The Use of Identified Hypoxia-related Genes to Generate Models for Predicting the Prognosis of Cerebral Ischemia‒reperfusion Injury and Developing Treatment Strategies

Abstract

Cerebral ischemia‒reperfusion injury (CIRI) is a type of secondary brain damage caused by reperfusion after ischemic stroke due to vascular obstruction. In this study, a CIRI diagnostic model was established by identifying hypoxia-related differentially expressed genes (HRDEGs) in patients with CIRI. The ischemia‒reperfusion injury (IRI)-related datasets were downloaded from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo), and hypoxia-related genes in the Gene Cards database were identified. After the datasets were combined, hypoxia-related differentially expressed genes (HRDEGs) expressed in CIRI patients were identified. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses of the HRDEGs were performed using online tools. Gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA) were performed with the combined gene dataset. CIRI diagnostic models based on HRDEGs were constructed via least absolute shrinkage and selection operator (LASSO) regression analysis and a support vector machine (SVM) algorithm. The efficacy of the 9 identified hub genes for CIRI diagnosis was evaluated via mRNA‒microRNA (miRNA) interaction, mRNA–RNA-binding protein (RBP) network interaction, immune cell infiltration, and receiver operating characteristic (ROC) curve analyses. We then performed logistic regression analysis and constructed logistic regression models based on the expression of the 9 HRDEGs. We next established a nomogram and calibrated the prediction data. Finally, the clinical utility of the constructed logistic regression model was evaluated via decision curve analysis (DCA). This study revealed 9 critical genes with high diagnostic value, offering new insights into the diagnosis and selection of therapeutic targets for patients with CIRI. : Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12035-024-04433-9.

Introduction

Stroke is one of the leading causes of human mortality, ranking second only to ischemic heart disease [1]. Stroke, which has a high incidence rate, is associated with mortality and disability, imposing a profound burden on patients, families, and society [2]. Eighty-seven percent of all strokes involve ischemia, and the primary clinical objective for patients is to restore cerebral blood flow to the greatest extent possible [3]. Common treatment methods include recombinant tissue plasminogen activator (rt-pa)-induced thrombolysis under certain conditions and interventions to retrieve emboli and restore the blood supply through reperfusion and thus salvage ischemic brain tissue. However, reperfusion can lead to tissue damage, causing cerebral ischemia‒reperfusion injury (CIRI) [3]. Investigating the mechanism underlying reperfusion injury and exploring methods to attenuate reperfusion injury are challenging and crucial topics in ischemic disease research.

CIRI involves complex pathophysiological mechanisms, including the disruption of energy metabolism, cell acidosis, exponential increases in excitotoxic amino acid synthesis or release, disruption of intracellular calcium homeostasis, free radical production, and apoptotic gene expression [4]. The interplay among these events establishes a complex regulatory network and subsequently triggers a series of pathological response cascades. Ultimately, these outcomes lead to neural cell death, including apoptotic death, and loss of blood‒brain barrier integrity, resulting in brain edema and neurological dysfunction.

Antioxidants that modulate the effects of free radicals to attenuate inflammation and inhibit neuronal apoptosis can significantly reduce the damage caused by CIRI [5–7]. However, the underlying mechanisms mediating these responses are not yet fully understood. Currently, standardized or effective treatments are unavailable for CIRI, and many underlying mechanisms and therapeutic targets for CIRI remain unidentified.

Hypoxia is a condition characterized by low oxygen levels in the body, and it profoundly affects various physiological cellular processes and contributes to several pathological conditions, including cancer, chronic inflammation, myocardial infarction, stroke, and ischemic acute kidney injury [8]. Hypoxia induces changes in the expression of hypoxia-responsive genes through the activation of transcription factors and inhibition of O₂-dependent histone demethylases, affecting the expression of hypoxia-responsive genes [9].

With the advancement of high-throughput sequencing technologies, the study of hypoxia-related genes (HRGs) has ushered in an era characterized by transcriptome analysis. The abundant information, including the identities of HRGs, obtained through these analyses can be used for exploring disease mechanisms and therapeutic targets [10].

In this study, we identified differentially expressed HRGs associated with CIRI to further investigate the mechanisms and therapeutic targets related to CIRI. Additionally, GO and KEGG enrichment analyses, gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA) were performed to identify the functions and pathways associated with these differentially expressed genes (DEGs). Moreover, the impact of gene expression levels on CIRI outcomes was examined. Least absolute shrinkage and selection operator (LASSO) regression analysis was performed with a support vector machine (SVM) algorithm to construct CIRI diagnostic models based on the identified HRGs, and 9 pivotal genes were identified. Logistic regression models were constructed for each gene, the diagnostic power of the models was evaluated via a nomogram analysis, and the clinical utility of each model was assessed via decision curve analysis (DCA). mRNA‒microRNA (miRNA), mRNA–transcription factor (TF), and mRNA–RNA-binding protein (RBP) interaction networks were constructed, infiltrating immune cells were identified with information from the CIRI dataset, and correlation analyses with these data were performed (Fig. 1). This research was performed to identify differentially expressed HRGs in patients with CIRI and provide a foundation for investigating the mechanisms underlying cerebral injury caused by ischemia‒reperfusion, with the ultimate goal of identifying clinical diagnostic and therapeutic targets.

Fig. 1.

Schematic showing the research plan

Materials and Methods

Downloaded Data

We utilized the GEOquery [11] package to download the ischemia–reperfusion injury (IRI)-related datasets GSE97537 [12], GSE82146 [13], GSE61616[14],GSE211253, GSE163614[15] and GSE201258[16]. The GSE97537 dataset is based on Rattus norvegicus samples, and we downloaded data for 12 samples, including IRI samples and normal samples. The data platform used was the GPL1355Rat230_2 Affymetrix Rat Genome 230 2.0 Array, and we included the expression profile data for 7 CIRI samples (comprising the IRI group) and 5 normal samples (comprising the control group) in this study. The GSE82146 dataset is also based on Rattus norvegicus samples, and we downloaded data from 27 samples, including CIRI samples and normal samples. The data platform used was Affymetrix Rat Gene 2.0 ST Array transcript (gene) version GPL17117RaGene-2_0-st, and we included the expression profile data for 15 CIRI samples (comprising the IRI group) and 12 normal samples (comprising the control group) in this study.

The datasets GSE97537 and GSE82146 were combined and used as the test set. The dataset GSE61616, includes information on a total of 15 brain ischemia–reperfusion injury samples and normal samples derived from Rattus norvegicus. The data platform file is GPL1355 Rat230_2 Affymetrix Rat Genome 230 2.0 Array. For this study, we included the expression profile data for 5 brain ischemia–reperfusion injury samples (group: IRI) and 5 normal samples (group: Control). The dataset GSE211253 includes information on a total of 12 brain ischemia–reperfusion injury samples and normal samples from Rattus norvegicus. The data platform file is GPL25947 Illumina NovaSeq 6000. We incorporated the expression profile data from 4 brain ischemia–reperfusion injury samples (group: IRI) and 3 normal samples (group: Control) into this study. The dataset GSE163614 includes information on 6 samples in total, including brain ischemia–reperfusion injury samples and normal samples from Rattus norvegicus. The data platform file is GPL25947 Illumina NovaSeq 6000. For this study, we included the expression profile data from 3 brain ischemia–reperfusion injury samples (group: IRI) and 3 normal samples (group: Control). The dataset GSE20125 includes information on 9 samples in total, including brain ischemia–reperfusion injury samples and normal samples from Rattus norvegicus. The data platform file is GPL25916 Arraystar Rat Epitranscriptomic microarray. We included the expression profile data from 3 brain ischemia–reperfusion injury samples (group: IRI) and 3 normal samples (group: Control) into this study. Table 1 presents information on these datasets. The datasets GSE61616, GSE211253, GSE163614, and GSE201258 were combined and used as the validation set.

Table 1.

List of information for the ischemia‒reperfusion injury (IRI) datasets

	GSE97537	GSE82146	GSE61616	GSE211253	GSE163614	GSE201258
Platform	GPL1355	GPL17117	GPL1355	GPL25947	GPL25947	GPL25916
Species	Rattus norvegicus	Rattus norvegicus	Rattus norvegicus	Rattus norvegicus	Rattus norvegicus	Rattus norvegicus
Samples in Control group	5	12	5	3	3	3
Samples in IRI group	7	15	5	4	3	3
Reference	[12]	[13]	[14]	/	[15]	[16]

IRI: ischemia‒reperfusion injury

We also retrieved data on HRGs from the GeneCards database (https://www.genecards.org/) [17] Specifically, the database included comprehensive human genetic information. Using "hypoxia" as the search term in the GeneCards database, only "protein-coding" and master regulator genes (MRGs) with a correlation score higher than 5 were retained, and a total of 123 HRGs were thus identified. Moreover, 228 HRGs were identified from references[18–20], and 307 HRGs were ultimately retained after the data were merged and duplicates were removed. The details are presented in Table S1.

Differentially Expressed Genes Associated With CIRI

We first merged the IRI datasets GSE97537 and GSE82146 and then normalized the merged dataset with the R package sva to eliminate batch effects and determine the underlying mechanisms mediated by different genes involved in IRI responses and related biological features and pathways. Then, the probability ratios were determined using the limma package [21] to generate the combined CIRI GEO dataset (hereafter, the GEO dataset). Differential expression analysis of the samples in the GEO dataset was performed to identify DEGs among the control and IRI subgroups. Genes with a log fold change (FC) > 0 and P value < 0.05 were chosen for further investigation. The DEGs with a logFC > 0 and P value < 0.05 were upregulated, while those with logFC < 0 and P value < 0.05 were downregulated.

We compared the DEGs in the GEO dataset with identified HRGs via an intersection analysis to identify hypoxia-related differentially expressed genes (HRDEGs) associated with IRI. The genes identified via this intersection analysis were visualized in a Venn diagram. These results from the differential expression analysis were also visualized using the R package ggplot2, which was used to create volcano plots, and pheatmap, which was used to generate heatmaps showing HRDEG expression levels.

DEG functional enrichment analysis (Gene Ontology, GO) and pathway enrichment analysis (Kyoto Encyclopedia of Genes and Genomes, KEGG) [22] are classification systems commonly used for large-scale functional enrichment studies. The GO annotations are based on biological process (BP), molecular function (MF), and cellular component (CC) terms. KEGG analysis [23], on the other hand, is a widely used database that provides information on genomes, biological pathways, diseases, and medications. In our study, we utilized the R package clusterProfiler [24] to perform the GO annotation analysis of HRDEGs. The inclusion criteria for this analysis were a P value < 0.05 and a false discovery rate (FDR) value (q value) < 0.05. We applied the Benjamini‒Hochberg (BH) method to correct the P value. The BH method is a technique used to control the false discovery rate (FDR) in multiple hypothesis testing. This method, introduced by Yoav Benjamini and Yosef Hochberg in 1995, involves ranking the p values from individual tests and then sequentially comparing these p values to specific thresholds to determine significance levels. By employing the BH method, researchers can better manage false positives resulting from multiple testing, thereby enhancing the reliability of the results.

Gene Set Enrichment Analysis

GSEA [25] was performed to evaluate the distribution pattern of genes of a predefined gene set within a ranked gene table associated with a phenotype to determine the relative impacts of the genes on the phenotype. In this study, we evaluated the enrichment of genes in the predefined gene set using the logFC values for ranking the genes. An enrichment analysis of all the genes associated with a phenotype was then performed using the clusterProfiler package. The following parameters were used for GSEA: 2020 seeds, 1000 calculations, a minimum of 10 genes and a maximum of 500 genes per gene set. The p values were corrected using the Benjamini‒Hochberg (BH) method. The gene set 'c2.cp.v7.2.symbols.gmt' from the Molecular Signatures Database (MSigDB) was used for the GSEA of genes in the GEO dataset. A p value of 0.05 and an FDR value (q value) of 0.25 were used as prescreening criteria to determine significant enrichment.

Gene Set Variation Analysis

GSVA [26] is a nonparametric, unsupervised analysis method to estimate the enrichment of gene pools in transcriptomes, as determined with microarray data by converting the between-sample gene expression matrix into a gene pool expression matrix. GSVA was performed to assess whether samples were enriched in different pathways. We obtained the gene set "m2.all.v2022.1.Mm.symbols.gmt" from the MSigDB database to perform GSVA with the GEO dataset and calculated the differences in functional gene enrichment between the normal (control) and IRI sample groups in the dataset.

Construction of the Hypoxia Diagnostic Model and SVM Analysis

We first performed LASSO based on the parameter family "binomial" and tenfold cross-validation [27] regression on HRDEGs using the glmnet software package [28] to obtain diagnostic models for the HRDEGs in the GEO dataset. We performed 1000 runs to prevent overfitting. LASSO regression is a widely utilized machine learning method primarily employed for constructing diagnostic models. The technique used in this study is based on linear regression with a penalty term (λ × the absolute value of the slope). Regularization was performed to address overfitting during the curve-fitting process. The $over\text{risk score is riskScore}={\sum }_{\text{i}}\text{Coefficient}\left({\text{hub gene}}_{\text{i}}\right)\ast \text{mRNA Expression}(\text{hub}{\text{gene}}_{\text{i}})$.

The LASSO regression diagnostic method was used to determine the penalty coefficient (λ) for each HRDEG. Subsequently, the risk score of the HRDEG diagnostic model was based on the HRDEGs.

Using an SVM algorithm for disease ontology classification [29], we constructed SVM models and screened HRDEGs by considering the number of genes with the highest accuracy and lowest error rate. SVM is a powerful supervised learning model used for classification and regression tasks. It works by finding the optimal hyperplane that separates data points of different classes with the maximum margin. In this study, SVM was employed to classify disease-related genes and identify the most significant genes. We then used LASSO20 to select genes obtained through joint screening of data from the regression and SVM analyses for use in subsequent analysis. The HRDEGs were analyzed and visualized using the FRIENDS plotting tool.

Clinical Evaluation of HRDEGs

We performed logistic regression analysis on the GEO dataset to examine the value of the HRDEGs as a clinical diagnostic tool for CIRI (based on the IRI group). We utilized the R package rms and the results of the logistic regression analysis to construct a nomogram [30], a graph based on the Cartesian coordinate system that describes the functional relationship between multiple independent variables that are visualized as a series of disjointed lines on the plane. Specific scales are assigned to each variable in a multivariate regression model, and then the total score, which enables the prediction of the probability of an event occurring, is calculated to perform this multivariate regression analysis.

A calibration graph is used to evaluate the accuracy of a prediction obtained with a model by comparing the prediction to the real-life results plotted on the graph. Specifically, the fit between the true probabilities and model-predicted probabilities under different situations is primarily evaluated via logistic regression analysis.

DCA is a widely used technique for evaluating the clinical value of prediction models, diagnostic tests, and molecular markers [31]. Ultimately, we utilized the R package ggDCA to generate DCA diagrams, which helped us assess the precision and discriminatory ability of the logistic regression models.

Analysis of Immune Cell Infiltration (CIBERSORT)

CIBERSORT (https://cibersort.stanford.edu/) [32] is a deconvolution tool available both as an R package and a web-based application used for analyzing human immune cell subtype expression matrices. It is based on a linear support vector regression method and is used to estimate immune cell infiltration in sequenced samples by utilizing a panel of gene expression signatures from 22 known immune cell subtypes. In this study, we utilized the CIBERSORT method to assess the status of immune cell infiltration in the GEO dataset. Initially, we compared the infiltration of 22 types of immune cells between the IRI group and the control group. Subsequently, we generated group comparison plots to visualize the correlation between IRI samples and the percentage of infiltrating immune cells and constructed a heatmap to show the expression of the DEGs. Only the expression levels of genes in immune cells that constituted a statistically significant population (p < 0.05) were included in the heatmap.

MRNA–miRNA and MRNA–RBP Interaction Networks

miRDB is a database [33] that is utilized for miRNA target gene prediction and functional annotation. In our study, we used miRDB to predict miRNAs that interact with hub genes (mRNAs). We then constructed an mRNA‒miRNA interaction network based on the prediction results.

The ENCORI database version 3.0 [34] (https://starbase.sysu.edu.cn/) is also known as the starBase database. It contains information on miRNA–noncoding RNA (ncRNA), miRNA‒mRNA, ncRNA–RNA, RNA‒RNA, RBP–ncRNA, and RBP–mRNA interactions. These interactions were identified by mining crosslinking immunoprecipitation sequencing (CLIP-seq) and degradome sequencing data. The ENCORI database includes various visual interfaces that can be used to explore miRNA targets. In addition to identifying miRNA targets, the ENCORI database was used to predict interactions between RBPs and hub genes (mRNAs). We then created an mRNA–RBP interaction network based on the results from analyses of these databases.

Receiver Operating Characteristic (ROC) Curve

The ROC curve is a graphical analysis tool commonly used for selecting an optimal model, eliminating the second-best model, or determining the optimal threshold value within one model. The area under the ROC curve (AUC) established with clinical data typically ranges between 0.5 and 1, with a higher AUC indicating a better diagnosis. The pROC package enables analyses at different levels of precision: results obtained with low precision are between 0.5 and 0.7, those obtained with increased precision are between 0.7 and 0.9, and those obtained with high precision are greater than 0.9. In our study, we utilized the pROC package to plot the ROC curves of the hub genes for various subgroups (in the control/IRI groups) in the GEO dataset. We also calculated the AUCs to determine the diagnostic relevance of hub gene expression with respect to CIRI.

Statistical Analysis

All data processing and analyses were performed using R software version 4.2.0. For comparisons of continuous variables between two groups, we performed independent Student’s t tests to estimate the statistical significance of differences between normally distributed variables, while we performed Mann‒Whitney U tests (i.e., Wilcoxon rank sum tests) to analyze the significance of differences between nonnormally distributed variables. We performed a chi-square test or Fisher’s exact test to determine the statistical significance of differences between the two groups of categorical variables. Spearman’s correlation analysis was performed to calculate the correlation coefficients of the expression of molecules being compared. All the statistical P values are two-sided unless stated otherwise. A significance level of 0.05 was used to determine statistical significance.

Results

IRI: ischemia‒reperfusion injury. HRDEGs: hypoxia-related differentially expressed genes. DEGs: differentially expressed genes. GSEA: gene set enrichment analysis. ssGSEA: single-sample gene set enrichment analysis. GSVA: gene set variation analysis. ssGSEA: single-sample gene set enrichment analysis. GO: Gene Ontology. KEGG: Kyoto Encyclopedia of Genes and Genomes. LASSO: least absolute shrinkage and selection operator. PCA: principal component analysis. ROC curve: receiver operating characteristic curve.

Data Aggregation and Correction

The IRI group (from the GEO dataset) was obtained by removing the batch effect of the combination using the R package sva. The expression data in the combined GEO dataset were then normalized using the limma package for the two IRI datasets obtained from GSE97537 and GSE82146 (Fig. 2a and b). The results indicated that after batch effect removal, the GEO dataset included information from 22 rats with CIRI and 17 rats without CIRI.

Fig. 2.

Dataset debatching

a and b. Boxplots showing the GEO dataset prior to (a) and following (b) normalization. Diagrams showing the results of PCA with the GEO dataset before (c) and following (d) batch effect removal. PCA: principal component analysis.

Moreover, we performed principal component analysis (PCA) on the expression matrix for the GEO dataset before and after the batch effect was removed according to the sources of the samples to verify the effects of removing the batch effect (Fig. 2c and d). After the batch effect was removed, the group effects in the GEO datasets attributable to different sources were substantially eliminated.

Analysis of Differentially Expressed Genes Associated With CIRI

We performed a differential expression analysis with the GEO dataset using the limma package to compare gene expression between the CIRI sample group (IRI group) and the normal sample group (control group). The analysis revealed a total of 4702 genes that met the criteria of a |logFC|> 0 and a P value < 0.05 in the IRI group, while 2569 genes were expressed at a low level (their expression was upregulated in the normal group, logFC < 0). We created volcano plots based on the results obtained from this differential expression analysis (Fig. 3a).

Fig. 3.

Analysis of differentially expressed genes in the brain ischemia‒reperfusion (IRI) injury dataset. a. Volcano plot showing the differentially expressed genes (DEGs) in the GEO dataset. b. Venn diagram showing the DEGs and CSRGs in the GEO dataset. c. Heatmap of DEGs in the GEO dataset. DEGs: differentially expressed genes. HRGs: hypoxia-related genes

We identified the overlapping genes that were differentially expressed with an absolute log FC greater than 0 and a p value less than 0.05 using both the GEO dataset and the identified HRGs to identify HRDEGs. Using the intersection gene set, 60 HRDEGs were identified and plotted in a Venn diagram (Fig. 3b).

We analyzed the differences in gene expression among various sample subgroups in the GEO dataset (Fig. 3c). We utilized the R package pheatmap to visualize the results of the analysis and selected 60 HRDEGs to construct a heatmap. The results showed that these 60 HRDEGs clustered more significantly in different subgroups of the GEO dataset.

GO Functional Enrichment Analysis and KEGG Pathway Enrichment Analysis of HRDEGs

We performed GO (Table 2) and KEGG (Table 3) enrichment analyses to analyze BP, MF, CC terms and biological pathways related to CIRI enriched with 60 HRDEGs. The enrichment of terms or pathways was considered significant when the P value was < 0.05 and the FDR value (q value) was < 0.05. The results of the GO functional enrichment analysis and KEGG pathway enrichment analysis are presented in bubble plots (Fig. 4a and b) and circular network plots (Fig. 4c and d). Additionally, bar graphs display the results of the GO functional enrichment analysis of the HRDEGs based on the logFC values (Fig. 4e and f). The results revealed that 60 HRDEGs were primarily enriched in the BP terms regulation of NIK/NF-κB signaling (GO: 1,901,222), response to hypoxia (GO: 0001666), circadian rhythm (GO: 0007623), and regulation of circadian rhythm (GO: 0042752); CC terms extracellular exosome (GO: 0070062), actin cytoskeleton (GO: 0015629), microbody part (GO: 0044438), and cell leading edge (GO: 0031252); and MF terms actin filament binding (GO: 0051015), lipase inhibitor activity (GO: 0055102), coenzyme binding (GO: 0050662), and phospholipid binding (GO: 0005543), among other terms.

Table 2.

Results of the GO enrichment analysis of HRDEGs

ONTOLOGY	ID	Description	Gene Ratio	Bg Ratio	p value	p adjusted	q value
BP	GO: 0001973	adenosine receptor signaling pathway	5/58	13/17962	3.71e-10	2.39e-07	1.54e-07
BP	GO: 0007189	adenylate cyclase-activating G protein-coupled receptor signaling pathway	9/58	133/17962	3.99e-10	2.39e-07	1.54e-07
BP	GO: 0019933	cAMP-mediated signaling	10/58	189/17962	4.41e-10	2.39e-07	1.54e-07
BP	GO: 0006164	purine nucleotide biosynthetic process	11/58	253/17962	4.42e-10	2.39e-07	1.54e-07
BP	GO: 0072522	purine-containing compound biosynthetic process	11/58	260/17962	5.91e-10	2.40e-07	1.55e-07
CC	GO: 0042383	sarcolemma	7/58	171/18446	1.05e-06	2.41e-04	1.59e-04
CC	GO: 0015629	actin cytoskeleton	9/58	485/18446	1.89e-05	0.002	0.001
CC	GO: 0098685	Schaffer collateral with CA1 synapses	5/58	119/18446	3.58e-05	0.003	0.002
CC	GO: 0045121	membrane lipid raft	7/58	396/18446	2.33e-04	0.008	0.005
CC	GO: 0098857	membrane microdomain	7/58	397/18446	2.37e-04	0.008	0.005
MF	GO: 0004016	adenylate cyclase activity	4/58	12/16882	6.08e-08	1.75e-05	1.26e-05
MF	GO: 0019239	deaminase activity	4/58	24/16882	1.27e-06	1.02e-04	7.35e-05
MF	GO: 0009975	cyclase activity	4/58	25/16882	1.50e-06	1.02e-04	7.35e-05
MF	GO: 0016849	phosphorus-oxygen lyase activity	4/58	25/16882	1.50e-06	1.02e-04	7.35e-05
MF	GO: 0016814	hydrolase acting on carbon–nitrogen (but not peptide) bonds in cyclic amidines	4/58	26/16882	1.77e-06	1.02e-04	7.35e-05

HRDEGs: hypoxia-related differentially expressed genes. GO: Gene Ontology. BP: biological process. CC: cellular component. MF: molecular function

Table 3.

Results of the KEGG enrichment analysis of HRDEGs

ONTOLOGY	ID	Description	Gene Ratio	Bg Ratio	p value	p adjusted	q value
KEGG	rno04270	Vascular smooth muscle contraction	10/44	145/9437	8.53e-10	1.25e-07	7.36e-08
KEGG	rno00230	Purine metabolism	7/44	138/9437	3.00e-06	1.55e-04	9.10e-05
KEGG	rno00071	Fatty acid degradation	5/44	50/9437	3.16e-06	1.55e-04	9.10e-05
KEGG	rno04927	Cortisol synthesis and secretion	5/44	73/9437	2.07e-05	6.95e-04	4.08e-04
KEGG	rno04918	Thyroid hormone synthesis	5/44	75/9437	2.36e-05	6.95e-04	4.08e-04

HRDEGs: hypoxia-related differentially expressed genes. KEGG: Kyoto Encyclopedia of Genes and Genomes

Fig. 4.

Functional enrichment analysis (Gene Ontology, GO) and pathway enrichment analysis (Kyoto Encyclopedia of Genes and Genomes, KEGG) of hypoxia-related differentially expressed genes (HRDEGs). a and b. Bubble diagrams showing the results of the GO functional enrichment analysis (a) and KEGG pathway enrichment analysis (b) of HRDEGs. c and d. Ring network diagrams showing the results of the GO functional enrichment analysis (c) and KEGG pathway enrichment analysis (d) of HRDEGs. e and f. Bar diagrams showing the results of GO functional enrichment analysis (e) and KEGG pathway enrichment analysis (f) of HRDEGs. HRDEGs: hypoxia-related differentially expressed genes. GO: Gene Ontology. BP: biological process. CC: cellular component. MF: molecular function. KEGG: Kyoto Encyclopedia of Genes and Genomes. The screening criteria for enriched GO terms and KEGG pathways were a P value < 0.05 and an FDR value (q value) < 0.05

The results of the KEGG enrichment analysis indicated that the genes were mainly associated with glycolysis/gluconeogenesis (rno00010), fatty acid metabolism (rno01212), vascular smooth muscle contraction (rno04270), purine metabolism (rno00230), dilated cardiomyopathy (rno05414), the cGMP-PKG signaling pathway (rno04022), the phospholipase D signaling pathway (rno04072), the oxytocin signaling pathway (rno04921), and other pathways.

GSEA and GSVA of the CIRI Dataset

GSEA was employed to assess the impacts of gene expression levels on CIRI by investigating the correlation between gene expression and BP, CC, and MF terms in different subgroups (in the control/IRI groups). All genes in the GEO dataset were found to be significantly enriched in the TGFBETA pathway (Fig. 5b), HYPOXIA pathway (Fig. 5c), JAK_STAT pathway (Fig. 5d), HEDGEHOG pathway (Fig. 5e), and other pathways (Table 4).

Fig. 5.

GSEA and GSVA of the GEO dataset. a. GSEA of the GEO dataset focused on 4 main biological features. b-e. DEGs in the GEO dataset were significantly enriched in the TGFBETA pathway (Fig. 5b), hypoxia pathway (Fig. 5c), JAK_STAT pathway (Fig. 5d), hedgehog pathway (Fig. 5e), etc. F. GSVA of the GEO dataset. Blue represents the normal sample (control) group, and red represents the ischemia‒reperfusion injury sample (IRI) group. IRI: ischemia‒reperfusion injury. GSEA: gene set enrichment analysis. GSVA: gene set variation analysis. The significant enrichment screening criteria for GSEA and GSVA were a P value < 0.05 and an FDR value (q value) < 0.25

Table 4.

GSEA of the merged dataset

Description	Set Size	Enrichment Score	NES	p value	p adjust	q value
WP_TGFBETA_SIGNALING_PATHWAY	102	0.573620578	2.122622133	0.001428571	0.028170319	0.021254311
REACTOME_CELLULAR_RESPONSE_TO_HYPOXIA	57	0.642703434	2.145095777	0.001560062	0.028170319	0.021254311
KEGG_JAK_STAT_SIGNALING_PATHWAY	99	0.37939375	1.40027319	0.049062049	0.173536042	0.130931744
REACTOME_HEDGEHOG_ON_STATE	66	0.463474299	1.58907174	0.013975155	0.075959755	0.05731111
KEGG_MAPK_SIGNALING_PATHWAY	219	0.450353416	1.828573904	0.001282051	0.028170319	0.021254311
WP_IL18_SIGNALING_PATHWAY	206	0.595138419	2.397303138	0.001302083	0.028170319	0.021254311
REACTOME_EXTRACELLULAR_MATRIX_ORGANIZATION	200	0.449802507	1.805547497	0.00130719	0.028170319	0.021254311
REACTOME_PROCESSING_OF_CAPPED_INTRON_CONTAINING_PRE_MRNA	179	0.529123523	2.098759595	0.00131406	0.028170319	0.021254311
REACTOME_TRANSLATION	189	0.581225543	2.316111903	0.001322751	0.028170319	0.021254311
KEGG_CYTOKINE_CYTOKINE_RECEPTOR_INTERACTION	165	0.468256454	1.842681018	0.001331558	0.028170319	0.021254311
REACTOME_HIV_INFECTION	167	0.474492434	1.870243659	0.001331558	0.028170319	0.021254311
REACTOME_SIGNALING_BY_ROBO_RECEPTORS	132	0.476172262	1.830685948	0.001356852	0.028170319	0.021254311
REACTOME_MRNA_SPLICING	137	0.549351727	2.113742853	0.001360544	0.028170319	0.021254311
REACTOME_PROGRAMMED_CELL_DEATH	134	0.495345671	1.899951295	0.001367989	0.028170319	0.021254311
REACTOME_S_PHASE	131	0.472876773	1.809898055	0.001369863	0.028170319	0.021254311

GSEA: gene set enrichment analysis

We performed GSVA with the GEO dataset (Fig. 5f), CIRI samples (IRI subgroups) and the corresponding normal samples (control subgroups) to investigate the variability of gene sets between the CIRI samples (IRI subgroups) and the corresponding normal samples (control subgroups). The results of the functional enrichment analysis revealed differences in the TSP1 pathway gene set between the ischemia–reperfusion injury samples (in the IRI group) and the corresponding normal samples (in the control group) in the GEO subset, as shown in Table 5.

Table 5.

GSVA of the merged dataset

	LogFC	Ave. Expr	t	P Value	adj. P Value
REACTOME_ATTENUATION_PHASE	1.018936881	-0.012990899	13.04520731	7.68E-17	2.01E-13
BIOCARTA_RNA_PATHWAY	1.009063559	0.00749099	11.04681637	2.31E-14	2.01E-11
REACTOME_EUKARYOTIC_TRANSLATION_ELONGATION	1.088578327	-0.025738363	10.65544926	7.49E-14	4.90E-11
BIOCARTA_TSP1_PATHWAY	1.063118432	0.000454537	9.027068908	1.24E-11	4.62E-09
TERAMOTO_OPN_TARGETS_CLUSTER_3	1.128634024	-0.060815482	7.706184307	9.69E-10	6.18E-08

GSVA: gene set variation analysis

Construction of the Diagnostic Models Based On HRDEGs

We performed a LASSO regression analysis to construct a diagnostic model based on the HRDEGs and to assess the diagnostic value of the 60 HRDEGs identified in the GEO dataset (Fig. 6a). Subsequently, we generated forest plots to visualize the expression of HRDEGs in different subgroups in the diagnostic model (Fig. 6b). As shown in Fig. 6b, 11 HRDEGs (ADD2, ADORA2A, Agxt, Ahr, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) were utilized to construct the HRDEG diagnostic model. LASSO regression analysis is an extension of linear regression analysis that incorporates a penalty term (the absolute value of the lambda slope) to correct for overfitting and thus increase the ability to make generalizations based on the model. In our study, we visualized the results of LASSO regression analysis and generated a graph illustrating the LASSO-derived variable trajectories (Fig. 6c). The graph indicates a correlation between the expression level of genes and the coefficient of the lambda penalty term (after log transformation). Furthermore, as the lambda value decreased, the number of genes with a coefficient of 0 gradually increased.

Fig. 6.

Construction of the CIRI diagnostic model with HRDEGs. a. Plot of the LASSO regression diagnostic model of HRDEGs in the GEO dataset. b. Forest plot of HRDEGs in the CIRI diagnostic model. c. LASSO variable trajectory plot of the diagnostic model. d. Number of genes with the lowest error rate determined via the SVM algorithm. e. Number of genes with the highest accuracy rate determined via the SVM algorithm. f. Forest plot of HRDEGs in the CIRI diagnostic model. Venn diagram showing the genes identified with the SVM algorithm and LASSO algorithm. HRDEGs: hypoxia-related differentially expressed genes. LASSO: least absolute shrinkage and selection operator

We constructed a model based on the 60 HRDEGs with the SVM algorithm and identified genes with the lowest error (Fig. 6d) or the highest accuracy (Fig. 6e). The SVM model showed the highest accuracy when 13 genes were included in the analysis (AMPD3, ANXA2, ANXA3, ADORA2A, BIN1, ADD2, APOBEC1, ADPRH, APCS, ACTG2, AKP3, ACTN4, and ALB). We identified the genes by both LASSO analysis and SVM modeling and identified 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) as hub genes. We plotted these genes in a Venn diagram (Fig. 6f).

We performed differential expression analysis on 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) among different subgroups (between the control/IRI groups) in the GEO dataset (Fig. 7a). The results revealed that the expression of six HRDEGs (ADORA2A, AMPD3, ANXA2, ANXA3, APOBEC1, and BIN1) in different subgroups of the GEO dataset was highly significantly different (p < 0.001). Additionally, the expression of two HRDEGs, ALB and APCS, in different subgroups of the GEO dataset (between the control/IRI groups) was highly significantly different (p < 0.01). The expression levels of the HRDEG ADD2 in different subgroups of the GEO dataset (between the control/IRI groups) were significantly different (P < 0.05).

Fig. 7.

Expression of HRDEGs in the GEO dataset of genes associated with cerebral ischemia‒reperfusion injury (CIRI). a. The results of comparisons of HRDEG expression in the CIRI diagnostic model among different subgroups (between the control/IRI groups) in the GEO dataset are shown. Blue represents the normal sample (control) group, and red represents samples in the GEO dataset (IRI) group. b. Correlation heatmap showing the expression levels of HRDEGs in the GEO dataset. C-K. The differentially expressed genes (DEGs) ANXA2 (c), APOBEC1 (d), ANXA3 (e), AMPD3 (f), BIN1 (g), ADORA2A (h), ALB (i), APCS (AUC = 0.67, Fig. 7j), and ADD2 (k) in the GEO dataset, as determined via a ROC curve analysis. The symbol * indicates P < 0.05 and represents statistical significance. The symbol ** is equivalent to P < 0.01, representing a high degree of statistical significance. The symbol *** is equivalent to P < 0.001, which represents the highest degree of statistical significance. The closer the AUC is to 1, the greater the diagnostic value. An AUC greater than 0.9 indicates high accuracy. An AUC between 0.7 and 0.9 indicates some accuracy. An AUC between 0.5 and 0.7 indicates low accuracy. IRI: ischemia‒reperfusion injury. HRDEGs: hypoxia-related differentially expressed genes. ROC: receiver operating characteristic curve

The correlations among the expression levels of 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) are shown in a heatmap (Fig. 7b), and the results indicate a significant correlation (p < 0.01) between the expression of the HRDEG ADD2 and the expression of ALB, APOBEC1, and BIN1 in the infiltration subgroup. Additionally, a significant correlation was observed between the expression levels of the HRDEGs ADORA2A and APOBEC1.

We plotted 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) to show their relationships among subgroups (between the control/IRI groups) in the GEO dataset. The ROC curve and the results are presented in Fig. 7c-k. As shown in the figure, the expression levels of the HRDEGs ANXA2 (AUC = 0.989, Fig. 7c), APOBEC1 (AUC = 0.960, Fig. 7d), ANXA3 (AUC = 0.971, Fig. 7e), AMPD3 (AUC = 0.955, Fig. 7f), and BIN1 (AUC = 0.930, Fig. 7g) in the GEO dataset were associated with the expression of ADORA2A (AUC = 0.832, Fig. 7h), ALB (AUC = 0.743, Fig. 7i), and APCS (AUC = 0.67, Fig. 7j), indicating high diagnostic value among different subgroups (between the control/IRI groups). ADD2 (AUC = 0.691, Fig. 7k) showed lower diagnostic value in different subgroups (in the control/IRI groups) in the GEO dataset.

Validation of the HRDEG-based Diagnostic Model

We performed differential expression analysis of eight HRDEGs (Add2, Adora2a, Alb, Ampd3, Anxa2, Anxa3, Apobec1, Bin1) between different groups (Control/IRI) in the combined validation set (Fig. 8a). The results showed that five HRDEGs (Add2, Ampd3, Anxa2, Anxa3, Apobec1) exhibited statistically significant differences in expression levels between the groups in the validation set (P < 0.001). There was a highly statistically significant difference in the expression level of Bin1 between the groups (Control/IRI) in the validation set (P < 0.01). The HRDEG Apcs was not included in the validation set; therefore, the validation results for the remaining genes are presented.

Fig. 8.

Expression of HRDEGs in the validation set for brain ischemia–reperfusion injury. a. Comparative analysis of HRDEGs between different groups (Control/IRI) in the validation set within the diagnostic model. Blue represents the normal sample group (Control), and red represents the sample group (IRI) from the GEO dataset. b-i. ROC curve analysis of differentially expressed genes Add2 (b), Adora2a (c), Alb (d), Ampd3 (e), Anxa2 (f), Anxa3 (g), Apobec1 (h), and Bin1 (i) in the validation set. ** indicates P < 0.01, representing high statistical significance. *** indicates P < 0.001, representing very high statistical significance. The closer the AUC value is to 1 in the ROC curve, the better the diagnostic performance. An AUC above 0.9 indicates high accuracy. An AUC between 0.7 and 0.9 indicates moderate accuracy. An AUC between 0.5 and 0.7 indicates low accuracy. IRI: ischemia–reperfusion injury. HRDEGs: hypoxia-related differentially expressed genes. ROC: Receiver operating characteristic curve

We also plotted the receiver operating characteristic (ROC) curves for the eight HRDEGs (Add2, Adora2a, Alb, Ampd3, Anxa2, Anxa3, Apobec1, Bin1) between different groups (Control/IRI) in the validation set and presented the results (Fig. 8b-i). The results indicated that the HRDEGs Ampd3 (AUC = 0.943, Fig. 8e), Anxa2 (AUC = 0.986, Fig. 8f), Anxa3 (AUC = 0.910, Fig. 8g), and Apobec1 (AUC = 0.971, Fig. 8h) demonstrated high diagnostic value in distinguishing between the different groups (Control/IRI) in the GEO dataset. Add2 (AUC = 0.857, Fig. 8b), Adora2a (AUC = 0.710, Fig. 8c), and Bin1 (AUC = 0.795, Fig. 8i) showed moderate diagnostic value, while Alb (AUC = 0.548, Fig. 8d) exhibited low diagnostic value in distinguishing between the different groups (Control/IRI) in the GEO dataset. The validation results indicated that the expression trends of the genes Add2, Ampd3, Anxa2, Anxa3, Apobec1, and Bin1 in the validation set were consistent with those in the test set and that differences in expression between the Control and IRI groups were statistically significant.

Clinical Evaluation of HRDEGs

Logistic regression analysis was performed to examine the expression of the 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1). Subsequently, a nomogram analysis was performed to assess the diagnostic power of the model. The results are presented in a nomogram (Fig. 9a). In addition, we performed a calibration analysis on the nomogram based on the logistic regression model and plotted the calibration curves (Fig. 9b). The DCA results were utilized to assess the clinical usefulness of the logistic regression-based model that was constructed. The DCA plot shows the threshold probability on the X-axis and the net benefit on the Y-axis, allowing the evaluation of diagnostic outcomes. The line representing the model consistently exceeded the X-value ranges of both the all-positive line and the all-negative line, indicating a stronger impact of the model as the X-value range increased. Subsequently, the clinical diagnostic value of the 9 HRDEGs was analyzed, and the results are visualized in a graph (Fig. 9c).

Fig. 9.

Diagnostic performance of HRDEGs in the CIRI diagnostic model

a. Nomogram (a), calibration curve (b), and DCA plot (c) showing the logistic regression model with HRDEGs. d. FRIENDS analysis of HRDEGs. HRDEGs: hypoxia-related differentially expressed genes. IRI: ischemia‒reperfusion injury. DCA: decision curve analysis.

Analysis of Immune Cell Infiltration Based On the CIRI Dataset (CIBERSORT)

We used the CIBERSORT algorithm to determine correlations among the expression profiles of 22 immune cell types in different sample groups (between the control/IRI groups) in the GEO dataset. Subsequently, we conducted an analysis of immune cell infiltration based on these results. The infiltration of the 22 immune cell types identified in each sample in the GEO dataset was visualized using a bar chart (Fig. 10a).

Fig. 10.

Analysis of immune cell infiltration using the GEO dataset (CIBERSORT). a. Histogram showing the results of the analysis of immune cell infiltration for 22 immune cells in the GEO dataset. b. Group comparison plot showing the abundance of infiltrating immune cells among different subgroups (between the control/IRI groups) in the GEO dataset. c. Heatmap showing the correlation between the HRDEG expression level and immune cell infiltration in the GEO dataset. The symbol ns represents no statistical significance, indicating that the result was not statistically significant (p > 0.05). The symbol * represents a statistically significant difference (p < 0.05). The symbol ** represents a high degree of statistical significance (p < 0.01), indicating a high level of confidence in the result. The symbol *** represents a very high degree of statistical significance (P < 0.001), indicating an extremely high level of confidence in the result. HRDEGs: hypoxia-related differentially expressed genes. IRI: ischemia‒reperfusion injury. In the correlation heatmap, red circles represent positive correlations among genes and the abundance of infiltrating immune cells. The size of the circle indicates the strength of the correlation. On the other hand, blue circles represent negative correlations between gene expression and the abundance of infiltrating immune cells, with larger circles indicating stronger correlations

We analyzed the abundances of 22 immune cell types in different subgroups (between the control/IRI groups) in the GEO dataset (Fig. 10b). Our findings revealed that 2 types of immune cells (M2 macrophages and CD8⁺ T cells) exhibited statistically significant differences in abundance among cell subgroups (P values less than those of 8 other types of immune cells, namely, naive B cells, activated dendritic cells, activated mast cells, resting mast cells, monocytes, activated natural killer (NK) cells, plasma cells, and quiescent memory CD4⁺ T cells), and the expression levels of the genes associated with these cells were highly significantly different among subgroups in the GEO dataset (p < 0.01). Moreover, three types of immune cells (memory B cells, naive CD4⁺ T cells, and regulatory T cells (Tregs)) showed differences in abundance with varying levels of significance among subgroups in the GEO dataset (P < 0.05).

We observed significant correlations (p < 0.05) among the expression levels of 9 highly relevant HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) in a correlation heatmap. Moreover, we observed a statistically significant increase in the infiltration of immune cells associated with HRDEGs in the GEO dataset. M2 macrophage abundance and ANXA3 expression were strongly correlated with the abundance of infiltrating immune cells.

mRNA‒miRNA, mRNA–TF, and mRNA–RBP interaction networks

We utilized the miRDB database to predict the miRNAs that interact with 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1). Subsequently, we extracted the relevant data from the database to construct an mRNA‒miRNA interaction network using Cytoscape software to visualize the results (Fig. 11a). The mRNA‒miRNA interaction network revealed the presence of 7 hub genes (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, and BIN1) and 71 miRNAs, resulting in a total of 77 pairs of mRNA‒miRNA relationships. Please refer to Table S2 for further details.

Fig. 11.

mRNA‒miRNA and mRNA–RBP interaction networks. a. Hub gene–miRNA reciprocal network. b. Hub gene–RBP reciprocal network. Yellow ovals represent mRNAs, blue ovals represent miRNAs, and pink ovals represent RBPs. RBP: RNA-binding protein

Using the mRNA–RBP data in the ENCORI database, we predicted the RBPs that interact with 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) and then visualized these interactions in an mRNA–RBP interaction network with Cytoscape software (Fig. 11c). The mRNA–RBP interaction network consisted of 8 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APOBEC1, and BIN1) and 16 RBPs, which constituted 43 mRNA–RBP interaction pairs. See Table S3.

Discussion

Ischemic stroke is a cerebrovascular disease characterized by high rates of disability and death. It is the most common subtype of stroke. However, reperfusion of obstructed blood vessels after ischemic stroke can lead to secondary brain damage, known as CIRI. Although mechanical thrombectomy has been shown to establish recanalization at a high rate as a treatment for ischemic stroke, the clinical outcomes for some patients remain poor [35].

Compared with other types of cells, neurons are highly vulnerable to hypoxia and hypoxia–ischemia [36]. Extensive research has demonstrated that mitochondrial damage is the primary factor contributing to hypoxic brain injury [37–39]. Under hypoxic conditions, cytochrome c is released from mitochondria, leading to mitochondrial permeability transition pore opening. This event initiates a caspase reaction cascade. The primary initiating caspase in the endogenous mitochondrial apoptotic pathway is caspase-9. Caspase-3, the executor of cell death, is also activated in neurons exposed to hypoxia. Caspase inhibitors decrease the neuronal mortality induced by hypoxia or hypoxia–ischemia [40].

Recently, a comprehensive study of hypoxic gene clusters associated with the prognosis was performed using univariate and multivariate Cox regression analyses. Researchers have developed time-dependent ROCs to evaluate the accuracy of using hypoxia-related risk profiles for predicting outcomes. Additionally, GSEA and exploration of immune cells infiltrating into the tumor microenvironment were performed to identify genes and cell types associated with hypoxia. This approach allowed the construction of hypoxia-associated risk models and the identification of molecular markers that can be used to distinguish the features of immune cells that infiltrate lung adenocarcinoma tumors [41]. Furthermore, similar to IRI, chronic intermittent hypoxia has been found to contribute to the abundance of reactive oxygen species (ROS) through repeated cycles of systemic hypoxia/reoxygenation. This process can increase blood pressure by modulating sympathetic nerve activity. In obstructive sleep apnea (OSA), researchers identified hypoxia-associated DEGs by identifying OSA-related genes that are also known hypoxia-associated genes. The authors evaluated the diagnostic value of these HRDEGs for patients with OSA using a ROC curve analysis. The diagnostic models were constructed using random forest (RF) and LASSO machine learning algorithms. Furthermore, GSEA was performed to investigate pathways associated with essential hypoxia-associated genes in patients with OSA. The researchers conducted a network analysis of protein‒protein interactions to explore the interactions among hypoxia-associated genes and assess their potential diagnostic value. Finally, researchers have developed a diagnostic model for OSA based on genes associated with hypoxia [42]. However, the mechanism underlying differences among HRDEGs in the context of CIRI remains unclear. Hypoxia plays a crucial role in the pathophysiology of CIRI, and further elucidation of its effect is essential for understanding CIRI pathophysiology. Prior to this study, no HRDEGs involved in CIRI had been reported.

In this study, DEGs were analyzed by combining 2 GEO CIRI datasets and intersecting them with a hypoxia-related gene dataset to identify 60 CIRI HRDEGs. Further analysis of the DEGs identified via GO enrichment analysis revealed 60 HRDEGs, and they were most enriched in the BP term circadian rhythm (GO: 0007623), CC term actin cytoskeleton (GO: 0015629), and MF term phospholipid binding (GO: 0005543). KEGG enrichment analysis revealed that the 60 HRDEGs were enriched in the vascular smooth muscle contraction (rno04270) pathway. GSEA revealed significant enrichment of all genes in the TGFBETA pathway (Fig. 5b), hypoxia pathway (Fig. 5c), JAK_STAT pathway (Fig. 5d), hedgehog pathway (Fig. 5e), etc. GSVA revealed that the expression of some gene sets, such as the set involved in the TSP1 pathway, differed between CIRI samples (in the IRI group) and corresponding normal samples (in the control group) in the GEO dataset. The HRDEG-based CIRI diagnostic model was generated using LASSO regression analysis and the SVM algorithm. The genes obtained from both the LASSO algorithm and SVM algorithm were compared, resulting in 9 common hub genes identified via intersection (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) as HRDEGs. ROC curves were then constructed to assess the diagnostic effectiveness of these HRDEGs used as targets. The differential expression levels of ANXA2, APOBEC1, ANXA3, AMPD3, and BIN1 in the GEO dataset indicated their significant diagnostic value among subgroups (between the control/IRI groups). ADORA2A, ALB, and APCS in the GEO dataset also exhibited high diagnostic value. However, ADD2 in the GEO dataset showed low diagnostic value when different subgroups (between the control/IRI groups) were compared. Subsequently, we performed logistic regression analysis and developed logistic regression models based on the expression of the nine highly correlated DEGs (HRDEGs). We analyzed a column–line plot of the models and generated a figure with these data to assess the diagnostic ability of the models (Fig. 8a). Furthermore, a calibration analysis was performed, and calibration curves were plotted (Fig. 8b). The constructed logistic regression model was evaluated via DCA to assess its clinical value. FRIENDS analysis was then performed with the 9 HRDEGs, and the resulting plot revealed that BIN1 plays an important role. The mRNA‒miRNA interaction network was established based on 7 hub genes (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, and BIN1) and 71 miRNAs, resulting in a total of 77 mRNA‒miRNA interaction pairs. In this study, we constructed an mRNA–RBP interaction network using 8 highly related HRDEGs, namely, ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APOBEC1, and BIN1, along with 16 RBPs. Thus, 43 pairs of mRNA–RBP interactions were identified. Additionally, our analysis of infiltrating immune cells in the GEO dataset revealed a significant correlation between the abundance of M2 macrophages and the ANXA3 expression level.

A total of 9 HRDEGs (ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1) were identified as hub genes. Combined with our diagnostic model based on HRDEGs, the ROC curve analysis revealed that the expression levels of the HRDEGs ANXA2 (AUC = 0.989, Fig. 7c), APOBEC1 (AUC = 0.960, Fig. 7d), ANXA3 (AUC = 0.971, Fig. 7e), AMPD3 (AUC = 0.955, Fig. 7f), and BIN1 (AUC = 0.930, Fig. 7g) in the GEO dataset among different subgroups (between the control/IRI groups) all showed high diagnostic value. The diagnostic model based on HRDEGs was then clinically assessed and was shown to have a better diagnostic accuracy. In addition, the FRIENDS analysis showed results similar to those obtained from the ROC curve analysis.

ANXA2, also known as membrane-associated protein A2, is a member of the calcium-dependent phospholipid-binding protein family. It plays a crucial role in regulating cell growth and signaling pathways. Additionally, this protein functions as an autocrine factor. Its expression is significantly increased in both CIRI mouse and rat datasets [43]. Furthermore, it is associated with anti-inflammatory responses [44]. Moreover, it plays a role in cerebral vascular integrity and retinal neovascularization and may be involved in angiogenesis after ischemic stroke [45]. Moreover, ANXA2 has been identified as a key hypoxia-associated gene involved in various diseases and physiological processes, such as endometrial cancer and glioblastoma, as well as bone development and bone repair [46–48].

APOBEC1, also known as apolipoprotein B mRNA-editing enzyme catalytic polypeptide 1, is a cytidine deaminase with sequence specificity. It was named because of its ability to catalyze C-to-U editing of the apolipoprotein B mRNA. APOBEC1 has multiple functions, including deaminase activity, enzyme-activating activity, and ribonucleoprotein complex-binding activity. It is involved in various processes, such as cellular responses to insulin stimulation, nucleic acid metabolic processes, and the positive regulation of mRNA modifications. APOBEC1 is predicted to be a component of the mRNA-editing enzyme complex of apolipoprotein B and is expected to exhibit activity in both the cytoplasm and nucleus. A previous study focused on familial hyperlipidemia and biomarkers for benign liver tumors, colon cancer, and obesity. The homologous relationship between APOBEC1 in humans and its involvement in lipoprotein metabolic pathways was assessed. Additionally, the study explored the interaction between APOBEC1 and 17α-ethynylestradiol. The findings revealed that APOBEC1 expression increased in cellular assays as the severity of injury due to oxygen–glucose deprivation (OGD) increased in vitro. Moreover, the overexpression of APOBEC1 intensified the severity of OGD-induced injury in vitro. In rats with middle cerebral artery occlusion (MCAO), APOBEC1 expression was elevated in the cortex, and its expression increased with the severity of injury due to OGD [49].

ANXA3 is a member of the membrane-bound protein family. This calcium-dependent phospholipid-binding protein plays a role in regulating cell growth and signaling pathways. It inhibits phospholipase A2 and catalyzes the formation of inositol 1-phosphate by cleaving inositol 1,2-cyclic phosphate. Additionally, ANXA3 may have an anticoagulant function. Isoforms of ANXA3 have been detected in the rat brain after ischemia, and they are produced by microglia in control tissues and activated microglia/macrophages in infarct areas [50]. Recent bioinformatics studies have shown that the serum ANXA3 level is a reliable marker for diagnosing ischemic stroke and that immune cell infiltration plays a crucial role in the development and progression of this condition [51]. Furthermore, studies have revealed that inhibiting ANXA3 and activating the PI3K/Akt pathway prevents cerebral CIRI [52]. ANXA3 is associated with hypoxia-inducible factor-1α and is involved in various disease-related processes [53–55].

The activity of AMPD3, an enzyme that converts AMP to IMP, is highly regulated in the adenylyl catabolic metabolic pathway. The AMPD3 gene encodes isoform E in humans and isoform C in rats, and both isoforms are produced in the brain [56]. AMPD3 is crucial for the adenylate catabolic pathway because it converts adenosine monophosphate into inosine monophosphate and uric acid. It plays a ubiquitous regulatory role in energy metabolism; it is involved in ATP-mediated lipid and protein synthesis, intracellular calcium buffering, apoptosis regulation, recovery pathway activation, and other biological processes. AMPD3 is closely associated with depression and plays an important role in the brain [57]. Furthermore, bioinformatics studies identified AMPD3 as a hypoxic immune cell-associated hub gene in patients with type 2 diabetes.

BIN1, also known as bridging integrating factor 1, encodes various isoforms of nucleoplasmic junction proteins. The isoforms expressed in the central nervous system interact with dynamin, synaptophysin, endothelin, and lattice proteins, suggesting their potential involvement in synaptic vesicle endocytosis. Moreover, the commonly expressed isoforms and those expressed in muscle are located in the cytoplasm and nucleus and are critical for activating apoptotic processes independent of cysteine asparaginase activity. Extensive research with mice has revealed the crucial role of this gene in the development of cardiac muscle. Alternative splicing leads to the generation of multiple transcript variants that encode distinct isoforms. A recent bioinformatics study further supported the differential expression of Bin1 in cerebral infarction [58]. These findings validate our data mining results, confirm the reliability of our diagnostic model, and underscore the significance of our study.

Our study was the first to report a dual association of ADORA2A with CIRI and other hypoxia-related genes. ADORA2A encodes a protein that plays a role in G protein-coupled adenosine receptor activity, actinin-binding activity, and metabotropic glutamate receptor type 5 binding activity. It is involved in various processes, including G protein-coupled receptor signaling pathways, chemical synaptic transmission, and secretory regulation. ADORA2A is located in multiple cellular components, including the axoneme, postsynaptic density complex, and synaptic membrane. It functions at glutamatergic synapses and is a component of both postsynaptic and presynaptic membranes. In cases of experimental stroke, thromboembolic stroke induces ADORA2A expression in the brain, which helps prevent brain damage [59]. Moreover, the adenosine A2A receptor (ADORA2A) inhibits inflammation mediated by microglia and astrocytes, thereby reducing white matter damage in cerebral small vessel disease [60]. Hypoxia induces changes in adenosine receptor–ADORA2A receptor expression, thereby promoting the transcription of genes involved in extracellular adenosine signaling [61]. Additionally, previous reports indicated that ADORA2A is a DNA methylation gene that is highly expressed in mouse cortical neurons and is expressed in response to OGD/reoxygenation [62]. The ROC curve analysis revealed that ADORA2A was a promising diagnostic target (AUC = 0.832, Fig. 7h), indicating its potential diagnostic value for identifying CIRI. Other studies have also documented the association of ADD2, ALB, and APCS with cerebral ischemia‒reperfusion, which aligns with the findings of our data mining analysis and validates our results [63–66].

The results of our GO enrichment analysis revealed that the DEGs were significantly enriched in circadian rhythms (GO: 0007623) and circadian rhythm regulation (GO: 0042752). Animal experiments related to cerebral ischemia have shown that circadian rhythms play roles in disease process cascades. Some findings have suggested a link between circadian rhythms and CIRI, and circadian rhythms, which are associated with depression, autism, stroke, and other brain diseases, regulate CIRI [65–67]. In rodent models of ischemic stroke, the cerebral infarct area is smaller during the night (the active phase) than during the day (the inactive phase) [65]. The core clock genes in the brain, which are widely distributed, influence various physiological processes, including stroke triggers such as arterial blood pressure, heart rate, coagulation homeostasis, and other rhythmic events [68]. The actin cytoskeleton plays a crucial role in shaping the morphology and polarity of all eukaryotic cells. It is critical for the assembly and disassembly of filamentous actin structures, which drive dynamic processes such as cell motility, phagocytosis, growth cone guidance, and cell division. During embryonic development, migrating cells or cell extensions, such as axons, undergo rapid and spatial alterations to the actin cytoskeleton organization in response to external cues [69]. Research has shown that hypoxia increases the actin polymerization rate, leading to cytoskeletal protein remodeling [70]. The associated actin-binding proteins are also related to the mechanism underlying CIRI [71]. Our identification of essential genes revealed that Anxa2 and Anxa3, members of the phospholipid-binding protein family, are enriched in phospholipid binding and are associated with ischemic stroke and hypoxia [51–55]. KEGG pathway enrichment analysis revealed a significant association between vascular smooth muscle contraction and hypoxia. This pathway regulates vascular smooth muscle cell proliferation, stimulates excessive vascular smooth muscle cell growth, and promotes vascular remodeling [72, 73]. Our data mining analysis and diagnostic model are validated by these related studies. Importantly, our study is the first to reveal that crucial genes are enriched in regulatory BP terms such as NIK/NF-κB signaling (GO: 1,901,222) and the hypoxia response (GO: 0001666).

Consistent with the results of our GSEA and related studies, the TGFBETA pathway has been previously reported to contribute to CIRI [74]. In the brain tissue of mice subjected to MCAO, JAK2 is highly expressed, and the JAK-STAT pathway affects the progression of CIRI through its regulation of apoptosis [75]. Furthermore, oxidative stress and inflammation promote the activation of the Hedgehog pathway in ischemic brain tissues [76]. These studies support our data mining results and confirm the significant roles of the enriched genes in disease development. Additionally, our data mining results revealed that the hypoxia pathway is a key factor in CIRI. Furthermore, the results of our GSVA revealed differences in the expression of genes in the TSP1 pathway and other gene sets between CIRI samples (in the IRI group) and corresponding normal samples (in the control group) in the GEO dataset (Fig. 5f). According to the literature, the platelet response element TSP-1 is a naturally occurring vasopressor that inhibits vascular generation in vivo and regulates cell behavior and extracellular matrix structure. Inhibition of endothelial cell migration and induction of apoptosis are the mechanisms by which TSP1 inhibits angiogenesis [77]. In vitro experiments with mouse brain endothelial cells (CECs) indicated a decrease in TSP-1 expression following OGD [78]. These findings align with our data mining results, indicating that IRI promotes the activation of the TSP1 pathway, leading to the inhibition of vascular regeneration. The enrichment analysis results further support the value of HRDEG-based diagnostics for patients with CIRI.

Increasing evidence suggests that CIRI is associated with an immune imbalance. Several hours after ischemic stroke, neuronal cells die, triggering the innate immune response in the brain. This response leads to the production of inflammatory cytokines, chemokines, ROS, nitrogen oxides, and other neurotoxic substances. Moreover, it disrupts the blood‒brain barrier and initiates a cascade of inflammatory reactions. Furthermore, immune inflammatory cells such as polymorphonuclear neutrophils, lymphocytes, and monocytic macrophages infiltrate the brain tissue after passing through the vascular endothelium. These cells identify antigens within the central nervous system (CNS) and trigger adaptive immune responses, exacerbating secondary neuronal damage and worsening any neurological dysfunction [79]. Our findings revealed a highly significant difference (P < 0.001) in the abundance of M2 macrophages and CD8⁺ T cells among the different subgroups in the GEO dataset. Microglia, which are resident macrophages in the CNS, play crucial roles in immune surveillance and maintaining CNS homeostasis under normal conditions [80, 81]. Microglia can be rapidly activated in response to CNS damage [82]. Like other macrophages, microglia can be classified as proinflammatory M1 or anti-inflammatory M2 microglia. Moreover, they transition between these two states through a highly controlled mechanism [83]. M1 microglia/macrophages produce proinflammatory mediators such as tumor necrosis factor-β (TNF-β), interleukin IL-6, and interleukin (IL)-1, which may worsen brain injury [84, 85]. On the other hand, M2 microglia/macrophages generate anti-inflammatory cytokines and neurotrophins, including IL-4, IL-10, transforming growth factor-β (TGF-β), insulin-like growth factor (IGF)-1, and brain-derived growth factor (BDNF); these substances contribute to the resolution of inflammation and promote brain repair [85, 86]. Isoforms of membrane-linked protein A3 (ANXA3), whose levels are increased in the postischemic rat brain, are specifically produced by microglia in control tissues and activated microglia/macrophages in infarct areas. This finding confirmed our data mining results, which indicated a strong correlation between the abundance of infiltrating M2 macrophages and high ANXA3 expression. Additionally, extensive reports have demonstrated that CD8⁺ T cells exert a detrimental effect on exacerbating neurological damage in the context of CIRI, with a 24-fold increase after transient cerebral ischemia [87]. Our study is the first to identify the differential and low abundances of naive B cells, resting mast cells, and plasma cells in the context of CIRI.

Despite these reported findings, our study also has several limitations. First, although this study successfully identified and constructed a diagnostic model based on nine key genes using bioinformatics methods, we must acknowledge that the lack of wet laboratory validation is a major limitation. The current conclusions primarily rely on the analysis of the GEO dataset and have not yet been validated in actual samples from patients with cerebral ischemia‒reperfusion injury (CIRI) through wet lab experiments such as qPCR and Western blot.

Our analysis indicated that CIRI may frequently occur in the following diseases or pathological processes in clinical settings:

Postreperfusion after cerebral infarction—Regardless of whether intravenous thrombolysis or arterial thrombectomy is used, restored blood flow may cause reperfusion injury to the brain.

Postcardiac arrest resuscitation—After successful resuscitation of a patient with cardiac arrest, the restored blood flow may result in reperfusion injury to the brain.

Postaneurysm clipping—In aneurysm surgery, temporarily blocking the blood flow to perform the surgery and then restoring it can lead to reperfusion injury.

Cardiac surgery—Certain cardiac surgeries, especially those requiring cardiopulmonary bypass, may induce cerebral ischemia‒reperfusion injury.

Postcarotid endarterectomy—After clearing the plaques that have accumulated in the carotid artery and restoring blood flow, reperfusion injury may occur.

The clinical manifestations may or may not support the clinical occurrence and development of cerebral ischemia‒reperfusion injury, thus limiting the available data on this pathophysiological process. We acknowledge the need for more experiments and clinical tests to validate the expression and functions of these key genes in real patients. However, the essential factor for CIRI is reperfusion. Clinically, whether CIRI has occurred is determined by comprehensive clinical presentations, changes in neurological scores, imaging findings, and laboratory tests. Therefore, many of the patients described above may have experienced CIRI, but CIRI is rarely definitively diagnosed. Using imprecise samples to validate research findings is not meaningful.

Second, our study’s dataset was primarily based on rat (Rattus norvegicus) samples rather than human samples. Although rat models are widely used in the study of cerebral ischemia‒reperfusion injury, the physiological and gene expression differences between species may limit the diagnostic value of these screened genes in actual human stroke patients. Future research should prioritize the use of datasets based on human samples to enhance the clinical relevance and applicability of the findings.

Furthermore, no additional clinical information was available in the GEO database; therefore, a prognostic model was built without considering the clinical setting. Additionally, in this study, we used standard dataset partitioning and cross-validation methods to train and evaluate the LASSO regression model, ensuring the model’s stability and effectiveness. Although our model performed well on the existing data and its robustness was increased through external validation using an independent dataset and ROC analysis, the HRDEG Apcs was not included in the validation set. Therefore, only the remaining genes were independently validated.

Nonetheless, we have not abandoned further exploration of this research. Future studies involving precise screening of clinical patients, accurate collection of relevant samples and data from CIRI patients, and experiments to confirm the expression levels of these genes in real patients are needed to further validate the diagnostic value and biological function of these genes.

In this study, we identified 9 relevant HRDEGs—ADD2, ADORA2A, ALB, AMPD3, ANXA2, ANXA3, APCS, APOBEC1, and BIN1—that may be used as diagnostic biomarkers for CIRI. Among these genes, ANXA2, APOBEC1, ANXA3, AMPD3, and BIN1 exhibited the highest diagnostic efficiency. We employed the LASSO method and other logistic regression techniques to establish a diagnostic model for CIRI, and the results were validated internally and externally, yielding promising diagnostic outcomes. Therefore, our genetic markers suggest a precise and reliable approach for predicting the prognosis of CIRI patients. Additionally, we analyzed immune cell infiltration in the context of CIRI using CIBERSORT and found a potential correlation between hypoxia and immune cell infiltration. Thus, this study presents a novel perspective on the diagnosis and treatment of CIRI.

Supplementary Information

Below is the link to the electronic supplementary material.

Authors’ contributions

All the authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Kaiwen Sun, Yang Dong, Lei Cao, Dongpeng Li, Jinghong Li, and Manxia Zhang. The first draft of the manuscript was written by Kaiwen Sun and Hongwei Li, and all the authors commented on previous versions of the manuscript. Bo Yang and Dongming Yan reviewed and revised the manuscript. All the authors have read and approved the final manuscript.

Funding

The authors declare that no funds, grants, or other support was received during the preparation of this manuscript.

Data Availability

Data are openly available in a public repository. The datasets generated and/or analyzed during the current study are available in the Baidu Cloud Inventory Repository https://pan.baidu.com/s/1g-xUc9BydOO8NzdDiRRTXw?pwd=5ay5 Extract code: 5ay5.

Code availability

The datasets generated and/or analyzed during the current study are available from the Baidu Cloud Inventory Repository. https://pan.baidu.com/s/1g-xUc9BydOO8NzdDiRRTXw?pwd=5ay5 Extract code: 5ay5.

Declarations

Ethical Approval

GeneCards and GEO are public databases. Ethical approval was obtained from the patients included in the database. Users can download relevant data for free for research and publish relevant articles. Our study is based on open source data, and thus there are no ethical issues or other conflicts of interest.

Consent to Participate

Not applicable.

Consent to Publish

Not applicable.

Competing Interests

The authors declare that they have no conflicts of interest.