Mitigation of liquid–liquid phase separation of a monoclonal antibody by mutations of negative charges on the Fab surface

Some monoclonal antibodies undergo liquid–liquid phase separation owing to self-attractive associations involving electrostatic and other soft interactions, thereby rendering monoclonal antibodies unsuitable as therapeutics. To mitigate the phase separation, formulation optimization is often performed. However, this is sometimes unsuccessful because of the limited time for the development of therapeutic antibodies. Thus, protein mutations with appropriate design are required. In this report, we describe a case study involving the design of mutants of negatively charged surface residues to reduce liquid–liquid phase separation propensity. Physicochemical analysis of the resulting mutants demonstrated the mutual correlation between the sign of second virial coefficient B2, the Fab dipole moment, and the reduction of liquid–liquid phase separation propensity. Moreover, both the magnitude and direction of the dipole moment appeared to be essential for liquid–liquid phase separation propensity, where electrostatic interaction was the dominant mechanism. These findings could contribute to a better design of mutants with reduced liquid–liquid phase separation propensity and improved drug-like biophysical properties.


Introduction
The number and variety of monoclonal antibodies (mAbs) approved for clinical trials and commercial release have been increasing continuously in recent decades [1][2][3]. However, unlike small-molecule drugs, manufacturing mAbs is challenging because their substantial complexity and labile conformational stability induce disorder in their higher-order structures. MAbs can be exposed to various stresses during their manufacturing processes, formulation, storage, delivery, and administration, which can result in undesirable fragmentation, aggregation, denaturation, and chemical modifications of the mAbs [4][5][6]. These alterations can affect the manufacturability as well as the efficacy and safety of therapeutic mAbs.
Therefore, all mAb drug candidates should undergo characterization of their physicochemical properties in order to mitigate the risks of significant instability, and this should be measurements and used its pH and ionic strength as parameters for in silico analyses, as described in the Analysis of full IgG and Fab model structures subsection. We also used mAbs in buffer B for dynamic light scattering measurements.

Analysis of full IgG and Fab model structures
Model structures of the full-length and antigen-binding fragment (Fab) of the WT mAb1 and its mutants were generated using the MODELLER implementation in Discovery studio 2017R2 software [27] (Dassault Systemes, Vélizy-Villacoublay, France) using crystal structures of an antibody (PDB ID, 1HZH) or a Fab (PDB ID, 5I19) as the template, respectively. The electrostatic potential of the surfaces was determined according to the Poisson-Boltzmann equation at pH 6.0 at an ionic strength of 43 mM, which corresponds to the ionic strength of buffer A (25 mM sodium phosphate/15 mM NaCl/pH 6.0) used for in vitro experiments. For force field assignment, the charge and bondi radii were applied as parameters. The +1 and −1 kBT/e isovalue surfaces were mapped onto the model structures, where kB is the Boltzmann constant. The angle θ (degree) of the Fab dipole moment between the WT and each mutant (Mi) was calculated with the component (Dipole X, Y, or Z) of each Fab dipole moment (S1 Table) according to Eq (1):

Viscosity measurement
The mAbs in buffer A were concentrated to 130-140 mg/mL using a Vivapore 5 Static Concentrator (Sartorius Stedim Biotech GmbH, Göttingen, Germany), and the concentration was adjusted to 120 mg/mL with buffer A. The viscosity of the mAb solution was measured at 25˚C using an m-VROC micro-viscometer (RheoSense Inc., San Ramon, CA, USA). Because of limited sample availability, all measurements were performed in one replicate.

Analytical ultracentrifugation-sedimentation equilibrium
Other than the following, all procedures were performed exactly as described in previous studies [28,29]. We determined apparent molecular weights by nonlinear least-squares fitting using Origin software (OriginLab Corporation, Northampton, MA, USA). B 2 was obtained from the slope of the plot between the inverse of the apparent molecular weight and protein concentration, as shown by Saito et al. [29,30]. Because of limited sample availability, the experiment was performed in one replicate.

Dynamic light scattering
The mAb samples at concentrations of 2.5, 5.0, 7.5, and 10.0 mg/mL (after filtration through a 0.22-μm filter) in buffer A or buffer B (25 mM sodium phosphate/150 mM NaCl/pH 6.0) were plated into wells of 384-well optically transparent-bottom microtiter plates. The diffusion coefficient D s was measured using a DynaPro Plate Reader version II (Wyatt Technology Corp., Santa Barbara, CA, USA) at 20˚C. The diffusion interaction parameter k D was calculated using the protocol described in a previous paper [29]. The experiment was performed in duplicate.

PEG-mediated relative solubility
Polyethylene glycol (PEG)-6000 solutions (0.5%-18.5% w/v) at 2% increments in buffer A were prepared from 50% concentrated PEG-6000 stock solutions. The mAb samples were diluted to 1 mg/mL in buffer A, and 12 μL of the mAb samples and 28 μL of PEG solution were mixed, followed by filtration 20 min later. The absorbance at 280 nm was then measured. The PEG m , i.e., the weight% of PEG in solution required to reduce the protein concentration by 50%, was calculated as the relative solubility [31]. The experiment was performed in duplicate.

Physicochemical analysis of WT mAb1
The WT mAb1 underwent LLPS at 50 mg/mL at 4˚C in buffer A and a relatively high viscosity of 15.43 mPa�s at a concentration of 120 mg/mL ( Table 1). The B 2 was −4.08 × 10 −5 mL�mol/g 2 and the k D was −33 mL/g for the WT. The sign for B 2 suggests the existence of attractive protein-protein intermolecular interactions [11,20,21,26,29,30]. The addition of 150 mM NaCl increased the k D value from −33 mL/g to −12 mL/g, indicating that the protein-protein self-attractive interactions in low-ionic-strength buffer A could essentially be mediated by electrostatic interactions.

Mutant design
The relationship between the surface charge distribution and the location of negatively charged surface residues was demonstrated on a full-length IgG model of WT mAb1 (Fig 1). The negative charges on the surface of the WT mAb1 were unevenly distributed in the surface electrostatic potential maps (Fig 1A), which was previously believed to induce protein selfinteractions and high viscosity [19,20]. As shown in Fig 1C, mAb has two Fabs and an Fc; the Fab is composed of a constant domain (CL and CH1) and an Fv domain (VL and VH). Major negative surface charges were located in both Fv domains of the full-length model. They primarily arose from five acidic amino acid residues: light chain (LC) Glu27, Asp28, Asp56, Asp93, and heavy chain (HC) Asp31. These five negatively charged amino acids were highlighted on the Fv domain facing forward in Fig 1B, while not highlighted on another Fv domain located on the right side. Their positions corresponded to positions of the major negative charge patches in panel A, left. Of these, LC Glu27, Asp28, and Asp93 formed a major continuous patch of negative charges on the molecular surface. We designed combinations of Ala-substituted mutants at the LC amino acids Glu27, Asp28, and Asp93 to disrupt the continuous negative charge patch. Specific designs were single-charge deletion mutants (M1, M2, and M3), double-charge deletion mutants (M4, M5, and M6), and a triple-charge deletion mutant (M7). In addition to these mutants, Ala-substituted mutants for HC Asp31 (M8) and LC Asp56 (M9) were designed.
The surface electrostatic potential of the designed Fabs are presented in Fig 2. The apparent sizes of the negatively charged patches were reduced in M1-M7 in proportion to the number of depleted negative charges. In M8 and M9, the apparent sizes of rather small negative charge patches were reduced, while the continuous patch was not reduced, as designed.

Physicochemical analysis of mutants
IEF was performed to confirm the charge deletions of the mutants ( Table 1). The isoelectric point (pI) of each mAb was distributed from 8.17 for the WT to 8.68 for M7. The pI values correlated well with the number of charge deletions as designed.
In addition to an alterlation in charges, mutations can often cause drastic changes in the protein structure [32], which might affect LLPS propensity, viscosity, and protein solubility. The structural identities were verified using DSC. Overall, DSC thermograms of the WT and mutants appeared almost identical and had a main peak T m of~80˚C (S2 Fig), indicating that mutations do not likely affect the conformational stability of proteins.

PLOS ONE
The LLPS propensity of all mAbs at 50 mg/mL in low-ionic-strength buffer A is shown in Fig 3A. Single-charge deletion mutants (M1, M2, and M3) still underwent LLPS at 4˚C however, they required more time to undergo LLPS than the WT. The WT underwent LLPS within 3 h of incubation at 4˚C, whereas the appearances of mutants M1, M2, and M3 remained cloudy and homogeneous for several hours. These mutants ultimately underwent LLPS following overnight incubation at 4˚C. LLPS was clearly mitigated in double-and triple-charge deletion mutants (M4, M5, M6, and M7), even after overnight incubation. M8 and M9 did not undergo clear phase separation at 4˚C; however, their solutions were opaque, unlike those of M4-M7. In general, the opaque appearance could be due to both LLPS and fine precipitation.
To explain the opacity of M8 and M9 solutions, we conducted a larger-scale LLPS experiment with centrifugation. M8 and M9 solutions showed opacity after 24 h incubation at 4˚C

PLOS ONE
( Fig 3C). Subsequent centrifugation for 48 h at 4˚C resulted in clear phase separation for M8 and M9, indicating that the opaque appearances of M8 and M9 solutions are due to LLPS (Fig 3D).
The B 2 values for these mAbs were determined using the AUC-SE method [28][29][30] (Table 1). A negative B 2 value denotes attractive protein-protein interactions, whereas a positive value denotes mutual repulsion. The B 2 values of all single-charge deletion mutants increased from that of the WT; however, their values were still negative. The double and triple- The k D value was determined by dynamic light scattering (Table 1) and was increased in proportion to the number of charge deletions. The k D values were −33 ml/g for WT, −20 to −25 mL/g for the other single-charge deletion mutants, −14 to −10 mL/g for the double-charge deletion mutants, and −12 mL/g for the triple-charge deletion mutant. Adding 150 mM salt increased the k D of the WT and mutants to a range of −11 to −19 mL/g (Fig 4). These results indicated that the deletion of negative charges effectively reduced attractive self-interactions and suggested that with regard to the WT, the electrostatic interaction could play a role as a a part of the soft interaction.
The relative solubility of the mAbs was evaluated by PEG precipitation. The value of PEG m increased proportionally with the number of Ala-substituted acidic residues, which ranged from 5.62% to 9.

Relationships between the dipole moment, B 2 , and LLPS
Our findings showed a strong correlation between B 2 and the Fab dipole moment of single-, double-, and triple-charge deletion mutants (Fig 5A), indicating that the association can be described as being driven by a Fab dipole-dipole attraction. Kanai et al. [33] described the Fab-Fab interaction for protein-protein self-attractive interactions. Gentiluomo et al. [34] characterized the native reversible self-association of mAbs using Fab and Fc fragments and confirmed that self-associations such as hydrophobic and electrostatic interactions are driven by the Fab fragment. In addition, Du et al. [24] and Chow et al. [25] identified the amino acids responsible for the LLPS propensity of mAbs, located in the Fab region. In this study, we calculated Fab dipole moments to investigate the Fab-dipole-related association. In addition, limited numbers of entire IgG crystal structures are available, making the calculation of the entire IgG dipole moment difficult because the full-length model is only a single snapshot and cannot be interpreted as a representation of multiple hinge-angle forms. The dipole moment of the entire mAb differs in magnitude and direction from the Fab dipole moment. The computational approach we followed using a Fab model structure provided insight into a better design of LLPS mitigation mutants. However, further in vitro analysis, such as X-ray crystal structure analysis, is required to obtain evidence of Fab-Fab interaction mediating protein network formation of the mAb1 heavy phase.
The 10 mAbs were divided into two groups. The first group comprised four mAbs containing double and triple mutants that did not show LLPS propensity and had positive B 2 . The second group comprised the remaining six mAbs that showed LLPS propensity and had negative B 2 . Therefore, the sign of B 2 could appear as an indicator of the LLPS propensity for mAb1. More precisely, B 2 should be normalized to B 2 � while considering the excluded volume, as shown in Eq (2) [35]. Here, B 2ex indicates the contribution of the excluded volume to the virial coefficient. Negative but the sign of B 2 appears to be well correlated with LLPS propensity in this study.
The LLPS propensity correlated with B 2 , and the B 2 value had a linear relationship with the dipole moment. While LLPS propensity is a macroscopic phenomenon observed at high protein concentrations, B 2 values are measured at relatively low protein concentrations. Accordingly, the dipole moment certainly contributes to B 2 , even at the high protein concentrations in this study. It is natural that not only the dipole moment but also other soft protein-protein interactions might play a role in determining B 2 . Saito et al. [36] also indicated a positive correlation between B 2 values and viscosity in highly concentrated mAb solutions. They speculated that a similar force component, such as the dipole moment, could work not only at low concentration but also at high concentration. This supported the findings of our study.

Significance of the dipole moment as an index for LLPS mitigation
The LLPS propensity correlated well with the magnitude of the Fab dipole moment, except for the outside-patch mutants M8 and M9. Despite exhibiting a Fab dipole moment of almost the same magnitude as the WT, the outside-patch M8 and M9 mutants did not undergo clear LLPS without centrifugation (Fig 3). The M8 and M9 LLPS propensity rank order could not be interpreted from the magnitude of the Fab dipole moment.
We further investigated the reason why M8 and M9 had lower LLPS propensities by plotting the angle between the WT dipole moment and each mutant dipole moment against the magnitude of the Fab dipole moment (Fig 5B). M8 and M9 had larger angles of the dipole moments than the other three single-charge deletion mutants (M1, M2, and M3). Our study indicated that the dipole moment could play a role in LLPS propensity. It is expected that the angle of Fab dipole moment affects the alignment of the molecule and protein-protein network formation in a crowded heavy phase. Therefore, the weak LLPS propensity of M8 and M9 could be explained by the angle to the WT Fab dipole moment. The results revealed that both the magnitude and the direction of the dipole moment could be essential for the LLPS propensity of mAb1, and the dipole moment could be the referential index for risk mitigation of the LLPS propensity.
Chow et al. [25] used X-ray crystal structure analysis to determine the amino acid residues of mutations that would mitigate the LLPS propensity. In the present study, we designed mutants based on tertiary structural models alone, without laborious in vitro experiments such as crystallization. We also calculated the Fab dipole of 10 mAbs from their primary structure within half a day. This in silico analytical approach could work in conjunction with highthroughput LLPS assessment in the course of drug candidate screening. In addition, the approach provides a reliable protocol for designing a mutant that does not show LLPS propensity in cases where electrostatic interaction is the dominant mechanism for LLPS. It reduces the number of candidate molecules required for in vitro LLPS assessment, thereby greatly improving operational efficiency.

Relationships among various biophysical parameters other than the dipole moment
The k D values strongly correlated with the B 2 value (Fig 5C; squared correlation coefficient = 0.89). A k D value of −20 mL/g was a threshold for LLPS where the B 2 value indicated zero in this case study. Previous studies [28][29][30][35][36][37][38] have also demonstrated a linear relationship between the k D and B 2 values. The investigations by Connolly et al. [28] indicated this relationship among eight different mAbs and corroborated our results.
All the negative surface charge mutations clearly reduced the viscosity of the WT. Moderate correlation between viscosity and the B 2 value was shown with a squared correlation coefficient of 0.58 (S3 Fig). Previous studies assessing viscosity [19,20] also suggested that the elimination of the uneven distribution of surface charges reduced viscosity. Because both LLPS and viscosity are a result of protein-protein self-attractive interactions, the two phenomena are correlated. LLPS is qualitatively correlated with B 2 . The viscosity of the WT and mutants is moderately correlated with B 2 , probably because of the nonlinear dependence of viscosity on the protein concentration. The viscosity of mAbs could be proportional to the exponential function of concentration [39].
The squared correlation coefficient between B 2 and PEG m was 0.84, indicating a strong correlation. The excluded volume theory predicts that PEG can trap H 2 O, thereby sterically excluding proteins from the solvent regions occupied by PEG and causing a phase transition of proteins [40]. Here, the electrostatic interaction could be the major interaction in a lowionic-strength buffer such as buffer A. The electrostatic interaction also reflected interaction parameters such as B 2 for our charge mutants in buffer A. Thus, the surface charge mutants showed correlations of B 2 with PEG m . (Fig 5D).

Conclusions
We demonstrated the mitigation of LLPS by deletion of negative charge patches from the surface of the mAb1 Fv region. LLPS propensity was eliminated for the mutants with positive B 2 values, with reducing theoretical Fab dipole moments. Therefore, the sign of B 2 may be useful for predicting LLPS propensity. Protein-protein soft interactions are complex; however, both the magnitude and the direction of the dipole moment could be part of the essential contributors to the LLPS propensity where the dominant mechanism for LLPS is electrostatic interaction. Further studies are required to apply these findings to other mAbs. The insights obtained in this study could open up new avenues for the design and selection of well-behaved therapeutic mAbs.