Introduction
Atherosclerosis is a chronic inflammatory disease which mainly results from the abnormal accumulation of macrophages, white blood cells and lipids in the artery wall, and eventually form mature plaques (1). It causes various types of cardiovascular disease, which account for ~50% of all mortalities worldwide (2). To date, the underlying molecular pathogenesis of atherosclerosis has remained largely elusive. An understanding of the molecular basis of atherosclerotic processes is crucial for the development of potential therapeutic strategies.
Atherosclerosis is characterized by endothelial dysfunction, vascular inflammation, plaque formation and diminished oxygen supply to target organs (3,4). Endothelial dysfunction is an initial step in atherosclerosis. Dysfunctional endothelial cells release a large variety of pro-inflammatory mediators, leading to the amplification of inflammatory response (5,6), and systemic inflammation may in turn contribute to endothelial dysfunction and accelerated atherosclerosis (7,8). Considering the crucial roles of inflammation in the pathogenesis of atherosclerosis, systemic markers of inflammation may be the most obvious candidates for potential biomarkers for early diagnosis, risk prediction and clinical prognosis of this disease (9). C-reactive protein has been implicated in multiple aspects of atherosclerosis and is currently the best validated inflammatory biomarker to predict the risk of atherosclerotic events (10,11). Fibrinogen, apolipoproteins and interleukins are associated with the progression of atherosclerosis and have a great relevance in risk prediction (12). Inflammatory molecules, including intercellular adhesion molecule-1 and vascular cell adhesion molecule-1, are also considered as biomarkers for risk prediction of atherosclerosis (11,13). Several studies have identified a large number of biomarkers with potential prognostic value for atherosclerosis; their overall clinical predictive value, however, is modest (14). Therefore, it is urgently required to identify a reliable, specific and cost-effective biomarker for facilitating the clinical diagnosis of atherosclerosis.
The microarray dataset no. GSE20129 from the Gene Expression Omnibus database (GEO; http://www.ncbi.nlm.nih.gov/geo/) has been utilized to reveal variations in gene expression associated with coronary artery calcification between Caucasian and African Americans based on a multi-ethnic study of atherosclerosis (15); furthermore, the dataset has been used to reveal systemic transcriptional alterations of innate and adaptive immune signaling pathways in atherosclerosis (16). The present study used the microarray dataset GSE20129 to identify differentially expressed genes (DEGs) associated with atherosclerosis using comprehensive bioinformatics methods. In addition, the recursive feature elimination (RFE) algorithm was applied for feature selection via iterative classification, and the support vector machine (SVM) classifier was used for the validation of prediction accuracy. The present study aimed to identify potential biomarkers for atherosclerosis via analysis of gene expression profiles.
Materials and methods
Microarray data
The array dataset GSE20129, which was provided by Huang et al (15), was downloaded from the GEO database. A total of 118 samples from the peripheral blood of females, including 47 atherosclerotic and 71 non-atherosclerotic patients, was used for expression profiling. The raw CEL data and annotation files were obtained based on the Illumina human Ref-8 v2.0 expression beadchip platform (Illumina Inc., San Diego, CA, USA) for subsequent analyses.
Data processing and screening of DEGs
All probe IDs of the dataset were transformed into gene symbols based on probe annotation files. If multiple probes corresponded to the same gene symbol, the mean value was calculated as the expression value of this gene. Z-score normalization of expression values was preformed and then the expression value (X) of each sample was corrected. The corrected expression value (X′) was calculated as follows: X′=(X-mean)/standard deviation.
The DEGs in samples from atherosclerotic patients compared with those in non-atherosclerotic patients were screened using the Limma package (http://www.biocon-ductor.org/packages/release/bioc/html/limma.html) (17) in R. P<0.01 was defined as the cut-off value for DEG screening.
Unsupervised hierarchical clustering analysis
Unsupervised hierarchical clustering analysis (18) for DEGs was performed to effectively distinguish atherosclerotic and non-atherosclerotic samples using cluster 3.0 (19) and to visualize these samples using the heatmap tool in TreeView (20). Filtering and normalization of all the expression data were performed using cluster 3.0. Genes with expression in at least 80% of all samples were retained, while others were discarded. All of the genes and samples were median centered (21) and normalized. The similarity matrixes were correlation centered (22) and normalized.
Functional enrichment analyses
The Gene Ontology (GO) tool (23) provides functional annotations for large-scale genomic or transcription data. The Kyoto Encyclopedia of Genes and Genomes database (KEGG; http://www.genome.ad.jp/kegg/) (24) holds information on pathways and networks of molecules or genes. The Database for Annotation, Visualization and Integrated Discovery (DAVID; http://david.abcc.ncifcrf.gov/) software (25) is an integrated biological database and analytic tool for systematic annotation of biological meaning for large lists of genes or proteins.
To understand the biological significance of DEGs, GO term and KEGG pathway analyses for up-and downregulated DEGs were performed using the DAVID online tool. The significance threshold was set at P<0.05.
Feature selection with the RFE algorithm
Feature selection offers classification models with high relevance by eliminating irrelevant features (26). In order to obtain candidates for diagnostic and prognostic biomarkers of atherosclerosis, the RFE algorithm (27,28) in python was applied for feature selection from the constellation of DEGs via iterative classification. Classification analyses were performed using a linear Support Vector Classification (SVC; linear kernel, step=1; cross-validation=5) provided by the python module scikit-learn (29). This recursive procedure adopted a backward elimination strategy to iteratively remove irrelevant features and was iterated until all relevant features were obtained.
Prediction with SVM classifier
All of the relevant features were processed with a linear SVM classifier (30) for classification prediction, which was repeated five times. Linear kernel and five-fold cross-validation of all samples were performed in this prediction procedure to assess classification performance. To further illustrate the prediction accuracy of this method, a receiver operating characteristic curve (ROC) was implemented and the area under curve (AUC) was calculated.
Results
DEG screening and hierarchical clustering analysis
Based on probe annotation files, a total of 18,196 gene symbols were acquired after transformation. A total of 430 DEGs with P<0.01 were screened using the Limma package, including 149 upregulated and 281 downregulated genes. The results of the unsupervised hierarchical clustering analysis were visualized in a heatmap (Fig. 1).
Functional enrichment analyses
Functional enrichment analyses for DEGs were performed using DAVID, and the significantly enriched GO terms and KEGG pathways are shown in Tables I and II, respectively. According to the enrichment results, the over-represented GO terms of upregulated genes were mainly associated with ribonucleoprotein complex biogenesis, ribosome biogenesis and positive regulation of T-cell proliferation, while downregulated genes were mainly associated with response to lipopolysaccharides, immune response and responses to molecules of bacterial origin (Table I). In addition, upregulated genes were significantly enriched in viral myocarditis and purine metabolism, while downregulated genes were significantly involved in leukocyte transendothelial migration and the Notch signaling pathway (Table II).
Feature selection with the RFE algorithm and prediction with SVM classifier
By applying the RFE algorithm, a total of 11 biomarkers were obtained, including APH1B, JAM3, FBLN2, CSAD and PSTPIP2 (Table III). These 11 biomarkers were then processed with the SVM classifier for classification prediction. The AUC of the ROC obtained with the SVM classifier was calculated as 0.92 (Fig. 2), indicating that the identified 11 biomarkers were representative. In order to verify that the biomarkers that we screened were representative for the DEGs in atherosclerotic patients, all 430 DEGs were processed with the SVM classifier, resulting in an AUC of the ROC of 0.95 (Fig. 3), which confirmed that the 11 biomarkers were representative as the prediction accuracy of the identified 11 biomarkers was markedly higher than the randomly selected 11 DEGs. In addition, to rule out any randomization, the 11 DEGs were randomly selected as features which were subjected to the process of classification for 100,000 times, and the AUC of every step was calculated. The results showed that the rank position of the AUC corresponding to 11 biomarkers screened was 33 (P=0.0033), suggesting that the prediction accuracy of the screened biomarkers was significantly higher than that of random results.
Discussion
Atherosclerosis is the proximate cause of occlusive arterial disease (31). Identification of potential biomarkers that have crucial roles in the progression of atherosclerosis is important for the development of therapeutic approaches. In the present study, a bioinformatics approach was used to predict potential biomarkers for the diagnosis and treatment of atherosclerosis. The screening of gene expression profiles of atherosclerotic patients identified 11 biomarkers, including APH1B, JAM3, FBLN2, CSAD and PSTPIP2, and the AUC of the ROC obtained using SVM classifier was calculated as 0.92, which indicated that these 11 biomarkers were representative. The crucial roles of these genes in the pathogenesis of atherosclerosis and their potential diagnostic and therapeutic values for this disease warrant further evaluation.
APH1B is a functional component of the gamma-secretase enzyme complex, which catalyzes the cleavage of transmembrane proteins such as Notch receptors and amyloid-b precursor protein (APP) (32). Notch activation induces senescence of endothelial cells and results in pro-inflammatory responses, and Notch signaling has pivotal roles in the pathogenesis of atherosclerosis (33). Blockade of Notch ligand delta-like 4 - Notch signaling attenuates the development of atherosclerosis (34). APP has also been found to have a potential pathogenic role in carotid atherogenesis (35). In addition, the single-nucleotide polymorphism Phe217Leu (rs1047552; T>G) in the APH1B gene is associated with premature coronary atherosclerosis (36). It can therefore be speculated that APH1B may be implicated in the progression of atherosclerosis and function as a biomarker for risk prediction.
Junctional adhesion molecule 3 (JAM3) is a member of vascular adhesion molecules, which mediate adhesion and interactions among cells or between cells and the extracellular matrix (37). Increased expression and activation of adhesion molecules in vascular endothelial cells and circulating leukocytes stimulate leukocyte recruitment into the vascular endothelium, which is an important step in the development of atherosclerosis (38,39). JAM-C has been shown to be involved in the control of inflammatory leukocyte recruitment in atherosclerosis (40). JAM3 may act as a counter-receptor for Mac-1, which mediates leukocyte-platelet interactions between vascular cells, and may thereby provide a molecular target for antagonizing interactions involved in several vascular pathologies, such as in atherothrombosis (37). In the present study, JAM3 was identified as a potential biomarker for atherosclerosis, which was in line with the abovementioned previous findings and suggested that JAM3 may be a potential therapeutic target in atherosclerosis.
The FBLN2 gene encodes an the extracellular matrix protein fibulin 2. Upregulation of FBLN2 may promote the arterial response to injury and accelerate atherogenesis among patients with diabetes (41). In addition, FBLN2 may lead to the aberrant regulation of activator protein-1 factors, which have been shown to be associated with the progression of atherosclerosis (42). Thus, FBLN2 may be implicated in the progression of atherosclerosis and serve as a biomarker for this disease.
Cysteine sulfinic acid decarboxylase (CSAD) can catalyze the conversion of cysteine sulfinic acid to hypotaurine, which may limit the generation of taurine (43). Taurine has been shown to have the potential protective effects against cardiovascular diseases and can effectively prevent the progression of atherosclerotic diseases (44). Therefore, it may be hypothesized that CSAD may have roles in the progression of atherosclerosis via limiting the generation of taurine.
Proline-serine-threonine phosphatase-interacting protein 2 (PSTPIP2) is expressed in macrophages and is tyrosine-phosphorylated in response to colony-stimulating factor-1 (CSF-1) (45). CSF-1 is thought to have a role in the accumulation of cholesterol-laden macrophages (foam cells) in atherosclerotic plaques (46). Chitu and Stanley (47) found that CSF-1 was involved in several inflammatory disorders, such as artherosclerosis. Although the causal roles of PSTPIP2 in the development of atherosclerosis have not been fully elucidated, it may be speculated that PSTPIP2 has a pivotal role in the progression of atherosclerosis via interaction with CSF-1. The results of the present study indicated that PSTPIP2 may be a potential biomarker for atherosclerosis, which requires further investigation.
Only gene expression profiles of atherosclerotic/non-atherosclerotic females were investigated in the present study. However, atherosclerosis may have different pathologies in males and females; therefore, sex differences should be explored. Furthermore, there was no experi mental validation in the present study; therefore, further investigations are required in order to verify the present findings and speculation. In particular, additional prospective biomarkers, including HIATL1, HOXB13, RNF148, RALB, CPEB3 and ABHD14B, were identified, which merit further investigation and discussion.
In conclusion, the bioinformatics analysis of the present study indicated that APH1B, JAM3, FBLN2, CSAD and PSTPIP2 may have important roles in the pathogenesis of atherosclerosis. They may be considered as potential biomarkers for early diagnosis and prognosis as well as therapeutic targets for this disease, which requires experimental validation.