Human Plasma N-glycosylation as Analyzed by Matrix-Assisted Laser Desorption/Ionization-Fourier Transform Ion Cyclotron Resonance-MS Associates with Markers of Inflammation and Metabolic Health^*

Participants

The LLS, described in detail previously (35, 42), is a family-based study comprising 1671 offspring of 421 nonagenarians sibling pairs of Dutch descent, and the 744 partners of these offspring. A total of 2144 individuals with clinical blood parameters available were included in the current analysis. The study protocol was approved by the Leiden University Medical Center ethical committee and an informed consent was signed by all participants prior to participation in the study.

All standard blood measurements were performed in nonfasting venous blood samples using fully automated equipment. Glucose, high-sensitivity C-reactive protein (hsCRP), triglyceride (TG), total cholesterol, and high-density lipoprotein cholesterol (HDL) levels were measured on the Hitachi Modular P800 (Roche Diagnostics, Mannheim, Germany). Free triiodothyronine (T3) levels were measured on the Modular E170 (Roche Diagnostics). Low-density lipoprotein cholesterol (LDL) levels were calculated using the Friedewald formula (43), and set to missing if plasma TG levels exceeded 4.52 mmol/L. Insulin levels were measured on the Immulite 2500 (DPC, Los Angeles, CA). Specific sandwich enzyme-linked immunosorbent assays (ELISA) were used for the determination of adiponectin (R&D Systems Europe, Abingdon, UK), leptin (Diagnostics Biochem Canada, Dorchester, Canada) and interleukin 6 (IL-6) levels (Sanquin Reagents, Amsterdam, The Netherlands). Alanine aminotransferase (ALT) and aspartate aminotransferase (AST) levels were measured using the NADH (with P-5′-P) methodology (Modular P800, Roche Diagnostics), and gamma-glutamyl transferase (GGT) levels using the l-gamma-glutamyl-3-carboxy-4-nitroanilide substrate methodology (Modular P800, Roche Diagnostics). Dehydroepiandrosterone sulfate (DHEA-S) levels were measured with an Architect delayed one-step immunoassay (Abbot, Wiesbaden, Germany). Hypertension was defined as having a systolic blood pressure > 140 and a diastolic blood pressure > 90. Antihypertensive medication included diuretics, beta-blockers, calcium channel blockers, and agents acting on the renin-angiotensin system. Cytomegalovirus (CMV) serostatus was determined by ELISA using the CMV-IgG ELISA PKS assay (Medac, Wedel, Germany).

N-glycan Preparation

N-glycans from total plasma proteins from participants of the LLS were released, labeled with 2-aminobenzoic acid (2-AA) (Sigma-Aldrich, Steinheim, Germany) to allow negative mode mass spectrometric detection, and purified using hydrophilic-interaction liquid chromatography (HILIC)-solid-phase extraction (SPE) as previously described (44). Specifically, 20 μl of 2% sodium dodecyl sulfate (SDS) (US BioChem, Cleveland, OH) was added to 10 μl plasma, randomly distributed across twenty-seven 96-well plates, followed by protein denaturation for 10 min at 60 °C and subsequent neutralization of the SDS by 10 μl 4% Nonidet P-40 substitute (Nonidet P-40) (Sigma-Aldrich). Then, after addition of 0.5 mU peptide-N-glycosidase F (PNGase F; Roche Diagnostics) in 10 μl 5x phosphate buffered saline solution, the N-glycans were released overnight at 37 °C. Without intermediate purification, the N-glycans were labeled for 2 h at 65 °C with the addition of 50 μl 48 mg/ml 2-AA 63 mg/ml NaCNBH₃ (Merck, Darmstadt, Germany) in a 10:3 (v/v) mixture of dimethyl sulfoxide (DMSO; Sigma-Aldrich) and glacial acetic acid (Merck). HILIC-SPE was subsequently performed using 40 mg microcrystalline cellulose (Merck) in 96-well 0.45 μm GHP-filter plates (Pall, Ann Arbor, MI). All wells of the filter plate were washed using water and subsequently equilibrated using 80:20 (v/v) acetonitrile (ACN; Biosolve, Valkenswaard, The Netherlands)/water. The labeled N-glycan samples were then applied to the wells in 80% ACN, and the wells were washed using ACN/water (80:20 v/v). Purified 2-AA labeled N-glycans were eluted in 0.8-ml-deep well collection plates (ABgene via Westburg, Leusden, The Netherlands) using 400 μl water.

Carbon-SPE

Prior to analysis by MALDI-FTICR-MS, samples were additionally desalted using carbon SPE. To achieve this, 100 μl of graphitic porous carbon (Grace, Deerfield, IL) was applied to each well of an OF1100 96-well polypropylene filter plate with a 10 μm polyethylene frit (Orochem Technologies, Lombard, IL) using a 96-well column loader (Millipore, Billerica, MA). The stationary phase was activated and conditioned with 2 × 200 μl ACN/water (80:20 v/v) and 3 × 100 μl 0.1% trifluoroacetic acid (TFA; Sigma-Aldrich) in water, respectively. Of the 2-AA labeled N-glycans, 100 μl were loaded into the wells and washed using 3 × 100 μl 0.1% TFA in water. Slight vacuum was applied to facilitate the procedure. The 2-AA labeled N-glycans were eluted into a V-bottom microtiter plate (Nunc, Roskilde, Denmark) using 3 × 30 μl of freshly prepared ACN/water (80:20 v/v) containing 0.1% TFA by centrifugation at 500 rpm (154 mm rotational diameter).

MALDI-FTICR-MS Analysis

One μl of 2-AA labeled N-glycans was spotted in quadruplicate on a 384-AnchorChip target plate (Bruker Daltonics, Bremen, Germany) and air-dried. Subsequently, 1 μl of 2,5-dihydroxybenzoic acid (2,5-DHB; Bruker Daltonics) matrix (20 mg/ml in ACN/water; 50:50 (v/v)) was applied to the spots and left to dry. To generate microcrystals, 2 × 1 μl of ethanol was applied to the spots for recrystallization prior to mass spectrometric analysis.

The 9.4 T FTICR APEX-ultra mass spectrometer was equipped with a dual electrospray ionization (ESI)/MALDI ion source (Apollo II) incorporating a quadrupole mass filter and a smartbeam laser system. Before analysis, the instrument was externally calibrated by peptide calibration standard (Bruker Daltonics). All experiments used a laser spot size of ∼150 μm, laser fluence slightly above threshold, and a laser repetition rate of 200 Hz. To allow the semiquantitative analysis for the range of expected N-glycans, all samples were analyzed using two methods: one optimized for lower m/z ions (approximately m/z 1000 to 2500) and another optimized for higher m/z ions (approximately m/z 2200 to 4000). The quadrupole was operated in rf-only mode with the selection masses set to m/z 1650 and 2500 for low-mass and high-mass measurements respectively. A customized experiment sequence (pulse program) was used, in which the multiple ICR-fill parameter was reconfigured to approximate a “random-walk” functionality (38). Briefly, the ions produced from 50 laser shots were accumulated in a hexapole and then transferred through the rf-only quadrupole to the collision cell. The sample stage was then moved 200 μm, and fresh sample interrogated with the next 50 laser shots. This cycle was performed nine times, effectively accumulating ions from 450 laser shots in the collision cell. The accumulated ions were then transferred to the ICR cell for a mass analysis scan. Each spectrum is the sum of eight such scans. All data were acquired using ApexControl 3.0.0 software (Bruker Daltonics) in expert mode, controlled by Hystar 3.8 software (Bruker Daltonics) for automatic measurement. In total, 20,736 spectra were recorded, these being for each biological sample a quadruplicate of low- as well as high-mass measurements.

Data processing

Following acquisition, representative low- and high-mass spectra were internally calibrated in DataAnalysis 4.2 (Bruker Daltonics) using a set of expected glycan masses (supplemental Table S1). Using the calibrated spectra, 37 glycan compositions were manually assigned within the low-mass measurements (H4N2 to H5N4S1), and 25 within the high-mass measurements (H5N4S1 to H7N6F1S4), using mass and parts-per-million (ppm) errors to validate the assignments (supplemental Table S1) (H = hexose; n = N-acetylhexosamine; F = deoxyhexose (fucose); S = N-acetylneuraminic acid). In addition, peak widths were assessed per composition to allow precise area integration (supplemental Table S1). The resulting 61 compositions (H5N4S1 being present in both the low- and high-mass spectra) were in agreement with previously reported observations, as well as knowledge of the biological synthesis of N-glycans (2, 27, 45, 46).

To achieve repeated extraction of the list of glycan compositions from the 20,736 MALDI-FTICR-MS measurements (10,368 low-mass and 10,368 high-mass), the spectra were converted to simple text based format (x,y) using msconvert from ProteoWizard 3.0.5622 (47). The raw mass spectrometric data has been made publicly available in the MassIVE repository (massive.ucsd.edu) titled “MALDI-FTICR-MS total plasma N-glycomics” with ID: MSV000080307. Spectrum calibration, analyte integration and spectrum curation was performed using MassyTools 0.1.5.0 (48). In short, spectra were calibrated by applying the least variance second degree polynomial fit through a set of calibration masses (supplemental Table S1). During this step, spectra were excluded from further analysis when not all calibrants were detected with intensities at least 3-fold higher than the maximum signal deviation within the local noise (MinMax; roughly corresponding to a root-mean-square (RMS) signal-to-noise ratio (S/N) of 9). This led to the exclusion of 470 low-mass and 372 high-mass spectra. Glycan compositions from the analyte list were subsequently integrated by summing and grouping the areas of 95% of the theoretical isotopic envelope belonging to that composition. Prior to summation, each isotope was integrated using the peak widths previously established, and an equal width local background (within a window of 50 Thomson) was subtracted from these. To further ensure data quality, spectra were removed if more than 5% of the total analyte area was below S/N 3 (MinMax), which led to the additional exclusion of 305 low-mass and 424 high-mass spectra. After additional curation of clinical samples with no available information on age and sex, we retained a total of 16,346 spectra (8194 low-mass and 8152 high-mass), yielding glycosylation information on 2144 individuals.

To arrive at one set of glycan values per individual, the replicate spectra for the low- and high-mass spectra were averaged for that individual. In case of low-mass spectra, 1878 averages were constructed from 4 spectra, 170 from 3, 76 from 2 and 20 from 1 (supplemental Fig. S1). For the high-mass spectra, 1857 averages were constructed from 4 spectra, 177 from 3, 83 from 2 and 27 from 1. To reconstruct the overview of the total plasma N-glycome, the low- and high-mass averages were normalized on the value of their overlapping composition H5N4S1, and subsequently combined. The resulting combined pattern was normalized by dividing each glycan value by the sum of all glycan values (total area normalization). Hereof, we calculated derived glycosylation traits on basis of enzymatic steps and protein groupings (supplemental Table S2). Glycan and derived trait variation within the replicate low-mass and high-mass measurements was assessed on the basis of mean, standard deviation (S.D.) and coefficient of variation (CV) (supplemental Fig. S2; supplemental Fig. S3). Mass spectrometric figures were exported from DataAnalysis 4.2 (Bruker Daltonics) and annotated with glycan depictions following the symbol nomenclature proposed by the Consortium for Functional Glycomics (CFG), created in GlycoWorkbench 2.1 build 146 (49, 50).

Data Analysis

Throughout data analysis we employed R 3.1.2 in an environment of RStudio 0.98.1091 (RStudio Team, Boston, MA) (51). In case of significance testing, a study-wide significance threshold was maintained of α = 1.00·10⁻⁵, values below or equal being considered statistically significant. The value arises from being the lower bound of the order of magnitude of an α = 0.05 significance threshold Bonferroni corrected for the total number of regression tests throughout the study (number of tests = 27 phenotypes·100 glycan features + 26 sex comparisons + 25 age comparisons = 2751; α = 0.05/2751 = 1.82 × 10⁻⁵ ≈ 1.00 ×10⁻⁵).

To limit the experimental component within the sample variability, batch correction was performed on the glycan and derived trait variables using the R package ComBat, using sample plate as batch (52). To limit outlier influence, individual glycosylation values exceeding a 5 times S.D. value from the mean of that variable were excluded from statistical analysis. Insulin, hsCRP, IL-6, TG, adiponectin, leptin, ALT, AST, GGT, and DHEA-S levels were transformed to the natural logarithm because of nonnormal distribution of the data. In addition, to obtain interpretable estimates, the glycosylation variables were scaled before analysis (i.e. mean subtraction and division by S.D.).

Association of Variables with Age and Sex

Linear and logistic regression analyses were performed to establish the association between age and sex (female = 0; male = 1) as outcome variables, and nonglycan clinical variables (Table I), glycans (Table II) and derived glycosylation traits (Table III) as predictor variables. As no glycosylation differences were found between LLS offspring and partners (a grouping to test predisposition for longevity), these individuals were grouped for all analyses. Furthermore, because the LLS contains multiple offspring from the same family, within-family (between-siblings) dependence was taken into account by using a sandwich estimator for the standard errors (53).

Association of Glycosylation with Clinical Variables

To eliminate possible confounding effects, age, sex and the interaction thereof were included as covariates in further models. For these analyses, the remaining nonglycan variables were used as outcome (using linear and logistic regression for respectively continuous and dichotomous variables), whereas glycans and derived traits were used as predictor (model: nonglycan ∼ β₁·age + β₂·sex + β₃·age*sex + β₄·glycan).

To visualize the association between nonglycans and (derived) glycan traits, the t-statistics (or Wald statistics in case of logistic regression, both β₄/S.E.₄) from the models were expressed in heatmap format. Sorting of the heatmap variables was performed using hierarchical clustering (Euclidean distance, complete linkage).

Measurement Variability

Following spectrum curation, we retained a total of 16,346 mass spectra originating from low- and high-mass measurements of MALDI spotting quadruplicates for each individual. In the measurement optimized for lower masses (m/z 1000 to 2500) 37 N-glycans could be detected, ranging from H4N2 to H5N4S1, with an average absolute ppm error of 2.24 (S.D. ± 3.18). In the measurement optimized for higher masses (m/z 2200 to 4000) 25 additional N-glycans were detected, from the overlapping composition H5N4S1 to H7N6F1S4, the average absolute ppm error being 3.65 (S.D. ± 3.98) (supplemental Table S1). Based on literature, the glycan compositions within the TPNG were presumed to have certain structural features (16, 27, 45, 46). Examples of this are the antennarity, judged as the number of N-acetylhexosamines minus two unless bisected, and bisection, judged to be the case if the number of N-acetylhexosamines equaled five and the number of hexoses five or less. Although these structural assignments are expected to represent the majority of structures contributing to an MS signal, additional structural isomers are likely to be present in the signals. For example, a composition assigned as tetraantennary may instead contain diantennary structures with two N-acetyllactosamine repeats, and the bisected species could be triantennary with incomplete galactosylation.

Assessing repeatability, an example quadruplicate measurement from a single individual yielded an average CV of 6.52% (S.D. ± 3.42%) for the 10 most abundant signals in the low-mass measurement (total area normalized for the mass range), and an average CV of 5.97% (S.D. ± 2.53%) for the 10 most abundant signals in the high-mass measurement (supplemental Fig. S2A; supplemental Fig. S2B). Combining the measurements by the overlapping composition H5N4S1 yielded for the 20 must abundant species an average CV of 9.29% (S.D. ± 5.88%) (supplemental Fig. S2C). Derived glycosylation traits, constructed to provide mathematical expressions of monosaccharide differences and groupings with structural similarity, showed a lower CV, a phenomenon previously observed for mass spectrometric plasma glycomics (54), i.e. on average 1.20% (S.D. ± 0.84%) for the 20 most abundant members (supplemental Fig. S2D). For a listing of derived traits and their calculations see supplemental Table S2.

In total, the glycosylation analysis workflow allowed for 2144 individuals the relative quantification of 61 N-glycan compositions and 39 derived traits. The LLS provided an additional 27 clinical variables to facilitate association analysis. Next to age and sex, measures were included on liver function (GGT, ALT, AST, AST/ALT), glucose metabolism (glucose, insulin, glucose/insulin), lipid metabolism (cholesterol, LDL-C, HDL-C, cholesterol/HDL-C, TG, lipid lowering medication, leptin, adiponectin), inflammation (hsCRP, IL-6), blood pressure (hypertension, antihypertensive medication), adrenal function (DHEA-S), thyroid function (free T3), as well as information on BMI, smoking and CMV infection, and familial propensity for longevity (Table I).

Association of Glycosylation with Age and Sex

Glycosylation was found to highly associate with age and sex by respectively linear and logistic regression analysis (Fig. 2; Table II; Table III). A GEE approach was used for all statistical analyses to adjust the standard errors (S.E.) for between-sibling dependence, and a study-wide significance threshold was maintained of α = 1.0·10⁻⁵. Changes with aging included a decrease of galactosylation of diantennary glycans, visible most specifically for the galactosylation of nonsialylated diantennaries with fucose (β_A2FS0G = −2.82 S.E. ± 0.13; p_A2FS0G < 2.2 × 10⁻¹⁶) and without fucose (β_A2F0S0G = −1.34 ± 0.15; p_A2F0S0G < 2.2 × 10⁻¹⁶). These changes were mainly driven by the increases in glycan compositions H3N4 (β_H3N4 = 1.21 ± 0.14; p_H3N4 < 2.2 × 10⁻¹⁶), H3N5 (β_H3N5 = 1.68 ± 0.14; p_H3N5 < 2.2 × 10⁻¹⁶), H3N4F1 (β_H3N4F1 = 1.50 ± 0.14; p_H3N4F1 < 2.2 × 10⁻¹⁶), H3N5F1 (β_H3N5F1 = 1.67 ± 0.14; p_H3N5F1 < 2.2 × 10⁻¹⁶) and the decreases in H5N4F1 (β_H5N4F1 = −1.60 ± 0.15; p_H5N4F1 < 2.2 × 10⁻¹⁶) and H5N4F1S1 (β_H5N4F1S1 = −0.81 ± 0.15; p_H5N4F1S1 = 4.5 × 10⁻⁸). Further increasing with age were the bisection of nonsialylated fucosylated diantennaries (β_A2FS0B = 0.77 ± 0.14; p_A2FS0B = 6.7 × 10⁻⁸), sialylation per galactose of fucosylated diantennaries (β_A2FGS = 1.14 ± 0.15; p_A2FGS = 9.8 × 10⁻¹⁵), and the fucosylation of both tri- and tetraantennary compositions (β_A3F = 0.82 ± 0.15; p_A3F = 7.4 × 10⁻⁸ and β_A4F = 0.77 ± 0.15; p_A4F = 2.0 × 10⁻⁷).

Fucosylation of triantennary and tetraantennary structures (A3F and A4F) proved also to be a major glycosylation difference between females and males (female = 0; male = 1) (β_A3F =0.73 ± 0.05; p_A3F < 2.2 × 10⁻¹⁶ and β_A4F = 0.74 ± 0.05;p_A4F < 2.2 × 10⁻¹⁶), driven by higher male values in all fucosylated tri- and tetraantennary compositions, such as H6N5F1S3 (β_H6N5F1S3 = 0.61 ± 0.05; p_H6N5F1S3 < 2.2 × 10⁻¹⁶) and H7N6F1S4 (β_H7N6F1S4 = 0.45 ± 0.05; p_H7N6F1S4 < 2.2 × 10⁻¹⁶), and significantly lower levels of all nonfucosylated tri- and tetraantennary compositions, such as H6N5S3 (β_H6N5S3 =−0.61 ± 0.05; p_H6N5S3 < 2.2 × 10⁻¹⁶) and H7N6S4 (β_H7N6S4 = −0.26 ± 0.05; p_H7N6S4 = 5.8 × 10⁻⁸). Furthermore, males proved to have lower bisection of diantennary fucosylated species (β_A2FB = −0.27 ± 0.04; p = 3.8 × 10⁻¹⁰) when compared with females, but did have a higher sialylation per galactose of tetraantennary nonfucosylated compositions (β_A4F0GS = 0.34 ± 0.04; p = 2.5 × 10⁻¹⁴).

Association Glycosylation With Inflammation and Metabolic Health

Regression analysis was used to establish the relationship between the glycosylation traits and clinical markers, adding age, sex and the interaction thereof as covariates to limit their confounding influence (supplemental Table S3). The t-statistics (or Wald-statistics; both β₄/S.E.₄) arising from the models were expressed in clustered heatmap format (Fig. 3; for a heatmap visualization of the results without age and sex adjustment see supplemental Fig. S4). Significantly associating with a selection of derived glycosylation traits were hsCRP (23 statistically significant associations out of a possible 39), GGT (14), BMI (13), leptin (13), smoking (10), TG (9), insulin (7), the ratio of total cholesterol and HDL (7), the ratio of glucose and insulin (5), adiponectin (5), HDL (4), IL-6 (3), ALT (2), the ratio of ASL and ALT (1) and lipid medication (1) (supplemental Table S4; supplemental Table S5). Only individual N-glycan associations could be proven for AST, glucose, cholesterol, LDL and free T3, whereas no associations were found for hypertension, the usage of antihypertensive medication, DHEA-S and CMV infection.

Inflammatory marker hsCRP showed the most associations with the total plasma N-glycome, including a positive association with tri- and tetraantennary glycans (β_A3 = 0.24 ± 0.03; p_A3 < 2.2 × 10⁻¹⁶ and β_A4 = 0.20 ± 0.03; p_A4 = 2.9 × 10⁻¹³) at the expense of high-mannose (β_M = −0.18 ± 0.02; p_M = 1.2 × 10⁻¹³), hybrid (β_Hy = −0.24 ± 0.03; p_Hy < 2.2 × 10⁻¹⁶), monoantennary (β_A1 = −0.17 ± 0.02; p_A1 = 1.8 × 10⁻¹²) and diantennary species (β_A2 = −0.25 ± 0.03; p_A2 < 2.2 × 10⁻¹⁶). Although fucosylation of diantennary species proved to decrease with higher hsCRP levels (β_A2F = −0.13 ± 0.02; p_A2F = 1.5 × 10⁻⁷), an increase was seen in the fucosylation of triantennary species (β_A3F = 0.13 ± 0.03; p_A3F = 1.0 × 10⁻⁶). Additional increases were found for sialylation of (most specifically) fucosylated diantennary (β_A2FGS = 0.25 ± 0.02; p_A2FGS < 2.2 × 10⁻¹⁶) and triantennary (β_A3FGS = 0.15 ± 0.02; p_A3FGS = 1.0 × 10⁻⁹) species, as well as an increase in average high-mannose size (β_MM = 0.13 ± 0.02; p_MM = 7.9 × 10⁻⁸) and a decrease in bisection of the nonfucosylated diantennaries in particular (β_A2F0B = −0.14 ± 0.02; p_A2F0B = 1.4 × 10⁻⁸). Notably, although galactosylation of nonsialylated diantennaries without fucose increased (β_A2F0S0G = 0.12 ± 0.02; p_A2F0S0G = 2.6 × 10⁻⁷), galactosylation of the same species, but with fucose, decreased instead (β_A2FS0G = −0.20 ± 0.03; p_A2FS0G = 1.7 × 10⁻¹²). Interestingly, the associations observed with hsCRP could only in part be translated to the upstream cytokine IL-6 (55, 56), which only showed reproduction of the decreased A2F galactosylation (β_A2FS0G = −0.19 ± 0.03; p_A2FS0G = 4.6 × 10⁻¹²) and increased sialylation per galactose thereof (β_A2FGS = 0.25 ± 0.02; p_A2FGS < 2.2 × 10⁻¹⁶).

BMI and leptin proved highly similar with respect to total plasma N-glycosylation associations, and showed considerable overlap with the aforementioned hsCRP as well. Taking an increase in BMI as example, galactosylation was decreased for fucosylated diantennary species (β_A2FS0G = −0.55 ± 0.10; p_A2FS0G = 9.6 × 10⁻⁹) and increased for nonfucosylated variants (β_A2F0G = 0.58 ± 0.09; p_A2F0G = 9.9 × 10⁻¹²). Changes were also seen with average high-mannose size (β_MM = 0.89 ± 0.08; p_MM < 2.2 × 10⁻¹⁶), bisection of nonfucosylated diantennary species (β_A2F0B = −0.39 ± 0.08; p_A2F0B = 2.0 × 10⁻⁶), and sialylation of diantennary glycans (β_A2GS = 0.60 ± 0.08; p_A2GS = 3.1 × 10⁻¹³). All of these effects were replicable within both leptin and hsCRP. However, shared with leptin but not seen with hsCRP, was the negative association of BMI with the sialylation of tetraantennary species, and specifically the nonfucosylated variants thereof (β_A4F0GS = −0.74 ± 0.08; p_A4F0GS < 2.2 × 10⁻¹⁶). When including hsCRP and leptin as variables in the model between BMI and glycosylation, most associations between the latter two are lost with the exception of high-mannose size (β_MM = 0.32 ± 0.07; p_MM = 4.0 × 10⁻⁷) and a strong remaining trend with glycan composition H4N4S1 (β_H4N4S1 = −0.24 ± 0.07; p_H4N4S1 = 1.8 × 10⁻⁴) (supplemental Table S6).

The liver marker GGT showed major associations with the total plasma N-glycome, whereas AST, ALT, and the ratio thereof were of only minor influence. GGT appeared similar to hsCRP in changes of increased galactosylation of nonfucosylated diantennary glycans (e.g. β_A2F0G = 0.08 ± 0.02; p_A2F0G = 2.5 × 10⁻⁷), sialylation of fucosylated diantennaries (e.g. β_A2FGS = 0.09 ± 0.01; p_A2FGS = 9.9 × 10⁻¹²), as well as an overall increased in antennarity (e.g. β_A3 = 0.07 ± 0.01; p_A3 = 3.7 × 10⁻⁸). However, the marker showed a decrease in tetraantennary sialylation similar to BMI and not hsCRP (β_A4F0GS = −0.11 ± 0.01; p_A4F0GS < 2.2 × 10⁻¹⁶), while lacking the decreasing galactosylation of nonsialylated fucosylated diantennaries (A2FS0G) prominently seen in both BMI and hsCRP.

Clinical markers considered of beneficial metabolic nature, in our study represented by glucose/insulin, HDL-C and adiponectin (57, 58), showed associations largely opposite to those established for hsCRP, BMI and GGT. Examples of this include, in case of adiponectin, the decreased galactosylation of nonfucosylated diantennaries (β_A2F0G = −0.06 ± 0.01; p_A2F0G =3.1 × 10⁻⁸), a decreased sialylation of diantennaries (β_A2GS = −0.05 ± 0.01; p_A2GS = 2.1 × 10⁻⁶), a decreased average high-mannose size (β_MM = −0.08 ± 0.01; p_MM = 1.7 × 10⁻¹⁵), and an increased sialylation of tetraantennary nonfucosylated species (β_A4F0GS = 0.06 ± 0.01; p_A4F0GS = 3.1 × 10⁻⁹).

Interestingly, the only clinical markers affecting tri- and tetraantennary fucosylation (after correction for sex) proved to be smoking with a positive association (β_A3F = 0.50 ± 0.08; p_A3F = 2.6 × 10⁻¹⁰ and β_A4F = 0.46 ± 0.08; p = 2.3 × 10⁻⁸), and TG levels with a negative association (β_A3F = −0.09 ± 0.01; p_A3F = 8.4 × 10⁻¹⁴ and β_A4F = −0.09 ± 0.01; p_A4F = 1.6 × 10⁻¹³). In addition, smoking was also the only clinical variable to positively associate with the bisection of fucosylated nonsialylated diantennary species (β_A2FS0B = 0.56 ± 0.07; p_A2FS0B = 4.7 × 10⁻¹⁵). Of particular interest is the high-mannose size trait (MM), which shows several of the strongest correlations, positively associating with cholesterol/HDL-C (β_MM = 0.08 ± 0.01; p_MM < 2.2 × 10⁻¹⁶), TG levels (β_MM = 0.11 ± 0.01; p_MM < 2.2 × 10⁻¹⁶), leptin (β_MM = 0.20 ± 0.02; p_MM < 2.2 × 10⁻¹⁶) and BMI (β_MM = 0.89 ± 0.08; p_MM < 2.2 × 10⁻¹⁶), whereas negatively associating with glucose/insulin (β_MM = −0.10 ± 0.02; p_MM = 3.0 × 10⁻¹⁰), HDL (β_MM = −0.09 ± 0.01; p_MM < 2.2 × 10⁻¹⁶), and adiponectin (β_MM = −0.08 ± 0.01; p_MM = 1.7 × 10⁻¹⁵). Considering the individual glycans contributing to the MM trait, the significant compositions proved to be H5N2 and H9N2.

Methodology

MS analysis of plasma N-glycosylation is not without challenges, particularly when employing MALDI ionization. A downside of this technique is the in-source and metastable loss of sialic acid residues, which is problematic given that plasma N-glycans are often highly sialylated (a notable exception being those from the fragment crystallizable (Fc) region of IgG) (16, 37). Although chemical derivatization prior to MALDI-MS has been found to solve these stabilization issues (e.g. by permethylation, methyl/ethyl esterification or amidation), this adds extra steps to the sample preparation workflow and often leads to byproducts (31, 64–69). Instead, here we have used MALDI-FTICR-MS equipped with an intermediate pressure source to decrease sialic acid decay (36, 38). The intermediate pressure source is known to promote ion integrity by cooling of the MALDI-generated ions (36, 39–41). In our study this has facilitated the relative quantification of N-glycan species containing up to four N-acetylneuraminic acids.

Although similarly sized investigations have analyzed plasma N-glycosylation by separation techniques like (U) HPLC or CGE-LIF (18–20, 23, 70), to our knowledge this is the largest study of its kind performed by MS (32–34). Glycan compositions as obtained by MS provide orthogonal information to chromatographic peaks. UHPLC, for example, can separate diantennary N-glycan isomers with α1,3- versus α1,6-arm galactosylation and can likewise distinguish an antennary from a bisecting GlcNAc, whereas MS can provide more precise groups of di-, tri-, and tetraantennary compositions, number of sialic acids, and clear separation of high-mannose type glycans (29, 71). As such, the construction of derived glycosylation traits making use of these features, while still biased on a compositional level, is simple to perform and provides additional insight into the complexity of glycan changes.

One example of the added information of derived traits can be found in the association of glycosylation with hsCRP. On an individual glycan level, we can only observe a relative increase in tri- and tetraantennary compositions together with the inflammation marker, and a corresponding relative decrease in all other compositions (which may be because of the total area normalization). However, when we mathematically take several individual glycans with shared biological features out of the total plasma N-glycome and compare them relative to each other in the form of a derived trait, we now additionally reveal, for example, a decrease in galactosylation within the subset of glycan compositions predominantly occurring on IgG-Fc (nonsialylated fucosylated diantennary species; A2FS0), a finding expected from literature (16, 72, 73).

Additional aspects of the analytical methodology need to be considered to allow valid interpretation of the presented findings. First, N-glycans are released from their protein backbones, and thus information is lost whether an observed glycan change originates from protein glycosylation or from glycoprotein abundance. Second, mass spectrometry does not distinguish isomers, meaning that a given monosaccharide mass, e.g. a hexose, is assigned differently based on literature knowledge of its compositional context (e.g. as mannose for H5N2 and as galactose for H4N4F1) (16, 27, 45, 46). Similarly, literature is the main source of information on the linkages between monosaccharides, for example presuming bisection for compositions having five N-acetylhexosamines but less than three galactoses (e.g. H5N5F1). Although this group indeed encompasses bisection, the information within will also be confounded by triantennary structures with incomplete galactosylation, even if these are not abundant in human plasma (16, 27). Third, the methodology presented here will not provide biologically true relative ratios of glycan compositions, as the profiles will, for example, be skewed by the ionization advantage of sialylated species in negative ion mode MS. Nonetheless, the relative signal differences will still be representative for the biological directions of change, as well as providing an estimate of the magnitude.

Clinical Findings

With the help of literature, we can speculate on the biological background of the observed associations. Acute inflammation, represented in our study by hsCRP and IL-6, shows to confirm previous glycomics studies with regard to the increase in antennarity, fucosylation of triantennary species, and a decrease in galactosylation of IgG glycans which is particularly well-established (9, 17, 21, 24, 59, 74). The first two may be explained by an inflammation-induced increase in plasma levels and glycosylation changes of acute phase proteins such as alpha-1-antitrypsin and alpha-1-acid glycoprotein (orosomucoid-1) (9, 75). Not only are these carriers of N-glycans with two or more antennae at baseline conditions, they furthermore display increased antennarity and sialyl-Lewis X upon acute inflammation (76–78). Additional observations include the increase of galactosylation and sialylation per galactose of diantennary fucosylated (A2F) species in general, which is notably different from the behavior of IgG. A contributor to this observation could be the level of IgM, an abundant immunoglobulin shown to increase with many autoimmune and inflammatory conditions and which carries the required highly galactosylated and sialylated A2F species (79, 80).

Of particular interest is the highly significantly increased size of high-mannose glycans (MM) which is not only observed with increased hsCRP, but as well with increasing BMI, non-HDL cholesterol and TG. This glycosylation change appears to largely stem from the increase in the single glycan composition H9N2, an analyte difficult to assess by commonly used liquid chromatography with fluorescence detection (23, 81). Within the total plasma N-glycome, this large high-mannose glycan may predominantly originate from apolipoprotein B, the main protein constituent of most non-HDL lipoproteins (e.g. VLDL, LDL) (16, 82, 83). The observed correlations with H9N2 and MM can represent differing apolipoprotein B levels or glycosylation thereof, and may be indicative of an unhealthy glycosylation profile with regard to lipid transport and metabolism. The association of MM with inflammatory marker hsCRP is likely a consequence of the connection between inflammation and obesity (84), as a model corrected for BMI no longer shows significant association between high mannose size and inflammation (data not shown). The effect size being less pronounced with hsCRP than with BMI is explainable by the aforementioned increase in IgM, which contains the smaller high-mannose compositions H5N2 and H6N2 at its Asn²⁷⁹ site (80). On the other hand, the association between MM and BMI remains true in a model adjusted for CRP and leptin, indicating the glycosylation trait may have potential to discriminate healthy from unhealthy obese.

Although largely following an inflammatory glycosylation profile, BMI, leptin, non-HDL cholesterol and TG do show a unique negative association with the sialylation of nonfucosylated tetraantennary compositions (A4F0GS), a change mainly occurring because of the relative increase of the lowly sialylated H7N6S1 and H7N6S2. Many proteins may contribute to this decrease in sialylation, a notable one being alpha-1-acid glycoprotein (85–87), which is known to bind lipophilic compounds with affinity modulated by its degree of sialylation (88–90). This effect also applies to binding of hydrophobic drugs, possibly explaining why we similarly observed trends of decreased A4F0GS with the usage of lipid- and antihypertensive medication (89, 90). Additionally, increased levels of glycoproteins with exposed galactose residues can reflect a modulation of protein turnover, e.g. recycling by the liver-based asialoglycoprotein receptors (5, 91–93). Low sialylation would induce rapid turnover of lipid scavengers like alpha-1-acid glycoprotein, a situation which is beneficial when the blood needs to be cleared of high levels of lipophilic compounds.

Smoking proved the only phenotype to positively associate with the fucosylation of tri- and tetraantennary species (A3F and A4F), as well as with the bisection of IgG-Fc type glycans (A2FS0B) (20, 94). Likely these reflect the chronic response to vascular injury obtained from smoking-induced shear stress and oxidative damage (95). Increased fucosylation of acute phase proteins, when antennary-linked in the form of sialyl-Lewis X or A, facilitates recruitment to sites of injury, e.g. by interactions with selectins presented on inflammation-activated endothelial cells (16, 76, 96).

Although it is invalid to interpret absence of statistical significance as an absence of association, we have nevertheless failed to identify derived traits or single glycans to be predictive of DHEA-S, CMV infection or the propensity for longevity. The latter is a peculiar absence, as plasma N-glycosylation analysis of the same cohort by HPLC had revealed two chromatographic peaks to be predictors of the phenotype (18). Reasons for this lack of biological reproduction may include the measurement error (HPLC tends to provide more robust measurements than mass spectrometry) (71), the chromatographic peaks representing a culmination of multiple mass spectrometric compositions that are not individually significant, the previous findings being incidental, or the inability of mass spectrometry to separate or detect the responsible analytes.

Glycosylation is of high interest for the development or improvement of biomarkers for disease and patient stratification (14, 97). In particular, the findings within this study may be of benefit to the detection and discrimination of metabolic syndrome or inflammatory disorders, and may bolster the predictability of existing biomarkers such as the Framingham Risk Score (98). Important glycosylation phenotypes in this regard would then be the derived traits MM, A3F, A4F, A2F0G, A2F0B, A2FS0G, A2FGS, and A4F0GS, as well as the various individual glycans comprising these groups, e.g. most tri- and tetraantennary compositions, high-mannose compositions H5N2, H6N2, and H9N2, and the truncated N-glycans suggested by compositions H3N3, H4N4, and H4N4S1. Interestingly, although this last example, H4N4S1, is a glycan composition difficult to characterize in a derived trait, it does show to be the major single glycan to positively associate with most beneficial markers of metabolic health (e.g. glucose/insulin, HDL-C, adiponectin) and negatively with most detrimental ones (e.g. hsCRP, BMI, GGT, hypertension, TG). The protein source of the N-glycan is as of yet unclear, but is of interest to study in more detail.

To summarize, we have reported a large number of associations between total plasma N-glycosylation as measured by MALDI-FTICR-MS and clinical markers of metabolic health and inflammation. By this, we have identified glycan compositions and derived traits indicative of overall metabolic health and inflammation, as well as finding glycosylation traits uniquely associating with single marker variables. With this knowledge, we hope to contribute to the interpretation of the plasma N-glycome as biomarker for health and disease, and to assist clinical translation of mass spectrometric glycosylation analysis.