Structural glycobiology: A game of snakes and ladders

Abstract

Oligo- and polysaccharides are infamous for being extremely flexible molecules, populating a series of well-defined rotational isomeric states under physiological conditions. Characterization of this heterogeneous conformational ensemble has been a major obstacle impeding high-resolution structure determination of carbohydrates and acting as a bottleneck in the effort to understand the relationship between the carbohydrate structure and function. This challenge has compelled the field to develop and apply theoretical and experimental methods that can explore conformational ensembles by both capturing and deconvoluting the structural and dynamic properties of carbohydrates. This review focuses on computational approaches that have been successfully used in combination with experiment to detail the three-dimensional structure of carbohydrates in a solution and in a complex with proteins. In addition, emerging experimental techniques for three-dimensional structural characterization of carbohydrate–protein complexes and future challenges in the field of structural glycobiology are discussed. The review is divided into five sections: (1) The complexity and plasticity of carbohydrates, (2) Predicting carbohydrate–protein interactions, (3) Calculating relative and absolute binding free energies for carbohydrate–protein complexes, (4) Emerging and evolving techniques for experimental characterization of carbohydrate–protein structures, and (5) Current challenges in structural glycoscience.

The complexity and plasticity of carbohydrates

Carbohydrates occupy a pivotal functional position in biological recognition processes (Sharon and Lis 1993; Varki 1993; Dwek 1996). The complex shape, functionality, and dynamic properties of oligo- and polysaccharides (hereafter denoted simply as “carbohydrates”) allow these molecules to function in intermolecular interactions as encoders of biological information. For instance, carbohydrate recognition is an integral part of normal biological development (Haltiwanger and Lowe 2004) and the immune defense against pathogens via the identification of exogenous carbohydrates (Brown and Gordon 2001; Cobb and Kasper 2005). Conversely, many bacterial and viral pathogens (such as Escherichia coli or Haemophilus influenza) initially adhere to host tissues by binding specifically to carbohydrates on the host’s cell surfaces (Karlsson 1986, 1989; Rostand and Esko 1997). Thus, there is an interest in developing therapeutic agents that can interfere with, modulate, or exploit carbohydrate-based host-pathogen interactions. Examples of therapeutic agents interfering with carbohydrate-specific interactions include the neuraminidase inhibitors zanamivir and oseltamivir used in the treatment of influenza infections (Dreitlein et al. 2001). Vaccines that employ bacterial polysaccharides, conjugated to carrier proteins, have been particularly effective, for example against H. influenza (Jennings 1992). Because abnormal glycosylation is also a marker for certain types of cancer (Hakomori 1989; Fukuda 1996) and other diseases, such as IgA nephropathy (Coppo and Amore 2004; Moura et al. 2004), inflammatory bowel disease (Campbell et al. 2001), and rheumatoid arthritis (Parekh et al. 1985; Malhotra et al. 1995), there is a growing interest in exploiting these variations in the development of therapeutics (Lo-Man et al. 2004; Buskas et al. 2005; Xu et al. 2005). In certain diseases, such as congenital disorders of glycosylation (Freeze 2001) or lysosomal storage diseases (Neufeld 1991), the origin of the observed glycosylation defects can be traced back to mutations in the glycan-processing pathway, suggesting a role for gene therapy and possibly glycosidase/transferase inhibition (Platt et al. 1994; Sly and Vogler 2002; Grabowski and Hopkin 2003).

Thus far, only rarely has the design of carbohydrate-based therapeutic agents made extensive use of 3D structural information, reflecting in part the difficulties of determining carbohydrate conformation, as well as a paucity of structural data for many carbohydrate—protein complexes. To help reverse this trend, computational approaches have emerged to complement experimental techniques in the analysis of structure–function relationships of carbohydrate–protein interactions.

A significant challenge in the characterization of the conformational properties of carbohydrates is that they are flexible, populating multiple (defined) conformational states under physiological conditions. This property necessitates a modification in the way we think about biological recognition processes. A rigid molecule can be fully characterized by a single conformational state, but not so for a flexible one. This raises an interesting question: How are flexible molecules recognized in nature? Does the receptor protein preferentially bind to the most frequently populated shape, or to the average shape, or to a relatively rare “bioactive” conformation, or does binding induce a unique conformation?

To help explore the concepts of the carbohydrate structure and recognition, let us compare carbohydrates to another flexible object, a snake. To the extent that a living snake is a flexible 3D object that is not random in its motional properties, it serves as a useful analogy for carbohydrate structure and recognition. The shape and motion (as well as color and sound) of a cobra are clearly distinct from those of a rattlesnake. Both are generally long, skinny, and wiggly, but each is recognizably different. Yet, the average shape of each snake would be remarkably similar; if each one were to wiggle to an equal extent to the right and left its average shape would be a straight line! So it goes with all flexible objects, including glycans; depending on the extent of the motion the average shape may be a very poor description of any instantaneous conformation. That is not to say that the average properties are not useful; most experimentally observable data are averages of a conformational ensemble. For example, NMR intensities are the average of contributions from all of the conformational states observed on the NMR timescale. This averaging means that NMR data must be used with care when deriving a 3D model for a flexible carbohydrate, as the data could point to a virtual conformation. Nevertheless, the NMR data are extremely important in characterizing the carbohydrate structure and dynamics, and for validating computational predictions (Sayers and Prestegard 2000; Kirschner and Woods 2001; Sayers and Prestegard 2002; Gonzalez-Outeirino et al. 2005, 2006).

If the average shape is not what is recognized, then a particular conformation must somehow be selected from the conformational ensemble. From a statistical perspective, recognition of the most frequently seen shape of the carbohydrate is more probable than recognition of a rarely populated conformation. Surveys of the Protein Data Bank (PDB) have found that in the majority of noncovalently associated carbohydrate–protein complexes (Petrescu et al. 1999; Imberty and Perez 2000) and glycoproteins (Petrescu et al. 1999), the glycosidic torsion angles were consistent with those displayed most commonly by glycans in a solution. Until recently, these types of statistical analyses were challenging to perform due to limited search software; however, a particularly convenient interface for this purpose (GlyTorsion) has been developed (Lütteke et al. 2005). Exceptions to the tendency to bind low energy conformations are found in carbohydrate-processing enzymes, in which ligand distortion in the active site may be integral to the enzyme’s function (Karaveg and Moremen 2005). Additionally, an induced fit in the carbohydrate has been observed in lectin binding (Casset et al. 1995; Imberty and Perez 2000). However, the induced fit may be the result of miss-matching a receptor with a ligand; that is, just because a protein is found to bind to a particular carbohydrate does not mean that it evolved to recognize that ligand. A miss-matched induced fit is potentially relevant to structures of plant lectins bound to mammalian glycans. These issues underlie challenges faced by computational predictions of carbohydrate epitopes.

Characterizing the extent of molecular flexibility is challenging for both experimental and theoretical methods. To include molecular flexibility accurately in a computational model requires a method that is able to generate states whose average properties are experimentally consistent. Current molecular dynamics (MD) simulations, when combined with force fields appropriate for carbohydrates, provide this capability (Vliegenhart and Woods 2006). A perpetual challenge to MD simulations of biomolecules is to achieve adequate sampling of molecular motions of the system under study. Fortunately, in the case of oligosaccharides, there are relatively few energetically accessible conformational states for each glycosidic linkage, and thus it is often possible to achieve good conformational sampling from an MD simulation of between 10 and 1000 ns, depending on the size and complexity of the oligosaccharide. Sampling efficiency may be enhanced by techniques such as replica exchange MD (Sugita and Okamoto 1999) and Monte Carlo (MC) simulations (Metropolis and Ulam 1949; Metropolis et al. 1953).

MD simulations can also provide insight into the dynamic properties of carbohydrate–protein complexes, given a reasonable initial structure for the complex. Crystallography continues to play a significant role in the characterization of carbohydrate–protein complexes; however, important new rungs have been added to the technological ladder. Most notably, advances in the a priori prediction of complexes by molecular docking have been made as well as NMR methods, such as saturation transfer difference (STD) NMR (Mayer and Meyer 2001; Meyer and Peters 2003). STD NMR experiments facilitate the identification of those regions of the carbohydrate that are proximal to the protein surface (the carbohydrate epitope). Thus, docking experiments may be validated by a qualitative comparison to STD NMR data (Haselhorst et al. 2004, 2007), or by quantitative computation of the NMR intensities in the complex (Wen et al. 2005). Integrating STD NMR data directly into docking algorithms is a potential next step. Emerging experiments included protein surface footprinting, based on either H/D exchange differences (King et al. 2002; Seyfried et al. 2007), or hydroxyl radical oxidation (Sharp et al. 2003, 2004; Hambly and Gross 2005; Takamoto and Chance 2006) that can identify carbohydrate interfaces on proteins, and as such may also provide a significant step up the technological ladder.

Here we summarize some recent reports that employ computational simulations to examine the 3D structures and dynamics of carbohydrates and carbohydrate–protein complexes. Our goal is to help promote awareness of the capabilities and limitations of current computational methods as applied to these systems and to illustrate how the integration of computation and experiment can advance efforts in the structure determination of carbohydrate–protein complexes (Figure 1).

Fig. 1.

The roles of computational methods (purple) alongside experimental methods (blue) in structural glycobiology.

Predicting carbohydrate–protein interactions

Given the development of high-throughput affinity assays, employing immobilized microarrays of biologically relevant glycans, there is now a considerable amount of data on carbohydrate–protein interactions in vitro (Fukui et al. 2002; Blixt et al. 2004; Paulson et al. 2006). Microarray technologies are able to identify carbohydrate ligands for a protein receptor, providing an unprecedented level of information pertinent to carbohydrate recognition. But the picture is incomplete. While microarray data may identify a core carbohydrate recognition sequence, the precise manner in which the protein recognizes and binds to a ligand remains undefined. To complement experimental structural studies of carbohydrate–protein complexes, computational docking of carbohydrates to proteins can contribute to the understanding of carbohydrate recognition.

Drug design

Often host cell-surface carbohydrates are the targets for invading microbes and viruses (Karlsson 1986; Smith et al. 2004) and conversely may form targets for the design of novel therapeutic agents (von Itzstein et al. 1993). In just such a case, the G_M1 binding site of the heat-labile enterotoxin produced by E. coli (a close structural analogue to the cholera toxin produced by Vibrio cholerae) was targeted for an inhibitor design (Minke, Diller, et al. 1999). Using crystal structures of the heat-labile enterotoxin with fragments of the carbohydrate head group of G_M1, potential inhibitors were identified (Minke, Roach, et al. 1999). The program AutoDock (Goodsell and Olson 1990) was used to predict the structure of the toxin–inhibitor complexes. After the docking studies were complete, the structures of the inhibitor–toxin complexes were solved by X-ray crystallography (Merritt et al. 1997), providing a direct test of the docking protocol. In general, the experimentally obtained structures validated the models predicted via the docking procedure (Figure 2). In the case of one inhibitor, melibionic acid (galactose-α-(1-6)-gluconic acid), the majority of the highest ranked binding modes from the docking simulations placed the galactose residue in a similar position to that found in the crystal structure. For the gluconic acid residue there was no consensus among the top-ranked binding modes, as the ligand adopted several poses. It is noteworthy that this residue was also unresolved in the crystal structure (Merritt et al. 1997) (a common occurrence when dealing with flexible molecules like carbohydrates). Melibionic acid thus bound to the toxin primarily through interactions made by the galactose residue while the gluconic acid provided non-specific interactions that enhanced the affinity relative to free galactose.

Fig. 2.

Inhibitors of the heat-labile enterotoxin from E. coli. Computationally docked galactose derivatives (white) were able to reproduce the experimentally observed binding modes (black) (reprinted from Minke, Diller, et al. 1999).

Identifying the protein contact residues

While the above example illustrates a successful use of docking in the characterization of the interaction between a protein and small carbohydrate derivatives, docking simulations have also been used to probe larger carbohydrate–protein interactions (Bitomsky and Wade 1999; Sachchidanand et al. 2002; Kadirvelraj et al. 2006). Heparin, a complex sulfated glycosaminoglycan, has several known protein-binding partners including antithrombin III and interleukin 8 (IL8). Lacking a structure of the heparin–IL8 complex, efforts have been undertaken to predict this carbohydrate–protein complex using a docking-based approach (Bitomsky and Wade 1999). The docking-based protocol included (1) using mono- and disaccharide heparin fragments to run a global docking search using several docking programs (GRID (Goodford 1985), AutoDock (Goodsell et al. 1996), and DOCK (Ewing and Kuntz 1997)); (2) calculating the probability that a particular protein residue would be found in the binding interface, based on the results from the global searches; and (3) docking a hexasaccharide fragment of heparin in the plausible binding regions. This method was first tested on three heparin-binding proteins with experimental structures, and the success of these test cases, along with available biological and spectroscopic data (Webb et al. 1993; Kuschert et al. 1998), supported the application of this protocol to the prediction of the heparin–IL8 complex. Using computational docking, it was possible to identify a shallow carbohydrate-binding site on the IL8 surface and construct 3D models of the binding mode of a complex and highly charged oligosaccharide.

Enzymatic pathway analysis

Computational docking can also assist in predicting carbohydrate–enzyme structures, and in defining enzymatic pathways. Recently the mechanism of action of a glycoside hydrolase in the N-glycan synthesis pathway, α-(1–2)-mannosidase I, was predicted using docking simulations (Mulakala et al. 2006), in line with previously predicted mechanisms, based on crystal structures of the enzyme with various substrates (Vallee et al. 2000; Karaveg et al. 2005). By computationally docking 16 α-D-manno-(1-2)-α-D-mannose conformers (generated by constraining relevant ring atoms and then minimizing the structure in vacuo) and evaluating binding energies and forces on the substrates (Figure 3) that were proposed to drive it toward the transition state (with AutoDock), a reasonable enzymatic pathway was deduced (Mulakala et al. 2006). The substrate of α-(1-2)-mannosidase I was predicted to populate conformations along the following pseudo-rotation pathway from its starting conformation (either ¹C₄ or ⁰S₂) through to the transition state (³E): ¹C₄ → ³H₂ → ⁰S₂ → ^3,0B → ³S₁ → ³E (Mulakala et al. 2006).

Fig. 3.

Docking used to compute hydrogen-bonding forces on the substrate α-D-mannopyranosyl-(1-2)-α-D-mannopyranose (³H₂) exerted by the active site of yeast α-(1-2)-mannosidase (measurements in pN) (reprinted from Mulakala et al. 2006).

Computational docking methods have evolved to help overcome difficulties in obtaining experimental 3D structures of multiple ligands to a common receptor. Docking strategies can utilize experimental or modeled protein structures to aid in the discovery new lead compounds for drug design. In the field of glycobiology docking simulations have several applications, in addition to drug design, as they provide computationally tractable methods to predict carbohydrate-binding sites on proteins, bound ligand conformations, and conformational pathways for carbohydrate-processing enzymes. Docking simulations are however far from infallible and benefit from experimental verification, as false positives and negatives are common. Typically, docking simulations do not generate a single model rather multiple plausible low energy poses. These results require ancillary information in order to select the most probable candidates. To facilitate selection, more rigorous post-docking approaches can be used to generate additional data for predicted carbohydrate–protein complexes; two such methods are highlighted in the following section.

Calculating relative and absolute binding free energies for carbohydrate–protein complexes

Characterization of the structural and thermodynamic properties of carbohydrate–protein interactions is desirable for many reasons, from understanding basic biological function to structure-based drug design. Two types of theoretical methods are commonly used to calculate the free energy of binding of a carbohydrate–protein complex: direct ΔG calculations and thermodynamic integration (TI) methods (Figure 4). Direct ΔG calculations use only the initial (free receptor and ligand) and final (complex) states of the cycle, which correspond to experimentally observable states. Information regarding the pathway is not required. Descriptions of the initial and final states are generated as structural ensembles collected from independent MD simulations, or from decomposition of a single MD simulation of the complex (reviewed in Swanson et al. 2004).

Fig. 4.

(A) The thermodynamic cycle for ligands (orange and green) binding to a protein receptor (blue). Direct ΔG calculations determine the free energy of binding of a ligand to a receptor (ΔG₁), whereas TI methods calculate the free-energy difference between receptor–ligand complexes (ΔG₄), where only the ligand is changed. (B) By considering the ΔG₄ pathway as a series of nonphysical intermediate states (0 < λ < 1), the free energy difference between the real states λ = 0 and λ = 1 can be computed by taking the ensemble average of the first derivative of the potential energy (V) with respect to λ.

In TI calculations the relative ΔG_binding is computed for closely related systems by slowly transforming or perturbing the initial state to the final state (Figure 4). By employing an appropriate (although nonphysical) thermodynamic cycle and by proceeding in small discrete steps, the relative ΔG_binding may be computed (Zwanzig 1954).

Each method has advantages and disadvantages (as highlighted in Table I). TI calculations are best suited to a set of structurally similar ligands or a set of mutations within a binding site. As for example in the refining of a lead compound into a high affinity binder or in the design of higher affinity receptor proteins. To screen structurally dissimilar ligands, or a single ligand against multiple receptors, direct ΔG calculations are preferred. Both of these theoretical methods can provide a detailed structural explanation of the system under study; notably, they can provide an estimate of the contributions of specific residues and atoms to the free energy, as well as decomposition of energetic contributions from steric and electrostatic components, something that is difficult to achieve experimentally.

Table I.

A comparison of methods used to predict binding free energies

Comparison	Automated docking	Direct ΔG calculations	Thermodynamic integration
Computational efficiency	High	Moderate	Low
Ligand set	Can be diverse	Can be diverse	Close structural analogues only
Binding free energy computed	Absolute ΔG	Absolute ΔG	Relative ΔG
Water model	Implicit	Implicit	Explicit
Accuracy strongly dependent on	Compounds used for calibration	Water model, force field, and sampling	Sampling and force field

Direct ΔG calculations

Carbohydrate-binding proteins, known as lectins, exhibit a range of carbohydrate-binding specificities. Galectin-3, a mammalian lectin with relatively narrow binding specificities, has affinity for blood group A-oligosaccharides. At the other end of the spectrum are lectins such as galectin-1 that can bind a relatively diverse group of oligosaccharides (Consortium for Functional Glycomics (CFG), http://www.functionalglycomics.org). How is it that lectins are able to display such broad binding specificities for structurally distinct carbohydrate epitopes? To address this question, the ability of galectin-1 to bind a variety of oligosaccharides with varying specificity was explored using a combination of computational docking, MD simulations, and binding free energy calculations (Ford et al. 2003). Galectin-1 is known to bind to Gal-β-(1-4)-GlcNAc (LacNAc)-containing oligosaccharides and there are several crystal structures of it in complex with ligands including LacNAc (Liao et al. 1994). MD simulations, using the GLYCAM force field (Woods et al. 1995), of galectin-1 complexed (via docking simulations) with carbohydrate ligands containing the LacNAc core were run to determine the binding mechanism of LacNAc derivatives (Ford et al. 2003). From these simulations, it was concluded that various substitutions at the nonreducing end of the LacNAc core were tolerated with little or no disruption of key-binding interactions. Hydrogen bonding and aromatic stacking that occurred between LacNAc and galectin-1 were also found to be present for all of the modified LacNAc ligands. Direct ΔG calculations based on the results from the MD simulations provided good qualitative agreement with experimentally determined binding affinities for the ligands under study. In addition to known ligands, two simulations were run with negative controls (GlcNAc and N-acetylmaltosamine) and in both cases the ligands diffused away from the binding pocket early in the simulation. The inclusion of negative controls in MD simulations provides considerable confidence in the MD timescale and in the force field used for the simulation. MD simulations in combination with direct ΔG calculations predicted binding conformations that explained the weakly selective binding behavior of galectin-1 and qualitatively ranked the predicted free energies of binding in accordance with the experimentally determined binding affinities.

Similarly, a direct ΔG computational study using the GLYCAM force field (Woods et al. 1995) exploring the specificity of concanavalin A, a plant lectin specific for mannose-containing oligosaccharides (common to N-linked glycans), was successful in structurally accounting for the ligand preferences of the protein (Bryce et al. 2001). The results were further de-convoluted by decomposing binding energy into entropic and enthalpic contributions. Binding energy contributions were analyzed on a per-residue basis, identifying key interactions for the general recognition strategy and specificity of concanavalin A. The specific interactions of high-mannose fragments that led to the observed binding differences, could thus be ranked.

In addition to examining biological ligands, direct ΔG calculations have unraveled the energetics of binding of carbohydrate inhibitors for disease-related enzymes. Neuraminidase, a viral coat protein from influenza, is responsible for cleaving terminal neuraminic acid (Neu5Ac, sialic acid) residues on the host cell surface to promote viral fusion and release of viral progeny (Palese et al. 1974; Palese and Compans 1976; Liu et al. 1995). Due to its central role in influenza infection, viral neuraminidase has been selected as a drug target. Inhibitors based on the natural ligand, Neu5Ac, boosted the affinity from millimolar for the natural ligand (Potier et al. 1979; Jedrzejas et al. 1995) to nanomolar for some of the better inhibitors (von Itzstein et al. 1993; Finley et al. 1999; Babu et al. 2000; Chand et al. 2001). Direct ΔG calculations, using the AMBER force field, revealed that the energetic component that resulted in such a jump in potency was caused by hydrophobic contacts (Masukawa et al. 2003). The neuraminidase inhibitors tamiflu and relenza have increased van der Waals interactions over the natural ligand resulting in more direct contacts with the enzyme and fewer water-mediated interactions. Electrostatic contributions were less important in determining the binding affinity but were useful in correctly orienting the ligand within the active site.

The direct ΔG method is not only applicable to the study of small ligands, such as oligosaccharide fragments or small molecule carbohydrate derivatives, but also for large complex carbohydrates. Antibodies against capsular polysaccharides (CPS) from one strain of Streptococcus, such as Group B Streptococcus agalactiae type III (GBSIII), rarely cross-react with other strains, such as the CPS from Streptococcus pneumoniae type 14 (Pn14). GBSIII and Pn14 share a similar backbone sequence, but Pn14 lacks α-Neu5Ac in its side chains. This minor difference attenuates the antigenicity of Pn14-type CPS as compared to the sialated GBSIII type-CPS for antibodies raised against the sialated antigen (Jennings et al. 1981). To quantify differences in immunogenicity and antigenicity of CPS from a Gram-positive bacterium (Group B S. agalactiae), Kadirvelraj et al. (2006) used a combination of computational tools, including molecular modeling, docking, MD simulations, and direct ΔG calculations. While the conspicuous difference between GBSIII and Pn14 is the loss of an acidic monosaccharide, the primary contributions to differences in affinity were found to be entropic in origin, as determined via MD simulations of the immune complex and direct ΔG calculations with the GLYCAM force field (Kadirvelraj et al. 2006). Pn14 exhibited greater flexibility in solution, and thus paid a greater entropic penalty upon binding to the antibody, as compared to GBSIII. The absence of sialic acid in the side chains of Pn14 also decreased contributions to van der Waals stabilization energy and the electrostatic interaction energy. These factors combined to yield differences in affinity that accounted for differential antibody recognition of bacterial CPS and provided a structural framework for interpreting the observed immunological data.

Thermodynamic integration calculations

The application of TI calculations to the analysis of binding free energies of inhibitors to carbohydrate-binding enzyme includes α-D-glucose-based inhibitors of glycogen phosphorylase (Archontis et al. 2005). A potent inhibitor of glycogen phosphorylase, hydan (a spirohydantoin of glucopyranose), and two of its analogues (methyl-hydan and NH₂-hydan) have been co-crystallized with the enzyme (Gregoriou et al. 1998; Oikonomakos et al. 2002; Watson et al. 2005). These structures provided a qualitative explanation for the differences in observed binding affinities (relative binding affinities: hydan > NH₂-hydan > methyl-hydan ≫ α-D-glucose (Bichard et al. 1995; Watson et al. 2005)). The addition of data from TI simulations using the CHARMM22 glucose force field parameters (MacKerell et al. 1998) provided a quantitative description of the binding interaction energies (Archontis et al. 2005). The addition of functional groups (methyl and NH₂ group) introduced a destabilizing van der Waals free energy component (the methyl group more so than the NH₂) through unfavorable interactions of the ligands with an aspartate residue and a water molecule in the binding site. NH₂-hydan improved with respect to electrostatic energies over hydan, but this was not sufficient to overcome the destabilizing van der Waals contacts. These results suggested alternative sites for hydan modifications that could yield more potent inhibitors.

Free energy calculations based on computational simulations of protein–carbohydrate complexes have also been employed to investigate antigenicitiy of carbohydrate–antibody interactions (Pathiaseril and Woods 2000). As a test of the ability of a method closely related to TI, free energy perturbation, to predict relative binding free energies for a series of haptens, the interaction energy of the trisaccharide epitope of the Salmonella serotype B O-antigen and a related monoclonal antibody fragment were analyzed. From calorimetric experiments, five structural analogues of the natural hapten were estimated to have similar binding affinities to the natural epitope, while one congener displayed very poor binding affinity (Bundle et al. 1994). Free energy perturbation simulations reproduced the experimentally determined relative free energies within an absolute error of 0.55 kcal/mol and revealed the likely protonation state of a histidine residue in the binding cleft. This computational study also demonstrated one of the challenges of free energy perturbation simulations, predicting relative free energies of ligands that differ only in their interactions with the solvent. This is also a challenge for the direct ΔG method, where accurately calculating the contributions of ligand–solvent and protein–solvent interactions tests the limits of the computational model.

Accurate representation of the effects of water as a solute is critical for the success of all of the computational methods discussed (MD, docking, and ΔG calculations). With an accurate force field and explicit water, MD simulations can generate conformational families that are consistent with experimentally observed structural data (Corzana et al. 2004; Gonzalez-Outeirino et al. 2006; Pereira et al. 2006). For docking studies, inclusion of conserved water molecules in the binding site can improve the success of the simulation (Minke, Diller, et al. 1999; Rarey et al. 1999; Österberg et al. 2002); however, the position of key waters is not always known. In the case of free energy simulations, the energetic contributions from direct interactions between the ligand and receptor can often be well modeled although the accurate treatment of contributions from ligand–solvent and protein–solvent interactions remains a challenge.

Emerging and evolving techniques for experimental characterization of carbohydrate–protein structures

Partially oriented NMR spectroscopy

Over the past decade, NMR structural constraints from anisotropies in spin interactions have greatly expanded the possibilities for biomolecular structure determination. By partially orienting a sample via medium- or field-induced alignment, observables that arise from anisotropic interactions such as residual dipolar couplings (RDCs), chemical shift anisotropy (CSA) offsets, and pseudocontact shifts in paramagnetic systems can be detected (reviewed in Prestegard 1998; Prestegard et al. 2004). In a recent application, a combination of NMR experiments, mass spectrometry (MS), and computational simulations, yielded a novel strategy to probe oligosaccharide conformation in a solution (Yu et al. 2007). In this experiment, a pentasaccharide fragment of chondroitin sulfate was isotopically enriched by replacing acetyl groups with ¹³C-labeled acetyl groups. The resulting ¹³C spectra provided structural constraints from the orientational dependence of RDCs and CSA offsets. To assign resonances, isotope ratios in the pentasaccharide, observed via the mass spectra, were correlated with enrichment levels seen in the NMR spectra. The ¹³C data alone were too sparse for structure determination, so additional structural constraints (NOEs, J-coupling constants, ¹H–¹H, and ¹³C–¹H RDCs) were determined. In the final structure determination steps, a combination of computational tools was used: REDCAT (residual dipolar coupling analysis tool) to calculate alignment parameters (Valafar and Prestegard 2004) and XPLOR-NIH (using the GLYCAM force field; Kirschner et al. 2008) to optimize the structure via simulated annealing (Schwieters et al. 2003, 2006). The ultimate goal of this novel strategy is to examine oligosaccharides in complex with proteins, in which case the ¹³C data can be augmented with principal order parameters determined from the protein and conformational restraints for the oligosaccharide from computational simulations (Yu et al. 2007). This combination of NMR, MS, and computation highlights the creative cross-disciplinary approach necessary to tackle oligosaccharide–protein structure determination.

Hydroxyl radical protein footprinting with mass spectrometry

The concept of “footprinting” generally refers to techniques that probe macromolecular surface changes upon complex formation of two or more molecules, by modifying solvent-exposed surfaces. This technique has existed since the late 1970s when it was first applied to probe DNA–protein complexes (Galas and Schmitz 1978; Schmitz and Galas 1980). Since that time the concept has evolved to include footprinting of RNA–protein (Motoki et al. 1991; Wang and Padgett 1989) and protein–protein complexes (Sheshberadaran and Payne 1988). In glycoscience, footprinting is drawing attention for its potential to define carbohydrate-binding surfaces on proteins. Footprinting data could be used to complement computational predictions for systems that are not amenable to NMR spectroscopy or crystallography.

The highest resolution protein footprint technique currently available uses hydroxyl-radical oxidation of amino acid side chains, which can be detected and quantified by MS techniques to map the solvent accessible surfaces of proteins and protein complexes (reviewed in Takamoto and Chance 2006). Since many amino acid side chains react with hydroxyl radicals, this technique is capable of relatively high-resolution mapping. In contrast to H/D amide exchange MS, oxidative footprinting has two critical advantages, namely it modifies the amino acid side chains, rather than the backbone and it is not a chemically labile modification. Hydroxyl radicals can be generated using metal-dependent chemical generation from peroxide, by radiolysis of dilute peroxide solutions or by direct radiolysis of water. Hydroxyl radicals are ideal reagents to probe solvent accessible surfaces, as they are small and highly reactive. The relative reactivity of amino acid side chains has been characterized (Xu and Chance 2005) and all but the amino acids Gly, Ala, Asp, Asn, Ser, and Thr are sufficiently reactive to be useful probes. Since 14 types of amino acids are potentially reactive (Xu et al. 2003), this method boasts considerably higher structural resolution than residue-specific techniques (such as lysine modification), which provide sparse sampling of the protein surface. In addition, initial studies demonstrate that the extent of oxidation (reactivity) is strongly dependent on the solvent accessible surface area (SASA) of the amino acid (Charvátová, Foley, Bern, Sharp, Orlando, and Woods, in preparation) (Figure 5).

Fig. 5.

Correlation between reactivity and SASA (calculated from MD simulation) for amino acids from galectin-1: (A) phenylalanine (high reactivity), (B) proline (medium reactivity), and (C) asparagine (control/nonreactive).

After brief exposure to hydroxyl radicals the sample is subjected to protease digestion. To determine the modification sites and to quantify the extent of oxidation, MS/MS techniques are employed. Due to the potential for multiple oxidation states for some residues and oxidation combinations within a given proteolytic peptide, a far larger and more complex data set is obtained than typical for a proteomics analysis. Efficient and reliable identification of the locations and relative amounts of oxidation requires sophisticated emerging computational analysis.

Recently, we have been exploring hydroxyl radical footprinting techniques for the determination of binding surfaces of protein–carbohydrate complexes. In particular, we were interested in assessing the potential of this method to resolve carbohydrate-binding sites for future studies where the protein–carbohydrate complex may not be amenable to traditional structure determination methods (Figure 1), such as antibody–polysaccharide complexes. Initially, we used human galectin-1-lactose as a model system and pulse laser irradiation of a 1% hydrogen peroxide solution for the generation of hydroxyl radicals. By comparison with MD data based on the crystal structure of the galectin-1–lactose complex (Lopez-Lucendo et al. 2004), hydroxyl radical footprinting was able to identify those residues defining the binding site (Figure 6). Differences in relative oxidation levels correlated well with the difference in SASA calculated from MD simulations of galectin-1 with and without lactose (Figure 6D) (Charvátová and Woods, unpublished data).

Fig. 6.

Extent of oxidation mapped onto the surface of the crystal structure of galectin-1 in the presence (A) and absence (B) of ligand. (C) The difference in the level of oxidation with and without ligand (with the ligand superimposed in green for reference) correlated with (D) the change in SASA of the protein from MD simulations with and without the ligand present.

Due to the large number of uncommon modifications that arose from the oxidation experiment, available proteomics (database-search) analysis software (Yates et al. 1995; Perkins et al. 1999) were poorly suited to assign the mass spectra of the modified peptides. A more efficient and sensitive identification of modified peptides was accomplished using the program ByOnic, which combines database searching with de novo analysis (Bern et al. 2007). Currently, the analysis of the MS data presents one of the most significant obstacles to hydroxyl radical footprinting technology, which must be overcome before footprinting makes the jump from “emerging” to established technology.

The ever-increasing pool of known carbohydrate-binding proteins and their preferred carbohydrate epitopes necessitates the development of efficient means to identify their interfaces; a challenge potentially well suited for hydroxyl radical footprinting and MS analysis.

Saturation transfer difference NMR

While footprinting may be used to identify the protein interface, STD NMR allows rapid identification of the regions of a carbohydrate ligand that are proximal to the protein surface in a complex (Mayer and Meyer 1999). STD NMR is now a mature technology; however, integrating STD NMR data with other methods, such as computational docking, is still evolving.

For the STD NMR experiment a single solution of the protein and carbohydrate ligand (with ~100-fold molar excess of ligand) is required, and two ¹H-NMR spectra, at different saturation frequencies, are measured. The STD spectrum, or the ‘on-resonance’ spectrum, is acquired by irradiating at a small window of frequency where only the protein resonates (Figure 7). The protons from a ligand in exchange between the bound and free forms can also become differentially saturated when bound to the protein. The reference spectrum, or the ‘off-resonance’ spectrum, is acquired with irradiation far from the protein’s or ligand’s frequency range. The difference between the on- and off-resonance spectra, the difference STD NMR spectrum, identifies carbohydrate protons in close proximity to protons from the protein, revealing the carbohydrate epitope (Figure 7). A lack of observed intensities for carbohydrate protons does not necessarily indicate that they are distant from the protein surface, since carbohydrate interactions with regions of the protein surface with low-proton density, or interactions mediated by water do not facilitate magnetization transfer. Computational modeling used in conjunction with STD NMR can help explain such ambiguous cases, improving resolution of and confidence in the carbohydrate-binding epitope, or STD NMR can be used to validate computational predictions.

Fig. 7.

STD NMR characterized the carbohydrate-binding epitope of the siglec sialoadehesin at high resolution. (A) The STD NMR spectrum of sialyllactose in the presence of sialoadhesin and (B) the corresponding reference ¹H NMR spectrum at 500 MHz. (C) The relative STD effects of sialyllactose bound to sialoadhesin. Percentages were calculated using the difference in individual signal intensities between the measured spectra and normalized to the largest STD effect (reprinted from Bhunia et al. 2004).

In addition to being a relatively straightforward experiment, STD NMR is advantageous because isotope labeling of the protein or ligand is not required (Meyer and Peters 2003). Some factors to consider when taking the spectra include the T₁ relaxation times of the ligand protons and the effect of temperature on the measured spectra (Yan et al. 2003). Most critically, a fast off-rate of the ligand is required in order to observe saturation transfer to the ligand (Mayer and Meyer 1999).

STD NMR has been employed to characterize the carbohydrate-binding epitopes for a variety of classes of protein receptors, including glycosytransferases (Biet and Peters 2001; Macnaughtan et al. 2007), bacterial toxins (Haselhorst et al. 2004), viral and antiviral proteins (Sandstrom et al. 2004; Haselhorst et al. 2007), parasitic sialadases (Todeschini et al. 2002), lectins (Bhunia et al. 2004) (Figure 7), and antibodies (Maaheimo et al. 2000; Johnson and Pinto 2004). With data from STD experiments, validation of atomic resolution computational models of the protein–carbohydrate interaction is possible. In addition, using programs, such as CORCEMA (Moseley et al. 1995), nuclear Overhauser effect intensities can be computed for the complex allowing back-calculation of STD NMR intensities. Calculated STD NMR intensities have recently been employed in the validation of docking simulations (Bhunia et al. 2004; Haselhorst et al. 2004, 2007) and for refinement of docking poses (Jayalakshmi and Rama Krishna 2004).

Varieties of footprinting and saturation transfer NMR techniques have existed for decades, yet the application and modification of these methods to investigate problems in glycoscience has come about relatively recently. In a field where obstacles in structural determination abound, the potential synergy between these two methods is an appealing prospect for obtaining medium-to-high resolution structural information. Combining these methods with computational tools, such as protein homology modeling, MD simulations or docking, could provide experimentally guided/validated atomic resolution descriptions of protein–carbohydrate interactions.

Current challenges in structural glycoscience

Carbohydrate structure prediction

Given only the sequence of a protein, scientists have several tools for predicting its 3D structural properties, including homology or comparative modeling, threading and secondary structure prediction algorithms based on sequence similarity to other known proteins (reviewed in Marti-Renom et al. 2000), and more recently, de novo fold prediction algorithms (Baker 2000). Unfortunately, strategies based on sequence homology are unlikely candidates for application to carbohydrates. Two carbohydrate molecules may share a similar core structure, yet carbohydrate-residue modifications via addition, substitution, deletion, change in linkage configuration or chemical modification may yield molecules with vastly different structural properties. An example of this phenomenon is seen in amylose versus cellulose; both are linear polymers of glucose found in plants. The building blocks of amylose and cellulose are glucose-α-(1-4)-glucose and glucose-β-(1-4)-glucose, respectively. Because the building blocks are diasteriomers, the polymers have different structural and chemical properties leading to very different functions within plants and different susceptibilities to hydrolysis (digestion) in humans. Cellulose polymers are linear and rigid (Nishiyama et al. 2002, 2003), whereas amylose polymers are helical in shape and have more internal flexibility than cellulose (Gessler et al. 1999). For carbohydrates a single change, such as modifying an anomeric configuration or linkage position, is sufficient to create two biologically distinct molecules.

For proteins, mutations in the sequence can be classified in two general groups: conservative, such as Leu → Val, or nonconservative, such as Leu → Asp. However, carbohydrate mutations are not easily classified in such terms. While carbohydrate residues can be grouped into distinct families (hexoses, hexosamines, acidic, etc.), most carbohydrate residues share similar properties, particularly the existence of polar and nonpolar groups. Thus, such grouping is of limited usefulness since it is the great diversity in topological presentation of these groups that give carbohydrate residues their complexity. As discussed in the introduction, structural prediction is often further complicated by the existence of several low energy conformations under biological relevant conditions. Two computational methods, MC and MD simulations, have been extensively employed in the prediction of carbohydrate conformational families. Over the past decade, explicit solvent MD simulations have been successful in predicting conformational families, and often the relative populations of these families (Naidoo et al. 1997; Picard et al. 2000; Corzana et al. 2002, 2004; Eklund et al. 2005; Landersjo et al. 2005; Pereira et al. 2006). Evolution of new computational tools to predict the carbohydrate structure will benefit from the knowledge gained through experiments and computational simulations, and the challenge will be in finding ways to apply what we have learned to generate more efficient and accurate structure prediction tools.

Nomenclature

As part of the set of tools to probe the carbohydrate structure, nomenclature systems that effectively and precisely communicate carbohydrate structure—from monosaccharide composition to 3D shape—are essential. For many years glycoscientists have relied on the International Union of Pure and Applied Chemistry and the International Union of Biochemistry and Molecular Biology (IUPAC-IUBMB) standard (Figure 8) (McNaught 1997). Unfortunately, the IUPAC nomenclature system for carbohydrates is extremely complex, and unlike the IUPAC system for amino acids, peptides, and proteins, it is not suitable for many glycomics applications. In the short term, a glycomics nomenclature “wish-list” might include a standardized single-letter code for monosaccharides and a nomenclature system consistent with the PDB file format. Due to the format restrictions of PDB files (discussed below), a continuing objective is the development and implementation of a comprehensive nomenclature system appropriate for current and future glycomics applications. It is likely that future nomenclature systems will evolve to be machine readable, for reasons of accuracy and completeness, at the expense of human readability. Nevertheless, within the foreseeable future, the PDB format will continue to be the most widely supported and accepted file format for structural data.

Fig. 8.

A sample of carbohydrate nomenclature systems and schematic representations using a human blood group A antigen as an example. Due to the lengthy formats of GLYDE-II and GlycoCT, readers are directed to relevant websites to view sample XML codes.

For a standardized code for monosaccharides, we propose a one-letter code for monosaccharides and a two-letter code for their common derivatives (Table II). This system could be generalized for other applications or used to simplify existing nomenclature systems (Figure 8). This system is currently utilized in the GLYCAM carbohydrate and lipid force field as the foundation of its three-character PDB-style nomenclature system (http://www.glycam.com).

Table II.

Monosaccharide codes that form the core of the PDB-style GLYCAM nomenclature

Symbol	Core monosaccharidesa	Symbol	Special cases and common derivativesb
	Pentoses		2-N-Acetylhexosamines
A	arabinose	GN	N-acetylglucosamine
R	ribose	LN	N-acetylgalactosamine
X	xylose	MN	N-acetylmannosamine
D	lyxose
			Uronic Acids
	Hexoses	GU	glucuronic acid
G	glucose	LU	galacturonic acid
M	mannose	MU	manuronic acid
L	galactose	IU	iduronic acid
I	idose
			Sialic Acids (9 carbon)
T	talose
		9N	N-acetylneuraminic acid
N	allose
		9G	N-glycolylneuraminic acid
E	altrose
		9O	3-deoxy-D-manno-oct-2-ulosonic acid (KDO)
K	gulose
	Hexuloses		Others
		8O	3-deoxy-D-glycero-D-galacto-non-2-ulosonic acid (KDN)
C	fructose
P	psicose
S	sorbose
J	tagatose
	6-Deoxyhexoses
F	Fucose
Q	quinovose
H	rhamnose

^aCase designates the isomeric configuration: D (upper case) and L (lower case).

^bThe case of the first character designates the isomeric configuration: D (upper case) and L (lower case), with the exception of sialic acids and KDN. For the second character, case designates the anomeric configuration: α (upper case) or β (lower case).

Since, at present, all molecular visualization programs as well as many other types of biostructural software depend on the PDB format, a three-character code for carbohydrates is necessary. Creating a three-character system that captures the structural diversity of carbohydrate structures presents many challenges but would greatly facilitate structural searches and glycomics analysis. One of these challenges is the desire to explicitly state linkage information. When dealing with proteins, the linkage between amino acids is implicit in its sequence. Linkage information for carbohydrate molecules must be determined experimentally at each position (which is not a trivial task) and then explicitly defined in the nomenclature. To add another layer of complexity, carbohydrate residues can exist as D- or L-isomers, α- or β-anomers, and as pyranose or furanose ring forms. In animal systems, most monosaccharides exist as a single isomer (e.g. D-mannose), yet there are sufficient counter examples from outside the animal kingdom (Fichtinger-Schepman et al. 1979, 1981) to make the case for the inclusion of isomeric configurations in the nomenclature. This means that any standard code for carbohydrates residues would ideally include, at a minimum, five types of information: (1) residue type, (2) isomer, (3) anomer, (4) linkage, and (5) ring structure.

While it would be preferable for the PDB format restriction if all this information could fit into an uppercase three-letter residue identifier, it is not possible to specify all potential permutations of carbohydrate residues using only three noncase sensitive characters. As such, machine-readable nomenclature systems that dispense with the three-character PDB-constraint, such as the linearly formatted LINUCS (Bohne-Lang et al. 2001) and LinearCode (Banin et al. 2002), and the XML programming language codes Glyco-CT (Herget et al. 2007) and GLYDE-II (Sahoo et al. 2005), have been proposed for bioinformatics applications and described elsewhere (Figure 8). To accommodate the numerous biostructural applications that still make use of the PDB format, the GLYCAM PDB-style carbohydrate nomenclature system was created.

The development of the GLYCAM PDB-style nomenclature has evolved over the past 14 years (Woods et al. 1995; Tessier et al. 2007; Kirschner et al. 2008), originating from the necessity of having a descriptive and logical three-character naming system for carbohydrate residues for MD simulations with AMBER (Case et al. 2005). Here we propose a case sensitive three-character code that provides a descriptive, but concise nomenclature system for all monosaccharides and several common monosaccharide derivatives (Tables II–V) that can still meet the restriction of the PDB file format. A complete description of the nomenclature system can be found on the GLYCAM website (http://www.glycam.com). In general, the PDB-style GLYCAM code uses the first character to encode the linkage position(s) (Table III), the second to encode the residue (via the character, Table II) and the isomer (via case), and the third to encode anomeric configuration (via the character) and the ring size (via case). For specific monosaccharide derivatives, the third-placed character encodes the derivative class (via the character, Table II) and anomeric configuration (via case) at the expense of defining the ring size. Given only three characters, we have attempted to generate a comprehensive, relatively readable and flexible nomenclature system (Tables II–V). This naming convention, as well as the GLYCAM parameters, can be used with molecular mechanics programs that support AMBER input files (Case et al. 2005), including NAMD (Phillips et al. 2005) and NWChem (Kendall et al. 2000). Additionally, the GLYCAM nomenclature system can be implemented in other molecular mechanics programs such as GROMOS (Scott et al. 1999) and CHARMM (Brooks et al. 1983).

Table V.

Examples of PDB-style GLYCAM residue identifiers for common monosaccharide derivatives (N-acetylhexosamines and uronic acids) specifying linkage position (first position), monosaccharide residue (second and third position), D- or L-configuration (second position case), and anomeric configuration (third position case)

Linkage	N-Acetylhexosamines				Uronic acids
	D-Sugars		L-Sugars		D-Sugars		L-Sugars
	α-D-Glcp-Nac	β-D-Glcp-NAc	α-L-Glcp-Nac	β-L-Glcp-NAc	α-D-Glcp-A	β-D-Glcp-A	α-L-Glcp-A	β-L-Glcp-A
3-	3GN	3Gn	3gN	3gn	3GU	3Gu	3gU	3gu
4-	4GN	4Gn	4gN	4gn	4GU	4Gu	4gU	4gu
3,4-	WGN	WGn	WgN	Wgn	WGU	Wgu	WgU	Wgu

Table III.

One-letter codes for specifying linkage positions in pyranoses and furanoses

Symbol	Linkage	Symbol	Linkage
0	Terminal	v	3,6-
1	1-	U	4,6-
2	2-	M	5,6-
3	3-	T	2,3,4-
4	4-	L	2,3,5-
6	6-	s	2,3,6-
Z	2,3-	R	2,4,6-
Y	2,4-	K	2,5,6-
O	2,5-	Q	3,4,6-
x	2,6-	J	3,5,6-
w	3,4-	p	2,3,4,6-
N	3,5-	I	2,3,5,6-

While suitable for many monosaccharides, no three-character system is sufficiently robust to cover all possible monosaccharide derivatives. Due to this format-imposed limitation, the GLYCAM nomenclature system includes monosaccharides and common monosaccharide derivatives, omitting the more elaborate and rare derivatives. Another issue arises due to potential overlaps with existing PDB residue identifiers. For these reasons, we are working to implement an extended version of the GLYCAM code suited for the newer Research Collaboratory for Structural Bioinformatics standard format, macromolecular Crystallographic Information File (mmCIF). This newer format dispenses with the three-character limitation, which will permit the development of a highly descriptive and comprehensive, yet relatively user-friendly, nomenclature system.

Conclusion

The ability of proteins to recognize and distinguish carbohydrate molecules is at the heart of many critical biological processes, such as cell adhesion, adaptive and innate immunity, and cell signaling. Carbohydrates flex and bend, taking on different 3D shapes, yet somehow nature has evolved to recognize these complex and pliable molecules. Protein receptors are able to differentiate between (1) the varied 3D shapes any given carbohydrate may adopt and (2) closely related carbohydrates that differ only in such structural properties as configuration or linkage position. To develop a better understanding how nature employs the peculiar structural and dynamic properties of carbohydrates to its advantage, 3D characterization of carbohydrate–protein complexes is required. However, structure determination of flexible and dynamic molecules is a great challenge. Beginning with more rigid systems (often small carbohydrates in complex with proteins, or glycoproteins with deliberately trimmed glycans), crystallography and NMR spectroscopy added the first few rungs of the technical ladder that leads to more high throughput methods for carbohydrate structure determination. For the next rung of the ladder, MD simulations were employed, alongside NMR techniques, as a means to identify the 3D shapes, and relative populations, of more complex systems.

The latest major advance has been the creation and widespread use of glycan microarrays for receptor screening. With the emerging wealth of microarray data comes the urgent need for high throughput methods that provide structural insight into the mechanisms of carbohydrate recognition. By characterizing the interacting interfaces between ligands and protein surfaces, experimental and computational methods offer a new route to deriving the 3D structures of carbohydrate–protein complexes. Current computational techniques for carbohydrate–protein structural characterization include docking simulations and absolute and relative free energy calculations. On the experimental front, we have highlighted oxidative footprinting of the receptor surface, which is emerging as a method to complement epitope-mapping data from techniques, such as STD NMR. Also, NMR strategies that use partially oriented samples, in combination with computational simulations, are a promising avenue for 3D oligosaccharide–protein structure determination.

Table IV.

Example PDB-style GLYCAM residue identifiers for glucose. Specifying linkage position (first position), monosaccharide residue (second position: one-letter code), D- or L-configuration (second position case), anomeric configuration (third position), and pyranose of furanose (third position case)

Linkage	Pyranoses				Furanoses
	D-Sugars		L-Sugars		D-Sugars		L-Sugars
	A-D-Glcp-	β-D-Glcp-	α-L-Glcp-	β-L-Glcp-	α-D-Glcf-	β-D-Glcf-	α-L-Glcf-	β-L-Glcf-
1-	1GA	1GB	1gA	1gB	1Ga	1Gb	1ga	1gb
4-	4GA	4GB	4gA	4gB	4Ga	4Gb	4ga	4gb
2,6-	XGA	XGB	XgA	XgB	Xga	XGb	Xga	Xgb
2,3,6-	SGA	SGB	SgA	SgB	Sga	SGb	Sga	Sgb

Abbreviations

CSA

chemical shift anisotropy

IL8

interleukin 8

LacNAc

N-acetyllactosamine (βGal(1-4)GlcNAc)

MC

Monte Carlo

MD

molecular dynamics

MS

mass spectrometry

RDC

residual dipolar coupling

SASA

solvent accessible surface area

STD NMR

saturation transfer difference nuclear magnetic resonance

TI

thermodynamic integration