final hub genes were obtained after taking the intersection of these feature genes screened by machine learning. Identification of biomarkers and construction of a nomogram In GSE221521 and GSE189005, the expression of final hub genes in control and DR groups was analyzed by a rank-sum test, where genes with consistent and significant expression trends in both datasets were used as candidate biomarkers. Afterwards, their ability to discriminate between DR and control samples was further assessed by ROC curves, where genes with AUC values greater than 0.7 in both datasets were used as biomarkers. Then, to explore the diagnostic role of the biomarkers, nomogram of the biomarkers was constructed in the training set using the rms (v 6.7.1) package to predict the incidence of DR [65], and their predictive performance was further evaluated by ROC curves, decision curve analysis (DCA) and calibration curves. Meanwhile, the differential expression of biomarkers was further analyzed by two-sided Wilcoxon rank-sum test in both GSE160306 and GSE60436 datasets (p < 0.05). Functional analysis of biomarkers In order to explore the biological pathways involved in biomarkers, the correlations between biomarkers and other genes were calculated separately in the training set to get their ranking. After that, GSEA of the biomarkers was performed by clusterProfiler (v 4.8.2) package based on KEGG gene set (c2.cp.kegg.v7.5.1.entrez.gmt) [60], and the top 5 results with the smallest p.adjust were selected for the display. Biological state or biological process correlation analysis In order to explore the relationship between the onset of DR and the biological state or biological process of the body, correlation analysis was carried out on them. In the training set, the ssGSEA algorithm in the GSVA package (v 1.48.3) [66] was employed to evaluate the biological state or biological process of all disease samples, which were subjected to Spearman correlation analysis with biomarkers based on the gene set (h.all.v2023.2.Hs.symbols.gmt). Immune infiltration analysis and biological process correlation analysis To explore the differences in immune cell infiltration between DR and control samples, the proportions of 22 immune cells in all samples of the training set were calculated based on the CIBERSORT algorithm and the immune cell LM22 gene set, and their correlation was also analyzed by the Spearman method. Afterwards, the differences in immune cells between DR and control groups were explored by Wilcoxon analysis, where the correlation between the differential immune cells and the biomarkers was also explored. Molecular regulatory network analysis In order to investigate the molecular regulatory mechanisms of biomarkers, upstream miRNAs of biomarkers were predicted in ENCORI and NetworkAnalyst databases. The predicted miRNAs in these two databases were intersected to obtain the target miRNAs. After that, the corresponding lncRNAs of the target miRNAs were also predicted in the ENCORI database with geneType=lncRNA and clipExpNum ≧ 5 as the screening criteria. Moreover, an lncRNA-miRNA-mRNA network was constructed by Cytoscape (v 3.7.1) [67]. In addition, RNA-binding proteins (RBP) associated with the biomarkers were predicted in the ENCORI database with the thresholds set to clusterNum greater ≧ 5 and clipExpNum ≧ 5. Similarly, the biomarker-RBP regulatory network was visualised via Cytoscape. Disease and drug prediction for biomarkers To explore the role of biomarkers in DR and other complications or metabolic diseases of DR, their relationship with biomarkers was analyzed using the CTD database. Besides, to obtain potential therapeutic drugs for DR, the biomarkers were uploaded to the DGIdb database to obtain the relevant drugs, and the biomarker-drug network diagrams were visualized using Cytoscape software. QRT-PCR The expression levels of biomarker were further verified by qRT-PCR. We analyzed 5 pairs of clinical samples from Hebei General Hospital using qRT-PCR, with each pair consisting of one DR sample and one normal sample. The study was approved by the Ethics Committee of Hebei General Hospital (No. 2025-LW-0149). All patients had signed an informed consent form. First, the total RNA of the frozen sample was extracted by TRIzol reagent. Then, an equal amount of mRNA was reverse transcribed to synthesize cDNA by SureScript-First-strand-cDNA-synthesis-kit. Next, the qPCR reaction was performed with 2 x Universal Blue SYBR Green qPCR Master Mix kit according to the reaction system. We calculated the relative gene expression using the 2‑ΔΔCT method. Statistical analysis Statistical analysis was executed by R software (v 4.1.0), and the data from different groups were compared by the Wilcoxon test. If not specified above, a p-value less than 0.05 was considered statistically significant (Table 1). Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Author contributions Jianmin Wang is responsible for conceptualization, methodology, writing– original draft, writing– review and editing, resources, funding acquisition. Liangjie Niu is responsible for software, validation, formal analysis, writing– original draft, writing– review and editing. Yifan Zhang is responsible for investigation, data curation, writing– original draft, writing– review and editing. Huika Xia is responsible for visualization, writing– original draft, writing– review and editing, supervision, project administration. Funding This work was supported by the Natural Science Foundation of Hebei Province (Grant No. H2020206650), Hebei Provincial Government funded Clinical Medicine Excellent Talents Project, and Hebei Province Medical Application Technology Tracking Project (Grant No. GZ2023091). Data availability The datasets (GSE189005 and GSE221521) supporting the conclusions of this article are available in the [GEO] repository, [https://www.ncbi.nlm.nih.gov/gds]. Declarations Competing interests The authors declare no competing interests. Ethics This study was performed in line with the principles of the Declaration of Helsinki and approved by the Ethics Committee of the Ethics Committee of the Hebei General Hospital. The approval number and date of approval are as follows: No. 2025-LW-0149. Consent to participate All patients provided written informed consent at the time of clinical sample collection for experiments, ensuring that the research process complied with ethical norms and that patients’ rights and wishes were fully respected. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Data Availability Statement The datasets (GSE189005 and GSE221521) supporting the conclusions of this article are available in the [GEO] repository, [https://www.ncbi.nlm.nih.gov/gds].