Study on the differential expression of disulfidptosis-related genes and their association with immune regulation in patients with diabetic retinopathy

Abstract

Diabetic retinopathy (DR), a frequently encountered microvascular complication of diabetes, currently lacks effective treatment options due to the complexity of its underlying pathophysiological mechanisms. The identification of disulfidptosis as a subtype of cell death opens up novel avenues for investigating the pathogenesis of several diseases. This study aims to identify and validate differentially expressed disulfidptosis-related genes (DRGs) in peripheral blood samples of patients with DR, and to explore their association with immune cell infiltration. Clinical patient datasets (GSE221521) were obtained from public online databases. Based on this dataset, differential expression, correlation, and enrichment analyses of DRGs were performed using R software to determine their potential mechanisms of action. False Discovery Rate (FDR) correction was employed to reduce false positive results (significance threshold at FDR < 0.05) based on the Benjamini-Hochberg method. Subsequently, the CIBERSORT algorithm was deployed to assess the infiltration levels immune cell associated with the differentially expressed DRGs, in order to explore immune dysregulation in the context of DR. Meanwhile, nomograms, calibration curves, ROC curves, nomograms, and decision curve analyses were conducted to validate the accuracy of key genes and construct a disease prediction model for assessing DR risks. Finally, the differentially expressed DRGs were validated using clinical samples from DR patients. Based on the GSE221521 dataset, significantly differential expressions of eight disulfidptosis-related genes were observed, and individual validation using clinical samples confirmed consistent expression patterns for FLNB, GYS1, FLNA, PRDX1, among which FLNB and GYS1 showed statistically significant differences. Immune infiltration analysis revealed that five DRGs (TLN1, FLNA, PRDX1, FLNB, and GYS1) were strongly correlated with macrophages, CD4 memory activated T cells, and M0 monocytes in DR patients. Functional enrichment analysis highlighted the involvement of platelet aggregation and activation, as well as Rap1 signaling, in the initiation and development of the disease. A joint predictive model was constructed based on eight differentially expressed DRGs, and achieved an AUC of 0.818, significantly outperforming single-gene models. This model was visualized as a nomogram to facilitate rapid assessment of individual risk based on gene expression patterns for early risk prediction and personalized intervention. However, this prediction model was built on a single dataset and required further validation in other independent queues. This is the first study to identify a strong association between DR and disulfidptosis, providing a novel perspective for identifying biomarkers and potential treatment strategies for DR.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38671-9.

Introduction

Diabetic retinopathy (DR), an irreversibly blinding disorder featured with retinal microvascular dysfunction^1,2, represents the most critical microvascular syndrome of advanced diabetes mellitus (DM). DR is projected to threaten the health of 191 million individuals by 2030³, and its prevalence is increasing globally, making it a major public health issue^4,5. It not only deteriorates the life quality for diabetic population but also brings a significant economic burden on society⁶. Current therapeutic strategies for DR are contingent on disease severity. These include intravitreal injection of anti-VEGF agents, retinal laser photocoagulation⁷ and vitreoretinal surgery⁸. However, many patients fail to achieve clinically significant visual improvement despite these interventions. The pathogenesis of DR is intricate, containing oxidative stress^9,10, inflammatory responses^11–13, autophagy^11,14, and various forms of programmed cell death¹⁵.

Programmed cell death (PCD), containing, pyroptosis^15,16, ferroptosis¹⁷, apoptosis¹⁸, cuproptosis¹⁹, necroptosis, and other forms²⁰, function in maintaining homeostasis and self-renewal^21,22, and is involved in DR progression. As a subtype of PCD, pyroptosis is featured with rapid plasma membrane permeabilization and secretion of intracellular contents alongside pro-inflammatory factors like IL-1β and IL-18²³ in driving the DR progression. Gu et al. demonstrated that in high glucose-treated retinal microvascular endothelial cells and clinical vitreous samples, both caspase-1 and IL-1β expressions were significantly unregulated, suggesting the involvement of pyroptosis in DR pathogenesis. Furthermore, in retinal microvascular endothelial cells, downregulation of miR-590-3p and upregulation of NLRP1/NOX4 suggest that decreased miR-590-3p promotes pyroptosis via a pathway involving NLRP1-mediated activation of NOX4²⁴. Pericytes contribute to structural integrity of the vascular wall and regulates tight junction proteins expressions in adjacent endothelial cells²⁵. Previous study demonstrated that high glucose triggers dose- and time-dependent activation of the NLRP3-caspase-1-GSDMD signaling pathway, culminating in plasma membrane pore formation and the secretion of inflammatory cytokines IL-1β and IL-18, along with lactate dehydrogenase from human retinal pericytes-hallmarks of pyroptosis²⁶. Earlier investigations linked necrosis to DR progression, as evidenced by enhanced necrotic death of retinal pericytes in diabetic rat models²⁷. Subsequent research has highlighted necroptosis as a critical pathogenic mechanism in blinding diseases containing age-associated macular degeneration, glaucoma, and retinitis pigmentosa, associated with the apoptosis of RPE cells, RGCs, and cone photoreceptors, respectively²⁸. Similarly, mechanistic studies in retinal ischemia-reperfusion injury models have shown that necroptosis mediates neuronal damage via the ERK1/2-RIP3 cascade, suppressed by necrostatin-1^29,30. Given that retinal ischemia underlies in DR, optic neuropathies, and retinopathy of prematurity. DR-related necroptosis studies have primarily focused on retinal ganglion cells (RGCs), with evidence showing that hyperglycemia promotes the expression and phosphorylation of RIPK1/RIPK3, and that necrostatin-1 effectively mitigates necroptotic death of RGCs induced by glucose³¹.

DR is a progressive condition with a long disease course, leading to vision loss in advanced stages. Thus, the determination of key genes for early-stage diagnosis and the exploitation of targeted therapeutic strategies remain of critical importance in managing DR. Recently, research by Xiaoguang Liu et al. determined a novel condition of PCD termed disulfidptosis³². The pathogenesis of DR involves multiple modes of cell death, suggesting that disulfidptosis may represent an additional mechanistic contributor to DR progression. Specifically, we noted that disulfide bonds could affect protein folding, oxidative stress, and redox homeostasis, all of which were related to cellular homeostasis and apoptosis. In the context of DR, chronic hyperglycemia and oxidative stress might disrupt disulfide bond dynamics, thereby contributing to retinal cell damage and vascular dysfunction. Accordingly, the validation and bioinformatic analysis of DRGs signatures may put new mechanistic insights on the diagnosis and treatment of DR. Currently, research on disulfidptosis remains in its early stages, with most studies focusing on cancer, Investigations in non-cancer diseases are limited, and the potential relationship between disulfidptosis and DR remains largely unexplored.

Through this investigation, we employed the GSE221521 dataset related to DR from the GEO database, integrating it with 16 DRGs previously reported by Liu et al.³². Based on these differentially expressed DRGs, we employed PPI analysis, GO annotation, and KEGG pathway enrichment analysis, and validation assays on clinical samples from DR patients. We further evaluated the levels of immune infiltration levels in active-phase DR samples to explore the immune microenvironment in DR patients. Additionally, the correlations of key gene expression and immune infiltration were also assessed. Finally, the potential diagnostic potential of these genes in DR was also evaluated by a risk-prediction model based on DRGs.

Methods

Data sources and preprocessing

Across this work, we analyzed relevant datasets from the Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo) via the keyword “diabetic retinopathy”. The GSE221521 dataset was selected as the discovery cohort, which is a prospective cohort comprising 50 healthy controls, 74 diabetic patients without DR, and 69 diabetic patients complicated with DR. Peripheral blood samples from 69 diabetic patients complicated with DR and 50 healthy controls were selected (Table 1). Gene expression data from GSE221521 were put into normalization and preprocessed via the “limma” package in R, including probe annotation and batch effect correction. Principal component analysis of the normalized data did not reveal any clustering driven by sub-groups or processing dates. Therefore, no additional batch correction was applied within this dataset, as the standard limma preprocessing pipeline (background correction, log2 transformation, quantile normalization) was deemed sufficient to address technical variations across arrays. A sum of 16 DRGs were retrieved from previously published studies³².

For the clinical validation, a total of 39 peripheral blood mononuclear cells (PBMCs) samples were enrolled from 21 healthy people and 18 DR patients in the department of ophthalmology at our hospital. Inclusion Criteria: (1) Age > 30 years, able to cooperate with all study-related examinations. (2) Diagnosed with T2DM according to American Diabetes. Exclusion Criteria: (1) Type 1 diabetes. (2) Poor systemic condition (e.g., cerebral infarction, myocardial infarction, uncontrolled hypertension, liver or kidney dysfunction). (3) Systemic diseases potentially affecting gene expression (e.g., malignancy, autoimmune diseases, viral infections). (4) Acute infection. (5) History of retinal laser photocoagulation, anti-VEGF injection, or vitrectomy. (6) History of intraocular surgery except cataract surgery. (7) Other retinal diseases affecting assessment (e.g., retinal vein occlusion, age-related macular degeneration). (8) Optic nerve diseases (e.g., glaucoma, ocular trauma). Besides, the detailed baseline characteristics (age, sex, BMI, diabetes duration, fasting glucose, HbA1c%, lipid profiles) for the screening cohort were now provided in Table 1.

Table 1.

Demographic and clinical characteristics of the subjects from the confirmation cohort.

Age (years)	Ctrl (n = 21)	DR (n = 18)	P value
	62.10 ± 4.34	65.44 ± 6.68	0.068
Sex (male/female)	(8/13)	(11/7)
BMI (kg/m2)	21.97 ± 3.18	23.46 ± 3.10	0.150
Fasting glucose (mmol/L)	5.20 ± 0.92	7.66 ± 3.25	0.006
HbA1c (%)	5.83 ± 0.58	7.62 ± 1.46	0.000
TC (mmol/L)	4.83 ± 0.82	4.77 ± 1.35	0.873
TG (mmol/L)	1.45 ± 1.07	2.12 ± 1.10	0.062
HDL-C (mmol/L)	1.25 ± 0.23	1.17 ± 0.15	0.242
LDL-C (mmol/L)	2.69 ± 0.67	2.28 ± 0.90	0.110

The data is presented as the mean ± SD, unless explicitly indicated otherwise.

Boldface values indicate statistically significant differences at p < 0.05.

DM, diabetes mellitus; BMI, Body Mass Index; HbA1c, glycated hemoglobin A1c; TC, total cholesterol; TG, triglyceride; HDL-C, high-density lipoprotein; LDL-C, low-density lipoprotein.

Expression analysis of DRGs

To determine differentially expressed DRGs between groups in GSE221521, the “limma” R package and the Wilcoxon test were carried out at a significance threshold of P < 0.05. To control for multiple comparisons and reduce false positive results, False Discovery Rate (FDR) correction was applied using the Benjamini-Hochberg method to adjust all p-values. The corrected p-values were considered statistically significant when FDR < 0.05 and identifying eight genes: FLNA, TLN1, PRDX1, FLNB, ACTB, NDUFS1, RPN1, GYS1 with log2FC values of 0.05, 0.05, −0.04, 0.05, 0.02, 0.02, −0.02, 0.057 respectively. The data was visualized using “pheatmap” and “ggpubr” R packages. To investigate the relationships among the differentially expressed DRGs, correlation analysis was performed and visualized via the “corrplot” and “circlize” R packages.

GO and KEGG pathway analysis

GO enrichment analysis was conducted to comprehensively investigate the biological processes of differentially expressed DRGs and to determine their roles in the development of DR. The analysis encompassed categories of Molecular Function (MF), Biological Process (BP), and Cellular Component (CC), which describe the biochemical activities, associated biological pathways, and subcellular localization of gene products, respectively. To determine the underlying molecular mechanisms of differentially expressed DRGs, KEGG pathway enrichment analysis was also performed. The Database for Annotation, Visualization, and Integrated Discovery (DAVID, https://david.ncifcrf.gov/) was utilized for functional annotation and pathway analysis. Eight differentially expressed DRGs were uploaded to the DAVID database and analyzed at P < 0.05.The data was visualized via R software (version 4.1.1), enabling the identification of key biological functions and signaling pathways related to differentially expressed DRGs in DR pathogenesis.

Immune infiltration analysis of differentially expressed DRGs

The CIBERSORT algorithm was applied for assessing the relative proportions of 22 immune cell types within each tissue sample. To assess the systemic immune signatures, immune scores were measured for each sample via the ESTIMATE algorithm. Furthermore, the relationships of signature genes with the abundance of infiltrated immune cells were predicted via Spearman’s rank correlation analysis. The graphical functions of the ggplot2 package were used to visualize the resulting correlations.

Establishment and validation of a nomogram model for DR diagnosis

After the selection of key differentially expressed DRGs a nomogram model for the prediction of DR occurrence was constructed via the “rms” package. “Points” represent the individual scores of each factor, and “Total Points” indicate the sum of all factor scores. To obtain the diagnostic predictive value of differentially expressed DRGs for DR patients, the “pROC” package was deployed to calculate the AUC for model accuracy assessment; logistic regression was applied for individual gene analysis, while LASSO regression was used for multiple-gene combination analysis. Specifically, this study employed L2-regularized logistic regression (ridge regression) to construct a joint predictive model. Initially, the input features of DRGs (FLNA, TLN1, PRDX1, FLNB, ACTB, RPN1, NDUFS1, GYS1) were standardized using z-score normalization. The logistic regression model was developed to learn the gene coefficients. And L2 regularization was applied to enhance the model’s generalization capability and reduce overfitting. To optimize performance, 5-fold cross-validation worked for evaluating the model’s stability and generalization. In each iteration, one subset was used as the validation set, and the remaining four subsets as the training set. This approach ensured the consistency of the model’s performance on different data partitions. A larger AUC value indicates higher predictive accuracy of the model. Finally, decision and clinical impact curves could predict the clinical efficacy of the model.

Subjects

The participants for this study were included from the Affiliated People’s Hospital of Ningbo University between 2019 and 2022, and each individual provided informed consent prior to enrollment. Ethical approval was granted by the Research Ethics Board (REB) of the hospital (approval number 2019-048). All procedures performed on participants were carried out by licensed physicians and ophthalmologists. The diagnostic principle of T2DM and DR in the selected patients adhered to the guidelines established by the American Diabetes Association. Exclusion criteria contained type 1 diabetes, prior eye disease or surgery, uncontrolled hypertension, liver or kidney dysfunction, autoimmune disorders, and malignancies. Ethical conduct throughout the study followed the principles outlined in the Declaration of Helsinki.

Extracting RNA and quantitative real-time PCR

RNA was isolated from PBMCs, aqueous humor, and human retinal microvascular endothelial cells (hRMECs) via TRIzol reagent (Sigma-Aldrich, St. Louis, MO, USA). Due to the small volume of aqueous humor samples, RNA extraction was conducted using 100 µl of pooled samples from three patients. The isolated RNA was then reverse transcribed into cDNA via he PrimeScriptTM RT reagent (Takara Bio Inc., Kusatsu, Shiga, Japan). Real Time-qPCR was processed via the SYBR Premix Ex Taq II kit (Takara Bio Inc., Kusatsu, Shiga, Japan) to quantify mRNA levels on a 7500 Fast Real-Time PCR System (Thermo Fisher Scientific, Waltham, MA, USA). The β-actin RNA level was used as the internal control, and the relative Gene mRNA expression was calculated by the 2^−ΔΔCt method. The primers are indicated in Table 2.

Table 2.

Primer set.

Gene name	Sequence (5’ -> 3’)
FLNA-F	AGAAGAAGATGCACCGCAAG
FLNA-R	GATGAGGCCCAGGATCAGCTT
TLN1-F	TATGGCTGGAGGCTGGGAAAG
TLN1-R	ACCATGATCGTCTTCACAGTTC
PRDX1-F	CCACGGAGATCATTGCTTTCA
PRDX1-R	CCCTGAACGAGATGCCTTCAT
FLNB-F	CTACTTGCCCATCACCAACTT
FLNB-R	CGTGCATTATCCACAGGCTTC
RPN1-F	CACAGTCAAGCTCCCAGTTGC
RPN1-R	GGTTTGTGTCTTCGTTGGAT
NDUFS1-F	TTAGCAAATCACCCATTGGACTG
NDUFS1-R	GCAAACCTGATGCAGCGAG
GYS1-F	GCGCTCACGTCTTCACTACTG
GYS1-R	ACAAACTCCTGGATTCGAGCC
β-actin-F	GGGCCAACCGCGAGAAGATGAC
β-actin-R	CAGGTCCAGACGCAGGATGG

Statistical analysis

Bioinformatics part: Statistical analyses were primarily addressed by conducting GraphPad Prism software (version 8.0.2), while all plots were generated by R software (version 4.1.1). For intergroup comparisons, the Student’s t-test was utilized for normally distributed data, whereas the Wilcoxon rank-sum test was deployed for non-normally distributed variables. For multiple group comparisons, one-way analysis of variance (ANOVA) combined with appropriate post hoc tests was employed. Correlation analyses were developed by Spearman’s rank correlation test. All statistical tests were subjected to FDR correction.

Experiment part: All data are presented as mean ± standard deviation (SD) for tabulated results and as mean ± standard error of the mean (SEM) in graphical presentations. All experiments were independently repeated three times. In the clinical validation section, a Student’s t-test was applied to compare gene expression differences between the DR group and the healthy control group. A p-value < 0.05 was deemed as statistically significance.

Effect size and confidence interval estimation: In addition to statistical significance (p-values and FDR), the magnitude of differential expression was assessed using standardized effect sizes. For pairwise comparisons between the DR and control groups, Cohen’s d was calculated as the unbiased estimator of the standardized mean difference, along with its 95% confidence interval (CI). The reporting of 95% CIs provides an estimate of the precision of the effect size measurement. All effect size calculations were performed using the effsize package in R (version 4.1.1).

Results

Identification of differentially expressed DRGs

The expression matrix of DRGs was constructed by integrating the GSE221521 dataset with a predefined list of DRGs. Eight differentially expressed DRGs were selected and visualized using boxplots (Fig. 1A) and heatmaps (Fig. 1B). In the DR group, FLNA, TLN1, FLNB, ACTB, NDUFS1, and GYS1 had the significant upregulation of expression (P < 0.05), whereas PRDX1 and RPN1 were significantly downregulated (P < 0.05). To validate these findings, the expression of these DRGs was further examined in clinical samples. The expression patterns of FLNB, GYS1, FLNA, and PRDX1 corresponded to those observed in the dataset (Fig. 1C), among which FLNB and GYS1 showed statistically significant differences, supporting the robustness and reliability of the dataset. Correlation analysis (Figs. 2A–B) demonstrated that the edge weights for FLNB and GYS1 were 3 and 2, respectively, and that these two genes were positively correlated. Additionally, FLNB was negatively correlated with RPN1, PRDX1, and LRPPRC, while GYS1 showed a significantly negative correlation with PRDX1, suggesting a close functional relationship among these genes.

Fig. 1.

Differential analysis and real-world validation of genes related to disulfidptosis. (A) In the boxplot, the x-axis indicates disulfidptosis-related genes, and the y-axis represents the expression of these genes; (B) For the heatmap, the x-axis indicates samples, and the y-axis means genes. Red denotes genes with high expressions, and blue denotes genes with low expressions; (C) Single-item comparison of DRGs expression in clinical samples. *P < 0.05, **P < 0.01, ***P < 0.001.

Fig. 2.

Correlation analysis of DRG. (A) is a correlation matrix plot where horizontal and vertical lines represent DRGs. Red represents the positive correlation, and blue represents the negative correlation. (B) is a PPI network. *P < 0.05, **P < 0.01, ***P < 0.001.

Although the absolute log2FC values of the identified DRGs were modest (ranging from 0.02 to 0.057), their consistent identification under a stringent FDR threshold (< 0.05) indicated statistically robust differential expression. To further assess their biological relevance, we evaluated the effect sizes. The standardized mean difference (SMD, Cohen’s d) for the most significant genes, FLNB and GYS1, were calculated as 0.42 (95%CI: 0.15–0.69) and 0.38 (95%CI: 0.11–0.65), respectively, which fall within the “small to medium” effect size range. This suggests that while the magnitude of expression change for individual genes is not large, the differences are consistent and detectable across cohorts. The biological significance of these coordinated subtle changes is explored through subsequent functional enrichment and predictive modeling.

Enrichment analysis of differentially expressed DRGs

GO and KEGG pathway analyses have been carried out to determine the biological processes and associated signaling axis of differentially expressed DRGs between healthy controls and DR patients. GO enrichment analysis identified that differentially expressed DRGs are involved in several key biological processes, including platelet aggregation, platelet activation, and circulatory system homeostasis (Fig. 3). Additionally, KEGG pathway analysis identified significant enrichment in the Rap1 signaling pathway and the MAPK signaling pathway (Fig. 4). Platelet aggregation is crucial for DR by promoting microvascular damage, ischemia, and neovascularization, while the Rap1 signaling pathway may contribute to vascular abnormalities in DR by participating in platelet activation and endothelial cell dysfunction³³.

Fig. 3.

GO enrichment analysis.

Fig. 4.

KEGG enrichment analysis (The pathway map of Amyotrophic lateral sclerosis (map05014)、N-glycan biosynthesis (map00510)、MAPK signaling pathway (map04010) 、Shigellosis (map05131) 、Thermogenesis pathway (map04714)、Rap1 signaling pathway (map04015)、Diabetic cardiomyopathy (map05415) 、 Salmonella infection (map05132)、Platelet activation༈map0461)et al. were adapted from the KEGG PATHWAY database [https://www.kegg.jp/kegg/pathway.html]

Specifically, the Rap1 signaling pathway regulate PI3K phosphorylation and activate Akt to inhibit the downstream substrate GSK-3β to protect the interendothelial adhesion function and vascular permeability by preventing the degradation of β-catenin and simultaneously raising the VE-cadherin expression³⁴. Besides, Akt phosphorylation could activate the activity of eNOS and promote the generation of nitric oxide (NO) and vasodilation³⁵. Furthermore, PI3K-Akt pathway can also exert anti-inflammatory and anti-apoptotic effects to ensure the survival and functional homeostasis of endothelial cells³⁶.

Immune infiltration analysis

To explore the systemic immune signatures in DR patients, the CIBERSORT algorithm was deployed to assess the immune-infiltrated condition in comparison to healthy controls. As plotted in Fig. 5A, the proportion of monocytes and macrophage M0 was higher in the DR samples, while the concentration of activated CD4 memory T cells was lower. In addition, we evaluated the correlations of key gene expression with immune infiltration by Spearman’s rank correlation analysis. Notably, several signature genes performed moderate correlations (|ρ| ≈ 0.38–0.45, p < 0.01) with specific immune populations. FLNA expression had a positive correlation with regulatory T cells (ρ = 0.45) and monocytes (ρ = 0.40), suggesting its potential role in immune suppression and inflammatory regulation. Similarly, FLNB showed a significant association with activated dendritic cells (ρ = 0.41), while GYS1 correlated with macrophages M0 (ρ = 0.39), highlighting a possible link between metabolic reprogramming and macrophage infiltration. These findings indicated that these identified DRGs were not only biomarkers but might also participate in shaping the systemic immune signatures. The graphical functions of the ggplot2 package were used to visualize these correlations, providing an intuitive view of the gene–immune cell interactions (Supplementary Table 1).

Fig. 5.

Comparison of the ratio of 22 immune cell subtypes between DR patients and normal individuals, and correlation analysis between DRG and immune cells (A) Comparison of the ratios of 22 immune cell subtypes between DR patients and normal controls, where the x-axis means the names of the immune cell subtypes and the y-axis means the relative proportions of the immune cell subtypes; (B) Correlation analysis of DRG and immune cells.

TLN1 is highly expressed in acute myeloid leukemia (AML) and is significantly positively associated with immune cells, concerning macrophages, effector memory T cells, as well as immature dendritic cells. Mechanistically, TLN1 physically interacts with myosin heavy chain 9 and FLNA, forming a complex that promotes Caspase3 expression and mediates the proliferation, differentiation, and apoptosis of AML cells³⁷. FLNA can promote the infiltrated of various immune cells (myeloid dendritic cells, CD8⁺/CD4⁺ T cells, neutrophils) through the FAK/AKT pathway. Mechanistically, FLNA promotes RhoA activation and stress fiber formation, supporting immune cell migration³⁸. FLNB can form a heterodimer with FLNA to antagonize RhoA activation; gene knockout of FLNB will impair cell migration, potentially affecting immune cell movement³⁹. There are currently no relevant reports on GYS1. The proportion of activated CD4 memory T cells was lower in the DR group, which may also be related to the decrease of these DRGs. The decrease of immune tolerance may finally promote disease progression (Fig. 5B).

Validation of the diagnostic model

We assessed the diagnostic performance of the nomogram model in predicting DR by ROC curve analysis. The calibration curve assessed the degree of consistency degree between the model’s predicted values and the actual values of model, while the decision curve analysis (DCA) further validated the model’s value in clinical decision-making. Specifically, the “Rms” package was carried out to construct the nomogram model for diagnosing DR based on eight differentially expressed DRGs (FLNA, TLN1, PRDX1, FLNB, ACTB, RPN1, NDUFS1, and GYS1). By locating the corresponding values on each variable axis and reading their scores, all variable scores were summed to yield the “Total Points,” which can then be converted into the probability of an individual developing DR based on the scale axis below (Fig. 6A). ROC curve analysis was deployed to identify the diagnostic performance of these differentially expressed DRGs for DR. The AUC values for FLNA, TLN1, PRDX1, FLNB, ACTB, RPN1, NDUFS1, and GYS1 were 0.75, 0.79, 0.64, 0.70, 0.68, 0.74, 0.63, and 0.74, respectively (Fig. 6B). However, since none of the genes were experimentally shown to induce disulfidptosis in DR models by direct disulfidptosis assays, thus this DRGs-related model was only hyothetical.

Fig. 6.

The receiver operating characteristic curve (A), Creating a nomogram to assess the risk of DR using differentially expressed DRGs (B-C), calibration curve (D), and decision curve analysis (E), and Clinical impact curve (F) of the screening nomogram in the training set.

We performed z-score normalization and constructed a joint prediction model using L2-regularized logistic regression, yielding Score = β0 + Σ (βi * zi), with individual probabilities calculated as 1/(1 + exp(-Score)). Compared to single-gene models, this model significantly improved overall discriminative ability (AUC = 0.82, Fig. 6C). In the latest study, a machine learning-based predictive model was set for proliferative DR via single-cell transcriptomics⁴⁰. This predictive model achieved an AUC of 0.81 in the test cohort and 0.83–0.96 in the validation cohorts, similar to our study. Our model demonstrated higher AUC and accuracy in both internal and external validations, and revealed the molecular basis of DR, offering new perspectives for early risk assessment and personalized intervention in DR. The calibration curve demonstrated that the solid and dashed lines were closely aligned, indicating a small error between actual and predicted risk, which suggested that the nomogram possessed good predictive accuracy (Fig. 6D). The DCA showed that the red line representing the nomogram was notably distant from the “all” curve, further supporting the model’s effectiveness (Fig. 6E). To more intuitively assess the clinical relevance of the nomogram, the clinical impact curve was generated by DCA. When the high-risk threshold was between 0.3 and 1, the “Number high risk” curve closely followed the “Number high risk with event” curve, suggesting strong predictive power of the model (Fig. 6F). Overall, the nomogram model demonstrated robust diagnostic value to distinguish DR from the normal controls.

Discussion

Epidemiological studies indicate that DR ranks as the most severe microvascular disorders of DM, featured with pathological features including neovascularization, pericyte loss, and endothelial dysfunction in diabetic patients⁴¹. Disulfidptosis is fundamentally triggered by gathering disulfide bonds, which promotes cytoskeletal collapse as well as subsequent PCD related to pathological processes¹¹. However, the precise molecular mechanisms of disulfidptosis across diverse diseases remain poorly characterized. Our investigation aims to clarify the involvement of DRGs in DR, develop an association of disulfidptosis with pathological events in DR, and identify candidate key genes through bioinformatics analysis to dissect the speculative mechanisms (Fig. 7). In pathogenic environments such as high blood sugar and oxidative stress, the abnormal expression of multiple genes related to dual cell death (such as FLNB, GYS1, etc.) promotes DR progression through different pathways. This process is accompanied by dysregulation of systemic immune signatures, including macrophage infiltration and CD4⁺ T cell activation, as well as platelet activation and activation of the Rap1 signaling pathway, leading to platelet aggregation and vascular dysfunction, ultimately causing retinal microvascular damage. The significant changes in the expression of FLNB and GYS1 genes in peripheral blood are highlighted in the figure, suggesting that they may serve as novel biomarkers and potential therapeutic targets for DR.

Fig. 7.

The speculative mechanism involving disulfidptosis-related genes in the occurrence and development of diabetes retinopathy.

We identified the expression of disulfidptosis-related regulatory factors across the peripheral blood of DR and healthy individuals by using the GEO database. Eight differentially expressed DRGs were identified: FLNA, TLN1, PRDX1, FLNB, RPN1, GYS1, ACTB, and NDUFS1. These differentially expressed DRGs (FLNB, GYS1, FLNA, and PRDX1) were consistent with the trends observed in the large dataset; notably, FLNB and GYS1 showed statistically significant differences, indicating high reliability of the dataset. PRDXs are a newly identified family of non-seleno glutathione peroxidases and have been reported to exist in diverse major tissues, such as the lens and retina. Strong evidence suggested that they acted as protective proteins by maintaining the redox balance through scavenging reactive oxidative species⁴². NDUFS1, a member of the 75 kDa subunit family of mitochondrial inner membrane complex I, played a significant role in glucose, energy, and amino acid metabolic pathways and was observed in the pathogenesis of clear cell renal cell carcinoma⁴³, lung cancer⁴⁴, and diabetes⁴⁵. However, its molecular mechanism in diabetes remained unclear, and rare study reported its role in DR. Additionally, NDUFS1 is an oxidative phosphorylation-related gene that encodes NADH dehydrogenase and contributes to the coordinated upregulation of mitochondrial respiratory chain expression⁴⁶. Under high-glucose conditions, intracellular ROS levels increased significantly, subsequently enhancing the activity of NADH oxidase⁴⁷. When NDUFS1 was downregulated, it led to the reduced mitochondrial respiration in the mitochondrial apoptosis pathway⁴⁸.

Correlation analysis suggested that there were significant interactions among these hub genes. In addition, we investigated the GO terms and KEGG pathways enriched by these differentially expressed DRGs. These differentially expressed DRGs were associated with platelet aggregation, activation, and retinal homeostasis, which potentially contributed to capillary occlusion and further retinal ischemia in DR pathogenesis⁴⁹. KEGG pathway^50,51 enrichment confirmed that they were mainly enriched in Rap1 and MAPK signaling axis⁵². Large amounts of cytokines and inflammatory mediators could activate the intracellular MAPK pathway, thereby triggering inflammation and apoptosis^53,54.

This study further investigated the immune infiltration levels in DR patients. We demonstrated that the proportions of monocytes and M0 macrophages were greater in DR samples than those in healthy controls, whereas the proportion of activated CD4 memory T cells was lower in DR samples. Meanwhile, we also analyzed the association of key gene expressions with systemic immune signatures. The activation of macrophages and monocytes was closely associated with the enhancement of local inflammatory responses⁵⁵. In a study on monocytes and DR⁵⁶, the monocyte count in DR patients was significantly increased compared to healthy controls. This suggests that an increase in monocytes may contribute to the advancement of DR. Meanwhile, evidence has shown that disrupting the interaction of IL-17 A with endoplasmic reticulum stress in M0 macrophages can inhibit neovascularization in DR⁵⁷.

As the key driver of disulfidptosis, disulfide stress may interact with the established oxidative stress pathways in DR, exacerbating retinal cell damage⁵⁸. In individuals with diabetic nephropathy (DN), four core genes related to RGCs were also identified: CXCL6, CD48, C1QB, and COL6A3⁵⁹. They were significantly related to immune cell infiltration, extracellular matrix (ECM) remodeling, and renal function decline⁶⁰. The chemokine CXCL6 can promote neutrophil recruitment, participating in leukostasis and vascular leakage in DR; immune regulatory molecules CD48 and C1QB are associated with complement system activation, potentially exacerbating the chronic low-grade inflammation of DR; and the collagen COL6A3 may be involved in ECM deposition, correlating with basement membrane thickening and vascular fibrosis in DR⁶¹. It suggests that these genes may engage in the advancement of DR by mediating systemic immune signatures.

One limitation of our study was the extrapolation of immune profiles only in peripheral blood but not retinal tissue. Peripheral blood gene expression could be influenced by diabetes duration, medications, inflammation, comorbidities and so on, all of which were not adjusted in the model. CIBERSORT provided valuable insights into the systemic immune signatures, while the immune microenvironment in retinal tissue might significantly differ from that of peripheral blood in terms of immune cell composition, function, and activation status. Therefore, immune profiles from peripheral blood may not fully reflect the retinal immune status. Future studies should consider using retinal tissue or vitreous fluid samples to directly assess the retinal immune microenvironment.

In the transcriptomic study based on an independent cohort of peripheral blood, the “signal pattern” composed of coordinated and subtle expression shifts of multiple genes is far more meaningful than the significant changes of a single gene. We note that the observed effect sizes (e.g., SMD ≈ 0.4 for FLNB and GYS1) for individual DRGs in peripheral blood are modest. This is consistent with the nature of blood transcriptomics in complex chronic diseases, where systemic signals are often diluted within a heterogeneous cell population. However, the critical insight from our study is that multiple genes exhibiting coordinated, albeit small, expression changes can collectively exert a significant biological impact. This is evidenced by: (1) the strong functional enrichment of these DRGs in DR-relevant pathways (e.g., platelet activation); (2) their significant correlations with disease-specific immune cell profiles; and (3) most importantly, the high diagnostic performance (AUC = 0.818) of the multi-gene model built from them.

This research tried to link disulfidptosis to DR for the first time, shedding light on its potential role in DR initiation and progression through its effects in immune regulation and vascular dysfunction. DRGs acted as novel biomarkers for early diagnosis and prognosis, enabling better risk stratification and earlier intervention. However, since none of the genes were experimentally shown to induce disulfidptosis in DR models by direct disulfidptosis assays, thus this DRGs-related model was only hyothetical. The model was very limited for its application only in one GEO dataset plus a small validation cohort. Besides, cross validation was only for internal use, as this model was exploratory or proof of concept.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Hao Liu: Conceptualization, Formal analysis, Writing - original draft, Funding acquisition. Yue Hao: Conceptualization, Formal analysis, Methodology, Resources, Writing - review & editing. Xi-Xi Zhang: Conceptualization, Formal analysis, Methodology, Resources, Writing - review & editing. Xin-Yi Wang: Methodology, Resources. Jun-Tao Zhang: Methodology, Resources. Li-Fen Guo: Methodology, Resources. Heng-Qian He: Methodology, Resources. Li-Qin Ying: Methodology, Resources. Si-Yu Xian: Methodology, Resources. Qin-Kang Lu: Conceptualization, Resources, Funding acquisition, Supervision, Writing - review & editing.

Funding

Thanks to Laboratory Animal Center of Ningbo University for the technical support. This study was supported in part by grants from Ningbo Natural Science Foundation Project (2024J276), Ningbo Clinical Research Center for Ophthalmology(2022L003), Ningbo Key Laboratory for neuroretinopathy medical research, the Project of Ningbo Leading Medical &Health Disipline (2016-S05), Technology Innovation 2025 Major Project of Ningbo (2021Z054) and Ningbo “Innovation Yongjiang 2035” Key Technology Breakthrough program (2024Z233).

Data availability

The original contributions presented in the study are included in the article.All sequencing data analyzed in the article were sourced from public databases (GSE221521, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE221521. And all code used for analysis is available upon request to the corresponding author. The validation data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics

This study was approved (or granted exemption) by the ethics committee of The Affiliated People’s Hospital of Ningbo University (approval no.MR-33-21-016772). We certify that the study was performed in accordance with the 1964 declaration of HELSNKI and later amendments.

Consent to participate

Written informed consent was obtained from all the participants prior to the enrollment of this study.