Atherosclerotic plaque vulnerability quantification system for clinical and biological interpretability

Summary

Acute myocardial infarction dominates coronary artery disease mortality. Identifying bio-signatures for plaque destabilization and rupture is important for preventing the transition from coronary stability to instability and the occurrence of thrombosis events. This computational systems biology study enrolled 2,235 samples from 22 independent bulks cohorts and 14 samples from two single-cell cohorts. A machine-learning integrative program containing nine learners was developed to generate a warning classifier linked to atherosclerotic plaque vulnerability signature (APVS). The classifier displays the reliable performance and robustness for distinguishing ST-elevation myocardial infarction from chronic coronary syndrome at presentation, and revealed higher accuracy to 33 pathogenic biomarkers. We also developed an APVS-based quantification system (APVSLevel) for comprehensively quantifying atherosclerotic plaque vulnerability, empowering early-warning capabilities, and accurate assessment of atherosclerosis severity. It unraveled the multidimensional dysregulated mechanisms at high resolution. This study provides a potential tool for macro-level differential diagnosis and evaluation of subtle genetic pathological changes in atherosclerosis.

Graphical abstract

Highlights

• •AI-powered atherosclerotic plaque vulnerable bio-signature (APVS)
• •APVS classifier accurately distinguishes between different pathological conditions
• •APVSLevel system unravel multidimensional dysregulated mechanism at high resolution

Cardiovascular medicine; Risk factor; Computational bioinformatics; Complex system biology

Introduction

Coronary artery disease (CAD) is the predominant cause of mortality worldwide,¹ which is often divided into chronic coronary syndrome (CCS) and acute coronary syndromes (ACS).² While the former is a clinically stable condition primarily caused by transient and reversible episodes of blood supply-demand mismatch, the latter clinically manifests as ST-elevation myocardial infarction (STEMI), non-ST-elevation myocardial infarction (NSTEMI), and unstable angina pectoris.³ Acute myocardial infarction (AMI) caused by unstable plaques significantly contributes to CAD-induced deaths, causing public health burdens clinically and economically.^4,5

To reduce CAD-related mortality, timely detection is the key, as each hour of prompt diagnosis and therapy could increase the survival rate by approximately 20%.⁶ The mechanism responsible for the sudden conversion of a stable to an unstable situation is typically plaque disruption, which may suddenly result in life-threatening coronary thrombosis.⁷ An effective biomarker signature is still necessary for monitoring plaque destabilization.

When clinical suspicion of CAD instability arises, biochemical markers of myocardial injury, such as cardiac troponin should be measured.² However, it is primarily indicative of myocardial injury rather than early myocardial perfusion abnormalities.⁸ Many pathologies other than AMI also result in troponin elevations, such as tachyarrhythmias, heart failure, hypertensive emergencies, and aortic dissection, leading to the false prediction.⁹ In elderly patients with renal dysfunction, CCS is often the most important contributor to troponin elevations.¹⁰ Moreover, AMI diagnosis based on unstable electrocardiogram patterns could be confounded by comorbid cardiomyopathies. Although several biomarkers, including exosomal RNAs and creatine kinase have been upregulated in AMI compared to CCS or health state, their diagnostic accuracy remains suboptimal.^11,12,13 Furthermore, many CAD risk prediction models were largely reliant on epidemiological data to perform macroscopic stratification and diagnosis, may not fully reflecting to incorporate the individual biological status of patients.^14,15,16

This study employed computational systems biology approaches to decode the transcriptome landscapes of patients across varying severity of atherosclerosis at bulk and single-cell levels. Using a machine learning (ML) program integrating nine learners, we identified an atherosclerotic plaque vulnerability biomarker signature (APVS) to construct a robust classifier. This classifier performed well in discriminating between STEMI and CCS, STEMI and healthy individual at presentation. We also established an APVS-based quantification system (APVSLevel) that could evaluate atherosclerosis severity from clinical and biological perspectives.

Results

Uncovering dysregulated gene co-expression pattern aggravating CCS

The study design was illustrated in Figure 1A. The weighted gene co-expression network analysis (WGCNA) was performed based on the cohort GSE28829. To meet the scale-free network for the biological hypothesis, we calculated the gene correlation matrix. The no scale R2 achieved 0.867 with the soft threshold β set to 12, which became the power of our adjacency matrix (Figure 1B). The topological overlap matrix (TOM) was next transformed from the adjacency matrix to make the segmentation of modules easier (Figure 1C). As a result, we identified 13 co-expression modules through hierarchical clustering (Figure S1A), among which four relevant modules (n = 2108) were highest involved in the human atherosclerosis progression from the early to the advanced stage (Figure 1D). Figure 1E confirmed the close associations among genes, module membership, and the deterioration of the disease. To seek the hub genes from the top four relevant modules, we filtered out those not significantly dysregulated between acute STEMI and CCS patients based on the cohort GSE59867 (Figure S1B), and identified 96 genes as the dysregulated co-expression pattern genes (DCPGs) (Figure S1C; Table S1). A comprehensive biological annotation applying all MSigDB datasets showed a prevailing association between DCPGs and immune-inflammatory activation and cellular differentiation (Figure S1D; Table S2).

Figure 1.

Uncovering dysregulated gene co-expression pattern

(A) Overview of the study methodology.

(B) Scale-free topological indices at various soft-thresholding powers (left); correlation analysis between the soft-thresholding powers and mean connectivity of the network. Setting the soft power of β to 12 resulted in a high scale-free R2 (>0.8), indicating a good fit to the scale-free topology model, and the stabilization of mean connectivity.

(C) Heatmap depicting the topological overlap matrix among all genes in WGCNA.

(D) Correlation between module eigengenes and the progression of early-stage to advanced-stage atherosclerosis.

(E) The significant associations among gene significances, module memberships, and the disease severity in the top four modules (light yellow: correlation coefficients (R) = 0.83, p value = 3e-08; dark red: correlation coefficients (R) = 0.69, p = 4e-05; magenta: correlation coefficients (R) = −0.71, p = 2e-05; grey60: correlation coefficients (R) = −0.47, p = 0.01).

Machine learning-based integrative program generates APVS

DCPGs were incorporated into our ML-based integrative program to establish an atherosclerosis plaque vulnerable signature (APVS) (Figure 2A). We employed nine classical learners via 10-fold cross-validation with 100 repetitions to fit models and evaluated their performance using five metrics, including the accuracy, C-index, F1-score, precision, recall, and root-mean-square of the residuals (RMSR), in the testing cohort. The random forest and back propagation neural network demonstrated superior classification capabilities based on these performance metrics (Figure 2B; Table S3). Next, we performed the recursive feature elimination cross-validation integrated with the random forest-based Gini coefficient method (Figure 2C), and 14 genes with the highest importance were selected as APVS gene panel (Figures 2D and 2E; Table S4).

Figure 2.

Machine learning-based integrative program generates APVS

(A) Schematic overview of the procedure for APVS development.

(B) Comprehensive performances of nine types of learners. Boxplots depicted the distribution of the residuals, with red highlighted dots representing root-mean-square of residuals (RMSR). Circles showed the distribution of recall, precision, F1-score, C-index, and accuracy of each learner.

(C) Recursive feature elimination was performed for 96 DCPGs using 10-fold 10-repeated cross-validation combined with decreasing accuracy method (random forest-based Gini coefficient), reducing the dimensionality of the feature space and avoiding overfitting. The error was minimized when the number of variables set to 14.

(D) Variable importance of DCPGs. Barplot showed the distribution of mean decreasing Gini coefficient, while line plot showed the mean decreasing accuracy. The top 14 genes were identified as APVS panel.

(E) Chromosomal positions of 14 APVS genes.

(F) Principle coordinate analysis of Bray-Curtis dissimilarities obtained for the APVS gene expression profiles in the GSE59867 (open circles) and GSE62646 (closed circles) cohorts. The circles and error bars indicated the mean and standard errors of the mean.

(G) Kinetics of dynamic expression was evaluated by fuzzy c-means clustering analysis using mfuzz, where Cluster2 and Cluster4 represented the fluctuation pattern of APVS. Temporal expression profiles were plotted with high membership values in warm colors (red and yellow), and lower membership values in cool colors (green and blue).

(H) Significantly enriched terms of mfuzz Cluster2 and Cluster4from Biological Processes (green), Reactome signalings (blue), Wikipathway (red).

Principal coordinates analysis (PCoA) of APVS gene expression profile in the GSE59867 cohort showed that significant shifts separated atherosclerotic patients with CCS from that with different severities (STEMI 1D: the first day of STEMI; 4–6D: after 4–6 days of STEMI; 1Mth: 1 month after STEMI; 6Mth: 6 months after STEMI) (p < 0.0001; PERMANOVA test with 10,000 permutations, Figure 2E). Similar results were observed from an independent cohort GSE62646, where there was a significant difference in the APVS profiles between patients with CCS and those with STEMI at presentation (p < 0.0001, Figure 2E). Furthermore, we performed mfuzz analysis to identify the association between the perturbation in the APVS profile expression pattern and disease status. The fuzzy c-means clustering grouped the trends of transcriptome expression change into 10 clusters, where Cluster2 and Cluster4 represented the two molecular perturbation patterns of APVS along the disease trajectory (Figure 2G). The Cluster2-pattern was prominently relevant to lipid oxidation and inflammation activation, whereas Cluster4-pattern was significantly enriched in the cellular biosynthesis and DNA repair processes (Figure 2H; Table S5).

Robust performance of APVS-based classifier

A neural network classifier was trained from APVS gene panel using backpropagation for patient diagnosis (Figure S2A; Table S6). To evaluate the discriminatory power and the generalization capacity of the APVS classifier, we constructed the receiver-operator characteristic (ROC) curves and confusion matrices. The area under the ROC curve (AUC) of the classifier achieved 0.955 [0.984–1.000] in the testing cohort (STEMIs: CCS). This was consistently observed in multiple independent external cohorts, with AUCs of 0.985 [0.971–0.998] in GSE59867 (STEMI: CCS); 0.997 [0.990–1.000] in GSE62646 (STEMI: CCS); 0.952 [0.884–1.000] in GSE28829 (Advanced stage: Early stage); 0.972 [0.895–1.000] in GSE41571 (Ruptured plaque: Stable plaque); 0.871 [0.766–0.976] in GSE48060 (STEMI: Health); 0.916 [0.805–1.000] in GSE60993 (STEMI: Health); 1.000 [1.000–1.000] in GSE141512 (STEMI: Health), respectively (Figure S2B). The confusion matrices provide detailed results of discriminant analysis and confirmed that the APVS classifier could stably differentiate individuals with different pathological conditions (Figure 3A). The higher F1-score, precision, specificity, kappa, accuracy, and AUC consistently suggested that the classifier might be better equipped to distinguish between pathological conditions, such as CCS and STEMI patients at presentation, CAD patients with advanced- and early-stage plaques, CAD patients with rupture and stable plaques than between healthy and pathological states (Figure 3A).

Figure 3.

Robust performance of the APVS-based classifier

(A) The confusion matrices showing the detailed results of discriminant analyses by APVS classifier in eight cohorts.

(B) Comparison of the C-indexes of the APVS classifier with those of 33 published crucial risk genes associated with coronary heart disease causation and susceptibility in eight cohorts.

Furthermore, we collected previously reported signature genes involved in the development and progression of coronary heart disease (Table S7) and compared the predicted efficacy of classifier with them in the eight cohorts. The results of C-indexes indicated that our classifier achieved improved accuracy to the other variables (Figure 3B). The APVS classifier has undergone optimization through an integrated program and obtained better extrapolation potential.

The clinical interpretability underlying APVSLevel

To address the lack of interpretability inherent in the ML classifier, we employed an unsupervised technique and established a quantification system based on APVS, called APVSLevel, which measured the atherosclerosis severity and plaque vulnerability at both bulk and single-cell resolution. We first assessed whether the APVSLevel could accurately predict clinical outcomes. The smooth HR curves for major adverse cardiovascular event (MACE) were constructed in the arteriosclerotic disease cohort GSE21545, suggesting a linear prognostic value of APVSLevel (Figure 4A). Univariate Cox regression analysis identified APVSLevel as a significant clinical factor associated with MACE (HR = 3.819, p < 0.01, Figure 4A). The Kaplan-Meier curve revealed that CCS patients with higher APVSLevel had a higher incidence of MACE (log rank p < 0.01, Figure 4B). To identify the associations between APVSLevel and clinical manifestations, we first measured the correlations between the APVSLevel and different types of clinical variables in cohort GSE90074 from molecular level. The hierarchical clustering of spearman correlations between all genes indicated that APVS gene profile strongly correlated with the severity of coronary stenosis (Figure 4C; Table S8). And APVSLevel were gradually elevated with an increasing coronary stenosis degree (level1: <10% stenosis in all major coronary arteries; level2: 10%–70% stenosis in one vessel and left main coronary artery excluded; level3: >70% stenosis in one vessel; level4: >70% stenosis in two vessels; level5: >70% stenosis in three vessels, p < 0.001, Figure 4D). The phenotype annotations of APVSLevel with the Human Phenotype Ontology (HPO) confirmed a strong association between APVSLevel and CAD deterioration (Figure S4).

Figure 4.

Clinical implications underlying APVSLevel

(A) Line chart displaying the estimated logarithm hazard ratios (HRs) represented by red lines, along with 95% confidence intervals indicated by shading, for the association between APVSLevel and Major Adverse Cardiovascular Events (MACE). This analysis was performed using the dfmacox function in a smooth HR optimal extended Cox-type additive hazard regression unadjusted model. The effect of APVSLevel on the risk of MACE was modeled with restricted cubic spline regressions. A value of 0.331 APVSLevel, as the optimal cutoff value, was used as the reference value for calculations.

(B) Cumulative (Kaplan-Meier) incidence of STEMI-related MACE in GSE21545. p value was derived by log rank test.

(C) Associations between gene profile and clinical variables in GSE90074. Genes were colored according to the signatures they belong to (APVS, DCPGs, top four relevant WGCNA modules identified in the first section of results). The color ranged for heatmap entries showed the p values for each association (enrichment and depletion) in a negative logarithmic scale.

(D–I) Line charts and boxplots depicted the mean and distribution of APVSLevel values in six independent datasets, illustrating their association with disease severity, progression, and remission timeline.

(J) Gantt chart showing the distribution of stenosis degree (level1: <10% stenosis in all major coronary arteries; level2: 10%–70% stenosis in one vessel and left main coronary artery excluded; level3: >70% stenosis in one vessel; level4: >70% stenosis in two vessels; and level5: >70% stenosis in three vessels) and Duke Coronary Artery Disease index across deciles of APVSLevel.

(K) Submap analysis of the APVSLevel-High and -Low groups and STEMI patients who either experienced restenosis or non-restenosis after PCI (left). Submap analysis of the APVSLevel-High and -Low groups and atherosclerotic patients who were either sensitive or not to cholesterol-lowering therapy with simvastatin (right). The p value indicates the degree of similarity between the paired gene expression profiles, with a lower p value indicating greater similarity.

Furthermore, our findings revealed that APVSLevel fluctuated with different scenarios of disease flares and remissions. The closer an individual is to having an active disease state, an elevated APVSLevel tends to be. Specifically, APVSLevel shifted significantly during various stages of CAD development, with the highest level observed in STEMI patients followed by NSTEMI, AMI after percutaneous coronary intervention (PCI), UA and healthy controls (p < 0.001, Figure 4E). During the acute phase, such as the first day of STEMI admission, patients exhibited significantly higher APVSLevel than those in the stable CCS phase. Subsequently, APVSLevel decreased after successful mechanical reperfusion or stenting of the infarct artery, and continued to progressively decline during the post-STEMI remission (Figures 4F–4I and S5A). We also assessed the association between APVSLevel and the intensification of pathophysiological process of coronary atherosclerosis. APVSLevel was significantly increased in atherosclerotic lesions compared to adjacent normal artery walls, as well as in the blood from atherosclerosis individuals compared to those from healthy individuals (Figures S5B and S5C). Stenosis degree (p < 0.001) and Duke Coronary Artery Disease Index¹⁷ (p < 0.01) both gradually increased according to APVSLevel deciles (Figure 4J).

Next, Subclass-Mapping analysis was applied to predict the likelihood of effective responses to therapy. Patients with higher APVSLevel exhibited expression patterns more closely resembling those of STEMI patients with restenosis after PCI, indicating a higher propensity for developing coronary restenosis in STEMI patients with high APVSLevel (FDR<0.01, Figure 4K; p < 0.05; Figure S5D). By comparing the expression profile of APVSLevel-High and -Low groups with another dataset containing 47 atherosclerotic patients who responded to cholesterol-lowering treatment with simvastatin, we found that patients with lower APVSLevel tended to exhibit better response and clinical benefits, suggesting a greater sensitivity to simvastatin in APVSLevel-Low state. (FDR<0.05, Figure 4K). Moreover, in a study involving high-cholesterol (HC) diet ApoE∗3Leiden transgenic mice treated with rosuvastatin and ezetimibe, alone and in combination,¹⁸ we found that mice receiving combination therapy with the most optimal therapeutic efficacy had the lowest APVSLevel. APVSLevel also showed a significant difference between mice with the highest and lowest cholesterol levels (p < 0.05, Figure S5E). These findings underscore the potential of APVSLevel as an indicator of the disease severity and progression, and suggested atherosclerotic individuals with low APVSLevel may be more likely to benefit from cholesterol-lowering therapy.

The biological implications underlying APVSLevel

We further computed the APVSLevel of over 1,400 atherosclerotic individuals across 10 bulk cohorts and classified them into APVSLevel-High and APVSLevel-Low subtypes by the median value (GSE123342, GSE61145, GSE59867, GSE62646, GSE29532, GSE20680, GSE20681, GSE28454, GSE90074, and GSE29111). The distribution pattern of APVSLevel was similar, and the proportions of two groups were comparable, indicating the stable generalizability of APVSLevel (Figure S5).

To understand the higher-order molecular phenotypes implicated with APVSLevel, we introduced a vissE framework and identified eight prioritized biological function themes that differed between the high and low subtypes. These themes were mainly related to antigen presentation, killer cell-mediated cytotoxicity, extracellular matrix (ECM) organization, and autoimmunity (Figures 5A andS6). Additionally, we utilized differential rank conservation analysis (DIRAC) to quantify conservation differences of 14 categories of relevant pathogenic pathways for APVSLevel-High and APVSLevel-Low samples (Figure 5B; Table S9). Most pathway terms had significantly lower rank conservation indexes (RCIs) in the APVSLevel-High samples (p < 0.05), suggesting a higher degree of dysregulation and higher variability in the transcriptome expression level in this subtype. Specific processes, such as necrotic core, DNA damage, senescence, and oxidative stress showed minimum RCIs in APVSLevel-High subtype (Figure 5B; Table S9). Gene set enrichment analysis (GSEA) revealed that APVSLevel-High subtype featured by activation of the coagulation cascade, collagen degradation, Notch and TLR signalings (Figure 5C). Gene set variation analysis (GSVA) analysis further identified APVSLevel-Low atherosclerotic populations as a stromal-rich subtype, distinguished by fibroblast migration, abundant fibrous content, and immune-suppression, while APVSLevel-High ones were characterized as an immune-promoting subtype, with a more robust inflammatory response, granulocyte activation, and macrophage M1 polarization (Figure S7).

Figure 5.

The biological implications underlying APVSLevel

(A) Left: The overlap network was created based on the gene-set similarity data and annotated using the MSigDB categories. Significantly dysregulated Gene-sets between APVSLevel-High and APVSLevel-Low subtypes were identified by the Qusage algorithm and the similarity was computed using the Jaccard index. The graph clustering approaches (Walktrap algorithm) identified eight higher-order biological functional communities. Right: VissE automatically assessed each community for biological similarities using text-mining approaches. Frequency analysis (adjusted for using the inverse document frequency) on the gene-set short descriptions was used to assess recurring biological themes in communities. The results were presented as word clouds.

(B) Top: Comparison of the RCI of pathogenic pathways between APVSLevel-High and APVSLevel-Low subtypes. p < 0.05 indicated differentially regulated pathways. An RCI of 1.0 indicated that the ranks of pathway genes were mostly stable among samples, whereas an RCI of 0.5 represented that the ranks of pathway genes varied greatly between samples of the same phenotype. Bottom: The number of pathways that was significantly dysregulated (p < 0.05) among each of 14 pathogenic categories.

(C) GSEA estimated differences in pathways activity across APVSLevel-High and APVSLevel-Low subtypes.

(D) The distribution of xcell quantification between APVSLevel-High and APVSLevel-Low plaques.

(E) The distribution of antigen presentation molecules (MHC), immunoinhibitors (SC), immunostimulators (EC), checkpoints (CP), and immunophenoscore (IPS) between APVSLevel-High and APVSLevel-Low plaques.

Immune landscape underlying APVSLevel

The aforementioned findings revealed that APVSLevel was relevant for promoting immune. We consequently investigated the immune landscape within the atherosclerotic lesions through deconvolution, and found that with elevating APVSLevel, the abundance of myeloid, lymphoid, and hematopoietic stem cell increased, while stromal cells decreased (Figure S8). The plaques with higher APVSLevel had increased infiltration of monocyte/macrophage, γδT cell, dendritic cell, neutrophil, and foam cell, but a lower proportion of fibroblast (Figure S8). The xcell quantification highlighted an overstimulated immune activity, microenvironment perturbation, and loss of stroma density in the APVSLevel-High atherosclerotic lesions (Figure 5D).

We also profiled four immunomodulators (162 molecules in total, Table S10), including antigen presentation molecules (MHC), immunoinhibitors (SC), immunostimulators (EC), and checkpoints (CP), to form immunophenoscore (IPS), which was an effective predictor of immune response and therapy. APVSLevel-High plaques had significantly higher levels of IPS and MHC but lower SC, suggesting superior immunogenicity and potential disorder of immunity regulation (Figures 5E andS9). Collectively, these results consistently suggested that APVSLevel-High subtype may benefit from immunotherapy.

Global regulatory landscape and therapeutic agent discovery

To systematically investigate the potential mechanisms underlying APVS in the disease progression, we screened for dysregulated components across multiple dimensions and integrated them into a broader interaction framework (Figure S10). With a threshold of |correlation coefficient| > 0.600 and p < 0.001, we established a global regulatory network using the co-expression trend analysis (Figure 6A). Our analysis revealed modulated relationships among prioritized regulators (APVS genes with the highest centrality, upstream transcription factors, downstream pathways, signaling hallmarks, biochemical reaction signatures, ands infiltrating immune cells) (Table S11). The co-expression pattern depicted the interactions of multi-dimensional components involved in regulatory mechanisms (Figure S11; Table S11).

Figure 6.

Global regulatory landscape and therapeutic drug benefits

(A) APVS-centric interactive network showing tight modulated relationships of critical regulators.

(B) Schematic diagram showing the APVS-driven therapeutic discovery.

(C) Results of eXtreme Sum signature matching method. Lower scores implied higher reversal potency and greater potential for application.

To identify potential targeted interventions for the prevention and treatment of acute arterial disruptive events, we leveraged a “signature reversion” strategy (Figure 6B). Specifically, two aberrant expression patterns were first identified, which were driven positively and negatively by APVSLevel, respectively (Figure S12). Next, we applied the eXtreme-Sum approach to match with these patterns with pharmacologic perturbation data. Out of 1,286 drug compounds, 466 with negative the similarity scores could reverse the expression levels of aberrant patterns (Table S12). Notably, several compounds, including MK.8686, AACOCF3, TTNPB, MS.275, and clofibrate, demonstrated the highest reversal potency and may serve as additional supplements to routine agents (Figure 6C).

Single-cell resolution interpretation of the biological significance of APVSLevel

To elucidate mechanisms underlying the impact of APVS-induced microenvironmental alterations on plaque vulnerability, we enrolled two single-cell RNA sequencing cohorts, comprising 43,964 cells from three atherosclerotic cores (AC) and three adjacent normal tissues (AN), as well as 2,237 cells from plaques of three STEMI patients and four CCS patients (Figure S13A). Consistent with previous bulk analyses, we found that the AC microenvironment presented a higher APVSLevel compared to AN (p < 0.0001, Figure 7A). Similarly, vulnerable plaques from STEMI patients had a significantly higher APVSLevel than stable plaques from CCS patients (p < 0.0001, Figure 7B). Specifically, APVSLevel were enhanced in monocyte/macrophage (Mono/Mac), mast cell, B cell, and plasma cell populations derived from AC (Figure 7C), suggesting their potential roles in the formation of atherosclerotic plaque. Moreover, we observed that a subset of macrophages in STEMI plaques exhibited higher APVSLevel than those in CCS plaques, and these cells predominantly secreted pro-inflammatory cytokines CXCL3 and IL1B (Figure 7D). Conversely, fibrotic macrophages expressing C1Q, a plaque-stabilizing phenotype showed lower APVSLevel in STEMI plaques (Figure 7D). The co-localization of APVSLevel and markers of macrophage-derived foam cells and proinflammatory markers were also confirmed (Figure S14A). These findings supported the use of APVSLevel to characterize pathological conditions at the single-cell level.

Figure 7.

Single-cell resolution interpretation of the biological significance of APVSLevel

(A) tSNE visualization of single cells captured from the atherosclerotic core (AC) and adjacent normal tissue (AN). Boxplot showing the distribution of APVSLevel between AC and AN.

(B) tSNE visualization of single cells captured from the vulnerable plaques from STEMI patients and stable plaques from CCS patients. Boxplot showing the distribution of APVSLevel between vulnerable and stable plaques.

(C) Distribution of APVSLevel for each cell type in AC and AN.

(D) Distribution of APVSLevel for each cell type in vulnerable and stable plaques.

(E) Distribution of published biological signatures in the AC between APVSLevel-High and APVSLevel-Low states.

(F) Cell-cell ligand-receptor (LR) and cytokine-related signaling network analysis. Cell-cell ligand-receptor (LR) and cytokine-related signaling network analysis. Left: Differences in inferred interaction number and strength of all cells between APVSLevel-High (lemon yellow) and APVSLevel-Low (mint green) states. Right: Circle plots displaying putative LR interactions between monocyte/macrophage (Mono/Mac) and other cell types from APVSLevel-High (top) and APVSLevel-Low (bottom) states.

(G) Signaling networks between APVSLevel-High (top) and APVSLevel-Low (bottom) states.

(H) Bar plot showing the 20 top-ranked ligands inferred to regulate Mono/Mac by the other cell types as senders according to NicheNet. Dot plot showing the expression percentage (dot size) and intensity (dot intensity) of top-ranked ligands in each sender.

(I) Heatmap showing regulatory potential of the top 20 ranked ligand and the top downstream target genes in myofibroblast.

(J) Bar plot showing the 20 top-ranked ligands inferred to regulate myofibroblast by the other cell types as senders according to NicheNet. Dot plot showing the expression percentage (dot size) and intensity (dot intensity) of top-ranked ligands in each sender.

(K) Heatmap showing regulatory potential of top 20 ranked ligand and the top downstream target genes in myofibroblast.

Then, we divided the 33,593 cells from AC into two vulnerability states based on the median APVSLevel (Figure S14B). An increased percentage of monocyte/macrophage (Mono/Mac), endothelial cell (EC) and a reduced percentage of myofibroblast, SMC was observed in APVSLevel-High area, suggesting heterogeneity in cell composition with the change of the plaque destabilization (Figure S14C). GSEA analysis revealed that coagulation activation, hypoxia, TNF signaling, heme metabolism, and cholesterol homeostasis dysregulation were significantly enriched in APVSLevel-High state, while WNT signaling, myogenesis, active proliferation and mesenchymal transition were more pronounced in the APVSLevel-Low state (Figure S14D). We further calculated 15 published biological gene-sets and found that APVSLevel-High state exhibited elevated immune response, inflammation promotion, cellular senescence, oxidative stress, and DNA damage processes. Conversely, fibrosis processes significantly contributed to the transitions to APVSLevel-Low state, which featured rich collagen components and was resistant to rupture. These indicated that APVSLevel could mediate the plaque microenvironment remodeling, supporting the bulk-level observations (Figures 7E and S14E).

Intercellular crosstalk within the atherosclerotic core was remodeled by APVS

The cell-cell communication networks in AC revealed that cell populations under APVSLevel-High had enhanced overall weighted incoming/outgoing signaling. Specifically, Mono/Mac under APVSLevel-High exhibited higher interaction weight (Figure 7F), while myofibroblast, which plays a critical role in fibrous cap formation, displayed reduced communication activities under APVSLevel-High (Figure S15). Additionally, under APVSLevel-High state, pro-inflammatory signaling networks (MIF and VCAM) were strengthened, whereas profibrogenic networks (SPP1 and fibronectin 1 [FN1]) were downregulated (Figure 7G).

We applied the NicheNet algorithm to further analyze active ligand-receptor pairs (LRs) and downstream targets driving the transition of atherosclerotic lesions from APVSLevel-Low to APVSLevel-High state. Our previous results had shown notable differences in abundance, communication weight and APVSLevel value of Mono/Mac and myofibroblast between two states, suggesting their significant potential contribution to plaque destabilization. We first designated Mono/Mac as “receiver” and ranked its ligands by the activity using the pearson correlation coefficient. Strikingly, the top-ranked ligands were specifically expressed in myofibroblast (Figure 7H), and we also identified top target genes strongly regulated by these ligands (Figure 7I). A group of APVS-induced active LRs, including IL34/CTSK, IL6/SOCS3, AGT/SPP1, and APOE/SPP1, were found to promote the crosstalk between Mono/Mac and myofibroblast during the plaque progression (Figure 7I). We further designated myofibroblast as “receiver”, and consistent with the previous result, we confirmed an APVS-induced strong interaction between myofibroblast to Mono/Mac through CXCL2/CDKN1A, ADM/CCL2, NRG1/FOS, and ITGB2/CCL2 (Figures 7J and 7K). Metascape analysis demonstrated that augmented crosstalk between Mono/Mac and myofibroblast under APVSLevel-High state significantly aggravated acute inflammation response and ECM degradation, increasing the risk of plaque rupture (Figure S16). These findings suggested that targeting the identified LRs interactions may offer promising therapeutic strategies for mitigating atherosclerosis progression.

Discussion

The vulnerability of atherosclerotic plaques is a critical determinant of the development of AMI in atherosclerotic patients with CCS.¹⁹ However, the progression of atherosclerotic plaque is a complex and dynamic process, and the detailed mechanisms underlying the formation, development and rupture of plaque are not fully understood.²⁰ In additional, the assessment of CAD progression and risk of AMI currently relies on a series of pathological indicators testing for CCS patients, which may fail to timely diagnose the early sign of insufficient myocardial blood supply before irreversible damage occurs.^21,22 Furthermore, differentiating between patients with different CAD subtypes also remains a challenge.^23,24 Therefore, the identification of a non-invasive biomarker signature for monitoring vulnerable plaques is essential for precise diagnose and prevention of STEMI.

In this study, we identified an atherosclerotic plaque vulnerability biomarker signature (APVS) and provided insights into potential pathological mechanisms and therapeutic targets for atherosclerosis progression at both bulk and single-cell resolutions. Our robust classifier, based on APVS, enables effective and stable differential diagnosis of STEMI patients from CCS or healthy individuals upon presentation. Furthermore, we developed a high-dimensional quantification system (APVSLevel) to evaluate atherosclerosis severity from clinical and biological perspectives, which may help prevent the transition from coronary stability to vulnerability and the occurrence of acute arterial disruptive events.

We have uncovered a dysregulated gene co-expression pattern that aggravates atherosclerotic plaque progression and incorporated it into our integrative framework. We conducted a comprehensive evaluation of nine classical learners on their ability to differentiate between STEMI and CCS patients, and the optimal learners were combined to generate an APVS-based classifier. Indeed, the selection of an optimal modeling algorithm is crucial; however, researchers might mostly choose the algorithms based on their preferences and knowledge limitations, leading to limited predictive power.²⁵ Some studies do not elaborate on the rationale and reasons for selecting such algorithms, which makes it difficult to find the best modeling approach to fit a specific situation. Our framework integrated nine ML algorithms, allowing each technique to be maximally exploited. After minimizing redundant information through this process, a 14 features-based consensus signature termed APVS was identified, which displayed accuracy for predicting clinically relevant outcomes. The ROC and confusion matrix demonstrated that our classifier maintained accurate and stable performance in eight external cohorts, suggesting a potential for extrapolation. Importantly, our classifier not only differentiated STEMI patients from healthy donors but also from CCS patients. Furthermore, it effectively eliminated false negative interference and reduced the misdiagnosis of acute arterial disruptive events, enabling timely treatment. We also retrieved 33 published causative risk genes, such as CYP7A1, TNNC1, TNNT2, MMP9, and TRIB1, which were reported to be significantly associated with the clinical strategies and outcomes.^26,27,28 Of note, APVS-based classifier outperformed these genes and achieved improved accuracy significantly in eight validation cohorts. Overall, our classifier displayed substantial diagnostic potential for distinguishing between different CAD subtypes and pathological conditions, leading to effective treatment and better outcomes for patients.

Currently, most risk-scoring systems for CAD rely on epidemiological evidence instead of biological pathogenesis.^14,15 Traditional models often perform macroscopic risk stratification without reflecting the actual atherosclerotic plaque state, which limits the potential of precision medicine.²⁹ To address this, we established a quantification system with both clinical and biological interpretability, termed APVSLevel, to measure the atherosclerosis severity and plaque vulnerability.

Atherosclerotic patients with elevated APVSLevel exhibited an enhanced procoagulant state and a greater risk of plaque rupture, ACS, and restenosis after PCI. APVSLevel was confirmed to undergo dynamic changes in response to different scenarios of disease flares and remissions, reflecting the severity of coronary stenosis during various stages of CAD development. The atherosclerotic core (AC) under the APVSLevel-High state had substantial infiltration of inflammatory cells, including B cells, macrophages, and neutrophils. B cell autoimmunity is known to aggravate atherosclerosis,³⁰ and elevated IgE immunoglobulins stimulate macrophages, promoting inflammation and production of proteases that attack the fibrous cap.²⁰ Neutrophils can accelerate artery thrombosis by promoting endothelial erosion via neutrophil traps and matrix metalloproteinase secretion.³¹ Fibrous cap thinning and necrotic core formation predominantly occur in the APVSLevel-High AC due to the degradation of ECM and the loss of SMC and fibroblasts.³² Furthermore, Notch signaling activation in the APVSLevel-High AC promoted proinflammatory response and senescence of EC.³³ APVSLevel-High AC was also associated with significant oxidative damage and immunogenicity. Immunogenic molecules, such as oxidized-LDL, found in APVSLevel-High plaque can lead to irreversible arterial wall damage by bridging the progression from chronic inflammation into an autoantibody response in an advanced stage.³⁴

APVSLevel is a powerful tool not only for characterizing the biological status at bulk level but also provided direct snapshots at single-cell resolution. Significant APVSLevel differences were revealed between AC and AN in various cell clusters, particularly monocyte/macrophage, SMC, and B cell. We found that APVSLevel promoted plaque destabilization in STEMI patients by activating CXCL3+/IL1B+ inflammatory macrophages and suppressing C1Q + fibrotic macrophages compared to CCS patients.³⁵ Moreover, cells in the AC under the APVSLevel-High state activated a series of pathways involving TNF, IFN response, oxidative damage, and heme metabolism. Heme aggravated the CAD through oxidative stress, EC metabolic dysfunction, and macrophage proinflammatory polarization.³⁶ On the other hand, enhanced Wnt pathway regulation of cellular cholesterol trafficking was observed in AC under APVSLevel-Low state, essential for plaque stabilization.³⁷ However, senescent cells triggered by DNA damage were implicated in immune aging and fibrous cap degeneration,^38,39,40 indicating cells under APVSLevel-High were subjected to DNA damage and oxidative stress and entered into a senescence state, contributing to lesion growth.

Under the APVSLevel-Low state, ECM family signaling network was enhanced within intercellular crosstalk, promoting atherosclerotic lesion regression by regulating ECM reorganization and decreasing lipid accumulation.⁴¹ In contrast, upregulated VCAM and MIF signaling networks within APVSLevel-High crosstalk further promoted atherogenic leukocyte recruitment and lesional inflammation, exacerbating plaque instability.^42,43 Notably, consistent with our observation of intensified SPP1 signaling network in APVSLevel-Low crosstalk, previous reports have suggested that the plaque macrophage-mediated paracrine signaling by SPP1 enhanced fibroblast collagen secretion, contributing to the formation of the desmoplastic region.^44,45 Moreover, a group of dysregulated LRs between macrophages and myofibroblasts was identified as therapeutic targets, including IL34/CTSK, IL6/SOCS3, AGT/SPP1, APOE/SPP1, CXCL2/CDKN1A, and ADM/CCL2, which were demonstrated to promote acute inflammation response and ECM degradation in the AC. Based on these findings, we screened for five compounds as additional supplements for more targeted interventions for plaque progression, including MK-886, AACOCF3, TTNPB, MS-275, and clofibrate.

In conclusion, this study comprehensively analyzed atherosclerotic plaque vulnerability using a computational biology approach, which would offer a broad biological and clinical perspective for future functional and therapeutic studies of atherosclerosis progression. The high-dimensional quantification system, APVSLevel, with the robust ability to measure the severity of atherosclerosis and plaque vulnerability, could potentially serve as a tool to optimize clinical decision-making and management for atherosclerotic patients.

Limitations of the study

Procedures bias from the system biology potentially cannot fully recapitulate the overall disease exacerbation. Secondly, samples from this work were retrospective; a prospective multicenter study is imperative to confirm the biological and clinical interpretability of APVSLevel. Thirdly, since the current study is based on a high-throughput cohort, it is necessary to compare the APVS-based classifier with existing biomarkers in the future study. Fourthly, the roles of most molecules from ASVS in atherosclerosis progression remain unknown, and further in vivo and in vitro experiments are needed to reveal their functions. Lastly, several clinical and molecular traits were inadequate, presumably masking the associations between APVS and certain factors. Thus, further studies should be conducted before the clinical translation.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Single-cell RNA-seq and clinical data	Alsaigh et al.46	GEO: GSE159677
Single-cell RNA-seq and clinical data	Emoto et al.35	GEO: GSE184073
Bulk microarray or RNA-seq and clinical data	Döring et al.47	GEO: GSE28829
Bulk microarray or RNA-seq and clinical data	Maciejak et al.48	GEO: GSE59867
Bulk microarray or RNA-seq and clinical data	Kiliszek et al.49	GEO: GSE62646
Bulk microarray or RNA-seq and clinical data	Lee et al.50	GEO: GSE41571
Bulk microarray or RNA-seq and clinical data	Suresh et al.51	GEO: GSE48060
Bulk microarray or RNA-seq and clinical data	Park et al.52	GEO: GSE60993
Bulk microarray or RNA-seq and clinical data	Osmak et al.53	GEO: GSE141512
Bulk microarray or RNA-seq and clinical data	Folkersen et al.54	GEO: GSE21545
Bulk microarray or RNA-seq and clinical data	Ravi et al.55	GEO: GSE90074
Bulk microarray or RNA-seq and clinical data	Park et al.52	GEO: GSE61145
Bulk microarray or RNA-seq and clinical data	Vanhaverbeke et al.56	GEO: GSE123342
Bulk microarray or RNA-seq and clinical data	Maouche et al.	GEO: GSE28454
Bulk microarray or RNA-seq and clinical data	Sinnaeve et al.57	GEO: GSE12288
Bulk microarray or RNA-seq and clinical data	SHARMA et al.	GEO: GSE46560
Bulk microarray or RNA-seq and clinical data	Puig et al.	GEO: GSE37824
Bulk microarray or RNA-seq and clinical data	Silbiger et al.58	GEO: GSE29532
Bulk microarray or RNA-seq and clinical data	Hägg et al.59	GEO: GSE40231
Bulk microarray or RNA-seq and clinical data	Huang et al.60	GEO: GSE20129
Bulk microarray or RNA-seq and clinical data	Verschuren et al.18	GEO: GSE38688
Bulk microarray or RNA-seq and clinical data	Milo et al.	GEO: GSE29111
Bulk microarray or RNA-seq and clinical data	Elashoff et al.61	GEO: GSE20680
Bulk microarray or RNA-seq and clinical data	Wingrove et al.61	GEO: GSE20681
Human transcription factors data	Cistrome database	http://cistrome.org/
Drug signatures and gene expression profiles for the drugs	Connectivity Map database	https://www.broadinstitute.org/
Human regulated signaling pathways information	Biocarta database	http://pid.nci.nih.gov/browse_pathways.shtml#biocarta/
Human Protein and biological interaction signatures	Reactome Pathway Database	https://reactome.org
C1-C8 and Hallmark biological datasets	Molecular Signatures Database	http://www.gsea-msigdb.org/gsea/msigdb
Software and algorithms
R version 4.1.3	https://www.r-project.org/	https://www.r-project.org/
APVS	github	https://github.com/DrZoggg/APVS
Seurat version 4.3.0	github	https://github.com/satijalab/seurat
vegan version 2.5.7	github	https://github.com/vegandevs/vegan
xCell version 1.1.0	Github	https://github.com/dviraran/xCell
AUCell version 1.16.0	github	https://github.com/simslab/AUCell
CellChat version 1.1.3	github	https://github.com/sqjin/CellChat
NicheNet version 3.14	github	https://github.com/saeyslab/nichenetr
limma version 3.50.3	github	https://github.com/cran/limma
survival version 3.3.1	github	https://github.com/therneau/survival
e1071 version 1.7.9	CRAN	https://cran.r-project.org/package=e1071
glmnet version 4.1.3	CRAN	https://cran.r -project.org/package=glmnet
randomForest version 4.7.1	CRAN	https://cran.r -project.org/package=randomForest
mboost version 2.9.5	CRAN	https://cran.r -project.org/package=mboost
kknn version 1.3.1	CRAN	https://cran.r -project.org/package=kknn
pls version 2.8.0	CRAN	https://cran.r -project.org/package=pls
neuralnet version 1.44.2	CRAN	https://cran.r -project.org/package=neuralnet
rpart version 4.1.16	CRAN	https://cran.r -project.org/package=rpart
ggplot2 version 3.3.5	Github	https://github.com/tidyverse/ggplot2
ggpubr version 0.4.0	github	https://github.com/kassambara/ggpubr
GSVA version 1.42.0	github	https://github.com/rcastelo/GSVA
WGCNA version 1.72.1	github	https://github.com/cran/WGCNA
clusterProfiler version 4.4.2	github	https://github.com/YuLab-SMU/clusterProfiler
igraph version 1.3.5	github	https://github.com/igraph/rigraph
pROC version 1.18.0	github	https://github.com/xrobin/pROC
DALEX version 2.4.0	github	https://github.com/ModelOriented/DALEX
rms version 6.2.0	github	https://github.com/harrelfe/rms
smoothHR version 1.0.4	CRAN	https://cran.r-project.org/web/packages/smoothHR/index.html
QuSage version 2.28.0	github	https://github.com/qusage/qusage
Mfuzz version 2.54.0	github	https://github.com/dholt/Mfuzz
Immunophenogram	github	https://github.com/icbi-lab/Immunophenogram
Submap	GenePattern	https://cloud.genepattern.org
xSum	PMID: 25606058	PMID: 25606058

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Junnan Tang (fcctangjn@zzu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Experimental model and study participant details

The study was approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University (2021-KY-0720). Especially, the source data we analyzed in this research were mainly acquired from the public databases, and the detailed information could be found in the method details part. The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Method details

Public datasets collection

The human and mouse datasets were selected from the Gene Expression Omnibus (GEO) database under the National Center for Biotechnology Information platform (NCBI), where the method of acquisition and application performed complied with relevant guidelines and policies. The analyzed data we used in this research were mainly acquired from the public database, and hence the need for the local ethics committee approval or patient informed consent was waived. In total, 2242 samples from 24 public cohorts were enrolled and reanalyzed in this study (Table S13).

Two single cell RNA-seq datasets from published studies of atherosclerotic patients were obtained from GEO. A total of 43964 single cells from three plaque samples of atherosclerotic core (AC) and three samples of adjacent normal tissue (AN) were collected from atherosclerotic patients undergoing artery endarterectomy using 10x genomics platform-based single-cell RNA-seq protocol (GSE159677).⁴⁶ Additionally, we analyzed 2237 cells from vulnerable plaques of 3 STEMI patients and stable plaques of 4 CCS patients (GSE184073).³⁵

22 microarray cohorts based on bulk transcriptome were obtained:

• 1.GSE28829⁴⁷ (13 early-stage plaques, 16 advanced-stage plaques);
• 2.GSE59867⁴⁸ (blood samples from 46 CCS patients, 111 patients on the first day of STEMI, 101 patients after 4-6 days of STEMI, 95 patients after 1 month of STEMI, 83 patients after 6 months of STEMI);
• 3.GSE62646⁴⁹ (blood samples from 14 CCS patients, 28 patients on the first day of STEMI, 28 patients after 4-6 days of STEMI, 28 patients after 6 months of STEMI);
• 4.GSE41571⁵⁰ (5 ruptured plaques, 6 stable plaques);
• 5.GSE48060⁵¹ (blood samples from 31 STEMI patients, 21 healthy individuals);
• 6.GSE60993⁵² (blood samples from 17 STEMI patients, 7 healthy individuals);
• 7.GSE141512⁵³ (blood samples from 6 STEMI patients, 6 healthy individuals);
• 8.GSE21545⁵⁴ (97 blood samples and 126 plaques from atherosclerotic patients with prognostic outcome information);
• 9.GSE90074⁵⁵ (blood samples from 18 patients with <10% stenosis in all major coronary arteries, 32 patients with 10%∼70% stenosis in 1 vessel, 31 patients with >70% stenosis in 1 vessel, 26 patients with >70% stenosis in 2 vessels, 36 patients with >70% stenosis in 3 vessels);
• 10.GSE61145⁵² (blood samples from 17 healthy individuals, 9 UA patients, 10 NSTEMI patients, 14 STEMI patients, 7 recovered STEMI patients after PCI);
• 11.GSE123342⁵⁶ (blood samples from 26 CCS patients, 65 patients on the first day of STEMI, 64 patients after 30 days of STEMI, 37 patients after 1 year of STEMI);
• 12.GSE28454 (blood samples from 26 patients within 6h of STEMI, 28 patients after 3 days of STEMI, 22 patients after 90 days of STEMI);
• 13.GSE12288⁵⁷ (blood samples from 110 CAD patients annotated with Duke CAD index);
• 14.GSE46560 (blood samples from 5 CAD patients with restenosis after insertion of stent, 6 CAD patients without after insertion of stent);
• 15.GSE37824 (blood samples from 47 atherosclerotic patients received simvastatin treatment 40mg once daily. 27 responders had good responses and clinical benefits with reduced LDL-C by 43%; 20 non-responders had poor responses and clinical benefits);
• 16.GSE29532⁵⁸ (blood samples from 8 STEMI patients: 8 samples collected at admission to the Emergency Unit before receiving any medication, 8 samples at 2h of STEMI, 8 samples at 12h of STEMI, 8 samples at 24h of STEMI);
• 17.GSE40231⁵⁹ (40 atherosclerotic arterial wall tissues, 40 adjacent normal tissues);
• 18.GSE20129⁶⁰ (blood samples from 48 atherosclerosis patients, 71 healthy individuals);
• 19.GSE29111 (blood samples from 44 STEMI patients, 8 UA patients);
• 20.GSE20680⁶¹ (blood samples from 52 atherosclerotic patients with ≥70% stenosis in >1 major vessel or ≥50% stenosis in >2 arteries, 56 atherosclerotic patients with luminal stenosis >25% but less than 50%, 87 atherosclerotic patients with luminal stenosis of <=25%);
• 21.GSE20681⁶¹ (blood samples from 99 atherosclerotic patients with ≥50% stenosis in >= 1 major vessel, 99 atherosclerotic patients with luminal stenosis of < 50%);
• 22.GSE38688¹⁸ (liver tissues from 32 ApoE∗3Leiden mice with High-Cholesterol (HC) diet. 8 mice received no treatment; 8 mice received Rosuvastatin treatment alone with poor response; 8 mice received Ezetimibe treatment alone with poor response; 8 mice received Rosuvastatin and Ezetimibe combination treatment alone with good response).

Omics data resource

The list of TFs was derived from the Cistrome database (http://cistrome.org/). C1-C8 and Hallmark biological datasets were retrieved from the Molecular Signatures Database (MSigDB) (http://www.gsea-msigdb.org/gsea/msigdb). Human protein and biological interaction signatures were derived from Reactome Pathway Database (https://reactome.org). Human regulated signaling pathways were extracted from the Biocarta database (http://pid.nci.nih.gov/browse_pathways.shtml#biocarta/). Drug signatures and gene expression profiles were downloaded from Connectivity Map (CMap) database (http://www.broadinstitute.org).

Single-cell and bulk data preprocessing

The quantified single-cell gene expression matrices were analyzed through the Seurat pipeline (Version: 4.1.3). Cells with more than 5% of reads from mitochondria genes and less than 500 or more than 5500 genes were removed, while genes expressing in more than 3 single cells were included. Using "FindIntegrationAnchors" and "IntegrateData" functions integrate cells from different samples. The top 2,000 variable genes were identified via "vst" selection, considered as the input features for dimensionality reduction using PCA. The first 20 significant PCs determined by jackstraw analysis were incorporated into t-SNE analysis for further dimensional reduction and clustering visualization. The findAllMarkers function with "wilcox" method was performed to identify DEGs from the top 2,000 variable genes.

To annotate cell types, highly expressed genes of all cell subclusters were used as the potential reference and combined with canonical cell-type-specific surface markers derived from CellMarker (http://yikedaxue.slwshop.cn/index.php). The computational tool scCATCH was used to confirm the inferred cell types in an unbiased fashion. Known cell surface biomarkers were selected for annotation, including T cell (CD3D, CD3E, CD8A, CD274, CD7), NK cell (KLRD1, NKG7, GZMA), B cell (CD79A, CD79B, MS4A1), dendritic cell (CD1C, CLEC10A, FCER1A), monocyte/macrophage (LYZ, CD68, CD14, CD163, FCGR3A), smooth muscle cell (TAGLN, ACTA2, CALD1, MYH11, MFAP4), myofibroblast (COL1A1, FAP, DCN, LUM), mast cell (KIT, HDC), endothelial cell (VWF, CD34, PECAM1, ICAM2), and plasma cell (JCHAIN, MZB1, IGHG3).

The bulk raw data were processed, normalized and corrected using limma and sva packages based on different platforms. Each probe set was annotated using official gene symbols from the Ensemble database and corresponding platform annotation information. In cases where multiple probe sets corresponded to the same gene, the probe set with the highest mean intensity across all samples was retained. The limma package was applied for the differential analysis.^62,63

Weighted correlation network analysis (WGCNA)

WGCNA⁶⁴ was used to screen the dysregulated gene co-expression pattern aggravating CAD. The gene expression levels were ranked in descending order based on the standard deviation, and the top 5000 genes were selected. The hierarchical clustering analysis excluded the outlier samples for the rationality. The Pearson correlation value between each gene pair was calculated to obtain a gene similarity matrix. We constructed an adjacency matrix using the formula, aij = | Sij | β (aij: adjacency matrix between gene i and j, Sij: similarity matrix of all gene pairs, β: the soft threshold). The optimal β was selected to satisfy the scale-free distribution by the “pickSoftThreshold” function, making the correlations more distinguishable. The adjacency matrix was transformed to topological overlap matrix (TOM) and 1-TOM, which reflected the similarity and dissimilarity among genes, respectively. Finally, we classified genes into different modules using the hierarchical clustering method and calculated the module eigengene (ME), which represented each module profile. The modules highly correlated with the progression of atherosclerotic plaques from the early to the advanced stage were identified as our dysregulated gene co-expression pattern. The settings of parameters were as follows: soft threshold β = 12, minModuleSize = 50, mergeCutHeight = 0.25, and deepSplit = 2.

APVS generated from an integrative program

A program integrating nine machine-learning (ML) learners was utilized to explore an atherosclerosis plaque vulnerable signature (APVS), including Backpropagation neural network (nnet), random forest (rf), boosted generalized linear model (glmboost), lasso and elastic-net-regularized generalized linear model (glmnet), bootstrap aggregation classification and regression trees (Bagged CART), Naive Bayes (NB), K-Nearest Neighbors (Knn), Partial Least Squares (pls), classification and regression trees (CART). The algorithm was implementation using the R package "e1071, glmnet, randomForest, neuralnet, mboost, kknn, pls, rpart".

The APVS signature generation procedure was as follows:

• 1.Dysregulated co-expression pattern genes (DCPGs) were identified as the intersection between genes dysregulated in STEMI and CCS patients and the WGCNA results.
• 2.The initial exploration of signature was conducted in cohort GSE59867, which was randomly divided into discovery and testing cohorts in a 7:3 ratio.
• 3.The Nine learners were performed on DCPGs to fit models separately. To prevent overfitting caused by the complex model, 10-fold 100-repeated cross-validation (cv) was adopted to improve the generalization ability of the discovery cohort.
• 4.The effectiveness and applicability of all models were evaluated in the testing cohort, based on a consensus assessment strategy that included the accuracy, Harrell's concordance index (C-index), F1-score, precision, recall, and the root mean square of the residuals (RMSR). The random forest and backpropagation neural network learners were considered the optimal schemes.
• 5.The APVS was generated using the random forest-Recursive Feature Elimination approach combined with 10-fold 10-repeated cv and the decreasing accuracy method (Gini coefficient method) for feature selection.

Construction of APVS-based classifier

The expression status of the selected features (APVS genes) was first transformed to "Feature Score" based on the expression level (above or below the median) :

Feature Score Matrix	Expression
	Low	High
Upregulated	0	1
Downregulated	1	0

The expression level of a specific feature was compared to the median of all sample expression values. If the expression value of the upregulated feature was higher than the median, it was given a value of 1, otherwise, it was given a value of 0. Similarly, if the expression value of downregulated gene is higher than the median, it was given a value of 0, otherwise, it was given a value of 1. The outcome variable was the occurrence of cases valued as 1 and controls valued as 0. Finally, the "Feature Score" sheet containing lines of samples, columns of selected features, and one column of the outcome variable was used for backpropagation neural network training to generate the classifier.

The generalization of the classifier in external independent cohorts was evaluated using a group of performance indexes, including accuracy, Harrell's concordance index (C-index), F1-score, precision, recall, Kappa, positive/negative predictive rate, sensitivity, and specificity.

Establishment of APVSLevel

One of the major challenges of employing the ML algorithm is the lack of interpretability. To address this issue, we employed an unsupervised approach that enables us to gain insights into an individual's disease state from both biological and clinical perspectives.

• 1.We performed principal coordinates analysis (PCoA) and analysis of similarities (ANOSIM) based on APVS genes. Bray-Curtis dissimilarity matrix was calculated by beta_diversity.py, and Bray-Curtis diversity was calculated using the R package Vegan with the function vegdist.
• 2.To determine if significant differences existed between groups, we used the PERMANOVA test (2-way adonis).
• 3.A quantification system named APVSLevel was developed, which can evaluate atherosclerotic plaque vulnerability and disease severity:

$$\mathrm{A}\mathrm{P}\mathrm{V}\mathrm{S}\mathrm{L}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{l}=\left(PCoA1score+PCoA2score\right)\times \sum {\mathit{exp}}_i$$

where exp_i represents the expression level of APVS genes.

Biological annotation

The hypergeometric test-based over-representation analysis (ORA) was employed to determine whether a specific gene set was overrepresented.⁶⁵ The results were sorted by Z-score. A cut-off value of 1.96 for the Z-score, a permuted p-value cut-off of 0.05, and 1000 gene permutations were used to ensure the robustness.

We employed Gene set enrichment analysis (GSEA) to decipher the biometric differences underlying APVSLevel at the bulk- and single-level, using 10,000 permutations per gene set to yield a normalized enrichment score (NES).⁶⁶ Gene-sets with an FDR<0.05 was considered statistically significant. Gene set variation analysis (GSVA) was performed to estimate the variation of gene-set via the samples of the expression profiles.⁶⁷

Functional themes summarization for APVSLevel through VissE framework

We leveraged the Visualising Set Enrichment Analysis Results (VissE) framework for the interpretation and analysis of results from a gene-set enrichment analysis using network-based and text-mining approaches. This approach exploited the relatedness between gene-sets and the inherent hierarchical structure that may exist in pathway databases and gene ontologies to cluster results. For each community of gene-sets vissE identifies, we performed text-mining to automate characterization of emerging themes represented by the community.

• 1.Quantitative set analysis of MsigDB gene-sets:

QuSage algorithm identified the significantly dysregulated MsigDB gene-sets between APVSLevel-High and APVSLevel-Low states, based on a based on an absolute logFC threshold of 0.5 and an FDR threshold of 0.05.⁶⁸

• 2.Compute the overlap network:

Gene set similarity was computed using the Jaccard index based on the overlap between dysregulated gene-sets. The overlap network plot was annotated using the MSigDB category.

• 3.Identify communities within the network:

Given a network of related gene-sets, we expected that related gene-sets would likely represent a common higher-order biological process. As such, graph clustering algorithm Walktrap, was used to perform clustering on gene-sets to identify gene-set communities.

• 4.Characterize communities:

Gene-set communities identified can be assessed for their biological similarities using text-mining approaches. We performed a frequency analysis (adjusted for using the inverse document frequency) on the gene-set names or their short descriptions to assess recurring biological themes in communities, which were presented as word clouds.

• 5.Visualize gene-level statistics for gene-set communities:

Gene-level statistics (logFC) for each gene-set community can be visualized to better understand the genes contributing to significance of gene-sets.

• 6.Identify protein-protein interactions (PPI) in each community.

Differential rank conservation analysis (DIRAC)

As in previous studies,⁶⁹ DIRAC was used to quantify differential expression variability among phenotypes using gene subsets based on individual transcriptomes. These subsets generally correspond to predefined gene networks or pathways. For each sample in each phenotype studied, DIRAC characterized the ordering of pathway genes by comparisons between their expression values of pairs of genes. Using the comparison statistics, it defines a rank template for each pathway and phenotype that represents the expected pairwise ordering of gene expression for that pathway in that phenotype. It then employed the rank matching score (RMS) to determine how well the pathway ranking in each sample matches the ordering defined in the rank template. Averaging the RMS over all samples within a phenotype generates a pathway-specific rank conservation index (RCI) which represents how well, on average, all samples in the same phenotype match the corresponding rank template. An RCI of 1.0 suggests the ranks of pathway genes are mostly unchanged among samples, whereas an RCI of 0.5 indicates the ranks of pathway genes are substantially varied between samples of the same phenotype. The pathway was defined as being (tightly or loosely) regulated based on the level of (high or low) conservation of transcript ordering (RCI) in the individual transcriptomes. However, a pathway can relatively be tightly regulated in one phenotype and loosely regulated in another, which indicates the deregulation of the given pathway. In this study, we applied 14 kinds of atherosclerosis-related pathogenic pathways comprising 91 gene sets from MsigDB to calculate the RCIs in APVSLevel-High and APVSLevel-Low samples. The pathway categories were summarized as follows: oxidative stress, coagulation, necrotic core, vulnerable plaque, acute coronary syndrome, thrombosis, aging & senescence, DNA damage, inflammaging, atherosclerosis stenosis, hematopoiesis, shear stress and lipid accumulation, foam cell, atherosclerotic lesion.

Cellular heterogeneity underlying APVSLevel

To quantify the relative fractions of different cell types in the plaque microenvironment, we employed a gene signature expression-based cell-type enrichment tool xCell to obtain cell type abundances was obtained through.⁷⁰ xCell algorithm analyzed transcripts per million for 64 immune and stroma cell types based on the previously learned information from thousands of pure cell types varying on the sources. This approach minimizes associations among closely related cell types and provides a reliable portrayal of the cellular heterogeneity landscape.

Immunogenicity and immunotherapeutic potential underlying APVSLevel

In this study, we utilized the immunophenoscore (IPS) to assess the immune state of atherosclerotic plaque.⁷¹ The IPS-related genes, classification, and associated weights were obtained through an R-script available on GitHub (https://github.com/icbi-lab/Immunophenogram). Using a panel of marker genes related to immune response or immune toleration, we quantified four different immunophenotypes, including antigen presentation molecules, effector cells, suppressive cells, and selected immunomodulators. Furthermore, we constructed an immunophenogram for each sample. The z-score was generated to summarize the four classes, and the higher z-score of IPS indicated a stronger immunogenicity and potential for immunotherapy of the sample.

Association between APVSLevel and restenosis as well as therapy response

Subclass Mapping (Submap) is an unsupervised approach focusing on the homological similarity of groups of samples in multiple datasets despite the technical differences.⁷² It estimates the enrichment of the expression profiles of predefined phenotypes in one dataset for markers identified in the other dataset. P<0.05 corrected by the Bonferroni method was considered as significant for different phenotypes that share similar underlying molecular properties. We applied the Submap to calculate the expression similarity between patients in the APVSLevel-High and APVSLevel-Low states and patients who responded/non-responded to the cholesterol-lowering treatment Simvastatin based on the GSE37824. We also evaluated the expression similarity between patients in the APVSLevel-High and APVSLevel-Low states and patients who experienced restenosis or not after PCI based on the GSE46560.

Global regulatory network

To elucidate potential regulatory mechanism underlying APVS, we constructed a multi-dimensional regulatory network.

• 1.We first identified multi-dimensional components dysregulated significantly (DS) between STEMI and CCS. This involved retaining upstream dysregulated transcription factors (DSTFs) with absolute logFC>0.5 and FDR<0.01, dysregulated hallmark signatures (DSHallmarks) with absolute t>2 and FDR<0.01, dysregulated pathways signatures (DSpathways) with absolute t>5 and FDR<0.01, dysregulated biochemical reactions (DS biochemical reactions) with absolute t>5 and FDR<0.01, and differences in cell compositions with absolute t>2 and FDR<0.01.
• 2.The Pearson correlations (normally-distributed data) and Spearman correlations (non-normally distributed data) were performed to co-analyze the interaction coefficients of upstream DSTFs, APVS genes, DSHallmarks, DScells, and downstream DSpathways and DS biochemical reactions. We used |correlation coefficient| > 0.600 and P<0.001 as the minimum screening criteria to select the most significant components that had the highest spatial correlation with APVS. Finally, we constructed regulatory patterns using the igraph package.

Computational novel therapy discovery and repurposing

• 1.Generation of query signature representing aberrant expression patterns: WGCNA identify aberrant expression patterns driven by APVSLevel. The top four modules positively and negatively associated with APVSLevel were selected as two kinds of aberrant expression pattern.
• 2.We applied the eXtreme-Sum signature matching approach to match these patterns with pharmacologic perturbation data.⁷³ eXtreme-Sum queries the pharmacologic perturbation datasets to find compounds with the ability to reverse these patterns. The sums of the change values in pharmacologic signatures relative to upregulated query/disease genes (sumup) and downregulated query/disease genes (sumdown) were calculated. Then, XSum is defined as the following:

$$\mathrm{X}\mathrm{S}\mathrm{u}\mathrm{m}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\mathrm{s}\mathrm{u}\mathrm{m}\mathrm{u}\mathrm{p}+\mathrm{s}\mathrm{u}\mathrm{m}\mathrm{d}\mathrm{o}\mathrm{w}\mathrm{n}$$

• 3.Randomization approach determines the statistical significance. Drug compounds with a predicted XSum score below the threshold were ranked in reverse order based on the score. Compounds with significant negative scores possess gene expression patterns that are anti-correlated or oppositional to the atherosclerosis-genesis and progression-associated expression pattern, therefore representing putative novel therapeutic indications.

Single cell RNA-seq data analysis

APVS genes were selected for AUCell scoring for each individual cell.⁷⁴ Based on the area under the curve (AUC) value, gene expression rankings of cells were generated to assess the highly expressed gene set proportion in each cell. Using "AUCell_exploreThresholds" function determines the threshold to recognize ASVS features active cells. The discrimination analysis and quantification were performed through the Cell Cycle Scoring function based on previously defined cell cycle-related genes, with cells projected onto t-SNE space for visualization and coloured according to cycle clustering for visualization.

NicheNet algorithm was used to infer cell communication networks following the official workflow of nichenetr R package.⁷⁵ Differentially expressed genes in receivers between APVSLevel-High and APVSLevel-Low atherosclerotic plaque were filtered to predict the top 20 potential regulatory ligands, their target genes on receivers, and the possible cell clusters as sources of ligand expression. Additionally, CellChat pipeline was conducted following the guidelines at https://github.com/sqjin/CellChat. The overall interaction, overall signalling pattern, outgoing/incoming signalling pattern, and ligandreceptor pair were checked in detail step by step.

Quantification and statistical analysis

All statistical tests were two-sided. P-value < 0.05 and FDR < 0.05 was suggested to be statistically significant. The mean ± standard deviation for descriptive statistics was used for continuous variables with a normal distribution. The Wilcoxon rank-sum test or T test was applied to compare continuous variables, and categorical variables were compared through the chi-squared or Fisher exact test. All data processing, statistical analysis, and plotting were conducted with R 4.1.3 software (Institute for Statistics and Mathematics, Vienna, Austria; www.r-project.org).

Published: August 9, 2023

Data and code availability

• •This paper analyzes existing, publicly available data. These accession numbers for the datasets are listed in the key resources table.
• •All original code is available at “https://github.com/DrZoggg/APVS”.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.