Multi-omics and experimental validation studies reveal key biomarkers of cellular senescence in diabetic retinopathy

Abstract

Objective

Diabetic retinopathy (DR) is a prevalent microvascular complication of diabetes, contributing to vision impairment and related retinal diseases. Growing evidence indicates that cellular senescence (CS) under high-glucose conditions plays a role in the pathogenesis of DR. This study aims to identify key biomarkers of CS in DR by integrating transcriptomics, single-cell sequencing data, and experimental validation, thereby offering insights for understanding the disease mechanism and developing novel therapeutic strategies.

Methods

DR-related datasets and CS-related genes (CSRGs) were retrieved from the Gene Expression Omnibus (GEO) and CellAge databases. The characteristic gene set for DR-CS was obtained by intersecting differentially expressed genes (DEGs), Weighted Gene Co-expression Network Analysis (WGCNA) results, and CSRGs. Subsequent analyses involved constructing protein-protein interaction (PPI) network, cytoHubba screening, enrichment analysis, and immune infiltration analysis. Machine learning methods were used to identify key biomarkers from the DR-CS characteristic gene set, which were then validated using external datasets. Single-cell sequencing and gene set enrichment analysis (GSEA) were employed to determine the cellular location and biological functions of DR-CS key biomarkers, and animal experiments further validated these biomarkers.

Results

A total of 67 DR-CS-related characteristic genes were identified. Enrichment analysis highlighted pathways like cellular senescence and the Advanced Glycation Endproducts-Receptor for Advanced Glycation Endproducts (AGE-RAGE) signaling pathway in diabetic complications as being closely related to DR development. A set of 13 characteristic genes was selected through a combination of PPI network and six cytoHubba algorithms. Further analysis using machine learning, expression analysis, and Receiver Operating Characteristic (ROC) analysis revealed MYC and LOX as key biomarkers of DR-CS. The expression characteristics of MYC and LOX in various cells were examined using single-cell RNA sequencing. Animal experiments demonstrated that the expression levels of MYC and LOX in the retina were significantly higher in DR group than in the control group (P < 0.05).

Conclusion

MYC and LOX were identified as key biomarkers of DR-CS. Thus, investigating these genes may provide new therapeutic targets for DR treatment by targeting cellular senescence.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13098-025-01899-y.

Introduction

Diabetic retinopathy (DR), the most common diabetic eye complication and retinal vascular disease, affected approximately 103.12 million adults globally in 2020, with projections indicating a 55.6% increase to 160.50 million by 2045 [1]. In the early stage, DR manifests as vascular endothelial cell damage within a high-glucose environment, characterized by leakage, microaneurysm formation, and punctate intraretinal hemorrhages. As the disease progresses, chronic inflammation and neurodegenerative changes drives the fibrous tissue proliferation. In a high-glucose environment, Müller cells release numerous pro-angiogenic, pro-inflammatory, and profibrotic factors, which contribute to angiogenesis and neurodegeneration. Moreover, the abnormal proliferation of Müller cells and the release of profibrotic mediators play a key role in the formation of fibrous tissue. Potentially, these processes can lead to blindness in severe cases [2, 3]. Current DR treatments mainly focus on anti-angiogenic therapies targeting vascular endothelial growth factor (VEGF) and its signaling pathways. However, due to the complex pathogenesis of DR, these therapies remain limited in efficacy [4]. Although genetic and environmental factors play important roles in DR development, several aspects of its general pathophysiology remain elusive. Thus, actively exploring potential DR pathomechanisms and biomarkers is crucial for its prevention and treatment strategies.

Recent studies have indicated that DR development involves various cell changes that regulate biological functions, including apoptosis, pyroptosis, and senescence [5, 6]. Particularly, cellular senescence (CS) in DR has gained significant research attention, although its underlying mechanisms in DR are not yet fully understood [7]. CS is a cell state triggered by endogenous or exogenous stimuli. It is characterized by a stable cell-cycle arrest and a complex senescence-associated secretory phenotype (SASP), which comprises inflammatory cytokines, chemokines, and growth factors [8, 9]. Studies have demonstrated that CS plays a role in the pathogenesis of early-stage DR and is positively correlated with disease progression. Research indicates that in high-glucose conditions, retinal endothelial cells, retinal pigment epithelial cells, and retinal ganglion cells undergo CS to varying degrees. This is evidenced by increased β-galactosidase activity, elevated SASP secretion, and upregulated expression of senescence-related proteins, which promote retinal pathological neovascularization and neurodegeneration [10–12]. In addition to hyperglycemia, the production of polyols and advanced glycation end products (AGEs), as well as the induction of oxidative stress and chronic persistent inflammation, collectively contribute to the development of a pro-senescence environment [13]. Furthermore, CS is associated with immune dysregulation. The interaction between the two factors creates a chronically exacerbated inflammatory environment in the retina, thereby accelerating retinal damage [14].

Based on this evidence, we hypothesized that specific CS-associated genes (CSRGs) are key drivers of DR progression and are linked to distinct patterns of immune cell infiltration in the retina. To test this hypothesis, we integrated DR-specific gene expression profiles with CS-related databases and employed bioinformatics analysis and animal experiments to identify and validate DR-CS genes as potential biomarkers for DR progression. Furthermore, immune infiltration analysis was conducted to elucidate the immune mechanisms of DR and the correlation between biomarkers and immune cells. This study aims to provide new insights to the complex interplay between CS, immune dysregulation, and DR pathogenesis.

Materials and methods

Data acquisition and processing

The Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) offers genetic, expression, and phenotypic data for the GSE102485 dataset (with retinal tissues from 3 control patients and 22 diabetic patients) and the GSE60436 dataset (with retinal tissues from 3 control patients and 6 diabetic patients). These two datasets were collectively employed as training set. The GSE94019 dataset (with retinal samples from 4 control patients and 9 diabetic patients) was designated as the validation set. The GSE165784 dataset, which includes fibrous membrane samples from 1 patient with proliferative vitreoretinopathy and 5 patients with DR, is utilized as the single-cell sequencing analysis. Additionally, CSRGs are collected from the CellAge database (https://genomics.senescence.info/cells/).

Differential expression analysis

R (V4.4.3) software was used for data analysis. The datasets GSE102485 and GSE60436 were normalized using the "normalizeBetweenArrays" function from the “limma” package. Subsequently, the "Combat" function from the "SVA " package was utilized to integrate the data and removed batch effects, with boxplots generated to illustrate the differences before and after this process. For the identification of differentially expressed genes (DEGs) with DR, the significance threshold was set at P value < 0.05 and |logFC| ≥ 1. Volcano plot and heatmap of the top 20 genes were constructed using the "ggplot2" and "pheatmap" packages. Finally, the "ClusterProfiler" package was employed to carry out Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses on the DEGs.

Weighted gene co-expression network analysis (WGCNA)

WGCNA is a systems biology method that clusters genes with similar expression patterns into modules and identifies key genes by analyzing the association between modules and specific traits [15]. After filtering out unsuitable genes and samples, the "WGCNA" package was used to construct a co-expression network. The β parameter, which emphasizes strong gene correlations and suppresses weak ones, was optimized using the pickSoftThreshold function. A topology overlap matrix and gene clustering tree were then generated to establish a weighted co-expression network model. Genes were divided into several related modules, and the module most associated with DR was selected for further analysis. Key WGCNA parameters were set as follows: minModuleSize = 30, mergeCutHeight = 0.25, and deepSplit = 3. Genes with Module Membership > 0.6 and overlapping with DEGs were designated as DR-related genes. Similarly, GO and KEGG analyses were performed on the genes of related modules.

Functional enrichment analysis of DR-CSRGs

The "ggvenn" package in R was used to intersect DEGs of DR with genes from the WGCNA-selected module and CSRGs, thereby identifying DR-CSRGs. Subsequently, GO and KEGG enrichment analyses were performed to explore the common functions and enrichment pathways of DR-CSRGs. Finally, the "ggplot2" package was utilized for visualization.

Protein-protein interaction (PPI) network and machine learning

The interaction relationships of DR-CSRGs were explored using the STRING database (https://cn.string-db.org/). Gene interactions with a confidence score less than 0.4 were excluded. Then, the PPI network was constructed using Cytoscape. Using the cytoHubba plugin’s six algorithms: Maximal Clique Centrality (MCC), Maximal Neighborhood Component (MNC), Degree, Edge Percolated Component (EPC), Stress, and Closeness, the top 20 genes related to DR-CSRGs were identified for each algorithm in the PPI network. The overlapping genes among these top 20 lists were selected as candidate genes.

Subsequently, four machine learning methods were applied to screen for key genes: Least Absolute Shrinkage and Selection Operator (LASSO), Support Vector Machine -Recursive Feature Elimination (SVM-RFE), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF). Specifically, the LASSO approach used the "glmnet" package to filter candidate genes, identifying those with the least error through cross-validation as disease-characteristic genes. SVM-RFE recursively removed unimportant features to enhance model performance, with genes showing the least error in the SVM-RFE model selected as characteristic genes. XGBoost captured complex data interactions and nonlinear relationships to identify feature genes potentially missed by LASSO. RF, an ensemble learning method, built multiple decision trees to boost model performance and stability, offering reliable estimates of feature importance. Genes common to all four machine learning models were regarded as candidate biomarkers.

Construction of nomogram and receiver operating characteristic (ROC) analysis

The "rms" package was utilized to construct a nomogram based on candidate biomarkers, facilitating DR risk prediction. The expressions of these biomarkers in the training set and validation set (GSE94019) were analyzed, and ROC curves were generated using the "pROC" package to assess their diagnostic efficacy for DR. Candidate genes exhibiting consistent expression trends in both sets, with significant differences between groups (P < 0.05), and an AUC > 0.7 were considered valid DR biomarkers.

Construction of gene-gene interaction (GGI) networks and gene set enrichment analysis (GSEA)

To explore interactions between biomarkers and their biological functions, the GeneMANIA database (http://www.genemania.org/) was employed to predict and construct GGI networks. For a comprehensive understanding of the functions and pathways of biomarkers, associations between biomarkers and other genes within the training set were calculated and ranked using the Spearman statistical method. Pathway gene sets (c2.cp.v2024.1.Hs.symbols.gmt) were downloaded from the Molecular Signatures Database (MSigDB) and utilized as a background set. Subsequently using this background set, GSEA was performed to search for the potential functions of the biomarkers using the "clusterProfiler" package. An adj. P < 0.05 was considered significant in the enrichment results.

Chromosome localization, subcellular localization, and single-cell sequencing analysis of key biomarkers

The "biomaRt" package was employed to retrieve gene positions for determining the chromosomal positions of the biomarkers, and a chromosome position map was generated using "RCircos" to illustrate specific gene locations. For functional and mechanistic insights, FASTA sequences of the biomarkers were obtained from the NCBI website (https://www.ncbi.nlm.nih.gov/). These sequences were analyzed using the mRNALocator database to predict subcellular localization.

The GSE165784 dataset was analyzed using the "Seurat" package. Cells expressing mitochondrial genes in over 10% of the cell population were excluded. Data normalization was performed using the "NormalizeData" function in "Seurat", followed by batch correction and dimensionality reduction with the "Harmony" package. Cells were identified using 30 principal components at a resolution of 0.6. Each cell cluster was annotated using cell type-specific marker genes, focusing on the distribution and expression patterns of key biomarker genes across different cell types. Additionally, intercellular communication was analyzed using the "CellChat" package.

Immune infiltration analysis

Given the strong association between DR’s pathophysiology and immune dysregulation, which can perturb the body’s normal homeostasis, we used "CIBERSORT" to assess the proportions of 22 types of immune cells. Immune cells with a relative proportion of 0 or those constituting less than 50% of the cells in the sample were excluded, and differences in immune cell infiltration between DR and control groups were compared (P < 0.05). The "ggplot2" and "ggpubr" packages in R were used to create stacked bar and box plots to visualize these differences. Additionally, the "linkET" package was employed to analyze the correlation between key biomarkers and related immune cells, the Mantel test was applied to calculate the correlation coefficient and P-value, with statistical significance set at P < 0.05.

Animal model establishment

The animal model was established using male C57BL/6J mice (SPF-grade), aged 5–6 weeks, purchased from Hangzhou Zi yuan Laboratory Animal Technology Company (Zhejiang, China, production license SCXK (Zhe) 2019-0004). After adaptive feeding for 1 week in ventilated cages under a 12-hlight/dark cycle, 50 ± 20% humidity, and 22 ± 2 °C temperature, they were randomly divided into normal control and diabetic groups. The diabetic group received intraperitoneal Streptozotocin (50 mg/kg, Macklin, Shanghai, S8817944) in sodium citrate buffer (pH = 4.3) for 5 days; the normal control group received the same volume of buffer. Mice with a tail-vein blood glucose level > 16.7 mmol/L after 1 week were considered diabetic models [16]. After 12 weeks, mice were anesthetized with sodium pentobarbital (50 mg/kg) and euthanized, and their eyeballs were enucleated. The study was approved by the Animal Ethics Committee of Anhui University of Chinese Medicine (Approval No. AHUCM-mouse-2024157).

Histopathological observation

Retinal tissues of mice were fixed in 4% formaldehyde, rinsed, dehydrated, embedded in paraffin. Sections of 5 μm thickness were cut. After hematoxylin staining, the tissues were dehydrated using ethanol gradients, stained with eosin, dehydrated with ethanol, and cleared with xylene. Finally, the sections were mounted with neutral gum. Morphological changes were observed under an optical microscope.

Detection of key biomarkers at the mRNA level

After constructing a PPI network, applying cytoHubba algorithms, using machine learning, and validating with external datasets to determine key biomarkers of DR-CS, these biomarkers were selected for RT-qPCR validation in animal models to assess their expression levels. Total RNA was extracted from mouse retinal tissue using TRIzol (G3013, Servicebio, China). The RevertAid First Strand cDNA Synthesis Kit (#K1622,Thermo, China) was used to reverse transcribe the RNA into cDNA. RT-qPCR was employed using the Hieff qPCR SYBR^® Green Master Mix (11201ES03, Yeasen, Shanghai) and Quantagene q225 system (KUBO, China). Relative mRNA expression levels were calculated using the 2^−ΔΔCt method. The primer information is listed in Table 1.

Table 1.

Sequences of primers used in RT-qPCR

Primers	Sequences (5’-3’)
M-GAPDH-F	CCTCGTCCCGTAGACAAAATG
M-GAPDH-R	TGAGGTCAATGAAGGGGTCGT
M-P53-F	GTCACAGCACATGACGGAGG
M-P53-R	TCTTCCAGATGCTCGGGATAC
M-LOX-F	ACTTCCAGTACGGTCTCCCG
M-LOX-R	GCAGCGCATCTCAGGTTGT
M-P21-F	CCTGGTGATGTCCGACCTG
M-P21-R	CCATGAGCGCATCGCAATC
M-MYC-F	CAACGTCTTGGAACGTCAGA
M-MYC-R	CTCGTCTGCTTGAATGGACA
M-VEGF-F	AGCACAGCAGATGTGAATGC
M-VEGF-R	AATGCTTTCTCCGCTCTGAA
M-TNFα-F	AGTCCGGGCAGGTCTACTTT
M-TNFα-R	GAGTTGGACCCTGAGCCATA
M-IL-6-F	AGTTGCCTTCTTGGGACTGA
M-IL-6-R	TCCACGATTTCCCAGAGAAC
M-MCP-1-F	CACTCACCTGCTGCTACTCA
M-MCP-1-R	GCTTGGTGACAAAAACTACAGC

Protein level detection of key biomarkers

To detect key biomarkers at the protein level, Western blotting (WB) was conducted using animal samples. Retinal tissues were lysed in RIPA buffer (Biosharp, China), then centrifuged at 13,000 g for 10 min at 4 ℃. The total protein concentration was determined using a BCA kit (Servicebio, China), and proteins were separated via 10% SDS-PAGE. The primary antibodies applied were: mouse anti-GAPDH (1:10000, Proteintech, 60004-1-Ig), rabbit anti-LOX (1:5000, ZEN-BIOSCIENCE, R381849), rabbit anti-P53 (1:2000, ABclonal, A0263), rabbit anti-P21 (1:2000, SAB, 41297) and mouse anti-MYC (1:2000, Proteintech, 60003-2-Ig). After overnight incubation at 4 ℃, membranes were incubated with secondary antibody (1:50000, Seracare, USA) for 1 h at room temperature. Bands were quantified using ImageJ.

Statistical analysis

Statistical analyses were performed using R (version 4.4.3) and GraphPad Prism (version 10.1.2). For comparisons between two groups, a t-test or Wilcoxon rank-sum test was used. A significance level of P < 0.05 was considered statistically significant.

Result

Differential gene expression analysis

Batch effects between the two datasets were removed using internal correction and standardization procedures, and boxplots displayed the results before and after batch removal (Fig. 1A-B). A volcano plot identified 1,401 DEGs in these datasets, with 828 upregulated and 573 downregulated (Fig. 1C). A heatmap showed the expression patterns of the top 20 genes ranked by|logFC| in the DR and control groups (Fig. 1D). Through enrichment analysis of DEGs, it was found that these DEGs are mainly enriched in biological processes such as collagen-containing extracellular matrix, extracellular matrix structural constituent, and cell-substrate junction, which are all hallmarks of CS (Fig. 1E). KEGG analysis also revealed that DEGs are associated with pathways such as CS, p53 signaling pathway, and necroptosis, indicating that these differentially expressed genes may play a crucial role in regulating the cell senescence process, thereby affecting the development of DR (Fig. 1F).

Fig. 1.

Data normalization and identification of DEGs. (A) Boxplots showing the expression profiles before batch removal. (B) Boxplots showing the expression profiles after batch removal. (C) Volcano plot of DEGs. (D) Heatmap of top 20 genes in DEGs. (E) Bar chart of enriched GO terms of DEGs. (D) Bar chart of enriched KEGG pathways of DEGs

WGCNA analysis

WGCNA constructs gene co-expression networks to identify modules strongly associated with specific phenotypes and selects key genes of potential biological importance. The sample dendrogram and trait heatmap showed no significant sample abnormalities (Fig. 2A), so no samples were excluded. The optimal soft threshold, determined by selecting an R² > 0.9 and minimal connectivity, was set at 12 (Fig. 2B). This identified eight co-expression modules (Fig. 2C). A heatmap displayed the correlation between these eight gene modules and DR/control groups (Fig. 2D). The brown module (R = 0.7, P < 0.05) and turquoise module (R = -0.66, P < 0.05) were most strongly correlated with DR. A total of 972 module genes were identified, highlighting their potential as targets for exploring DR’s molecular mechanisms. The scatter plot of gene-module membership in the brown module and turquoise module demonstrated the strength of co-expression relationships among intramodular genes (Fig. 2E-G). Enrichment analysis of the genes in this module also revealed associations with pathways such as CS, p53 signaling pathway, and AGE-RAGE signaling pathway in diabetic complications. This further confirms that CS plays an important role in the occurrence and development of DR.

Fig. 2.

WGCNA analyses of datasets. (A) Sample dendrogram and trait heatmap. (B) Selection of the optimal soft threshold power (β). (C) Cluster dendrogram showing the identified 8 modules. (D) Module − trait relationships between Normal and DR group. (E) Scatter plot of gene-module membership in the brown module. (F) Scatter plot of gene-module membership in the turquosise module. (G) Eigengene adjacency heatmap representing the pairwise correlations between different modules, highlighting co-expression relationships

Functional enrichment analysis and PPI network construction

The overlap of DEGs, WGCNA module genes, and CSRGs identified 67 candidate DR-CS-related signature genes (Fig. 3A). GO enrichment analysis revealed that these genes were enriched in 937 GO terms, including 860 BPs (e.g., myeloid cell differentiation, gliogenesis, and positive regulation of interleukin-8 production), 31 CCs (e.g., cell-substrate junction, collagen-containing extracellular matrix, and focal adhesion), and 46 MFs (e.g., cytokine receptor binding, type II transforming growth factor beta receptor binding, and Toll-like receptor binding) (Fig. 3B, Supplementary Table 1). KEGG pathway analysis showed significant enrichment of these genes in 85 functional pathways, including cellular senescence, AGE-RAGE signaling pathway in diabetic complications, MAPK signaling pathway, and p53 signaling pathway (Fig. 3C, Supplementary Table 2). Using Cytoscape, a PPI network was constructed, comprising 67 nodes and 348 edges (Fig. 3D), highlighting strong interactions of genes like MYC, MMP9, JUN, and CCL2. Six cytoHubba screening methods identified 13 candidate genes: CCND1, PTGS2, MYC, SERPINE1, LOX, HRAS, IRF1, CDKN1A, MMP9, JUN, CCL2, MYD88, and TLR4 (Fig. 3E). These genes play key roles in DR-CS and show potential as biomarkers or therapeutic targets.

Fig. 3.

Screening and functional analysis of candidate genes. (A) Venn diagram displaying the overlap of genes identified through intersection analysis. (B) PPI of candidate genes. (C) Bar chart of enriched GO terms of candidate genes. (D) Bar chart of enriched KEGG pathways of candidate genes. (E) UpSet plot displaying the results of six cytoHubba screening methods identified 13 candidate genes

Multi-algorithm feature selection for identifying key biomarkers of DR-related cellular senescence

The 13 candidate genes identified by cytoHubba were inputted into LASSO, SVM-RFE, RF, and XGBoost algorithms for feature selection. In the LASSO model, six genes (MYC, LOX, HRAS, IRF1, JUN, and CCL2) were identified as key features (Fig. 4A-B). Using the SVM-RFE algorithm, the top seven genes (MYC, LOX, SERPINE1, CCL2, MMP9, HRAS, and JUN) were selected based on maximum accuracy (Fig. 4C-D). The RF algorithm ranked genes by importance (> 0.5), yielding seven core genes: CCL2, SERPINE1, LOX, MMP9, JUN, MYD88, and MYC (Fig. 4E-F). XGBoost calculated partial dependence and feature importance for 13 DR-CSRGs after multiple iterations, identifying SERPINE1, LOX, CCND1, IRF1, and MYC as core genes (Fig. 4G-H). A Venn diagram (Fig. 4I) showed LOX and MYC as key biomarkers for DR-CS, overlapping across all four algorithms.

Fig. 4.

Machine learning screening of Key Biomarkers. (A) LASSO cross-validation plot showing model performance for different lambda values. (B) LASSO regression path plot showing the change of coefficients under different lambda values. (C) The number of genes with the highest accuracy obtained using the SVM-RFE algorithm. (D) The number of genes with the lowest error rate obtained via the SVM-RFE algorithm. (E) Evaluation of RF Model Performance. (F) Relative importance of candidate genes calculated in RF. (G) XGBoost machine learning algorithm for the selection of core DR-CSRGs. (H) Gene importance ranking from the XGBoost model. (I) Venn diagram related to key biomarkers

Nomogram development and ROC analysis

A nomogram was constructed to evaluate the contribution of key biomarkers LOX and MYC in predicting DR occurrence. The nomogram was developed based on a logistic regression model, which is a widely used statistical method for binary classification problems. In this model, the presence or absence of DR served as the dependent variable, while the expression levels of key biomarkers (LOX and MYC) were included as independent variables. The logistic regression model allows us to estimate the probability of DR occurrence based on the combined effects of these biomarkers. The nomogram visually presents the relationship between each biomarker and the predicted probability, represented as a column-line graph (Fig. 5A). The calibration curve, demonstrating a slope close to the ideal, underscores the nomogram’s predictive accuracy (Fig. 5B). Moreover, ROC curves revealed that LOX (AUC = 0.804), MYC (AUC = 0.964), and the Nomogram (AUC = 0.994) all exhibited AUC values exceeding 0.70, indicative of good predictive value (Fig. 5C).

Fig. 5.

Predictive ability and expression analysis of key biomarkers for DR. (A) Nomogram for assessing the risk of DR occurrence using LOX and MYC. (B) Calibration curve evaluating the predictive ability of the nomogram model. (C) ROC curves of key biomarkers in training set. (D) MYC expression in the training set. (E) LOX expression in the training sets. (F) MYC expression in the validation set GSE94019. (G) LOX expression in the validation set GSE94019. (H) ROC curves of LOX and MYC in validation set GSE94019

Expression analysis indicated that LOX and MYC were consistently upregulated in both the training set and the validation set (GSE94019), with statistically significant differences between the DR and control groups (P < 0.05) (Fig. 5D-G). ROC analysis of the validation set GSE94019 showed AUC values for LOX and MYC above 0.70 (Fig. 5H), suggesting their potential as accurate DR biomarkers.

Functional exploration of key biomarkers

The GeneMANIA database was employed to predict the top 20 genes related to MYC and LOX functions, including RB1, PETN, and LOXL1. A gene-gene interaction network was constructed based on these genes, which are primarily involved in physical interactions. The functions of LOX and MYC were found to be associated with processes such as extracellular matrix organization, protein oxidation, oxidoreductase activity, and protein-lipid complex assembly (Fig. 6A). GSEA analysis revealed that MYC was enriched in pathways such as extracellular matrix (ECM) proteoglycans, signaling by interleukins, and extracellular matrix organization. LOX was enriched in pathways like ECM receptor interaction, miRNA targets in ECM and membrane receptors, and collagen formation (Fig. 6B).

Fig. 6.

Specific signaling mechanisms of MYC and LOX. (A) Co-expression network mapping of coregulated genes. (B) GSEA enrichment analysis of MYC-associated and LOX-associated pathways

Correlation of potential biomarkers with different immune cells

A heatmap illustrated the distribution of 22 immune cell types across samples from the DR and normal groups (Fig. 7A). After excluding immune cells with a relative proportion of 0 or less than 50% in the samples, differences in immune cell composition between the DR and normal groups were compared. The analysis revealed that naive B cells, activated NK cells, monocytes, and M2 macrophages were more abundant in the DR group, whereas resting mast cells were more prevalent in the normal group (Fig. 7B). Furthermore, the correlation of LOX and MYC expression levels with immune cells was calculated. LOX expression showed significant correlations with memory B cells, plasma cells, regulatory T cells (Tregs), resting NK cells, and resting mast cells resting (Fig. 7C). MYC expression exhibited significant correlations with memory B cells, resting NK cells, and resting mast cells resting (Fig. 7D).

Fig. 7.

Immune infiltration analysis. (A) Heatmap displaying the relative abundance of 22 immune cell types in each sample from DR and normal groups. (B) Boxplot illustrating the differences in immune cell infiltration between the DR and normal groups. ^*P < 0.05, ^**P < 0.01, ^***P < 0.001. (C) Correlation analysis between LOX expression and immune cell types. (D) Correlation analysis between MYC expression and immune cell types

Key biomarker chromosome localization, subcellular localization, and single-cell sequencing analysis

LOX and MYC were located on chromosomes 5 and 8 (Fig. 8A). LOX mRNA and MYC mRNA were predominantly expressed in the cytoplasm (Fig. 8B). The GSE165784 dataset comprised fibrovascular membranes (FVMs) from five patients with proliferative diabetic retinopathy (PDR). After batch correction and unsupervised clustering, 8 clusters corresponding to 11 cell types were identified (Fig. 8C), including microglia (with 4 subclusters), monocytes, dendritic cells, macrophages, fibroblasts, endothelial cells, pericytes, and T cells. Microglia were the predominant cell type in FVMs in PDR. MYC was mainly expressed in microglia, macrophages, and pericytes (Fig. 8D), whereas LOX was highly expressed in fibroblasts and pericytes (Fig. 8E). Intercellular interaction analysis revealed strong interactions between pericytes and other cell types, including fibroblasts, macrophages, and endothelial cells (Fig. 8F-G).

Fig. 8.

Key biomarker chromosome localization, subcellular localization, and Single-Cell sequencing analysis. (A) Chromosomal localization map of LOX and MYC. (B) Subcellular localization prediction map of LOX mRNA and MYC mRNA. (C) The result of cell subpopulation distribution in two groups. (D) Feature and box plot showing the distribution of MYC in various cell types. (E) Feature and box plot showing the distribution of LOX in various cell types. (F) Chordal diagram showing the interaction strength based on ligand-receptor pairs. (G) Heatmap showing the number of ligand-receptor interactions between different cell types

Experiment validation

In the diabetes-induced mouse model, the two groups showed gradual weight gain over time. However, after 8 weeks, mice in the model group exhibited significantly lower weight than the control group (P < 0.05). Fasting blood glucose levels in the model group also showed a significant increase compared to the control group (P < 0.05) (Fig. 9A), confirming the successful establishment of the diabetes model. HE staining revealed normal retinal structure in the control group, with a thick retina, clear layer boundaries, and tight cell arrangement. In contrast, the DR group showed retinal thinning, loose structure, increased intercellular spaces, and vascular dilation and congestion in the ganglion cell layer (Fig. 9B), indicating the successful modeling of DR. WB and RT-qPCR assays were used to measure the expression of key biomarkers in retinal tissues. RT-qPCR showed higher mRNA levels of P53, P21, SASP (including TNF-α, IL-6, MCP-1 and VEGF) in the DR group than in the normal control group. As P53, P21 and SASP are markers of CS [17], this suggests a link between CS and DR. Additionally, MYC and LOX expression levels were significantly higher in the DR group than in the normal control group (P < 0.05) (Fig. 9C-D), confirming the reliability of the findings.

Fig. 9.

Experiment validation. (A) Changes in body weight and fasting blood glucose of mice during rearing. (B) Pathological staining images and each layer of retinal thickness in different groups of mice. (C) The mRNA expression levels of P53, P21, SASP (including TNF-α, IL-6, MCP-1, and VEGF), MYC and LOX in different groups of mice. (D) Protein expression levels of MYC and LOX in different groups of mice

Discussion

DR, a chronic diabetic complication, is a leading cause of blindness and visual impairment in the working-age population [18]. Current research on DR pathogenesis focuses on oxidative stress, inflammation, neurodegeneration, and vascular dysfunction. These factors exacerbate endothelial damage and the release of inflammatory cytokines, leading to retinal cell injury and apoptosis [19]. SASP comprises a variety of inflammatory cytokines, chemokines, growth factors, and small-molecule metabolites. Chronic inflammation, a key feature of DR, is also an effective inducer of CS [20, 21]. Specifically, in DR, chronic inflammation accelerates CS, and the resulting SASP exacerbates the inflammatory environment. From a therapeutic perspective, senescence-induced SASP is an important participant in DR pathogenesis. Thus, targeting senescent cells or modulating SASP represents a novel therapeutic approach for DR [7, 10]. To improve the prognosis of DR patients, it is necessary to identify specific key biomarkers related to CS and study the immune cell infiltration patterns associated with DR. Such efforts will enhance our understanding of how CS impacts DR development.

Bioinformatics approaches are increasingly applied in disease diagnosis and therapeutic target discovery, previous bioinformatics analyses have explored biomarkers associated with pyroptosis, inflammation, and ferroptosis in DR [22–24]. In this study, we identified 67 key hub genes. Subsequently, GO analysis revealed that these DR-CS genes play crucial roles in numerous cellular processes, and KEGG analysis identified 85 enriched pathways, among which the AGE-RAGE signaling pathway in diabetic complications, the PI3K-Akt signaling pathway, the p53 signaling pathway, and CS related pathways were closely associated with the pathogenesis of DR. The interaction between AGEs and their receptor RAGE is a key pathological mechanism in diabetic complications, associated with retinal pathologies like capillary basement membrane thickening, blood-retinal barrier disruption, and neuronal apoptosis, promoting DR development [25, 26]. Throughout the different stages of DR, the PI3K-Akt pathway exhibits varying levels of activation. During the early stages of DR, the PI3K/AKT-Nrf2 pathway reduces ROS formation, enhances antioxidant activity, and decreases retinal pigment epithelium apoptosis and senescence. In the late stages, the VEGFA/PI3K/Akt pathway significantly contributes to cellular apoptosis and compensatory angiogenesis in DR [27, 28]. Studies have demonstrated a substantial increase in p53 expression within retinal endothelial cell clusters of db/db mice. These findings confirm that p53 accelerates endothelial CS and exacerbates DR progression [29]. By constructing a PPI network and utilizing key Cytoscape plugins and four machine-learning methods, we screened out two closely linked core genes, MYC and LOX, which exhibited close associations with DR. ROC curve analysis revealed that they hold promising diagnostic value for DR.

Lysyl oxidase (LOX), one of the five members of the LOX family (LOX, LOXL1-LOXL4), plays a crucial role in the assembly of the ECM by catalyzing the cross-linking of collagen and elastin chains [30]. Guido Kroemer has suggested that ECM changes should be recognized as a hallmark of aging, given that the ECM significantly impacts mitochondrial homeostasis, CS, and stem cell exhaustion. ECM changes may both contribute to and result from aging processes [31]. Research has indicated that LOX not only physiologically remodels the ECM but also induces CS under oxidative stress conditions [32], thereby emphasizing the relationship between LOX and CS. Furthermore, LOX is a key enzyme responsible for the maturation and development of the retinal capillary basement membrane and is associated with regulating its ultrastructural integrity [33]. Previous studies have shown that LOX levels are significantly increased in DR [34, 35]. In a high-glucose environment, the upregulation of LOX can damage the ultrastructure and function of the basement membrane, leading to retinal endothelial cell dysfunction and increased permeability. Additionally, the increased expression and activity of LOX can promote AKT signaling inactivation, thereby activating caspase-3 and consequently inducing apoptosis in retinal microvascular endothelial cells [36]. Moreover, LOX overexpression is associated with retinal vascular inflammation and degeneration in diabetes, resulting in increased mechanical regulation of retinal vascular stiffness. Studies have demonstrated that LOX expression is regulated by several cytokines in inflammatory pathways, particularly TNF-α. High concentrations of TNF-α increase LOX expression through PI3K/Akt and Smad3 signaling pathways, as well as TGF-β-mediated signaling. In turn, LOX, as a downstream effector of RAGE, can mediate the activation of endothelial cells and matrix remodeling via AGE-RAGE pathway, thus serving as a key determinant of DR vascular inflammation [37]. In this study, we found that LOX was upregulated in the DR model, and its mechanism of activating CS might involve inflammatory responses and the promotion of vascular basement membrane thickening.

The MYC gene family, including c-MYC, N-MYC, L-myc, m-MYC, MYC-R, and MYC-D, belongs to the bHLH-Zip DNA-binding protein superfamily. These genes regulate approximately 10–15% of the genome, controlling transcriptional programs that drive cellular processes such as metabolism, growth, proliferation, differentiation, and senescence [38]. MYC can promote cell proliferation and alter cell-cycle dynamics. When expressed at high levels, it may cause replicative stress, genomic instability, and increased apoptosis sensitivity [39]. In retinal pigment epithelium cells, N-MYC activation can cause nucleolar fusion and cytoplasmic granule formation, increase p21 and p53 protein levels, thereby inducing a strong senescence-like phenotype [40]. In lens epithelial cells, downregulation of N-MYC downstream-regulated gene 2 can inhibit UVB radiation-induced CS by regulating pyroptosis via the caspase-1/NLRP3 pathway. c-MYC is closely associated with angiogenesis and plays a critical role in DR progression. When Müller glial cells are exposed to high glucose, c-MYC can enhance the release of proinflammatory cytokines such as IL-1β, TNF-α, and IL-6 through the MIAT/TXNIP pathway [41, 42]. FBXW7, an E3 ubiquitin ligase for c-MYC, is activated by Rg1 through the inhibition of miR-100-3p. This process leads to c-MYC ubiquitination and degradation, thereby reducing proliferation, migration, invasion, and angiogenesis in human retinal microvascular endothelial cells cultured under high-glucose conditions. It can also alleviate vascular leakage and capillary degeneration in DR [43].

Immune dysregulation is a key driver of DR pathogenesis, yet immune system modulation remains poorly understood. CIBERSORT analysis reveals significant differences in immune cell composition between DR patients and healthy individuals. Specifically, in DR tissues, naive B cells, activated NK cells, monocytes, and M2 macrophages are elevated, whereas normal tissues have higher levels of resting mast cells. Studies have shown that during DR progression, an increasing number of B cells infiltrate the retinal macula [44]. Additionally, more monocytes and monocyte-derived macrophages are recruited to the retina in DR patients, with monocyte infiltration being a cause of increased retinal vascular permeability. During tissue remodeling, infiltrating monocytes differentiate into pro-repair (M2) macrophages. These macrophages facilitate inflammation resolution through phagocytosis, pinocytosis, and the production of pro-resolving mediators [45–47]. Natural killer (NK) cells, a type of innate lymphocytes, control the initiation of inflammatory responses via the production of interferon-γ (IFN-γ). In diseased choroidal tissue, aggregated NK cells promote neutrophil NETosis, leading to the formation of neutrophil extracellular traps (NETs) and aiding vascular repair [48]. Few studies have explored the role of mast cells in DR pathogenesis. However, research indicates that mast cell deficiency in the bursa premacularis may be associated with the pathogenesis of PDR [49]. Extensive experimental research shows that targeting immune cells with antibodies, recombinant proteins, or pharmacological and gene therapy approaches can reduce neurovascular degeneration during DR progression and promote tissue remodeling. However, further studies are needed in this field [50]. Furthermore, LOX correlates with memory B cells, plasma cells, Tregs, resting NK cells, and resting mast cells. MYC shows correlations with memory B cells, resting NK cells, and resting mast cells. Regulatory T (Treg) cells, a specific CD4⁺ T cell subpopulation, perform immunosuppressive functions in chronic inflammatory conditions like aging and diabetes [51]. These findings underscore the complex interplay between CS and immune responses in DR pathogenesis.

We utilized single-cell datasets to analyze the localization of specific genes in different cells. MYC is widely present in microglia. Hyperglycemia activates microglia, inducing the release of inflammatory and chemokine factors, thereby triggering retinal inflammation [52]. Studies have shown that the upregulation of Myc can mediate the transformation of microglia into dendritic cells, driven by the ERK/Myc signaling pathway. This highlights its role in neuroinflammatory responses [53]. LOX is predominantly highly expressed in fibroblasts and pericytes. In a high-glucose environment, fibroblasts with high LOX expression are activated to secrete multiple fibrotic factors and collagen, promoting retinal vascular basement membrane thickening, retinal cell inflammation, and fibrosis, thereby contributing to DR development [54, 55]. Pericytes are crucial support cells for retinal vessels [56]. Inhibiting LOX expression can prevent pericyte loss and vascular leakage in diabetic retinopathy [57]. In cell communication studies, pericytes showed strong interactions with fibroblasts and endothelial cells, highlighting the importance of intercellular interactions in DR pathology. When pericytes undergo apoptosis, they release cytokines and matrix metalloproteinases, which activate surrounding fibroblasts and promote their proliferation and ECM deposition. Additionally, LOX-PP, released during LOX processing, is associated with apoptosis in various pathological tissues. In retinal endothelial cells cultured under high glucose, LOX-PP expression increases, potentially impairing AKT activity and inducing apoptosis [58]. Thus, intercellular communications may be associated with LOX functions. In this study, we established a model of DR and detected the expression of SASP, P53 and P21 to confirm the correlation between CS and DR. We also validated the expression of LOX and MYC in this model. These findings, consistent with bioinformatics analyses, suggest that LOX and MYC may serve as potential therapeutic targets for DR.

Current treatments for DR targeting CS mainly focus on senolytics, including Bcl-2 inhibitors, heat shock protein 90 inhibitors, BCL-xL inhibitors, dasatinib, and quercetin. These can eliminate senescent cells from various tissues without damaging healthy cells, thereby slowing disease progression and improving overall health. However, related clinical trials require large-scale validation [59]. In this study, we conducted bioinformatics analyses and animal experiments to uncover the potential roles of MYC and LOX in DR-CS. However, our study has limitations. Firstly, the small sample size of some datasets might have affected the statistical power and generalizability of the results. The limited sample size could increase the risk of false-negative results in gene expression difference analysis and subsequent correlation studies. Consequently, some genes closely related to DR pathogenesis might not have shown significant differences. Moreover, due to individual differences, environmental factors, and the clinical heterogeneity of DR, the key biomarkers identified in this study may not fully represent all DR patients, and their generalizability may be limited. Secondly, due to experimental limitations, we did not carry out further in-vitro experiments. Future research will employ gene silencing and/or overexpression experiments to further elucidate the functional roles in DR-CS, for instance, siRNA knockdown of LOX will be conducted, followed by quantification of senescence marker and the profiling of SASP-related gene expression. These steps aim to gain a more comprehensive understanding of their roles in DR.

Conclusion

This study integrated transcriptomic and single-cell data, identifying MYC and LOX as key DR-CS biomarkers. We characterized their functional roles and related pathways, and explored the association between CSRGs and DR pathophysiology. These two biomarkers showed marked differential expression in DR model and a strong correlation with immune cell infiltration, and exhibited high clinical value for disease diagnosis. Collectively, these findings enhance our understanding of DR pathogenesis and suggest novel therapeutic targets, thereby providing significant insights for future research and clinical strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, Jinju Li and Zhaohui Fang; methodology, Jinju Li and Hao Yang; data curation, Jinju Li and Yixuan Lin; formal analysis, Jinju Li and Hao Yang; writing—original draft, Jinju Li; writing—review and editing, Yixuan Lin and Zhaohui Fang; funding acquisition, Yixuan Lin and Zhaohui Fang. All the authors participated in planning, execution, and analysis and have read and approved the final submitted version.

Funding

This study was supported by the Project of the Research Institute of Health of Hefei Comprehensive National Science Centre (2023CXMMTCM003), the National Natural Science Foundation of China (82174153), the Anhui Higher Education Research Project (2023AH050867) and the Clinical Research Project of Anhui University of Chinese Medicine (2024YFYLCZX11).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was approved by the Animal Ethics Committee of Anhui University of Chinese Medicine (Approval No. AHUCM-mouse-2024157).

Competing interests

The authors declare no competing interests.