Pregnancy-associated serum N-glycome changes studied by high-throughput MALDI-TOF-MS

Pregnancy requires partial suppression of the immune system to ensure maternal-foetal tolerance. Protein glycosylation, and especially terminal sialic acid linkages, are of prime importance in regulating the pro- and anti-inflammatory immune responses. However, little is known about pregnancy-associated changes of the serum N-glycome and sialic acid linkages. Using a combination of recently developed methods, i.e. derivatisation that allows the distinction between α2,3- and α2,6-linked sialic acids by high-throughput MALDI-TOF-MS and software-assisted data processing, we analysed the serum N-glycome of a cohort of 29 healthy women at 6 time points during and after pregnancy. A total of 77 N-glycans were followed over time, confirming in part previous findings while also revealing novel associations (e.g. an increase of FA2BG1S1(6), FA2G1S1(6) and A2BG2S2(6) with delivery). From the individual glycans we calculated 42 derived traits. With these, an increase during pregnancy and decrease after delivery was observed for both α2,3- and α2,6-linked sialylation. Additionally, a difference in the recovery speed after delivery was observed for α2,3- and α2,6-linked sialylation of triantennary glycans. In conclusion, our new high-throughput workflow allowed the identification of novel plasma glycosylation changes with pregnancy.

immunity and inflammation 18,19 . In contrast, targeted analysis of IgG glycosylation would miss the observation of a sialyl-Lewis motif, for the simple reason that none of the IgG glycoforms contain this epitope 20 .
The profiling of a complex sample, such as TSNG, can be performed using a combination of separation and detection methods. For example, HILIC (hydrophilic interaction liquid chromatography) may be combined with either CGE-LIF (capillary gel electrophoresis -laser induced fluorescence) or HPLC-FL (high-performance liquid chromatography -fluorescence) 12,21 . Alternatively, glycans can be permethylated and measured by LC-MS (liquid chromatography -mass spectrometry) 22 . A different technique extensively used for high-throughput profiling is MALDI-TOF-MS (matrix-assisted laser desorption/ionization -time-of-flight -mass spectrometry), which allows for the acquisition of a glycan spectrum within seconds, while LC based methods generally require much more time 23,24 . Moreover, technical challenges posed by the instability and ionization bias observed for sialylated glycans have been successfully overcome by permethylation as well as by selective stabilisation of sialic acids by e.g. ethyl esterification [25][26][27][28][29][30] . The latter method introduces a mass difference between α 2,3-and an α 2,6-linked sialic acids, allowing their discrimination by mass spectrometry, and is well suited for high-throughput N-glycan profiling using MALDI-TOF-MS 27 .
An important issue that arises when analysing larger sample sets by MALDI-TOF-MS is data processing, which is often the rate-limiting step. The recently developed processing software package, MassyTools, offers a valid solution to this problem by rapidly integrating a predefined list of analytes from any number of mass spectra 31 . Additionally, MassyTools calculates a wide variety of spectral and analyte quality criteria that allow for an efficient and reliable automated curation of (clinical) data.
In this study, we applied high-throughput MALDI-TOF-MS profiling together with a newly developed, automated data processing tool to the TSNG analysis of a longitudinal cohort of healthy women (n = 29) encompassing time points during pregnancy and after delivery 31 . Pregnancy-associated glycosylation changes have previously been studied in this cohort for specific proteins such as IgG, AAT and immunoglobulin A (IgA), as well as for serum glycans using CGE-LIF methodology 12,32 . The use of derived traits allowed us to obtain additional biological insight, as the calculated derived traits are mainly based on single enzymatic step monosaccharide modifications. Furthermore, the method allowed us to identify more glycans with a higher confidence, including differentiation of some isomers, than with the previously mentioned CGE-LIF based study. The utilised methods for sample preparation, as well as data processing, proved to be fast and reliable. Data analysis revealed previously unnoticed N-glycomic changes with differential dynamics in sialylation and abundance of triantennary glycans.

Results
The TSNG profiles of 29 healthy Caucasian women, screened at three time points during pregnancy and three time points after delivery, were analysed using MALDI-TOF-MS after enzymatic glycan release and linkage-specific sialic acid derivatisation ( Fig. 1; Supplementary Fig. S1) 27 . The entire workflow is displayed in Fig. 2 Table S3). The retained compositions included 56 sialylated structures, of which 31 contained α 2,3-linked and 45 contained α 2,6-linked sialylation. Furthermore, we extracted 5 oligomannosidic, 7 hybrid, 32 fucosylated and 13 bisected compositions. All calculated glycosylation traits were compared pairwise using a Wilcoxon signed-rank test to highlight differences between time points (Table 1; Supplementary  Table S4). Additionally, all (77) individual glycan traits were compared pairwise using a Wilcoxon signed-rank test (Supplementary Table S5). A summary of the individual glycan results can be found in the Supplementary  Table S6. Bonferroni correction was used to address the multiple testing problem, by adjusting the significance threshold (e.g. a significance level α < 0.0012 for 42 tests per time point) 33 . TSNG changes during pregnancy. When comparing TSNG glycosylation features between the 1 st and 3 rd trimester (trim) of pregnancy, statistical analyses revealed a rise in the level of triantennary glycans (A3; 10.5% to 12.6%; p = 0.0005; Fig. 3B) as well as complex changes in sialylation patterns: we observed increased levels of α 2,3-linked sialylation of complex glycans (AL; 9.4% to 10.8%; p = 0.0003; Fig. 3K), especially in diantennary glycans (A2L; 6.2% to 7.1%; p = 0.0004; Fig. 3L). Furthermore, there was an increase of α 2,6-linked sialylation in triantennary fucosylated glycans (A3FE; 54.4% to 56.5%; p = 0.0001; Fig. 3T). The increase of sialylation in diantennary glycans is in line with an increase of galactosylation of diantennary fucosylated glycans (A2FG; 86.8% to 90.1%; p < 0.0001; Fig. 3E). However, we observed only a trend towards an increase in the related traits of galactosylation of all complex type glycans (AG; 96.0% to 97.0%; p = 0.0019; Fig. 3C) and galactosylation of diantennary glycans (A2G; 95.8% to 96.9%; p = 0.0014; Fig. 3D). Interestingly, we also observed a trend towards an increase in antenna fucosylation, indicative of the presence of a sialyl-Lewis epitope, although this was found non-significant after multiple testing correction (0.07% to 0.09%; p = 0.0019; data not shown).

Discussion
The total serum N-glycome of a set of 174 serum samples from 29 individuals taken at 3 time points during pregnancy and 3 time points after delivery was analysed using a recently developed glycan derivatisation technique able to differentiate between α 2,3-and α 2,6-linked N-acetylneuraminic acids (sialic acids) 27 . This was used in combination with high-throughput MALDI-TOF-MS detection and a high-throughput data extraction software package 31 . The sample set has previously been studied by CGE-LIF, showing an increase in highly sialylated di-, tri-and tetraantennary glycans and a decrease in diantennary structures with zero or one sialic acid, which was confirmed in the current study 12 . While CGE-LIF potentially differentiates more isomers than the method applied herein, the usefulness of the separation is often compromised by the overlap of glycan species due to a necessary compromise between resolution and throughput 34 . MALDI-TOF-MS normally cannot separate isomers without fragmentation. However, by combining it with a recently developed sialic acid derivatisation method, we did obtain information on the sialic acid linkage 27 . Furthermore, we applied statistics to the derived glycosylation traits instead of individual glycans, as derived traits offer a higher robustness and more directly provide insights into changes in glycosylation pathways 35 . The CGE-LIF study could not use derived traits as the potential overlap of glycan species with diverging properties compromises the construction of derived traits with unambiguous    Supplementary  Table S2.
biological meaning. We compared our results with the CGE-LIF study at the individual glycan level. The comparison shows that we reproduced many of the previously measured CGE-LIF results and additionally found many novel findings, including glycosylation changes during pregnancy (1 st trimester vs. 3 rd trimester), and that various additional glycoforms, including isomers, change with delivery (3 rd trimester vs. 6 WPP) and long term recovery (3 rd trimester vs. 26+ WPP; Supplementary Table S7). For instance, we observed an increase of H5N5E2, H4N5F1E1 and H4N4F1E1 with delivery, while all of these glycans were not detected using CGE. However, there were also results that we were not able to reproduce, such as an increase in H5N4 with the delivery. In the previous paper the responsible CGE-LIF peak had been assigned to two potential glycoforms, namely H5N4 and H4N5F1. In the current study by MALDI-TOF-MS, however, we did not identify the H5N4, while the H4N5F1 peak we did detect matched the CGE-LIF data. Therefore, in the light of the mass spectrometric data, one possibility would be that the previously detected CGE-LIF peak did not consist of both H5N4 and H4N5F1, but rather of only H4N5F1. The applied MALDI-TOF-MS method showed no changes in either oligomannosidic or hybrid glycans, confirming previous results by CGE-LIF 12 . However, a variety of glycosylation changes was observed in complex glycans over time (Fig. 3). The glycosylation changes in serum are caused in part by changes in protein-specific glycosylation and in part by changes in glycoprotein levels due to pregnancy (e.g. decreasing levels of AAT and IgG) 3,12,13,32,36 .
We observed an increase in galactosylation of diantennary fucosylated glycans during pregnancy, followed by a sharp decrease with delivery (A2FG; Fig. 4). In addition, we observed a trend for increasing α 2,6-linked sialylation of diantennary fucosylated glycans (A2FE; 46.4% to 48.9%), but this was not significant after multiple testing correction (p = 0.0141). IgG is one of the most prominent serum proteins carrying diantennary fucosylated glycans, of which only a small fraction contains α 2,6-linked sialic acids 6 . The levels of IgG decrease with pregnancy from 9.5 mg/mL (7-17 weeks of pregnancy) to 7.8 mg/mL (34-38 weeks), resulting in an increase in the contribution of other proteins, such as IgM, to the A2FE trait 36 . Together with the known increase in α 2,6-linked sialylation of IgG N-glycans during pregnancy this results in a major contribution of IgG-related changes to the observed increase in A2FE 3,32 . The fact that this increase is hardly visible in TSNG analysis may indicate that changes in other glycoproteins, with respect to concentration and glycosylation, counteract the IgG-associated changes. Interestingly, IgM shows an opposite trend with pregnancy with its levels increasing from 1.2 mg/mL during 7-17 weeks of pregnancy to 1.46 mg/mL during 34-38 weeks of pregnancy 36 . While IgM is known to carry diantennary core-fucosylated N-glycans with rather high levels of α 2,6-linked sialylation, its glycoform-related pregnancy-associated glycosylation changes are not known 6,37 . Hence, changes in IgM glycosylation over the course of pregnancy should be investigated.
Our results show specific longitudinal glycosylation changes in serum triantennary glycans. The levels of triantennary structures and their α 2,3-or α 2,6-linked sialylated variants show an increase during pregnancy (1 st trimester to 3 rd trimester) followed by a decrease after delivery (3 rd trimester to 6 WPP). This could reflect changes in the protein concentrations or glycosylation of abundant glycoproteins that contain triantennary glycans, such as AAT, HPT and AGP 6,[38][39][40] . The protein concentration of AGP remains stable throughout pregnancy, but whether the glycosylation levels change with pregnancy is unknown 36 . In contrast, AAT and HPT both increase in concentration during pregnancy and could therefore be responsible for the observed increase in triantennary glycans 36 . The main function of AAT is to protect tissue from enzymatic attacks 41 . One may speculate that an increase in AAT could be attributed to a mechanism where the developing embryo and its supporting structures (placenta and endometrium) are protected from enzymatic attacks by the immune system of the mother. Interestingly, the previously mentioned CGE-LIF study on the same cohort also showed increased levels of triantennary glycans on AAT during pregnancy 12 . Therefore, it is likely that a combination of the increasing concentration of AAT and the increase in triantennary glycans of AAT make a major contribution to the observed increase in triantennary glycans in the TSNG.
The linkage of the sialic acid is of major importance to determine if a sialyl-Lewis epitope is present on a glycan. A sialyl-Lewis X or sialyl-Lewis A epitope requires the presence of both a fucose (α 1,3-linked for sialyl-Lewis X or α 1,4-linked for sialyl-Lewis A) and an α 2,3-linked sialic acid 42 . In diantennary glycans the fucose is mostly present in the core, while in tri-and tetraantennary glycans it is generally on the antenna, resulting in a sialyl-Lewis X or A epitope 6 . However, we did not observe any changes in relation to the percentage of sialyl-Lewis epitopes during and after pregnancy. Our results do show changes in α 2,6-linked sialylation of fucosylated triantennary glycans (A3FE; Fig. 3T). In addition, we found changes in α 2,3-linked sialylation of triantennary glycans (A3L; Fig. 3O), but this could be further narrowed down to changes in α 2,3-linked sialylation of non-fucosylated triantennary glycans (A3F0L; Fig. 3P). Therefore, it is apparent that there are no changes related to sialyl-Lewis epitopes, or that the changes are obscured by changing glycoprotein levels, such as AAT 36 .
Many physiological processes are involved in the postpartum recovery of the mother. Homeostasis needs some time to re-establish and this is reflected by the changes in the serum glycome. Some traits rapidly change (between the 3 rd trimester and 6 WPP) and then stabilise (similar levels at 6 and 26+ WPP), while variations in others emerge only later in time (significant at 26+ WPP; Fig. 4). For IgG it has been shown for RA patients that the glycosylation restores postpartum within the first 6 WPP to a similar level as before the pregnancy 3 . Accordingly, we observed in the current study a similar time course for the galactosylation of diantennary fucosylated glycans. Similarly, also α 2,3-linked sialylation of both di-and triantennary glycans, clearly non-IgG related, are examples of fast responses 32 . However, other traits did not recover within the first 6 WPP but only changed after 26 weeks or more, namely the number of antennae on complex type glycans and α 2,6-linked sialylation of triantennary fucosylated glycans (Fig. 4).
A major advantage of our study design is that patient samples are collected from a longitudinal cohort, thereby reducing the impact of biological variance. Consequently, our results can be attributed purely to the pregnancy-associated changes. However, a minor drawback is represented by isomeric glycan structures that we were unable to resolve. Typical examples include H6N5E2, which could be either a triantennary glycan or a diantennary glycan with a LacNAc repeat, and H5N5, which could be either a truncated triantennary glycan or a bisected diantennary glycan. Consequently, the use of compositional information to assign a derived trait may to a certain extent lead to glycan misclassifications. Therefore, the derived traits are calculated based on the most common structure assigned to a distinct composition 6 . In addition, many glycomic studies deal with age, sex and medication as major confounders, but none of these are applicable here as the study encompassed only healthy women of roughly the same age (32.1 years, SD ± 4.4 years) 4 . However, we did not take into account other confounders such as breastfeeding, as the resulting group sizes would lead to low statistical power. Despite the limitations of our study, the results should be indicative of changes that occur during and after pregnancy, as they agree with previous literature 12, 32 .
In conclusion, we observed changes in galactosylation, sialylation and antennarity of the TSNG during pregnancy. Furthermore, differences were observed in the galactosylation, bisection, fucosylation, sialylation and antennarity of the TSNG after delivery. Lastly, a difference in the recovery speed was observed between α 2,3-and α 2,6-linked sialylation of triantennary glycans. However, the biological relevance of the observed changes in the sialylation of triantennary structures and their linkages is at present unknown and will require further study.
Samples. The sera of 29 healthy Caucasian women (mean age 32.1 years, SD ± 4.4) were obtained at three time points during pregnancy (1 st , 2 nd and 3 rd trimester), and three time points after delivery (6, 12 and 26+ WPP). Sera were collected within the framework of the nationwide (The Netherlands) prospective Pregnancy-associated Amelioration of Rheumatoid Arthritis (PARA) study, designed to investigate the improvement of RA during pregnancy 44 . The study was in compliance with the Helsinki Declaration, informed consent was obtained from Scientific RepoRts | 6:23296 | DOI: 10.1038/srep23296 all subjects, and the study was approved by the Erasmus University Medical Center Ethics Review Board. A total of 174 PARA (6 * 29) serum samples from the healthy control group of the PARA cohort were analysed. Furthermore, 10 replicates of a commercially available pooled plasma sample of healthy volunteers were used as a technical control, yielding an RSD of the main peak (H5N4E2) of 7% ( Supplementary Fig. S4). The use of derived traits reduced the technical variation, for instance the RSD of the derived trait 'galactosylation of all glycans' (AG) was found to be 0.8% for the technical control.
Sample preparation. In order to analyse the TSNG of healthy women during pregnancy, 10 μL serum was mixed with 20 μL 2% SDS, prior to denaturation by 10 min incubation at 60 °C. N-glycans were released by addition of 20 μL release mixture (1:1 4% NP-40: 5x PBS) containing 1 U PNGase F, followed by overnight incubation at 37 °C. Of the 50 μL released glycans, 1 μL was derivatised by ethyl esterification as reported in literature 27 . Briefly, the 1 μL sample was added to 20 μL 250 mM EDC + 250 mM HOBt in pure ethanol and incubated for 1 h at 37 °C. Following derivatisation, the glycans were purified by hydrophilic interaction liquid chromatography (HILIC)-solid phase extraction (SPE) using cotton thread tips, as described previously 32 . The cotton material, packed into pipette tips, was pre-conditioned by pipetting three times 20 μL water, followed by equilibration with three times 20 μL 85% ACN. The samples were mixed with 20 μL ACN and then pipetted through the pre-equilibrated cotton material extensively. The tips were sequentially washed three times with 20 μL 85% ACN + 1% trifluoroacetic acid and three times with 85% ACN. Glycans were eluted in 10 μL water. Afterwards, 1 μL of the TSNG SPE eluate was applied to an AnchorChip MALDI target plate, mixed with 1 μL of 5 mg/mL super-DHB in 50% aqueous ACN containing 1 mM NaOH, and dried at room temperature. Data acquisition. Derivatised glycan samples were measured using the reflectron positive ion mode of an Ultraflextreme MALDI-TOF-MS (Bruker Daltonics), equipped with a Smartbeam-II laser, and operated by flex-Control 3.4 build 135. Prior to the measurement, external calibration was performed using a peptide calibration standard. Ions were accelerated at 25 kV with 140 ns delayed extraction. Per sum spectrum, a total of 20,000 shots were accumulated between m/z 1,000 and m/z 5,000.
MassyTools data processing. Using the flexAnalysis Batch Process program (3.3, build 65), the obtained spectra were exported into a text-based format ((x,y) file) consisting of an m/z value coupled with an intensity value per line, separated by a tab. In order to perform processing by MassyTools; software freely available at https://github.com/Tarskin/MassyTools), we applied the 'free' reducing end mass modifier and 'sodium' charge carrier as general parameters 31 . Additionally, a single letter code was utilized by the program to refer to monosaccharides: H for hexose, N for N-acetylhexosamine, F for fucose, E for an α 2,6-linked N-acetylneuraminic acid and L for an α 2,3-linked N-acetylneuraminic acid.
Internal calibration was performed using a minimum of 5 out of a list of 7 potential calibrants passing a signal-to-noise (S/N) ratio threshold of 9. Potential calibrants were H5N4E1 (1,982.708 Da), H5N4F1E1 (2,128.766 Da), H5N4E2 (2,301.835 Da), H6N5E2L1 (2,940.052 Da), H6N5F1E2L1 (3086.110 Da), H7N6E1L3 (3,532.227 Da) and H7N6E2L2 (3,578.269 Da; [M + Na] + throughout). Additionally, we required a minimum of 4 calibrants between m/z 1,000 and m/z 2,333 as well as a minimum of 1 calibrant between m/z 2,333 and m/z 3,667. The program was allowed to detect calibrants within a window of m/z 0.4 around the exact mass. The calibration was assessed by evaluating the mass error of the major calibrant glycan (H5N4E2) and major non-calibrant glycan (H5N4E1L1) for the technical control samples. H5N4E2 showed an average mass error of − 0.3 ppm, SD ± 1.9 ppm while H5N4E1L1 showed an average mass error of − 3.1 ppm, SD ± 3.0 ppm.
Extraction was performed using a ± m/z 0.49 window. A list of 109 analytes was used for initial extraction (Supplementary Table S1). The software was allowed to use an m/z region of 15 around the analyte to identify the background area and noise values. Subsequently, spectral and analyte curation was performed. Firstly, the "fraction of spectrum in analytes" and "fraction of analyte area above S/N cut-off " parameters were used for spectral curation. The spectral quality criteria of all spectra were found to be within three times SD of the average, therefore all spectra were kept for further analysis. Secondly, a quality control (QC) parameter and a S/N cut-off were used to curate the analyte list: analytes with an average QC value above 1.8 × 10 −4 or an average S/N of below 6 were discarded, while analytes with an average QC value below 1.8 × 10 −4 and an average S/N of above 6 ( Supplementary Fig. S2) were retained. A total of 77 analytes passed these quality criteria and were used for further analysis.
Derived traits calculation & statistical analysis. Glycans were grouped according to their compositional features. All derived traits are based on single enzymatic steps (e.g. AG for the attachment of a galactose to an N-acetylhexosamine). The total set of 42 calculated derived traits can be found in the Supplementary Table S2. Furthermore, statistical analysis was performed using IBM SPSS Statistics 20 (IBM, New York, USA). All derived traits were tested for normality using a Shapiro-Wilk test 45 . As part of the derived traits were not normally distributed, a Wilcoxon signed-rank test was used to test pairwise differences observed in the cohort 46 . All 42 traits were tested by comparing the 1 st trimester with the 3 rd trimester (pregnancy), the 3 rd trimester with 6 WPP (short term recovery) and, lastly, the 3 rd trimester with 26+ WPP (long term recovery). Bonferroni correction was used to address the multiple testing problem, by adjusting the significance threshold (significance level α < 0.0012 for 42 tests per time point) 33 .