Autoantibodies to Oxidatively Modified Peptide: Potential Clinical Application in Coronary Artery Disease

Coronary artery disease (CAD) is a global health issue. Lipid peroxidation produces various by-products that associate with CAD, such as 4-hydroxynonenal (HNE) and malondialdehyde (MDA). The autoantibodies against HNE and MDA-modified peptides may be useful in the diagnosis of CAD. This study included 41 healthy controls (HCs) and 159 CAD patients with stenosis rates of <30%, 30–70%, and >70%. The plasma level of autoantibodies against four different unmodified and HNE-modified peptides were measured in this study, including CFAH1211–1230, HPT78–108, IGKC2–19, and THRB328–345. Furthermore, feature ranking, feature selection, and machine learning models have been utilized to exploit the diagnostic performance. Also, we combined autoantibodies against MDA and HNE-modified peptides to improve the models’ performance. The eXtreme Gradient Boosting (XGBoost) model received a sensitivity of 78.6% and a specificity of 90.4%. Our study demonstrated the combination of autoantibodies against oxidative modification may improve the model performance.


Introduction
Coronary artery disease (CAD) is a global health care issue which affected nearly 1.72% of individuals worldwide [1]. The progression of CAD consists of three stages: fatty streak, plaque progression, and disruption. Starting from inflammatory cells recruitment and lipid accumulation, the inflammation may lead to plaque formation and eventually cause atherosclerosis [2]. Inflammation occurs with the accumulation of free radicals, which may accelerate the process of CAD progression and create a positive feedback loop to form additional free radicals [3]. Inflammation has been considered as the major factor of CAD progression. Hence, the diagnosis of CAD in the clinic focuses on the measurement of inflammation markers, such as C-reactive protein (CRP), cytokines, and adhesion molecules [4].
In addition to inflammation, lipid peroxidation can also generate free radicals; it has been considered as a process of oxidants' attack on the lipids containing carbon-carbon double bonds [5]. Lipid peroxidation produces several by-products such as malondialdehyde (MDA), propenal (acrolein), hexanal, and 4-hydroxynonenal (4-HNE), which are associated with chronic diseases [5][6][7][8][9]. These by-products may modify macromolecules, and cause damage to the molecular functions of the macromolecules [6]. Furthermore, since ≥ 130 mg/dL, TG ≥ 200 mg/dL, and HDL-C < 40 mg/dL, or receiving lipid lowering drug prescription. Patients were considered as having hypertension with one of following criteria: systolic blood pressure ≥140, diastolic blood pressure ≥90, or receiving an antihypertensive drug prescription. Patients were considered as having diabetes with one of following criteria: diagnosis with diabetes or receiving a hypoglycemia drug prescription. Healthy controls (HCs) were excluded if they suffered from hyperglycemia, hyperlipidemia, hypertension, or angina pectoris. Ten mL of blood samples were collected from patients and HCs. After the blood samples were centrifuged 3000 rpm for 10 min, the plasma was stored at −80 • C until analyzed. This study was approved by the institutional review board of the study hospital, and all volunteers provided informed consent before participating. Patient samples were randomly selected. The demographic characteristics of patients are summarized in Table 1 artery disease, patients were separated into the group of coronary stenosis rate of 30-70% and the group of coronary stenosis rate of >70%. CAD patients who also were diagnosed as RA were excluded ( Figure 1). As for comorbidity, patients were considered as having hyperglycemia with one of following criteria: total cholesterol ≥ 200 mg/dL, LDL-C ≥ 130 mg/dL, TG ≥ 200 mg/dL, and HDL-C < 40 mg/dL, or receiving lipid lowering drug prescription. Patients were considered as having hypertension with one of following criteria: systolic blood pressure ≥140, diastolic blood pressure ≥90, or receiving an antihypertensive drug prescription. Patients were considered as having diabetes with one of following criteria: diagnosis with diabetes or receiving a hypoglycemia drug prescription. Healthy controls (HCs) were excluded if they suffered from hyperglycemia, hyperlipidemia, hypertension, or angina pectoris. Ten mL of blood samples were collected from patients and HCs. After the blood samples were centrifuged 3000 rpm for 10 min, the plasma was stored at −80 °C until analyzed. This study was approved by the institutional review board of the study hospital, and all volunteers provided informed consent before participating. Patient samples were randomly selected. The demographic characteristics of patients are summarized in Table 1     and incubated at 37 • C for 2 h. For blocking, 3% of a BSA solution was added and incubated at 37 • C for 1 h after the plates were washed with PBS containing 0.05% Tween 20 (PBST). The plates were washed with PBST, and a rabbit anti-HNE antibody (ab46545, Abcam, Cambridge, MA, USA) was added. After incubating for 2 h at room temperature, the plates were washed with PBST. A mouse anti-rabbit antibody conjugated with horseradish peroxidase (HRP) was added into plates and incubated for 1 h. We washed the plates with PBST and detected the HRP with SureBlue Reserve™ TMB (Kirkegard and Perry Laboratories, Gaithersburg, MD, USA) for 30 min. The color reaction was stopped with 1N HCl, and the absorbance was measured at 450/620 nm. The concentration of HNE-protein adducts in plasma was calculated according to a standard HNE-modified BSA curve.

Detection of Plasma Autoantibodies against Unmodified and HNE Modified Peptide
The BSA and four peptides with 1mg/mL were modified with 4-Hydroxynonenal (Sigma, Neustadt, Germany). The BSA or peptides (10 µg/mL) were added into a 96-well plate and incubated at 37 • C for 2 h. For blocking, after washing with PBST, 3% of a BSA solution was added and incubated at 37 • C for 1 h. The plates were washed with PBST and the 100-fold diluted plasma samples were added. The plates were incubated at room temperature for 2 h. After the plates were washed, rabbit anti-human IgG-HRP (A80-118P, 1:30,000, BETHYL) or rabbit anti-human IgM-HRP (A0420, 1:10,000, Sigma) were diluted and added into plates. The plates were incubated at room temperature for 1 h. After that, the plates were washed completely and added with the SureBlue ReserveTM Peroxidase Substrate (Kirkegard & Perry Laboratories, Gaithersburg, MD, USA). The plates were incubated at room temperature for 15 min. The color reaction was stopped by adding 1N HCl, and the absorbance was measured at 450/620 nm. All ELISA experiments were conducted following the ELISA Guidebook. The quality controls samples were prepared with two replicates in each plate to calculate the percent coefficient of variation (CV%) across wells and plates. An experiment was repeated if the CV% was calculated above 20%.

Statistical Analysis and Machine Learning
The significance of IgG and IgM autoantibodies between HCs and RA, RA-CAD, or CAD patients were determined by Student's t-test. The Student's t-test was calculated by GraphPad Prism (v.8.0; GraphPad software, San Diego, CA, USA). The significance level of all statistical tests was set to p < 0.05. The feature ranking was conducted by WEKA (vers.3.8.5). The models we built in this study were based on eXtreme Gradient Boosting (XGBoost) and the Light Gradient Boosting Machine (GBM) with 5-fold cross validation with scikit-learn (vers.0.21.3). A confusion matrix was applied in this study to calculate the accuracy, precision, sensitivity, specificity, and f1 score. The value of AUC was calculated with scikit-learn (vers.0.21.3). The comparison of models was evaluated by an ANOVA test.

Measurement of Autoantibodies against HNE Modified BSA
Plasma samples were subjected to ELISAs for measuring IgG and IgM autoantibodies against unmodified and HNE-modified BSA. Plasma levels of IgG and IgM against HNEmodified BSA were found to be increased in patients with RA and RA-CAD (Supplementary Figure S1A,B). Plasma levels of IgM against BSA were found significantly different between HC and RA (Supplementary Figure S1B).

Measurement of Autoantibodies against HNE-Modified Peptides and HNE Adducts
Plasma samples were analyzed with ELISA to detect IgG and IgM autoantibodies against unmodified and HNE-modified peptides (Supplementary Figure S2). Plasma levels of IgG against the CFAH 1211-1230 unmodified peptide and HNE-modified peptide in CAD patients with a stenosis rate >70% were notably higher than HCs (p = 0.008, p = 0.0002). Plasma levels of IgM against the CFAH 1211-1230 unmodified peptide and HNE-modified peptide in CAD patients with a stenosis rate >70% were significantly lower than HCs (p < 0.0001, p < 0.0001) and CAD patients with a stenosis rate <30% (p = 0.0092, p = 0.0211). Plasma levels of IgG against the HPT 78-108 unmodified peptide and HNE-modified peptide in CAD patients with a stenosis rate >70% were significantly decreased compared to HCs (p = 0.0002, p = 0.0001). Plasma levels of IgM against the HPT 78-108 unmodified peptide and HNE modified peptide in CAD patients with a stenosis rate >70% were notably lower than HCs (p < 0.0001, p < 0.0001). Plasma levels of IgG against the IGKC 2-19 unmodified peptide in CAD patients with a stenosis rate >70% were significantly lower than HCs (p < 0.0001). In contrast, plasma levels of IgG against the IGKC 2-19 HNE-modified peptide in CAD patients with stenosis rates of <30%, 30-70%, and >70% were decreased compared to HCs (p = 0.02, p = 0.001, p = 0.0002). Plasma levels of IgM against the IGKC 2-19 unmodified peptide in CAD patients with a stenosis rate >70% were lower than HCs (p < 0.0001) and CAD patients with a stenosis rate <30% (p = 0.0056). Plasma levels of IgM against IGKC HNE-modified peptide in CAD patients with stenosis rates of <30%, 30-70%, and >70% were lower than HCs (p = 0.0001, p = 0.0005, p < 0.0001). Plasma levels of IgG against the THRB 328-345 HNE-modified peptide in CAD patients with a stenosis rate >70% were higher than HCs (p < 0.0001) and CAD patients with stenosis rate <30% (p = 0.0039). Plasma levels of IgM against the THRB 328-345 unmodified and HNE-modified peptide in CAD patients with a stenosis rate >70% were significantly higher than HCs (p = 0.0003, p = 0.0002) and CAD patients with stenosis rate <30% (p = 0.00498, p = 0.00486). Furthermore, HNE-modified protein adducts in patients with CAD were higher than HCs (Table 1).

Optimization of Machine Learning Algorithms with Autoantibodies against HNE Modified Peptides
To increase the performance of LightGBM and XGBoost, we firstly performed feature ranking with InfoGain + Ranker to list the features from the most important to less important in HCs against patients with stenosis rates of <30%, 30-70%, and >70% ( Table 2). The autoantibody, IgG anti-IGKC 2-19 HNE, was ranked as the top feature in each analysis. Next, we performed forward selection to further optimize LightGBM and XGBoost. In the LightGBM model, the features that were selected included IgG anti- In differentiating HCs and CAD patients with a stenosis rate of <30%, the model received an accuracy of 77%, a precision of 73%, a f1 score of 72.6%, a sensitivity of 78.2%, a specificity of 76.4%, and an AUC value of 0.832. In discriminating HCs and CAD patients with a stenosis rate of 30-70%, the model received an accuracy of 72.7%, a precision of 72.5%, a f1 score of 69.1%, a sensitivity of 71.5%, a specificity of 75.8%, and an AUC value of 0.816. As for HCs and CAD patients with stenosis rates of >70%, the model received an accuracy of 73%, a precision of 61%, a f1 score of 57.8%, a sensitivity of 62%, a specificity of 79.6%, and an AUC value of 0.819. In the XGBoost model, the features that were selected included IgG anti-IGKC 2-19 HNE, IgG anti-CFAH, IgG anti-HPT 78-108 , IgM anti-HPT 78-108 HNE, IgM anti-IGKC 2-19 , IgG anti-IGKC [2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19] , and IgM anti-IGKC 2-19 HNE (Table 3). In discriminating HCs and CAD patients with stenosis rates of <30%, the model received an accuracy of 77.7%, a precision of 74.5%, a f1 score of 71.4%, a sensitivity of 74.2%, a specificity of 79.2%, and an AUC value of 0.845. As for HCs and CAD patients with stenosis rates of 30-70%, the model received an accuracy of 75.3%, a precision of 71.7%, a f1 score of 71.4%, a sensitivity of 75.6%, a specificity of 75.4%, and an AUC value of 0.825. In differentiating HCs and CAD patients with stenosis rates of >70%, the model received an accuracy of 77.3%, a precision of 66.9%, a f1 score of 62.8%, a sensitivity of 64.5%, a specificity of 83.2%, and an AUC value of 0.856 (Table 3).  LGBM: Light Gradient Boosting Machine (lightGBM), XGB: eXtreme Gradient Boosting (XGBoost).

Optimization of Machine Learning Algorithms with Autoantibodies against HNE and MDA-Modified Peptides
In our previous study, we had reported that the autoantibodies against four different MDA modified peptides may serve as biomarkers in diagnosing patients with CAD [23].  Table 4. The autoantibody, IgG anti-IGKC 1-18 HNE, was selected as the first feature. We then performed forward selection with autoantibodies against HNE and MDA-modified peptides (Table 5, Figure 2). In the LightGBM model, the features selected included IgG anti-IGKC 1-18 HNE, IgM anti-A1AT MDA, IgM anti-IGKC 1-18 MDA, IgG anti-A2M MDA, IgG anti-A1AT MDA, and IgM anti-CFAH HNE. In differentiating HCs and CAD patients with a stenosis rate of <30%, the model received an accuracy of 75.7%, a precision of 74.8%, a f1 score of 72%, a sensitivity of 75.1%, a specificity of 77.2%, and an AUC value of 0.848. As for HCs and CAD patients with a stenosis rate of 30-70%, the model received an accuracy of 76.1%, a precision of 72.2%, a f1 score of 71.3%, a sensitivity of 75.7%, a specificity of 76.6%, and an AUC of 0.845. In discriminating HCs and CAD patients with a stenosis rate of >70%, the model received an accuracy of 82.7%, a precision of 74.5%, a f1 score of 71.9%, a sensitivity of 75.2%, a specificity of 86.4%, and an AUC value of 0.904. In the XGBoost model, the features selected included IgG anti-IGKC 1-18 HNE, IgM anti-A1AT MDA, and IgM anti-IGKC 1-18 MDA. In the discrimination of HCs and CAD patients with a stenosis rate of <30%, the model received an accuracy of 78.2%, a precision of 76.8%, a f1 score of 75%, a sensitivity of 78.4%, a specificity of 78.8%, and an AUC value of 0.847. As for HCs and CAD patients with a stenosis rate of 30-70%, the model received an accuracy of 78.6%, a precision of 76.6%, a f1 score of 74.2%, a sensitivity of 77.1%, a specificity of 79.9%, and an AUC value of 0.881. In the differentiation of HCs and CAD patients with a stenosis rate of >70%, the model received an accuracy of 86.1%, a precision of 81.7%, a f1 score of 77.2%, a sensitivity of 78.6%, a specificity of 90.4%, and an AUC value of 0.935. Statistical tests were performed to validate the improvement.

Discussion
In this study, we examined the diagnostic performance of four HNE-modified peptides in CAD that had been previously reported in patients with RA [20]. Autoantibodies found in RA have been related to cardiovascular events [24]. Thus, we speculated that the autoantibodies we discovered previously may be useful in the diagnosis of CAD. We firstly examined the IgG and IgM autoantibodies against BSA and HNE-modified BSA (Supplementary Figure S1). The results indicated the diagnostic potential of IgG and IgM autoantibodies against HNE-modified adducts. Hence, we measured the IgG and IgM autoantibodies against peptides, including CFAH, HPT, IGKC 2-19 , THRB, and HNE-modified peptides. We found that IgG and IgM autoantibodies 2-19
The proteins were discovered to be associated with the development of CAD. For instance, CFAH has been found in early human coronary atherosclerotic lesions [25]. Lee et al. suggested that HPT may be elevated in the plasma derived from CAD patients [26]. Although no study has reported the elevation of IGKC in CAD patients, MDA-modified IgG was found to be significantly elevated in CAD patients [27]. The elevation of THRB fragments was reported in patients with peripheral artery disease (PAD), which is also a cardiovascular risk factor [28,29]. Increasing the level of 4-HNE and above-mentioned proteins in CAD patients may accelerate the increment of oxidized proteins, which play a crucial role in the formation of atherosclerosis. Therefore, the 4-HNE modified peptides may also trigger inflammation and strengthen the development of atherosclerosis. Studies indicated that autoantibodies may be associated with the severity of CAD [30,31]. The decreased level of autoantibodies against 4-HNE modified peptides possibly indicates the clearance of 4-HNE or immune dysregulation [32,33]. However, the role of autoantibodies against MDA and 4-HNE modified peptides in the development of CAD requires further investigation.
To date, no other study has investigated autoantibodies against HNE-modified peptides in patients with cardiovascular disease. However, HNE increment has been associated with oxidative stress and vascular disease [34,35]. In addition, the formation of HNEmodified adducts may induce the generation of autoantibodies [36]. Hence, we speculated that the decreasing level of IgG and IgM against HNE-modified peptides may result from the immune dysfunction of patients with CAD [37].
To further explore the diagnostic ability, we conducted feature selection (feature ranking + forward selection) and machine learning models. IgG anti-IGKC 2-19 HNE was ranked as the most important feature during each ranking (Table 2). Thus, the following forward selection was initiated with IgG anti-IGKC 2-19 HNE. Various types of machine learning algorithms have been applied into medical studies. Examples include decision tree, support vector machine, random forest, and neural network [38]. Recently, two advanced tree-based models, Light Gradient Boosting Machine (lightGBM) and eXtreme Gradient Boosting (XGBoost), have been popular in other studies. Wang et al. built their miRNA classifier with LightGBM and identified hsa-mir-139 as an important feature for breast cancer diagnosis [39]. Joo et al. analyzed the cohort data from Korean National Health and developed various machine learning prediction models to estimate 2-year and 10-year risks of cardiovascular disease (CVD). LightGBM owned the highest AUC value in estimating 10 year follow-ups without medication features [40]. Al'Aref et al. incorporated clinical features and coronary artery calcium scores (CACs) with an XGBoost model to estimate the pretest of obstructive CAD on coronary computed tomography angiography (CCTA). They received the best performance with an AUC of 0.866 [41]. Kim et al. enrolled 1312 patients with obstructive CAD on coronary angiography. They built a model with the XGBoost algorithm and received an AUC of 0.820 as the best model in their experiments [42].
In our study, we performed forward feature selection with LightGBM and XGBoost. We received the highest performance with features including IgG anti-IGKC 2-19 HNE, IgM anti-A1AT 284-298 MDA, and IgM anti-IGKC 76-99 MDA. The XGBoost model received a sensitivity of 0.786 and a specificity of 0.904. Many biomarkers have been utilized in the clinic. For instance, the C-reactive protein (CRP) has been considered as a useful biomarker to predict cardiac death, AMI, and heart failure [43]. In addition, high sensitivity-CRP (hs-CRP) was found mildly elevated (up to 15 mg/L) in suspected ACS patients, which may be a meaningful prognostic marker in the clinic [44]. Furthermore, inflammation markers were studied in CAD. For example, the elevation of interleukin-6 (IL-6) was found in induced myocardial infarction [45]. Cyclophilin A was found elevated significantly in CAD patients with type 2 diabetes. It served as a high specificity biomarker for diagnosis of CAD patients with type 2 diabetes [46]. Furthermore, other researchers focus on circulating protein and RNA [47]. Together, a multiplex biomarker panel may improve the diagnosis with a comprehensive analysis [48]. In this study, we demonstrated a combination of autoantibodies against two types of oxidative modified peptides (MDA and HNE). The XGBoost model was improved the most after we incorporated autoantibodies against MDA and HNE-modified peptides together (Table 5). Our future work may incorporate other oxidative stress related markers to comprehend our oxidative model.
An in vitro diagnostic multivariate index assay (IVDMIA) combines multiple values with an algorithm to reach an improved accuracy compared to a single biomarker [49]. It has been applied to improve the diagnosis of ovarian cancer [50]. In addition, autoantibodies have been considered as valid biomarkers for various diseases such as cancer [51], RA [52], neurodegenerative disease [53], and CAD [54]. Together, these studies indicate the potential of improving diagnostic ability with an IVDMIA combined with an immunoassay in the clinic. However, several limitations should be noted. The machine learning models had been improved after we incorporated multiple autoantibodies against oxidatively modified peptides. Nevertheless, a larger sample size is required for further validation. Furthermore, other biomarkers from the clinic that may improve the diagnosis should be included, such as hs-CRP.

Conflicts of Interest:
The authors declare no conflict of interest.