Gene patterns linked to cell aging that may help diagnose heart attacks using multiple data types and machine learning

Transcriptome datasets associated with AMI were obtained from the Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/↗), including GSE34198↗ (49 AMI cases, 48 controls), GSE48060↗ (31 AMI cases, 21 controls), GSE60993↗ (26 AMI cases, 7 controls), and GSE29111↗ (37 AMI cases, 15 controls). Subsequent missing value handling and normalization were performed: in the expression matrix, rows (genes) and columns (samples) with a missing value (NA) proportion exceeding 50% were removed. The impute.knn function from the R package “impute” (11) was employed to fill in missing values using k-nearest neighbors, followed by log2(x + 1) normalization of the data. The datasets GSE34198↗, GSE48060↗, and GSE60993↗ were combined, and batch effects were removed using the removeBatchEffect function from the R package ‘limma’ (version 3.42.2) (12), producing a unified expression profile for the training set.

Protein expression profiles for 48 AMI cases and 50 controls were obtained from the iProx database (https://www.iprox.cn/page/home.html↗) under accession PXD062794 (13), followed by missing value imputation and normalization.

The senescence gene set (Supplementary Table 1, n = 307) used in this study was downloaded from the CellAge database (https://genomics.senescence.info/cells/↗).

Differential expression gene (DEG) analysis was performed using the “limma” package. Differentially expressed genes were identified using a threshold of |fold change (FC)| ≥ 1.2 and P < 0.05. Notably, the selection of |FC| ≥ 1.2 was based on comprehensive consideration of transcriptomic analytical practices, statistical principles, and the biological characteristics of peripheral blood samples in acute myocardial infarction. There is no universal fixed standard for FC threshold selection in transcriptomic studies, which is often determined by research context and objectives. FC reflects the magnitude of expression difference (effect size), while p-value evaluates statistical significance; the combined use of FC and statistical significance helps balance biological effect and statistical reliability. A moderate FC threshold (|FC| ≥ 1.2) facilitates the retention of biologically meaningful genes with moderate expression changes, which is particularly suitable for the “low-amplitude, polygenic synergistic regulation” expression pattern commonly observed in peripheral blood transcriptome studies. To further reduce potential noise and ensure result robustness, the differentially expressed genes were subsequently subjected to multi-step screening including weighted gene co-expression network analysis (WGCNA) and four machine learning algorithms. Volcano plots and expression heatmaps were generated using the R packages “ggplot2” (14) and “heatmap” to visualize the identified differential genes.

Enrichment analyses for Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways were performed using the enrichGO and enrichKEGG functions from the “clusterProfiler” package (15). Gene set sizes were constrained between 5 and 5000. Statistical significance was defined as a p-value less than 0.05 and a false discovery rate (FDR) below 0.1. The results were visualized as bubble plots using the “ggplot2” package.

The gene co-expression network was constructed, and modules associated with AMI were screened using the R package “WGCNA” (16) was used to construct a gene co-expression network and screen modules associated with AMI. The specific steps were as follows: first, genes with a small median absolute deviation (approximately 50%) were excluded. The pickSoftThreshold function was used to select the soft threshold (β = 16), with the selection criterion of a fitting index > 0.8 (at β = 16, the scale-free topology fit index reached R² = 0.86, and the mean connectivity was 0.11; Supplementary Figures 1A, B). The adjacency matrix was derived using the adjacency function and subsequently transformed into a topological overlap matrix (TOM) through the TOMsimilarity function. Gene modules were identified through unsupervised clustering, employing the function ‘hclust’ for hierarchical clustering. The clustering tree branches were segmented into distinct modules using the cutreeDynamic function (Supplementary Figure 1C). The mergeCloseModules function was applied to combine modules with highly correlated eigengenes, resulting in four gene modules. To elucidate the biological relevance of the modules, a correlation analysis was conducted between the eigengenes of the four modules and the AMI phenotype, followed by the creation of a module-phenotype association heatmap using the labeledHeatmap function. Gene modules significantly linked to AMI were chosen for further analysis, with module membership and gene significance scores assessed to determine their importance. We selected modules that exhibited significant positive correlations with AMI, as positive associations are more closely relevant to the pathological and diagnostic mechanisms of AMI. Although the correlation coefficients were moderate (r = 0.24 and r = 0.23), both modules showed highly significant statistical significance (P < 0.001), supporting their reliability in this peripheral blood transcriptomic study, where sample heterogeneity commonly limits phenotypic correlation coefficients from reaching high levels. Of note, WGCNA was conducted using differentially expressed genes rather than the whole transcriptome. This design reduces noise from low-variation background genes, strengthens disease-relevant signals, and improves the stability of module detection related to AMI.

A univariate logistic regression analysis (using P < 0.10 as the inclusion criterion) assessed the crude association between each SRG and AMI, with a forest plot illustrating the effect sizes. Four algorithms were independently executed: eXtreme Gradient Boosting (XGBoost) with a learning rate of 0.05 and max depth of 3 using the ‘xgboost’ package (17), Random Forest with 500 trees and 5 variables tried at each split using the ‘randomForest’ package (18), Lasso with 10-fold cross-validation using the ‘glmnet’ package (19), and Gradient Boosting Machine (GBM) with an interaction depth of 3 and shrinkage of 0.05 using the ‘gbm’ package (http://cran.rproject.org/web/packages/gbm/index.html↗). Variable importance scores were output, and candidate genes were selected according to the “maximum-gap cut-off” strategy based on the ranking of importance scores. The “maximum-gap cut-off” determines the threshold based on the adjacent jumps in the ranked importance scores, with no subjective presets, which can reduce overfitting and retain the most informative variables. Finally, the SRGs screened by each algorithm were intersected to obtain common genes, which were defined as diagnostic SRGs. To evaluate model stability and avoid overfitting, 10 fold cross validation was performed for each machine learning algorithm. The mean AUC, standard deviation (SD), and 95% confidence interval (CI) across all folds were calculated and reported. The mean AUCs of XGBoost, Random Forest, Lasso, and GBM were 0.810 (SD = 0.025), 0.796 (SD = 0.027), 0.775 (SD = 0.030), and 0.803 (SD = 0.028), respectively. All models showed small SD and narrow 95% CI, indicating good generalization ability and low overfitting risk.

A binary multivariate logistic model was fitted using glm (family = binomial) (20), and the unstandardized coefficient β of each SRG was extracted as the weight. The combined dataset (GSE34198↗, GSE48060↗, and GSE60993↗) was randomly divided into training set and test set according to a ratio of 7:3, and the diagnostic model construction and model validation was used respectively. Linear predictors were calculated, and the rmda:: decision_curve function evaluated the net benefit rate across a threshold range of 1%–99% using these predictors.

A nomogram was constructed using the “rms” package (v6.7.0) (21). In the nomogram, each gene is assigned a score, and the cumulative score of the SRGs predicts an individual’s AMI risk probability. A calibration curve was plotted to verify the consistency between predicted probabilities and observed results.

A risk score was constructed based on the regression coefficients (β_i) of the above model. The risk was calculated using the formula: Risk = Σ β_i × Expr_i, where Expr_i denotes the standardized expression level of the i-th SRG, and β_i is the associated regression coefficient. The risk score for each sample was calculated by summing the products of each gene’s expression level and its corresponding regression coefficient. This risk score represents a supervised modeling output derived from cross-sample expression patterns, primarily designed to quantify the discriminative capacity of the model for distinguishing AMI from control samples. As a linear combination of gene expression weighted by regression coefficients, it emphasizes predictive performance at the cohort level.

Decision Curve Analysis (DCA) was used to evaluate the risk score. With “All” (all patients diagnosed with AMI) and “None” (all patients diagnosed as normal) as the baselines, the net benefits of the prediction model under the two threshold probabilities were compared to assess the benefit of the risk score in clinical decision-making. The DCA was performed using the “rmda” package (v1.6) (22), and visualization was conducted in combination with the “ggplot2” package (v3.4.4). In parallel, a logistic regression model was constructed with a binary outcome (1 = AMI, 0 = Control) and the SRGs risk score as the predictor. The results demonstrated a statistical association between the risk score and AMI occurrence, indicating a predictive capability of the model. Importantly, the integration of DCA and logistic regression provides complementary evidence from both statistical and clinical perspectives. While logistic regression quantifies the strength and significance of the association between the risk score and disease occurrence, DCA evaluates the net clinical benefit under practical decision-making scenarios. Together, these analyses comprehensively demonstrate that the SRGs-based model is not only statistically reliable but also clinically valuable, thereby overcoming the limitations of relying on a single analytical approach.

Using the SRG diagnostic model, the ‘pROC’ package was employed to generate the Receiver Operating Characteristic (ROC) curve and compute the Area Under the Curve (AUC) (23), considering AMI as the true positive. For the logistic regression diagnostic model, the input variable for ROC analysis was the calculated risk score derived from the model, combined with the corresponding clinical labels (AMI vs. normal). The AUC was calculated based on the combined dataset (GSE34198↗, GSE48060↗, and GSE60993↗).

To further ensure model robustness and minimize the risk of overfitting and data leakage, a 10-fold cross-validation strategy was implemented within the training set. In this process, the training data were partitioned into ten subsets, where in each iteration, 90% of the data were used as the training fold and the remaining 10% as the validation fold. The model was independently trained within each training fold and evaluated on the corresponding validation fold, and the AUC values were averaged across all folds to estimate the stability and generalizability of the model.

Eight-week-old male C57BL/6 mice and 18-month-old male C57BL/6 mice were sourced from Henan Skobes Biotechnology Co., Ltd. All procedures received approval from Nankai University’s Institutional Animal Care and Use Committee (IACUC; approval number 2025-SYDWLL-000757). Following a one-week acclimatization to reduce environmental stress, mice were randomly allocated to experimental groups.

For the 8-week-old cohort, infarct-zone cardiac tissues were collected 24 h after LAD ligation for exploratory RNA-seq analysis.

To further evaluate the age-related relevance of the identified genes, an additional 18-month-old cohort underwent the same MI/sham procedure. In this aged cohort, peripheral blood leukocytes and infarct-zone cardiac tissues were collected 24 h after LAD ligation for qPCR and western blot validation.

The myocardial infarction (MI) model was established by ligating the left anterior descending (LAD) coronary artery, a commonly used technique for inducing MI in mice. The mice were randomly assigned to either the Sham group or the MI group. Both groups were anesthetized using isoflurane and positioned supine. After removing the chest hair, the surgical area was disinfected with 75% ethanol. A left thoracotomy was conducted between the 3rd and 4th intercostal spaces to access the heart. The LAD was ligated using a 6–0 silk suture approximately 2–3 mm below the left atrial appendage. Following ligation, the chest was immediately closed, and any remaining air was removed to avoid pneumothorax. The surgical incision was sutured, and sham-operated mice underwent the same procedure without LAD ligation.

Successful induction of MI was verified 24 hours post-surgery via transthoracic echocardiography using a high-resolution small-animal ultrasound system. Heart samples and plasma were collected and stored in liquid nitrogen until further analysis.

RNA sequencing (RNA-seq) was performed by Majorbio (Shanghai, China). Poly(A)-enriched mRNA was reverse-transcribed and used to construct strand-specific cDNA libraries, which were sequenced on an Illumina NovaSeq 6000 platform. Clean reads were aligned to the Mus musculus reference genome (GRCm38/mm10, NCBI) using HISAT2. Gene expression levels were quantified as fragments per kilobase of transcript per million mapped reads (FPKM) for downstream visualization and expression comparison.

For differential expression analysis, raw read count data were used as input for DESeq2 (24). Infarct-zone cardiac tissues collected 24 h after LAD ligation from sham mice (n = 3) and MI mice (n = 3) were analyzed. Differentially expressed genes were identified using the criteria of |FC| ≥ 1.5 and P < 0.05. Volcano plots were generated using ggplot2 and ggrepel (25).

Given the limited sample size, this RNA-seq analysis was intended primarily as exploratory and supportive in vivo evidence for the expression trends of the identified genes.

Peripheral blood was collected from sham and MI mice anesthetized with isoflurane. After red blood cell lysis using RBC lysis buffer (Solarbio, 4 °C, 15 min), the remaining peripheral blood leukocytes were harvested by centrifugation. Total RNA was extracted using TRIzol (Solarbio), and 1 μg of RNA was reverse-transcribed into cDNA. Gene expression was measured using PerfectStart Green qPCR SuperMix (2×, Transgen). The primer sequences are listed in. 1

For protein analysis, hearts were excised 24 h after LAD ligation, and infarct-zone tissues from the anterior wall were carefully dissected on ice. Tissues were snap-frozen and lysed in RIPA buffer supplemented with Complete™, PhosSTOP™, and 1 mM PMSF. Equal amounts of protein (20 μg) were separated by 10% SDS-PAGE, transferred onto a 0.22 μm PVDF membrane, and blocked with 5% skim milk in TBST. Membranes were incubated with primary antibodies overnight at 4 °C, followed by HRP-conjugated secondary antibodies (1:5,000) for 1 h at room temperature. Signals were detected using ECL (Thermo Fisher) on a Tanon 5200 system. Band intensities were quantified using NIH ImageJ and normalized to GAPDH. A total of six independent biological samples per group (Sham n = 6, MI n = 6) were included for western blot validation. The antibodies used are listed in. 1

The ImmuCellAI algorithm (26) (https://guolab.wchscu.cn/ImmuCellAI/↗) was employed to analyze expression profiles and quantify immune cell infiltration abundance.

Consensus clustering analysis was performed using ConsensusClusterPlus (27). Agglomerative pam clustering utilized the 1-Spearman correlation coefficient as the distance metric, conducting 10 repeated samplings on 80% of the samples. The optimal number of clusters was determined based on empirical cumulative distribution function (CDF) plots. Based on the expression profiles of senescence-related genes (SRGs), AMI patients were stratified into distinct molecular subtypes.

Subsequent analyses were conducted at the whole-transcriptome level. Immune infiltration was estimated using the ImmuCellAI algorithm based on predefined immune cell marker genes, and functional enrichment analyses were performed using genome-wide differentially expressed genes between clusters. This analytical framework enables the characterization of subtype-specific biological features beyond the initial clustering features.

AMI senescence-related scores (ASS) were constructed based on the pre-defined diagnostic SRGs. The gsva() function from the R package “GSVA” (method = “ssgsea”) (28) was used to quantifies the coordinated activation based on the relative expression ranks within each sample. The obtained scores were normalized by z-score to facilitate inter-sample comparison and visualization.

For the ASS-based senescence scoring model, the input consisted of the ASS calculated via ssGSEA and the corresponding clinical labels (AMI vs. normal). The AUC analysis was performed on the combined dataset (GSE34198↗, GSE48060↗, and GSE60993↗). Additionally, a 10-fold cross-validation was performed to evaluate the stability and generalizability of the AUC model. In this process, the training data were partitioned into ten subsets, with 90% of the data used as the training fold and the remaining 10% as the validation fold in each iteration. The model was independently trained within each training fold and evaluated on the corresponding validation fold, with the AUC values averaged across all folds to assess the robustness and generalizability of the model.

ASS represents an unsupervised pathway-level metric that reflects the intrinsic activity of cellular senescence programs, independent of phenotype labels, thereby providing a mechanistically interpretable measure of senescence status.

The single-cell transcriptome dataset of AMI (GSE163129↗, n=5, cell=40,314) was used. Referring to the analysis workflow from the original study (https://github.com/Junglab-CMC/Macrophage-heterogeneity-after-MI↗), the Seurat tool (29) was employed to process, cluster, and annotate cells, resulting in their classification into nine primary cell types. UMAP was employed to reduce the dimensionality of single-cell gene expression data, enabling visualization of cell distribution and ASS scores in two dimensions.

The single-cell ASS was calculated using the gsva() function from the GSVA R package (version 1.34.0) with parameters: method = “ssgsea”, kcdf = “Gaussian”. This algorithm treats each individual cell as an independent sample and quantifies the enrichment level of the four-gene diagnostic signature via the running-sum algorithm, which integrates gene rank order and coordinated expression patterns. The obtained single-cell ASS values were normalized by z-score to facilitate inter-cell comparison and visualization.

Dot plots were used to show the expression differences of diagnostic SRGs (GRN, MXD1, FOS, SOD2) and ASS scores among different celltypes. The dot size indicates the percentage of cells that express the gene, while the color shows the gene’s average expression level.

Statistical analyses were conducted using R software (v4.2.2). Based on the distribution characteristics of RNA-seq data and the prevalent analytical paradigms in the field of bioinformatics, we employed the Wilcoxon test for the analysis of different groups, followed by multiple testing correction. Spearman’s correlation test was used to calculate the correlation between different variables. Odds ratio (OR) and their 95% confidence intervals were reported. A p-value below 0.05 was deemed statistically significant. P-values were categorized as: not significant (ns), *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.

After integrating human AMI transcriptome cohorts (GSE34198↗, GSE48060↗, GSE60993↗) and correcting for batch effects, a total of 313 differential genes were identified by comparing AMI and normal samples (|FC| ≥ 1.2, P < 0.05; Figure 1A). GO and KEGG analyses indicated that the differentially expressed genes (DEGs) were predominantly associated with pathways related to senescence and inflammation, such as cell cycle arrest, inflammatory response, and TNF signaling pathway. This suggests a link between senescence, inflammation, and the development of AMI (Figure 1B). Gene co-expression network analysis was subsequently conducted using the DEGs. The network achieved a scale-free topology with a soft threshold power β of 16, indicated by R² exceeding 0.8 (Figure 1C), and 4 stable modules were identified (Figure 1D).The blue and teal modules exhibited the strongest correlation with AMI status, with correlation coefficients of 0.24 and 0.23, respectively (Figures 1E, F).By intersecting the DEGs, the hub genes of the two modules, and the senescence gene set from the CellAge database, 13 high-confidence senescence-related differential genes (SRGs) were identified (Figure 1G; Supplementary Table 4), and a heatmap showed the expression of these SRGs (Figure 1H).

To evaluate the discriminative ability of individual SRGs for AMI, we first conducted univariate logistic regression and plotted a forest plot to visualize the effect sizes. The results showed that higher log-normalized SOD2 expression was significantly associated with increased odds of AMI (OR = 3.410), indicating a positive association between SOD2 expression and the AMI phenotype (Figure 2A). To further screen SRGs with universal roles, four algorithms—XGBoost, Random Forest, Lasso, and GBM—were used to identify feature genes with potential diagnostic value for AMI (Figures 2B–E). Among these, FOS, SOD2, MXD1, and GRN were the intersecting genes of the four feature gene sets (Figure 2F) and were designated as diagnostic SRGs. We developed a diagnostic model using a multivariate logistic approach to assess the clinical significance of SRGs, determining each gene’s weight and calculating the AMI risk score for each sample. A nomogram visually represented the significance of each gene within the AMI diagnostic model (Figure 2G). Based on the above integrative analysis, four feature genes (FOS, SOD2, MXD1, and GRN) were ultimately identified, and a diagnostic model was constructed based on this four-gene signature. To validate the performance of the diagnostic model, we conducted a decision curve analysis (DCA). It was found that within a wide range of diagnostic risk score thresholds, the risk score curve was significantly higher than the “All” (all patients diagnosed as AMI) and “None” (all patients diagnosed as normal) curves, indicating that the model had obvious clinical net benefits and showed acceptable discriminatory ability for AMI, with high practical value in clinical diagnostic decision-making (Figure 2H). Furthermore, the model’s ROC curve demonstrated an AUC of 0.808 (Figure 2I), and these results suggest that the four-gene model has acceptable discriminatory performance and potential utility as a composite diagnostic tool for AMI. We also performed independent validation of the diagnostic model using the external independent dataset GSE29111↗. The results yielded an AUC value of 0.749 and CI of 0.611-0.886 in the external validation cohort, demonstrating favorable diagnostic performance (Supplementary Figure 2).

To further validate the diagnostic role of the diagnostic SRGs in AMI, we first confirmed that these 4 genes were consistently highly expressed in the AMI group in the training dataset (Figure 3A). We then used GSE29111↗ as a validation set and found that these genes were also highly expressed in the AMI group (Figure 3C). The ROC curves showed that all 4 genes exhibited good diagnostic performance and could effectively distinguish the AMI group from the control group (Figure 3B). Analysis of the AMI single-cell dataset GSE163129↗ revealed significantly elevated expression levels of these four genes in the AMI group (Figure 3D). These findings suggest that the identified SRGs may serve as candidate AMI-associated biomarkers with potential diagnostic relevance across multiple datasets.

To further validate the role of the four diagnostic SRGs in AMI, we first performed RNA-seq using infarct-zone cardiac tissues collected 24 h after LAD ligation in 8-week-old sham and MI mice. Differential expression analysis showed that Fos, Mxd1, Sod2, and Grn were upregulated in the MI group (Figure 4A). The heatmap further showed that these four SRGs could distinguish MI samples from sham controls (Figure 4B). Given the limited sample size of this young mouse RNA-seq cohort (Sham n = 3, MI n = 3), these results should be interpreted as exploratory and supportive transcriptomic evidence from a standard acute MI model.

To further address the age-related relevance of these findings, we additionally performed validation experiments in 18-month-old mice with MI and age-matched sham controls. qPCR analysis of peripheral blood leukocytes showed that Mxd1, Fos, Sod2, and Grn were significantly elevated in the aged MI group (Figures 4C–F). At the protein level, western blot analysis of infarct-zone cardiac tissues, together with densitometric quantification, confirmed the upregulation of GRN, c-FOS, MXD1, and SOD2 in aged MI mice (Figure 4G; Supplementary Figures 3A–D). These results provide additional support for the age-related relevance of the identified genes in an in vivo MI setting.

Furthermore, analysis of the human peripheral blood proteomic dataset (PXD062794) showed that SOD2 protein abundance was significantly increased in AMI patients (Figure 4H), and the ROC curve yielded an AUC of 0.774, supporting its potential diagnostic relevance (Figure 4I). However, among the four diagnostic SRGs, only SOD2 had available and analyzable protein expression data in the public plasma proteomic dataset PXD062794. Future studies based on larger-scale or self-established proteomic datasets are needed to verify the protein expression levels of the other three diagnostic SRGs in acute myocardial infarction. In addition, the qPCR and western blot results obtained from 8-week-old mice were retained as supplementary data (Supplementary Figures 3E–I), providing supportive evidence that the identified genes also showed concordant dysregulation in a standard acute injury model.

We utilized the ImmuCellAI algorithm to analyze the relationship between immune cell infiltration and risk scores derived from the SRG-based diagnostic model. Our findings indicated a higher prevalence of anti-inflammatory cells, like iTreg cells, in the low-risk group, while pro-inflammatory cells, including Th17 cells, macrophages, and neutrophils, were notably more abundant in the high-risk group. Neutrophil infiltration was significantly greater in the high-risk group than in the low-risk group (Supplementary Figure 4A). Analysis revealed a moderate positive correlation between the risk score and neutrophil infiltration (Spearman R = 0.456, P < 0.001), indicating significant recruitment and activation of neutrophils in high-risk patients. The risk score showed a positive correlation with macrophages, Tfh cells, and Th17 cells, while exhibiting a negative correlation with CD8⁺ T cells, exhausted T cells, iTreg cells, Th1 cells, Th2 cells, and NKT cells (Supplementary Figures 4B–K, all P < 0.001). Given the known involvement of neutrophils in acute inflammation and tissue injury (30), these findings suggest that higher risk scores are associated with increased neutrophil infiltration and a more pro-inflammatory immune microenvironment in AMI. This further validated the effectiveness of the SRG-based diagnostic model in characterizing differences in the immune microenvironment.

To investigate the role of SRGs in stratifying AMI patients for personalized MI treatment, we employed consensus clustering (Figure 5A). When the number of clusters k = 2 (Figure 5B), additional clusters yielded only marginal gains (Figure 5C). Together, these results strongly support k=2 as the statistically optimal number of clusters. The clustering results showed relatively high consistency and stability. Furthermore, since we found that diagnostic SRGs (FOS, MXD1, GRN and SOD2) were significantly highly expressed in the C2, we initially determined that the C2 was defined as a cluster with high senescence-related signature (Figure 5D). Subsequently, through DEGs analysis between C1 and C2, we obtained DEGs (|FC| > 1.5, P < 0.05) (Figure 5E), and intersected these genes with the SRGs to obtain six key genes (Figure 5F). All six key genes exhibited high expression levels within the C2 (Figure 5G), we concluded that the C2 exhibited high cellular senescence-related characteristics in AMI patients. We also found that the diagnostic SRGs (FOS, MXD1, GRN and SOD2) were also among the cluster-specific DEGs, which further illustrated the universality of these 4 genes. These results all supported the effectiveness of the diagnostic model based on SRGs, especially diagnostic SRGs, in identifying AMI clusters.

To deeply explore the immune cell infiltration and functional enrichment of AMI clusters, we performed immune infiltration analysis using ImmuCellAI algorithm. The C1 exhibited a high infiltration of anti-inflammatory cells, such as nTreg cells and iTreg cells, whereas the C2 showed a significant increase in proinflammatory cells, including Th17 cells, macrophages, and neutrophils. Neutrophil infiltration was notably higher in the C2 compared to the C1 (Figure 6A). The correlation heatmap showed that the correlation between SRGs and various immune cells was generally consistent in both C1 and C2. However, in the C2, the positive correlation between SRGs and some proinflammatory immune cells (e.g., neutrophils) was more significant (Figures 6B, C), indicating the immune response is crucial in AMI pathogenesis.

Functional enrichment analyses using GO and KEGG revealed that DEGs in the C1 are predominantly associated with anti-inflammatory pathways, including negative regulation of leukocyte activation, inflammatory response, TGF-β signaling, and Treg cell differentiation (Figures 6D, E), whereas the C2 DEGs were associated with proinflammatory pathways, including myeloid leukocyte activation, neutrophil activation, TNF signaling, and Toll-like receptor signaling pathways (Figures 6F, G). This suggests that the C2 shows increased inflammation and immune response traits in AMI patients, indicating that targeting cellular senescence could offer novel strategies for AMI prevention and treatment.

We first constructed the ASS based on the diagnostic SRGs and evaluated its diagnostic value in AMI. The ASS represents an alternative manifestation of the same gene set features as the previously constructed diagnostic model, and therefore exhibits a certain degree of similarity in the results. The ASS was significantly higher in AMI samples than in control samples (Figure 7A), and the ROC curve showed that it had high discriminative ability for AMI (Figure 7B). A comparative analysis of senescence-related gene expression between high and low ASS groups showed a significant upregulation of diagnostic SRGs (FOS, MXD1, GRN, and SOD2) in the high-score group (Figure 7C), suggesting that the ASS can effectively characterize the gene features related to senescence. Immune infiltration analysis showed that neutrophils, macrophages, and some T cell subsets were significantly enriched in samples with high ASS, while dendritic cells and B cells showed a decreasing trend (Figure 7D), suggesting a strong link between ASS and immune microenvironment activation. The correlation heatmap showed that the correlation between SRGs and various immune cells was generally consistent in both high and low ASS group (Supplementary Figures 5A, B). These results indicated that in high-score samples, inflammation-related signals were associated with stronger senescence-related transcriptional features, suggesting a close link between inflammatory activation and senescence-associated signatures in this group.

Single-cell transcriptome data of AMI was used and annotated to further analyze senescence in immune cells (Figure 7E), and the ASS scores were significantly higher in proinflammatory cell populations such as neutrophils and macrophages (Figure 7F). To corroborate these findings, we analyzed the distribution of ASS signature genes within the single-cell data. The findings indicated that the four core genes were predominantly located in neutrophils, aligning with the immune infiltration analysis trend (Figure 7G). These results suggested that, in AMI, senescence-related transcriptional signals are associated with neutrophil activation and pro-inflammatory responses, and that the ASS may reflect an immune microenvironment linked to both inflammatory activation and senescence-related features.