A Perspective on the Confident Comparison of Glycoprotein Site-Specific Glycosylation in Sample Cohorts

Abstract

Protein glycosylation, resulting from glycosyl transferase reactions under complex control in the secretory pathway, consists of a distribution of related glycoforms at each glycosylation site. Because the biosynthetic substrate concentration and transport rates depend on architecture and other aspects of cellular phenotypes, site-specific glycosylation cannot be predicted accurately from genomic, transcriptomic, or proteomic information. Rather, it is necessary to quantify glycosylation at each protein site and how this changes among a sample cohort to provide information about disease mechanisms. At present, mature mass spectrometry-based methods allow for qualitative assignment of the glycan composition and glycosylation site of singly glycosylated proteolytic peptides. To make such quantitative comparisons, it is necessary to sample the glycosylation distribution with sufficient coverage and accuracy for confident assessment of the glycosylation changes that occur in the biological cohort. In this Perspective, we discuss the unmet needs for mass spectrometry acquisition methods and bioinformatics for the confident comparison of protein site-specific glycosylation among sample cohorts.

Graphical Abstract

Approximately 30% of the vertebrate transcriptome passes through the secretory pathway, the majority of which becomes glycosylated.¹ Glycoproteins located on the plasma membrane, in the pericellular matrix and extracellular matrix, define the cellular antigenicity and the microenvironments in which vertebrate biology occurs. Numerous protein–glycan binding interactions govern multicellular phenotypes and host–host and host–pathogen interactions.² Micro- and macroheterogeneous protein and lipid glycosylation enables long-lived multicellular species to survive in the presence of rapidly evolving pathogens.³ In addition, glycoproteins display accelerated evolution during vertebrate radiation, driven by the need to elaborate the protein interaction properties of plasma membranes and extracellular matrices.⁴ Therefore, to understand the roles of glycoproteins in the numerous disease mechanisms that depend on glycan–protein binding, it is necessary to quantify the distribution of glycosylation that resides at each glycosite. Beyond this, it is necessary to define the molecular populations that tune the functional properties of glycoproteins that exist in a given biological context. Viewed simply, the theoretical numbers of such glycosylated proteoforms are multiples of the number of glycans that exist at each protein site. However, for proteins with several sites of glycosylation, this results in theoretical numbers of glycoprotein structural variants that are too large to exist biologically.⁵ It is therefore important to define the subset of such glycosylated proteoforms that exist among biological cohorts.

We address the challenge of comparing how glycoprotein proteoform populations differ among sample cohorts in this Perspective. Mass spectrometry-based methods can now identify the glycan composition and glycosylation site for singly glycosylated peptides. Methods for acquiring MS data to quantify site-specific glycosylation with sufficient confidence for identification of changes that occur during disease mechanisms are now emerging.

SUMMARY OF LC-TANDEM MS METHODS FOR ASSIGNMENT OF SITE-SPECIFIC GLYCOSYLATION

Mass spectrometry-based omics studies rely on statistical analysis to identify the most confident data assignment. This requires careful consideration of rates of true and false identifications and sources of experimental error. The glycoproteome is orders of magnitude more complex than the proteome, owing to the non-template-driven glycan biosynthetic pathways. Given this complexity, what is the limit of sample complexity that can reasonably be analyzed with the present technology? We summarize the state of the glycoproteomics field with this question in mind.

Discovery-based bottom-up proteomics often uses data-dependent acquisition (DDA) whereby ions are selected automatically during data acquisition based on abundance for subsequent tandem MS acquisition. In such experiments, a finite instrument acquisition speed limits quantitative reproducibility due to the stochastic nature of the ion selection. Proteomics database searching can be used to identify small post-translational modifications, such as those that occur most often in relation to intracellular processes, including phosphorylation, acetylation, methylation, and acylation.⁶ It is not well suited, however, to analysis of glycosylated peptides for a few reasons. The glycan undergoes preferential collisional dissociation, resulting in the need for specialized MS acquisition methods, including the use of collision energy ranges higher than those for unmodified peptides of a similar amino acid composition. For complex glycans, the glycan size is often >1000 Da and it is heterogeneous. Thus, glycosylation heterogeneity divides the total ion signal, necessitating the enrichment of the glycosylated molecules. Even with enrichment, the high number of glycoforms at a given site (often >30) that elute within a small retention time window during an LC-MS experiment makes it impossible to select all precursors for tandem MS quantification of the complete glycopeptide distribution. Such heterogeneity also limits the number of precursor ions that can be quantified using tandem MS inclusion lists. Therefore, some of the glycopeptide glycoforms cannot be identified unambiguously in the absence of a tandem MS scan.

During a tandem MS event, glycopeptides dissociate to form low-molecular weight oxonium ions, high-molecular weight ions from loss of saccharide units from the precursor ions (here termed peptide + Y ions), and peptide backbone product ions (Figure 1).

Figure 1.

Types of glycopeptide product ions produced by collisional dissociation. Oxonium ions are colored green, peptide b_n ions blue, y_n ions red, and peptide + Y ions goldenrod. In the fragmentation diagram, the glycosylated peptide residue is colored green. Peptide y_n ions that are glycosylated are colored green.

The relative abundances of these product ions depend on the analyzer and dissociation method. For example, with beam-type collisional dissociation, including so-called higher-energy dissociation (HCD), peptide backbone product ions are observed in relatively low abundances provided that sufficient collision energy is used.^7–9 When resonant collisional excitation is used, for example with ion traps, product ions formed from the precursor are cooled rapidly and peptide backbone dissociation product ions may not be observed. For electron transfer dissociation-based methods, it is typically necessary to use supplemental vibrational excitation to mitigate nondissociative electron transfer, whereby product ions are held together by noncovalent forces in the gas phase. Examples are electron transfer higher-energy collisional dissociation (EThcD)^10–12 and activated ion electron transfer dissociation (AI-ETD).^13,14 Thus, although ETD cleaves primarily the peptide backbone, there is a degree of dissociation of the glycan from vibrational excitation when using it for analysis of glycopeptides.

The presence of oxonium ions is diagnostic for a glycosylated peptide. The abundances of peptide + Y ions depend on the collision energy. If the collision energy is relatively low, then ions resulting from loss of one or a few saccharide units will be abundant. If the collision energy is relatively high, then peptide + Y ions corresponding to peptide plus mono- to trisaccharides may be abundant.^15,16 The tandem mass spectra that result in the most confident glycopeptide identifications include all three types of product ions, being able to accurately sequence the peptide and characterize the glycan composition. To produce such product ion patterns for the highest percentage of precursor ions, it is necessary to use a stepped collision energy.^7,9,17,18 The obvious disadvantage to this approach is that it reduces the number of precursor ions that can be analyzed in a given time window.

For glycopeptides, the most confident assignments come from tandem mass spectra with product ions with resolved isotopic clusters due to the propensity for product ions to be multiply charged. Such high-resolution LC-tandem mass spectral data can be acquired using QTOF or Orbitrap analyzers. Note that many proteomics data sets acquired using ion trap–Orbitrap hybrid instruments employ the low-resolution ion trap for product ion mass spectra. Such low-resolution tandem mass spectra yield low-confidence glycopeptide assignments because the product ion compositions cannot be assigned from a list of theoretical structures with high accuracy.^{15,16,19–21} The advantage of high-resolution glycopeptide LC-tandem MS analysis is that the same chromatography columns, mobile phases, and collisional tandem MS instruments that were used widely in proteomics laboratories can be used for glycopeptides.

Activated electron dissociation results in preferential dissociation of the peptide backbone with minimal dissociation to glycopeptide glycans.²² Electron transfer dissociation has undergone several cycles of engineering improvements on commercial mass spectrometry systems and has been applied to biological glycoscience studies.^10,11,23,24 The advantage to ETD is that it is capable of producing peptide backbone dissociation for larger peptides than for collisional dissociation. Some commercial instruments are capable of both collisional and electron transfer dissociation in the same LC-MS run. Electron transfer dissociation requires considerably lower scan frequencies than does collisional dissociation, with the result that ETD should be used in LC-MS workflows, for which efficient use of analyzer time is paramount, only for precursor ions that do not dissociate effectively using HCD. Newer instruments can select among activation methods and parameters based upon precursor mass and charge state, which can be used to perform ETD only on sufficiently large and charged precursors for a subset of samples with multiply glycosylated peptides.²⁵ Not all multiply glycosylated peptides are substantially larger than the singly glycosylated peptides in a sample, particularly when considering O-glycopeptides, which can make it difficult to determine which precursors to dissociate with ETD on-the-fly for such samples.

Collisional dissociation typically dissociates the glycan and therefore does not distinguish sites of glycosylation on peptides with multiple glycosylation sites. ETD methods are more suited for this purpose. One disadvantage to ETD is that positive charge states are diminished by the addition of an electron, with the result that small glycopeptides may not carry sufficient charge for efficient dissociation. Another problem is that ETD products may be observed as gas phase pairs with m/z values equivalent to the loss of charge from the precursor, a phenomenon known as ETnoD. Dissociating such ions requires a supplemental dissociation step using collisional dissociation, known as electron transfer high-energy dissociation (EThcD), a method that is available on a commercial instrument. Another approach requires supplemental excitation with an infrared laser, which requires a customized instrument.¹³

QUANTIFICATION OF SITE-SPECIFIC GLYCOSYLATION

Multiple-reaction monitoring (MRM), the most sensitive method for quantification of compounds using mass spectrometry, requires use of stable isotope-labeled internal standards for absolute quantification. For best practices in proteomics, three tiers of targeted quantitative assays have been defined.²⁶ For exploratory studies (tier 3), no reference standards are used and the precision is similar to that of label free discovery proteomics. For research use quantification (tier 2), limited reference standards are used and a moderately high precision is achieved with a 20–35% coefficient of variation. For clinical bioanalysis (tier 1), reference standards are used for every peptide quantified and a high precision is achieved (<20–25% CV). For glycopeptides, there are few such standards available. Thus, MRM methods generally produce relative quantification that is useful for following biological changes to glycosylation.²⁷ Mechref et al. introduced the approach of using oxonium ion MRM transitions for glycopeptides, simplifying the task of identifying transitions at the possible expense of specificity.²⁸ Goldman et al. used MRM to quantify plasma haptoglobin glycopeptides related to liver disease.^29,30 These authors purified the haptoglobin to alleviate concerns about the specificity of the MRM transitions in a complex glycoprotein background. They also used neuraminidases to limit the glycan heterogeneity. Lebrilla et al. have used MRM to quantify 64 immunoglobulin glycopeptides from serum without enrichment.³¹

Parallel reaction monitoring (PRM) has the advantage that a full-range tandem mass spectrum is acquired for every isolated precursor ion, allowing users to optimize the specificity of the assay. Larsson et al. quantified an amyloid β-glycopeptide modified with a core 1 (Hex-HexNAc) disaccharide using a glycopeptide standard synthesized for this purpose.³² Yoo et al. used a glycopeptide PRM method to determine extent of fucosylation of α-fetoprotein immunoprecipitated from human serum as a biomarker for hepatocellular carcinoma.³³ These authors extended this approach to analyze glycopeptides from several glycoproteins enriched using multiplexed immunoprecipitations.³⁴

While data-independent acquisition (DIA) allows tandem MS of precursor ions without the need for ion isolation, spectral libraries are used widely for interpretation of overlapping patterns of precursor and product ions in DIA data sets.³⁵ Spectral clustering³⁶ and spectral library searching use experimentally collected mass spectra identified by a separate mass spectrometry experiment.³⁷ This method works well for peptides, and for small molecules where there are no canonical fragmentation pathways that can be generalized to all molecules.³⁸ It has also been used for glycopeptides in GPQuest³⁹ and SweetNET.⁴⁰ Spectral libraries lack portability, as the library spectra are specific to the collision energies used during acquisition. Thus, this approach becomes less sensitive as the acquisition conditions differ from those from the original library construction. This is particularly important for glycopeptides, for which the relative abundances of oxonium ions, peptide + glycan ions, and peptide backbone dissociation ions vary considerably as collision energy is increased. Fortunately, the use of stepped collision energies helps mitigate this variability.⁹ Spectral libraries have seen extensions for clustering both identified and unidentified spectra⁴¹ and for MRM identification and quantification with SWATH.⁴²

Tandem MS results in similar monosaccharide residue losses and oxonium ions for the series glycoforms present for a given glycopeptide. Given this, the extent to which DIA will generate acceptable specificity for quantification of such series of glycopeptide glycoforms is not clear. DIA methods have been used for analysis of immunoglobulin G glycoforms, for which there is very limited glycan heterogeneity due to steric constraints.⁴³ Ye et al. also showed DIA with in silico-augmented libraries may be tractable for sufficiently small glycans with limited heterogeneity, as found for common smaller O-GalNAc glycans, though requiring additional metrics to calibrate identification confidence.⁴⁴ Molloy et al. compared PRM and DIA for analysis of targeted glycopeptides from blood plasma.⁴⁵ They reported that PRM was effective for quantification but limited by duty cycle. They analyzed DIA results using a theoretical targeted glycopeptide database rather than a spectral library. While this quantified more glycopeptides than with the PRM method, the specificity whereby targeted glycopeptides can be identified in the presence of complex and uncharacterized plasma glycoproteins was not defined. There is a general challenge in achieving appropriate fragmentation for molecules spanning several neutral mass ranges falling within the same wide isolation window in DIA, where a single collision energy is applied to the entire window.^46,47 This incomplete fragmentation exacerbates the problem discussed earlier in acquiring high-quality assignments.

While DIA and targeted methods borrowed from proteomics may be applicable to glycoproteomics samples, they still suffer from the increased complexity of the sample, and the greater redundancy among product ions of related glycoforms. Chromatogram libraries are particularly vulnerable to corruption by co-isolation of adjacent glycoforms overlapping in elution time, and the problem is resistant to the solution for small modifications.⁴⁸ However, it has been shown that relative elution time shifts are induced by differences in monosaccharide composition. Recently, tools have been developed to exploit these relationships^49–52 for estimating confident identifications of lower-quality structural assignments relying on only precursor mass and apex elution time. This method shows potential for overcoming some of the limitations of DIA for glycopeptides.

In summary, DIA has the potential to improve the ability to quantify targeted peptides, but its specificity for quantification of targeted glycopeptides in complex biological samples with or without spectral libraries remains undefined.

QUANTIFICATION OF GLYCOPEPTIDES USING ISOBARIC LABELS

Lee et al. demonstrated that TMT-labeled N-linked glycopeptides produce HCD reporter ions.⁵³ They used this approach to quantify vitronectin N-linked glycopeptides as candidate markers for hepatocellular carcinoma and identified glycoforms with significantly increased levels in diseased plasma. In further work, this group used their approach for simultaneous qualitative and quantitative comparison of site-specific N-glycopeptide glycoforms.⁵⁴ Zhang et al. used TMT tagging to quantify fucosylated glycopeptides enriched using lectin affinity chromatography as part of a study of changes in glycosylation associated with aggressive prostate cancer cell lines.⁵⁵ For these studies, proteins were extracted from cells, digested with trypsin, enriched using lectin chromatography, labeled with TMT, and then analyzed using LC-HCD-MS. This group also developed a workflow for simultaneous identification of glycoand phosphopeptides co-isolated from breast tumor patient-derived xenografts.⁵⁶ Palmisano et al. mapped N- and O-glycoproteins from urine to discriminate prostate cancer from benign prostatic hyperplasia.⁵⁷ For these studies, the total extracted proteins were digested and labeled with TMT10plex. Glycopeptides were enriched using a combination of hydro-philic interaction and titanium dioxide solid phase extraction and quantified using LC-HCD-MS.

Viner et al. demonstrated that ETD produces four distinct isobaric product ions when using the TMT-6 plex amino-specific label.⁵⁸ They then demonstrated quantification of OGalNAc-modified glycopeptides from crystallin protein tryptic digests based on the ETD-produced reporter ions. Such peptides undergo dissociation of the O-GalNAc residue during HCD, making it impossible to quantify with confident identification of the glycosylation site. Ye et al. compared glycosylation among 10 lots of human chorionic gonadotropin drug products using a combination of HCD and ETD.⁵⁹ The protein tryptic digests were labeled with TMT10plex using high-resolution LC-MS with alternating HCD and ETD methods. They used partial least-squares discriminant analysis to identify the glycopeptides that differed among the protein lots.

THE QUESTION OF APPROPRIATE SAMPLE COMPLEXITY IN GLYCOPROTEOMICS: HOW COMPLEX IS TOO COMPLEX?

As shown in Figure 2, assignment of glycopeptides from tandem mass spectra entails defining a search space consisting of the theoretical protein sequence, the peptides generated therefrom, and theoretical glycan compositions.^15,16 The most rigorous approach is to measure the proteome and glycome of the sample and use the data to build a glycopeptide search space. A typical glycoprotein glycosite has ≥30 glycoforms, vastly increasing the complexity of the search space relative to that of a proteome with no glycosylation sites. While this effect is appreciated generally, there is no consensus regarding the use of confidence measures to ensure that glycopeptide assignments are correct in a given biological background. We demonstrated that the ability to assign glycopeptides confidently decreases as the sample complexity increases.¹⁹ We suggest that investigators spike in a well-characterized glycoprotein from a different species to complex biological samples from which glycopeptides are assigned. In doing so, they will be able to determine the extent to which assignment confidence deteriorates for the well-characterized glycopeptides in the presence of a complex background relative to the pure glycoprotein.

Figure 2.

Schematic diagram of a glycoproteomics search engine, including inputs, outputs, and intermediate steps.

Software for identification of glycopeptides can be broken into the following categories: peptide-centric, glycan-centric, combined, and complete. Peptide-centric⁶⁰ methods focus on identifying peptide backbone sequences. While they may use peptide + Y ions, they do not control for the FDR of the glycan component. Glycan-centric methods^61,62 focus on identifying the attached glycan, without sequencing the peptide. A combined method^51,63,64 uses a single score to capture both peptide and glycan components and controls the total uncertainty but does not control the uncertainty of both components separately. A complete method^18,65,66 works by controlling the uncertainty of both components separately and jointly. Though one might desire to use a complete method for any experimental data, it requires that the experiment produces the appropriate fragmentation. Complete methods usually require relatively complex fragmentation reactions, in some cases spanning multiple spectra, limiting their throughput.

Some data-dependent tandem MS methods^51,63,64,67 look at the oxonium ions in a tandem mass spectrum to constrain which glycans could realistically be attached to a given peptide. These tools can be used on most glycoproteomics data as oxonium ions are expected to be present in all collisional dissociation spectra, and they provide orthogonal signature ions to the large intact peptide + Y ions for assigning glycan compositions. They are vulnerable to contamination caused by co-isolation, such as when a glycan with a sialic acid is co-isolated with an unsialylated glycan. The prevalence of co-isolation is expected to be a function of sample complexity. While there is value in using these signals for the determination of the class (N, O, or GAG linker) and subclass (e.g., NeuAc-bearing, NeuGc-bearing, high-mannose N-glycans), isolation purity has yet to be addressed either experimentally or algorithmically.

BIOINFORMATICS DEVELOPMENTS NEEDED FOR QUANTITATIVE GLYCOPROTEOMICS

In an LC-MS experiment, the limited number of precursor ions that can be analyzed in a time window and the stochastic nature of precursor ion isolation result in missing values that limit the quantitative value of DDA methods.⁶⁸ Likewise, there are a limited number of ions that can be quantified using a tandem MS inclusion list. Until recently,^63,64,67 most software for glycopeptide identification rarely included built-in support for label free quantification (LFQ), and when they did, the features of these methods were not always complete. Byonic can integrate glycopeptide identification and quantification because of the underlying Byologic platform for LFQ of normal proteomics. The glyXtool was built on top of OpenMS,⁶⁹ which includes a collection of different LFQ implementations and support tools for match between runs (MBR), feature finding and alignment, and integration, providing a bridge between glycopeptide representation within their application code and OpenMS’s glyco-agnostic algorithms. These techniques can help to diminish the effect of missing values caused by stochastic precursor selection. While these steps move toward streamlining quantification of site-specific glycosylation, we argue that unsolved problems remain.

The most popular approach to reducing missing values, MBR, requires that an analyte be fragmented and identified at least once in a study to transfer that identification to other acquisitions in that study.⁷⁰ Recent work has shown that transferred identifications that were identified in only a small fraction of the acquisitions were more likely to be false positives.⁷¹ Integration of techniques for validating glycopeptide identifications by glycan-induced retention time shifts discussed earlier would be expected to reduce these false transferred identifications. When an analyte is never fragmented across any data acquisition, a larger issue for smaller studies, we may still be able to infer its identity from its mass and retention time offsets from other identified glycopeptides. While this may be done in an ad hoc manner for manual annotation, a high-throughput approach that incorporates some amount of statistical rigor has not been developed.

Quantitative targeted proteomics software for building acquisition lists and analyzing the results is designed for unmodified peptides or those that have small, well-defined, modifying groups.⁴² They do not at present address the problem of defining large, heterogeneous glycan post-translational modifications for peptides. As a result, it is necessary for users to construct long precursor target lists and MRM/PRM transitions using time-consuming manual effort or approximate in silico methods.^44,72 There is a need to adapt existing tools and algorithms from quantitative proteomics to account for heterogeneous glycosylation as a post-translational modification. Part of the problem is that formats for defining and representing small molecule protein modifications differ from those needed to define the distribution of glycoforms that exist at a given protein site. This effort is made challenging by the lack of proteomics data exchange formats that can be applied to glycosylation (see the discussion below). Even if the existing tools were used to encode such large modifications, they would not be able to accurately account for the non-peptide-derived fragments unless the user manually entered an exhaustive list.

There is a clear, unmet, need to communicate glycoproteomics results to data repositories and knowledge bases, as well as between software. This is a challenge due to the extreme complexity of the domain, with an effectively intractable number of biosynthetically feasible glycan structures and the ways in which they may be represented.⁷³ In proteomics, the mzIdentML data standard^74–76 was developed by the Proteome Standards Initiative in collaboration with instrument and software vendors to facilitate storage and transfer of proteomics identification data. This standard provides a stable vendor-independent means for data exchange that has vastly improved the dissemination of proteomics data, and a growing ecosystem inter-operating software. Skyline can read mzIdentML files produced by compliant identification software to understand which peptides were identified when in which samples when building targets or transition lists.⁷⁷ OpenMS similarly can read and write mzIdentML, mzQuantML, and mzTab files that, as previously mentioned, lets pipeline builders weave together tools from different authors, where each piece of software needs to know how to do only one thing rather than everything.⁷⁸ PRIDE can automatically infer which scans were used to identify which peptides and which raw data files they were taken from by inspecting an mzIdentML file, automating the project submission.⁷⁹ PRIDE can also generate annotated spectra from such submissions, obviating the requirement for authors to submit manually annotated spectra with articles.

The mzIdentML format was designed for representing peptides and a user-specified number of modifications such as those defined in the Unimod database,⁸⁰ or free-text descriptions. As we have emphasized here, glycans at a given protein site exist as a set of related glycoforms, the ontologies for representing which need to be incorporated into an exchange format. In the absence of such an exchange format, each piece of glycoproteomics software must create a method for encoding glycans, making it difficult to compare or combine their results. As a result, exchange of glycoproteomics data remains a serious barrier to incorporation of glycopro teomics results into repositories and knowledge bases. Such an exchange format should use an externally hosted glycan database, such as GlyTouCan,^81,82 represented in Open Biomedical Ontology (OBO) format, or another acceptable exchange format. An example where this has begun to happen is in GlyConnect, a database of glycoproteins based on manual curation from published research and high-throughput studies.⁸³ It still relies on those high-throughput studies being able to provide enough information to connect the glycan notation used by their software to the glycan represented in GlyTouCan. However, GlyConnect is linked to UniProt⁸⁴ and NeXtProt^85,86 knowledge bases that are used by experts in many branches of biology, not just glycobiology. Through this connection, these Web sites now communicate to their users that the function of the protein that they are browsing is not one uniform molecule but dozens to hundreds of different structures inpotentia depending upon biological conditions.

To create an exchange format, the field must reach some consensus on the level of detail with which to represent glycans. While it might be ideal to specify explicitly all structures exactly,⁸⁷ including stereoisomeric identities and linkage, current experimental and computational methods are incapable of reliably and uniformly discriminating these properties, such as separating GlcNAc and GalNAc or 2–3-and 2–6-linked NeuAc. Should it still be an option to explicitly specify these properties, does this mandate that any consumer of glycoproteomics results must be able to recognize the subsumption relationship⁸⁷ between GlcNAc2Man6 and HexNAc2 Hex6? Both could be valid identifications of the same experimental evidence, and without the ability to recognize that one is just a specialization of the other. If subsumption is a feature of the standard, then the cost to implement the standard grows considerably, but leaving it off will lead to noncomparable results. A fair middle ground might be to define the standard with the freedom to be specific, but to recommend that generic names be used wherever appropriate, and for subsumption resolution to be required at the monosaccharide level but not at the aggregate level.

CONCLUSIONS

Tandem mass spectrometry methods have advanced to the point that singly glycosylated peptides can be assigned in LC-MS workflows using software available in the commercial and academic domains. The most appropriate use of these methods is for characterization of glycoproteins from enriched samples. This is because the extreme heterogeneity of glycoproteins from complex biological samples limits the ability to understand sources of error in glycopeptide assignments. Fortunately, the field is moving rapidly toward defining operating procedures and data exchange formats to allow widespread dissemination of accurate and reproducible glycoproteomics data.