The changes of immunoglobulin G N-glycosylation in blood lipids and dyslipidaemia

Written informed consent was obtained from each subject at the beginning of the study, and the study has been approved by the Ethics Committee of the Capital Medical University, Beijing, China. The ethics approval was given in compliance with the Declaration of Helsinki [32].

An observational cross-sectional study was conducted during 2012, with a total of 913 participants of Chinese Han ancestry who were recruited from a community-based survey in Beijing [24, 27]. All participants were required to meet the following inclusion criteria: (1) age equal to or greater than 18 years; (2) no medication history during the previous 2 weeks; and (3) signed informed consent prior to participation. Individuals were excluded based on the following criteria: (1) pregnant or lactating women; (2) history of mental illness or infectious disease; and (3) history of malignant tumor, stroke or other cerebrovascular disease, congenital heart disease, acute myocardial infarction, liver disease, renal failure, chronic obstructive pulmonary disease, rheumatoid arthritis or other disease.

According to the Chinese guidelines for the prevention and control of dyslipidaemia in adults [33], the participants were grouped into cases with TC ≥ 6.2 mmol/L (defined as hypercholesteremia), TG ≥ 2.3 mmol/L (defined as hypertriglyceridemia), HDL < 1.0 mmol/L (defined as the decreased HDL), or LDL ≥ 4.1 mmol/L (defined as the elevated LDL). The controls were those who did not have dyslipidaemia at the time of the study.

Each participant was required to complete a thorough health examination, which included anthropometric measurements and an analysis of the physical and chemical properties of a blood sample. Detailed information about subject recruitment and enrolment has been previously described [24, 27]. As in our previous study [24], 12 clinical traits (age, sex, body mass index (BMI), waist–hip ratio (WHR), systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood triglycerides (FBG), resting heart rate (RHR), TC, TG, HDL and LDL) were assessed in the current study [24].

To obtain the anthropometric trait data, physical examinations and interviews were carried out by trained nurses and physicians. The BMI was calculated by the formula weight (in kilograms)/height² (in metres squared). The SBP and DBP were measured twice on the right arm using a standard mercury sphygmomanometer after the subjects had rested for at least 10 min in a sitting position [34]. For the blood parameters (haematology and biochemical traits), fasting blood samples were collected in the morning after an overnight fast by venepuncture. All collected blood samples were processed within 8 h and blood plasma which was separated from whole blood was stored in freezers (− 80 °C) before the measurements. Haematology and biochemical parameters were measured by standard blood chemistry and haematology assays (Hitachi Automatic Analyzer, Model-7600, Tokyo, Japan), which were conducted at the Medical Laboratory of Beijing Xuanwu Hospital, Beijing.

As previously reported, IgG was first isolated from human plasma [24, 35]. The IgG N-glycans were cleaved and analysed by hydrophilic interaction chromatography (HILIC)- ultra-performance liquid chromatography (UPLC) as previously described [35]. In brief, after washing and equilibrating the protein G monolithic plates, 50 µL of plasma was diluted 10× with binding buffer (1× phosphate-buffered saline, pH 7.4), applied to the protein G plates, and immediately washed. The IgGs were eluted with 1 mL of 0.1 M formic acid and immediately neutralized with 1 M ammonium bicarbonate. Next, the IgG N-glycans cleavage and labelling were performed. The released N-glycans were labelled with 2-aminobenzamide, a fluorescent dye used to visualize glycans by UPLC, by multistage mixing with 2-aminobenzamide, dimethyl sulfoxide, glacial acetic acid, and 2-picoline borane. Finally, the IgG N-glycans were analysed by HILIC–UPLC in 24 IgG glycan peaks (GPs). The glycan structures of the glycans per peak were previously reported, and a detailed description is shown in Additional file 1: Table S1 [35]. The normalization and batch correction of the UPLC data was detailed in a previous study [24, 35]. The total area normalization was applied where the amount of glycans in each peak was expressed as a percentage of the total integrated area. The calculation formula is shown in Additional file 1: Table S1.

Among the 24 initial GPs, GP3 was excluded as it did not pass quality control tests. An additional 54 derived traits were calculated for the remaining 23 initial glycans [35]. Of the derived traits, 34 were of interest and were included in the analysis describing the relative abundances of galactosylation, sialylation, bisecting GlcNAc and core fucosylation (Additional file 2: Table S2).

The normal distribution of glycans was tested by the Kolmogorov–Smirnov test where P < 0.10 was considered to be statistically significant. Continuous variables underlying the normal distribution were represented as the mean ± standard deviation (SD), otherwise the medians (interquartile ranges) were used. The difference of continuous variables between two groups was tested using Student’s t test or the Wilcoxon rank-sum test. Categorical variables were represented as n (proportion), and the between-group differences were tested by the Chi square test. For the normally distributed variables, including age, BMI, WHR, SBP, DBP, FBG and RHR, t-tests were used to compare the differences between two groups. A generalized linear model (GLM) [36] was fitted to estimate the effect of blood lipids on glycosylation when adjusting for confounding factors, including age, sex, BMI, WHR, SBP, DBP, FBG and RHR, because most of the glycans (dependent variables) were not normally distributed. There were internal associations among the independent variables, which could have induced multi-collinearity in the statistical models. Therefore, principal component analysis (PCA) [37] was used to combine variables to principal components before the regression analysis. Moreover, these principal components were used in the multiple variables association analysis instead of the original variables. The internal associations among the independent variables were analysed using by the R package “corrplot” [38].

Before correlation and regression analysis, the z-score of normalized transformations for IgG N-glycans was applied to add the consistent comparability among IgG N-glycans. Canonical correlation analysis (CCA) [40] was used to determine two sets of variables of the initial glycan structures (x) and the blood lipids (y) and to find the overall correlation between the two sets of variables. The identified variables with a statistically significant impact on the canonical variables were judged by the canonical loadings. Generally, absolute values greater than 0.30 are considered to be significant loadings [39].

To explore IgG N-glycans as biomarkers to diagnose dyslipidaemia, 23 initial glycans were used as predictors. Logistic regression was performed to identify the 23 glycans that were respectively related to cases’ status after adjusting for the effect of age, sex, BMI, WHR, SBP, DBP, FBG and RHR. Considering the internal associations among these variables, we chose to correct sex, Prin.1 and Prin.2 instead of the original variables. The significant glycans were presented as the odds ratio (OR) with a 95% confidence interval (CI) and were graphed by a forest plot, which was generated using the R package “forestplot” [40]. Next, a classification model was established based on the significant glycans. Considering that there were internal associations in the glycans that might have induced multicollinearity in the model, the least absolute shrinkage and selection operator (LASSO) method was used to select glycans to reduce the dimension of data. The LASSO method was carried out by the R package “lars” [41]. The logistic regression model was then used to assess the discrimination of dyslipidaemia by combining the glycans left by the LASSO method. The classification model was only used to assess the ability of IgG N-glycans to be used as diagnostic biomarkers for disease; therefore, in the classification model there was no adjustment for the effect of age, sex, BMI, WHR, SBP, DBP, FBG and RHR. A receiver operating characteristic (ROC) curve was developed for the calculation of the area under the cure (AUC) with a 95% CI.

Data analysis was performed using SPSS Statistics version 21.0 for Windows (IBM Corp., Armonk, NY, USA), SAS software version 9.2 (SAS Institute, Chicago, IL, USA) and R version 3.3.2 (R Core Team 2016). All reported P values were two-sided, and P < 0.05 was considered statistically significant. Particularly, P < 0.05/57 (0.0009) was considered to be statistically significant in the analysis of glycans to correct for multiple comparisons.

Among the 913 recruited participants, 598 (64.50%) had a complete dataset of blood measurements and glycan traits and were included in the further analysis. In total, 8 clinical traits were described and compared between the dyslipidaemia and control groups (Table 1). Most of the traits (6 out of 8) were significantly different between the two groups (P < 0.05). The number of males and the values of age, BMI, WHR, FBG, and SBP in the dyslipidaemia group were significantly higher than those in the control group.

As shown in Additional file 3: Figure S1, the significant correlation coefficients in independent variables ranged from 0.08 to 0.89. PCA was used to combine a few variables to several principal components in order to reduce the multicollinearity of independent variables. The resulting Prin. 1 and Prin. 2 explained up to 54.78% of variance in SBP, DBP, FBG, RHR, age, BMI and WHR (Prin.1 = 0.773 × SBP + 0.783 × DBP + 0.575 × FBG + 0.106 × RHR + 0.352 × Age + 0.688 × BMI + 0.733 × WHR, Prin.2 = 0.361 × SBP + 0.364 × DBP − 0.176 × FBG + 0.739 × RHR − 0.370 × Age − 0.246 × BMI − 0.329 × WHR). The resulting Prin. 3 and Prin. 4 explained up to 85.65% of variance in TC, TG, HDL and LDL (Prin.3 = 0.985 × TC + 0.134 × TG + 0.444 × HDL + 0.923 × LDL, Prin.4 = 0.133 × TC + 0.889 × TG − 0.755 × HDL + 0.092 × LDL). Therefore, in the analysis of the association between IgG N-glycans and blood lipids, the principal components of blood lipids (Prin. 3 and Prin. 4) were used as independent variables instead of blood lipids, and they were adjusted for confounding factors (sex, Prin. 1 and Prin. 2).

The results of the normal distribution tests of glycans are shown in Additional file 4: Table S3. Of the 23 directly measured glycans, 16 were not abnormally distributed (P < 0.10). Therefore, a GLM was used to estimate the effect of the principal components of blood lipids (Prin. 3 and Prin. 4) on glycosylation, adjusting for confounding factors (sex, Prin. 1 and Prin. 2). In total, 2 GPs in the IgG N-glycans were significantly associated with blood lipids after adjusting for the effects of sex, age, BMI, WHR, FBG, RHR, SBP and DBP (P < 0.05/57). As shown in Table 2, TC, TG and LDL were positively correlated with GP6, while they were negatively correlated with GP18. In addition, HDL negatively correlated with GP6 but positively correlated with GP18. Furthermore, 9 initial traits and 7 derived traits in the IgG glycome were significantly associated with blood lipids after adjusting for the effects of sex, age, BMI, WHR, FBG, RHR, SBP and DBP (P < 0.05). Detailed information is shown in the supplemental content (Additional file 5: Table S4).

The results of CCA showed that there are 4 pairs of canonical variables, with canonical correlations of 0.390 (F = 2.44, P < 0.001), 0.324 (F = 1.88, P < 0.001), 0.240 (F = 1.39, P = 0.053) and 0.197 (F = 1.16, P = 0.282) for each successive pair by the CCA. The first and the second canonical sets were statistically significant, indicating that N-glycan structures were significantly correlated with the blood lipids. As shown in Fig. 1, 10 initial traits (GP4, GP6, GP9-12, GP14, GP17-18 and GP23) tended to be significantly associated with TC, TG and LDL levels in the first canonical set. In addition, the level of GP11 was strongly associated with canonical variables with a loading of 0.645, while the response variable with the highest canonical loading was 0.781 (TC).

Among the 23 GPs, 9 GPs (GP1, GP4, GP5, GP6, GP11, GP14, GP18, GP20 and GP21) were found to be significantly different in participants with dyslipidaemia than in controls after adjusting for confounding factors (sex, Prin. 1 and Prin. 2) (Fig. 2 and Additional file 6: Table S5). As shown in Additional file 7: Figure S2, the significant correlation coefficients in independent variables ranged from 0.11 to 0.79. The LASSO method was used to select IgG N-glycans in order to reduce the dimension of data. The remaining glycans used by the LASSO method are listed in Additional file 8: Table S6. The classification model, incorporating significant IgG N-glycans, was able to distinguish dyslipidaemia from controls, and the AUC was 0.692 (95% CI 0.644–0.740) (Fig. 3).