Novel network biomarkers profile based coronary artery disease risk stratification in Asian Indians.

BACKGROUND: Multi-marker approaches for risk prediction in coronary artery disease (CAD) have been inconsistent due to biased selection of specific know biomarkers. We have assessed the global proteome of CAD-affected and unaffected subjects, and developed a pathway network model for elucidating the mechanism and risk prediction for CAD.
MATERIALS AND METHODS: A total of 252 samples (112 CAD-affected without family history and 140 true controls) were analyzed by Surface-Enhanced Laser Desorption/Ionization Time of Flight Mass Spectrometry (SELDI-TOF-MS) by using CM10 cationic chips and bioinformatics tools.
RESULTS: Out of 36 significant peaks in SELDI-TOF MS, nine peaks could do better discrimination of CAD subjects and controls (area under the curve (AUC) of 0.963) based on the Support Vector Machine (SVM) feature selection method. Of the nine peaks used in the model for discrimination of CAD-affected and unaffected, the m/z corresponding to 22,859 was identified as stress-related protein HSP27 and was shown to be highly associated with CAD (odds ratio of 3.47). The 36 biomarker peaks were identified and a network profile was constructed showing the functional association between different pathways in CAD.
CONCLUSION: Based on our data, proteome profiling with SELDI-TOF MS and SVM feature selection methods can be used for novel network biomarker discovery and risk stratification in CAD. The functional associations of the identified novel biomarkers suggest that they play an important role in the development of disease.

INTRODUCTION

Coronary artery disease (CAD) is the principal cause of death in most countries and despite of major advances in treatment, a large number of victims die apparently healthy and suddenly without prior symptoms. The major challenge in cardiovascular medicine is to find a way of predicting the risk that an individual will suffer from the disease.[1] Most risk prediction algorithms screen using the Framingham risk score (FRS), which considers conventional risk factors such as total cholesterol, high-density lipoprotein (HDL), smoking, hypertension, age, and gender in the algorithm. However, Kanjilal et al.[2] have shown that in Asian Indians, the Framingham model defined only 5% of their study cohort to be at high risk, which was an underestimation of CAD risk in the genetically predisposed population. Addition of new biomarkers may add a better value proposition to the risk prediction.[3] Furthermore, it was also shown that use of inflammatory markers such as C-reactive protein.[4] And coagulation factor-VII expression and genetic markers can add value for risk prediction in Asian Indians.[5] All data so far suggest that there is a need to identify better biomarkers to develop a comprehensive model for CAD risk prediction, especially for Indians.

Global proteome analysis can provide an overall understanding of the disease changes and contribute to the field of clinical cardiovascular science.[67] Biomarker discovery using Surface-Enhanced Laser Desorption/Ionization Mass Spectrometry (SELDI-TOF MS) is a novel approach and widely used in biomarker detection and identification.[8] This method is highly advantageous due to the sensitivity of the assay and the low sample volume requirement. Recent advances in the use of SELDI-TOF MS in CAD have been highlited by Wang,[6] in the chinese population.

In the present study we used the SELDI-TOF MS technology for identifying differentially expressed protein patterns in subjects with and without CAD. Furthermore, we have used three different techniques namely Support Vector Machine (SVM), Discriminant Analysis (DA), and Multilayer perceptron Artificial Neural Networks (ANN) for risk prediction. We identified that the SVM models can give better classification and therefore can be used along with protein profiles in risk prediction. Of the 9 m/z peaks, which could significantly discriminate affected and unaffected subjects, one of the peaks was HSP27 and was validated as a potential risk prediction biomarker in this study.

There are approximately 30,000 articles on cardiac biomarkers on PubMed. However, only a small number of these studies have yielded useful biomarkers for clinical purposes. Genes or proteins usually work collaboratively and involve several pathways. Protein-protein interactions and sub-networks play a major role in modulation of specific pathways and by using this information the predictive value of algorithms could be improved to higher levels. Based on the network profile developed from the biomarkers, we identified interaction of several pathways like stress (HSP27, DAOA), metabolic stress (ROMO1, QRFP), inflammation (INFA2, PLDN, CDKN2B, APP, FAU, and ENSG00000235915), coagulation (PLG, FGA, C3), obesity (APOC2, INSL4), hypertension (VIP), calcium binding (CALML4), and cell adhesion (VTN, MPZL3) as interacting members in the disease. The modulation of one or more of these pathways can lead to a chain reaction of changes in the pathways leading to the onset of CAD. Therefore, use of these novel biomarkers may give better risk prediction for CAD in Indians.

MATERIALS AND METHODS

Study participants and samples

The study comprised of 252 population based subjects out of which 112 probands without family history of CAD and 140 true controls were included. The baseline characteristics of study participants are shown in Table 1. The affected subjects were selected based on the following criteria: (1) Patient is a male ≤60 and female ≤65 on the onset of CAD, diagnosis of CAD via ECG/echo/biochemical or angiogram, patients posted for Percutaneous Transluminal Coronary Angioplasty (PTCA) and Coronary Artery Bypass Surgery (CABG) as diagnosed and given in the physicians report and also as answered in the questionnaire. The control subjects were enrolled above the age of 18 and should not have cardiovascular disease and other major illness like caner, liver failure according to the World Health Organization (WHO) guidelines. All the patient samples were collected after required ethics review board assessment and individual consent.

Table 1

Baseline characteristics of study participants

Biochemical assays

Blood was collected from the participants after a 12-h fasting period. Serum cholesterol and triglycerides were estimated by standard enzymatic analyze following manufacturer's guidelines (Randox Laboratories, London, UK). HDL cholesterol was estimated after precipitation of non-HDL fractions with a mixture of 2.4 mmol/l phosphotungstic acid and 39 mmol/l magnesium chloride, and LDL cholesterol was estimated using the Friedewald formula[9]. A normal human serum pool (NHP) prepared in-house was run with each batch. The inter-assay coefficients of variation (CVs) for commercial controls and NHP ranged from 4.9% to 7% for total cholesterol, 6.1% to 7.7% for triglycerides, and 7.1% to 12.2% for HDL cholesterol.

Reagents and instruments

Sinapinic acid (SPA) and CM10 chip were purchased from Bio-Rad, Hercules, CA (USA) and all other reagents from Sigma Aldrich, St. Louis, MO (USA). The serum samples (in duplicates) were analyzed using CM10 chip followed by the Ciphergen Express Client software. Serum samples were thawed on ice and centrifuged at 14,000 r.p.m. for 5 min at 4°C. A 5-μl volume of supernatant of each sample and 10 μl of U9 buffer (9 M urea, 2% CAHPS, 1% dithiothreitol (DTT)) were added into a tube, which was mixed for 30 min on a platform shaker at 4°C. Next, 185 μl of sodium acetate (100 mM, pH 4) was added to the U9/serum mixture and mixed at 4°C for 2 min on the shaker. A 200-μl volume of sodium acetate was added and mixed for 5 min to activate the CM10 chips. Diluted samples (100 μl) were spotted onto a bioprocessor (Ciphergen Biosystems, Fremont, CA, USA) containing the ProteinChip arrays and then mixed on a platform shaker for 60 min at 4°C. The excess serum was discarded and the chips were washed three times with 200 μl of sodium acetate and twice with 5 μl of dH2O. The chips were then removed from the bioprocessor and air-dried. Before SELDI analysis, 1 μl of a saturated solution of SPA (Bio-Rad) was applied on to each chip twice and air-dried.

Protein chip array analysis

A set of different protocols was used on each spot with varied laser intensities. Pre-processing was done using ProteinChip© software 3.1. The peaks with less background noise were considered for further analysis after baseline subtraction. After normalization peaks with standard deviation of ±2 were deleted. Finally clusters were made by Expression Differential Matrix (EDM) within a range of 1500-30,000 Da. We considered m/z less than 1500 as matrix noise. A first pass of 20% anda 0.3% mass window and a second pass of 2 were given. Mass accuracy was calibrated to <0.1% by all-in-one peptide molecular mass standard (Ciphergen Biosystems, Fremont, CA, USA).

Bioinformatics and biostatistics analysis

The baseline for the study participants were carried out using SPSS version 17 software. The continuous data were analyzed and calculated by Student's t-test and cross-tabs for the categorical data. We clustered the spectra and considered those spectra, which had a significant P value (P < 0.05) for further analysis. To better discriminate the CAD and control subjects based on the peak intensities for diagnostic profiling, we considered three methods, Support Vector Machine (SVM), Multilayer perceptron Artificial Neural Networks (ANN), and Discriminant Analysis (DA).[1011] The type of SVM model we used was C-SVM and the kernel function used was RBF (radial basis function). Optimal values for parameters were found by SVM grid and pattern search with search criterion to minimize the total error. Each combination of peak was analyzed by 10-fold cross-validation. For ANN, architecture was made with an input layer with 36 neurons, a single hidden layer with nine neurons, and output layer with two neurons and four-fold cross-validation.

The 36 m/z peaks were determined as 31 potential biomarkers using proteomics tools from SWISSPROT (www.expasy.org) based on the mass and pI (standard deviation of ± 1% of the overall mass of the protein).[12131415]. We selected one peak (m/z 22859) corresponding to HSP27 for performing enzyme-linked immunosorbent assay (ELISA) assays in new set of affected (n = 125) and unaffected (n = 431) subjects. HSP27 ELISA (R and D Systems, Minneapolis, MN, USA; cat. no. DYC1580-2) was performed in serum samples of the subjects. The biomarkers identified above were given as input into STRING database (http://string-db.org/)[16] to generate the network of biomarkers for assessing functional association.

RESULTS

Out of 252 subjects, 91.5% of CAD-affected subjects were diabetic and 77.1% were suffering from hypertension [Table 1]. Furthermore the conventional risk factors hypertension, diabetes, smoking, total cholesterol, HDL, and age were found to be significant between cases and controls.

Differential protein pattern in controls and CAD-affected

The spectra of 112 subjects with CAD and 140 controls were analyzed. Fifty-six CAD samples and 70 control samples were used as test set and same number of samples in the as validation set for blind test. A total of 67 m/z clusters were obtained of which 36 were significantly (P < 0.05) differentially expressed. The specific proteins for each m/z were listed in Supplementary Table 1 after the SWISSPROT database search. We obtained nine peaks that could discriminate the cases and controls in the SVM model. The descriptive statistics of these nine peaks are shown in Table 2. Biomarkers with m/z 22,859 [Figure 1], 9284, 14,660, 9481, and 14,720 were highly expressed in CAD-affected subjects, and m/z 5896, 8922, 8600, and 19,251 were highly expressed in controls [Table 2].

Supplementary Table 1

m/z peaks identified after SELDI-TOF analysis. Significant peaks used for discrimination of CAD and controls are marker in bold.

Table 2

Mean intensity±SE levels in CAD and controls for biomarkers in the test data

Figure 1

Representative spectrum report of average m/z 22,859 in CAD and control samples

Comparison of three different approaches of model building

The 36 peaks were further analyzed by different techniques to obtain the best set of peaks and algorithm for risk prediction. We compared performance of three algorithms Discriminative Analysis (DA), Multilayer perceptron Artificial Neural Networks (ANN), and Support Vector Machine (SVM) based on accuracy, sensitivity, specificity, and area under the receiver operating curve (ROC) [Table 3a]. SVM was found to be the best model for classification using our data with an area under the curve (AUC) of 0.807 and better specificity, sensitivity, and accuracy. Furthermore, the test set also gave good classification data with SVM [Table 3b] with AUC of 0.785 and other features. Also when we consider the overall misclassification the least values were observed for SVM with 23.02% and 26.19% for training and test data, respectively.

Table 3a

Classification of CAD and controls using three different methods, SVM, ANN, and DA, suggesting that SVM model is the best classifier for training data

Table 3b

Classification of CAD and controls using three different methods, SVM, ANN, and DA, suggesting that the SVM model is the best classifier for training data

Use of SELDI biomarkers and modulation of seven different pathways for risk stratification

As we know that FRS is used widely for risk prediction; however we also know that the use of FRS is limited for Asian Indians. Therefore, we considered the nine peaks identified by SELDI-TOF-MS as potential biomarkers along with the FRS model for risk stratification. FRS alone gave an AUC of 0.888, which improved to 0.963 on addition of the nine potential biomarkers [Figure 2]. These nine biomarkers represent seven different pathways, stress and stress/immunity (m/z s 22,859: HSP27, 5896: Leukocyte-specific transcript-1), coagulation (m/z s 8922: Plasminogen precursor activating peptide, 9284: Vitronectin-10, 8600: Pallidin gene isoform-2), infection and inflammation (m/z 19,251: Interferon α-2), mitochondrial damage (m/z 14,660: Farataxin chain-3), calcium binding (m/z 9481: Calmodulin-like protein-4 isoform-3), and cell cycle (m/z 14,720: Cyclin-dependent kinase-4 inhibitor-B). Modulation of these biomarkers results in the change in the functional implications of the pathways, which may result in the disease.

Figure 2

Receiver operating curves for FRS and for addition of biomarker expressions in discriminating CAD vs. controls. The improvement of AUC curve suggests that addition of SELDI-TOF-based feature selection biomarkers may add value in CAD risk stratification

Networking biomarkers and pathways

The proteins identified [Supplementary Table 1] were further taken to generate the functional association network among themselves and with the other proteins. These proteins are from multiple pathways like inflammation, cell signaling, cell adhesion, immunity, obesity, lipid metabolism, coagulation, stress, membrane transport, protein degradation, coagulation, and cell cycle. Our data suggest that CAD is a multi-factorial process and deregulation of these factors may lead to the disease. As seen in Figure 3a, the network of the proteins identified suggests that 16 of 31 proteins have minimal or no linkage among themselves (MPZL3, INSL4, SCOC, ROMO1, CALML4, ENSG0000023591, PSAP, SRP9L1, ANKDD1A, CMTM1, CDN2B, HMSDV, DSC10, KRTDAP, VGLL4, FXN, and APOC2). These proteins need to be proven further to understand their biological role suggesting that they might be novel biomarkers for CAD in this study. Fifteen proteins were networked with at least one more protein (HSPB1, FGA, PLG, VTN, APP, QRFP, C3, POMC, CHGA, VIP, CRH, FAU, C11orf10, IFNA2, and CDKN2B).

Figure 3a

Functional association of proteins identified by SELDI-TOF MS

Further, when we extended this network by adding other interacting or functionally associated proteins [Figure 3b], we saw that the individual proteins, which were not in the network in Figure 3a, had changed from 17 to 9 proteins. These nine proteins (ANKDD1A, ROMO1, SCOC, SRP9L, MPZL3, CALML4, CMTM1, KRTDAP, and ENSG0000023591) are potentially novel members and further analysis may be needed to identify their networks and associations. However, most of the other biomarkers were directly or indirectly were associated.

Figure 3b

Extended network and association of potential biomarkers

CAD-associated networks

Our data [Figure 3a and b] suggest that multiple pathways are associated and networked together in the CAD subjects for the onset of the disease. The most networked proteins [Supplementary Table 1] were identified based on number of edges for each protein in the network [Figure 3b]. The biomarkers FAU, CRH, APP, VIP, CHGA, POMC, HSPB1, C3, FGA, VTN, INFA2, FXN, CDKN2B, and PLDN are from different pathways suggesting that these pathways interact in the disease condition. These association studies suggest that coagulation, cell signaling, kinase inhibitors, stress, protease inhibitors, and obesity are major pathways leading to CAD in Asian Indians in our studies. It is understood that stress leads to several changes, which might play a major role in early pathogenesis of CAD. Therefore, markers like ROMO1 and HSP27 might be the first candidate markers, which need to be evaluated, and we evaluated HSP27 as potential marker for the same.

The stress-related protein HSP27 is highly associated with CAD in Asian Indians

HSP27 is a member of the small heat-shock protein (HSP) (sHSP) family and is involved in diverse range of functions in addition to its chaperoning function. We identified m/z 22,859 as HSP27 (molecular weight 22,783 Da). HSP27 is also a member associated with five different partners, which in turn are regulating multiple pathways such as PLG (plasminogen precursor-activating peptide), CDKN1A, CTNNB1, CCND1 (cell-cycle proteins), and kinases such as MAPKAPK2 and 5. These associations suggest that many cellular responses may be triggered along with HSP27 in CAD and therefore this biomarker was further validated. We performed ELISA assay for HSP27 in 431 subjects (125 CAD-affected and 306 unaffected). We found that the CAD-affected subjects had higher expression levels that unaffected [Table 4a]. The odds ratio of HSP27 alone [Table 4b] was not significant; however, after addition of conventional risk factors (age, gender, body mass index, waist circumference, and hypertension), the odds ratio of fourth quartile in comparison to first quartile improved to 2.81 (95% confidence interval (CI): 1.18-6.79, P = 0.019). Furthermore, upon adjustment with lipids (triglycerides, total cholesterol, HDL, and LDL), the odds ratio of the fourth quartile improved to 3.47 (95% CI: 1.41-8.56, P = 0.007).

Table 4a

Mean expression levels of HSP27 in CAD-affected and unaffected subjects

Table 4b

Association of HSP27 based on odds ratio

DISCUSSION

As CAD is a major killer in India, it is very important to identify the ways of improving risk prediction. At present diagnosis or risk prediction is dependent on clinical history, physical examination, and other tests, which do not look at the biochemical or molecular changes, which might give early risk prediction to CAD. In our present study, we have explored and validated the process of identification of biomarkers using SELDI-TOF-MS and further using the patterns of m/z to diagnose the risk of CAD in Asian Indians. It has been well established that SELDI-TOF-based CAD diagnosis can be used;[678910] our attempt to use novel biomarkers for risk prediction can add value for early diagnosis and prevention of CAD endemic.

SELDI-TOF-based biomarker detection and use of protein patterns derived from serum to differentiate between CAD and no CAD is of major interest[1117] as it allows complete proteome profiling in a high-throughput format. Using feature selection techniques such as SVM helps in a more robust method of classifying the subjects.[1819] Our analysis showed that the power of each biomarker to discriminate between the cases and controls was best for the SVM model by estimating the ROC. The greater the AUC value for the biomarkers shows the relative importance value of the ability to accurately distinguish between the different groups.[20] Apart from using the SVM to identify the best biomarkers and their discrimination between CAD and no CAD, we have also established that the biomarkers add a very important value for FRS risk prediction method.

The 31 proteins identified and their networking suggests that multiple pathways are associated with CAD and in specific inflammation, cell signaling, coagulation, cell adhesion, stress, and obesity are the major pathways. It is also very interesting to note novel proteins, which are not associated with any of the known pathways to be identified [Figure 3a and b]. These proteins may have no earlier data in relation to coronary artery disease and more studies may be needed to understand their role. Of the 15 proteins identified as highly networked FAU, CRH, POMC, VIP, VTN, and PLG are prominent with more than 10 associations suggesting that further analysis with these proteins may yield better understanding of the pathways involved in the CAD.

Our data suggest that the stress- and immunity-related protein HSP27 might play a major role in CAD for Asian Indians [Table 4a and b]. The functional role of several HSPs in atherosclerosis has been well studied as they represent the response of the cells to blood vessel to different stress signals.[21] It is also well known that HSPs are potential targets for immune response and contribute to inflammatory process.[22] The smooth muscle cells (SMCs) play a important role in atherogenesis as they can over express HSPs as a part of survival mechanism following exposure to variety of stressors (example: High blood pressure). Most research on HSPs was focused on HSP60/65 and 70; however recently evidence of role of HSP27 in CAD is becoming evident.[23242526] In our study we have analyzed the serum levels of 431 subjects and found that HSP27 alone is not associated, but when the model was adjusted for conventional risk factors and lipids, higher association to CAD was seen [Table 4b]. Our data suggest that HSP27 might play important role in risk prediction and further studies are needed to evaluate the value addition by this biomarker. It was also suggested that phosphorylated HSP27 may have a protective effect in atherogenesis;[27] therefore further studies are needed to evaluate the functional role of HSP27 versus the use of expression levels in risk prediction.

Our findings revealed that SELDI-TOF-MS technique can be used for risk stratification of CAD-affected and unaffected subjects using the SVM method. The networking of proteins and the pathways indicate that several pathways such as stress, inflammation, coagulation, cell adhesion, signaling, and obesity are interlinked and might crosstalk in the development of the disease. Our approach has resulted in understanding the network and modulation of pathways that contain specific sub-networks and novel biomarkers that may help in improving the risk prediction. Further, we used SELDI-TOF-MS not only in identification of new biomarkers, but also as a means of understanding the mechanism of CAD development by network construction.