Identification and validation of efferocytosis-related biomarkers for the diagnosis of atherosclerosis based on bioinformatics and machine learning Authors: Xingyu Fu, Ao Yin, Chao Wang, Xinxin Liu, Min Li, Dan Liu, Xue Guan, Xiuru Guan Published: 2025 ABSTRACT AbstractBackgroundAtherosclerosis is a primary contributor to worldwide morbidity and mortality. Failure to timely clear apoptotic cells can trigger a cascade reaction, where the necrotic core expands until the fibrous cap is ruptured, and atherosclerotic plaques become vulnerable. Efferocytosis is an important method for recognizing and eliminating apoptotic cells. Nevertheless, the specific effect of efferocytosis on atherosclerosis remains uncertain. This study aimed to identify and verify the relevant characteristics of efferocytosis for detecting atherosclerosis.MethodsThe data of gene expression patterns of atherosclerosis were sourced from the Gene Expression Omnibus (GEO) database, and the differential expression analyses of efferocytosis-related genes (EFRGs) were performed between the atherosclerosis samples and the control samples. Subsequently, protein–protein interaction (PPI), correlation analysis, and functional enrichment analysis were performed to reveal the interaction between molecules as well as their pathways. Machine learning (ML) was employed to determine hub genes to construct a clinical prediction model. At the same time, immune infiltration, single-cell transcriptome analysis, and cell experiments were conducted in both atherosclerosis and control samples to provide a reference for the immune cell landscape and the cell heterogeneity under this condition.ResultsThe study revealed that 14 genes were closely related to efferocytosis in atherosclerosis. Among them, an ML model was used to screen 5 potential diagnostic biomarkers, including tumor necrosis factor (TNF), apolipoprotein E (ApoE), neutrophil cytosolic factor 1 (NCF1), triggering receptor expressed on myeloid cells 2 (TREM2), and chitinase-3 like-protein-1 (CHI3L1). Subsequent external validation indicated that, except for TNF, the other 4 genes were all upregulated. From the cell-type identification by estimating relative subsets of RNA transcripts (CIBERSORT) analysis, those 5 genes were all significantly associated with various immune cells. Further single-cell RNA sequencing (scRNA-seq) analysis demonstrated that those 5 genes were selectively upregulated in the macrophages of atherosclerosis lesions, which was supported by mRNA levels in cell experiments.ConclusionsThis study clarified the association between atherosclerosis and efferocytosis, and established an effective diagnostic model. Moreover, potential treatment targets for atherosclerosis were identified, offering new insights into the potential mechanism of atherosclerosis.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40001-025-03669-y.Keywords: Atherosclerosis, Efferocytosis, Bioinformatics, Machine learning Introduction Atherosclerosis serves as the chief pathological basis of cardiovascular diseases, such as coronary heart disease and stroke [1]. Its pathogenesis involves endothelial dysfunction, low-density lipoprotein (LDL) oxidation modification, monocyte recruitment, and macrophage polarization [2, 3]. Macrophages become foam cells after the uptake of oxidized LDL (ox-LDL) via scavenger receptors, release pro-inflammatory cytokines, and worsen the inflammation at the affected regions. With worsening of the lesion, the apoptosis of foam cells within the plaque increases. If not removed in time, it will expand the necrotic core and eventually form an unstable plaque, increasing the risk of rupture. Efferocytosis is a process that phagocytic cells (especially macrophages) identify, phagocytose, and eliminate apoptotic cells. This mechanism is critical for sustaining tissue homeostasis and inhibiting inflammatory responses [4]. Efferocytosis is accomplished by specialized phagocytes such as macrophages through the coordinated action of the "Find-Me", "Eat-Me", and "Don't Eat-Me" signaling systems [5]. This process relies on multiple molecular mechanisms, including signal recognition, bridging proteins, and intracellular degradation [6]. Nevertheless, as atherosclerosis is exacerbated by such factors as the intensified oxidative stress and the deteriorated pro-inflammatory microenvironment, the efficiency of efferocytosis declines accordingly. Impaired efferocytosis may lead to obstacles in clearing apoptotic cells, increase plaque vulnerability, and ultimately result in adverse cardiovascular diseases [7]. Over the past few decades, numerous studies have expounded on the significance of efferocytosis in various diseases. However, up to now, the role of efferocytosis in atherosclerosis remains unclear. Current articles mainly elaborate on the association between efferocytosis and atherosclerosis through in vivo and in vitro experiments, but there is still no bioinformatics analysis and machine learning (ML). Therefore, using bioinformatics analysis and ML to further investigate efferocytosis in atherosclerosis may provide new diagnostic biomarkers and treatment approaches. This study adopted systematic bioinformatics and ML approaches to identify the key biomarker genes of atherosclerosis. Subsequent immune infiltration and single-cell transcriptome analyses corroborated that the expression levels of those key genes in immune cells changed. To verify the accuracy of the findings, external data sets (GSE28829) and cell experiments were utilized for validation to better understand the potential molecular processes in the development of atherosclerosis. Therefore, the current research could reveal new targets for the detection and therapeutic interventions of atherosclerosis. Materials and methodologies Data source The data were sourced from the database of Gene Expression Omnibus (GEO) (official website: https://www.ncbi.nlm.nih.gov/geo/). Based on the platform GPL17077, the GSE100927 [8] data set included 69 atherosclerotic carotid artery plaques and 35 carotid arteries without atherosclerotic lesions, and was used as controls. The GSE28829 [9] data set comprised 13 early samples and 16 late samples of atherosclerotic plaques of human carotid arteries, and it served as the validation set. The samples in the GSE159677 [10] data set were obtained from 3 patients, including the carotid atherosclerotic plaques and the proximal adjacent parts that were matched with the patients. The data of single-cell RNA sequencing (scRNA-seq) in the GSE159677 were used to depict the cell groups that expressed key biomarkers in atherosclerosis. To collect the efferocytosis-related genes (EFRGs), the GeneCards database (official website: https://www.genecards.org/) was searched using the keyword “efferocytosis”. A total of 145 EFRGs were found. Construction of a weighted gene co-expression network analysis (WCGNA) This study performed a WGCNA utilizing the software R [11]. The samples of this study were hierarchically clustered, and abnormal samples were excluded. The soft powers with β = 8 were identified using the PickSoft Threshold function to construct a scale-free network. The adjacency matrix was established and then was transformed into a topological overlap matrix (TOM). A dendrogram of genes and the module colors were created through dissimilarity measurement. Analysis of differential expression genes (DEGs) In R (version 4.4.3), DEGs derived from the GSE100927 data set were determined utilizing the Limma package on normalized count data. |DEGs were identified based on Log2 FC|> 0.5 and p value < 0.05. A volcano plot of DEGs was constructed using the ggplot2 package. After intersecting DEGs with WGCNA and EFRGs, EFRGs in atherosclerosis were identified. The Venn diagram package and ggplot2 package were employed to generate Venn diagrams. Functional enrichment analysis After the identification of the EFRGs, a series of enrichment analyses were performed using the clusterProfiler package (version 4.10.1) [12], encompassing Gene Ontology (GO) enrichment analysis for biological process (BP), cellular component (CC), and molecular function (MF). Moreover, the enrichment analysis of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway was also conducted. The top-ranked results (p value < 0.05) were ultimately filtered and visualized. Construction of a protein–protein interaction (PPI) network The Search Tool for the Retrieval of Interacting Genes (STRING) data set was adopted to develop a PPI network [13]. The Cytoscape was employed to visualize these results. ML ML methods, such as least absolute shrinkage and selection operator (LASSO) [14], support vector machine–recursive feature elimination (SVM–RFE) [15], random forest (RF) [16], and logistic regression, were applied to screen for novel diagnostic biomarkers of atherosclerosis, to further select hub genes used for diagnosing atherosclerosis. From 14 atherosclerosis-related EFRGs, five hub genes were identified through the intersection of these algorithms. In the LASSO analysis, ‘alpha = 1’ and ‘lambda = NULL’ were utilized, while the ‘lambda.min’ was selected as the optimal parameter through fivefold cross-validation. For SVM–RFE, fivefold cross-validation was employed to evaluate different feature subsets, and the optimal feature was selected based on cross-validation error. The default parameter ‘mtry’ was used, and the ‘ntree’ value when the error curve became stable was employed for the final model. The univariate logistic regression analysis was conducted to screen for significant features with a p value less than 0.05 and a confidence interval (CI) not including 1. Construction of receiver operating characteristic (ROC) curve and box plot In GSE100927, the ROC curve was used to appraise the diagnostic value of hub genes through pROC (version 1.18.5). The ROC curve showed the area under the curve (AUC), its specificity, and sensitivity. For validating its diagnostic capability, the external data set (GSE28829) was used. Box plots were generated by the “ggplot2” in the R software to display the expression levels of hub genes in atherosclerosis samples and control samples. Construction and verification of comprehensive performance metrics A nomogram model was designed using the rms package (V6.8.0) to predict atherosclerosis. Each hub gene in the nomogram received a specific score, and subsequently, the total score was obtained by summing all the scores of the predictive factors involved. The calibration curve was utilized to appraise the predictive accuracy of the model. The pROC package (V1.18.5) was utilized to plot the ROC curves and calculate the AUC values with 95% CI, to evaluate the ability of the model to distinguish patients with atherosclerosis from healthy individuals across different thresholds. The rmda package (V1.6.0) was employed to generate the decision curves, demonstrating that the prediction model provided significant net benefit for patients in clinical practice. Assessment of immune infiltration The 22 types of immune cell genes (LM22) were used as the reference set. The cell-type identification by estimating relative subsets of RNA transcripts (CIBERSORT) (V0.1.0) package was employed to evaluate the relative abundance of immune cells (22 types) in the atherosclerosis group and the control group according to the expression profile of genes. The relative abundance of infiltrating immune cells was determined using a threshold of p value < 0.05 [17]. A correlation heatmap related to the infiltrating immune cells was plotted utilizing the package (corrplot) in the R software. Then, the Wilcoxon rank sum test was used to investigate the differences in the immune cell infiltration between the atherosclerosis group and the control group. Then, the Spearman correlation analysis was conducted between biomarkers and infiltrating immune cells. Ultimately, the results were visualized through ggstatsplot (version 0.9.1) in the R software. Data analysis of single-cell transcriptome The transcriptome data of atherosclerotic core (AC) plaques and patient-matched proximal adjacent (PA) tissue in the data set GSE159677 were combined. Those data were processed using the Seurat package (V4.0.0) through RStudio for single-cell analysis [18]. Functions, including RunPCA, FindClusters, and RunUMAP, were utilized to decrease the dimensionality of those data through principal component analysis (PCA), clustering analysis (CA), uniform manifold approximation and projection (UMAP). Subsequently, the UMAP plot generated by the DimPlot function was employed to visualize the cell clusters, and the violin plot was drawn using the VlnPlot function. Construction of a regulatory network of TF–miRNA–mRNA One regulatory network, including transcription factor (TF), microRNA (miRNA), and mRNA was established through NetworkAnalyst (official website: https://www.networkanalyst.ca/NetworkAnalyst/). The findings were visualized using Cytoscape (V3.10.2). Construction of an in vitro foam cell model THP-1 monocyte-derived macrophages were cultured in 10% cell culture medium at 37 °C with 5% CO₂. When they grew to the logarithmic phase with a density of 5 × 105/ml, they were treated with an additional 2% cell culture medium. Then, 0.1 mg/ml PMA was added to them. The mixture was placed in a cell culture plate for 48 h to induce THP-1 cells into macrophages. The macrophages were subsequently treated with ox-LDL (50 μg/ml) for 48 h to form foam cells. Oil Red O staining Oil Red O staining was used to observe the formation of lipid droplets, so as to validate whether the foam cell model had been constructed or not. Cell slides were placed in a 24-well plate, and THP-1 macrophages and foam cells were cultured as described above. After 48 h, the medium was discarded, and those cells were washed with phosphate-buffered saline (PBS). They were immobilized by paraformaldehyde for 10 min, treated with 60% isopropanol for 3–5 s, and then stained with Oil Red O for 10–20 min. Following staining, 60% isopropanol was used for differentiation (3–5 s) to clarify the cell matrix. The reaction was stopped with distilled water. After sealing with PBS, the slides were placed under a microscope to observe the formation of lipid droplets. Lipid droplet staining Foam cells and THP-1 macrophages were cultured according to the procedures above. The culture medium was discarded, and those cells were washed twice with PBS. They were digested with 400 μL trypsin and transferred to 1.5 ml light-protected EP tubes. After a centrifugation at 800 rpm for 5 min, the supernatant was discarded. Each tube was treated with 1 ml of 1 μg/ml BODIPY and stained at an indoor temperature for 20 min (dark condition). Following centrifugation, the supernatant was discarded, and then the cells were washed 2–3 times with PBS. Ultimately, after resuspending in 300 μL 1640 complete medium, the cells were evaluated by flow cytometry. RNA extraction and quantitative real-time polymerase chain reaction (qRT-PCR) THP-1 macrophages were cultured as described above. The expression of tumor necrosis factor (TNF), apolipoprotein E (ApoE), neutrophil cytosolic factor 1 (NCF1), triggering receptor expressed on myeloid cells 2 (TREM2), and chitinase-3-like protein-1 (CHI3L1) was measured by qRT-PCR. The total RNA was extracted from the cells that were treated with ox-LDL (50 μg/ml) for 48 h (the foam cell group) and the cells that were treated with 1640 medium for 48 h (the control group). PrimeScript™ RT Master Mix (Perfect Real Time) was employed to reverse transcribe the total extracted RNA into cDNA. Seven 2 × SYBR Green qPCR MasterMix Ⅱ, forward and reverse primers were used for two-step PCR amplification. GAPDH served as an internal reference for qRT-PCR of mRNA. The independent sample t tests were employed to verify significant differences. Western blot (WB) for protein expression The protein expression of TREM2 was assessed by standard WB. Primary antibodies were diluted as follows: TREM2 (1:1000; Cat#510482; ZhengNeng),;GAPDH (1:2000; Cat#60004-1-Ig; Proteintech). The secondary antibodies (goat anti-mouse IgG, Cat#S0002; goat anti-rabbit IgG, Cat#S0001; Affinity) conjugated with horseradish peroxidase (HRP) were diluted at a 1:10000 ratio to detect the primary antibodies. ImageJ was used to quantify all the band intensity, and relative protein expression levels were normalized. The full-length, uncropped original western blots are presented in Supplementary Material 2. Statistical analysis Data were displayed as mean ± standard deviation, and unpaired t tests were employed for comparative analyses. A two-tailed p value < 0.05 was defined as statistical significance. GraphPad Prism V10 and R 4.4.3 were employed for statistical analysis and data visualization. Results Identification of differentially expressed EFRGs Figure 1 illustrates the data analysis process. The differential expression analysis of EFRGs was conducted using the GSE100927 data set, and 441 DEGs in total were obtained. Among them, 321 genes were down-regulated, while 120 were up-regulated. The results are shown in a volcano plot (Fig. 2A). In addition, using WGCNA analysis with default recommended parameters, we selected the optimal soft threshold power of 2 based on a scale-free network (Fig. 2B). Five modules with significant co-expression were determined (Fig. 2C). According to the analysis of the correlation between modules and traits, the positive association between turquoise color and atherosclerosis was most significant (Fig. 2D). Next, DEGs and WGCNA, were interacted with EFRGs, and 14 differentially expressed EFRGs were eventually identified (Fig. 2E). Flowchart of the study Recognition of the differences in EFRGs between the atherosclerosis sample and the normal control sample. A Volcano plot of DEGs based on the GSE100927 data set. B Process of WGCNA (topological analysis of various soft threshold powers: the x-axis represents the soft threshold power; the y-axis represents the fitting index of scale-free topology model and average connectivity). C Clustering dendrogram of DEGs related to atherosclerosis based on topological overlap, and with specified module colors. D Module–trait correlation (a row means a module, a column means a feature, and each cell includes the corresponding correlations and the p values. It is coded based on the color legend according to the correlation). E Venn diagram of DEGs, WGCNA, and EFRGs Functional enrichment analysis of differentially expressed EFRGs To investigate the correlations between the 14 differentially expressed EFRGs, a gene correlation heatmap was drawn, and a PPI network was constructed (Fig. 3A, B). The 14 differentially expressed EFRGs were closely correlated, suggesting their potential influence on atherosclerosis. To deeper understand their biological significance, GO and KEG functional enrichment analyses were performed. The GO enrichment analysis indicated that in BP, pathways positively associated with cytokine production were upregulated, while pathways related to receptor-mediated endocytosis and inflammatory response were downregulated. In CC, these genes were significantly enriched in special granule tubes and cell membranes. In MF, activities, such as phosphatidylserine binding and protein tyrosine kinase binding, were upregulated, while activities, such as SH3-domain binding and amyloid-β protein binding, were downregulated. In addition, the KEGG enrichment analysis suggested that these genes were significantly enriched in the signaling pathway of mitogen-activated protein kinase (MAPK) and the interaction pathway of the cytokine–cytokine receptor, followed by phagosome, necroptosis, and efferocytosis (Fig. 3C). Functional enrichment analysis of differentially expressed EFRGs. A Correlation heatmap of differentially expressed EFRGs. B PPI network for differentially expressed EFRGs. C Bubble chart for GO enrichment analysis of differentially expressed EFRGs (bubble charts for BP, CC, MF, and KEGG included) ML for identifying hub genes To further screen diagnostic genes capable of distinguishing patients with atherosclerosis from healthy controls, ML algorithms were leveraged to further screen the 14 differentially expressed EFRGs. 13 candidate genes were determined by the SVM–RFE (Fig. 4C). In contrast, according to the LASSO regression algorithm, 6 of the 14 differentially expressed EFRGs were determined (Fig. 4A, B). Through the RF algorithm, 14 genes were chosen as key genes according to the feature weight Mean Decrease in Gini (MDG) (Fig. 4D, E). In addition, as shown in the forest plot for univariate logistic regression analysis, 14 genes significantly associated with atherosclerosis were identified, and all these genes were regarded as risk factors for atherosclerosis (all OR > 1, P < 0.05) (Fig. 4F). Based on the above results, a total of 5 hub genes, encompassing TNF, ApoE, NCF1, TREM2, and CHI3L1, were determined through cross algorithm (Fig. 4G). Those genes were promising diagnostic targets. Usage of ML to identify hub genes. A, B Identification of 6 genes through LASSO regression algorithm. C Identification of 13 genes through SVM–RFE algorithm. D Selection of differentially expressed EFRGs through RF algorithm. E Identification of 14 key genes through RF algorithm (MDG displays the hierarchy of genes based on their relative importance). F Forest plot for univariate logistic regression analysis of 14 differentially expressed EFRGs. OR > 1, and the entire confidence interval is to the right of 1 (excluding 1), with a p value of < 0.05. The result is statistically significant. G 5 hub genes identified by intersecting the algorithms of SVM–RFE, RF, LASSO, and logistic regression Expression of hub genes, analysis of ROC curve, and validation of hub genes In GSE100927, the expression levels of these genes were investigated between the atherosclerosis samples and the control samples. In the atherosclerosis samples, the expression levels of these genes were elevated (Fig. 5A). In addition, the AUC values of TNF, ApoE, NCF1, TREM2, and CHI3L1 were 0.957, 0.851, 0.943, 0.859, and 0.887, respectively. The AUC values for the 5 hub genes were all > 0.7, suggesting that they could be used as diagnostic markers (Fig. 5B). The samples in the validation set GSE28829 were categorized according to early and advanced stages of atherosclerosis. The results showed that in the advanced stage of atherosclerosis, the expression levels of the other four genes were higher, except for TNF (Figure S1A). The ROC curve confirmed their specificity and sensitivity in diagnosing atherosclerosis. Notably, in the GSE28829 data set, the AUC values for all genes, except TNF, were greater than 0.7 (Figure S1B). This phenomenon might be associated with the stage of atherosclerosis. Identification of AUC values of 5 hub genes through ROC curve and usage of box plot for heterogeneity and external validation. A Expression levels of hub genes in GSE100927 data set. B ROC curve analysis of hub genes in GSE100927 data set (* refers to p value < 0.05, ** refers to p value < 0.01, *** refers to p value < 0.001, **** refers to p value < 0.0001, and ns refers to p value ≥ 0.05) Construction and verification of comprehensive performance metrics To improve the prediction performance of hub genes, a comprehensive nomogram model was constructed based on CHI3L1, TREM2, NCF1, ApoE, and TNF. A distinct score was assigned to the corresponding biomarker (Fig. 6A). Through summarizing those scores, this model can predict the risk of atherosclerosis. Notably, TNF, CHI3L1 and TREM2 demonstrated outstanding diagnostic performance. Subsequently, a calibration curve was plotted to appraise the precision of the prediction probability. It can be observed that the model had good calibration in most prediction intervals (Fig. 6B). In addition, an ROC curve was supplemented to evaluate the accuracy of the model in differentiating atherosclerotic diseases. The AUC of this model was 0.991, suggesting that it was effective in clinically identifying high-risk populations (Figure S2A). The decision curve analysis (DCA) clearly demonstrated that across a probability range of [0–0.95], this prediction model provided clinically significant net benefit for patients, thereby strongly supporting recommending interventions at specific risk thresholds (Figure S2B). Construction and verification of comprehensive performance metrics. A Nomogram for diagnosing atherosclerosis based on hub genes. B Calibration curve for assessing the predictive performance of the model in diagnosis (ideal curve: the scenario, where predicted probabilities perfectly match observed probabilities; apparent curve: the relationship between the predicted probability and the observed probability when the model is not calibrated; bias-corrected curve: the relationship between the predicted probabilities and the observed probabilities after the model is not calibrated) Immune cell infiltration analysis To clarify whether efferocytosis promoted the progression of atherosclerosis by regulating the immune cell infiltration, a CIBERSORT analysis was implemented to ascertain the distribution proportions of immune cells in GSE100927 (Figure S3). We attempted to ascertain the link between key hub genes in atherosclerosis and infiltrating immune cells. According to the results from the correlation analyses, TNF, ApoE, NCF1, TREM2, and CHI3L1 were positively correlated with memory B cells, γ-δ T cells, and M0 macrophages. They were negatively correlated with naive B cells, plasma cells, resting memory T cells, and M2 macrophages (Fig. 7A). In addition, through the analysis of the proportions of 22 different immune cell types, considerable differences in the expression of 16 immune cell subpopulations were observed (Fig. 7B). For instance, the proportions of memory B cells, M0 macrophages, and activated mast cells in the atherosclerosis group were remarkably elevated compared to the control group. In contrast, the proportions of M1, M2 macrophages, and activated or resting CD4 T cells in the atherosclerosis group were remarkably lower than those in the control group. Usage of immune cell infiltration analysis to identify the heterogeneity of immune cells and analysis of the correlation with genes. A Correlation of TNF, ApoE, NCF1, TREM2, and CHI3L1 with infiltrating immune cells. B Box plot of the distinctions in immune infiltration between the atherosclerosis samples and the control samples Validation by scRNA-seq data set To more accurately depict the expression of key features in immune cells, the scRNA-seq GSE159677 database was utilized to determine the cell clusters that expressed the 5 hub genes in atherosclerosis. The distributions of these key genes among the 7 integrated cell clusters were identified (Fig. 8A). After the quality control and cell filtration, unbiased clustering was performed on the cells according to their gene expression patterns, as shown in the UMAP plot (Fig. 8B). UMAP was used for visualization. It was found that TNF, ApoE, NCF1, TREM2, and CHI3L1 were all expressed in M1 and M2 macrophages. Notably, CHI3L1 expression was inapparent. TNF and NCF1 were notably significantly expressed within T and B lymphocytes. In addition, ApoE exhibited significant expression in vascular smooth muscle cells. However, the expression of TREM2 was more concentrated, to some extent reflecting its higher specificity. Subsequently, the expression levels of those genes were evaluated in the PA group (Control) and the AC group (Atherosclerosis), and the obtained results were consistent with the previous analyses (Fig. 8C). Usage of the Gse159677 data set for the single-cell analysis of 5 hub genes. A UMAP plot (the 7 identified cell clusters of the scRNA-seq data set GSE159677). B UMAP visualization results of 5 hub genes. C Violin plots of the expression levels of patients’ 5 hub genes in the PA group and AC group in the GSE159677 database Construction of a TF–miRNA–mRNA network Figure 9 depicts the underlying transcription factors and miRNAs regulating the expression of those hub genes. Three transcription factors, including SPI1, SP1, and JUN, and one miRNA (miR-665), were determined. They were crucial for regulating the expression of the hub genes. In general, the influence of those regulatory elements and hub genes on atherosclerosis needs to be further investigated. Construction of a TF–miRNA–mRNA network Verification of cellular experiments Oil Red O staining and lipid droplet staining disclosed significant lipid accumulation in the foam cells, indicating the successful construction of the in vitro foam cell model (Fig. 10A, B). To ascertain the role of EFRGs in atherosclerosis, qRT-PCR was employed to assess the mRNA levels of key hub genes. The analysis demonstrated that compared to the control group, the mRNA levels of TNF, ApoE, NCF1, TREM2, and CHI3L1 genes in the foam cell group (ox-LDL-treated group) were all upregulated, consistent with the previous results (Fig. 10C). The scRNA-seq analysis revealed that TREM2 was highly and specifically expressed within specific macrophage subpopulations in atherosclerotic plaques, suggesting its potential key role in the local pathological microenvironment. The differences in TREM2 expression at the mRNA level were confirmed by qRT-PCR. To verify whether this transcriptional change was concurrently reflected at the protein level, quantitative analysis was subsequently performed on the protein expression of TREM2 through the Western blot (Fig. 10D). These results emphasized a significant role of TREM2 in the pathogenesis and progression of atherosclerosis. Construction of foam cell model and verification of hub genes. A Oil Red O staining of the foam cell group and the control group. B Average fluorescence intensity ratio of the foam cell group and the control group in lipid droplet staining. C mRNA levels of 5 hub genes in the foam cell group and the control group (normalized by GAPDH). D Relative protein level of TREM2 in the foam cell group and the control group (normalized by GAPDH) (** represents p value < 0.01, *** represents p value < 0.001, and **** represents p value < 0.0001) Discussion For preventing and treating atherosclerotic cardiovascular diseases, early identification, preventive intervention, and timely treatment play a decisive role in reducing the mortality and incidence rates, thus alleviating its significant social and economic impacts [19, 20]. Systematically screening susceptibility markers and analyzing their mechanisms in atherosclerosis pathogenesis can provide a scientific foundation for precise prevention and personalized treatment for atherosclerosis. The efferocytosis mediated by macrophages is a critical physiological mechanism for eliminating apoptotic cells. It has become a research hotspot due to its key role in various pathological conditions, such as cardiovascular diseases, ischemic stroke, and malignant tumors [21–23]. Notably, the specific molecular mechanism of efferocytosis during the pathological process of atherosclerosis and its therapeutic effect have not been fully elaborated. Therefore, this study intends to systematically analyze and screen EFRGs, aiming to evaluate their diagnostic efficacy and prognostic predictive value in atherosclerosis, and reveal possible molecular intervention targets. The BP analysis of the GO enrichment analysis indicated that these genes were negatively correlated with the receptor-mediated endocytosis pathway and impaired efferocytosis. This damage might be due to the abnormal function of endocytosis-related proteins or receptors when macrophages fail to effectively recognize and engulf apoptotic cells [24–26]. When apoptotic cells are not promptly cleared and gradually accumulate, they can induce the release of dangerous signals, such as pro-inflammatory cytokines [27]. This situation corresponds to the positive correlation with the regulatory pathways involved in cytokine production. Pro-inflammatory cytokines play a crucial role in the pathogenesis and progression of atherosclerosis by promoting endothelial activation, macrophage infiltration, foam cell formation, and plaque instability [28–30]. The upregulation of activities such as phosphatidylserine binding and protein tyrosine kinase binding in MF analysis is usually closely related to the regulation of immune cell function, inflammatory signal transduction, and efferocytosis [31, 32]. The SH3 domain proteins and amyloid-β protein (Aβ) are closely associated with the regulation of inflammation and cell phagocytosis [33, 34]. This suggests that efferocytosis participates in the pathological progression of atherosclerosis through multiple pathways and is influenced by immune cells, cytokines, and the body’s microenvironment. The molecules involved in efferocytosis may contribute to this process, but the specific mechanism still needs to be further elaborated. This study identified 14 EFRGs significantly associated with atherosclerosis. In medical research, ML is increasingly being used for accurate diagnosis, prognosis, and prediction for treatment outcomes [35–38]. Therefore, in this study, LASSO, SVM–RFE, RF, and logistic algorithms were utilized to detect the genes crucial for the diagnosis of atherosclerosis. A total of 5 hub genes, including TNF, ApoE, NCF1, TREM2, and CHI3L1, were determined. It has been reported that hub genes with an AUC > 0.7 are considered helpful to diagnose diseases [39]. Subsequently, during the validation using external data sets and cell experiments, the TNF with the highest AUC value in the GSE100927 data set turned out to have the lowest value (< 0.7) in the external validation. The other hub genes all demonstrated good diagnostic performance. This difference in the results might be due to the different effects of cytokine secretion in different stages of atherosclerosis and the different mechanisms mediated by immune cells and cytokines [40, 41]. Therefore, in the GSE100927 data set, significant differences in the expression of TNF were observed between normal individuals and atherosclerosis patients. Nevertheless, in the external data set GSE28829, the effects and participation mechanisms of TNF in the early and late stages of atherosclerosis were different. Hence, the results were not significant. From a functional perspective, the 3'UTR region of TNF mRNA maintains stable mRNA levels in both early and late stages of atherosclerosis through a dynamic balance between TTP (promoting degradation) and HuR (promoting stabilization) [42, 43]. Mechanistically, during the stage of lipid streak formation, TNF in endothelial cells primarily rapidly activates the NF-κB pathway, inducing transient expression of endothelial adhesion molecules and activating the inflammatory microenvironment. Moreover, TNF in macrophages, by synergizing with ox-LDL via toll-like receptors, initiates M1 polarization of macrophages, enhances CD36 expression, and stimulates foam cell formation [44–46]. In the advanced stage, the necrotic core expands. TNF upregulates matrix metalloproteinases via the NF-κB pathway, compromising plaque structural stability. Furthermore, TNF activates the caspase-8 pathway via TNFR1, inducing apoptosis of smooth muscle cells in the fibrous cap and necrosis of macrophages [47, 48]. This indicates that the stage-specific functions of TNF rely primarily on downstream co-activators rather than its own expression level. In addition, as the disease progresses, some cells grow and others decline. The expression of TNF at different stages of the disease may not change linearly, which may also lead to no difference in its overall expression. These findings may also be related to various factors, such as limited sample size of the GSE28829 data set, technical limitations of the detection methods, plaque heterogeneity, dynamic changes in TNF expression, and distinct mechanisms of functional compensation. The calibration curve confirmed that this model could predict the risk of atherosclerosis. In conclusion, there may be potential limitations in identifying core genes by employing multiple ML methods. In the future, multi-dimensional cross-methods should be adopted to verify this model. Atherosclerosis serves as a chronic inflammatory disease of the vascular wall, and the stability of plaques is associated with the levels of inflammatory cells and the thickness of the fibrous cap. A thin-cap plaque filled with immune cells is called a fragile plaque and is more prone to thrombosis [49]. Given the importance of immunity in atherosclerosis, we attempted to explore the association between the genetic characteristics of atherosclerosis and immune cells. By employing the CIBERSORT algorithm and single-cell RNA sequencing to quantify immune cell proportions and precisely identify functional cellular subsets, it may be more effective to investigate the specific mechanisms linking impaired efferocytosis to plaque vulnerability throughout the progression of atherosclerosis [50, 51]. This study employed the CIBERSORT to investigate the infiltration of immune cells in atherosclerosis. Correlation analysis demonstrated that in comparison with the control samples, the proportion of M0 macrophages in atherosclerosis samples increased, while the proportions of M1 and M2 macrophages decreased. This result reflects the dynamic plasticity and functional heterogeneity of macrophages in the atherosclerosis microenvironment. That is, early ox-LDL is phagocytosed by macrophages in large quantities through scavenger receptors, and at this time, macrophages mainly present a pro-inflammatory M1 phenotype [52]. As the disease progresses, the phenotypic transformation during the inflammatory resolution period and the reprogramming of the lipid microenvironment cause macrophages to be in an intermediate state of the transformation from M1 to M2, and some M1 macrophages continue to increase but are functionally dysregulated [53, 54]. At the end stage of the disease, the barrier of efferocytosis is impaired, and lipotoxic apoptosis or necrosis of foam cells occurs, ultimately contributing to a reduction in the number of M1 and M2 [55, 56]. This is consistent with previous research conclusions [57]. This study also confirmed an association between hub genes and various types of immune cells. Besides, scRNA-seq analysis also suggested different expression patterns of 5 hub genes in specific immune cell subsets, including vascular endothelial cells, T cells, and B cells. These findings emphasize that the immune system is indispensable in the pathological process of atherosclerosis. Furthermore, these hub genes are crucial in shaping the immune microenvironment of patients with atherosclerosis and may also be critical in regulating the atherosclerosis pathological pathway, especially in affecting the fibrous caps and lipid core of the plaques by regulating efferocytosis. Through various ML algorithms and experimental verification, it can be proved that the 5 hub genes have excellent diagnostic performance, and the research model can predict atherosclerosis. Since ApoE, TNF, and NCF1 are common biomarkers for atherosclerosis and have been extensively studied, we will not deeply discuss them here to avoid redundancy. ROC curve confirmed that all five hub genes exhibited favourable diagnostic performance. The scRNA-seq analysis revealed that TREM2 was highly expressed and significant in M1 and M2 macrophages. Notably, TREM2 is an emerging biomarker in recent years and is relatively popular and controversial in atherosclerosis-related studies. TREM2 is an immunoregulatory receptor highly expressed in myeloid cells (including macrophages, microglia, and osteoclasts), belonging to the immunoglobulin superfamily [58]. It binds to the adaptor protein DAP12 to regulate cell behavior through signal pathways, such as SYK–PI3K–AKT, and recognize apoptotic cells by identifying exposed phospholipids [59, 60]. Numerous studies have elucidated that TREM2 participates in the pathological physiological processes of various diseases by inhibiting inflammation and promoting phagocytosis [61–64]. Nevertheless, the correlation between TREM2 and atherosclerosis remains unclear. Some studies have shown that when TREM2 is overexpressed in vascular smooth muscle cells and macrophages, it upregulates scavenger receptors, accelerating the uptake of ox-LDL, causing lipid accumulation and the formation of foam cells [65]. In addition, soluble TREM2 (sTREM2) is activated by NF-κB signaling and promotes the inflammatory cytokine production and the morphological changes of microglia [66]. While other studies have shown that inflammation activation can cause the instability of plaques, and TREM2 may inhibit macrophage inflammatory activation by regulating cytokines, macrophage polarization, and inflammatory bodies, maintaining plaque stability [59, 67, 68]. Further studies have confirmed that in cardiac-specific sepsis injury, TREM2hi macrophages maintain mitochondrial homeostasis, indirectly protecting cardiomyocytes [69]. In addition, using the agonistic TREM2 antibody 4D9 to treat mice, it was found that the accumulated dead cells in the necrotic core decreased [70]. This suggests that TREM2 plays a pivotal role in maintaining the balance between foam cell death and cell clearance and controlling the necrosis of plaques in atherosclerosis lesions. Moreover, many studies have shown that the impact of TREM2 on the disease at different stages of atherosclerosis may lead to different outcomes. In the lipid streak formation stage, TREM2 directly drives plaque growth through CD36-dependent lipid uptake, promoting the uptake of lipids and increasing the size of lesions [65]. In the advanced plaques, the TREM2low foam macrophage subpopulation in the plaque is highly sensitive to ferroptosis. Enhancing the TREM2 signal can block mitochondrial damage mediated by HMOX1–LDHB, reducing cell death caused by lipid peroxidation [71]. In addition, TREM2 clears apoptotic foam cells around the necrotic core by recognizing phosphatidylserine released by apoptotic cells. This suggests that TREM2 exerts dual effects in atherosclerosis. Our results indicate that TREM2 is closely associated with efferocytosis and is selectively upregulated in atherosclerosis lesions in macrophages. This finding is subsequently verified in in vitro cell models, suggesting that TREM2 has good diagnostic performance for atherosclerosis. In summary, TREM2 can be regarded as a new targeted biomarker for atherosclerosis. However, how to transform TREM2 into a more beneficial form is a new issue in current clinical research. Since the hub genes related to efferocytosis in atherosclerosis have been identified, it is essential to determine the regulating factors controlling those genes, thereby providing guidance for future studies. Therefore, a TF–miRNA–mRNA regulatory network was established, emphasizing that SPI1, SP1, and JUN were key transcription factors, and miR-665 was an important regulatory factor. SPI1 is a myeloid cell-specific transcription factor that can regulate macrophage polarization and participate in promoting macrophage glycolysis through the SHP2/SPI1 axis, and is involved in metabolic reprogramming [72, 73]. In addition, overexpression of SPI1 can weaken the efferocytosis of macrophages, leading to difficulty in clearing apoptotic cells and expanding the necrotic core of the plaques [74]. SP1 is a ubiquitous zinc finger transcription factor. It promotes the activation of inflammasome and the releasing of pro-inflammatory factors by binding to the promoter of target genes, exacerbating vascular endothelial damage [75]. Moreover, SP1 is capable of inhibiting the expression of mitochondrial protein PHB1, decreasing the mitochondrial membrane potential, accumulating the ROS, and inducing macrophage apoptosis and lipid deposition [75]. JUN is the core component of the AP-1 transcription factor and forms a dimeric complex with FOS, promoting extracellular matrix degradation and monocyte infiltration [76, 77]. In summary, the above 3 transcription factors are closely correlated, but further research is still needed to understand how they regulate atherosclerosis by influencing the hub genes. miRNA is a small non-coding RNA that acts as a post-transcriptional regulatory factor for protein-coding genes [78]. The discovered miR-665 is considered a promising target for the prevention and treatment of atherosclerosis [79]. This miRNA is a component of the complex regulatory network involving TF–miRNA–mRNA interactions, highlighting the complex molecular interactions underlying the pathogenesis of atherosclerosis. Understanding this regulatory network will provide valuable insights for interventions and treatments. Previous research has mainly focused on exploring the specific mechanisms between atherosclerosis and efferocytosis through in vivo and in vitro experiments. The novelty of this study lies in its focus on the underexplored perspective of atherosclerosis and efferocytosis. By utilizing bioinformatic analyses and ML approaches, we assessed the practical utility and clinical value of this prediction model from multiple dimensions and identified novel diagnostic biomarkers, thereby establishing a theoretical foundation for future clinical practice. Nonetheless, certain limitations are observed in this study. Due to constraints in the study period and experimental resources, in vivo experiments could not be conducted, nor were new clinical samples obtained for validation. In addition, the absence of detailed clinical information, such as medication history of patients and comorbidities, may potentially influence the performance of the model. For instance, the widely prescribed statins, known for their anti-inflammatory properties, could lead to a significant underestimation of M1 or M2 macrophage ratios and overall inflammatory activity. Meanwhile, inflammatory responses resulting from varying comorbidities of patients (such as diabetes mellitus) may introduce unquantified confounding factors, thereby potentially causing the generalization ability of this model to be either overestimated or underestimated. To address these deficiencies and translate the current findings into practical clinical value, a subsequent verification strategy is formulated for future research. Future research should establish an ApoE-/- mouse atherosclerosis model, and directly intervene in the core target through gene overexpression/knockdown techniques in the living body. In this way, alterations in plaque burden can be observed, and immune cells within plaques can be precisely quantified via flow cytometry. To tackle the challenges in clinical translation, samples of peripheral blood from clinical patients can be collected to measure the expression levels of core targets, evaluating their potential as biomarkers for non-invasive diagnosis. Furthermore, small-molecule agonists or inhibitors targeting these molecules should be developed, laying the groundwork for subsequent drug development. It is expected that these foundational discoveries can ultimately be advanced into clinical practice, providing new strategies for a precise diagnosis and treatment of atherosclerosis. Conclusion Through comprehensive bioinformatics methods, this study identified key EFRGs related to atherosclerosis, providing new targets for predicting the progression of atherosclerosis plaques. TREM2 has promising diagnostic values for efferocytosis in atherosclerosis. However, it still remains necessary to combine other factors to formulate effective strategies for the prevention of cardiovascular diseases. Supplementary Information Supplementary Material 1.Figure S1. Expression and ROC Curve Analysis of Hub Genes in GSE28829 Dataset (Validation Set). A Expression Levels of Hub Genes in GSE28829 Dataset. B ROC Curve Analysis of Hub Genes in GSE28829 Dataset(* refers to P-value < 0.05, ** refers to P-value < 0.01, *** refers to P-value < 0.001, **** refers to P-value < 0.0001, ns refers to P-value ≥ 0.05). Figures S2. ROC Curve and DCA Curve Construction. A ROC curve analysis was used to evaluate the accuracy of the prediction model in differentiating atherosclerotic diseases. The AUC = 0.991 (95% CI: 0.980–1.000) indicates that the prediction model has a high discriminatory ability. B DCA reveals that across the threshold probability range from 0.05 to 0.95, the predictive model yields positive net benefits. The predictive model's net benefit value, ranging from 0.487 to 0.658, suggests excellent clinical utility. In the same threshold probability range, the models for TNF, APOE, NCF1, TREM2, and CHI3L1 exhibit net benefit values ranging from 0.311 to 0.650, −0.026 to 0.647, 0.183 to 0.649, 0.154 to 0.626, and 0.212 to 0.648, respectively. Figures S3. Distribution of the Proportion of Immune Cell Types (The height of the bar stands for the proportion of each cell type in the corresponding sample, and different colors indicate a specific type of cell.) Supplementary Material 2. Acknowledgements Not applicable. Author contributions FXY performed the conceptualization, methodology and Software in this research. YA finished the data curation and original draft writing. WC did the visualization and investigation work. LXX performed the supervision work. LM performed the software and validation of this work. LD performed formal analysis. GX performed resources and project administration of this study. GXR performed the study writing, Reviewing and Editing. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Funding Not applicable. Data availability All the data were obtained from Gene Expression Omnibus (GEO) (official website: https://www.ncbi.nlm.nih.gov/geo/). Declarations Ethics approval and consent to participate GEO belongs to public databases. The patients involved in the database have obtained ethical approval. Users can download relevant data for free for research and publish relevant articles. This study is based on opensource data, so there are no ethical issues. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1.Figure S1. Expression and ROC Curve Analysis of Hub Genes in GSE28829 Dataset (Validation Set). A Expression Levels of Hub Genes in GSE28829 Dataset. B ROC Curve Analysis of Hub Genes in GSE28829 Dataset(* refers to P-value < 0.05, ** refers to P-value < 0.01, *** refers to P-value < 0.001, **** refers to P-value < 0.0001, ns refers to P-value ≥ 0.05). Figures S2. ROC Curve and DCA Curve Construction. A ROC curve analysis was used to evaluate the accuracy of the prediction model in differentiating atherosclerotic diseases. The AUC = 0.991 (95% CI: 0.980–1.000) indicates that the prediction model has a high discriminatory ability. B DCA reveals that across the threshold probability range from 0.05 to 0.95, the predictive model yields positive net benefits. The predictive model's net benefit value, ranging from 0.487 to 0.658, suggests excellent clinical utility. In the same threshold probability range, the models for TNF, APOE, NCF1, TREM2, and CHI3L1 exhibit net benefit values ranging from 0.311 to 0.650, −0.026 to 0.647, 0.183 to 0.649, 0.154 to 0.626, and 0.212 to 0.648, respectively. Figures S3. Distribution of the Proportion of Immune Cell Types (The height of the bar stands for the proportion of each cell type in the corresponding sample, and different colors indicate a specific type of cell.) Supplementary Material 2. Data Availability Statement All the data were obtained from Gene Expression Omnibus (GEO) (official website: https://www.ncbi.nlm.nih.gov/geo/).