Abstract
Background:
Unstable carotid artery plaques are an important risk factor for ischemic stroke, and their clinical prognosis is poor. The present study to systematically investigate the metabolic changes of carotid plaques and use machine learning methods to identify and screen metabolic biomarkers in unstable carotid plaques for helping diagnosis of stroke risk caused by unstable plaques.
Method:
A non-targeted metabolomics analysis was performed on 67 cases (40 stable and 27 unstable) of carotid artery plaques. Specific metabolic signatures were identified in unstable plaques. Four machine learning algorithms, including random forest (RF), support vector machine (SVM), least absolute shrinkage and selection operator (LASSO), and logistic regression (LR), were used to construct feature analysis models for unstable carotid artery plaques and predict the associated metabolic biomarkers.
Results:
A total of 98 metabolites significantly differentially associated with unstable plaques were identified. Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis revealed that the cGMP-PKG signaling pathway, glucagon signaling pathway, central carbon metabolism in cancer, and lipolysis regulation in adipocytes are metabolic pathways significantly associated with unstable plaques. The network diagram of metabolites and metabolic pathways revealed the relationship between 43 metabolites and their corresponding pathways. Furthermore, some metabolites that may serve as biomarkers for unstable plaques were screened.
Conclusion:
Different metabolite patterns associated with unstable plaque tissue were identified and characterized. This study identified some potential metabolic biomarkers significantly associated with unstable carotid artery plaques, which can predict metabolic products and further improve the prediction of stroke risk in unstable plaques.
1 Introduction
Atherosclerotic plaques can cause carotid artery lumen stenosis, reduce blood flow to brain cells, or rupture and become dislodged, leading to stroke events. Stroke is one of the causes of adult disability (1). Early lesion symptoms are not obvious and rarely receive attention in the process of carotid artery plaque formation (2). Intervention, such as carotid endarterectomy, typically takes place when the degree of carotid artery stenosis is high (3). However, cerebrovascular events may also occur in patients with low-grade carotid artery stenosis (<30%) and no other obvious causes of stroke (4). Therefore, early detection of carotid atherosclerotic disease and reliable identification of plaque instability are of great significance for stroke prevention.
Carotid artery plaques are divided into stable and unstable plaques (5). Unstable carotid artery plaques are characterized by thin fibrous caps and large lipid cores (1), which are the initial changes leading to plaque rupture (5, 6). Advances in imaging technology have allowed us to detect carotid plaque features, and Doppler ultrasound imaging can identify stable and unstable plaques (7, 8). Unstable plaques include hypoechoic lipid plaques, mixed plaques with significantly uneven echoes, ulcerative plaques with uneven surfaces, and bleeding plaques without internal echoes (9). In addition, there are reports suggesting that metabolic disorders of carbohydrates, lipids, and amino acids can lead to plaque instability. Therefore, the structure, composition, and production of metabolites within plaques may be more direct markers of plaque rupture compared to the degree of luminal stenosis and stroke risk. Thus, the biological metabolic basis of unstable plaques needs to be explored further (10).
Metabolomics is an emerging omics technique that involves detailed quantification of small molecule metabolic markers in biological samples that can then be used to identify new biomarkers (11), diagnose and monitor diseases, and characterize metabolic pathways in disease pathogenesis (12). At present, the analysis of related metabolites in human atherosclerotic plaque tissue using metabolomics has been reported (10), but scarce studies have focused on the metabolomic characterization of stable/unstable plaques. Our research incorporates clinical data and applies machine learning techniques.
To comprehensively study the metabolic characteristics of unstable carotid plaques, the present study used multivariate analysis to identify 98 potential metabolic biomarkers of unstable plaques based on untargeted metabolomics. Metabolic pathway enrichment analysis was performed to explore the metabolic pathways and metabolites enriched in unstable plaques. Subsequently, four machine learning algorithms, including RF, SVM, LASSO, and LR, were used to further identify metabolic biomarkers associated with unstable plaques, which can further improve the prediction of stroke risk in unstable plaques.
2 Methods
2.1. Study design and participants
The present case-control study was carried out at the Xi'an Tangdu Hospital as a part of two concurrent multicenter cohort studies (registered on ClinicalTrails.gov, NCT06027463 and NCT06120478). A total of 67 patients with carotid artery stenosis were included in the study between June 2022 and August 2023. All participants met the following selection criteria: (1) patients over 18 years old diagnosed with carotid artery stenosis and receiving carotid endarterectomy (CEA) for the first time; (2) preoperative imaging examination confirmed carotid artery stenosis of >70%; (3) informed consent form was provided; and (4) follow-up after surgery was performed at the treatment institution. Participants with the following characteristics were excluded from the study: (1) age of <18 or >85 years; (2) refractory hypertension; (3) severe bleeding tendency; (4) complete occlusion of the responsible carotid artery (100%); (5) multiple stenosis of the unilateral carotid artery beyond the surgical range; (6) untreated intracranial large aneurysm; (7) severe surgical contraindications; and (8) other malignant diseases, such as malignant tumors, or expected death within one year. The study was approved by the Medical Ethics Committee of Tangdu Hospital (K202211-29). All carotid artery plaques were divided into stable and unstable plaque groups after combining the patients' carotid artery ultrasound examination results. The unstable group included ulcerative, hemorrhagic, and echolucent plaques, while the stable group included echogenic plaques. Mixed echogenic plaques and plaques that were difficult to classify were excluded. The plaque characteristics were reviewed and confirmed by two senior professional carotid and cerebrovascular ultrasound physicians. A total of 67 patients were included in the final metabolomics analysis: 40 in the stable group and 27 in the unstable group. The patients and/or their caregivers signed informed consent in accordance with the standards of the Declaration of Helsinki.
2.2. Sample collection and preparation
Carotid artery plaque tissue was used to identify metabolites in the present study. Carotid artery plaques obtained during carotid endarterectomy were immediately processed and stored in polypropylene tubes at −80 ℃. Then, 50 mg of solid sample was added to a 2-mL centrifuge tube together with a grinding bead 6 mm in diameter. Next, 400 μL of extraction solution [methanol: water = 4:1 (v:v)] containing 0.02 mg/mL of the internal standard (L-2-chlorophenylalanine) was used for metabolite extraction. Samples were ground using the Wonbio-96c (Shanghai wanbo biotechnology Co., Ltd.) frozen tissue grinder for 6 min (−10 °C, 50 Hz), followed by low-temperature ultrasonic extraction for 30 min (5 °C, 40 kHz). The samples were then kept at −20 °C for 30 min, after which they were centrifuged for 15 min (4 °C, 13,000 g). The resulting supernatant was transferred to the injection vial for liquid chromatography-tandem mass spectrometry (LC-MS/MS) analysis.
2.3 Quality control sample
As a part of the system conditioning and quality control (QC) process, a pooled QC sample was prepared by mixing equal volumes of all samples. The QC samples were disposed and tested in the same manner as the analytic samples. The whole sample set was injected at regular intervals (every 5–15 samples) in order to monitor analysis stability.
2.4 Ultra-high-performance liquid chromatography tandem mass spectrometry (UHPLC-MS/MS) analysis
The LC-MS/MS analysis was conducted using a Thermo UHPLC-Q Exactive HF-X system equipped with an ACQUITY HSS T3 column (100 mm × 2.1 mm i.d., 1.8 μm; Waters, USA) at Majorbio Bio-Pharm Technology Co. Ltd. (Shanghai, China). The mobile phases consisted of 0.1% formic acid in water:acetonitrile (95:5, v/v; solvent A) and 0.1% formic acid in acetonitrile:isopropanol:water (47.5:47.5, v/v; solvent B). The flow rate was 0.40 mL/min and the column temperature was 40 ℃. The injection volume was 3 μL.
2.4.1 MS conditions
Mass spectrometry data were collected using a Thermo UHPLC-Q Exactive HF-X mass spectrometer equipped with an electrospray ionization (ESI) source operating in positive and negative modes. The optimal conditions were set as follows: source temperature at 425 ℃; sheath gas flow rate at 50 arb; aux gas flow rate at 13 arb; ion-spray voltage floating (ISVF) at −3,500 V in negative mode and 3,500 V in positive mode; and normalized collision energy of 20–40–60 eV rolling for MS/MS. Full MS resolution was 60,000, and MS/MS resolution was 7,500. Data acquisition was performed in the Data-Dependent Acquisition (DDA) mode. The detection was carried out over a mass range of 70–1,050 m/z.
2.5 Machine learning models
2.5.1 RF
An RF model was constructed using the Majorbio Cloud Platform (cloud.majorbio.com) to identify biomarkers most likely associated with unstable carotid artery plaques. Receiver operating characteristic (ROC) analysis for the RF models was performed on the same platform.
2.5.2 SVM
ROC curve analysis was used to evaluate the diagnostic performance of potential biomarkers. The SVM classification model and ROC analyses for multiple biomarkers were performed using the Majorbio Cloud Platform (cloud.majorbio.com).
2.5.3 LASSO and logistic regression
The LASSO regression is a penalty function that finds the most representative molecules by differentially scaling down the variables. LASSO regression in the present study was used to screen metabolites specifically expressed in unstable carotid artery plaques. Logistic stepwise regression was utilized to construct a diagnostic model. In addition, ROC curves were used to assess the model's ability to differentiate between stable and unstable carotid artery plaque groups.
2.6 Data analysis and statistics
The study design and data analysis workflow are shown in Figure 1. The LC/MS raw data pretreatment was performed using Progenesis QI software (Waters Corporation, Milford, USA). A three-dimensional data matrix in CSV format was then exported. It contained sample information, metabolite name, and mass spectral response intensity. Internal standard peaks, as well as any known false positive peaks (including noise, column bleed, and derivatized reagent peaks), were removed from the data matrix, deredundant, and peak pooled. At the same time, the metabolites were identified by searching the HMDB (http://www.hmdb.ca/), Metlin (https://metlin.scripps.edu/), and Majorbio databases.
Figure 1
The data matrix obtained by searching the databases was uploaded to the Majorbio cloud platform (https://cloud.majorbio.com) for data analysis. The data matrix was then pre-processed. First, at least 80% of the metabolic features detected in any set of samples were retained. After filtering, the minimum value in the data matrix was selected to fill the missing value and each metabolic signature was normalized to the sum. To reduce the errors caused by sample preparation and instrument instability, the response intensities of the sample mass spectrometry peaks were normalized using the sum normalization method to obtain the normalized data matrix. The QC sample variables with a relative standard deviation (RSD) of >30% were excluded and log10 logarithmicized to obtain the final data matrix for subsequent analysis.
Then, the R package “ropls” (version 1.6.2) was used to perform principal component analysis (PCA), orthogonal least partial squares discriminant analysis (OPLS-DA), and seven-cycle interactive validation evaluating model stability. The metabolites with variable importance on projection (VIP) > 1 and p < 0.05 were determined to be significantly different based on the variable importance in the projection (VIP) obtained using the OPLS-DA model and the p-value generated by Student's t test.
Differential metabolites among two groups were mapped into their biochemical pathways using metabolic enrichment and pathway analysis based on the KEGG database (http://www.genome.jp/kegg/). Cytoscape version 3.7.0 was used to create a network diagram of metabolites and their related pathways. These metabolites were classified according to the pathways they were involved in or the functions they performed. Enrichment analysis was used to evaluate a group of metabolites in a function node. The annotation analysis of a single metabolite was developed into annotation analysis of a group of metabolites. Python package “scipy.stats” (https://docs.scipy.org/doc/scipy/) was used to perform enrichment analysis to obtain the most relevant biological pathways for experimental treatments.
Statistical analyses were performed using SPSS 25.0 (IBM Corporation, New York, NY, USA). Two-tailed p < 0.05 was considered statistically significant.
3 Results
3.1 Patients' clinical characteristics and subsequent analysis
In order to explore the differences in metabolic characteristics of stable and unstable plaques in patients with carotid artery stenosis, a total of 67 subjects were selected and divided into stable (40 cases) and unstable (27 cases) plaque groups according to preoperative carotid artery ultrasound results. A total of 49 general clinical characteristics of the subjects are shown in Supplementary Table S1. There are certain differences in biological gender, TIA, hypertension, and levels of GLO and TP between the two groups (P < 0.05). Next, metabolomics analysis was performed on plaque tissue samples obtained during carotid endarterectomy. A total of 98 differential metabolites were identified, including 82 up-regulated and 16 down-regulated metabolites. Finally, machine learning algorithms were used to analyze the differential expression of metabolites in two groups of plaque tissues, select potential metabolic biomarkers, and provide more meaningful targets for prediction of unstable plaques (Figure 1).
3.2 Metabolic biomarker identification
The metabolomics results for carotid plaque tissue samples from 67 patients with carotid artery stenosis were analyzed using bioinformatics tools. As a result, 657 differential metabolites associated with unstable plaques were annotated, of which 387 were up-regulated and 270 were down-regulated in the unstable plaque group (Supplementary Figure 1A). The VIP values were calculated for each metabolite using PCA and OPLS-DA models. Metabolites with VIP > 1 were considered potential candidate metabolites. Out of 657 differential metabolites, 98 metabolites with VIP > 1 and p < 0.05 were screened, of which 82 were up-regulated and 16 were down-regulated (Figure 2A, Supplementary Table S2).
Figure 2
PCA analysis revealed that the inter group repeatability effect and inter group dispersion were unsatisfactory (Supplementary Figure 1B). Therefore, further calculation of the cumulative R2Y and Q2 values in the OPLS-DA model was used to estimate the goodness of fit and predictive ability of the model. The cumulative R2Y of the OPLS-DA model was 0.950, and the cumulative Q2 was 0.481. The OPLS-DA score chart shows that the stable plaque group is separated from the unstable plaque group (Figure 2B).
3.3 Metabolic pathway enrichment analysis
Metabolic pathway analysis was conducted based on 98 differential metabolites to reveal metabolic pathways associated with unstable plaques. These metabolites were involved in 89 metabolic pathways (Supplementary Table S3). Among the 89 metabolic pathways, 38 had p < 0.05. The top 20 of these pathways are shown in Figure 2C. In order to further understand the functions of KEGG enriched pathways, KEGG functional pathway enrichment analysis was conducted on 98 differential metabolites. Six of seven KEGG metabolic pathway categories were enriched, including metabolism, genetic information processing, environmental information processing, cellular processes, organizational systems, and human diseases. Among them, the lipid metabolism pathway was enriched with seven related metabolites (palmitoylcarnitine, LysoPC (15:0/0:0), 2-phospho-D-glyceric acid, L-palmitoylcarnitine, arachidonic acid, 5 (S)-HETE, and 3-phosphoglycerate; Figure 2D, Supplementary Table S4). The search for 98 differential metabolites revealed that only 43 metabolites were present in the KEGG pathway database (Supplementary Table S5). Therefore, the relationship between these 43 metabolites and the pathways they enriched was analyzed and the network of metabolites and metabolic pathways was plotted (Figure 3).
Figure 3
3.4 Association between differential metabolites and clinical indicators
The Pearson correlation coefficient was calculated to investigate the potential relationship between the top 50 differential metabolites and 49 clinical indicators (Supplementary Table S6). The correlation coefficient matrix is shown in Figure 4. Among the 50 metabolites, 15 differential metabolites were associated with MONO, 13 with TC, 12 with GLU and hypertension, 11 with NEUT, 10 with age and LDLC, eight with APTT and ALT, seven with DBIL and PCT, six with diabetes, five with WBC and HDLC, four with D-D and TP, three with PCA, sex, AST, PSA, ALB, weight, EO, and GLO, two with stroke and Fib, and one with drinking, UA, symptom, TBIL, IBIL, RBC, HGB, and HCT. The correlation coefficient was between −0.422 and 0.413. Methylene bisacrylamide had the highest correlation with sex and was also associated with four of the 49 clinical indicators (sex, D-D, TP, and GLO; Figure 4, Supplementary Table S6).
Figure 4
3.5 Predicting metabolic biomarkers and clinical indicators of unstable plaques based on machine learning algorithms
To investigate the potential of metabolomics parameters as diagnostic biomarkers, RF, SVM, and LASSO logistic regression classifiers were used to distinguish stable and unstable carotid artery plaques. Different models based on metabolites or combinations of metabolites and clinical indicators were also constructed. The RF model predicted 10 important metabolic biomarkers from 98 differential metabolites (Figure 5A). The model's ROC curve is shown in Figure 5B. After correlating 49 clinical indicators, 17 important metabolic biomarkers and three clinical indicators were predicted (Figure 5C). The model's ROC curve is shown in Figure 5D. The SVM model predicted 10 important metabolic biomarkers from 98 differential metabolites (Figure 5E). The model's ROC plot is shown in Figure 5F. After association with 49 clinical indicators, 13 important metabolic biomarkers and seven clinical indicators were predicted (Figure 5G). The model's ROC curve is shown in Figure 5H. LASSO regression predicted 10 important metabolic biomarkers (Supplementary Table S7) from 98 differential metabolites. Figure 5I shows the cross-validation error curve for the LASSO regression model. The coefficient distribution plot is shown in Figure 5J. A logistic model was constructed based on the above 10 metabolic biomarkers, which were also predicted to be the most important. The model's ROC curve is shown in Figure 5K. After correlating 49 clinical indicators, 12 important metabolic biomarkers and two clinical indicators (Supplementary Table S7) were predicted using LASSO regression. Figure 5L shows the cross-validation error curve for the LASSO regression model. The coefficient distribution plot is shown in Figure 5M. A logistic model was constructed based on the above 12 metabolic biomarkers and two clinical indicators and used to predict seven metabolic biomarkers. The ROC curve for the model is shown in Figure 5N. The diagnostic performance of these machine learning models is shown in Table 1, with an AUC range of 0.756–0.883. A metabolic ensemble of 98 differential metabolites was generated. Then, 30 metabolic biomarkers were screened using RF, SVM, and LASSO logistic regression analysis, identifying 22 metabolic biomarkers associated with unstable plaques after removing duplicate metabolites. In addition, 37 metabolic biomarkers and 12 clinical indicators were screened using RF, SVM, and LASSO logistic regression by associating 49 clinical indicators. Then, 31 metabolic biomarkers and 10 clinical indicators were screened out after removing duplicate values (Figures 5C,G, and Supplementary Table S7). Finally, 14 metabolic biomarkers (Table 2, Supplementary Figure S3) that were most relevant to unstable plaques were identified by intersecting the above 22 metabolic sets with 31 metabolic sets. The distribution of each metabolite's abundance between the two groups is shown in Supplementary Figure 2.
Figure 5
Table 1
| Machine learning algorithm models | Classification | Sensitivity | Specificit | ALIC (95% CI) |
|---|---|---|---|---|
| RF | Metabolites | 1.000 | 0.600 | 0.811 (0.620–1.000) |
| Metabolites + Clinical indicators | 1.000 | 0.730 | 0.883 (0.737–1.000) | |
| SVM | Metabolites | 1.000 | 0.800 | 0.878 (0.727–1.000) |
| Metabolites + Clinical indicators | 0.830 | 0.730 | 0.756 (0.532-0.979) | |
| LASSO + LR | Metabolites | 1.000 | 0.750 | 0.880 (0.722–1.000) |
| Metabolites + Clinical indicators | 0.780 | 0.920 | 0.880 (0.718–1.000) |
Diagnostic performance of four machine learning algorithms
AUC, area under curve; Cl, confidence interval.
Table 2
| Metabolites | Unstable group | Stable group | Regulate | FC | VIP | P value | FDR | ||
|---|---|---|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | ||||||
| N-Glycolylneuraminic acid | 4.681 | 0.5057 | 4.058 | 0.5188 | Up | 1.1535 | 3.6292 | 7.62E-06 | 0.01131 |
| LysoPS (18:0/0:0) | 4.503 | 0.4716 | 3.962 | 0.6013 | Up | 1.1365 | 2.9011 | 0.000208 | 0.04838 |
| 5'-Guanylic Acid | 5.813 | 0.5204 | 5.083 | 0.9104 | Up | 1.1436 | 3.1636 | 0.00036 | 0.04838 |
| 1-[(2S,5S)-3,4-Dihydroxy-5-(hydroxymethyl)oxolan-2-yl]-1,3-diazinane-2,4-dione | 4.473 | 0.5299 | 3.778 | 0.9784 | Up | 1.184 | 4.5081 | 0.001292 | 0.1811 |
| N-Acetyl-D-tryptophan | 6.4 | 0.3379 | 6.642 | 0.2908 | Down | 0.9636 | 2.518 | 0.002578 | 0.1811 |
| 2-(N-(7-Nitrobenz-2-oxa-1,3-diazol-4-yl)amino)-2-deoxyglucose | 4.073 | 0.5862 | 3.619 | 0.5796 | Up | 1.1254 | 3.3916 | 0.002608 | 0.1811 |
| 2,5-Dihydroxybenzenesulfonic Acid | 4.246 | 0.7643 | 3.66 | 0.7587 | Up | 1.1601 | 3.8007 | 0.002961 | 0.1237 |
| SM (d18:2(4E,14Z)/16:0) | 6.448 | 0.1837 | 6.594 | 0.1958 | Down | 0.9779 | 1.885 | 0.003111 | 0.1898 |
| 9H-Purine-9-ol | 6.596 | 0.3085 | 6.814 | 0.2802 | Down | 0.968 | 2.3501 | 0.003709 | 0.194 |
| Vidarabine | 5.356 | 0.2697 | 5.547 | 0.2445 | Down | 0.9656 | 2.2856 | 0.003785 | 0.194 |
| 2,3-Dihydroxypropyl dodecanoate | 4.157 | 0.07921 | 4.23 | 0.1085 | Down | 0.9827 | 1.2786 | 0.004076 | 0.1999 |
| PE (18:1(9Z)/O:O) | 4.702 | 0.5132 | 4.331 | 0.5305 | Up | 1.0857 | 3.0679 | 0.005902 | 0.2303 |
| CMPF | 3.804 | 0.9657 | 3.272 | 0.7254 | Up | 1.1626 | 2.9535 | 0.0124 | 0.2362 |
| D-Ribulose 5-phosphate | 5.473 | 0.2488 | 5.035 | 0.863 | Up | 1.087 | 2.0218 | 0.01278 | 0.2393 |
Metabolic biomarkers based on machine learning prediction
4 Discussion
At least 20% of ischemic strokes are thromboembolic and are caused by atherosclerotic plaques at the bifurcation of carotid or internal carotid artery (13, 14). Today, clinical guidelines for the primary and secondary prevention of stroke in patients with carotid artery stenosis caused by atherosclerotic plaques still rely on the patients’ general characteristics (hypertension, smoking status, diabetes, and hypercholesterolemia) and the static measurement of arterial stenosis degree (14). However, the degree of arterial stenosis is not the best estimate of stroke risk, and a new method of assessing stroke risk is needed to improve the risk prediction for individual patients (14). The plaque characteristics are crucial for defining their stability and stroke risk (15–17). Currently, computed tomography (CT), magnetic resonance (MR), coronary computed tomography angiography (CCTA), and positron emission tomography (PET) imaging can identify plaque instability features and directly measure disease activity in coronary arteries and large blood vessels (18). However, the application value of these methods in the early diagnosis of unstable plaques needs to be further validated. It is also important to find new methods to predict unstable plaques for stroke prevention. Metabolomics has emerged as a promising method of identifying potential biomarkers of cerebrovascular disease (19). In the present study, metabolic biomarkers and pathways of unstable plaques were screened using metabolomics with the goal to provide a convenient and low-cost solution for the early diagnosis of unstable plaques.
Metabolomics, together with machine learning algorithms, has become a powerful tool for identifying metabolic biomarkers for disease diagnosis (20). A recent study suggested that metabolomic profiles derived from nuclear magnetic resonance spectroscopy analysis of blood samples can simultaneously inform the risk of multiple diseases (21). Many metabolites in blood samples may have different tissue origins, and the analysis results may be biased. Therefore, unstable and stable carotid artery plaque tissues were selected for metabolomics analysis in the present study. Many studies have reported blood differential metabolic biomarkers and pathways related to risk prediction, early diagnosis, and prognosis of cerebral ischemia (22, 23). Several studies have also elucidated the role of microbial metabolite trimethylamine-N-oxide (TMAO) and lipopolysaccharides in cardiovascular disease and stroke (24, 25), which may trigger atherosclerosis by affecting thrombosis, inflammation, and oxidative stress (26).
In the present study, carotid artery plaques were collected from 67 patients in order to carry out a case-control study to identify 98 differential metabolites. Different models were constructed using machine learning to preliminarily screen 22 important metabolites. To increase the predictive value of plaque metabolomics, metabolomic information was matched with 49 clinical indicators to further predict biomarkers and key clinical indicators associated with unstable plaques. Finally, 14 metabolic biomarkers (with nine upregulated and five downregulated metabolites) and 10 clinical indicators were identified (Supplementary Table S7). These findings to some extent reflect the potential clinical practicality of plaque metabolomics analysis and improve the comprehensive risk assessment of unstable plaques.
N-glycylneuraminic acid (Neu5Gc) is one of the most common forms of sialic acid (27, 28). In mammals, the de novo Neu5Gc synthesis pathway begins with the activated precursor cytidine-5′- monophosphate-N acetylneuraminic acid (CMP Neu5Ac), which is catalyzed by CMP Neu5Ac hydroxylase. Very low levels of Neu5Gc can be detected in healthy human tissue (29). Studies have shown that Neu5Gc has a negative impact on the cardiovascular system and leads to the aggravation of atherosclerotic plaques (30). In the present study, Neu5Gc levels were significantly increased in unstable plaque tissues. Since these results were consistent with previous reports, Neu5Gc can serve as a potential metabolic biomarker for unstable plaques.
A variety of lysophospholipids (LPLs) are present in the body, including sphingosine 1-phosphate (S1P), lysophosphatidic acid (LPA), lysophosphatidylserine (LysoPS), and lysophosphatidylinositol (LPI). S1P and LPA have been identified as bioactive substances and are known to be implicated in the pathogenesis of atherosclerosis. In addition, LPA and LysoPS have recently been recognized as bioactive LPLs (31, 32). LysoPS is an emerging lipid mediator, possibly derived from phosphatidylserine, which is exposed to the cell membrane surface during platelet activation (33) or apoptosis (34). It is thought to be implicated in the pathogenesis of acute coronary syndrome (35). Additional studies have shown that LysoPS has several factors involved in the pathogenesis of atherosclerosis. LysoPS enhances the uptake of oxidized low-density lipoprotein, reduces the expression of inflammatory mediators, and alleviates endoplasmic reticulum stress in RAW 264.7 cells. These results suggest that LysoPS has atherogenic properties in foam cell formation and anti-atherosclerotic properties in reducing inflammation, implying that LysoPS has bidirectional properties in the pathogenesis of atherosclerosis (36). In addition, basic studies have shown that LysoPS has a variety of biological activities, such as attenuating the expression of inflammatory mediators in macrophages (36), inhibiting T lymphocyte proliferation (37), and obstructing the development and function of regulatory T lymphocytes (38). However, LysoPS had a significantly increased expression in unstable carotid plaque tissue in the present study. It may also promote the formation of foam cells in unstable plaques, thereby leading to the decline of plaque stability. Therefore, it can be recommended as a potential metabolic marker of unstable plaques.
N-acetyl tryptophan (NAT) is a neurokinin-1 receptor (NK-1R) antagonist that can disrupt the binding of substance P (SP) to NK-1R. In addition, NAT is included in the Neurodegenerative Drug Screening Alliance library consisting of 1,040 compounds compiled by the National Institute of Neurology and Stroke in the United States (39), indicating its potential as a therapeutic agent. There are three NAT isomers: N-acetyl-L tryptophan (L-NAT), N-acetyl-D tryptophan (D-NAT), and N-acetyl-DL tryptophan (DL-NAT) (39). A previous study showed that D-NAT levels in non-small cell lung cancer patients significantly increased after treatment with epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) or programmed death 1 (PD-1)/programmed cell death ligand 1 (PD-L1) inhibitors, indicating that D-NAT may have anti-tumor therapeutic effects (40). Metabolic syndrome (MetS) significantly increased the risk of atherosclerotic cardiovascular disease and type 2 diabetes (T2DM), while D-NAT level was significantly reduced in MetS patients. In addition, D-NAT at the physiological level has anti-inflammatory and antioxidant properties, making it an attractive biomarker for the occurrence of MetS due to the critical role of inflammation in MetS (41). The present study results indicated that D-NAT level was significantly reduced in the unstable plaque group compared to the stable plaque group, which strongly suggests its potential as a metabolism-related biomarker.
Iron death is a cell death mechanism characterized by intracellular iron accumulation and lipid peroxidation (42). There are reports that 3-carboxy-4-methyl-5-propyl-2-furanopanoic acid (CMPF) can further induce cell death by increasing reactive oxygen species and lipid peroxidation levels (43). Previous studies have demonstrated that the furan fatty acid metabolite CMPF level increases in patients with gestational diabetes mellitus, T2DM, and impaired glucose tolerance, and damages the pancreatic β-cell function. Mechanistically, CMPF directly acts on β-cells, leading to impaired mitochondrial function, reduces glucose-induced ATP accumulation, and induces oxidative stress, leading to dysregulation of key transcription factors and ultimately decreasing insulin biosynthesis (44). Liu et al. found that CMPF can promote the development of diabetes by inducing metabolic remodeling to encourage preferential fatty acid use compared to glucose (45). Diabetes is also an important risk factor for atherosclerosis and stroke (46). The present research indicated that CMPF was significantly upregulated in unstable carotid artery plaques, which fully confirms its reliability as a metabolic biomarker.
The present study had some limitations. Only patients who underwent carotid endarterectomy at a single advanced stroke center were enrolled in the study. The sample size was also relatively small, which may have affected statistical analysis and generalization of results. In addition, this was a retrospective study with a potentially lower evidence accuracy compared to that of prospective studies. In addition, the present case-control study design may have been affected by selection bias, making it difficult to determine the chronological order of exposure and disease. Therefore, the data cannot directly determine causal relationships. Nevertheless, the present study was the first to use metabolomics combined with machine learning to predict metabolic biomarkers of unstable carotid artery plaques. It was relatively accurate in identifying metabolic biomarkers for studying the characteristics of unstable plaques and serving as a reference for their early diagnosis.
5 Conclusion
In summary, metabolomics analysis of stable and unstable carotid artery plaques was used to identify 14 metabolic biomarkers for early diagnosis their involvement in stroke risk assessment. In addition, it was discovered that metabolites closely related to unstable plaques, especially Neu5Gc, LysoPS, CMPF, and D-NAT, it could help to identify plaque instability and, in turn, contribute to stroke risk stratification. Despite some limitations, the present hypothesis-generating study suggests the existence of metabolic biomarkers and pathways specific of unstable arterial plaques with potential clinical applications upon validation in bigger cohorts.
Statements
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Ethics statement
The studies involving humans were approved by Medical Ethics Committee of Tangdu Hospital. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.
Author contributions
J-QM: Formal analysis, Writing – original draft, Writing – review & editing. LW: Formal analysis, Methodology, Writing – review & editing. X-PQ: Conceptualization, Writing – review & editing. YZ: Writing – review & editing. L-JS: Writing – review & editing. G-DG: Writing – review & editing. CW: Writing – review & editing. L-LZ: Writing – review & editing. Q-XF: Writing – review & editing. YQ: Writing – review & editing. L-LS: Writing – review & editing. BL: Funding acquisition, Writing – review & editing.
Funding
The author(s) declared that financial support was received for this work and/or its publication. This work was financially sponsored by grants from the National Natural Science Foundation of China (81971206, 82073053, 81630027, 82130038), Shaanxi Youth Science and Technology New Star Project (2021KJXX-18), Shaanxi Innovative Research Team for Key Science and Technology (S2022-ZC-TD-0,065), Xi'an Science and Technology Project (22YXYJ0149), China General Technology Group Project (2023610101000028), The State Key Laboratory of Cancer Biology Project (CBSKL2019ZZ12, CBSKL2022KF11).
Acknowledgments
The authors acknowledge the participants and their families who donated their time and effort in helping to make this study possible. Metabolomics analysis was performed by Shanghai Meiji Biomedical Technology Co., Ltd.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcvm.2026.1646067/full#supplementary-material
References
1.
AbdulsalamMFengJ. The composition of vulnerable plaque and its effect on arterial waveforms. J Mech Behav Biomed Mater. (2021) 119:104491. 10.1016/j.jmbbm.2021.104491
2.
TanLXuQShiRZhangG. Bioinformatics analysis reveals the landscape of immune cell infiltration and immune-related pathways participating in the progression of carotid atherosclerotic plaques. Artif Cells Nanomed Biotechnol. (2021) 49(1):96–107. 10.1080/21691401.2021.1873798
3.
Carballo-PerichLPuigoriol-IllamolaDBashirSTerceñoMSilvaYGubern-MéridaCet alClinical parameters and epigenetic biomarkers of plaque vulnerability in patients with carotid stenosis. Int J Mol Sci. (2022) 23(9):5149. 10.3390/ijms23095149
4.
LiRXApostolakisIZKemperPMcGarryMDJIpAConnollyESet alPulse wave imaging in carotid artery stenosis human patients in vivo. Ultrasound Med Biol. (2019) 45(2):353–66. 10.1016/j.ultrasmedbio.2018.07.013
5.
FinnAVNakanoMNarulaJKolodgieFDVirmaniR. Concept of vulnerable/unstable plaque. Arterioscler Thromb Vasc Biol. (2010) 30(7):1282–92. 10.1161/ATVBAHA.108.179739
6.
PeetersWHellingsWEde KleijnDPVde VriesJPPMMollFLVinkAet alCarotid atherosclerotic plaques stabilize after stroke: insights into the natural process of atherosclerotic plaque stabilization. Arterioscler Thromb Vasc Biol. (2009) 29(1):128–33. 10.1161/ATVBAHA.108.173658
7.
XiaZGuMJiaXWangXWuCGuoJet alIntegrated DNA methylation and gene expression analysis identifies SLAMF7 as a key regulator of atherosclerosis. Aging. (2018) 10(6):1324–37. 10.18632/aging.101470
8.
SabaLSaamTJägerHRYuanCHatsukamiTSSalonerDet alImaging biomarkers of vulnerable carotid plaques for stroke risk prediction and their potential clinical implications. Lancet Neurol. (2019) 18(6):559–72. 10.1016/S1474-4422(19)30035-3
9.
ChenS-JLiuR-RShangY-RXieY-JGuoX-HHuangM-Jet alAn ultrasound model to predict the short-term effects of endovascular stent placement in the treatment of carotid artery stenosis. Front Cardiovasc Med. (2020) 7:607367. 10.3389/fcvm.2020.607367
10.
SeeleyEHLiuZYuanSStroopeCCockerhamERashdanNAet alSpatially resolved metabolites in stable and unstable human atherosclerotic plaques identified by mass spectrometry imaging. Arterioscler Thromb Vasc Biol. (2023) 43(9):1626–35. 10.1161/ATVBAHA.122.318684
11.
FiehnO. Metabolomics–the link between genotypes and phenotypes. Plant Mol Biol. (2002) 48(1-2):155–71. 10.1023/A:1013713905833
12.
HarshfieldELMarkusHS. Association of baseline metabolomic profiles with incident stroke and dementia and with imaging markers of cerebral small vessel disease. Neurology. (2023) 101(5):e489–501. 10.1212/WNL.0000000000207458
13.
PettyGWBrownRDJrWhisnantJSicksJPO'FallonJDWiebersWMet alIschemic stroke subtypes: a population-based study of incidence and risk factors. Stroke. (1999) 30(12):2513–6. 10.1161/01.STR.30.12.2513
14.
SkagenKSkjellandMZamaniMRussellD. Unstable carotid artery plaque: new insights and controversies in diagnostics and treatment. Croat Med J. (2016) 57(4):311–20. 10.3325/cmj.2016.57.311
15.
HowardDPJvan LammerenGWRothwellPMRedgraveJNMollFLde VriesJ-PPMet alSymptomatic carotid atherosclerotic disease: correlations between plaque composition and ipsilateral stroke risk. Stroke. (2015) 46(1):182–9. 10.1161/STROKEAHA.114.007221
16.
YangQGuoHShiXXuXZhaMCaiHet alIdentification of symptomatic carotid artery plaque: a three-item scale combined angiography with optical coherence tomography. Front Neurosci. (2021) 15:792437. 10.3389/fnins.2021.792437
17.
de WaardDde BorstGJBulbuliaRPanHHallidayA. Pre-operative carotid plaque echolucency assessment has no predictive value for long-term risk of stroke or cardiovascular death in patients undergoing carotid endarterectomy. Eur J Vasc Endovasc Surg. (2017) 54(2):135–41. 10.1016/j.ejvs.2017.05.013
18.
AndrewsJPMFayadZADweckMR. New methods to image unstable atherosclerotic plaques. Atherosclerosis. (2018) 272:118–28. 10.1016/j.atherosclerosis.2018.03.021
19.
NicholsonJKLindonJC. Metabonomics. Nature. (2008) 455(7216):1054–6. 10.1038/4551054a
20.
ShenXWangCLiangNLiuZLiXZhuZet alSerum metabolomics identifies dysregulated pathways and potential metabolic biomarkers for hyperuricemia and gout. Arthritis Rheumatol (Hoboken, NJ). (2021) 73(9):1738–48. 10.1002/art.41733
21.
BuergelTSteinfeldtJRuyogaGPietznerMBizzarriDVojinovicDet alMetabolomic profiles predict individual multidisease outcomes. Nat Med. (2022) 28(11):2309–20. 10.1038/s41591-022-01980-3
22.
QureshiMIVorkasPACouplandAPJenkinsIHHolmesEDaviesAH. Lessons from metabonomics on the neurobiology of stroke. Neuroscientist. (2017) 23(4):374–82.
23.
ShinTHLeeDYBasithSManavalanBPaikMJRybinnikIet alMetabolome changes in cerebral ischemia. Cells. (2020) 9(7):1630. 10.3390/cells9071630
24.
WangZZhaoY. Gut microbiota derived metabolites in cardiovascular health and disease. Protein Cell. (2018) 9(5):416–31. 10.1007/s13238-018-0549-0
25.
KasaharaKReyFE. The emerging role of gut microbial metabolism on cardiovascular disease. Curr Opin Microbiol. (2019) 50:64–70. 10.1016/j.mib.2019.09.007
26.
ChenWZhangSWuJYeTWangSWangPet alButyrate-producing bacteria and the gut-heart axis in atherosclerosis. Clin Chim Acta. (2020) 507:236–41. 10.1016/j.cca.2020.04.037
27.
KoonerASYuHChenX. Synthesis of N-glycolylneuraminic acid (Neu5Gc) and its glycosides. Front Immunol. (2019) 10:2004. 10.3389/fimmu.2019.02004
28.
PeriSKulkarniAFeyertagFBerninsonePMAlvarez-PonceD. Phylogenetic distribution of CMP-Neu5Ac hydroxylase (CMAH), the enzyme synthetizing the proinflammatory human Xenoantigen Neu5Gc. Genome Biol Evol. (2018) 10(1):207–19. 10.1093/gbe/evx251
29.
WangJShewellLKDayCJJenningsMP. N-glycolylneuraminic acid as a carbohydrate cancer biomarker. Transl Oncol. (2023) 31:101643. 10.1016/j.tranon.2023.101643
30.
KawanishiKCokerJKGrunddalKVDharCHsiaoJZenglerKet alDietary Neu5Ac intervention protects against atherosclerosis associated with human-like Neu5Gc loss-brief report. Arterioscler Thromb Vasc Biol. (2021) 41(11):2730–9. 10.1161/ATVBAHA.120.315280
31.
YaginumaSOmiJUwamizuAAokiJ. Emerging roles of lysophosphatidylserine as an immune modulator. Immunol Rev. (2023) 317(1):20–9. 10.1111/imr.13204
32.
TakagiYNishikadoSOmiJAokiJ. The many roles of lysophospholipid mediators and Japanese contributions to this field. Biol Pharm Bull. (2022) 45(8):1008–21. 10.1248/bpb.b22-00304
33.
TanXShiJFuYGaoCYangXLiJet alRole of erythrocytes and platelets in the hypercoagulable status in polycythemia vera through phosphatidylserine exposure and microparticle generation. Thromb Haemostasis. (2013) 109(6):1025–32. 10.1160/TH12-11-0811
34.
NagataSSuzukiJSegawaKFujiiT. Exposure of phosphatidylserine on the cell surface. Cell Death Differ. (2016) 23(6):952–61. 10.1038/cdd.2016.7
35.
KuranoMKanoKDohiTMatsumotoHIgarashiKNishikawaMet alDifferent origins of lysophospholipid mediators between coronary and peripheral arteries in acute coronary syndrome. J Lipid Res. (2017) 58(2):433–42. 10.1194/jlr.P071803
36.
NishikawaMKuranoMIkedaHAokiJYatomiY. Lysophosphatidylserine has bilateral effects on macrophages in the pathogenesis of atherosclerosis. J Atheroscler Thromb. (2015) 22(5):518–26. 10.5551/jat.25650
37.
BelliniFBruniA. Role of a serum phospholipase A1 in the phosphatidylserine-induced T cell inhibition. FEBS Lett. (1993) 316(1):1–4. 10.1016/0014-5793(93)81724-E
38.
BarnesMJLiCMXuYAnJHuangYCysterJG. The lysophosphatidylserine receptor GPR174 constrains regulatory T cell development and function. J Exp Med. (2015) 212(7):1011–20. 10.1084/jem.20141827
39.
SirianniACJiangJZengJMaoLLZhouSSugarbakerPet alN-acetyl-l-tryptophan, but not N-acetyl-d-tryptophan, rescues neuronal cell death in models of amyotrophic lateral sclerosis. J Neurochem. (2015) 134(5):956–68. 10.1111/jnc.13190
40.
YanCWuDGanLWangJYangWXuB. Significant metabolic alterations in non-small cell lung cancer patients by epidermal growth factor receptor-targeted therapy and PD-1/PD-L1 immunotherapy. Front Pharmacol. (2022) 13:949745. 10.3389/fphar.2022.949745
41.
ShimKGulharRJialalI. Exploratory metabolomics of nascent metabolic syndrome. J Diabetes Complications. (2019) 33(3):212–6. 10.1016/j.jdiacomp.2018.12.002
42.
JiangXStockwellBRConradM. Ferroptosis: mechanisms, biology and role in disease. Nat Rev Mol Cell Biol. (2021) 22(4):266–82. 10.1038/s41580-020-00324-8
43.
ParkJSKimD-HChoiH-IKimCSBaeEHMaSKet al3-carboxy-4-methyl-5-propyl-2-furanpropanoic acid (CMPF) induces cell death through ferroptosis and acts as a trigger of apoptosis in kidney cells. Cell Death Dis. (2023) 14(2):78. 10.1038/s41419-023-05601-w
44.
PrenticeKJLuuLAllisterEMLiuYJunLSSloopKWet alThe furan fatty acid metabolite CMPF is elevated in diabetes and induces β cell dysfunction. Cell Metab. (2014) 19(4):653–66. 10.1016/j.cmet.2014.03.008
45.
LiuYPrenticeKEversleyJHuCBatchuluunBLeaveyKet alRapid elevation in CMPF may act as a tipping point in diabetes development. Cell Rep. (2016) 14(12):2889–900. 10.1016/j.celrep.2016.02.079
46.
PaolissoPFoàABergamaschiLAngeliFFabrizioMDonatiFet alImpact of admission hyperglycemia on short and long-term prognosis in acute myocardial infarction: mINOCA versus MIOCA. Cardiovasc Diabetol. (2021) 20(1):192. 10.1186/s12933-021-01384-6
Summary
Keywords
biomarkers, carotid artery, machine learning, metabolomics, unstable plaque
Citation
Ma J-Q, Wang L, Qu X-P, Zhang Y, Song L-J, Gao G-D, Wang C, Zheng L-L, Fang Q-X, Qu Y, Shen L-L and Liu B (2026) Integrating untargeted metabolomics and deep learning approaches to identify specific metabolic signatures and new mechanisms in unstable plaques. Front. Cardiovasc. Med. 13:1646067. doi: 10.3389/fcvm.2026.1646067
Received
12 June 2025
Revised
18 March 2026
Accepted
22 April 2026
Published
12 May 2026
Volume
13 - 2026
Edited by
Ashton Faulkner, Royal Veterinary College (RVC), United Kingdom
Reviewed by
Erik J. Behringer, Loma Linda University, United States
Kaijian Sun, Guangzhou First People’s Hospital, China
Updates
Copyright
© 2026 Liu, Ma, Wang, Qu, Zhang, Song, Gao, Wang, Zheng, Fang, Qu, Shen and Liu.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Bei Liu liubei206@163.com
†These authors have contributed equally to this work and share first authorship
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.