Proteomic Atlas of Atherosclerosis: The Contribution of Proteoglycans to Sex Differences, Plaque Phenotypes, and Outcomes

BACKGROUND: Using proteomics, we aimed to reveal molecular types of human atherosclerotic lesions and study their associations with histology, imaging, and cardiovascular outcomes. METHODS: Two hundred nineteen carotid endarterectomy samples were procured from 120 patients. A sequential protein extraction protocol was employed in conjunction with multiplexed, discovery proteomics. To focus on extracellular proteins, parallel reaction monitoring was employed for targeted proteomics. Proteomic signatures were integrated with bulk, single-cell, and spatial RNA-sequencing data, and validated in 200 patients from the Athero-Express Biobank study. RESULTS: This extensive proteomics analysis identified plaque inflammation and calcification signatures, which were inversely correlated and validated using targeted proteomics. The inflammation signature was characterized by the presence of neutrophil-derived proteins, such as S100A8/9 (calprotectin) and myeloperoxidase, whereas the calcification signature included fetuin-A, osteopontin, and gamma-carboxylated proteins. The proteomics data also revealed sex differences in atherosclerosis, with large-aggregating proteoglycans versican and aggrecan being more abundant in females and exhibiting an inverse correlation with estradiol levels. The integration of RNA-sequencing data attributed the inflammation signature predominantly to neutrophils and macrophages, and the calcification and sex signatures to smooth muscle cells, except for certain plasma proteins that were not expressed but retained in plaques, such as fetuin-A. Dimensionality reduction and machine learning techniques were applied to identify 4 distinct plaque phenotypes based on proteomics data. A protein signature of 4 key proteins (calponin, protein C, serpin H1, and versican) predicted future cardiovascular mortality with an area under the curve of 75% and 67.5% in the discovery and validation cohort, respectively, surpassing the prognostic performance of imaging and histology. CONCLUSIONS: Plaque proteomics redefined clinically relevant patient groups with distinct outcomes, identifying subgroups of male and female patients with elevated risk of future cardiovascular events.

D espite decades of research, there remain ambiguities on the mechanisms that induce atherosclerotic plaque instability. According to the response-toretention hypothesis, 1 the binding of cholesterol-containing lipoprotein particles to intimal proteoglycans is the central pathogenic process in early atherogenesis. Once retained, lipoproteins get oxidized, accumulate in foam cells, and provoke a cascade of inflammatory processes that drive the formation of atherosclerosis. In addition to its critical role in the retention of lipoprotein particles, ECM (extracellular matrix) degradation processes are causally linked to the development of vulnerable plaques. 2 The ECM, in particular of the fibrous cap, ultimately defines the propensity of plaques to rupture leading to atherothrombotic events. There is a need to further understand how ECM composition differs across plaque phenotypes. The matrisome refers to the ensemble of ECM and ECMassociated proteins. 3 Proteomics offers an opportunity to define the protein composition of human atherosclerosis that can advance current assessments by histology and imaging. Till now, mapping initiatives lack information on large blood vessels and atherosclerotic plaques.
According to the modified American Heart Association classification, there are 8 types of atherosclerotic lesions: intimal thickening, fibroatheroma, late fibroatheroma, healed plaque rupture, fibrocalcific plaques, erosions, thin-capped atheroma, and ruptured plaques. 4 Calcification is supposedly a late feature of advanced atherosclerotic lesions. 5 According to ultrasound measurements, atherosclerotic plaques are classified as echolucent, mixed, or echogenic. Echolucent plaques are soft, unstable plaques with a thin fibrous cap, a necrotic lipid core, and inflammation. Echolucent plaques are more prone to rupture and cause clinically overt disease. 6 In contrast, echogenic plaques are less likely to cause clinical symptoms due to reduced inflammation and increased ECM content. 7 Until now, attempts to molecularly phenotype atherosclerotic plaques have primarily relied on transcriptomics. 8 We developed novel mass spectrometry (MS)-based methods to characterize extracellular proteins. 9, 10 We successfully applied these methods to a small cohort of atherosclerotic plaques from patients with symptomatic and asymptomatic carotid stenosis, revealing a molecular signature with known and novel inflammatory biomarkers that could be measured in plasma and predicted future cardiovascular events. 11 Combining large-scale tissue proteomics with bioinformatics, we strove to reveal molecular types of human atherosclerotic lesions, study their associations with the current classifications by histology and imaging techniques and relate them to cardiovascular outcomes in 2 independent cohorts.

Novelty and Significance
What Is Known?
• Atherosclerotic plaques are currently characterized using imaging and histology. • Transcriptomic signatures have been associated with plaque subtypes.

What New Information Does This Article
Contribute?
• This proteomics study doubles the coverage of the plaque proteome in large cohorts. • Calcification and inflammatory protein signatures of atherosclerosis were identified and integrated with bulk, single-cell, and spatial transcriptomics. • A sex-specific signature, enriched in large-aggregating proteoglycans, contributed to grouping patients into clusters with different outcome trajectories.
We generated the largest proteomics data set of fresh-frozen human plaques to date, combining untargeted and targeted proteomics with network reconstruction and clustering techniques to gain molecular insights into protein changes related to clinical characteristics. The proteomic signatures of plaque composition were found to predict long-term, adverse cardiovascular events, improving risk stratification for patients with advanced atherosclerosis. Notably, the proteoglycan content of plaques was inversely associated with serum estradiol levels and contributed to the stratification of predominantly female patients into a subgroup associated with worse cardiovascular outcomes.

Data Availability
Summary data and additional analyses that support the findings of this study are available from the corresponding authors on reasonable request. Proteomics and spatial transcriptomics data were deposited at the Proteomics Identifications (PRIDE) and Gene Expression Omnibus (GEO) repositories, respectively.

Patient Cohorts
One hundred twenty patients with carotid artery stenosis undergoing carotid endarterectomy were included in the discovery cohort from Medical University of Vienna, Austria. The validation cohort consisted of 200 patients undergoing carotid endarterectomy from the Athero-Express Biobank study. 12

Statistical and Bioinformatics Analysis
The proteins identified in the GuHCl and NaCl extracts were filtered for ECM and extracellular proteins. The data set was further filtered for missing values. The limma package was used to compare different phenotypes using the Ebayes algorithm correcting for selected covariates. For discovery proteomics nominal P values were used to identify dysregulated proteins. P values corrected for multiple testing are provided in the Supplemental Material. For validation in the Athero-Express cohort, correlation, and enrichment analyses, significance was based on Benjamini-Hochberg corrected P values. Proteins belonging to the revealed proteomics signatures were considered potential inputs to classification models to predict the primary end points. An ensemble dimensionality reduction technique was used deploying a multiobjective evolutionary algorithm to identify the optimal feature subset to be used as input to a support vector machines model and optimize the support vector machine's parameter values. A more detailed methods description is provided in the Supplemental Material.

Expanding the Coverage of the Human Plaque Proteome
Our discovery cohort consisted of 219 human carotid endarterectomy specimens from 120 patients (Table S1).
No statistically significant differences were observed between symptomatic and asymptomatic patients based on ultrasound or histology plaque classification. We validated our findings using carotid plaques from 200 patients from the Athero-Express Biobank. 12 Statistical comparisons were corrected for statin use, which was higher in symptomatic patients. With proteomics data collected from 419 carotid endarterectomy specimens, this study surpasses previous carotid plaque proteomics studies 11,14-19 ( Figure S1). Quantitative, multiplexed proteomics with tandem mass tags (TMT) was conducted on 3 extracts resulting in consistent quantification of 1459 predominantly cellular proteins in SDS (Summary

Distinct Cellular Proteome Characteristics of the Plaque Core and Periphery
The carotid endarterectomies from the discovery cohort were divided into peripheral and core specimens. The proteomics analysis of cellular protein extracts (SDS) differentiated core from periphery of the plaques (Figure 1A). Spatial RNAseq data (Differential Expression Results Comparing Clusters of Spatial RNASeq Data From Carotid Plaques in the Supplemental Material) supported the findings on regional differences. Enrichment analysis of proteomics and spatial transcriptomics data revealed 3 positional clusters: cluster 1 representing changes in the periphery, cluster 2 in the core, and cluster 3 corresponding to an intermediate area (Figure 1B). Changes in cellular protein content in plaque cores were associated with clinical characteristics, particularly calcification, symptomatic status, hyperlipidemia, and sex ( Figure 1C). The significant changes in the SDS fraction were primarily related to lysosomal degradation, ECM, and ECM-associated proteins ( Figure 1D). Quantification of cell markers for smooth muscle cells Proteomic Atlas of Atherosclerosis antigen CD14) and CD45 (PTPRC [receptor-type tyrosine-protein phosphatase C]) showed positive correlations with inflammation-related proteins and an inverse relationship to structural ECM proteins, especially collagens. In contrast, alternative macrophage markers such as C163A (scavenger receptor cystein-rich type 1 protein M130) and MRC1 (mannose receptor 1) did not show an association with the loss of structural ECM proteins.

Extracellular Protein Changes in Calcified and Symptomatic Plaques
To explore the impact of calcification on plaque proteome composition, we compared cores of calcified and noncalcified plaques for soluble and core matrisome proteins, respectively (Figure 2A and 2B). Among the most upregulated proteins in calcified plaques were FETUA (fetuin-A), OSTP (osteopontin), and gamma-carboxylated proteins such as MGP (matrix Gla protein), GAS6 (growth arrest-specific 6), FA9 and FA10 (coagulation factors IX and X), prothrombin (THRB), and vitamin-Kdependent proteins C and Z (PROC, PROZ). Additionally, an increase in C apolipoproteins (APOC1-4) was observed in calcified plaques. Conversely, inflammatory proteins, such as S10A8/A9 (calprotectin) and PERM (myeloperoxidase), and proteins associated with iron transport, including TRFL (lactotransferrin), FRIL and FRIH (ferritin light and heavy chain) and HPT (haptoglobin), were reduced. No corresponding changes were observed at transcript level in a publicly available bulk RNAseq data set comparing calcified (n=11) and noncalcified fibroatheromas (n=9; Figure 2C). The proteomic signatures associated with calcification and inflammation were inversely correlated (r=−0.615 and r=−0.595 for soluble and core matrisome, respectively, both P<0.001). By conducting a comparative analysis between changes related to calcification (Figure 2A and 2B) and changes between symptomatic and asymptomatic patients ( Figure 2D and 2E), we observed overlapping protein changes, which were labeled in bold font. Symptomatic plaques showed increased levels of FRIL and FRIH, CD44 (the hyaluronan receptor), and CD97 (a receptor essential for leukocyte migration).
After excluding calcified plaques ( Figure S3), FRIL, FRIH, CATD (cathepsin D), and MMP9 (matrix metalloproteinase 9) remained the most upregulated proteins in symptomatic plaques. The upregulation of MMP9 and cathepsins is consistent with the findings of our previous proteomics study. 16 The increase in FRIL and FRIH in symptomatic plaques was confirmed at the transcript A and B, Volcano plots of significantly dysregulated proteins in the calcified (n=60) vs noncalcified (n=46) core plaque comparison for soluble (NaCl; A) and core (GuHCl; B) matrisome extracts with 283 and 286 consistently quantified extracellular proteins, respectively. Proteins significant in both calcified vs noncalcified and symptomatic vs asymptomatic comparisons are labeled in bold. C, Heatmap displays the log2 fold changes (FC) between calcified (n=11) and fibroatheroma (n=9) plaques and nominal P values from the RNA-sequencing data set GSE104140 for the corresponding transcripts of proteins significantly changing in both calcified vs noncalcified and symptomatic vs asymptomatic comparisons. D and E, Volcano plots of significantly dysregulated proteins in the symptomatic (n=36) vs asymptomatic (n=69) comparisons of the plaque cores for the soluble matrisome (NaCl; D) and the core (GuHCl; E) matrisome extracts with 283 and 286 consistently quantified extracellular proteins, respectively. F, Heatmap displays the log2 FC between unstable (n=4) and stable (n=4) carotid plaques and nominal P values from the RNAseq experiment GSE120521 for the corresponding transcripts of proteins significantly changing in both symptomatic vs asymptomatic and calcified vs noncalcified comparisons. Differential protein analysis was conducted using the Ebayes method of the limma package correcting for age, sex, and statins. Differential expression analysis for RNAseq data was performed without corrections. Proteins with 0.01≤P<0.05, 0.001≤P<0.01, and P<0.001 are highlighted in green, orange, and red colors, respectively. Proteomic Atlas of Atherosclerosis level using a publicly available bulk RNAseq data set comprising stable and unstable carotid plaque samples (n=4 per group, Figure 2F). Additionally, gene expression of GAS6, MGP, IBP3 (insulin-like growth factor binding protein 3), and LTBP4 (latent-transforming growth factor beta-binding protein 4) was reduced in symptomatic plaques, consistent with the results of our present proteomics study. Further correlation analysis was conducted between FRIL and FRIH with other serum proteins ( Figure S4). In the plaque matrisome extracts, ferritin levels neither correlated with hemopexin and haptoglobin nor with albumin (r=−0.21 and r=−0.24, with P>0.05 for FRIH and FRIL, respectively). In the cellular SDS extracts, ferritin levels moderately correlated with hemoglobin (r=0.43 and r=0.39 for FRIH and FRIL, respectively). For a subset of the cohort (n=57), we also had measurements of serum ferritin. Ferritin levels in plaques only weakly correlated with ferritin levels in the circulation (r=0.28 and r=0. 33 for FRIH and FRIL, respectively). This correlation was of similar strength as the correlation seen for hemoglobin and hematocrit. Thus, the increase in ferritin in plaques of symptomatic patients is not a result of increased contamination by serum proteins.

Validation by Targeted Proteomics
To validate the findings from our untargeted proteomic analysis, we conducted targeted proteomics (parallel reaction monitoring [PRM]) to quantify 135 matrisome proteins by MS in the discovery cohort (Summary Details of Proteins and Peptides Quantified using Targeted Proteomics PRM Method and Summary Results and Statistics of the PRM Targeted Proteomics Analysis Using the GuHCl Extract in the Supplemental Material). Targeted proteomics results showed good agreement with those of the untargeted discovery approach, with correlations exceeding r=0.6 (P<0.001).
Ranking changes in the inflammation signature in the plaque cores, based on their effect size revealed that CD14 and PERM showed the most pronounced changes ( Figure 3A). Furthermore, PERM, TIMP1 (metalloproteinase inhibitor 1), S10A8/A9, and DEF1 (neutrophil defensin 1) were most strongly associated with neutrophils ( Figure 3B), whereas the abundance of cathepsins (CATD, CATB) demonstrated a stronger correlation with macrophage markers ( Figure 3B). This cellular distribution was further supported by network reconstruction analysis of the cellular proteome filtered for cell receptorassociated proteins, which revealed 9 significant clusters ( Figure S5). Consistent with the M1 and M2 paradigm, the M1 cluster contains CD14, while the M2 cluster contains C163A and MRC1 (CD206). The strong association of cathepsins with CD14 suggests that cathepsins may be predominantly expressed in M1 macrophages. The neutrophil cluster contained several proteins that constituted the inflammatory signature of symptomatic plaques, including S10A8/A9, MMP9, and TIMP1. 11 Next, we ranked the proteins of the calcification signature and identified OSTP and PROC as among the most upregulated proteins in cores of calcified plaques ( Figure 3C). Notably, although most proteins of the calcification signature were positively correlated with SMCs and negatively with neutrophils, no significant correlations were observed between FA9, FETUA, and OSTP and any cellular marker proteins, indicating that these proteins were deposited rather than expressed in calcified plaques.

Sex Differences in Plaques Attributed to Calcification
Aside from calcification and symptomatic status, sex was identified as another important covariate in the proteomics analysis ( Figure 1C). Given that previous studies have highlighted sex-specific gene expression changes in atherosclerosis, 20 we compared sex-specific protein abundance changes in our study (Table S2). We reconstructed a matrisome network for the plaque cores and used hierarchical clustering to analyze the changes between females and males in all clusters (outer circle) for soluble proteins and the core matrisome, respectively ( Figure 4A and 4B). Comparing sex differences to protein changes in calcified versus noncalcified plaques (middle circle), we found that the changes between plaques from female and male patients were highly correlated with the changes in calcification (r=0.808 and r=0.795 for log2 fold changes in soluble and core matrisome, respectively, both P<0.001). Conversely, sex differences were inversely associated with changes in symptomatic plaques (inner circle).
Four clusters of soluble matrisome proteins (Figure 4A: C4, C6, C17, C19) and 4 clusters of the core matrisome ( Figure 4B: C16, C17, C19, C20) were significantly altered with sex. However, most of these protein changes were associated with calcification and inflammation, such as those in C4 containing gamma-carboxylated proteins or in C19 containing neutrophil-derived proteins. Only the sex-associated protein changes constituting C17 and C20 appeared to be independent of calcification and inflammation. C17 and C20 contained the 2 large-aggregating proteoglycans PGCA (aggrecan) and versican (CSPG2) and members of the HPLN (hyaluronan and proteoglycan link) protein family (HPLN1, HPLN3). As a result, the top significantly enriched pathways or GO (Gene Ontology) functional terms were growth-factor binding for C17 and cell adhesion and hyaluronic acid binding for C20 ( Figure S6).
We validated the sex-related changes in clusters 17 and 20 using label-free proteomics data from 200 carotid plaques of the Athero-Express Biobank (Table  S3)  Proteomic Atlas of Atherosclerosis to sex, 11 were confirmed, including SODE (superoxide dismutase), HPLN3, and the 2 large-aggregating proteoglycans CSPG2 and PGCA ( Figure 5A). Moreover, both cohorts exhibited an inverse correlation between sex and inflammation ( Figure 5B). Estradiol levels were available in a subset of female patients in the Atheroexpress Biobank (n=22), and analysis showed an inverse correlation with the CSPG2, HPLN3, and PGCA content of atherosclerotic plaques ( Figure 5C), providing further evidence of the involvement of sex hormones in regulating plaque ECM composition.
by increased expression of SELENOP (selenoprotein P), macrophage inflammatory protein 1 beta (CCL4L2) and several complement factors. High OSTP (SPP1 gene) expression characterized the macrophages in cluster 5. The CellChat tool was used to further study cellcell communication patterns. OSTP was predominantly implicated in the outgoing communication of Cluster 5 macrophages via CD44 ( Figure S10C and S10D). The hyaluronan receptor CD44 was increased in symptomatic lesions ( Figure 2D). The macrophages in cluster 8 were characterized by an upregulation of calprotectin (S10A8/A9).

Comparison to Imaging Classification by Ultrasound
The reconstructed matrisome correlation network in Figure 7A depicts the plaque proteins associated with the inflammation and calcification signatures and their relationship to the current ultrasound classification. Higher levels of CATD and NID2 (nidogen 2), a basal lamina protein, were observed in echolucent plaques, whereas echogenic plaques were characterized by increased levels of calcification-related proteins (PROC, OSTP), fibrillar collagens (CO1A1 [collagen alpha-1(I) chain], CO1A2 [collagen alpha-2(I) chain]), DERM (dermatopontin), and SDF1 (stromal cell-derived factor 1; Figure 7B). Comparing calcified lesions versus fibroatheroma based on histology (Figure 7C) confirmed an increase in PROC, OSTP, and SDF1 levels. Thus, echogenic plaques reflect predominantly calcified lesions.

Identifying Molecular Plaque Phenotypes Based on Proteomics
Through the integration of proteomic measurements of both the cellular and matrisome proteome, principal component analysis and clustering analyses were   Figure 8B. Clusters 1 and 2 included predominantly plaques from male patients with a high degree of inflammation. Yet, cluster 2 had markedly worse outcomes ( Figure 8C). Cluster 2 was separated from cluster 1 by a loss of SMC markers, reduced abundance of CSPG2, and an increase in the platelet-specific ITA2B (integrin alpha IIB). Clusters 3 and 4 included most female patients, had less inflammation, but a higher SMC and CSPG2 content. Cluster 3, however, differed by the higher abundance of proteins linked to calcification. Although classification based on ultrasound and histology did not achieve an area under the curve (AUC) >50.0% in the discovery cohort, unsupervised clustering based on proteomics achieved an AUC of 60.4% (Table  S4). Cluster 2 had the strongest association with adverse cardiovascular outcomes ( Figure 8C) and consisted only of males with echolucent plaques and reduced structural ECM proteins ( Figure 8B and 8D). However, most females were grouped into cluster 3, which was also strongly associated with adverse cardiovascular events despite representing calcified, echogenic plaques with high CSPG2 content. Cluster 1, representing symptomatic mixed plaques, and cluster 4, which represented SMC-rich plaques, were associated with a better prognosis.
To explore whether plasma measurements could identify plaque clusters, we expanded our previously identified inflammation signature 11 (MMP9, CATD, FN1 [fibronectin], calprotectin, or S10A8/9) with PERM, which was one of the most upregulated proteins in symptomatic plaques in our targeted proteomics analysis. We trained and tested elastic-net logistic regression models to predict the plaque proteomic clusters by combining the inflammatory plasma biomarkers, with the ultrasound classification of plaques and the available clinical information from the patients. Among the clinical, ultrasound, and plasma biomarkers, all 5 inflammatory plasma biomarkers were selected using the elastic-net approach. Results of 10-fold cross-validation revealed that the logistic regression models were able to predict plaque clusters with 82.29% AUC.
Finally, we utilized a multiobjective evolutionary optimization algorithm to identify the minimum plaque signature that performs best as input for a nonlinear classifier (support vector machines). A 4 protein signature, consisting of CNN1, PROC, the collagen-binding protein SERPH1 (serpin H1), and CSPG2, returned an AUC of 75.0% using 10-fold cross-validation. When tested on the Athero-Express Biobank study, the proteomics signature outperformed the histology-based Plaque Vulnerability Index with an AUC of 67.5% compared with 51.0% (Table S4).

DISCUSSION
This study represents the most comprehensive proteomics analysis of human atherosclerotic plaques to date, incorporating proteomics analyses of 419 freshfrozen plaque samples from 320 patients, integrated with bulk, scRNAseq, and spatial RNAseq, carotid ultrasound, and histology. The analysis linked inflammation and calcification signatures to clinical characteristics, such as sex, and validated prognostic signatures for cardiovascular outcomes in an independent cohort. Unlike previous studies that relied on bulk RNAseq data, 8 our study is the first to use the plaque proteome to define plaque clusters and annotate them based on clinical characteristics and cardiovascular outcomes.

Proteomics of Human Atherosclerosis
Prior proteomics studies on carotid plaques 11,[14][15][16][17][18][19] were limited to less than 20 patients and conducted labelfree proteomics analysis. Matic et al 19 were among the first to apply TMT multiplexing to carotid plaques from 18 patients. Larger studies are needed to identify plaque subtypes. The largest previous proteomic study 21 on coronary plaques used autopsy samples. The use of fresh, frozen samples and the focus on ECM in the present study doubled the coverage of the human plaque proteome. Core and periphery samples were separated as the 2 regions differed in their proteomic profile and periphery samples can have a variable contribution from the normal vessel wall.

Inflammation Signatures of Atherosclerosis
Inflammatory proteins play an important role in atherogenesis. 22 In line with our previous results, the inflammatory signature of symptomatic plaques included MMP9 and cathepsins. 11 MMP9 was correlated with neutrophils, whereas cathepsins were associated with macrophages. The importance of neutrophils in advanced atherosclerosis is further highlighted by PERM as being one of the most significantly upregulated proteins in our targeted proteomics analysis of symptomatic plaques. As shown by Silvestre-Roig et al, 22  PTPRC, receptor-type tyrosine-protein phosphatase C; PROC, vitamin K-dependent protein C; SMC, smooth muscle cells; S10A8, protein S100-A8; S10A9, protein S100-A9; TAGLN, transgelin; and TIMP1, metalloproteinase inhibitor 1. Proteomic Atlas of Atherosclerosis the rise in ferritin along with other proteins involved in iron transport. Ferritin can promote plaque destabilization through the oxidation of lipids by iron overload. 23 Elevated body iron stores have been previously associated with increased risk of carotid atherosclerosis. 23 Ferritin changes in plaques, however, were not due to serum contamination. Both scRNAseq and spatial RNAseq data confirmed that ferritin is expressed by plaque macrophages. Although RNA reflects gene transcription at the time of tissue collection, extracellular proteins tend to accumulate over time, and protein deposition is better captured by proteomics.

Calcification Signatures of Atherosclerosis
Consistent with the notion that calcification dampens vascular inflammation, the proteomics signatures linked to plaque calcification and inflammation were inversely correlated. Vascular calcification results from an imbalance between stimulatory mediators and calcification inhibitors, such as MGP 24 and FETUA. 25 The upregulation of both proteins in calcified plaques suggests a compensatory increase of calcification inhibitors. In addition to MGP, proteomics identified the upregulation of other gamma-carboxylated proteins in calcified plaques, such as FA9 and FA10. Gamma-carboxylation is a key calcification regulator at cellular level. 26 This vitamin K-dependent post-translational modification replaces glutamate residues with gamma-carboxyglutamate. Since no patient was under warfarin treatment, an oral vitamin K antagonist, the observed differences in gamma-carboxylated proteins are not related to medication. Several gamma-carboxyglutamate residues are contained within the Gla domain of certain proteins, responsible for the high-affinity binding of calcium ions. Thus, Gla domain-containing proteins may be retained in calcified plaques. This is consistent with the observation that transcriptomics data revealed no corresponding changes in gene expression for FA10 and FETUA. Typically, there is a poor yield of RNA from calcified lesions. Karlöf et al 27 performed a high versus low calcification comparison on carotid plaques using microarray data from 40 patients. When comparing their top significant transcriptomic changes to our proteomic changes, abundant collagens (CO1A1 and CO1A2) and PRG4 (proteoglycan 4) were found to be significantly upregulated both at the transcript and the protein level. In contrast, APOC1 was increased in the proteome but reduced in the transcriptome of calcified plaques, consistent with a potential plasma origin. When exploring changes between symptomatic and asymptomatic plaques using only noncalcified plaques, PROC and FETUA were upregulated in asymptomatic plaques, suggesting a role in the early calcification response or plaque stability.

Sex Differences in Atherosclerosis
Although the pathobiology of atherosclerotic cardiovascular disease is generally similar between both sexes, there are unique aspects specific to women, with the coronary artery calcium scoring, for instance, predicting mortality better in women than in men. 28 In a recent study 20 using RNAseq, sex-stratified gene regulatory network analysis identified GAS6 and serpin G1 as potential key SMC driver genes in plaques from females compared with males. In our cohort, GAS6 was also more abundant in plaques of female patients, but serpin G1 was unchanged at the protein level. GAS6, however, was part of the calcification signature, and calcification was more common in carotid plaques from female patients. When further comparisons were restricted to noncalcified plaques, GAS6, and other calcification-related proteins were not different between plaques from male and female patients. Thus, these changes with sex might be explained by a higher susceptibility to calcification in age-matched female patients. When excluding protein changes associated with calcification, only clusters containing the 2 large-aggregating proteoglycans, CSPG2 and PGCA, were found to be sex-related. When calcified plaques were excluded, differences in PGCA and CSPG2 remained as the protein changes most prominently associated with sex along with hyaluronic acid-binding link proteins HPLN1 and HPLN3 (data not shown). These sex-related matrisome changes were validated in the Athero-Express cohort. 12 The Role of Large, Aggregating Proteoglycans in Atherosclerosis CSPG2 and PGCA are expressed in a subset of SMCs in human carotid plaques. Large, aggregating proteoglycans play an important role in regulating the mechanical properties of the vessel wall 29 by controlling tissue hydration and vascular wall viscosity. CSPG2 is known to contribute to lipoprotein retention in atherosclerosis. In murine atherosclerosis, Hartmann et al 30 identified a transition state SMC phenotype linked with plaque stability that expressed ECM proteins and PGCA. We have recently uncovered an increase in CSPG2 and PGCA in stented porcine coronary arteries. 31 Proteases of the ADAMTS (a disintegrin and metalloproteinase with thrombospondin motifs) family partly control the accumulation of large proteoglycans after vascular injury. 29 Proteoglycan deposition could represent a missing link between lipoprotein retention and hemodynamics, consistent with the known predilection sites for atherosclerosis. In this study, we report differences in hyaluronan-binding proteoglycans in carotid plaques between males and females. This observation was further supported by an inverse association of the hyaluronan-binding proteoglycan content in plaques to circulating estradiol levels in female patients.

Plaque Classification by Ultrasound
Carotid duplex ultrasound is commonly used as a firstline diagnostic tool to assess extracranial carotid artery stenosis. 32 Echolucent plaques are associated with an increased risk for future ischemic cerebrovascular events. 33 Carotid plaque area as well as plaque neovascularization detected in carotid ultrasound are also predictive for future cardiovascular events. 34 However, this imaging prediction is mainly based on the separation of advanced from early stages of atherosclerosis. In our study, all patients had advanced atherosclerosis. The proteome changes corresponding to echogenic plaques were predominantly fibrillar collagens and few proteins of the calcification signature. Although ultrasound imaging captures the collagen and lipid content of the plaque, it is not able to distinguish between complex lesions with high inflammation and stable fibroatheroma with lipid deposits and low-grade inflammation. Thus, proteomic signatures could complement carotid imaging read-outs to improve prognostication.

Plaque Classification by Histology
Histological evaluation with American Heart Association classification is performed after surgery to assess morphological properties of atherosclerotic plaques. 35 Although there are some data regarding histological plaque properties and patient outcome in contemporary cohorts, such as the ability of a histology-based Plaque Vulnerability Index to predict outcomes, 36 symptomatic carotid stenosis is defined as the occurrence of symptoms ipsilateral to stenosis during the preceding 6 months. 32 Given that atherosclerotic plaque morphology is a dynamic process, most plaques undergo several subclinical ruptures and healings at the stage of advanced lesions. 37 Therefore, the in-depth molecular characterization offered by proteomics may better capture the evolution of plaques over time. Expanding on the traditional classification of vulnerable and nonvulnerable plaques, male patients with molecular features of vulnerable plaques (cluster 2) had the worst prognosis followed by a subgroup of predominantly female patients with calcified lesions (cluster 3) with large-aggregating proteoglycans driving most of the clustering. Histology captured only a subset of these proteomic changes. Thus, assessment of plaque proteome content even at a single location, together with clinical risk stratification may improve identification of vulnerable patients at risk of future cardiovascular events.

Risk of Recurrent Events
Atherosclerosis is a systemic disease. Patients with atherosclerosis at multiple vascular beds are at particularly high risk for atherothrombotic events, and each affected territory increases the risk not only of local organ damage but also the total risk of cardiovascular events. 32 This warrants more aggressive treatment of cardiovascular risk factors for secondary prevention in these patients. 32 In a previous study, plaque calcification and plaque eccentricity were positively correlated to plaque burden in both peripheral and coronary artery disease. 38 We show here that proteomic characteristics of carotid plaques are associated with long-term cardiovascular events. Echolucent plaques with reduced structural ECM proteins had the worst prognosis (cluster 2), followed by calcified plaques (cluster 3). Although calcified plaques are traditionally considered stable lesions with low risk of embolization, the incidence of preoperative neurological symptoms was similar in patients with calcified and noncalcified plaques. 39 Similarly, recent findings suggest that stable calcified plaques in coronary arteries are associated with high cardiovascular risk. 40 In our cohort, calcified plaques were more common in female patients, highlighting potential sex differences also in advanced stages of atherosclerosis. A signature of 4 proteins, as revealed by machine learning predicted the composite cardiovascular end point with high accuracy in both the discovery and validation cohort over a 9 and 3-year follow-up periods, respectively. The signature included the SMC marker CNN1, the gamma-carboxylated protein PROC, the collagen-binding protein SERPH1 and the hyaluronan-binding proteoglycan CSPG2. Although plaque ECM proteins may not be readily detectable in the circulation, inflammatory plasma proteins may aid in classifying patients. The identified plaque clusters could be predicted using inflammatory plasma biomarkers (MMP9, CATD, FN1, calprotectin or S10A8/9, PERM) in combination with clinical characteristics (82.29% AUC, using 10-fold cross-validation).

Study Limitations
Carotid endarterectomies represent advanced atherosclerotic plaques, which differ from early atherosclerosis. Moreover, all patients in the study were white patients. In addition, future comparisons should include plaques from different locations within the arterial bed, which have been associated with distinct genetic susceptibility loci. 41 Coronary plaques, however, are more difficult to attain than fresh-frozen samples. Furthermore, MS has limitations in detecting very low-abundant proteins. Finally, our spatial RNAseq analysis was limited by the small sample size (n=2).

Conclusions
In the largest proteomics data set of fresh-frozen human plaques to date, we combined untargeted and targeted proteomics with network reconstruction and clustering techniques to provide molecular insights into protein changes related to clinical characteristics. Our findings demonstrate the potential of proteomics to complement