Identification and validation of oxidative stress-related genes in primary open-angle glaucoma by weighted gene co-expression network analysis and machine learning Authors: Lu-Qian Yang, Yue Song, Feng Huang, Wei Wu, Hong-Pei Ji Published: 2024 ABSTRACT Primary open-angle glaucoma (POAG) is a common ocular disease, and there is currently no effective treatment for POAG therapy. Thus, identifying some effective diagnostic markers is beneficial to the treatment of patients. The expression profile was obtained from Gene Expression Omnibus (GEO) database. The functional enrichment was analyzed using Gene Ontology, Kyoto Encyclopedia of Genes and Genomes and gene set enrichment analysis. Co-expressed genes were identified using weighted gene co-expression network analysis (WGCNA). Hub genes were screened through Lasso regression, support vector machine-recursive feature elimination (SVM-RFE) and Random Forest, and receiver operating characteristic curve was used to assess diagnostic value. Immune cell infiltration was calculated using IOBR package. The regulatory network was constructed through STRING, miRactDB and Cytoscape. The oncoPredict package was employed to predict the candidate chemotherapy agents. According to GSE27276 database, 541 differentially expressed genes were identified. Five oxidative stress-related genes with high area under the curve value, namely HBB, MAOA, ACOX2, ALDH7A1 and TYMP, were determined using WGCNA and machine learning. Infiltration level of NK cells, CD4 T cells and dendritic cells were significantly increased in POAG group compared with normal group, while CD8 T cells and Tregs cells were significantly decreased. HBB was closely related to most immune cells. Hub genes were all targeted by 16 miRNAs. Drug sensitivity analysis exhibited that some drugs were more sensitivity for POAG, such as Acetalax_1804, Ibrutinib_1799 and OSI_027_1594. We identified 5 oxidative stress-related genes with high diagnostic value for POAG. 1. Introduction Primary open-angle glaucoma (POAG), a prevalent global etiology of visual impairment, is characterized by the gradual decline of retinal ganglion cells and the irreversible damage inflicted upon the optic nerve.[1] This stealthy form of glaucoma remains asymptomatic during incubation period.[2] Presently, it is widely acknowledged that elevated intraocular pressure (IOP) constitutes the primary and most prevalent risk factor for the onset of POAG,[3] and it may also precipitate oxidative stress.[4] Additionally, topical ocular medications aimed at glaucoma represent the predominant therapeutic strategy for attenuating IOP, albeit with attendant adverse effects.[5] Emerging research has identified certain mRNAs as playing critical roles in the genesis of glaucoma, including CPAMD8, ASB10, and WDR36.[6–8] Furthermore, advancements in bioinformatics and molecular biology techniques have led to the identification of an increasing number of biomarkers for the early diagnosis of the condition, with remarkable efficacy. Consequently, there is an urgent imperative to delve deeper into the molecular mechanisms underpinning glaucoma pathogenesis, in order to devise early diagnostic and therapeutic targets. Oxidative stress, an environment engendered by a disequilibrium between oxidants and antioxidants within the organism, culminates in the excessive accumulation of deleterious free radicals and superoxide.[9–11] Moreover, the aberrant or excessive production of these entities wreaks cellular havoc and incites inflammation.[12,13] Preceding studies have established that oxidative stress plays a pivotal role in numerous diseases, encompassing cancers, cardiovascular ailments, and diabetes.[14–16] Additionally, recent investigations have revealed incipient indications of oxidative stress within the ocular tissues of individuals afflicted by POAG. Investigators have astutely discerned a diminution in antioxidant enzyme activity and an augmentation in oxidative stress markers within POAG.[17,18] With further investigation, the potential mechanisms of oxidative stress in glaucoma have been preliminarily explored. Primarily, oxidative stress triggers an excessive production of ROS, leading to the demise of retinal ganglion cells.[19] Furthermore, trabecular meshwork cells, the most sensitive cells in the eye to ROS, function as regulators of aqueous humor outflow.[20] However, studies have shown that excessive ROS levels lead to damage in the trabecular meshwork, increasing resistance to aqueous humor outflow and subsequently causing a rise in IOP.[21] Importantly, the elevation in IOP also induces oxidative stress, further triggering the development of glaucoma.[22] Nevertheless, the potential mechanisms by which oxidative stress-related genes contribute to the development of POAG warrant further exploration. In light of this, we harnessed bioinformatics techniques to analyze gene expression profiles obtained from the Gene Expression Omnibus (GEO) database, with the objective of identifying central genes of diagnostic significance and further elucidating the role of oxidative stress-related genes in the progression of POAG. 2. Methods 2.1. Data collection In this study, the gene expression profile (GSE27276) was obtained from the GEO database (http://www.ncbi.nlm.nih.gov/geo), including 19 normal samples and 19 POAG patients. Besides, 1483 oxidative stress-related genes with relevance score > 7 were obtained from GeneCards (https://www.genecards.org/).[23] 2.2. The analysis of differentially expressed genes The differentially expressed genes (DEGs) from GSE27276 were identified using the limma R package after the raw data were log-transformed. The |fold change| > 1.5 and P < .05 were used as the screening criteria for DEGs. Moreover, the DEGs were visualized using ggvolcano R packages. 2.3. Gene Ontology and Kyoto Encyclopedia of Genes and Genomes analysis To explore the potential biological functions and mechanisms of genes, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed using the “clusterProfiler” R package. GO analysis includes biological process, cellular component and molecular functions. The terms of GO and KEGG with P < .05 were considered statistically significant. 2.4. Weighted gene co-expression network analysis The weighted gene co-expression network analysis R package was used to identify the gene modules highly related to POAG. At first, the outliers and abnormal samples were removed. Secondly, a scale-free network was constructed after the soft power was determined. Then, the matrix was converted to a topological overlap matrix after the co-expression similarity and adjacency were calculated. Next, cluster analysis is carried out to identify gene modules, and each module has at least 50 genes. Finally, the characteristic genes were determined, modules were clustered, and similar modules were merged (abline = 0.25). 2.5. Machine learning Lasso, Random Forest, and SVM-REF are widely applied in the selection of key genes in genomic studies. Lasso utilizes sparse regularization to identify crucial genes that have a significant impact on the target variable. Random Forest constructs multiple decision trees and integrates their predictions to evaluate the importance of genes and select key ones. SVM-RFR, based on support vector machine regression, handles nonlinear relationships and contributes to key gene selection. In this study, glmnet, randomForest and e1071 R packages were respectively used for Lasso, Random Forest and SVM-REF analysis. 2.6. Gene set enrichment analysis Gene set enrichment analysis (GSEA) was applied to determine the potential function of hub genes. The reference gene set was acquired from the Molecular Signature Database (MSigDB). 2.7. Immune infiltration and correlation analysis Immune infiltration analysis was performed using the quantiseq algorithm of the IOBR R package. Besides, the psych R package was used to calculate the Pearson correlation between infiltration levels of immune cells and hub genes. The correlation results were visualized using the ggplot2 R package. 2.8. The construction of interaction network The protein–protein interaction network was constructed in Search Tools for the Retrieval of Interacting Genes (STRING) based on hub genes. Besides, the miRNA targeted hub genes were predicted in miRactDB (https://ccsm.uth.edu/miRactDB/index.html). Then, the regulatory network was visualized using Cytoscape. 2.9. Drug sensitivity The training data of the GDSC2 database was first downloaded using the oncoPredict R package. Then, after the expression profile of GSE27276 was imported in R software and set as test data, the half-maximal inhibitory concentration (IC50) of all drugs in POAG and normal samples could be calculated using the oncoPredict package. 2.10. Statistics analysis All statistical analyses in this study were performed in R software (version 4.3.1). Then independent-sample t-test was used to calculate the differences between the 2 groups. The P < .05 was considered as statistically significant difference. 3. Results 3.1. The identification of DEGs At first, the DEGs analysis was performed based on the expression profile of GSE27276. A total of 541 DEGs were obtained, of which 257 DEGs were upregulated and 284 were downregulated (Fig. 1). The volcano plot of DEGs. FC = foldchange. 3.2. Enrichment function analysis of DEGs Then, to preliminary elucidate the potential mechanisms involved in the pathogenesis of POAG, a functional enrichment analysis of DEGs was performed. The function analysis exhibited that DEGs were involved in cell adhesion molecules, antigen processing, ECM-receptor interaction, IL-17 signaling pathway and Th17 cell differentiation (Fig. 2A). Besides, based on the GO analysis, these genes were enriched in response to chemical, regulation of response to stimulus, system development, transport and establishment of localization (Fig. 2B). In term of molecular functions, the results showed that DEGs were mainly played roles in structural molecule activity, signaling receptor binding, identical protein binding, protein-containing complex binding and molecular function regulator (Fig. 2C). In addition, in the aspect of cellular component, DEGs mainly existed in the extracellular region, vesicle, endomembrane system, cytosol and organelle membrane (Fig. 2D). The enrichment analysis by GO and KEGG. (A) KEGG analysis of DEGs (B–D) GO analysis of DEGs. (A) KEGG analysis. (B) Biological process (BP) terms of GO analysis. (C) Molecular function (MF) terms of GO analysis. (D) cellular component (CC) terms of GO analysis. 3.3. Construction of co-expression networks At first, the optimal soft-threshold power of 7 was determined based on the scale-free network construction (Fig. 3A and B). Then, 19 modules were determined by WGCNA analysis (Fig. 3C and D). Moreover, the correlation analysis showed that the pink module was strongly and positively correlated with POAG (cor = 0.86, P < .001) (Fig. 3E). Therefore, 379 genes in pink modules were used for the next analysis. Finally, a number of 15 intersecting genes were obtained by the Venn diagram (Fig. 3F). WGCNA co-expression network of POAG. (A and B) The optimal soft-threshold power. (C) Dendrogram based on a dissimilarity metric for DEGs. (D) The heatmap of correlation between WGCNA modules. (E) The correlation between WGCNA modules and POAG. (F) Venn diagrams of DEGs, pink modules and oxidative stress-related genes. DEGs = differentially expressed genes, POAG = Primary open-angle glaucoma, WGCNA = weighted gene co-expression network analysis. 3.4. Screening hub genes by 3 machine learning methods Next, SVM-RFE, LASSO regression and RF machine learning algorithms were used to screen hub genes with diagnostic value. Firstly, the RF algorithm ranked the importance of genes, among which 9 genes with importance >0 were used for subsequent analysis, and MAOA was the most important (Fig. 4A). Besides, 15 genes were identified by the SVM-RFE algorithm (Fig. 4B and C). Additionally, 7 genes were extracted using the LASSO regression (Fig. 4D and E). Subsequently, the results of 3 algorithms were overlapped by the Venn diagram and finally 5 hub genes were obtained, namely HBB, MAOA, ACOX2, ALDH7A1, and TYMP (Fig. 4F). Identification of the hub genes by machine learning methods. (A) The importance of key genes was calculated using RF algorithms. (B and C) The number of hub genes was determined using SVM-RFE algorithms. (D and E) The hub genes were screened out using Lasso regression algorithms. (F) Venn diagrams for 3 algorithms. SVM-RFE = support vector machine-recursive feature elimination. 3.5. The expression and diagnostic value of hub genes As shown in Figure 5A to E, it was obvious that ALDH7A1, HBB, and ACOX2 were significantly upregulated in the POAG group compared with the normal group, while TYMP and MAOA were downregulated. Besides, the receiver operating characteristic (ROC) curves exhibited that all hub genes possessed a high diagnostic value for POAG (Fig. 5F). Expression of hub genes and ROC curve analysis. (A) The expression of ALDH7A1 in normal and POAG groups. (B) The expression of HBB in normal and POAG groups. (C) The expression of TYMP in normal and POAG groups. (D) The expression of MAOA in normal and POAG groups. (E) The expression of ACOX2 in normal and POAG groups. (F) ROC curve showed the diagnostic value evaluation of 5 hub genes. ** P < .01, *** P < .001. ROC = receiver operating characteristic, POAG = Primary open-angle glaucoma. 3.6. The GSEA of hub genes Then, GSEA was applied to determine the potential function of hub genes in the occurrence of POAG. The results exhibited that TYMP and ALDH7A1 were both enriched in the adherens junction (Fig. 6A and B). HBB and MAOA were both involved in the WNT signaling pathway, GNRH signaling pathway and calcium signaling pathway (Fig. 6C and D). Besides, ACOX2 was enriched in histidine metabolism, p53 signaling pathway and glycerophospholipid metabolism (Fig. 6E). GSEA results for hub genes in patients with POAG. (A) GSEA results of TYMP. (B) GSEA results of ALDH7A1. (C) GSEA results of HBB. (D) GSEA results of MAOA. (E) GSEA results of ACOX2. GSEA = gene set enrichment analysis, POAG = Primary open-angle glaucoma. 3.7. The immune infiltration analysis According to the KEGG and GSEA results, we found that DEGs and hub genes were both closely related to the immune response, thus, immune infiltration analysis were performed. As shown in Figure 7, the infiltration level of NK cells, CD4 T cells and dendritic cells (DCs) were significantly increased in POAG group compared with normal group, while CD8 T cells and Tregs cells were significantly decreased in the POAG group. The immune infiltration level in normal and POAG group. The blue indicated normal group, and red indicated POAG group. ns, not significant, * P < .05, ** P < .01. POAG = Primary open-angle glaucoma. 3.8. The correlation between hub genes and immune cells To explore the function of hub genes in immune response, the correlation analysis was performed. It was obvious that ACOX2 was closely related to the infiltration level of B cells, while ALDH7A1 was negatively related to B cells (Fig. 8A and B). MAOA was positively and negatively related to B cells and DC cells, respectively (Fig. 8C). On the contrary, HBB was positively related to CD4 T cells and DC cells, but negatively related to M2 macrophages, CD8T cells and Tregs (Fig. 8D). Besides, TYMP was closely related to Tregs and B cells (Fig. 8E). The immune infiltration level was related to hub genes. (A) The relationship between ACOX2 and immune infiltration levels. (B) The relationship between ALDH7A1 and immune infiltration levels. (C) The relationship between MAOA and immune infiltration levels. (D) The relationship between HBB and immune infiltration levels. (E) The relationship between TYMP and immune infiltration levels. 3.9. The construction of miRNA-mRNA network To further determine the potential mechanism of hub genes, we constructed an interaction network. At first, the STRING network was used to construct a protein–protein interaction interaction network based on hub genes (Fig. 9A). Then, the miRactDB database was used to predict the miRNA targeted to hub genes, and the overlapping gene was screened out through Venn diagram (Fig. 9B). Finally, the miRNA-mRNA network was constructed using Cytoscape (Fig. 9C). The construction of miRNA-mRNA network. (A) The protein–protein interaction network of hub genes. (B) Venn diagram of miRNAs targeted hub genes. (C) The miRNA-mRNA network constructed by hub genes and common miRNAs. The orange octagon indicated hub genes; the blue triangle indicated genes related to hub genes; the green rectangle indicated common miRNAs. 3.10. The drug sensitivity analysis Finally, the drug sensitivity analysis was performed and the top 9 drugs were shown in Figure 10A to I. It indicated that these drugs could be candidate drugs for POAG, and it will aid in the selection of the most suitable drugs for clinical practice. Drug sensitivity in normal and POAG groups. (A) The sensitivity of Acetalax_1804 in normal and POAG groups. (B) The sensitivity of Ibrutinib_1799 in normal and POAG groups. (C) The sensitivity of OSI_027_1594 in normal and POAG groups. (D) The sensitivity of EPZ5676_1563 in normal and POAG groups. (E) The sensitivity of AZD5991_1720 in normal and POAG groups. (F) The sensitivity of Leflunomide_1578 in normal and POAG groups. (G) The sensitivity of Sapitinib_1549. (H) The sensitivity of TAF1_5496_1732 in normal and POAG groups. (I) The sensitivity of IRAK4_4710_1716 in normal and POAG groups. POAG = Primary open-angle glaucoma. 4. Discussion Glaucoma, a ubiquitous ophthalmic affliction, remains recalcitrant to efficacious therapeutic interventions.[24] The etiology of glaucoma is frequently ascribed to genetic predispositions, with an expanding cohort of genes implicated in the genesis of glaucoma being unveiled in recent years.[25] Furthermore, extant studies posited that oxidative stress may exert a substantial influence on the pathogenesis of glaucoma.[26–28] Consequently, our objective was to discern genes associated with oxidative stress that contribute to the development of glaucoma. In the present study, a total of 541 DEGs (257 upregulated and 284 downregulated) were identified, predicated upon the GSE27276 dataset. Enrichment analysis revealed that these DEGs were intimately associated with immune-related signaling pathways. As previous investigations have elucidated, alterations in immune function are likewise implicated in oxidative stress.[29–31] This suggested that the immune response may modulate the progression of POAG through the agency of oxidative stress-related genes. Subsequently, 5 hub genes, namely HBB, MAOA, ACOX2, ALDH7A1, and TYMP, were identified through WGCNA and machine learning. Additionally, the ROC curve demonstrated that the area under the curve values of these hub genes surpassed 0.75, thereby indicating their high diagnostic potential. A plethora of evidence underscored the pivotal role of aberrant energy metabolism and immune responses in the genesis of diseases. Given that immune system cells are particularly vulnerable to oxidation, oxidative stress may impinge upon immune responses.[32] In our study, both KEGG and GSEA enrichment analyses revealed that oxidative stress-related hub genes may contribute to the development of POAG by mediating metabolism and immune responses. ACOX2 primarily participates in the degradation of branched fatty acids.[33] Due to the competition for nutrients between cancer cells and immune cells, metabolic alterations may impact the functions of immune cells such as CD8 T cells.[33] However, it has been observed that ACOX2 and the infiltration levels of CD8 T cells exhibited an inverse correlation in lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC).[34,35] This suggests that ACOX2 may function as a double-edged sword within the immune system. Furthermore, ALDH7A1, an aldehyde dehydrogenase, plays a significant role in energy metabolism and oxidative stress.[36] Previous studies have identified mutations in ALDH7A1 as being implicated in the development of osteoporosis and various cancers, including colorectal and oral cancers.[37–39] Additionally, Stoll et al discovered that the upregulation of ALDH7A1 may inhibit the production of immune effectors in nonsmall cell lung carcinoma, thereby impacting the immune microenvironment and inflammation.[36] However, we found that ACOX2 and ALDH7A1 were highly correlated with B cells, which exhibited no significant differences between the normal and POAG groups. GSEA enrichment revealed that both genes were upregulated and negatively correlated with metabolism-related signaling pathways. This suggested that ACOX2 and ALDH7A1 may not contribute to the development of POAG by regulating immune responses but rather by influencing energy metabolism. Experimental evidence has demonstrated that MAOA could target T and macrophage cells to modulate immune responses and immunotherapy in tumors.[40,41] However, our results revealed that MAOA was only negatively correlated with DCs, which can activate immune responses by presenting antigens to T cells.[42] Additionally, previous studies have shown that TYMP can suppress immune responses by promoting the secretion of IL-10, and IL-10 can inhibit CD8 T cells through Tregs.[43,44] We identified that TYMP was positively correlated with Tregs in POAG, which could secrete IL-10 to inhibit CD8 T cells.[45] This indicated that the downregulated MAOA may promote the infiltration of DCs, while TYMP may inhibit the infiltration of Tregs. However, DCs were increased and Tregs were decreased in POAG, but CD8 T cells were decreased. This implies that the infiltration levels of CD8 T cells are primarily affected by other factors. Luo et al posited that HBB may play different roles in different tumors based on its disparate prognostic effects in various tumors.[46] In our study, the upregulation of HBB was positively correlated with CD4 T cells and DCs but negatively correlated with CD8 T cells and Tregs in POAG. This indicated that HBB may play a major role in the infiltration of CD8 T cells in POAG. Furthermore, the ROC curve demonstrated that HBB exhibited the highest area under the curve value in the diagnosis of POAG. Therefore, we inferred that HBB and CD8 T cells may be essential for the occurrence and development of POAG. In conclusion, we identified 5 hub genes with significant diagnostic value, and these genes may contribute to the development of POAG through metabolism and immune responses. This may provide foundational theoretical insights for the diagnosis and treatment of POAG patients. However, several limitations are inherent in this study. Firstly, the sample size of our data is small, necessitating validation in a cohort with a larger sample size. Secondly, these analyses were predicated solely on bioinformatics. Thus, the underlying mechanisms required elucidation through in vivo and in vitro experiments. Author contributions Conceptualization: Lu-Qian Yang. Data curation: Lu-Qian Yang, Yue Song, Feng Huang, Wei Wu. Formal analysis: Lu-Qian Yang. Funding acquisition: Lu-Qian Yang. Investigation: Lu-Qian Yang. Methodology: Lu-Qian Yang. Project administration: Lu-Qian Yang. Resources: Lu-Qian Yang. Software: Lu-Qian Yang. Supervision: Lu-Qian Yang, Hong-Pei Ji. Validation: Lu-Qian Yang. Visualization: Lu-Qian Yang. Writing – original draft: Lu-Qian Yang.