Extension of the CHARMM Classical Drude Polarizable Force Field to N- and O-Linked Glycopeptides and Glycoproteins

Abstract

Molecular dynamic simulations are an effective tool to study complex molecular systems and are contingent upon the availability of an accurate and reliable molecular mechanics force field. The Drude polarizable force field, which allows for the explicit treatment of electronic polarization in a computationally efficient fashion, has been shown to reproduce experimental properties that were difficult or impossible to reproduce with the CHARMM additive force field, including peptide folding cooperativity, RNA hairpin structures, and DNA base flipping. Glycoproteins are essential components of glycoconjugate vaccines, antibodies, and many pharmaceutically important molecules, and an accurate polarizable force field that includes compatibility between the protein and carbohydrate aspect of the force field is essential to study these types of systems. In this work, we present an extension of the Drude polarizable force field to glycoproteins, including both N- and O-linked species. Parameter optimization focused on the dihedral terms using a reweighting protocol targeting NMR solution J-coupling data for model glycopeptides. Validation of the model include eight model glycopeptides and four glycoproteins with multiple N- and O-linked glycosylations. The new glycoprotein carbohydrate force field can be used in conjunction with the remainder of Drude polarizable force field through a variety of MD simulation programs including GROMACS, OPENMM, NAMD and CHARMM and may be accessed through the Drude Prepper module in the CHARMM-GUI.

Introduction

Atomistic molecular dynamics (MD) simulations of biomolecular systems have increasingly been part of biophysical studies to gather insights on various biological processes,¹ including multiple studies targeting carbohydrates.² The majority of MD studies to date have used empirical potential energy functions based on nonpolarizable, additive force fields due to the computational efficiency as well as availability. ^{3, 4} However, with the rapid improvements in the computational power and highly tailored algorithms we can now opt for higher accuracy in modeling of atomistic interactions, such as improved treatment of electrostatic interactions. To this end, there are multiple ongoing efforts to develop force fields that provide an explicit electronic polarization response. These include the AMOEBA model, the classical Drude oscillator model developed in our lab as well as other models developed in a number of laboratories.^{5, 6} Although requiring additional computational resources, polarizable force fields have proved that certain limitations of additive forces can be resolved by the explicit inclusion of polarizability.^{6, 7} Toward developing a comprehensive polarizable force field, we have extended the Drude force field to include parameters for carbohydrates,^8–11 molecular ions,^{12, 13} small molecules,¹⁴ nucleic acids,^{15, 16} atomic ions,¹⁷ lipids,¹⁸ and proteins.¹⁹ Details of the Drude energy function, parametrization strategies, and methodological information have been reported in previous studies.^{12, 16, 20}

Protein-carbohydrate interactions occur frequently in various biological processes such as metabolic activities, cell recognition and signaling, inflammation, among others.²¹ For example, enzymatic hydrolysis of lignocellulosic biomass is the most abundant protein-carbohydrate interaction in nature that recycles carbon. In addition, protein-protein interactions are also mediated through glycosylated entities on either one or both the proteins, for example, playing important roles in the immune system.²² Simulations of glycosylated proteins using nonpolarizable force fields have already been performed in multiple studies to get better insights about catalytic mechanisms, binding specificities and kinetic and thermodynamic preferences.²³ N- and O-linked glycosylations are the two most common types of glycosylation where the glycan is covalently linked to a protein via either the nitrogen of the Asn side chain (N-linked) or the oxygen of the Ser or Thr side chains (O-linked).

Post-translational protein modification by N-linked glycosylation, which occurs in the endoplasmic reticulum, plays a critical role in cell surface expression and is often required for protein stability and biological function.²⁴ It has been found that N-linked glycosylation generally occurs at the sequence Asn-X-Ser/Thr, where X is any amino acid except proline.²⁵ This type of glycosylation is found in nearly all eukaryotes,²⁴ and in most cases the linkage occurs between Asn and N-acetylglucosamine (GlcNAc), replacing the alcohol moiety of GlcNAc C1 with an amide linkage to the Asn side chain. The O-glycosidic carbohydrate–protein bond is an important biological linkage present, for example, in mucin glycoproteins²⁶ and as a post-translational modification for cytosolic proteins in signaling pathways.²⁷ O-linkages do not have a preferred sequence motif. To allow for polarizable simulations of glycopeptides and glycoproteins, parameters for these linkages are required.

In this work, parameters for the glycosidic linkage between carbohydrates and proteins in the context of the Drude polarizable force field are developed and validated. The linkages build on the previously parametrized polysaccharides ⁹ to include N- or O-glycosidic linkages to proteins, thereby allowing the simulation of carbohydrates that are important in biomolecular function and molecular recognition. In addition to the parameter optimization and validation presented in this study, the newly developed parameters were applied to characterize the structure and properties of O-glycosidic linkages in the absence of restraint data, thereby indicating their utility in molecular modeling of larger glycoproteins.

Computational Methods

Model Systems

Model compounds with a hexopyranose monosaccharide connected to N-acetyls or dipeptides were selected for parametrization. For O-linked models both Ser and Thr dipeptides were used (Figure 1, left panel). For the N-linked parameters, parametrization was initiated with the previously parametrized N-acetylglucoseamine model compound¹¹ with the N-acetyl group at the C1 position (Figure 1, top of right panel) or at both the C1 and C2 positions ((Figure 1, bottom of right panel). Quantum mechanical (QM) calculations were performed with the PSI4²⁸ and Gaussian 03²⁹ programs. Adiabatic dihedral potential energy scans (PES) were performed by geometry optimization at the MP2/6–31G(d) model chemistry with the target dihedral constrained followed by RIMP2/cc-pVQZ single point energies (i.e., RIMP2/cc-pVQZ//MP2/6–31G(d) QM model chemistry). Analogous adiabatic MM PES were then computed with the target dihedral constrained to the QM values with the targeted dihedral parameter force constants set to zero.⁹ The energy difference between the MM and QM (MM-QM) PES was then targeted for the initial fitting of the dihedral parameters using the least squares fitting program developed in this laboratory specifically for use in parameter optimization.³⁰ During fitting four multiplicities, n of 1, 2, 3, and 6, were included and the phase angle, δ, was limited to 0° and 180° as required to ensure applicability of the parameters to stereoisomers about a chiral center.

Figure 1.

Model compounds used in dihedral fitting of the O- and N-linkages. For O-linkage, an initial simple model compound based on THP-Thr/Ser without the hydroxyls on the hexopyranose ring was used (top of left panel) to perform the dihedral potential energy scans. Subsequently the full model compounds with the hydroxyls (bottom of left panel) were created based on conformations of the THP-based model compounds. Both α and β anomers of both the N- and O-glycan linkages were included explicitly in this study, therefore the linkages are represented with dotted line.

For validation eight O-linked model systems were selected based on availability of experimental data and previous studies using the additive CHARMM36 force field.^{31, 32} α and β conformers of N-acetylgalactosamine, β conformer of N-acetylglucosamine and β conformer of D-glucose sugars were built with a glycosidic linkage to both Ser and Thr dipeptides (Table S1). The systems were solvated in a 36 Å cubic water box. The Solution Builder and Drude Prepper modules in CHARMM-GUI were used to generate the simulation inputs for both nonpolarizable and polarizable systems.^{33, 34} Additionally, four glycoproteins with high glycosylation content were selected for bulk testing as shown in Table 1. The glycoproteins were also built using the CHARMM-GUI and were simulated in the NPT ensemble at 300 K for 500 ns each with the Drude FF using OpenMM 7.4.2. ³⁵ Standard MD inputs from Drude Prepper were used to run the simulations.

Table 1.

Glycoprotein Systems and Their Glycosylation Sites.^a

PDB	SER residues with O-linkage	THR residues with O-linkage	ASN residues with N-linkage
1K7C 36	-	-	104 (CARB), 182 (CARA)
1MYR 37	-	-	21(CARE), 90(CARF), 218(CARA), 244(CARG), 265(CARB), 292(CARC), 346(CARH), 361(CARD), 482(CARI)
1RMG38	368(CARC), 373(CARK), 380(CART), 383(CARM), 400(CARN), 409(CARG), 418(CARH)	367(CARS), 370(CARD), 371(CARQ), 372(CARL), 374(CARI), 376(CARP), 385(CARE), 386(CARR), 398(CARF), 405(CARJ), 408(CARO)	32(CARA), 299(CARB)
3GLY 39	443(CARC), 444(CARD), 453(CARE), 455(CARF), 459(CARI), 460(CARJ)	452(CARG), 457(CARH), 462(CARK), 464(CARL)	171(CARA), 395(CARB)

^aThe glycan at each glycosylation site is defined by segment identifiers in parentheses generated by the CHARMM-GUI, such as CARA, CARB, CARC, etc.

MD Simulation Settings

All molecular mechanics calculations for parametrization were carried out using the CHARMM program version c41a2 and with the polarizable Drude force field.^{10, 11, 40, 41} The water model is the polarizable SWM4-NDP model. ⁴² The validation MD simulations were carried out using OpenMM 7.4.2 with both additive CHARMM36 force field ^{4, 31, 43} and polarizable CHARMM Drude force field. CHARMM modified TIP3P water model was used in the additive simulations.⁴⁴ All polarizable MD simulations used a 1 fs time step while the additive simulations used 2 fs time step. All MD simulations were performed at 300 K, unless otherwise stated. All minimizations and MD equilibrations were performed following the previously presented protocol for lipids^{10, 41} and details of the integrators and treatment of nonbond interactions were those presented in our recent protein force field optimization study.¹⁹ Validation MD simulations to calculate NMR coupling constants were run for 0.5 μs each using both the additive and polarizable force fields.

J Coupling Equations

NMR heteronuclear three-bond proton-carbon coupling constants (in Hz) for the glycopeptide angle rotations ³J_HαHβ, ³J_HNHα, ³J_HΝH2 were computed from the simulations using modified Karplus eqs 1–5. ⁴⁵

$$\begin{gathered} 3J_{HnH\alpha }=6.51co{s}^2\varphi -1.76cos\varphi +1.60 \\ \varphi =(CY-N-C\alpha -C)-{60}^{°} \end{gathered}$$ 1

$$\begin{gathered} 3J_{HnH2}=9.6co{s}^2\varphi -1.51cos\varphi +0.99 \\ \varphi =(HN-N-C2-)H2+{60}^{°} \end{gathered}$$ 2

$$\begin{gathered} 3J_{H\alpha H\beta }=9.5co{s}^2\varphi -1.6cos\varphi +1.80 \\ \varphi =(N-C\alpha -C\beta -O1)-{120}^{°}forThr \end{gathered}$$ 3

$$\begin{gathered} 3J_{H\alpha H\beta 2}=9.5co{s}^2\varphi -1.6cos\varphi +1.80 \\ \varphi =(N-C\alpha -C\beta -O1)forSer \end{gathered}$$ 4

$$\begin{gathered} 3J_{H\alpha H\beta 3}=9.5co{s}^2\varphi -1.6cos\varphi +1.80 \\ \varphi =(N-C\alpha -C\beta -O1)-{120}^{°}forSer \end{gathered}$$ 5

To compare J coupling between published experimental data and ensemble averages from the MD simulations, the value of the dihedral angle of interest was first calculated for each snapshot of each glycopeptide in the simulation for which the corresponding J coupling value was calculated using the above equations. The resulting J coupling values for the dihedral angle of interest were then averaged for all snapshots from the 0.5 μs MD simulations (coordinates saved every 10 ps) yielding the presented ensemble average. The convergence of J coupling was evaluated by dividing the trajectories into ten 50 ns blocks.

Parametrization Using Reweighting Targeting Experimental Condensed Phase Data ⁴⁶:

In parameter optimization through reweighting we minimize the target function, f,

$$f=\frac{1}{\beta }ln\tau +w\mathrm{*}{RMSE}_{\lambda ,\Delta \lambda }$$ 6

where τ represents the probability of conformations reproducing the experimental observable A, RMSE is the root-mean-square error between the initial and new parameters, λ and Δλ, respectively, and w is an assigned weighting factor to balance the contributions of reproduction of the experimental observables and the change in the parameters during the reweighting procedure. β is the Boltzmann constant multiplied by the temperature. The ensemble average, indicated by brackets, < >, of the targeted experimental properties, A, associated with a new force field parameter set λ + Δλ is determined from

$$\langle A_{\lambda +\Delta \lambda }\rangle =\frac{\langle A_{\lambda }{e}^{-\beta (E_{\lambda +\Delta \lambda }-E_{\lambda })}\rangle }{\langle e^{-\beta (E_{\lambda +\Delta \lambda }-E_{\lambda })}\rangle }$$ 7

, in which E_λ and E_{λ + Δλ} are the potential energies with parameters λ and λ + Δλ for each sampled conformation. Sampling of parameter changes is performed using Metropolis Monte Carlo sampling and the RMSE between the initial and new parameter is calculated based on eq 8 where n is the number of parameters being optimized.

$${RMSE}_{\lambda ,\Delta \lambda }=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left({\lambda }_i^{new}-{\lambda }_i^{initial}\right)}^2}$$ 8

Results and Discussion

Parametrization Strategy

The initial electrostatic parameters, including the partial atomic charges, atomic polarizabilities (ALPHA values) and atomic Thole scaling factors (THOLE values), were transferred directly from the full monosaccharides, the available glycosidic linkages, and the proteins. Details of the optimization of the electrostatic parameters applied to the N- and O-linkage nitrogen and oxygen atoms, respectively, were presented in previous work.^{9, 11}

N-Glycans.

To develop the Drude force field parameters for the N- linkage, hexopyranose with an N-acetyl moiety substituted at C1 was used. Both α and β anomers of the linkage were treated explicitly. Most of the initial bond, angle, and dihedral parameters were readily transferred from the N-acetyl substituent at the C2 position, as previously developed for GlcNAc and GalNAc. The missing dihedral parameters (C–N–C1–C2/O5, N–C1–C2/O5-C3/C5) were optimized targeting a 1D QM scan at the MP2/6–31G*//RIMP2/cc-pVTZ model chemistry of the C–N–C1–C2 dihedral. The dihedral parameters were optimized using a least squares fitting protocol to minimize the root-mean-square (RMS) error between the QM and MM energy surfaces. Energy surfaces were offset to zero at the lowest energy conformation over all model compound anomers, and the conformations with relative QM energies of > 12 kcal/mol were excluded during the fitting. The refined dihedral parameters obtained from the least squares fitting protocol were able to reproduce the low-energy profiles for both α- and β-anomers of N-acetyl derivative (Figure 2). The final RMSE for the Drude force field was 1.46 kcal/mol versus 2.70 kcal/mol for the additive C36 force field indicating the ability to yield improved conformational properties due to the presence of explicit polarization in the Drude model. Potentially contributing to the difference in the level of agreement of the Drude and C36 FFs with the QM data is the use of different model compounds and energy scans as target data for the optimization, although both efforts used the QM (RI)MP2/cc-pVTZ//MP2/6–31G(d) model chemistry with the only difference being the use of canonical MP2 with C36 versus RIMP2 with the Drude for the treatment electron correlation in the single point calculations.

Figure 2.

Relative conformational energies of selected conformations of hexopyranose with N-acetyl substituted at C1 for both the α and β anomers. All the model compound conformations with relative QM energies below 12 kcal/mol are included. RMSE of Drude and additive C36 relative energies with respect to QM are included in parentheses shown in the inset.

O-Glycans.

To develop Drude force field parameters for O-glycosidic linkages, the α and β anomers of Ser- and Thr-linked hexopyranose monosaccharide were used. Initial values for bonded and nonbonded parameters were transferred from the carbohydrates ^8–11 and proteins ⁴⁰. For each linkage type there are a total of 32 possible diastereomers associated with the eight possible diastereomers for the monosaccharide and two possible amino acids (namely, Ser and Thr) times the α and β anomers for the monosaccharide (8×2×2×1=32). Since the number of possible 2D scans is huge and computationally expensive, it was necessary to prepare target data on a practical number of energy surfaces. Accordingly, QM scans were performed on the THP-Thr/Ser model compounds shown in Figure 1 (top left panel). Besides the computational cost of omitting the hydroxyl groups, use of this simple compound can reduce the occurrence of gas phase hydrogen bonds between the carbohydrate and amino acids that can complicate the energy surfaces. In addition to the THP-Thr/Ser QM 2D φ/ψ scan, QM data were also generated by scanning the χ glycosidic dihedrals every 15°. Since the amino acid φ/ψ dihedrals are important for the conformation of the protein, for both 2D φ/ψ and χ, the protein φ/ψ dihedrals were kept close to the values from the distributions shown in Figure 3. The carbohydrate ring oxygen was constrained in three different orientation (at −60°, 0°, and 60° as shown in Figure S1) with respect to the amino acid HN for the χ scan and χ was constrained at −60°, 0°, and 60° for the glycopeptide linkage φ/ψ scan with THP-Thr/Ser. The QM calculation involved geometry optimization in the presence of the constraints for the model compounds at the ab initio MP2/6–31G(d) model chemistry. The final QM conformations of the THP-Thr/Ser model compounds with relative energies lower than 8.0 kcal/mol were then used to construct the full glycopeptide linkage model compound with hydroxyls shown in Figure 1 (bottom left panel). All generated conformers were then subjected to QM optimization and single point energy calculations to generate the final target data for the fitting shown in Figure 4. Since the O linkage based on Ser shares the same atom types with p1→6p glycosidic linkages that were previously parametrized, the parameters were transferred from existing glycosidic linkage ⁹ and only missing parameters were optimized. Table S2 includes a list of the optimized parameters along with the final values obtained in this study. As with the N-glycan conformational energies, there is significant improvement with the RMSE for the Drude FF being 1.38 kcal/mol versus 2.45 kcal/mol for the additive C36 FF (Figure 4). This again points to the presence of explicit polarization in the Drude FF leading to improved treatment over the additive FF.

Figure 3.

2D protein φ/ψ distribution of target data used during the O-linked dihedral fitting. The distribution is overlaid on the structure of the full O-linked Thr model compound (Figure 1, bottom left panel).

Figure 4.

Relative conformational energies of selected conformations of the hexopyranose monosaccharide linked to the Ser and Thr dipeptides for both the α and β anomers. All the model compound conformations with relative QM energies below 12 kcal/mol are included. RMSE of Drude and additive C36 relative energies with respect to QM are included in parentheses shown in the inset.

Reweighting-Based Parameter Optimization Targeting Solution NMR Data

After the initial optimization of the dihedral parameters targeting the QM data, MD simulations of the target glycopeptides were performed. NMR J-coupling data calculated from these simulations did not reproduce J values very well (shown as Drude_QM in Figure 5). We therefore performed a reweighting optimization on the new parameters targeting the experimental J-coupling data associated with the protein backbone. The reweighting procedure was adopted from a previously used reweighting protocol as described in eqs 6–8.⁴⁶ Using the target dihedral and corresponding J values from the initial trajectories small perturbation to the dihedral force constant was undertaken and then converted to the observable J values using equations 1–5. A Monte Carlo simulated annealing (MCSA) protocol was combined in the context of reweighting eqs 6–8 to drive the optimization against the target experimental value. The reweighting follows the standard Metropolis acceptance criteria for updating dihedral force constants. An additional constraint was included as a penalty to avoid overfitting. This was based on the RMSE between the perturbed and original values of the dihedral force constants, eq 8, thereby limiting the change in the parameters from the initial values from the QM fitting. The final parameter sets therefore represent an agreement of improved reproduction of the experimental J values and minimum perturbation from initial QM-based dihedral parameter values (Table S2).

Figure 5.

³J_HαHβ vs φ (protein) distribution obtained from MD. ³J_HαHβ calculated based on eq 1 for every 1° form −180° to 180°. “QM” indicates the initial QM-based dihedral parameters.

Validation Runs

Final parameters for the O-linkages were validated by calculating J-coupling values based on trajectories generated using reweighted dihedral parameters. The eight model compounds with O-linkages to Ser/Thr were simulated with the optimized Drude force field as well as CHARMM36 for 0.5 μs each. The average J coupling values calculated using eqs 1–5 over the trajectories are shown in Table 2. The Drude J coupling values for backbone φ are in agreement with the experimental values with the RMSE increased slightly compared to C36. Figure 6 shows the distribution of the φ dihedral (C-N-CA-C) in SER/THR-BGLCNA along with the average values calculated every 50 ns. Similar distributions for other model compounds are provided in the Supporting Information (Figures S2–S4). In the case of glycopeptides with N-acetyl side chains, the J coupling values for the θ dihedral (HN-N-C2-H2) are consistently lower than experimental values with RMSE increased significantly. However, the majority of the conformations sampled are between 120° and 240° (Figures 7, S5, and S6) that correspond to the anti conformation observed in the experimental data. Note that the φ and θ dihedral parameters were optimized in a prior studies¹¹ and are only analyzed for validation. Most importantly, the J coupling values for the χ dihedrals (N-CA-CB-OG/OG1) optimized in this study show good agreement with the experimental data. Lower RMSE was obtained with the Drude FF relative to C36 in the case of J coupling calculated with eq 4 for the SER models. Figure 8 shows the dihedral distribution for SER/THR-BGLCNA, while distributions for other six model compounds are shown in Figure S7–S9.

Table 2.

J Coupling Constants for Different Dihedrals from the Model Compounds from Experiment, the Drude Force Field and the C36 Additive Force Field.^a

		ϕ (CY-N-CA-C)			θ (HN-N-C2-H2)			χ (N-CA-CB-OG)
compd	name	expt	C36	Drude	expt	C36	Drude	expt	C36			Drude
									Eq4	Eq5	Eq3	Eq4	Eq5	Eq3
1	SER-AGALNA	6.2	7.6	7.0	9.2	9.9	7.2	5.5, 4.5	9.4	5.6	-	10.4	3.0	-
2	THR-AGALNA	8.8	7.5	6.4	9.5	9.9	7.5	2.5	-	-	3.4	-	-	2.7
3	SER-BGALNA	6.6	7.7	6.9	9.6	9.6	7.3	6.8, NA	9.6	5.7	-	5.6	2.8	-
4	THR-BGALNA	7.4	7.6	5.4	9.9	9.9	7.5	3.5	-	-	3.3	-	-	9.8
5	SER-BGLC	6.9	7.7	6.9	-	-	-	4.6, NA	9.3	5.3	-	4.4	2.8	-
6	THR-BGLC	7.5	7.5	5.8	-	-	-	3.4	-	-	3.7	-	-	8.2
7	SER-BGLCNA	6.8	7.7	6.8	9.4	9.8	7.0	6.6, 5.2	9.6	5.9	-	4.8	2.9	-
8	THR-BGLCNA	7.5	7.5	5.8	9.4	10.0	7.3	-	-	-	5.6	-	-	7.4
	RMS diff b	-	0.88	1.42	-	0.43	2.21	-	3.67	-	-	2.69	-	-

^aNA – Not available

^bRoot mean square difference was calculated for values only if experimental data are available for all relevant systems.

Figure 6.

J-coupling values for the protein (φ) dihedral C-N-CA-C for the model compounds SER-BGLCNA and THR-BGLCNA. (A) and (C) represent the Drude FF results. (B) and (D) represent the additive C36 FF results. The results for the other six compounds are provided in the Supporting Information.

Figure 7.

J-coupling values for the dihedral HN-N-C2-H2 for the model compounds SER-BGLCNA and THR-BGLCNA. (A) and (C) represent the Drude FF results. (B) and (D) represent the additive C36 FF results. The results for the other four compounds are provided in the Supporting Information.

Figure 8.

J-coupling values for the dihedral N-CA-CB-OG for the model compounds SER-BGLCNA and THR-BGLCNA. (A) and (C) represent the Drude FF results. (B) and (D) represent the additive C36 FF results. With THR-BGLCNA experimental data is not available. The results for the other four compounds are provided in the Supporting Information.

Four glycoproteins were studied with multiple O- and N-linked glycosylations by starting from the crystal structures reported in Table 1. Analysis of the trajectories was performed on the conformational variability of the different carbohydrates and proteins. Trajectories were aligned to starting structure with respect to protein backbone atoms before RMSD calculations. RMSD of all carbohydrates grouped together compared to RMSD of all protein showed that the glycan portions are sampling large regions of conformational space around the glycosylation site. Figure S10 shows the RMSD of protein, carbohydrate, and protein+carbohydrate for all four systems over the 0.5 μs simulations. RMSD of the proteins remain lower than 3 Å for all systems except for PDB code 1K7C where it increased to 4 Å gradually from 2 Å during 100–300 ns and then stabilized. The RMSD of the carbohydrates varied from 6 to 16 Å across the systems. Such high flexibility of the glycans is consistent with high conformational diversity of carbohydrates as observed in NMR and crystallographic studies. Contributing to the magnitude of the RMSD is the presence of multiple glycans on each protein, where the RMSD is overall the glycans listed in Table 1. Analysis was also performed on the dihedral angle distributions of the glycosidic linkages by pooling the MD simulations for all systems. Figure 9 shows that for the N-glycans the glycosidic dihedral angle distributions are in good agreement with the values from the crystal structures. The glycosidic linkage dihedral O5-C1-ND2-CG shows two minima around 60° and 300°, while closer to the protein the dihedral C1-ND2-CG-CB shows a narrow distribution around 180°. The only region sampled in the MD simulations but not observed in the crystal structure occurs with the O5-C1-ND2-CG dihedral in the region around 60°.

Figure 9.

Boltzmann inverted glycosidic dihedral angle distributions associated with all the N-glycans from all four glycoproteins. A) C1-ND2-CG-CB vs. O5-C1-ND2-CG, B) ND2-CG-CB-CA vs. C1-ND2-CG-CB, and C) CG-CB-CA-N vs. ND2-CG-CB-CA. Black dots indicate the values of the dihedrals observed in crystallographic structures.

Similarly for the O-glycans dihedral sampling near the glycosidic linkage is in very good agreement with the experimentally observed values in the crystal structures. Figure 10 shows the Boltzmann inverted dihedral distributions for the pooled data of the dihedrals for the O-linked glycans from PDB codes 1RMG and 3GLY. Both the SER and THR linked O-glycans show three clear minima for the dihedral OG-CB-CA-N sampling the g(+), g(−), and anti conformational states. Importantly, the glycosidic linkage dihedral O5-C1-OG-CB for both SER and THR linked glycans shows a minima around 60° for α anomers and another minima at 300° for β anomers. A broader sampling of the dihedral C1-OG-CB-CA was observed in SER linked glycans compared the same in THR linked glycans. Overall, these results indicate that the developed parameters are sampling conformations of the glycosidic linkages consistent with those observed in glycoprotein crystal structures.

Figure 10.

Boltzmann inverted glycosidic dihedral angle distributions associated with all the O-glycans from all four glycoproteins. (A) C1-OG-CB-CA vs. O5-C1-OG-CB in Ser O-linkages, (B) OG-CB-CA-N vs. C1-OG-CB-CA in Ser O-linkages, (C) C1-OG1-CB-CA vs. O5-C1-OG1-CB in Thr O-linkages, and (D) OG1-CB-CA-N vs. C1-OG1-CB-CA in Thr O-linkages. Black dots indicate the values of dihedrals observed in crystallographic structures.

The presence of explicit polarization in the Drude FF allows for variations in the dipole moments beyond those due to changes in intramolecular geometry. Accordingly, dipole moment analysis was performed on each monosaccharide by selecting the groups of atoms with integer charge. Figure S11 shows the atom selection scheme. More details about selection of atoms appropriate calculation of dipole can be obtained in the CHARMM-GUI Drude Prepper article.³⁴ We observed that most of the cases with N-glycans containing the N-acetyl glucosamines (BGLCNA) at positions 1 and 2 showed slightly different dipole moments going from 1 to 2. In some cases, as shown in Figure 11 for N-glycosylation at Asn265 for PDB code 1MYR, the BGLCNA residues show a bimodal distribution of the dipole moment. Analysis of the time series shows the dipole moment of the BGLCNA monosaccharides to undergo multiple transitions between the lower and higher dipole moments. Similar behavior is observed for the majority of monosaccharides for which a bimodal distribution exist, though exceptions are present (Figure S11). Such dipole variability indicates that the Drude FF will be of utility to investigate the role of changes in the electronic structure of glycopeptides and glycoproteins in biologically relevant scenarios.

Figure 11.

Monosaccharide dipole moment analysis for the N-glycan of PDB code 1MYR at Asn265 (segment ID: CARB assigned by CHARMM-GUI, see Table 1). Changes in dipole moment across the simulation are shown on left. On right, the normalized probabilities of dipole moment are shown for each monosaccharide.

Conclusions

The present study performs force-field parameter optimization of glycopeptide moieties extending the Drude polarizable force field to glycopeptides and glycoproteins. Reweighting of parameters allowed for fine-tuning of the force field to reproduce the target solution properties such as J coupling with more accuracy than direct transfer of parameters based on gas phase QM data. This extension will allow for studies of the role of post-translations modifications involving the O- and N-linked glycosylation on various biological phenomena including protein-protein interactions and protein stability using molecular simulations. Currently the PDB database has more than ~9400 protein structures that exhibit O- and N-glycosylation at multiple sites indicating the range of systems to which the Drude force field may now be applied. In addition, this extension will allow use of a polarizable force field in the study of intrinsically disordered peptides that use glycosylation to protect against proteolysis.⁴⁷