Toward Biotherapeutics Formulation Composition Engineering using Site-Identification by Ligand Competitive Saturation (SILCS)

Abstract

Formulation of protein-based therapeutics employ advanced formulation and analytical technologies for screening various parameters such as buffer, pH, and excipients. At a molecular level, physico-chemical properties of a protein formulation depend on self-interaction between protein molecules, protein-solvent and protein-excipient interactions. This work describes a novel in silico approach, SILCS-Biologics, for structure-based modeling for protein formulations. SILCS Biologics is based on the Site-Identification by Ligand Competitive Saturation (SILCS) technology and enables modeling of interactions among different components of a formulation at an atomistic level while accounting for protein flexibility. It predicts potential hotspot regions on the protein surface for protein-protein and protein-excipient interactions. Here we apply SILCS-Biologics on a Fab domain of a monoclonal antibody (mAbN) to model Fab-Fab interactions and interactions with three amino acid excipients, namely, arginine HCl, proline and lysine HCl. Experiments on 100 mg/ml formulations of mAbN showed that arginine increased, lysine reduced, and proline did not impact viscosity. We use SILCS-Biologics modeling to explore a structure-based hypothesis for the viscosity modulating effect of these excipients. Current efforts are aimed at further validation of this novel computational framework and expanding the scope to model full mAb and other protein therapeutics.

INTRODUCTION

Biotherapeutics (aka protein therapeutics, biologics, biopharmaceuticals) represent an important and growing class of highly specific biologically active macromolecules. ^1–3 During drug product development of biotherapeutics, maintaining physical and chemical stability is of paramount importance during long term storage. The solution properties and long term stability of the biotherapeutics depend on several factors such as 3D conformational structure, ¹ flexibility, ⁴ global and local dynamics, ^5–7 dispersive interactions, ^8–10 volume, ^11,12 hydration, ¹³ compressibility, ^5,11,13–15 and stability ¹⁶ of the protein in solution. These solution properties and protein stability can be modulated by temperature, ⁵ pressure ¹⁷ and the components ⁵ of a formulation such as buffer(s) and excipient(s) ¹⁸ and its compositions. ^5,19 Long term stability is particularly challenging for high concentration protein formulation for high dose administration. Additional consideration at high concentration proteins is to achieve low solution viscosity for ease of production and patient administration. ^16,20,21 Currently formulation of protein-based therapeutics in pharmaceutical industry employs advanced experimental designs and high throughput experimental technologies for formulation screening and analytics. ^22–24 Systematic addition of one or more excipients ¹⁹ in varying concentrations is a part of advanced protein formulation development, ²⁵ which invariably involves expensive and time-consuming screening experiments.

Currently, formulation screening experiments are designed empirically based on historic knowledge with limited understanding of molecular level interactions of proteins with its formulation components at early stages of development. At a molecular level, however, physico-chemical properties, stability (physical and chemical) and solution viscosity of a protein in a formulation composition depends on the intrinsic three-dimensional (3D) structure and dynamics as well as interactions with other protein, ^26–28 solvent and excipient ^29–31 molecules. Multiple computational tools have been developed to correlate sequence and structure of a protein with solution properties such as aggregation ³² and viscosity. ³³ Recently molecular dynamics simulations have been employed to directly model protein-excipient interactions while accounting for solvent dynamics and protein flexibility using a fully atomistic ^34–36 or a coarse-grained³⁷ description of the protein. The preferential interaction parameter (Γ) quantifies the extent by which the presence of a protein perturbs the concentration of an excipient in solution due to protein-excipient interaction and excipient-excipient interaction.^38,39 While Γ may be calculated directly from MD simulations, ^35,40,41 this as well as other MD-based methods typically require individual, extensive simulations for each excipient-protein pair as well as combination of excipients making them impractical for a screening campaign.

This manuscript introduces a novel computational framework, SILCS-Biologics, to build a physics-based model of a protein formulation accounting for the 3D structure and dynamics of protein and its interactions with other protein, excipient and solvent molecules in the formulation. It builds on the SILCS (Site-Identification by Ligand Competitive Saturation) technology ^42–44 which has been successfully applied for affinity ranking of small molecules binding to protein targets ^45–49 as well as to identify protein surface residues likely to be involved in protein-protein interaction with self or other proteins. ⁵⁰ SILCS simulations involves explicit water mixed solvent simulations which combine oscillating chemical potential Grand Canonical Monte Carlo and molecular dynamics simulations. ^42–44 In the SILCS framework, the simulations to capture the dynamics of the protein need to be performed only once and the information generated from these simulations enables rapid calculations of interactions of other proteins ⁵⁰ and small molecules ⁵¹ with the whole solvent exposed regions of the protein.

The work presented in this manuscript employs SILCS-Biologics to study protein-protein and protein-excipient interactions using the Fab domain, a prototype model system, of a therapeutic monoclonal antibody (mAbN). mAbN was targeted for high concentration protein development and solution viscosity was one of the challenges in its development. Viscosity of mAbN formulations was measured with respect to mAbN concentration in the absence and presence of potential viscosity modulating excipients, arginine HCl, proline and lysine HCl. At the formulation concentration of 100 mg/ml of the mAbN, arginine HCl increased, lysine HCl reduced and proline had no impact on solution viscosity. In the absence of any excipients (control) the viscosity of mAbN was measured at ~ 5.9 cP at 100 mg/ml, higher than traditional well-behaved monoclonal antibodies. We compute and overlay the Fab-Fab interaction hotspots and the surface interaction (“differential coating”) pattern for the three excipients. SILCS Biologics provides a computationally efficient method for in silico excipient screening to identify all potential binding sites to generate a global map of how an excipient may noncovalently interact with the surface of a protein. ^50,51 This information is used to explore a molecular mechanism for the observed differential effect of the excipients on mAbN surface. The SILCS-Biologics framework provides a systematic approach for advanced protein formulation development by enabling high throughput in silico screening of a library of excipients, alone and in combinations, aimed at modulating protein surface chemistry and/or protein-protein interactions for biotherapeutics formulation composition engineering.

METHODS

SILCS Simulations and FragMap Generation

The SILCS simulations were initiated using an in-house X-ray crystal structure of the Fab domain of mAbN. SILCS simulations were performed using the SilcsBio software package (provided by SilcsBio, LLC) with GROMACS ⁵² as a molecular dynamics (MD) simulation engine. The SILCS simulation protocol is previously described. ⁴⁸ Briefly, the simulation systems involving protein, water and eight probe molecules, namely benzene, propane, methanol, formamide, acetaldehyde, imidazole, methylammonium and acetate were prepared. The probe molecules are selected to represents diverse functional groups that are common in small organic molecule to protein interaction, such as apolar, hydrogen bond donor and acceptor, and positive and negative charge interactions. The probe selection and the validation of the SILCS approach in reproducing protein-ligand interaction has been investigated in other publications.^43,53,54 Ten independent simulation systems with the probe molecules at approximately 0.25 M concentration each were prepared, where each probe molecule is distributed randomly, and the sidechains of solvent-exposed residues were systematically varied to improve convergence. Protein protonation states were determined at the experimental pH of 5.6. Each system was equilibrated using 1 ns of MD simulation, followed by 100 cycles of grand canonical Monte Carlo (GCMC)/MD simulations. During each cycle, 200,000 steps of oscillating μ_ex GCMC simulation, ^44,55 which drives the sampling of probe and water molecules, followed by 1 ns of MD simulation was performed. The cycle was repeated over 100 cycles making the aggregated simulation time 1 μs (10 ⨉ 100 ns). Weak harmonic restraints with a force constant of 0.12 kcal/mol/Ų were kept on all Cα atoms of protein during MD simulations. For the Fab domain studied here, each individual 100 ns GCMC/MD simulation took three days on a NVIDIA GTX 1080Ti GPU card, with the full collection of simulations completed in that time frame using a cluster of 10 machines.

The protein conformations and distributions of water and solutes were saved every 10 ps during the MD simulation phase for analysis. The protein, solutes and water were described using the CHARMM36m protein force field, ^56,57 the CHARMM General force field (CGenFF) ⁵⁸ and the TIP3P water model modified for the CHARMM force field, ⁵⁹ respectively. For MD simulations, hydrogen bond lengths were constrained using the LINCS algorithm ⁶⁰ with an integration time step of 2 fs. Long-range electrostatic interactions were handled with the particle-mesh Ewald method ⁶¹ with cutoff of 8 Å and maximum grid spacing 𝜅 = 1.2 Å with a fourth-order spline. Long-range Lennard-Jones interactions were handled with an isotropic long-range dispersion correction method ⁶² with a cutoff of 8 Å, and a force-switching function was applied between 5 and 8 Å cutoff for a smooth transition. During the simulation, the temperature and the pressure of the system were maintained at 298 K and 1 bar using a Nose-Hoover thermostat ^63,64 and Parrinello-Raman pressure control ⁶⁵.

The FragMaps were generated by binning the selected solute atoms into voxels of a 1 Å spaced grid spanning the simulation system. The voxel occupancy can be used to measure a grid-based free energy (GFE) ⁵³ using the equation $\Delta G_i=-RT\ln (N_i/N_0)$, where N_i is the observed voxel occupancy of the probe at grid point i, and N₀ is the expected voxel occupancy of the probe alone in the bulk for a given concentration. Therefore, GFE is a measure of the free energy change of moving an atom from the bulk state to the grid point i. For example, if the GFE of a voxel near protein is −1.5 kcal/mol, then the probe atoms are about 12 times more likely to be found in that voxel than in the voxels that are far away from the protein (“bulk”) at the simulation temperature. In addition to the FragMaps for the probe atoms, an Exclusion map is also computed to capture the core of the protein that is the region of the protein never sampled by the non-hydrogen atoms of the water or probe molecules.

SILCS-PPI: Protein-Protein Interaction Using SILCS FragMap

A global protein docking was performed using the SILCS FragMaps to determine the PPI preference of the Fab using the SilcsBio software package (SilcsBio, LLC). The SILCS simulation protocol, namely SILCS-PPI, is designed to maximize the complementarity between the FragMaps from the receptor protein and the sidechain probability maps from the ligand protein and vice versa. ⁵⁰ For the SILCS-PPI protocol, two set of maps are required, SILCS FragMaps and the protein side chain probability grid maps (PPGMaps), which can be computed from the same MD trajectories. The PPGMaps are computed following the same protocol as FragMaps, i.e., binning the selected atoms (see Table 1 from ref ⁵⁰ for the selected atoms) into 1 ⨉ 1 ⨉ 1 ų voxels, but they are normalized by the maximum occupancy values so they have ranges from 0 to 1, whereas FragMaps are Boltzmann weighted. Both types of maps are pre-processed to reduce sampling noise as follows. For the FragMaps, all voxels that are 5 Å away from the excluded area (i.e., protein interior) as defined by the SILCS Exclusion map are removed since those grids are influenced by bulk-phase behavior and may only add noise to the sampling. For the ligand PPGMaps, all voxels that are overlapping with the Exclusion map are removed since they represent the repulsive core of the protein and only PPGMaps that are accessible to the solvent or SILCS probe molecules are useful for PPI determinations.

Table 1.

Viscosity at 100 mg/ml as a function of excipient and statistics from SILCS analysis

	Actual Conc. (mg/mL)	True Viscosity (cP)	SILCS #PPI blocking poses	SILCS #Exclusive Binding Sites
Control	101.0	5.8	-
Arginine	105.6	8.6	10	2
Lysine	102.2	4.6	14	2
Proline	105.0	5.8	7	1

The protein-protein docking is performed globally using an FFT-based algorithm. The receptor FragMaps and PPGMaps are fixed in space and the orientation of ligand FragMaps and PPGMaps are systematically sampled over all possible orientations with 10˚ rotational angle interval by default. The rotation angles are uniformly distributed to avoid the biases around the poles, ^66,67 which resulted in 14,904 possible orientations with 10˚ rotational angle interval. For each orientation of ligand protein, all possible translations of ligand FragMaps/PPGMaps can be efficiently evaluated using FFT operations. The overlapped voxels from FragMaps and PPGMaps are multiplied and summed to yield the protein grid free energy (PGFE) score. For each orientation, the top 10 scoring solutions are stored for further analysis by default, which resulted in 149,040 docked conformations. Details of the scoring function and FFT procedure can be found in reference ⁵⁰.

FragMaps and the PPGMaps of the Fabs are used to calculate the a Fab-Fab self-interaction preference map. A clustering analysis was performed on the total 149,040 docked poses. Due to the large number of poses, the center of mass (i.e., x, y, and z position) and the orientation (i.e., Euler angles) of the “ligand” protein were used to compute the distance between the ligand poses. A stepwise clustering analysis was performed where the clusters were identified using the center of mass of the ligand pose followed by another clustering analysis using the three Euler angles of decoys that belong to the same cluster. The distance in angular space is measured by $d_{12}=\sqrt{{\sum }_{\theta }^{\alpha ,\beta ,\gamma }{\left(\cos \left({\theta }_1\right)-\cos \left({\theta }_2\right)\right)}^2+{\left(\sin \left({\theta }_1\right)-\sin \left({\theta }_2\right)\right)}^2}$, where 𝛼, 𝛽, and 𝛾 represent the three Euler angles. Cluster cutoffs of 10 Å and 0.5 (approximately 30˚) were used for the center of mass and the Euler angles, respectively.

For the present system, maps derived from the simulations of the Fab of mAbN was used for both receptor and ligand. This results in docked conformations that are sterically incompatible with the full-length antibody. These conformations are filtered by counting steric clashes between two full-length IgG1 antibody models (PDB:1HZH) ⁶⁸ after they are overlaid with respect to the corresponding domains in the decoy conformation. This filtering further reduces the number of poses to 2,045 which were selected for further analysis. At this stage, to calculate the PPI preference for each pose, the contacts between the receptor and ligands atoms within a 5 Å cutoff are counted and summed over poses. The final PPI preference value is assigned by normalizing the number of contacts per residue by the maximum number of contacts. This value is called the PPI preference (PPIP) and higher values suggests a residue is more likely to be involved in a PPI.

SILCS-Hotspots: Excipient Binding Site Identification Using SILCS-MC

Excipient docking and screening in SILCS-Biologics uses a modified version of the a SILCS-MC, ⁴³ docking algorithm which has recently been extended into the SILCS-Hotspots approach. ⁵¹ Briefly, SILCS-MC algorithm involves Monte-Carlo sampling of the ligand in translational, rotational, and torsional space in the field of FragMaps. The energy of a ligand conformation is evaluated by the combination of CGenFF intramolecular energies and Ligand Grid Free Energy (LGFE) score, which is the sum of atomic GFEs. The atomic GFE for each ligand atom is assigned by the FragMap voxel that the atom occupies. The protein structure is not explicitly included, but the Exclusion map prevents the ligand from sampling the region where no probe or water molecules visited during the SILCS simulations, i.e., the interior of the protein. This allows for rapid docking of the ligand while accounting for protein side chain flexibility in a mean-field fashion, as that information is embedded in the FragMaps and the Exclusion map. LGFEs have been shown to correlate well with the binding affinities of small, drug-like molecules to a range of proteins. ⁵⁴

In the excipient docking procedure, the FragMap space is divided into 14.14 × 14.14 × 14.14 ų boxes. In each box, the following SILCS-MC sampling is performed for every excipient molecule. The excipient is first randomly positioned within a sphere of radius 10 Å centered at the middle of the sampling box. The excipient is then subjected to 10,000 steps of MC sampling at 300 K. During this phase of sampling, the molecular translations, orientation, and the torsion angle of an internal rotatable bond can change by up to 1 Å, 180˚, and 180˚, respectively. This is followed by 40,000 steps of simulated annealing MC sampling, where the temperature of the system is lowered from 300 K to 0 K. During the annealing phase, the molecular translations, orientation, and the torsion angle of an internal rotatable bond can change by up to 0.2 Å, 9˚, and 9˚, respectively. This process is repeated 1,000 times for each excipient in each sampling box. The results from individual sampling boxes are pooled together and clustering analysis was performed as follows. A center-of-mass based clustering with a 3 Å cluster radius is performed with a simple clustering algorithm ⁶⁹. The cluster radius of 3 Å was determined empirically. Once the cluster centroids for each excipient are identified, a second round of clustering is performed to identify the sites populated by different excipients. This is performed using the same clustering algorithm and a radius of 4 Å from which binding sites are identified that contain one or more members from the collection of excipients under study, e.g., the binding sites where one or more types of excipient bind. To rank the clusters, we used the average LGFE score of the cluster, which was computed by the average of the cluster members. This yields a list of excipient binding sites ranked by average LGFE scores of the excipients at the binding site. This list can be further analyzed to classify binding sites that are favored by different excipients (mutual binding site) as well as the sites that are exclusive to certain excipients (exclusive binding site). It should be noted that the site identification and the excipients included in each site are sensitive to the two clustering radii. Cluster radii that are too small will result in too many binding sites and cluster radii that are too large may lump distinctive clusters into one, resulting in loss of information. The default clustering radii was used in our previous validation study. ⁵¹

SILCS-Biologics for mAbN

Figure 1 shows the computational workflow that comprises SILCS-Biologics. The calculations were initiated with the crystal structure of the Fab domain of mAbN. If an experimental 3D structure of the Fab or Fc is not available, homology models can be used. While Fc domain can be reliably modeled due to high sequence conservation and large number of available structures, homology modeling of Fab domains is steadily improving ^70,71. The Fab crystal structure was used to perform SILCS simulations and FragMap generation as described in Section B.1. This was followed by SILCS-PPI simulations to compute the Fab-Fab self-interaction preference map (Section B.2). Once the SILCS FragMaps are calculated, calculation of the binding distribution of excipients is rapid. The FragMaps were used for the SILCS-Hotspots simulation to identify the binding sites of arginine, lysine and proline, the three excipients of interest here (Section B.3). The PPI preference and excipient hotspot maps were combined to identify the excipient binding sites overlapping with high PPI preference regions. The FragMap generation step was the bulk of the computation taking roughly three days of 10 GPU nodes, whereas the subsequent SILCS-PPI calculation took around 8 hours on a 32 CPU node and SILCS-Hotspots took roughly 2 hours/excipient on single CPU nodes.

Figure 1.

Computational workflow of SILCS Biologics. The workflow is initiated using the crystal structure of mAbN Fab. SILCS-Hotspots simulations are performed for arginine, lysine and proline and the Fab is used as the ligand protein for SILCS-PPI simulations.

Viscosity measurements

Viscosity measurements were performed on mAbN (pI ~8.7) in a histidine-based formulation composition at pH 5.6. 0.1 M arginine HCl (arginine), proline, and lysine HCl (lysine) purchased from Sigma-Aldrich (St. Louis, MO) solutions were prepared by adding the excipient of interest into the primary buffer to screen for their effect on formulation characteristics. The protein was buffer exchanged into the corresponding buffers by spin filtration using Amicon Ultra 15mL Centrifugal Filters (Millipore, Burlington, MA) with a 30 or 50 kDa cutoff. Sample viscosities were measured at ~100 mg/mL at ambient temperature as a function of shear rate with a m-VROC viscometer (Rheosense, San Ramon, CA) at a rate of 100 μL/min for 20s. The results are reported in centipoise (cP).

RESULTS

The fundamental information from the SILCS simulations are the FragMaps and the Exclusion map. Figure 2 shows the apolar, H-bond donor and H-bond acceptor FragMaps overlaid on the mAbN Fab. Figure 2b focuses on the complementarity-determining region (CDR) view of the Fab and includes the Exclusion map as well as the FragMaps. In addition to the FragMaps shown in Figure 2, specific FragMaps can be visualized with individual functional groups, e.g., the hydroxyl group in methanol probe or oxygen in water probe, which are not included in the donor or acceptor maps.⁵⁵

Figure 2.

FragMaps for the mAbN Fab. Shown are the generic apolar, H-bond donor and H-bond acceptor maps at cutoffs of −1.0, −0.7 and −0.7 kcal/mol, respectively, overlaid on a A) a ribbon diagram of the protein and B) a ribbon diagram and SILCS Exclusion map focused on the CDR region of the Fab. Apolar maps comprise of those from benzene and propone, hydrogen bond donor maps from Formamide H and Imidazole N-(H) and hydrogen bond acceptor maps from Formamide O, Acetaldehyde O and Imidazole N maps. FragMaps were also computed for the individual functional groups including those in acetate, methylammonium, methanol and water (not shown).

The FragMaps encompass sites around the entire protein surface including the CDRs. This allows mapping of excipients over the entire protein surface, which is important because excipients may simultaneously interact at multiple sites on the protein surface. While the FragMaps presented in Figure 2 are at negative contours, indicating favorable interactions of the different types of functional groups, the FragMap GFE scores range from negative to positive, with the latter indicating regions of the protein where interactions of the functional groups are unfavorable with respect to being in aqueous solution. Concerning the CDR region, the number of well-defined FragMap densities below the GFE cutoff is relatively small, indicating potentially low likelihood in interactions with excipients as well as other proteins. Overlaying the Exclusion Map on the crystal structure showed that tripeptide segments of the heavy chain CDR1 and CDR3 loops were outside the Exclusion map indicating local flexibility and greater solvent exposure for these residues. Residues lying outside the Exclusion Map may be more susceptible to various post-translational modifications.

The FragMaps and Exclusion map were used to perform the SILCS-PPI and SILCS-Hotspot simulations. SILCS-Hotspot simulations were performed for arginine, lysine and proline. Figure 3 shows the interacting poses of each excipient using a LGFE cutoff of −2 kcal/mol based on small molecule benchmarking.⁵¹ The excipients bind multiple sites with some sites being exclusive to each excipient. Interestingly, there are differences in the interaction pattern of arginine and lysine despite both being positively charged aliphatic amino acids. This information alone may be of utility in the context of excipient formulation. For example, the number of binding sites occupied by a given excipient may yield information on its potential to stabilize the Fab.

Figure 3.

Binding poses of arginine (dark blue), lysine (cyan) and proline (green) on the mAbN Fab, which is shown in ribbon diagram with the heavy chain on the right. Binding poses are those with LGFE < −3 kcal/mol.

SILCS-PPI simulation was performed using the Fab as a “ligand” protein to identify Fab-Fab self-interaction hotspot regions. Figure 4 shows the Fab surface colored by the PPI preference score and show patches with high PPIP scores. Consistent with the weak density for favorable interactions based on the FragMaps, the CDR region of the Fab shows low PPIP scores compared to the framework residues. We are working on patch analysis algorithms to quantify the strength, geometry and residue composition of the PPIP patches. Such patch analysis can help identify candidate residues for surface engineering for improving solution properties.

Figure 4.

Predicted Excipient Binding Site and Fab-Fab PPI Preference pattern. Residues PPI probability of less than 20% are white, those with probability greater than 50% are red and the intermediate are orange. In the left panel, the heavy chain is on the right. Lysine, arginine and proline molecules are rendered in cyan, blue and green balls and the exclusive sites for each excipient are indicated by arrows of corresponding colors.

Figure 4 combines the results from SILCS-Hotspots and SILCS-PPI by overlapping the excipient hotspots sites on the PPIP map. The interacting site of an excipient was defined as protein residues within 5 Å of the excipient docking pose. A interacting site was identified as a PPI interacting site if any residue in the interacting site had a PPI preference value of over 40%. Figure 3 shows the sites where excipient interaction can block Fab-Fab interactions. The number of such sites were 14 for lysine, 10 for arginine and 7 for proline. Figure 4 shows the excipient interacting poses overlaid on the PPIP surface. Figure 4 also indicates the sites that are exclusive to the excipient – two for lysine, two for arginine and one proline site. Other sites were occupied and were common for any two or all three excipients. The maximum PPI preference for residues in the two sites exclusive to lysine were 75% and 82%, indicating that these residues were involved in the dominant mode of PPI. The two sites exclusive to arginine overlap and the same residue has the highest PPI preference (49%) for both sites. In case of proline, the sole exclusive site has a maximum PPI preference of 48%.

Table 1 lists the measured viscosity values along with the SILCS statistics discussed above. A potential mechanism by which excipients may alter viscosity is by disrupting the dominant protein-protein interaction modes by occupying the sites involved in the PPI. Assuming this mechanism, given that proline did not alter viscosity, suggests that the proline interacting sites are not involved in the dominant PPI. On the other hand, the viscosity reducing effect of lysine suggests that the dominant PPI mode involves the lysine interacting sites. Out of the three excipients, since only lysine reduces viscosity, the lysine-exclusive interacting sites are candidates for dominant PPI sites. Residues in the lysine-exclusive sites, particularly ones with high PPI preference values, would then be candidates to mutate for engineering solution properties. The current model where excipient interacting is assumed to disrupt PPI, however, does not explain the increase in viscosity due to arginine and other mechanisms may be involved as discussed in the next section. Comparison of the surface interacting pattern for more excipients with viscosity modulating effects can help further narrow down the relevant sites and corresponding residues. We note that, in addition to occupying a PPI interacting site, the interacting affinity in the site may also be relevant for disrupting PPI. In the SILCS framework, the LGFE score for an interaction pose correlates with the interaction affinity. However, we did not explore the interaction affinity aspect in this work since the greater experimental data would be likely be required to extract meaningful correlations.

DISCUSSION

This work presented a novel in silico approach, SILCS Biologics, for atomistic modeling of protein formulations with potential applications to biotherapeutics formulation composition screening and engineering. The method is a combination of SILCS-Hotspots and SILCS-PPI analysis that allows for a comprehensive mapping of excipient binding sites to the entire protein surface as well as mapping of the potential of the entire protein surface to interact with self and/or other proteins in solution. This modeling framework can drive rational selection of excipients for biotherapeutics drug product development. We demonstrated the SILCS Biologics approach on the Fab domain of a monoclonal antibody and modelled three amino acid excipients, namely, arginine, lysine and proline. This work was motivated by the experimental observation of the differential impact of the three excipients on the viscosity of antibody formulations. Key observations from the various components of the workflow are as follows. The Exclusion Map from the SILCS simulation highlighted segments of two CDR loops which showed higher flexibility and solvent exposure. Such residues which lie outside the Exclusion Map may be more susceptible to chemical modification. Simulations of Fab-Fab interaction with SILCS-PPI showed greater PPIP scores for the solvent exposed residues of the framework as opposed to the CDRs. This is interesting since the framework region is more amenable for engineering as it is typically not involved in direct interactions with the antigen and favorable residue mutations in the framework may be transferrable to other mAbs. SILCS-Hotspots simulations for the three amino acids (arginine, lysine and proline) excipients identified multiple surface interaction sites for each excipient, including sites that unique to each excipient. Overlaying the excipient interaction sites on the PPI map from SILCS-PPI simulations can help identify interaction sites which can potentially block protein-protein interaction. By correlating the excipient interaction patterns from simulations with the experimental observed impact on viscosity, we propose that the two interaction sites exclusive to lysine may be involved in the dominant interaction mode between two Fab molecules to lower mAbN viscosity. This hypothesis may be tested by (1) mutating residues in the two lysine-exclusive interaction sites and (2) excipients which also occupy the lysine-exclusive interaction sites. SILCS-Hotspots and other small molecule in silico docking approaches can be employed to identify other excipients which can occupy the lysine-exclusive interaction sites. Similarly, arginine-exclusive interaction sites were also identified for PPI sites.

Here, PPI blocking mechanism does not explain the viscosity enhancing effect of arginine on mAbN and other mechanisms of actions may need to be explored.^72,73 We note that excipients can alter solution properties of protein formulations by mechanisms that indirectly modulate protein-protein interactions. For example, recent studies indicate possible formation of arginine clusters of 8–10 molecules in solution ⁷⁴. Such a phenomenon may change the local hydration around the surface ⁷⁵ of mAbN and potentially induce conformational changes that can contribute to the increased viscosity observed for mAbN seen in arginine composition, though further studies are required to confirm this hypothesis. Any such clustering phenomenon is not accounted for in the SILCS-Biologics workflow described here. Excipients may also modulate charge distribution ⁷⁶ on a protein and the dipole moment of the mAb which is known to influence solution viscosity ^76,77.

Multiple enhancements of the current implementation of SILCS-Biologics can be envisioned. In this work, only the Fab domain was modelled as a model system for SILCS-Biologics framework. Modeling just the Fab may be physically relevant as first approximation for studying viscosity of high concentration antibodies as there are many examples of viscosity and other solution properties being modulated by changes solely in the Fab domain which is the variable domain of an antibody. Nevertheless, there are examples ⁷⁸ where Fab-Fc interactions mediate aggregation and viscosity behavior. SILCS-Biologics workflow can be extended to model the full mAb. In this context, large scale domain-domain motions, for example, hinge motion in a mAb, may be addressed by assembling FragMaps computed independently for the Fab and Fc domains using one or more reference conformations for the complete mAb. Protein concentration, molecular crowding, and excipient concentration are not explicitly modelled in this specific study; however, these may be accounted for using the LGFE scores of the excipients to estimate percent saturation at different concentrations. Finally, the present study is executed at a constant pH. However, the computational efficiency of the SILCS approach allows for all accessible protonation states of the excipients to be considered and preliminary results indicate that SILCS FragMaps calculated at protein protonation states representing different pHs can be interpolated to model them as a function of pH. Another potential direction for the SILCS-Biologics analysis is to generate physics-based descriptors which can be coupled with machine learning algorithms to increase the predictive ability of the computational models. We are working on building a database of SILCS FragMaps for a diverse range of Fc and Fab scaffolds as well as collecting experimental data for different biophysical properties of interest. Subsequent publications would highlight additional nuances of application of the novel SILCS-Biologics in understanding 3D conformational structure, flexibility, dynamics, dispersive interactions, stability and viscosity of proteins. This work opens several avenues to better understand protein-protein and protein-excipient interactions for other mAb based fragments and modalities using SILCS-Biologics. These integrated features are expected to build and validate computational – experimental correlations for predictability in biotherapeutics formulation composition engineering.

Abbreviations:

3D

three-dimensional

PPI

protein-protein interactions

PPIP

protein-protein interaction preference

MC

Monte Carle

MD

molecular dynamics

GFE

Grid Free Energy

LGFE

Ligand Grid Free Energy

GCMC

Grand Canonical MC

CDR

Complementarity Determining Region

Γ

preferential interaction parameter

FragMaps

Fragment maps

PPGMaps

protein side chain probability grid maps