Identification of Biomarkers for Oxidative Stress in Age-Related Macular Degeneration: Combining Transcriptomics and Mendelian Randomization Analysis

Abstract

Purpose

Oxidative stress has long been recognized as a significant influence in the pathophysiology of age-related macular degeneration (AMD). Therefore there is a need to explore the relationship between oxidative stress-related biomarkers and AMD.

Methods

Based on Gene Expression Omnibus database–Gene Expression Omnibus Series (GSE)29801 and GSE135092 datasets, three machine learning methods were used to screen biomarkers. The Wilcoxon test was used to compare the percentage of immune cells in control and AMD samples. The causal relationship between biomarkers and AMD was explored in a series of Mendelian randomization (MR) analyses. Ultimately, the expression levels of biomarkers were validated by quantitative real-time polymerase chain reaction (qRT-PCR) in the simulated AMD cell model.

Results

A total of 16 differentially expressed oxidative stress-related genes (DE-OSRGs) were screened. Functional enrichment analysis indicated that DE-OSRGs participated in cellular senescence, cell cycle regulation, and PPAR signaling pathways. Machine learning methods were used to screen for five biomarkers (GFAP, Stearoyl-CoA desaturase [SCD], BCKDHB, GPX8, and MSRB2). The qRT-PCR results showed that the expression levels of five biomarkers were significantly different between the simulated AMD cell model and control groups. Spearman correlation analysis showed that GPX8 had the highest positive correlation with M2 macrophages (correlation coefficient [cor] = 0.36, P < 0.01), and SCD had a strong negative correlation with eosinophils (cor = −0.28, P < 0.05). MR results revealed that BCKDHB played a crucial role as a risk factor for AMD (odds ratio > 1, P < 0.05).

Conclusions

This study screened the biomarkers related to oxidative stress in AMD, providing a certain theoretical basis for the prevention and clinical diagnosis of AMD.

Translational Relevance

Identifying biomarkers with diagnostic value for AMD could provide new understanding of its pathogenesis, and open up potential targets for clinical intervention.

Introduction

Age-related macular degeneration (AMD) stands as a prominent cause of visual impairment and significant vision loss.¹ Forecasts indicate that the population of individuals suffering from AMD will surge to 288 million by 2040,² which would be a heavy burden to the health system. According to the AMD classification system, the clinical phenotype includes neovascular AMD, geographic atrophy, or both, all of which associated with initial manifestations of accumulation of drusen and pigmentary abnormalities.³ Currently, anti-vascular endothelial growth factor was the mainly treatment for neovascular AMD (nAMD) with limited effectiveness, which also made an economical and mental burden on patients.⁴ No exact treatment can affect the progression of geographic atrophy.¹ Therefore there is a pressing need to intensify efforts toward the prevention of advanced AMD. Identifying causal and modifiable risk factors for advanced AMD is paramount for implementing effective preventive interventions.

In the pathogenesis of AMD, oxidative stress has long been thought to have a significant impact on the retinal pigment epithelium (RPE),⁵ supporting by both environmental and genetic factors. There is an etiologic role for genes that encode oxidative stress-related proteins.⁵ For example, complement factor H, the primary regulator of the alternative pathway within the complement system, may play a pivotal role in the pathogenesis of AMD by modulating oxidative stress responses in the outer blood-retinal barrier.^6,7 Although the discovery of these genetic variants has led to help predict the risk of developing AMD, more details remain investigated. A novel research approach is necessary to investigate the relationship between oxidative stress-related biomarkers and AMD.

Mendelian randomization (MR) is a method that determines whether a risk factor has a causal effect on a health outcome by using genetic variants as instrumental variables. By using genetic variations that are thought to meet the instrumental variables (IVs) assumptions, it provides an alternate method of examining the question of causation in epidemiological research.⁸ This technique has been widely used in previous AMD studies, including evaluating the causal relationships between eight serum lipid biomarkers,⁹ investigating into associations between genetic predictors of lipid fractions, smoking and AMD risk,^10,11 verifying high-density lipoprotein cholesterol a causal risk factor for AMD risk. And in a study exploring the association between sleep duration and AMD, researchers found that there is an association between insufficient sleep and AMD through cross-sectional studies and MR analysis, and insufficient sleep increases the risk of early AMD onset.¹²

Although a series of studies have already revealed a wide range of molecular mechanisms involved in AMD, the genetic investigation of oxidative stress-related biomarkers for AMD remains largely unexplored.^13–16 In this study, based on the transcriptome data of AMD, we conducted a systematic research by integrating machine learning algorithms, immune infiltration analysis, MR analysis, and experimental verification methods. The study not only screened out biomarkers related to oxidative stress but also verified the causal association between biomarkers and AMD at the genetic level. It is worth noting that previous MR analyses in AMD research mostly focused on traditional risk factors such as lipid fractions and smoking.^10,11 However, in this study, it was applied to verify the causal relationship between oxidative stress-related genes (OSRGs) and AMD, exploring potential targets for the early prevention and clinical treatment of AMD.

Methods

Data Acquisition

The datasets-related to AMD were derived from Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). The analysis was conducted on a total of 41 AMD and 50 normal macular RPE-choroid tissue samples from the GSE29801 cohort, as well as 26 AMD and 105 normal macular RPE-choroid tissue samples from the GSE135092 cohort. Moreover, the genome-wide association studies (GWAS) data of outcome and eQTL GWAS data of exposure factors in MR analysis were all gathered from the OpenGWAS database (https://gwas.mrcieu.ac.uk/). The finn-b-H7_AMD dataset within AMD's GWAS comprised 3763 samples with AMD and 205,359 normal samples (European race), encompassing a total of 16,380,424 single-nucleotide polymorphisms (SNPs). Meanwhile, 1104 OSRGs were retrieved from GeneCards database (https://www.genecards.org/, v5.20) with oxidative stress as keywords and score > 7.5.

Differential Expression and Functional Enrichment Analyses

The limma package (v3.52.4)¹⁷ was used to identify differentially expressed genes (DEGs) between AMD and control groups in the GSE29801 and GSE135092 datasets, using a threshold of |log₂FC| > 0 and P < 0.05. The DEGs were visualized using the volcano plot and heat map generated by the ggplot2 (v3.3.6) and ComplexHeatmap (v2.12.1) packages, respectively.¹⁸ Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were implemented through the clusterProfiler package (v4.7.1.001).¹⁹ The biological pathways that exhibited significant enrichment were visualized by GOplot package (v1.0.2; P < 0.05).

Revealing Potential Biomarkers for AMD

Three machine learning algorithms—least absolute shrinkage and selection operator (LASSO), support vector machine recursive feature elimination (SVM-RFE), and Boruta—were used to select biomarkers. The LASSO algorithm was implemented using the glmnet package (v4.1-6),²⁰ SVM-RFE was implemented using the e1071 package (v1.7-12),²¹ and the Boruta algorithm was used with the assistance of the Boruta package (v8.0.0).²² The feature genes obtained by above machine learning algorithms were intersected using ggvenn package (v0.1.9) to derive biomarkers. After that, the RCircos package (v1.2.2) was used for the analysis of biomarker chromosome locations in this study.²³ The Spearman correlation among the expression levels of the biomarkers was analyzed and demonstrated through the psych (v2.2.9) and ggcor (v0.9.8.1) package, respectively.

Gene Set Enrichment Analysis (GSEA)

To conduct functional analysis, we initially performed Spearman correlation analysis between each biomarker and all genes. Subsequently, the results were sorted based on the correlation coefficient for further GSEA analysis using the clusterProfiler package (|NES| > 1, P < 0.05). The gene set used for GSEA was c2.cp.kegg.v7.0.symbols.gmt, which was downloaded from the GSEA website (http://www.gsea-MSigdb.org/gsea/msigdb).

Immune Infiltration Analysis

The proportion of 22 immune cell types in the GSE29801 dataset were determined using the CIBERSORT algorithm and LM22 gene set 11 and samples with the P > 0.05 were excluded from further analysis. The ggplot2 package was used to visualize the infiltration abundance percentages of these immune cell types in both control and AMD samples. Differences in immune cell proportions between control and AMD samples were assessed using the Wilcoxon test.

Establishment of Regulatory Network and Prediction of Potential Therapeutic Drugs

We predicted the transcription factors (TFs) targeting biomarkers in the ChEA3 database and screened for TFs supported by ChIP-seq data in the ENCODE database. The microRNAs (miRNAs) associated with biomarkers were identified using the miRTarBase database. Network mapping was performed using Cytoscape software to illustrate the regulatory relationships between the biomarkers, TFs, and miRNAs. In addition, the DGIdb database was used to forecast drugs that target specific biomarkers, aiming to discover potential therapeutic interventions for AMD.

Study Design for MR Analysis

Using a two-sample MR analysis, we evaluated the causal association between biomarkers and AMD. Hence, the choice of effective IVs must adhere to three fundamental assumptions: (i) IVs must exhibit a strong association with the exposure; (ii) IVs were not associated with any potential confounding factors, including outcomes; (iii) IVs only exert an impact on the outcome through exposure factors, and there are no alternative routes of influence.

MR Analysis of AMD and Biomarkers

Using the extract_instruments function in TwoSampleMR package (v0.5.7),²⁴ IVs (SNPs) with significant correlation with biomarkers (exposure factors) were read and screened with P < 5 × 10⁻⁸ as the screening criteria. Subsequently, the clump parameter should be set to TRUE, and IVs in linkage disequilibrium should be eliminated by adjusting the parameters accordingly as follows: r² = 0.001; kb = 100. To address the issue of IVs that show significant correlation with AMD (outcome), we standardized the effect alleles and effect sizes using the harmonise_data function in the TwoSampleMR package. This ensured that the effect alleles of the IVs were consistent between the exposure factors and the outcome, and that the units and direction of the effect sizes were also standardized. The F value of each IVs was shown in Supplementary Table S1. Subsequently, five algorithms (MR egger, weighted median, inverse variance weighted [IVW], simple mode, and weighted mode) were used for MR analysis. Notably, the IVW result was considered the primary outcome, with a P < 0.05 indicating a significant causal relationship. Finally, we visualized the results of the MR analysis by plotting forest map, scatter plot, and funnel plot. The causal effect estimation of each SNP on AMD were presented through the forest plot. Scatter plot could visually demonstrate the influence of IVs on exposure factors and outcome. The funnel plot was used to evaluate whether the IVs in MR analysis were symmetrically distributed.^25,26

Sensitivity Analysis

To assess the robustness of the findings of MR analysis, we performed a sensitivity analysis. The heterogeneity was detected by calculating the Cochran's Q statistic and its corresponding P value. A P value > 0.05 indicated no significant heterogeneity.^27,28 Furthermore, the MR egger intercept was assessed to examine potential horizontal pleiotropy.^26,29 It is important to highlight that if the calculated P value of the MR egger intercept is greater than 0.05, it indicates no heterogeneity.³⁰ However, if the calculated P value of the MR egger intercept is lower than 0.05, it requires a reconsideration of the research design.^29,31,32 Ultimately, a leave-one-out method was used for recalculating results after systematically removing individual SNPs. ³³ And each SNP was removed in sequence, and the remaining SNPS were used to calculate the causal effect of exposure on the outcome. ^34,35 In this analysis, it was crucial to ascertain the accuracy of the direction, which could be achieved by using the TwoSampleMR package to complete the Steiger directivity analysis.^26,36

Quantitative Real-Time Polymerase Chain Reaction (qRT-PCR)

The human retinal pigment epithelial cell line (ARPE-19) was purchased from the American Type Culture Collection (CRL-2302) and cultured at 37° C in a medium containing D/F12, 10% fetal bovine serum and PS (control group). In order to simulate the pathophysiology of AMD in ARPE-19 cells, we performed a dose-dependent pre-test to determine the optimal sodium iodate (SI) concentration (Supplementary Fig. S1). And, the concentration that significantly reduced cell activity but did not kill the cells completely was then selected (the cell activity was reduced to about 50% of the control group). Finally, we induced oxidative stress with 3mM SI (MCE, HY-Y0628), cultured with Dulbecco's modified Eagle medium base medium for 24 hours, then replaced with D/F12 complete culture medium for 24 hours (AMD group). We used the FastPure Complex Tissue/Cell Total RNA Isolation Kit (Vazyme, Nanjing, China) to extract total RNA. Then, the total RNA was reverse-transcribed into cDNA using the ABScript III RT Master Mix for qPCR with gDNA Remover (ABclonal, Wuhan, China). Thereafter, qRT-PCR was performed using the Genious 2X SYBR Green Fast qPCR Mix (ABclonal). To guarantee the accuracy of the experimental findings, each experimental group was subjected to three separate replicates. In the data analysis stage, we used the widely used 2^−ΔΔCt method for relative quantification, using GAPDH as an internal control to guarantee the precision and comparability of the findings. The primer sequences were provided in Table 1.

Table 1.

The Primer Sequences of Biomarkers

Primer Name	Sequence 5′-3′
H-GAPDH	F:5′-GGAGTCCACTGGCGTCTTCA -3′
	R:5′-GTCATGAGTCCTTCCACGATACC -3′
GFAP	F:5′-AGGTCCATGTGGAGCTTGAC-3′
	R:5′-GCCATTGCCTCATACTGCGT-3′
SCD	F:5′-TTCCTACCTGCAAGTTCTACACC-3′
	R:5′-CCGAGCTTTGTAAGAGCGGT-3′
BCKDHB	F:5′-GATTTGGAATCGGAATTGCGG-3′
	R:5′-CAGAGCGATAGCGATACTTGG-3′
GPX8	F:5′-CCGCCCAAGCAAGGAAGTAG-3′
	R:5′-TCTAACCAGAGCTGCTATGTCAG-3′
MSRB2	F:5′-CGAGGCTCATGGTACGTCTG-3′
	R:5′-GCTGATCCTAACGAGGTATCCAG-3′

Statistical Analysis

All statistical analyses were conducted using R package (v4.2.1). Various variables, including expression quantity and infiltration ratio, were assessed using either the Wilcoxon test or t-test to compare differences between groups. In the graphical representation, ns denotes P > 0.05, * indicates P < 0.05, ** represents P < 0.01, *** signifies P < 0.001, and ****implies P < 0.0001.

Results

Identification of DEGs Between AMD and Control Groups

The GSE29801 dataset revealed a total of 1508 DEGs between the AMD and control groups, comprising 904 down-regulated and 604 up-regulated genes. Figures 1A and 1B illustrated the volcano plot of DEGs and the heat map displaying the top 20 up- and downregulated DEGs. Similarly, in the GSE135092 dataset, we identified 2231 DEGs (1200 upregulated genes and 1031 downregulated genes) between AMD patients and controls (Figs. 1C, 1D). To identify DEGs exhibiting consistent expression patterns in both the GSE29801 and GSE135092 datasets, we performed an intersection analysis of DEGs separately for upregulated and downregulated genes. As shown in Figure 1E, a total of 121 upregulated intersecting genes and 127 downregulated intersecting genes were identified, resulting in a set of 248 DEGs that were subsequently used for further analysis.

Figure 1.

Identification of DEGs in the GSE29801 and GSE135092 datasets. (A, B) Volcano plot and heatmap of DEGs in the GSE29801 dataset. (C, D) Volcano plot and heatmap of DEGs in the GSE135092 dataset. (E) Venn diagram of up-and down-regulated DEGs in GSE29801 and GSE135092 datasets.

Screening of DEGs Associated With OSRGs and Mining of Their Potential Functions

The intersection of 248 DEGs identified above and 1104 OSRGs yielded a set of 16 intersecting genes, which were designated as differentially expressed OSRGs (DE-OSRGs) (Fig. 2A). To elucidate the molecular function and mechanism of DE-OSRGs, we performed GO and KEGG enrichment analyses to identify shared functions and associated pathways among genes. The findings demonstrated that these DE-OSRGs were implicated in the regulation of secretion, cellular apoptotic processes, and enzymatic activities, such as positive regulation of secretion, muscle cell apoptotic process, lipase activity (Fig. 2B). The DE-OSRGs were found to be associated with crucial cellular processes such as cellular senescence, cell cycle regulation, PPAR signaling, and p53 signaling pathways in accordance with KEGG annotations (Fig. 2C).

Figure 2.

Identification and enrichment analysis of DE-OSRGs. (A) Venn diagram of 248 DEGs and 1104 OSRGs; (B) GO enrichment analysis of DE-OSRGs; (C) KEGG enrichment of DE-OSRGs.

Screening of Potential Biomarkers Linked to Oxidative Stress in AMD

To identify biomarkers associated with the progression of AMD, we performed an analysis on 16 DE-OSRGs using three machine learning algorithms. The LASSO model with the least error identified eight genes, including GFAP, Stearoyl-CoA desaturase (SCD), BCKDHB, GPX8, HADHB, PLA2G7, MSRB2, and CHEK2 (Fig. 3A). At the same time, the SVM-RFE model demonstrated the lowest error rate (0.225) when incorporating 12 genes, including GFAP, MSRB2, SCD, PLA2G7, BCKDHB, CHEK2, OXTR, TGFB2, AOX1, CDKN2A, GPX8, and HADHB (Fig. 3B). In addition, six genes (GFAP, SCD, CDKN2A, BCKDHB, GPX8, and MSRB2) were identified by Boruta (Fig. 3C). Subsequently, five genes (GFAP, SCD, BCKDHB, GPX8, and MSRB2) were selected by cross-referencing the genes identified by three machine learning algorithms (Fig. 3D). The results of gene expression analysis indicated that the expression levels of GFAP, SCD, and BCKDHB were significantly up-regulated in AMD samples, while GPX8 and MSRB2 exhibited significant down-regulation in both GSE29801 and GSE135092 datasets (Figs. 3E, 3F). Consequently, these five genes have been identified as biomarkers for AMD. Chromosome localization analysis revealed the genomic positions of five genes: GPX8 on chromosome 5, BCKDHB on chromosome 6, MSRB2 and SCD on chromosome 10, and GFAP on chromosome 17 (Fig. 3G). Based on Spearman correlation analysis, GPX8 and GFAP exhibited a strong negative correlation (cor = −0.71, P < 0.001), whereas GFAP and MSRB2 showed a significant positive (cor = 0.58, P < 0.001) (Fig. 3H).

Figure 3.

Machine learning screening and expression validation of biomarkers for AMD. (A–C) Feature genes screening by the LASSO, SVM-RFE, and Boruta algorithms. (D) The intersection Venn diagram of feature genes screened by three machine learning algorithms. (E, F) Expression analysis of biomarkers in the GSE29801 and GSE135092 datasets; (G) chromosomal distribution of biomarkers; and (H) correlation analysis of biomarkers.

Functional Analysis of Biomarkers

To preliminarily predict the potential mechanism of action for biomarkers in AMD, we conducted single-gene GSEA enrichment analysis for biomarkers. It was found that GFAP was associated with focal adhesion, adherens junctions, and regulation of actin cytoskeleton pathways (Fig. 4A). Additionally, BCKDHB and SCD significantly enriched peroxisome, whereas BCKDHB specifically enriched the nod-like receptor signaling pathway (Figs. 4B, 4C). In addition, MSRB2 was related to the Wnt signaling pathway, GPX8 was related to the proteasome, and they were collectively association with ribosome (Figs. 4D, 4E). To further investigate the gene interactions and co-functions of the biomarkers, we used GeneMANIA database to analyze and construct the network. As depicted in Figure 4F, a network comprising five biomarkers and 20 interacting genes was established, which exhibited associations with amino acid metabolic process and tricarboxylic acid cycle enzyme complex.

Figure 4.

GSEA enrichment analysis and GeneMANIA network of five biomarkers. (A–E) Construction of GeneMANIA network for five biomarkers; and (F) GSEA analysis enrichment of five biomarkers.

Immune Microenvironment Analysis in AMD Patients

To gain insights into the disparities in the immune microenvironment between the control and AMD groups, we conducted an immune infiltration analysis. First, the samples in the GSE29801 dataset was initially screened based on the proportion of 22 immune cells, resulting in the retention of 65 samples for subsequent analysis (P < 0.05). The percentage of infiltration abundance of 22 immune cell types in control and AMD samples was illustrated in Figure 5A. According to the Wilcoxon test, statistically significant variations were observed in the infiltration proportions of two different immune cell types, memory B cells and resting NK cells, between the control group and the AMD group (Fig. 5B). To explore the relationship between biomarkers and 22 distinct immune cell types, we conducted a Spearman correlation analysis. Our findings indicated that GPX8 demonstrated the strongest positive correlation (cor) with M2 macrophages (cor = 0.36, P < 0.01) (Fig. 5C). Conversely, we observed a strong negative correlation between SCD and eosinophils (cor = −0.28, P < 0.05) (Fig. 5C).

Figure 5.

Immune infiltration, regulatory network and drug prediction analysis. (A) Percentage of immune cell infiltration in normal and AMD samples; (B) differences in immune cell infiltration between normal and AMD groups; (C) correlation analysis of biomarkers and immune cells; (D) TF-mRNA regulatory network; (E) miRNA-mRNA regulatory network; and (F) drug-gene interaction.

Exploring the Regulatory Mechanism of Biomarkers and Prediction of Potential Drugs

Depending on the prediction in the ChEA3 and ENCODE databases, 70 TFs targeting five biomarkers were acquired. The network including 75 nodes and 102 edges was constructed, with EBF1 could simultaneously regulate the expression levels of SCD, BCKDHB, and MSRB2 (Fig. 5D). Similarly, the mRNA-miRNA network consisted of four biomarkers (GFAP, SCD, BCKDHB, and GPX8) along with 175 miRNAs they targeted (Fig. 5E). To enhance the prediction of potential drugs for AMD, we conducted a biomarker-targeted drug screening. Our analysis revealed two biomarkers associated with 29 drugs, including SCD linked to five drugs and the GFAP gene related to 25 drugs. Notably, our findings suggested that colchicine had the potential to target both GFAP and SCD (Fig. 5F).

Causal Effect of Biomarkers on AMD

To investigate the causal relationship between biomarkers and AMD, a two-sample MR analysis was completed. Because only four biomarkers had GWAS data (GFAP, SCD, BCKDHB, and MSRB2), we ultimately selected these four biomarkers for MR analysis. Utilizing 4 biomarkers with GWAS data as exposure variables and AMD as the outcome. The findings from the IVW analysis revealed a statistically significant correlation between BCKDHB and AMD, with an odds ratio (OR) greater than 1, P < 0.05, indicating that BCKDHB was a risk factor for AMD (Fig. 6A). However, the other three genes (GFAP, SCD and MSRB2) had no significant causal relationship with AMD (P > 0.05) (Fig. 6A). Then, the forest map was used to display each SNP site of BCKDHB for the purpose of predicting its impact on AMD, which was computed using the Wald ratio method. Figure 6B demonstrated that the overall effect size of the BCKDHB on the outcome variable was significantly greater than 0, indicating an elevated risk of AMD associated with increased levels of BCKDHB. The scatter plot indicated that BCKDHB was a risk factor for AMD (Fig. 6C). The results of the funnel plot was approximately symmetrically distributed from left to right, indicating that the MR analysis conformed to Mendel's second law of random grouping (Fig. 6D). To estimate the credibility of the MR analysis results, a sensitivity analysis was performed on the results. According to the heterogeneity test, the P value for heterogeneity of BCKDHB was found to be larger than 0.05, indicating no significant heterogeneity (Table 2). Additionally, the horizontal pleiotropy test of BCKDHB revealed that the P > 0.05, indicating an absence of horizontal pleiotropic effects (Table 3). This suggested that there were no confounding factors present in the study. The absence of any abnormal SNP results regarding to the leave-one-out results (Fig. 6E). Finally, and more significantly, Steiger test demonstrated that the expression of BCKDHB posed a significant risk factor for AMD, with no evidence of reverse causality (Direction = TRUE) (Table 4).

Figure 6.

The causality of biomarkers and AMD. (A) Causal analysis of four biomarkers and AMD; (B) Forest plot of MR analysis; (C) scatter plot of MR analysis; (D) funnel plot of MR analysis; and (E) leave-one-out analysis of MR analysis.

Table 2.

The Heterogeneity Test of BCKDHB and AMD

Exposure	Outcome	Method	Q	Q_df	Q_Pval
BCKDHB	AMD	MR Egger	0.747212839	6	0.993414927
BCKDHB	AMD	Inverse variance weighted	1.206257217	7	0.990782112

Q, Cochran's Q test estimate; Q_df, Q_degree of freedom.

Table 3.

The Horizontal Pleiotropy Analysis of BCKDHB and AMD

Exposure	Outcome	Egger_Intercept	SE	Pval
BCKDHB	AMD	0.095061492	0.140306339	0.523322711

Table 4.

Steiger Directivity Test

Exposure	Outcome	snp_r2.exposure	snp_r2.outcome	Correct_Causal_Direction	Steiger_Pval
BCKDHB	AMD	0.039504519	5.13E-05	TRUE	5.03E-204

Experimental Verification of Biomarkers Expression Levels

In this study, qRT-PCR was used to analyze the differential expression of GFAP, SCD, BCKDHB, GPX8, and MSRB2 in the control and simulated AMD cell model groups. As shown in Figure 7, compared with the control group, the expression levels of GFAP, SCD, BCKDHB and GPX8 were significantly increased, while the expression levels of MSRB2 were significantly decreased, which was consistent with the results of bioinformatics analysis (P < 0.05). These findings suggested that the expression levels of biomarkers potentially influence the development of AMD.

Figure 7.

Experimental verification of biomarkers expression levels. *P < 0.05; **P < 0.01; ***P< 0.001.

Discussion

AMD is one of the leading causes of vision loss in the elderly, and is becoming a global crisis, with 288 million people expected to be affected worldwide by 2040.³⁷ Genetics, environmental damage and age-related problems are risk factors for the development of the disease. All of these risk factors are related to the induction of oxidative stress.³⁸ We conducted a systemic framework to identify and investigate the relationships between oxidative stress-related biomarkers and AMD based on GEO database. Based on the GSE29801 and GSE135092 datasets, A total of 16 DE-OSRGs were screened. The results of the functional enrichment analysis suggested that these DE-OSRGs participated in cellular senescence, cell cycle regulation, and PPAR signaling pathways. Through machine learning methods, five biomarkers were filtered namely GFAP, SCD, BCKDHB, GPX8, and MSRB2. Bioinformatics and qRT-PCR results showed that the expression levels of five biomarkers were significantly different between AMD and control groups. Spearman correlation analysis showed that GPX8 had the highest positive correlation with M2 macrophages (cor = 0.36, P < 0.01), and SCD had a strong negative correlation with eosinophils (cor = −0.28, P < 0.05). MR results demonstrated that the expression of BCKDHB posed a significant risk factor for AMD (OR > 1, P < 0.05), whereas others probably expressed as post disease.

It has been shown that cellular senescence, cell cycle regulation, and PPAR signaling pathways are closely related to the development of AMD. Oxidative stress can elicit cellular senescence, and stress-induced premature senescence resulting from oxidative stress has been associated with the development of AMD.³⁹ The transcription factors known as peroxisome proliferator-activated receptors (PPARs) belong to the superfamily of steroid hormones and are triggered by ligands.⁴⁰ It has been suggested that PPARα activation might have beneficial effects on neovascular AMD.⁴¹ PPARγ is a member of the PPARs family, and its activation can regulate lipid metabolism, inflammatory response and cell proliferation.⁴² PPARγ is situated in the neuroretina and retinal pigment epithelium, which are critical to photoreceptor degradation and visual impairment.⁴³ In the pathogenesis of AMD, PPARγ may affect the progression of the disease by regulating lipid metabolism and inflammatory responses in RPE cells.⁴⁴ The activation of PPARγ can reduce the expression of inflammatory factors, which may have a protective effect on AMD.⁴⁴ Thus DE-OSRGs might influence the development of AMD by participating in these signaling pathways.

Actually, some of biomakers-GTAP, SCD and MSRB2-in our study have been verified by conventional studies, which can support our results. Methionine sulfoxide reductases (MSRs), including MSRA, MSRB1, MSRB2, and MSRB3, are antioxidant repair enzymes family and thought to safeguard various organs from oxidative damage induced by reactive oxygen species.⁴⁵ Prior research showed that the expression of MSRA and MSRB2 in human RPE cells is regulated by chemically induced hypoxia.⁴⁶ One of the human tissues with the highest expression levels of MSRB2 mRNA is the retina. Furthermore, the expression of MSRB2 is most prominent in the outer plexiform layer of the macula, as well as in the foveal regions. Some studies have indicated that MSRB2 plays a crucial role in protecting cones from various types of oxidative stress and may be essential for the preservation of central vision.⁴⁷ In our study, we also find that MSRB2 down-regulation in both GSE29801 and GSE135092 datasets AMD samples, which was in accordance with previous studies.

Glial fibrillary acidic protein (GFAP) is a well-recognized marker of retinal stress. GFAP, expressing in retinal astrocytes and Muller cells, has been demonstrated to be modulated by AMD.⁴⁸ The expression of GFAP increased in Muller cells within the neural retina, in conjunction with changes in the RPE and photoreceptor degeneration.⁴⁹ Then GFAP immunoreactivity in astrocytes from aging group was higher than astrocytes from younger retinas.⁵⁰ Palko et al.⁵¹ reported that human donor maculae from patients with wet-AMD also exhibit the typical endfeet localization of hypercitrullinated GFAP. Actually, GTAP was verified up-regulated in AMD group in our research as well.

SCD, a protein located within the endoplasmic reticulum, functions as an essential enzyme that regulates the production of monounsaturated fatty acids. An elevated expression level of SCD is strongly indicative of cellular proliferation.⁵² SCD inhibitors may exhibit anti-inflammatory properties and offer protective effects in AMD, albeit with the underlying mechanisms still largely unknown. For example, sterculic acid, a widely reported SCD inhibitor, has been demonstrated to effectively counteract the inflammatory and cytotoxic responses induced by 7-Ketocholesterol in both in vivo and in vitro models of choroidal neovascularization.⁵³ Similar result was revealed in our study that SCD up-regulated in AMD group. These three biomarkers testified our technique. Furthermore, we filtered other two biomarkers that have not been demonstrated by previous studies to explore the relationships between AMD.

Glutathione peroxidase 8 (GPX8) serves as a pivotal regulator of redox homeostasis. The absence of GPX8 led to oxidative stress in oral cancer cells.⁵⁴ No articles have been published that reveal the association between GPX8 and AMD. Our study underscores the potential significance of GPX8 in the progression of AMD and suggests a novel research direction for AMD.

Branched chain keto acid dehydrogenase E1, beta polypeptide (BCKDHB) is a gene that codes for the subunits of a multimeric mitochondrial complex responsible for decarboxylating α-keto acid derivatives of the branched-chain amino acids.⁵⁵ Mutations in the BCKDHB can lead to the development of maple syrup urine disease, an autosomal recessive metabolic disorder.⁵⁶ Recently, some studies observed that BCKDHB was identified as lipid metabolism-related genes and decreased in diabetic retinas.⁵⁷ Similar to GPX8, this gene has not been verified as AMD biomarker before. According to our MR study, we suggested that BCKDHB gene played a major risk in AMD patients, which increases before a person disgnosed as AMD. Through the horizontal pleiotropy test and the heterogeneity test, BCKDHB as a novel biomarker is reliable. Maybe we provide a new direction in AMD researches. Meanwhile, in our biomarker-targeted drug screening suggested that colchicine had the potential to target both GFAP and SCD. Colchicine is an anti-inflammatory alkaloid and primarily used for the treatment of gout flares.⁵⁸ Sha et al.⁵⁹ presented genetic evidence using MR technique to suggest a causal association between certain immune-mediated inflammatory conditions and AMD. McGeer and Sibley⁶⁰ first observed that their rheumatoid arthritic (RA) patients, after being treated with hydroxychloroquine, showed a lower likelihood of developing AMD, indicating a potential protective benefit of hydroxychloroquine against AMD. Similar to our result, it is suggested that anti-inflammatory colchicine maybe a potential medication for AMD therapy.

Prior research has already established that the levels of M2 macrophages exhibit a marked increase in tandem with age and the severity of AMD; furthermore, macrophage dysfunction plays an essential role in the development and progression of AMD.^61,62 Other studies have investigated the impact of CSF1 secreted by human HCVECs on macrophage migration and M2 polarization, mediated through the CSF1R/PI3K/AKT/FOXO1 signaling pathway.⁶³ Spearman correlation analysis revealed that one of our biomarkers, GPX8, exhibited the highest positive correlation with M2 macrophages, this provides an important foundation for future immunotherapy research. As for eosinophils, according to our results, a strong negative correlation existed between SCD and eosinophils, while SCD was regarded as a risk biomarker. However, some studies observed eosinophils significantly increased in patients with nAMD,⁶⁴ even with a positive correlation with severity,⁶⁵ which are opposite of our results. We need to do more experiments to verify these two results.

Our study filtered five biomarkers through machine learning and MR technique, which is a novel approach to use independent large GWAS summary data sets to use for AMD biomarkers associated with oxidative stress. In addition, this study first illustrated a clear causal relationship between AMD and BCKDHB. BCKDHB gene played a major risk in AMD patients, which was verified that it increased before a person was diagnosed with AMD by both the horizontal pleiotropy test and Steiger test. Although BCKDHB did not occurred before in other researches as AMD biomarkers, our results show a possible direction for prediction and therapy. To identify these results, we conducted qRT-PCR, which was consistent with the results of bioinformatics analysis. Certain limitations should be notes in our work. Considering the exploratory nature of this study, we aim to screen out as many potential biomarkers as possible to provide more clues and directions for further research, rather than strictly controlling false positives. Therefore we did not conduct multiple comparison correction. It is particularly important to note that there is a random chance of 5% that a marker has an uncorrected P value of 0.05. In addition, the datasets were of European descent, which limit results are possible lack of the association to other ethnic groups. And we did not divide into geographic atrophy and nAMD, which cannot declare these five biomarkers’ effects, respectively in these two subtypes. Finally, the datasets did not provide detailed information for patients such as age, which may inevitably introduce bias into our results.

Conclusions

In this study, we systematically screened the AMD biomarkers related to oxidative stress. Based on the GSE29801 and GSE135092 datasets in the GEO database, we screened out a total of 16 DE-OSRGs. Then, five biomarkers of AMD were screened through three machine learning methods (LASSO, SVM-RFE and Boruta), including GFAP, SCD, BCKDHB, GPX8 and MSRB2. Moreover, the results of the qRT-PCR experiment indicated that there were significant differences in these 5 biomarkers between the AMD group and the control group. MR analysis indicated that BCKDHB was a risk factor for AMD (OR > 1, P < 0.05). In conclusion, the discovery of these biomarkers not only contributes to a deeper understanding of the pathogenesis of AMD, but also provides potential targets for clinical intervention.