Mechanistic and predictive formulation development for viscosity mitigation of high-concentration biotherapeutics

ABSTRACT

Proteins are an important class of therapeutics for combatting a wide variety of diseases. The increasing demand for convenient, patient-centric treatment options has propelled the development of subcutaneously delivered protein therapies and increased the interest in novel formulations and delivery methods. However, subcutaneous delivery of protein therapeutics remains a challenge due to the high protein concentrations ( >100 mg/mL) required to circumvent lower bioavailability and the smaller injection volumes required to enable the use of mature and cost-effective devices, such as standard prefilled syringes and autoinjectors. At high concentrations, protein solutions exhibit elevated viscosity, which poses injectability and manufacturing challenges. Here, we review the state of the art in experimental and computationally predictive formulation development approaches for viscosity mitigation of high-concentration protein solution therapeutics, and we suggest new directions for expanding the utility of these approaches beyond traditional monoclonal antibodies. Innovative approaches should leverage and combine advances in both experimental and computational methods, including machine learning and artificial intelligence, to rapidly identify formulation compositions for viscosity reduction, and subsequently facilitate the development of patient-centric biotherapeutics.

GRAPHICAL ABSTRACT

Introduction

Protein therapeutics, which have a worldwide market valuation exceeding 200 billion dollars, are now routinely developed by the pharmaceutical industry.^1,2 Over 150 antibodies are approved for use in patients and many more are in the clinical pipeline.³ Most marketed monoclonal antibody (mAb) drugs are packaged as a solution-in-vial and approved for immuno-oncology indications.^1,2,4 Moreover, there is continued growth in the development of new biotherapeutic modalities, such as multispecific antibodies, chimeric proteins, and protein degraders.⁵ Protein therapies are also expanding into additional therapeutic areas, including neuroscience, immunology, and cardiometabolic disorders.⁵ These indications may require high doses or large injection volumes to improve the pharmacokinetic profile or to facilitate crossing of the blood–brain barrier.⁶

While historically the most common route of administration for protein therapeutics has been intravenous (IV) infusion, subcutaneous (SC) injection can enable faster drug delivery, as well as at-home, self-administration.^7,8 In one example, patient chair time for Rituximab IV administration ranged from approximately 3 to 5 h, but was reduced to approximately 10 min to 2.5 h by SC administration, with high variability depending on the country of the administration.^9–11 At-home therapy may also expand patient access and reduce healthcare system burden and costs.^12,13 Accordingly, developing biotherapeutics for SC injection has become a key enabler for patient-centric drug product development, as it can improve patient convenience and access and facilitate flexible dosing options and site-of-care shifts, including at-home treatment.^12,13 Furthermore, for already marketed biotherapeutics, commercializing a new presentation for SC administration represents a valuable life-cycle management opportunity. To support SC delivery and self-administration, it may be favorable to develop low-volume drug–device combination products, such as prefilled syringes (PFS) or autoinjectors (AI), which often require high protein concentrations. As a result, there has been a recent increase in the commercialization of high-concentration ( >100 mg/mL) protein drug products (DPs), with the majority of these products being integral drug–device combination products, such as AIs (Figure 1A,B). While many high-concentration protein DPs are approved for immunology indications, other therapeutic areas, such as oncology, are becoming increasingly common (Figure 1C,D). High protein concentrations are often required for SC administration due to the generally low potency of antibody drugs, decreased bioavailability, and lower dose volumes.^14,15 While large-volume injection devices, such as on-body injectors and infusers, enable large-volume, low-protein concentration formulations for SC delivery, they are more costly and less mature technologies relative to low-volume devices such as PFS and AI. Other strategies include the addition of delivery enhancers, such as hyaluronidase. Hyaluronidase is an enzyme that can degrade parts of the hyaluronic acid matrix in the SC space, allowing more rapid injection of larger volumes. The benefits and concerns regarding hyaluronidase are extensively reviewed elsewhere.^{1,2,8,16–21}

Figure 1.

Trends in the product presentations and therapeutic areas of marketed high-concentration biotherapeutics. A) number of marketed products with a given presentation (prefilled syringe, PFS; autoinjector, AI; on-body injector, OBI). B) change in marketed product presentations over time. C) number of marketed products indicated for a given therapeutic area. D) change in the indicated therapeutic areas over time.

Chemical and physical stability are key attributes in the development of biotherapeutics, and for high-concentration protein therapeutics, viscosity is an additional consideration.^20,22 During clinical development, pharmaceutical companies expend great effort to characterize the behavior and properties of a given therapeutic candidate and to optimize the molecule’s formulation to mitigate physicochemical liabilities and maintain product quality throughout the established shelf-life.²³ At high protein concentrations, aggregation kinetics can increase due to decreased intermolecular distances and increased protein–protein interactions (PPIs). Characterization and mitigation of aggregation throughout the DP manufacturing process and shelf-life are important for maintaining product safety and quality. Similarly, DP viscosity may need to be characterized and reduced to help maintain product quality, since high solution viscosity poses significant challenges for DP manufacturability and injectability.^19,24,25 High viscosity is often unique to high protein concentration solution DPs. Protein solutions exhibit exponential increases in viscosity with increasing protein concentrations such that small changes in concentration within the high-concentration regime dramatically affect viscosity. Yet, for complex modalities, such as bispecific antibodies, high-viscosity solutions may be observed at lower protein concentrations.²⁶ Thus, ‘high-concentration’ is relative to the protein solution properties rather than a specific mass per volume concentration.^1,18,27 As such, high-concentration DPs can be defined as those that exhibit distinct solution and stability properties, such as high viscosity, particle formation, and opalescence, relative to DPs at low concentrations ( < 30 mg/mL).¹⁸ However, in the case of mAbs, these properties are typically observed for concentrations exceeding 100 mg/mL.

In this review, we discuss experimental and computationally predictive formulation development approaches for viscosity mitigation in high-concentration therapeutic protein solutions and suggest new directions for expanding the utility of these approaches beyond traditional biotherapeutics. Recent reviews have focused on elements of high-concentration biologics’ development individually, including the origins of viscosity,²⁸ manufacturability,²⁷ existing computational methods,²⁹ developability assessment,²² and formulation compositions.¹ In comparison, we herein address the challenges and opportunities that arise from combining computational methods with experimental approaches to innovate on formulation composition design for viscosity reduction. While we discuss the development of high-concentration protein therapeutics generally, most research has prioritized monoclonal antibodies (mAbs) owing to their prevalence. For example, most multispecific antibodies are currently administered IV, with two approved for SC administration (Hemlibra and Vabysmo).⁴ However, a growth in SC-administered multispecifics is anticipated. More research on new modalities, such as multispecific antibodies and fusion proteins and challenges in their development for SC administration will be forthcoming. Recent reviews regarding the development of low-concentration, large-volume protein drug products for additional context outside of the high-concentration, low-volume solutions we focus on here are available.^17,30,31

High viscosity is a concern for drug product and process design

The development of high-concentration, high-viscosity protein solutions, compared to low-concentration, low-viscosity protein solutions, presents challenges in manufacturing, injectability, and drug–device combination product performance.¹⁷ In the manufacturing process, drug substance (DS) and DP may undergo a series of unit operations before final packaging, including ultrafiltration/diafiltration (UF/DF), filtration, transfer, mixing, and filling.^19,32 Process time must be carefully considered accounting for DS stability at processing temperatures to ensure product quality and safety (i.e., to minimize temperature excursion and to prevent protein aggregation and degradation).¹⁹ High viscosity can negatively impact any of these steps and extend the overall processing time. High protein concentrations and viscosity can slow unit operations involving membranes (UF/DF and filtering) due to decreasing filtrate flux with increasing protein concentration.³³ Interactions between the protein and oppositely-charged excipient molecules lead to charge imbalance across the membrane since charge–charge interactions cause excipient molecules to be retained, resulting in attraction of protons to one side of the membrane and thus pH shifts per the Gibbs–Donnan effect.³⁴ UF/DF also results in high local concentrations at the membrane, causing high protein concentration solutions to be more prone to filter fouling and aggregation. During formulation and compounding steps, homogenization of viscous solutions, especially mixing viscous DS with a low-viscosity diluent, requires higher mixing power per unit volume and/or increased mixing durations.²⁷ This can result in increasing shear stress, cavitation, and the introduction of air–water interfaces that might lead to physical instability and lengthen total process times. Additionally, since shear stress increases with increasing viscosity, it may be useful to slow pumping rates to reduce shear stress. Slower pumping speeds also lengthen total process times. Finally, once formulated, the DP solution is filled into the appropriate primary container-closure system (CCS), such as vials or PFS. In this filling step, performance robustness becomes more challenging for high-concentration biologics due to increased viscosity and the propensity for the filling needle to clog at high concentrations.^35,36 Moreover, the viscosity of protein solutions is not only sensitive to protein concentration but also to temperature, with viscosity rapidly increasing as temperature decreases.³⁷ As a result, small changes in protein concentration or temperature can result in large changes in viscosity and variation in filling and other unit operations sensitive to viscosity (e.g., UF/DF). Furthermore, process yields are reduced for high-concentration, viscous protein solutions.^27,32 These and other challenges in DP process are the subject of ongoing investigation.^27,32 The extensive challenges that arise during manufacturing provide additional motivation for the development of low-viscosity DPs to improve process robustness and reduce process times.

High viscosity has implications for route of administration and final product presentation

Viscosity plays a key role in the selection of the final drug product presentation and formulation composition of high-concentration biotherapeutics. To improve patient convenience and adherence, there has been increased interest in drug–device combination products that enable SC and/or at-home self-administration, including PFS, autoinjectors, and on-body injectors.² The flexibility and portability that drug–device combination products provide are of specific utility to patients managing chronic diseases, where repeated dosing may occur for months to years. The functionality and robustness of drug–device combination products requires control of DP solution viscosity to maintain device performance. The final product presentation (e.g., autoinjector, vial, and syringe) and route of administration (e.g., SC vs. IV) are also dependent on the solution viscosity, considering that a solution may be too viscous for administration via a device, or too difficult to draw and inject from a vial.³⁸ Similar to manufacturing robustness, product robustness must be established given the potential variability in solution viscosity due to protein concentration variability, as well as fluctuations in product temperature. A product with high viscosity can also have drawbacks upon administration.³⁹ If the solution viscosity is high (>20 centipoises, cP), the force required to depress a syringe plunger may cause pain for patients and healthcare providers, and it may result in delivering a partial dose through incomplete injection.^31,40,41 Similarly, partial doses may be administered when the viscosity of the drug product solution requires an injection force that exceeds the capability of the injection device, resulting in incomplete or failed injection. Interestingly, a small-scale study using a qualitative, self-reported scale of pain perception suggested that patients perceive more injection pain with solution viscosity of 1 cP than with solution viscosities ranging from 15–20 cP.⁴⁰ However, this study also used a syringe pump, which may not accurately capture patient experience during standard SC injection.⁴⁰ It should be noted that water has a viscosity of 1 cP at 20°C.⁴⁰ Nonetheless, a failed or incomplete injection due to high solution viscosity is of particular concern due to impact on dose-efficacy and patient safety.

Although 20 cP is a common viscosity threshold for development,^42–44 many considerations must be made as to what DP viscosity is allowable for the target product presentation (e.g., vial, PFS).⁴³ The 20 cP limit arises from the linear relationship between syringeability (ability to draw up the solution) and viscosity for thin needles (e.g., 26 G, 27 G, 30 G) used to minimize the pain of SC injection.^42,43,45 Generally, important points of consideration are the needle gauge to be used, the limits of the actuation force (e.g., hand-strength, spring, motor), and the flow rate required to administer the dose within the required time-frame. Additionally, molecule-specific considerations may require lower viscosities for the DP process than required for the device alone to prevent the previously stated manufacturing challenges and potential critical quality attribute impact.²⁷ Thus, parameterizing the acceptable viscosity for a given DP presentation and manufacturing process may not always be transferrable to a different molecule, process, or presentation. What viscosity the patient tolerates best upon injection is a qualitative question that remains under active study. Thus, the upper limit for target DP viscosity may be a consequence of process and product design needs, in addition to injectability concerns.

Electrostatic and hydrophobic interactions drive high protein concentration solution viscosity

Viscosity is an intrinsic solution property that describes the resistance of the fluid to flow, and it depends on the mechanical stress the solution experiences (shear and/or strain).⁴⁶ As such, solution viscosity may vary under varying shear rates. Solution viscosities also exhibit temperature-dependence, such that increases in temperature generally result in decreases in viscosity, particularly within high-viscosity solutions (Figure 2).³⁷ High-concentration protein solutions (i.e., above 100 mg/mL) often exhibit non-Newtonian, shear-thinning behavior.⁴⁷ In the simplest description, non-Newtonian fluids exhibit shear rate-dependent viscosity, while Newtonian fluids exhibit shear rate-independent viscosity.⁴⁶ Non-Newtonian behavior is common in self-interacting systems, but even at high packing fractions, non-interacting hard-sphere fluids can exhibit non-Newtonian behavior.⁴⁸ A number of models have been proposed to explain the non-Newtonian behavior of protein molecules, though this remains an active area of study.^25,48–50 Given the potential complexities behind a single viscosity value, it is important when assessing viscosity data to consider its relationship to the solution temperature, the applied shear rate(s), the method used, the various unit operations during manufacturing, and injectability.

Figure 2.

The temperature dependence of viscosity.³⁷ Viscosity increases with increasing protein concentration and decreasing temperature (4°C to 40°C from light to dark). Different shapes correspond to formulation conditions tested in the original publication (NaCl, sucrose, both, or neither). Reprinted with permission from Yuan, G. et al. Flow activation energy of high-concentration monoclonal antibody solutions and protein–protein interactions influenced by NaCl and sucrose. Molecular Pharmaceutics 21, 4553–4564 (2024). https://doi.org/10.1021/acs.Molpharmaceut.4c00460. Copyright 2024 American chemical society.

The literature attests to a model wherein high viscosity of high-concentration protein solutions arises from PPIs. In one widely accepted model of self-association, weak interactions between protein monomers may lead to the formation of reversible, transient clusters that resist deformation under shear stress, resulting in high solution viscosity.^47,51–54 Specifically, the interactions between protein monomers facilitate the formation of reversible clusters, which can grow up to percolated networks. These networks may change the solution microstructure and trap solvent molecules to create immobile fluid volumes that resist shear deformation.^55,56

One of the driving protein properties behind PPIs and viscosity is the protein’s charge in the formulation condition.⁵⁷ Most mAbs have isoelectric points around 7 and 8, while they are typically formulated at a pH of 5–6.⁵⁸ In this pH range, proteins exhibit a net positive charge and, due to long-range repulsion (LRR), are generally not strongly self-interacting in dilute conditions. The extent of PPIs varies with the magnitude of the protein’s net charge rather than only on the relative sign (positive or negative). A more positively charged protein is less self-interacting due to long-range repulsions than a modestly, yet still positively charged protein. Nonetheless, long-range repulsion may not be sufficient to prevent charge–charge interactions since antibodies are comprised of a multidomain structure, whose different domains may be oppositely charged, and form local intermolecular electrostatic attractions. Additionally, with increasing protein concentrations, intermolecular distances decrease. At short distances, distinct charges in protein subdomains or individual charged residues on the protein surface may engage in electrostatic attractions. As a result, both net charge and local charged residues contribute to the solution viscosity (Figure 3). The groupings of surface charged residues on distinct structural regions are referred to as patches.

Figure 3.

Multiple interactions and effects contribute to solution viscosity.⁵⁹ Panels A-D show various protein-excipient interactions which may have viscosity modulating effects and how given excipient types modulate those interactions. A) excluded volume, B) electrostatic repulsion, C) electrostatic attraction, and D) hydrophobic attraction. Figure reproduced from Hong et al.^59.

One approach to exploring the driving forces of the viscosity of concentrated protein solutions is to vary the pH and ionic strength. In one example, Liu et al. tested if the addition of sodium chloride to a mAb solution would reduce solution viscosity.⁴⁷ Their results suggest that salt shielding of electrostatics and reducing overall repulsion contributes to decreasing the solution viscosity.⁴⁷ This may be due to the increase in ions (positive and negative) available in solution to interact with charged groups on the protein molecule and shield the charged amino acid residues from interacting. The increased number of ions changes the total ionic strength of the solution. This observation of a reduction in mAb viscosity with increasing ionic strength supports the premise that electrostatic interactions play a role in the high viscosity of mAb solutions. Furthermore, Liu et al. suggested that the increase in viscosity with increasing protein concentration is due to reversible, electrostatically mediated PPIs since analytical ultracentrifugation results show increases in the molecule’s apparent molecular weight.⁴⁷ Their results also show the apparent molecular weight decreases and correlates with the decrease in viscosity upon the addition of sodium chloride.⁴⁷ However, increasing ionic strength is not the panacea for viscosity reduction since other examples showed increases in viscosity with increasing ionic strength.^23,28 These studies suggest the high viscosity of some protein solutions may be electrostatically driven, while others may not. One potential explanation is that the PPIs driving high viscosity for a given solution may be due primarily to a protein’s short-range attractions (SRA) which are not electrostatic. Thus, increasing ionic strength and screening charges may prevent electrostatic attraction, but not affect clustering due to van der Waals or hydrophobic interactions.

In addition to the net or apparent charge of a protein molecule contributing to PPIs and high viscosity, available data suggest that the patterning of charged residues on the protein surface also contributes to PPIs and viscosity. In one study, Apgar et al. introduced a scheme to generate mutant antibodies with distinct surface charge properties, such as through charge swapping. Their data show that simple charge swap mutants do not always exhibit viscosity reduction.⁶⁰ Upon mutation to oppositely charged amino acids at some positions, viscosity increases, demonstrating that the chemistry of surface-exposed residues affects PPIs and viscosity. Additional work from Lesins⁶¹ and Roberts and Blanco,⁶² among others,⁶³ further suggests that PPIs may be due to electrostatic interactions between ‘charged patches’ or ‘clusters’ of charged amino acids. Lesins’s work shows that ‘non-uniform’ charge distributions, which result in the same net charge, impact the interactions between proteins and chromatography resins. Similarly, Roberts and Blanco modeled proteins as spheres with patches of charge or hydrophobicity and simulated PPIs. Their data show that patch size can affect PPIs such that there is a minimum patch size where forming PPIs becomes energetically favored. A follow-up publication to the Apgar study, from Rai et al., utilized a machine learning (ML) approach to identify portions of an antibody surface that are associated with high viscosity.⁶⁴ This model decomposes the input protein’s electrostatic surface potential to identify the drivers of viscosity and finds that negatively charged patches of aspartate and glutamate residues strongly contribute to high viscosity.⁶⁴ These data further support the model of charged patches driving self-association and high viscosity.

While charge-based interactions contribute to PPIs and viscosity, proteins also experience non-electrostatic SRA. That is, at small intermolecular distances, hydrophobic interactions, van der Waals forces, and hydrogen bonding can drive protein self-association.⁶⁵ To parse the relative contributions of electrostatic and hydrophobic patches to protein solution viscosity, Nichols et al. introduced mutations to hydrophobic patches in an IgG1 antibody and compared the variants’ viscosity and other molecular properties. The mutations decrease viscosity in a manner that is comparable to charge neutralizing mutations, demonstrating that hydrophobic patches also contribute to solution viscosity.⁶⁶ Similarly, Esfandiary and colleagues added free aromatic amino acids to a test IgG1 mAb solution to determine if hydrophobic interactions drive mAb self-association.⁶⁷ Then, they measured the antibody’s hydrodynamic diameter (D_H) to assess PPIs and found that the addition of free aromatic amino acids reduces the protein’s D_H. Since free aromatic amino acids could interact with hydrophobic surfaces, these data suggest a role of hydrophobic patches/residues in driving self-association.⁶⁷ While D_H may not be predictive of viscosity at high concentrations, D_H as a measurement of self-association does follow a similar trend where viscosity decreases with smaller D_H values.⁶⁸ It must be noted that the effect of uncharged interactions may be stronger in crowded environments, where intermolecular distances begin to approach the molecular size (e.g., ~10 nm for mAbs) and the local concentrations of hydrophobic residues increase.⁶⁹ In combination with electrostatic interactions driving PPIs and high viscosity, the literature supports hydrophobic interactions also playing a significant role, demonstrating the complex origin of high-concentration protein solution viscosity.

The properties of a protein solution, such as viscosity, depend on the protein’s properties (e.g., charge, PPI propensity) as well as hydrodynamic and crowding effects.⁴⁶ Hydrodynamic effects are the effects on a molecules’ diffusion due to other molecules’ motions in solution.⁷⁰ In the context of protein solutions, crowding effects are the changes to the thermodynamics and kinetics of a protein’s interactions due to steric constraints that arise at high volume fractions.^70,71 Zidar and colleagues estimated the contribution of hydrodynamic effects and other steric constraints related to crowding to be approximately 30% of observed viscosity at 150 mg/mL of their test mAb.⁶⁵ Comparably, they estimated electrostatic and hydrophobicity effects contribute 20% and 30%, respectively. In that report, they attributed the final 10% to ‘other’ effects driving PPIs, such as van der Waals interactions. Yet, the relative contribution of PPIs or crowding to solution viscosity may also be related to the protein concentration in the solution.⁷²

In addition to surface chemistry, the unique shape of antibodies make possible unique configurations for the resulting protein clusters, which contribute to high viscosity. Antibodies are composed of light and heavy chains that assemble into a characteristic ‘Y’ shape composed of crystallizable fragment (Fc) and antigen-binding fragment (Fab) domains. The latter domain contains the variable fragment subdomain (Fv) and the complementary-determining regions (CDRs). The sequences of the Fc domains are a principal difference between the IgG1, IgG2, and IgG4 subclasses, which are the most common subclasses among therapeutic antibodies. Determining the molecular and structural origins of PPIs underlying high viscosity requires understanding the geometry of mAb self-interaction. For two interacting mAb molecules, the interactions can be subdomain specific: Fc-Fc, Fc-Fab, or Fab-Fab. By combining computational modeling and experimental measurements, one study suggested that Fab-Fc interactions dominate protein clusters, with significant differences between IgG subclasses.⁵³ In that work, Chowdhury et al. used coarse grained (CG) molecular simulations and light scattering experiments to probe the arrangement of molecules in viscosity-inducing reversible aggregates. Their approach with vedolizumab shows that IgG1 exhibits stronger interaction energy between the CDRs and Fc than IgG4. These PPI differences between IgG subclasses likely arise due to sequence differences in the constant domains on both the Fab and Fc.⁵³ Chowdhury et al. noted that charged residues strongly contribute to the differences in interaction energy by IgG subclass, whereas hydrophobic interactions are approximately equal. Kanai and colleagues provided additional evidence indicating which regions of antibodies are involved in PPIs.⁷³ They generated full-length IgG1 and, via proteolytic cleavage, generated samples of the Fab alone and of a dual Fab without the Fc (i.e., F(ab’) ₂). Then, they measured solution viscosity for mixtures of these components. Their data show that the Fab alone exhibits a concentration dependent, exponential increase in viscosity. They also showed that simply mixing the full-length mAb with its corresponding Fab increases solution viscosity, suggesting that Fab-Fab interactions are sufficient to drive viscosity increases. To probe Fab-Fc interactions, they mixed the Fab from one antibody with a solution containing a different, full-length mAb of the same class (IgG1) that also had an identical Fc sequence. This solution did not exhibit a significant concentration-dependent increase in viscosity, suggesting Fab-Fc interactions may not drive viscosity for the test molecules in this study. In comparison, Tang et al. showed that the IgG4 Fc is more negatively charged than that in the IgG1 subclass, which may lead to increased Fab-Fc interactions.⁷⁴ In this work, they compared IgG1, IgG2, and IgG4 molecules with identical variable domains (Fv) by measuring the viscosity of these directly comparable antibody solutions. Overall, solutions of IgG4 molecules with a given Fv exhibited higher viscosity than the corresponding IgG1 solutions. It is important to note, however, that even within a given subclass, the identity of the Fv had a significant impact on the observed viscosity. These results suggest that the extent of Fv-Fc interactions may vary from mAb to mAb, even within one subclass, since changing the CDR changes the antibody’s properties, including propensity for PPIs. The CDRs are unique to each therapeutic mAb and determine antigen binding and mAb efficacy and can play a significant role in driving protein interactions in addition to their biological function in antigen binding. Work from Gupta et al. further emphasized the importance of the CDRs in PPIs by showing that varying the sequence of the test mAb’s CDRs modulates PPIs.⁵⁸ While they did not measure viscosity, they observed large changes in the diffusion interaction parameter (k_D) upon CDR sequence changes, supporting the role of the CDRs in mAb PPIs. Accordingly, the configuration of the solution structure (i.e., the arrangement of molecules in solution) at elevated concentrations is unlikely to be the same for different mAbs, making viscosity reduction strategies equally complex and unique for every drug molecule.^53,75

The existence of reversible native protein clusters in solution may contribute to viscosity, but the relationship between viscosity and the size of the clusters remains somewhat unclear. In one report, Dandekar and Ardekani constructed a molecular physics-based theoretical framework to identify the effect of clusters on solution viscosity.⁵⁵ This approach modeled the solution structure as protein clusters of variable sizes and orientations. Their model suggests that the formation of clusters can lead to the ‘trapping’ of solvent in ‘immobile fluid fractions’ within the cluster structure. This immobile fluid fraction cannot flow as readily under shear as the free fluid fraction, giving rise to solution viscosity.⁵⁵ This model is able to accurately fit viscosity measurements, suggesting that larger cluster sizes would lead to higher solution viscosities. One implication is that molecules that are susceptible to multiple kinds of interactions (e.g., electrostatic, hydrophobic, Fc-Fc, Fab-Fc) may form more extensive and larger clusters akin to networks rather than dimers and oligomers leading to higher viscosities. However, it must be noted that this model assumes different protein cluster configurations are equally likely. Since some cluster configurations might be more thermodynamically favored, their results may overestimate the effect of cluster size on viscosity. Kastelic et al. have also suggested that both protein monomers and clusters may form extended networks.⁷⁶ In their approach, which models antibodies as patchy hard spheres, Kastelic et al. described the capacity for mAbs to form extended networks in solutions using statistical thermodynamic approaches.⁷⁶ This model suggests that whether a mAb forms networks is related to its capacity to form Fab-Fc interactions in addition to Fab-Fab interactions, which leads to higher viscosities. This result may be related to that of Dandekar and Ardekani since networks could also trap solvent, creating immobile fluid fractions. Consequently, these different theoretical reports suggest that protein interactions may lead to the formation of reversible protein clusters and networks, which trap solvent and cause the fluid to resist deformation under shear stress, resulting in elevated viscosity.⁷⁷

Current formulation composition design, drug development, and viscosity mitigation approaches

In parallel to the increased understanding of the basis of viscosity in highly concentrated protein solutions, advances continue to be made in the formulation development, characterization, and viscosity mitigation of high-concentration biologics.^16,78 These are primarily novel formulation strategies actively under investigation, including nanoparticles and spray dried, spray freeze dried or crystalline protein particles for inhalation or injectable suspensions,^79,80 while some novel viscosity-reducing excipients have also been recently described.^{65,72,79,81,82} Given that viscosity scales exponentially with protein concentration, if the efficacious dose can be maintained through lowering the DP concentration, viscosity mitigation through formulation composition changes may not be necessary. Otherwise, viscosity is a key developability consideration and much literature supports the adjustment of the liquid composition to mitigate high viscosity.^72,81–84

Generally, to optimize and maintain product quality and safety during drug development, scientists must adjust and define the composition of the DP solution, which may contain buffer, pH, tonicity modifiers, anti-oxidants, surfactants, stabilizers, and viscosity-reducing additives.¹⁶ Many summaries of commercially available products, their formulations, and the rationale behind typical formulation components are available.^1,16,85–88 Critically, viscosity-reducing additives may be added to the protein formulation composition to reduce the viscosity of the solution.

Compiled in Table 1 are the formulation compositions, product presentations, protein concentrations, approval years, and therapeutic areas and indications for currently marketed antibody drug products, with concentrations exceeding 100 mg/mL as of March 2025. Trends identified in a 2023 paper by Ghosh et al.¹ remain consistent (Figure 1). Overall, the most common product presentations for high-concentration antibodies are PFS and vial, with each accounting for ~33% of the products. One DP’s label describes the product as a ‘pen’ which, combined with those labeled as AI, comprises most of the remaining products. Only two products are on-body injectors (OBI), likely due to the possibility of delivering low-concentration, large-volume formulations with such devices. Interestingly, from 2003 to 2025, the proportion of vial products has decreased, with PFS and AI becoming increasingly common. Among these high-concentration drug products, approximately half are marketed for immunology indications. Oncology and neurology products account for another quarter, with the final quarter composed of hematology, infectious disease, ophthalmology, cardiology, and ‘other’ therapeutic areas. Over time, high-concentration DPs across therapeutic areas other than immunology have become more common, most notably starting around 2015. Finally, among high-concentration protein DPs, arginine, usually in the presence of histidine buffer, is the most common viscosity-reducing excipient, though most products do not contain any known viscosity modifier.¹ A majority of these DPs are formulated at pH less than 6.0, although approximately one quarter are formulated between pH 6.0 and 6.5.

Table 1.

A summary of commercial mAb or fusion protein drug products with protein concentrations ≥100 mg/mL. Included drug products are FDA and/or EMA approved. Some drug product protein concentrations may be ≥100 mg/mL after reconstitution of a lyophilized powder but are not presented in this table.

Drug Product	Protein Concentration	mAb Subclass	Formulation	FDA First Approval Year	Indication	Presentation
Piasky Genentech, Inc.	170 mg/mL	IgG1	histidine (47 mg/ml) arginine hydrochloride (21.1 mg/ml) poloxamer 188 (0.5 mg/ml) pH 5.8	2024	Hematology, Paroxysmal Nocturnal Hemoglobinuria	Vial
Hympavzi Pfizer Inc.	150 mg/mL	IgG1	histidine (1.12 mg/ml) histidine hydrochloride monohydrate (2.67 mg/ml) edetate disodium (0.05 mg/ml) polysorbate 80 (0.2 mg/ml) sucrose (85 mg/ml) pH 5.8	2024	Hematology, Hemophilia	PFS, AI
Alhemo Novo Nordisk Inc.	100 mg/mL	IgG4	histidine (5.12 mg/ml) arginine hydrochloride (5.27 mg/ml) phenol (3.5 mg/ml) polysorbate 80 (0.25 mg/ml) sodium chloride (1.46 mg/ml) sucrose (51.3 mg/ml) pH 6.0	2024	Hematology, Hemophilia	Pen
Spevigo Boehringer Ingelheim Pharmaceuticals, Inc.	150 mg/mL	IgG1	arginine hydrochloride (5.3 mg/ml) glacial acetic acid (0.32 mg/ml) polysorbate 20 (0.40 mg/ml) sodium acetate (3.25 mg/ml) sucrose (51.4 mg/ml) pH 5.2–5.8	2024	Immunology, Generalized Pustular Psoriasis	Vial, PFS
Ebglyss Eli Lilly and Company	125 mg/mL	IgG4	histidine (3.1 mg/ml) glacial acetic acid (0.9 mg/ml) polysorbate 20 (0.3 mg/ml) sucrose (59.8 mg/ml) pH 5.4–6.0	2024	Immunology, Atopic Dermatitis	PFS, AI
Leqembi Eisai, Incorporated	100 mg/mL	IgG1	histidine (0.18 mg/ml) histidine hydrochloride monohydrate (4.99 mg/ml) arginine hydrochloride (42.13 mg/ml) polysorbate 80 (0.50 mg/ml) pH 5.0	2023	Neurology, Alzheimer’s	Vial
Rystiggo UCB, Inc.	140 mg/mL	IgG4	histidine (1.05 mg/ml) histidine hydrochloride monohydrate (4.87 mg/ml) proline (28.78 mg/ml) polysorbate 80 (0.30 mg/ml) pH 5.6	2023	Immunology, Myasthenia Gravis	Vial
Beyfortus AstraZeneca AB	100 mg/mL	IgG1	histidine (2.2 mg/ml) histidine hydrochloride monohydrate (3.3 mg/ml) arginine (17 mg/ml) polysorbate 80 (0.2 mg/ml) sucrose (41 mg/ml) pH 6	2023	Infectious Disease, Respiratory Syncytial Virus (RSV)	PFS
Veopoz§ Regeneron Pharmaceuticals, Inc.	200 mg/mL	IgG4	histidine (0.575 mg/ml) histidine hydrochloride (1.325 mg/ml) arginine (10.5 mg/ml) polysorbate 80 (0.75 mg/ml) sucrose (10 mg/ml) pH 5.8	2023	Immunology, CHAPLE Syndrome	Vial
Bimzelx UCB, Inc.	160 mg/mL	IgG1	sodium acetate (2.83 mg/ml) glacial acetic acid (1.23 mg/ml) glycine (16.5 mg/ml) polysorbate 80 (0.4 mg/ml) pH 5.1	2023	Immunology, Plaque Psoriasis	PFS, AI
Omvoh Eli Lilly and Company	100 mg/mL	IgG4	citrate (2.1 mg/ml) citric acid (0.4 mg/ml) polysorbate 80 (0.3 mg/ml) sodium chloride (8.8 mg/ml) pH 5–6	2023	Immunology, Ulcerative Colitis	PFS
Vabysmo Genentech, Inc.	120 mg/mL	IgG1	histidine (3.1 mg/ml) methionine (1.044 mg/ml) polysorbate 20 (0.4 mg/ml) sodium chloride (73.1 mcg) sucrose (1.462 mg/ml) acetic acid (pH adjustment) pH 5.5	2022	Ophthalmology, Macular Degeneration	Vial, PFS
Adbry* LEO Pharma A/S	150 mg/mL	IgG4	sodium acetate trihydrate (6 mg/ml) acetic acid (0.3 mg/ml) polysorbate 80 (0.1 mg/ml) sodium chloride (5 mg/ml) pH 5.5	2021	Immunology, Atopic Dermatitis	PFS
Skyrizi Abbvie Inc.	150 mg/mL	IgG1	sodium acetate (0.75 mg/ml) glacial acetic acid (0.054 mg/ml) polysorbate 20 (0.2 mg/ml) trehalose (63.33 mg/ml) pH 5.7	2021	Immunology, Plaque Psoriasis	PFS, AI, OBI
Aduhelm§ Biogen Inc.	100 mg/mL	IgG1	histidine (0.60 mg/ml) histidine hydrochloride monohydrate (3.39 mg/ml) arginine hydrochloride (31.60 mg/ml) methionine (1.49 mg/ml) polysorbate 80 (0.50 mg/ml) pH 5.5	2021	Neurology, Alzheimer’s	Vial
Evkeeza Regeneron Pharmaceuticals, Inc.	150 mg/mL	IgG4	histidine (0.74 mg/ml) histidine hydrochloride monohydrate (1.1 mg/ml) arginine hydrochloride (14.8 mg/ml) proline (30 mg/ml) polysorbate 80 (1 mg/ml) pH 6	2021	Cardiology, Familial Hypercholesterolemia	Vial
Saphnelo AstraZeneca AB	150 mg/mL	IgG1	histidine (1.5 mg/ml) histidine hydrochloride monohydrate (3 mg/ml) lysine hydrochloride (9 mg/ml) polysorbate 80 (0.05 mg/ml) trehalose dihydrate (49 mg/ml) pH 5.9	2021	Oncology, Systemic Lupus Erythematosus	Vial
Susvimo§ Genentech, Inc.	100 mg/mL	IgG1	histidine hydrochloride (1 mg/ml) polysorbate 20 (0.1 mg/ml) sucrose (82 mg/ml) pH 5.5	2021	Ophthalmology, Macular Degeneration	Vial
Darzalex Faspro Janssen Biotech, Inc.	120 mg/mL	IgG1	histidine (0.3267 mg/ml) histidine hydrochloride monohydrate (1.2267 mg/ml) methionine (0.9 mg/ml) polysorbate 20 (0.4 mg/ml) sorbitol (49.001 mg/ml)	2020	Oncology, Multiple Myeloma	Vial
Vyepti Lundbeck Seattle BioPharmaceuticals, Inc.	100 mg/mL	IgG1	histidine (1 mg/ml) histidine hydrochloride monohydrate (2.8 mg/ml) polysorbate 80 (0.15 mg/ml) sorbitol (40.5 mg/ml) pH 5.8	2020	Neurology, Migraine	Vial
Enspryng Genentech, Inc.	120 mg/mL	IgG2	histidine (3.1 mg/ml) arginine (26.1 mg/ml) poloxamer 188 (0.5 mg/ml) aspartic acid (as needed for pH adjustment) pH 6.0	2020	Immunology, Neuromyelitis Optica Spectrum Disorders	PFS
Phesgo Genentech, Inc.	120 mg/mL	IgG1	histidine (0.44 mg/ml) histidine hydrochloride monohydrate (3.61 mg/ml) α, α-trehalose (39.7 mg/ml) methionine (1.49 mg/ml) polysorbate 20 (0.4 mg/ml) sucrose (34.2 mg/ml) pH 5.5	2020	Oncology, Breast Cancer	Vial
Ronapreve‡ Roche Registration GmbH	120 mg/mL	IgG1	histidine (0.74 mg/ml) histidine hydrochloride monohydrate (1.09 mg/ml) polysorbate 80 (1.00 mg/ml) sucrose (80.0 mg/ml) pH 6.0	2020	Infectious Disease, Covid-19	Vial
Beovu Novartis Pharmaceuticals Corporation	120 mg/mL	scFv	sodium citrate (2.58 mg/ml) polysorbate 80 (0.2 mg/ml) sucrose (58 mg/ml) pH 7.2	2019	Ophthalmology, Macular Degeneration	Vial, PFS
Xembify Grifols Therapeutics LLC	200 mg/mL	IgG	glycine (12.0–19.52 mg/ml) polysorbate 80 (10–40 mg/ml) pH 4.1–4.8	2019	Immunology, Primary Humoral Immunodeficiency	Vial
Herceptin Hylecta Genentech, Inc.	120 mg/mL	IgG1	histidine (0.39 mg/ml) histidine hydrochloride monohydrate (3.67 mg/ml) methionine (1.49 mg/ml) polysorbate 20 (0.4 mg/ml) α, α-trehalose dihydrate (79.45 mg/ml)	2019	Oncology, Breast Cancer	Vial
Aimovig Amgen Inc.	140 mg/mL	IgG2	acetate (2.0 mg/ml) polysorbate 80 (0.10 mg/ml) sucrose (65 mg/ml) pH 5.2	2019	Neurology, Migraine	PFS, AI
Takhzyro Takeda Pharmaceuticals U.S.A.	150 mg/mL	IgG1	histidine (7.8 mg/ml) citric acid monohydrate (4.1 mg/ml) polysorbate 80 (0.1 mg/ml) sodium chloride (5.3 mg/ml) sodium phosphate dibasic dihydrate (5.3 mg/ml) pH 6.0	2018	Other, Hereditary Angioedema	Vial, PFS
Ajovy Teva Branded Pharmaceutical Products R&D, Inc.	150 mg/mL	IgG2	histidine (0.543 mg/ml) histidine hydrochloride monohydrate (2.62 mg/ml) disodium ethylenediaminetetraacetic acid dihydrate (EDTA) (0.136 mg/ml) polysorbate 80 (0.2 mg/ml) sucrose (66 mg/ml) pH 5.5	2018	Neurology, Migraine	PFS, AI
Trogarzo Theratechnologies Inc.	150 mg/mL	IgG4	histidine (1.55 mg/ml) polysorbate 80 (0.45 mg/ml) sodium chloride (3.04 mg/ml) sucrose (52.03 mg/ml) pH 6.0	2018	Infectious Disease, Human Immunodeficiency Virus	Vial
Ilumya* Sun Pharmaceutical Industries Limited	100 mg/mL	IgG1	histidine (0.495 mg/ml) histidine hydrochloride monohydrate (1.42 mg/ml) polysorbate 80 (0.5 mg/ml) sucrose (70.0 mg/ml) pH 5.7–6.3	2018	Immunology, Plaque Psoriasis	PFS
Emgality Eli Lilly and Company	120 mg/mL,	IgG4	histidine (0.5 mg/ml) histidine hydrochloride monohydrate (1.5 mg/ml) polysorbate 80 (0.5 mg/ml) sodium chloride (8.8 mg/ml) pH 5.3–6.3	2018	Neurology, Migraine	PFS, AI
Dupixent Regeneron Pharmaceuticals, Inc.	150 mg/mL, 175 mg/mL	IgG4	histidine (3.1 mg/ml) arginine hydrochloride (5.25 mg/ml for 150 mg/ml or 10.53 mg/ml for 175 mg/ml) polysorbate 80 (2 mg/ml) sodium acetate (1 mg/ml) sucrose (50 mg/ml) pH 5.9	2017	Immunology, Atopic Dermatitis	PFS
Tremfya Janssen Biotech, Inc.	100 mg/mL	IgG1	histidine (0.6 mg/ml) histidine hydrochloride monohydrate (1.5 mg/ml) polysorbate 80 (0.5 mg/ml) sucrose (79 mg/ml) pH 5.8	2017	Immunology, Plaque Psoriasis	PFS, AI
Kevzara Sanofi-Aventis U.S. LLC	131.6 mg/mL, 175.4 mg/mL	IgG1	histidine (3.25 mg/ml) arginine (7.84 mg/ml) polysorbate 20 (2 mg/ml) sucrose (50 mg/ml) pH 6.0	2017	Immunology, Rheumatoid Arthritis	PFS, AI
Siliq* Bausch Health Ireland, Limited	140 mg/mL	IgG2	glutamate (4.3 mg/ml) polysorbate 20 (0.10 mg/ml) proline (24 mg/ml) pH 4.8	2017	Immunology, Plaque Psoriasis	PFS
Rituxan Hycela Genentech, Inc.	120 mg/mL	IgG1	histidine (0.53 mg/ml) histidine hydrochloride monohydrate (3.47 mg/ml) methionine (1.49 mg/ml) polysorbate 80 (0.6 mg/ml) α, α-trehalose dihydrate (79.45 mg/ml)	2017	Oncology, Non-Hodgkin’s Lymphoma	Vial
Hemlibra Genentech, Inc.	150 mg/mL	IgG4	histidine (3.1 mg/ml) arginine (26.1 mg/ml) poloxamer 188 (0.5 mg/ml) aspartic acid (pH adjustment) pH 6.0	2017	Hematology, Hemophilia	Vial
Benlysta GlaxoSmithKline LLC	200 mg/mL	IgG1	histidine (0.65 mg/ml) histidine monohydrochloride (1.2 mg/ml) arginine hydrochloride (5.3 mg/ml) polysorbate 80 (0.1 mg/ml) sodium chloride (6.7 mg/ml) pH 6.0	2017	Oncology, Systemic Lupus Erythematosus	PFS, AI
Anthim§ Elusys Therapeutics, Inc.	100 mg/mL	IgG1	histidine (6.2 mg/ml) polysorbate 80 (0.1 mg/ml) sorbitol (36 mg/ml) pH 5.5	2016	Infectious Disease, Anthrax	Vial
Cosentyx Novartis Pharmaceuticals Corporation	150 mg/mL	IgG1	histidine, histidine hydrochloride monohydrate (3.103 mg/ml) methionine (0.746 mg/ml) polysorbate 80 (0.2 mg/ml) trehalose dihydrate (75.67 mg/ml) pH 5.8	2015	Immunology, Plaque Psoriasis	PFS, AI
Humira AbbVie Inc.	100 mg/mL	IgG1	mannitol (42 mg/ml) polysorbate 80 (1 mg/ml) pH 5.2	2015	Immunology, Rheumatoid Arthritis	PFS, AI
Praluent Regeneron Pharmaceuticals, Inc.	150 mg/mL	IgG1	histidine (0.931 mg/ml) polysorbate 20 (0.1 mg/ml) sucrose (100 mg/ml) pH 6.0	2015	Cardiology, Hypercholesterolemia	PFS, AI
Repatha Amgen Inc.	140 mg/mL (PFS/AI), 120 mg/mL (OBI)	IgG2	acetate (1.2 mg/ml) polysorbate 80 (0.1 mg/ml) proline (25.43 mg/ml) pH 5.0	2015	Cardiology, Hypercholesterolemia	PFS, AI, OBI
Nucala GlaxoSmithKline LLC	100 mg/mL	IgG1	sodium phosphate dibasic heptahydrate (4.16 mg/ml) citric acid monohydrate (0.95 mg/ml) edta disodium dihydrate (0.019 mg/ml) polysorbate 80 (0.20 mg/ml) sucrose (120 mg/ml) pH 6.3	2015	Immunology, Asthma	PFS, AI
Hyqvia Takeda Pharmaceuticals U.S.A., Inc.	100 mg/mL		glycine (18.8 mg/ml) pH 4.6–5.1	2014	Immunology, Immunodeficiencies	Vial
Actemra Genentech, Inc.	180 mg/mL	IgG1	histidine (1.69 mg/ml) histidine hydrochloride monohydrate (1.93 mg/ml) arginine hydrochloride (21.11 mg/ml) methionine (4.48 mg/ml) polysorbate 80 (0.2 mg/ml) pH 6.0	2013	Immunology, Rheumatoid Arthritis	PFS, AI
Orencia Bristol-Myers Squibb Company	125 mg/mL	IgG1 Fc	dibasic sodium phosphate anhydrous (0.838 mg/ml) monobasic sodium phosphate monohydrate (0.286 mg/ml) poloxamer 188 (8 mg/ml) sucrose (170 mg/ml) pH 6.8–7.4	2011	Immunology, Rheumatoid Arthritis	PFS, AI
Hizentra CSL Behring AG	200 mg/mL	IgG	proline (28.8 mg/ml) polysorbate 80 (8 to 30 mg/l) pH 4.6–5.2	2010	Immunology, Immunodeficiencies	Vial, PFS
Ilaris Novartis Pharmaceuticals Corporation	150 mg/mL	IgG1	histidine (2.1 mg/ml) histidine hydrochloride monohydrate (1.3 mg/ml) mannitol (49.2 mg/ml) polysorbate 80 (0.4 mg/ml)	2009	Immunology, Periodic Fever Syndromes	Vial
Simponi Janssen Biotech, Inc.	100 mg/mL	IgG1	histidine, histidine hydrochloride monohydrate (0.87 mg/ml) polysorbate 80 (0.15 mg/ml) sorbitol (41.0 mg/ml) pH 5.5	2009	Immunology, Rheumatoid Arthritis	PFS, AI
Cimzia UCB, Inc.	200 mg/mL	Fab’ fragment	sodium acetate (1.36 mg/ml) sodium chloride (7.31 mg/ml)	2008	Immunology, Rheumatoid Arthritis	PFS
Gammagard Liquid Takeda Pharmaceuticals U.S.A., Inc	100 mg/mL	IgG, IgA, IgM	glycine (18.8 mg/ml) pH 4.6–5.1	2005	Immunology, Immunodeficiencies	Vial
Synagis AstraZeneca AB	100 mg/mL	IgG1	histidine (3.9 mg/ml) chloride (0.5 mg/ml) glycine (0.1 mg/ml) pH 6.0	2004	Immunology, Respiratory Syncytial Virus	Vial
Xolair Genentech, Inc.	150 mg/mL	IgG1	histidine (1.37 mg/ml) histidine hydrochloride monohydrate (2.34 mg/ml) arginine hydrochloride (42.1 mg/ml) polysorbate 20 (0.4 mg/ml)	2003	Immunology, Asthma	PFS, AI

FDA: United States Federal Drug Administration, PFS: prefilled syringe, AI: autoinjector, Pen: a combination device that does not involve triggered needle insertion and could be adjustable-dose, OBI: on-body injector. ^‡Ronapreve is only EMA approved and has Emergency Use Authorization (EUA) in the United States. *Adbry, Siliq, and Ilumya are marketed as Adtralza, Kyntheum, and Ilumetri respectively, in the EU. ^§Veopoz, Aduhelm, Susvimo, and Anthim are only marketed and approved in the United States. If a drug product has multiple indications, a representative one was chosen, usually corresponding to the first approved indication for a specific drug product.

Viscosity reduction mechanisms of commonly used additives

Since common viscosity-reducing excipients cannot be assumed to mitigate viscosity for every protein solution, the optimal formulation composition must be determined empirically, usually by focusing on known viscosity-reducing excipients such as arginine, sodium chloride, histidine, and proline.^{72,82,84,89–95} For many of these excipients, the mechanism of viscosity reduction is not well understood, though some hypotheses have been proposed. Arginine, which is the most commonly used viscosity-reducing excipient in approved DPs, is thought to modulate PPIs and reduce viscosity by interacting with hydrophobic regions of the protein and providing some charge shielding through its charged moieties (Table 1).^{89,93,96–98} That is, arginine may help modulate electrostatic forces and potentially disrupt hydrophobic interactions. Arginine has also been known to suppress protein aggregation by interacting with protein folding intermediates during the refolding processes used to generate purified proteins.⁹³ The observation that arginine prevents aggregation suggested its utility in protein formulation to mitigate PPIs driven by electrostatic and/or hydrophobic interactions.^96,99 In one study with lysozyme, Ito and colleagues showed that arginine forms specific interactions with aromatic residues at the protein surface.¹⁰⁰ Related work showed that arginine and lysine both reduce the viscosity of human serum albumin (HSA) to similar levels as sodium chloride and guanidine hydrochloride, supporting the utility of arginine as a viscosity-reducing excipient for some protein formulations.⁹⁶ To further test arginine’s viscosity reduction capacity, Inoue et al. compared the performance of arginine on a panel of globular or antibody proteins.⁸⁹ They concluded that arginine might act as a viscosity-reducing excipient for antibody proteins, and not globular proteins. The authors speculate that this selectivity may come from the antibodies’ unique shapes and sequences. However, some antibodies do not exhibit viscosity reduction upon the addition of arginine. For example, work from Somani and colleagues combined molecular dynamics (MD) simulations and protein docking to suggest that arginine likely forms specific interactions with the target protein, although the result is an increase in the solution viscosity of their test mAb. They emphasized that excipients can also have indirect effects on PPIs, making their mechanism of viscosity reduction complex.⁹⁷ For example, arginine may self-associate in solution, which may be related to the observation of protein dehydration in the presence of arginine.⁹⁷ In another example with the NISTmAb, arginine does not act as a viscosity reducer, but other amino acid excipients do.²³

Other amino acids, such as histidine and proline, exhibit viscosity-reducing effects in high-concentration protein solutions.^{80,84,90–92} Zidar and colleagues evaluated the effect of combining different excipients for viscosity reduction and show that the combination of arginine and histidine resulted in moderate viscosity reduction.⁶⁵ Comparatively, histidine shows viscosity reduction similar to arginine alone.⁶⁵ Additionally, as shown in Table 1, histidine is the most common buffer component, which is likely a consequence of both its pKas and potential capacity for viscosity reduction.¹⁰¹ Histidine may act as a viscosity-reducing excipient through its capacity to form a variety of interactions with the protein, including cation-$\mathit{\pi }$, $\mathit{\pi }-\mathit{\pi }$ stacking, hydrogen bonding, and electrostatic interactions.¹⁰¹ These interactions are possible due to the aromaticity of the imidazole ring and the charged amino and carboxylic acid moieties.¹⁰¹ Wang and colleagues observed that free amino acids (e.g., arginine, histidine, proline) decrease solution viscosity to a different extent for two mAbs. This finding suggests the mAb’s sequence and its specific chemical properties, such as charge patterning or hydrophobicity, affect which excipient can reduce viscosity.^90,96,102 Wang et al.’s study focused on the measured viscosity of two antibodies (both IgG1) in the presence of different amino acid salts.⁹⁰ They found the tested antibodies show no change in viscosity upon the addition of glycine, while viscosity decreases upon the addition of arginine or sodium chloride. This is in contrast to the NISTmab, which did not exhibit arginine-mediated viscosity reduction.²³ In the Wang et al. study, one of the tested antibodies exhibits a biphasic viscosity change upon the addition of proline, where viscosity decreases until 150 mM proline and increases at concentrations of 150–300 mM proline. In comparison, the second antibody only shows a slight decrease in viscosity with proline, which again emphasizes that the properties of the mAb under investigation affect the viscosity-reducing effect of a given viscosity-reducing excipient. By contrast, Hung and colleagues observed only a monotonic decrease in viscosity in the presence of proline.⁹¹ Proline is amphipathic and has a cyclic hydrophobic hydrocarbon side chain, but retains the polar carboxylic acid and amino groups. These characteristics, in theory, allow proline to suppress PPIs by interacting with hydrophobic, polar, and charged regions on a protein.

While arginine seems to both modulate electrostatic attractions and SRA like hydrophobicity, salts like sodium chloride are limited to screening charge–charge interactions.^{47,52,65,95,96,103,104} Of the 55 products listed in Table 1, only 9 contain sodium chloride, although many contain hydrochloride as counterions for the histidine buffer components. Thus, charge screening due to chloride ions may still be present and strongly contributing to DP properties. In probing the basis of a mAb’s high viscosity, Liu et al. showed that increasing sodium chloride concentration led to an exponential decrease in viscosity.⁴⁷ Additionally, Yadav and colleagues showed that ionic strength modulated apparent charge, affected PPIs, and reduced viscosity.⁵⁷ In that work, PPIs, as measured by either solution storage modulus or k_D reach their maximum at the pH at which net charge equals zero and in the absence of ions. At 150 mM ionic strength, PPIs only appear to change modestly across pH levels.⁵⁷ Those results suggest the strong contributionelectrostatics may have to the tested mAb’s interactions, whereby sodium chloride reduces viscosity presumably due to charge screening. However, this effect may not be observed for all mAbs and, in fact, sodium chloride may increase the viscosity of some mAb solutions.^23,28 One theoretical approach suggested that it may only be possible to reduce electrostatic attractions to a certain extent via salt addition alone, and thus only reduce viscosity to an unspecified limit.¹⁰³ Since salts shield charge, protein particles with increasing salt concentrations may begin to behave as uncharged or net neutral spheres, which may enable increased viscosity via hydrodynamic interactions.¹⁰³ Another possibility is that non-electrostatic interactions such as hydrophobic interactions are the primary drivers of PPIs in the case of mAbs, and therefore solution viscosity is unaffected with higher ionic strengths.

Novel viscosity-reducing excipients include hydrophobic salts and caffeine, which may interact with hydrophobic patches on proteins to disrupt PPIs (akin to arginine). As an initial demonstration, Du and Klibanov showed that adding hydrophobic salts, at concentrations up to 500 mM, reduces the viscosity of high-concentration protein solutions independent of the solution pH.⁸¹ They assessed benzenesulfonate, tetramethylammonium, tetraethylammonium, and camphor sulfonate as salts with various counter ions, including chloride, bromide, fluoride, and iodide. These hydrophobic cations can reduce protein solution viscosity, particularly in the presence of anions that disrupt hydrogen bonding with water molecules (i.e., chaotropic anions). The authors hypothesized that the hydrophobic cation forms contacts with hydrophobic portions of the protein. A follow-up study from Guo et al. tested numerous hydrophobic salts with increasing carbon chain lengths, different salt forms, and alternate connectivity (e.g., moving a moiety from the para to the ortho substituted position on an aromatic ring).¹⁰⁵ By testing a series of molecules with related structures, that work suggests that the structure of the anion matters in as much as it is a hydrophobic moiety with large molar volume (i.e., ‘aliphatic and bulky’), while the structure of the cation may not be a strong contributor to the salt’s viscosity decreasing effect. These studies suggest the potential of using hydrophobic salts to reduce viscosity, though the mechanism remains unclear. It must be noted, however, that novel excipients, such as these hydrophobic salts, require extensive toxicology and immunology studies to assess their safety prior to use in clinical and commercial DPs. Accordingly, the inclusion of novel excipients in drug products poses substantial timeline and regulatory risks.

Caffeine may also have utility in reducing protein solution viscosity potentially via hydrophobic interactions with protein surfaces.⁸² Caffeine is both a weak acid and a weak base with pKa values of 14.0 and 0.7, and thus it is not ionized under common pH conditions for mAb formulation. Zeng et al. showed that 75 mM caffeine induces significant viscosity reduction with infliximab and ipilimumab, exceeding the effect of arginine or sodium chloride.⁸² Measurements of the k_D showed that caffeine increases k_D, indicating a reduction of attractive forces. Conversely, arginine and sodium chloride reduce k_D, indicating increased attraction. Furthermore, their intrinsic protein fluorescence data suggest that caffeine forms direct, but weak, interactions with the protein molecules. They hypothesized that this direct interaction leads to viscosity reduction. Other work suggested that caffeine can form $\mathit{\pi }-\mathit{\pi }$ interactions with surface-exposed phenylalanine residues and interact with other hydrophobic residues such as isoleucine.^106,107 Finally, Zeng et al. showed that caffeine did not affect protein physical stability, suggesting that it may be a useful viscosity-reducing excipient when hydrophobicity is known to be a driver of high viscosity. However, the generalizability of caffeine as a viscosity-reducing excipient remains untested. Interestingly, while caffeine acted as a viscosity-reducing excipient in the work on IgG1 mAbs from Zeng et al., it also can induce crystallization of pembrolizumab, an IgG4 mAb.^108,109 Peak shifts in the NMR spectra indicated a potential direct binding of caffeine to pembrolizumab.¹⁰⁸ The phase behavior of proteins (i.e., soluble vs. crystalline), especially at high concentrations, can be quite complex, and excipients that act as viscosity reducers in one case may induce dramatic changes to the solubility and viscosity of a different protein. These complexities further highlight the difficulty of generalizing excipient choices for viscosity reduction in high-concentration protein solutions. While caffeine shows some promise, it is not currently an excipient or additive in any marketed high-concentration antibody DPs (Table 1). Biotechnology companies have also developed proprietary viscosity-reducing excipients and combinations thereof, as discussed by Braun et al.⁷²

In addition to the identification of novel viscosity-reducing excipients, alternative formulation approaches are in development. For ultra-high-concentration (UHC) (>200 mg/mL) biologics, there is increasing interest in novel formulation strategies, such as suspensions of spray dried or spray frozen dried proteins, or protein crystals for SC administration.^110–114 Antibody microparticles formed by spray or spray freeze-drying have been suggested to be physically and chemically stable, active, and can be formulated in low-volume, aqueous and non-aqueous suspensions (e.g., IgG1 in water for injection or propylene glycol).^80,114,115 It may also be possible to use spray (freeze) drying to generate protein particles for inhalation.^80,112,113 Similarly, batch crystallization methods may be used to form protein crystals for subsequent suspension.^114,116,117 Initial data suggest that these approaches may yield potent, active, stable, and low-viscosity DPs.¹¹⁴ However, the safety, reproducibility, long-term stability, and manufacturing processes for these formulations are not well characterized or established, limiting their widespread application in clinical or commercial DP development. For many of these formulation strategies, the optimal composition of the protein solution prior to processing (e.g., drying, freeze-drying, crystallization) must still be identified. Additionally, processing and stability concerns make these approaches more complex than ‘traditional’ solution formulations. Alternative formulation strategies are reviewed in further detail elsewhere.^21,118

Finally, for solution protein formulations, it is of interest to ask, what is the maximum achievable mass per volume? Notably, as shown in Table 1, the highest protein concentration in any DP is 200 mg/mL. Garidel and colleagues attempted to identify the theoretically maximum mAb concentration based on the proposed arrangement of a densely packed fluid.¹⁸ Here, Garidel et al. consider a face-centered cubic packing arrangement and model mAbs as spheres. In this arrangement, they calculated that mAbs can have a theoretical maximum packing fraction of 0.74, corresponding to approximately 413 mg/mL.¹⁸ This calculation is truly an upper limit since a fluid will unlikely adopt crystalline packing. This calculation also accounts only for steric considerations, while intramolecular and intermolecular interactions, such as charge–charge repulsion, can result in larger “effective” molecular volumes. Furthermore, mAbs do not actually pack as non-overlapping spheres in the close packing that crystal lattice models assume.¹⁸ These caveats would further limit the maximum achievable concentrations to well below 400 mg/mL.^75,119,120 Thus, for cases where such high concentrations may be required, UHC approaches may enable the development of DPs at concentrations exceeding 200 mg/mL.

Rheological characterization of drug product solutions provides insight for improved formulation composition design

The viscosity of a DP solution is robustly assessed during development given its potential impact on manufacturability, injectability, and drug–device combination product functionality. Given that viscosity measurements are highly dependent on the method, temperature, and shear rate, studies assessing formulation composition require careful experimental design and data interpretation (Figure 2). Viscometers are the most widely used instruments for viscosity characterization, given that most protein DP solutions exhibit Newtonian behavior in the shear rate ranges relevant for injection, device functionality, and manufacturing. Typically, the zero-shear viscosity is extrapolated from viscometer measurements at shear rates varying across an order of magnitude, assessing if the fluid exhibits Newtonian behavior. If the measured viscosity is observed to be independent of the shear rate, the solution is assumed to be Newtonian, and the average of the measured shear-independent viscosity values is taken as the zero-shear viscosity. The different viscometer technologies and their application to injectability and formulation development, in the context of rheology fundamentals, are reviewed elsewhere.^25,46,121

It must be noted that, at high concentrations, antibodies may also exhibit non-Newtonian behavior (e.g., shear thinning) at high shear rates.⁴⁹ The exact concentration and shear rates at which antibody solutions exhibit non-Newtonian behavior are dependent on protein sequence and formulation composition. Both shear thinning and exponential increases in viscosity with increasing protein concentrations suggest the solution may be non-Newtonian. Proteins’ non-Newtonian behavior is due, in part, to the non-spherical shape of therapeutic proteins like antibodies, as well as PPIs. Nonetheless, at sufficiently high packing fractions, even isotropic systems, such as hard sphere particles exhibit non-Newtonian behavior.^48,122 As a result, understanding what protein concentration may lead to non-Newtonian behavior is of interest for DP design. For example, a shear thinning solution may exhibit longer than anticipated injection time at low shear rates due to relatively higher viscosity at lower shear rates.¹²³ Accordingly, protein solution viscosity should be characterized at the shear rates that correspond to the injection shear rates, whether via manual injection or delivery devices. In general, it should be noted that the high packing fraction at which mAbs begin to exhibit exponential viscosity increases is approximately the same for all mAbs and can be predicted from theory.⁵⁵

By definition, non-Newtonian fluids exhibit shear rate-dependent viscosity, demonstrating the potential utility of rheometers to control the shear rate and shear stress for deeper understanding of solution properties, such as the storage (G’) and elastic/loss (G’’) moduli,^51,57 which are descriptions of the energy stored and dissipated during fluid deformation, respectively.⁵¹ Changes in G’ at different formulation conditions may indicate whether a change in viscosity is due to changes in PPIs or other mechanisms, such as hydrodynamics.⁵¹ Comparably, changes in G’’ may also reflect changes in PPIs, but are harder to interpret since G’’ can both increase and decrease as the solution relaxation time (τ) increases.¹²⁴ This level of characterization often falls outside the scope of formulation and drug development efforts but has been useful for understanding the origins of viscosity.⁵⁷

Pham and colleagues showed that the elastic (loss) modulus is linearly correlated with protein physical stability; specifically, visible particle counts.¹²⁵ This study measured G’ and G’’ of two IgG1 mAbs, three IgG4 mAbs, and one bispecific mAb to connect interfacial properties to physical stability. Measurement of G’ and G’’ may be useful to determine if a given formulation composition modulates PPIs. Pham et al.’s work shows that changes in G’ are correlated with product attributes such as the formation of both visible and subvisible particles. This may suggest that particle formation is also a consequence of PPIs, though more studies are needed. Measurement of G’ and G’’ may require complex rheological techniques, but such parameters are relevant for the development of high-concentration DPs because they may be predictive of important DP attributes. Furthermore, since proteins generally exhibit non-Newtonian behavior at high packing fractions (volume fraction) rheological techniques that can appropriately assess solution properties such as G’ and G’’ of non-Newtonian fluids can be useful.

Experimental techniques for characterization of high-concentration protein solutions and understanding PPIs

Although not widely used in formulation composition design, the available literature demonstrates the utility of additional biophysical techniques, such as light scattering, self-interaction nanoparticle spectroscopy (SINS), mass spectrometry (MS), and nuclear magnetic resonance (NMR), to improve the efficiency of formulation design. These techniques enable assessment of PPIs and solution structure and provide insight about product viscosity and stability.

Scattering techniques for assessing PPIs

Small-angle X-ray or neutron scattering (SAXS/SANS) involves subjecting a sample to an incident X-ray or neutron beam and detecting the resulting scattering intensities over a wide range of wavevectors, q, or scattering angles. These measurements allow for capturing the structure of the molecule and solution over a wide range of length-scales (from 10⁻¹ to 10² nm), enabling detection of short- and long-range interactions and providing information about the shape and size associated with the molecule.

The scattering intensities from SAXS measurements of antibodies at q values smaller than 0.01 Å⁻¹ (equivalent to ~6 nm or about the size of the Fab or Fc domains) correspond to information related to interactions between mAb domains, and thus it also provides information about the structural arrangement of molecules in solution (i.e., the solution structure).¹²⁶ For protein non-mAb molecules, the range of q values that inform about solution structure depends on the size of the molecule and effective range of the PPIs, but it similarly represents the equivalent distance between nearest neighbor interacting molecules. The molecular arrangement of an antibody solution is most evident from analysis of the structure factors of high-concentration protein solutions. For example, Chowdhury and colleagues measured scattering profiles for different mAbs up to 200 mg/mL in the presence of arginine or sodium chloride¹²⁷ and calculated the corresponding structure factors.¹²⁷ Thereafter, they fit 12-bead, CG mAb simulations to the experimental structure factors and extracted changes in PPIs, nearest neighbor interactions, and the cluster size distribution (in terms of both radial dimension and number of monomers within a cluster). Notably, the cluster size distribution is used to predict the solution viscosity for their test antibodies with a high degree of accuracy relative to experimental values.¹²⁸ Importantly, in this work, Chowdhury et al. identified particular features of their test mAb’s structure factor to gain insight into PPIs and solution structure for characterizing the origin of solution viscosity across length scales (from monomers to oligomers to solution scale). The value of the structure factor at low q values (long length scales) corresponds to the overall repulsion (S(0) less than 1) or attraction (S(0) greater than 1) that the molecule exhibits.¹²⁹ At the intermediate q range, near 0.06 Å⁻¹ (~10 nm) and 0.1 Å⁻¹ (~6 nm), mAbs may exhibit two characteristic peaks, P1 and P2, respectively. P1 in S(q) is thought to correspond to local molecular arrangements at the first coordination shell, while P2 is thought to characterize domain–domain structuring in the case of mAbs.¹²⁷ Thus, Chowdhury et al. suggested that future work could analyze S(q) in terms of appearance/disappearance of P1 and P2, as well as changes in the q value over which P1 and P2 are centered. They suggest that changes in q value for P2 may indicate changes in the inter-domain spacing. Chowdhury and colleagues then applied this understanding to interpret changes in the structure factor as a function of formulation conditions (arginine versus sodium chloride) and IgG subclass.⁵³ These studies suggest that there are few differences in the structure factor between mAb formulations in terms of clustering and intramolecular interactions. Instead, most changes in structure factor occur at long length scales, suggesting that arginine and sodium chloride might mainly modulate long-range interactions. Similarly, they showed that the cluster size distribution can be used to estimate the solution viscosity with relatively high accuracy.¹³⁰ Together, their data suggest that clusters of increasing size give rise to high viscosity in this test system and formulations that reduce viscosity shift the cluster size distribution toward smaller clusters overall.

SAXS’s ability to simultaneously probe multiple-length scales could enable screening for formulation composition design. At low protein concentrations, scattering profiles can be analyzed to calculate an estimate of the radius of gyration (R_g) of the protein in solution to support the identification of formulation compositions that may exhibit PPIs unfavorable for further development.^131,132 Increases in R_g beyond the typical size of a mAb may suggest that PPIs are giving rise to higher-order species (dimers, clusters) in a given solution condition, which might be used for screening stable, viscosity-reducing formulations. Many additional studies have suggested the utility of SAXS for formulation screening and development.^{23,126,131–134} At dilute conditions, multiple parameters in addition to R_g can be derived from a single sample’s scattering data, including the pair-distance profile P(r) and the maximum particle dimension (D_max). In one example, Fukuda et al. collected scattering profiles of mAb solutions with citrate or histidine buffers at medium and high salt conditions. Their data show that viscosity increases with citrate buffer versus histidine and with low salt versus high salt. From the scattering profiles in these same conditions, they computed the R_g, P(r), and D_max. R_g and D_max at 60 mg/mL exhibited strong linear correlations with log-transformed viscosity at 200 mg/mL in all formulation conditions tested, suggesting that SAXS-derived parameters may be used to screen formulation compositions for viscosity reduction. Conversely, SAXS analysis of high-concentration formulations can be used to extract information about protein structure factor at zero wavevector (S(0)). Xu et al. showed that S(0) exhibits a strong correlation with viscosity and stability (T_onset) across a wide variety of formulation conditions with the NISTmAb (Figure 4C).²³ Thus, it may also be possible to screen for stable, viscosity-reducing compositions by collecting scattering intensities, and only selecting the subset of the formulations that exhibit low levels of PPIs by R_g, D_max, and S(0) for confirmatory measurements.²³ This approach could be data rich, providing similar information to other methods, such as chromatography, viscosity, and other light scattering techniques. However, additional efforts are necessary to establish comparability between SAXS and traditional methods used for drug product characterization.

Figure 4.

Correlations of sample viscosity with A) k_D, B) B₂₂, and C) S(0). Only S(0) exhibits clear correlation with viscosity. All panels reprinted with permission from “studying excipient modulated physical stability and viscosity of monoclonal antibody formulations using small-angle scattering” Amy Yuanyuan Xu, Maria Monica Castellanos, Kevin Mattison, Susan Krueger, and Joseph E. Curtis molecular Pharmaceutics 2019 16 (10), 4319–4338 DOI: https://doi.org/10.1021/acs.Molpharmaceut.9b00687 copyright 2019 American chemical Society.²³

Similar to SAXS, static and dynamic light scattering (SLS/DLS) experiments allow measurement of the second virial coefficient (B₂₂ or A₂) and the (k_D), respectively. Both SLS and DLS are commonly deployed for formulation composition design since both B₂₂ and k_D are functions of PPIs and colloidal stability, where positive (negative) values of either parameter indicate repulsive (attractive) PPIs.^52,135,136 Notably, the definitions of both parameters are for dilute solution conditions, which limits the provided information about PPIs to pairwise interactions. Consequently, both SLS and DLS measurements occur at low protein concentrations ( <20 mg/mL) to assess B₂₂ and k_D, respectively.¹³⁷ Despite being dilute solution-derived parameters, many studies show simple statistical correlations of both B₂₂ and k_D to viscosity at high concentrations, suggesting that these dilute measurements may have use in candidate selection and formulation screening.¹³⁸ The theoretical basis for these parameters relates them to self-interaction, which contributes to both viscosity and physical stability. However, neither parameter captures all of the interactions a protein can experience at high-concentration conditions (e.g., multi-body interactions, molecular crowding).^52,137 Yet these parameters are still useful since qualitatively, if the solution exhibits negative B₂₂ or k_D values, then the PPIs correspond to net attractive interactions, whereas positive B₂₂ or k_D values indicate net repulsive interactions.¹³⁶ However, unlike B₂₂, for k_D the switch from the repulsive regime to the attractive regime at zero does not reflect net zero PPIs since k_D includes a hydrodynamic contribution.¹²⁶ While formulation composition changes can affect the thermodynamics of PPIs (as measured by B₂₂), k_D includes both thermodynamic and hydrodynamic contributions to PPIs as described by the Stokes-Einstein-Sutherland equation for diffusion.^139–143 While both B₂₂ and k_D may be statistically correlated to viscosity, careful interpretation of the correlation between B₂₂ and viscosity must be taken as this parameter describes only thermodynamic interactions. Comparatively, k_D captures the effect from hydrodynamic forces, but it is possible that the contribution of hydrodynamics to viscosity at high concentrations is more significant than what k_D captures in the dilute regime. Hydrodynamic effects are important to consider in the development of viscosity-reducing formulation compositions because changes in formulation conditions may affect protein-solvent/excipient interactions and solvent dynamics. Because k_D implicitly accounts for hydrodynamic effects, researchers should be careful in their interpretation of k_D values if their goal is to better understand PPIs. At a given solution condition, the contribution of the hydrodynamic effects to k_D may exceed the contribution of PPIs, making the use of k_D as a measurement of PPIs potentially misleading.¹³⁹ As a result, there is no unified theoretical relationship between either of these two parameters and viscosity at high concentrations. Yet both of these experimental approaches can be informative to evaluate how formulation composition modulates PPIs, and thereby, viscosity, and aggregation.¹⁴⁴

Many studies have assessed the correlation between B₂₂ or k_D and viscosity. The general trend is that viscosity increases as B₂₂ or k_D decreases. That is, as PPIs become less repulsive or more attractive, viscosity is more likely to increase for a given DP formulation. The extent of this negative statistical correlation between viscosity and k_D varies with the molecule and formulation, with a correlation coefficient ranging as high as 0.96 for some datasets.⁵² This trend appears across IgG subclasses at different pH and ionic strength conditions, and within sequence variants of one starting molecule.^{52,138,145–150} The quantity of molecules tested for correlation of B₂₂ or k_D with viscosity has been as high as 83 unique IgG1s.¹⁵¹ In disagreement with other studies, Xu and colleagues’ tests on the NISTmAb in a variety of formulations did not find an overall correlation of B₂₂ or k_D with viscosity (Figure 4A,B).²³ However, they did find that S(0) is correlated with viscosity (Figure 4C). In this study, they tested several types of excipients including surfactants, sugars, and amino acids. The lack of correlation of B₂₂ and k_D with viscosity is interesting considering that all three parameters (S(0), B₂₂, k_D) are related to nearest neighbor interactions. However, while S(0) from the structure factor should report on overall repulsion/attraction similarly to B₂₂, it is a high-concentration measurement unlike B₂₂. Xu et al.’s results suggest that the PPIs measured in dilute conditions are not representative of the interactions proteins exhibit at high concentrations due to multibody and crowding effects.

In general, it is informative to determine B₂₂, k_D, and S(0) as early as possible during candidate selection and formulation development stages. Assessment of B₂₂ and k_D via SLS/DLS requires little protein and can be rapidly measured for multiple formulations in a multiwell plate format with the latest technologies.¹⁵² On the other hand, SAXS/SANS not only offer the capability to extract S(0) but also to probe the solution structure and protein–protein interactions over a wide range of length scales. As such, information drawn from SAXS/SANS analysis might provide further mechanistic understanding of the impact of formulation composition on protein intermolecular interactions and solution viscosity. However, SAXS/SANS data collection and analysis can be more time-consuming than SLS/DLS, and accessibility to synchrotron and nuclear facilities with suitable instruments is significantly limited. Benchtop SAXS instruments can also be used to evaluate protein solutions, though data collection time greatly increases from a few minutes in a synchrotron-based instrument to several hours in a benchtop instrument. As a result, more targeted use of SAXS/SANS is appropriate, as its implementation during early-phase candidates and formulation screening might be challenging.

Nanoparticle spectroscopy may have utility for assessing PPIs

The SINS approach is increasingly used to quantify protein PPIs. There are two main versions of SINS in the literature: affinity capture SINS (AC-SINS) and charge stabilized SINS (CS-SINS).^153–158 Generally, SINS involves the functionalization of nanoparticles that can undergo a change in plasmon resonance upon binding of a partner protein to the immobilized probe protein.^154,155 This change in plasmon resonance can be measured as a shift of the solution’s absorbance maximum. SINS-based methods allow quantification of the relative extent of PPIs. As the first SINS iteration, AC-SINS was thought to potentially be an alternative to measuring B₂₂ for screening protein crystallization solution conditions. Yet AC-SINS is limited in its utility for formulation development because the required pH of the test solution is higher than typical mAb formulation conditions. As such, SINS measurements are tested only in limited buffer and ionic strength conditions. Starr and colleagues demonstrated that introducing poly-lysine to the AC-SINS protocol incorporates charge–charge repulsion into the antibody-coated particles and increases the nanoparticle colloidal stability.¹⁵³ Stable nanoparticles allow measurement of the SINS score at representative formulations of pH and ionic strength conditions.¹⁵³ In a dataset of more than 50 commercially available mAbs, they found a strong correlation of SINS score with viscosity across IgG subclasses. However, the correlation between SINS score and viscosity was weaker when computed for the IgG4 subclass alone, though this test dataset had significantly fewer IgG4 molecules than IgG1 and IgG2. Similarly, Mock and colleagues recently showed that SINS results are correlated with protein solution viscosity.¹⁵¹

One limitation of the SINS approach is that the SINS score is based on the relative absorbance shift of antibodies used in the study rather than an absolute thermodynamic measure of association. However, it is worth noting that the antibodies used in the study from Starr et al. are representative of pharmaceutical antibodies.¹⁵³ In their work, Starr and colleagues compared the charge, hydrophobicity, and charge symmetry for their mAb samples to the distribution of these properties in the Therapeutic Antibody Database (https://tabs.craic.com). They found that the antibodies used to test and calibrate the SINS method were representative of molecules in the Therapeutic Antibody Database. This suggests that the SINS score, while bracketed in relative terms, can still be useful if the calibration and reference data are representative of truly high and low self-associating proteins. However, if the reference data are not sufficiently diverse, this method cannot provide a measure of PPIs since the score is only relative to the reference data. To improve the utility of SINS in formulation screening and development, three main barriers must be addressed: 1) the difficulty of establishing controls, 2) unclear relationships between the SINS score and physical stability, and 3) the lack of a clear relationship to the physical principles underlying viscosity. Firstly, for formulation development, SINS may be most useful to screen for compositions that reduce PPIs. Because screening involves formulation components of a wide variety of chemistries and spectral properties, it is possible that some formulation compositions may not have acceptable absorbance properties to establish the necessary control values that are used to calculate the SINS score. Secondly, a useful formulation composition screening readout would have a well-established relationship to both physical stability and solution viscosity, but the relationships between the SINS score and physical stability as well as viscosity are unclear both theoretically and heuristically. Finally, a major limitation of SINS is that the resulting score reflects PPIs in a physically unrepresentative state: one partner in an interacting pair being immobilized on a particle an order of magnitude larger than the probe molecule itself, and the interacting target. Furthermore, while the origins of plasmon shifts are well described, the solution condition limitations associated with SINS make relating the observed self-interaction to its physicochemical basis difficult.

Protein structural techniques for understanding protein-excipient interactions and the effect of formulation composition on PPIs

While experiments that measure solution properties are useful, researchers may need to rely on higher resolution structural methods, such as MS and biomolecular NMR to better understand the underlying PPIs driving high viscosity. Both NMR and MS can be used to identify the regions of the protein involved in protein–protein or protein-excipient interactions.¹⁵⁹ While current formulation composition development approaches rely on empirical assessments and tuning of protein solution viscosity via composition changes, understanding the protein structural determinants of viscosity can help drive more efficient formulation screening. For example, if the relative contributions of charge and hydrophobicity to protein self-association are understood, it may be possible to rationally select viscosity-reducing excipients or combinations thereof. However, this idealized workflow is further limited by poor understanding of the viscosity-reducing mechanisms of known viscosity-reducing excipients.

Mass spectrometry as an approach for mapping PPIs

MS offers high spatial and structural resolution insights, which can be used to understand the effect of formulation composition on PPIs.¹⁶⁰ While it is theoretically possible to use MS to understand the effect of excipients on protein structure and conformation, most approaches leverage footprinting techniques to understand the structural determinants of protein self-association. Non-footprinting based approaches include crosslinking,¹⁶¹ native, and ion mobility MS,¹⁶² and are best suited for characterizing protein structure and chemical stability as a function of formulation. Footprinting approaches involve generating a state of interest (monomer or complexed protein) and applying a footprinting reagent that will undergo a reaction with the complex to shift the mass of the protein or protein fragment. The comparison of the mass spectrum of the state of interest to a reference state could identify the regions of the protein that contribute to forming the state of interest. One of the most common footprinting approaches is hydrogen-deuterium exchange (HDX-MS) in which the mass shift due to the uptake of deuterium ions can be followed with a mass spectrometer and ‘mapped’ back onto the protein structure. This allows for the identification of the exact residues or structural regions that are involved in PPIs. In one such case, HDX-MS was used to identify the groups of residues underlying the PPIs of a therapeutic mAb with intractable solution properties (aggregation and high viscosity at low concentrations).¹⁶¹ Then, in combination with in-silico analyses, such as comparing structural regions protected from deuterium uptake with an aggregation propensity calculator, the researchers identified mutations that remediated the aggregation and viscosity concerns. Such an approach illustrates the utility of HDX-MS for mapping PPIs and its potential for elucidating the interactions between excipients and proteins, which would have direct applicability to developability considerations and formulation composition design. Hu and colleagues further applied HDX-MS to map the conformational dynamics of the protein backbone in the presence and absence of different charged amino acids.¹⁶³ Their results suggest that viscosity-reducing additives such as arginine, lysine, glutamate, and aspartate result in deprotection of self-association sites in the test mAb. Specifically, HDX data show protection of the variable domains of the Fab and part of the Fc. Upon the addition of arginine, the Fab variable regions uptake more deuterium than in arginine’s absence, indicating a change in the solvent accessibility of those structural regions. These results suggest that these regions are involved in PPIs, with arginine disrupting interactions at those sites. By contrast, negatively charged excipients such as glutamate and aspartate were less effective at reducing PPIs at these sites. Ultimately, this study provided proof of concept for the applicability of HDX-MS for selecting excipients that reduce PPIs and viscosity. Using MS to explore the protein structure determinants of PPIs and protein-excipient interactions remains an active area of research.^{160,161,163–167}

Nuclear magnetic resonance suggests the basis of protein-excipient interactions and PPIs

Similar to MS-based approaches, NMR offers high levels of protein structural resolution which can be of use in formulation composition design. To understand the relationship between formulation composition and solution properties, one study used NMR to examine the interaction of sucrose with a multivalent antibody fragment protein.¹⁵⁹ Two-dimensional spectra showed no observable interactions between sucrose and the protein molecule, supporting the excluded volume effect as the origin for sucrose stabilization.^59,168,169 One study examined the effect of buffer molecules on the spectra of a Fab domain and showed that buffer components may induce changes to specific residues’ local environments, shifting their peaks positions.¹⁷⁰ These residue-specific changes may be due to direct interactions with the protein or the consequence of local structural rearrangements the buffer molecules otherwise induce. Comparatively, a separate study examined the interaction of the viscosity-reducing combination excipient, arginine glutamate, with a test mAb.¹⁷¹ In this work, arginine glutamate reduced sample viscosity the most at ~150 mM. Proton spectrum intensity, translational diffusion coefficients, and transverse relaxation rate measurements suggest that arginine glutamate directly interacts with the antibody with a dissociation constant of 90 mM.¹⁷¹ Dose-dependent changes in the protein spectra upon the addition of arginine glutamate suggest that the molecule forms direct interactions with the protein (changing the local environment as reflected in a shift in peak intensity) to reduce viscosity. Furthermore, this study also suggests that the addition of arginine glutamate has little to no effect on protein diffusion, but does induce a decrease in protein tumbling (R₂), potentially reflecting a decrease in antibody self-association, a reduction in solution viscosity, and a decrease in antibody hydrodynamic radius (R_H). These NMR data suggest that the solution effects of viscosity-reducing excipients may be the consequence of direct interactions of the excipients with the protein molecule, which could shift the distribution of cluster sizes (reducing R_H). Accordingly, NMR may be well-suited for exploring the fundamental mechanisms driving viscosity and viscosity reduction.

It is also theoretically possible to screen viscosity-reducing excipients via NMR in a manner similar to the screening of new, small-molecule ligands for protein targets. This remains to be shown as a generalized approach, but if the directly observable interaction of an excipient with a protein is a signature of a viscosity-reducing excipient, NMR screening could provide novel insights into the mechanisms of viscosity-reducing excipients for formulation composition design.¹⁷² For drug discovery, NMR-based fragment screening is often performed in pools wherein multiple-drug fragments are added to one sample and screened for signatures of binding to the protein.¹⁷³ These pools can then be split and analyzed to identify the fragment that binds to the protein. In theory, this workflow could be adapted to identify viscosity-reducing excipients in pools. NMR harbors an additional strength in that data can be collected at representative DP concentrations since NMR peak height increases with increasing protein concentration.¹⁷⁴ However, the analysis of NMR data is not straightforward, making subsequent studies and the adaptability of NMR experiments limited. Yet the mechanistic insights possible suggest that NMR could become a useful tool for the development of viscosity-reducing formulation compositions.

Computational tools aid in predicting PPIs and identifying formulation conditions

Theoretical and computational tools can aid in predicting PPIs and identifying phase-appropriate formulation conditions by offering additional insights to experimental approaches.^29,175 Computational tools have many capabilities that could be of use to researchers developing protein DPs. We have recently reviewed the computational approaches for high-concentration protein instability analyses and will discuss some of those tools here in the context of formulation composition development and viscosity reduction.²⁹ These tools roughly fall into two categories based on whether modeling approaches explicitly incorporate solution conditions. The tools that consider solution conditions usually involve computing protein properties or dynamics at different pH and ionic strength values. Broadly speaking, these tools can indicate the approximate pH and ionic strength conditions that would optimize protein properties such as conformational stability and physical stability. Given the molecular basis of PPIs that drive viscosity, calculating these properties as a function of pH and ionic strength may suggest viscosity-reducing solution conditions.

Protein descriptors may aid in viscosity-reducing formulation composition design

The calculation of protein descriptors in-silico may help identify formulation conditions that minimize viscosity by allowing efficient selection of screening conditions based on physical and chemical principles. Tools that enable computation of various protein geometric and chemical properties might facilitate correlation of molecular protein descriptors with solution properties and high viscosity. Descriptors of charge and hydrophobic patterning on the protein surface may be useful in identifying viscosity-reducing formulation compositions, as electrostatic and hydrophobic interactions are believed to drive high viscosity.^176,177 Additional protein descriptors investigated in the literature include Fv charge, zeta potential, isoelectric point, and patch number and area.^60,177 Furthermore, descriptors that capture these properties in structurally distinct regions and in different ways (e.g., total accessible hydrophobic surface area or the number of anionic or cationic patches) may have additional utility in further developing our understanding of the molecular origins of viscosity. In one example for developability screening, Apgar et al. designed mutants of a study molecule, calculated surface charge map (SCM) scores, and measured viscosity of the mutants.⁶⁰ The SCM score accounts for the distribution of charged residues across the Fv surface, and, unlike the net charge on one domain, it implicitly accounts for charge patches or unequal distribution of charged residues across the surface.^44,178 They found that SCM scores better correlate with solution viscosity than other charge-related descriptors, such as Fv charge, and it even performs better than other similar parameters like Sharma and Tomar scores.⁶⁰ Thus, SCM correlations with viscosity offer some molecular insights into the experimental data showing the role charge plays in driving viscosity. Interestingly, different charge-related descriptors (Sharma, Tomar, SCM)^44,149,176 exhibit differential correlations with viscosity, emphasizing that there is a non-trivial relationship between these charge descriptors and solution viscosity. Furthermore, the study from Apgar et al. emphasized the relevance and complexity that underlays sequence design for viscosity reduction, since disrupting negative charge patches or increasing net charge did not always result in the expected decrease in viscosity. Finally, this work showed that having an estimate of certain protein properties, which may drive self-association, is not sufficient to rationally design sequences that are less prone to self-associate and form clusters at high concentrations.

Hydrophobicity also contributes to solution viscosity, suggesting that information about protein hydrophobicity may also be helpful in selecting candidates with potentially low viscosity for drug development. Nichols et al. identified a hydrophobic patch on the Fab of an antibody, which they mutated with a charged amino acid to disrupt the hydrophobic patch.⁶⁶ They found that disrupting that hydrophobic patch results in lower viscosity relative to the control antibody, suggesting that the hydrophobic patch was involved in PPIs and contributed to viscosity.⁶⁶ Another hydrophobicity-related descriptor is the high-viscosity index (HVI), which has been successful in classifying molecules based on viscosity in combination with a charge-based decision tree.¹⁷⁹ The HVI simply indicates the ratio of the number hydrophobic and hydrophilic residues in the Fv to the length of the Fv. Given that the HVI can be used to classify molecules based on their sequence, it could be well suited for screening drug candidates. One limitation of the HVI is that it does not change in different pH or ionic strength conditions. While hydrophobicity is generally unaffected by pH and ionic strength, the solvent accessibility of a hydrophobic residue may change due to alterations in local side chain and protein dynamics at different pH. In another study, Park and Izadi tested a measure of hydrophobicity they call ‘CDR_HPATCH_WW,’ which is modestly correlated with viscosity and calculated including the antibodies’ conformational dynamics.¹⁸⁰ These data suggest that considering hydrophobicity is important for predicting viscosity and that including conformational dynamics may improve this prediction.

The patch view of protein surface chemistry driving high viscosity may not fully consider the contribution of protein conformational dynamics to protein surface properties and thus to viscosity. Work from Park and Izadi on the mutants from Apgar et al. suggests that the dynamic nature of the CDRs could modulate the ionization constants of residues in the CDR local environment, altering the charge patterning on the protein surface.¹⁸⁰ Specifically, they conducted MD simulations of the antibody Fv domain and calculated custom charge and hydrophobicity surface descriptors, which they used to build a model that predicts antibody developability. By calculating the CDR dynamics information, pKas of ionizable residues were recalculated based on the local environment, which could be a more physically accurate representation of the charge patterning across the domain surface. In that study, the correlation between a given charge descriptor and viscosity improved after recalculating and averaging the descriptor for all of the simulated structural conformations, suggesting that ‘static’ calculations of charge-based descriptors may be poor representations of the actual ensemble of charge states that the protein adopts.¹⁸⁰ Notably, an improved correlation was also obtained for a hydrophobicity-related descriptor after ensemble-averaging based on the antibody’s conformational dynamics. One possible explanation is that the CDRs can form intramolecular interactions and make hydrophobic regions solvent inaccessible and less likely to form intermolecular PPIs. Therefore, an ensemble view of the protein surface might provide a deeper understanding on how patches could evolve and drive self-association.¹⁸⁰ Although this approach is oriented at identifying candidate sequences for development and represents the findings for a specific solution condition, the electrostatic surface calculation could be performed at different pH levels. Since changes to pH alter the protein’s net charge and can modulate viscosity, recalculating the electrostatic surface potential at different pH levels may have utility in the drug development and formulation composition design process. Existing work has already suggested that disrupting charge and hydrophobic patches could drive viscosity reduction, and therefore the consideration of protein conformational dynamics may be a necessary step to further assess whether all or only some surface patches contribute to high viscosity.⁶⁶

The correlations of these various protein descriptors with viscosity are dependent on the given protein sequence, structure, and charge assignment to the different residues. That is, there is no generalized relationship between a given descriptor and viscosity, as such correlations strongly depend on the molecules tested in each study. Furthermore, the values of some of these descriptors can change depending on the pH ranges and ionic strengths at which they are computed, as well as the software/approach used to calculate them. It is important to note that the computed value for a given descriptor from two different software packages may not exactly agree and may not represent the ‘true’ value of that descriptor. There are practical differences in how different software packages compute properties such as charge. This includes such choices as charged atoms being modeled as point charges when in reality a quantum approach is necessary to accurately model charge.¹⁸¹ Similarly, even in a point-charge model, the charge assigned to residues with complex pKas, such as histidine, may vary depending on the algorithm used to compute the residue’s pKa in its local structural environment.^182,183 Accordingly, a particular calculation of a charge-related descriptor may not completely capture the protein’s properties in solution. This kind of theoretical constraint exists for the calculations of many of the descriptors reported to be correlated with viscosity. Additionally, the calculation of in-silico descriptors usually occurs for an isolated single molecule and does not account for the effects of neighboring molecules that would be present in a high-concentration solution. While some studies have shown statistical correlations between charge-based descriptors and viscosity, the theoretical or fundamental connection between such correlations and formulation composition remains unclear. The judicious choice of test molecules may bias correlations between descriptors and viscosity, in addition to variations in how descriptors are calculated between different software packages.¹⁸⁴

Molecular simulations offer detailed insights to aid formulation composition design

Molecular simulation approaches can probe PPIs at different pH and ionic strength conditions, making them useful for formulation composition designs to minimize PPIs and viscosity.^185–189 Of particular interest are the structural elements (atoms, residues and residue types, domains, as well as protein tertiary and quaternary structure) that drive PPIs. Molecular simulations are commonly set to run at user-specified pH and ionic strength levels, providing a formulation condition-related link. Molecular simulations have been shown to be able to sample the configuration proteins may adopt at a given pH and ionic strength combination. Then, analysis of these configurations with distance metrics, solvent accessibility, and so on could inform what parts of the protein may be involved in PPIs. Two main simulation methods are commonly deployed for biomolecular research: MD and Monte Carlo simulations (MC). MD relies on iteratively applying a small perturbation to the model and calculating the next position, after some time, and velocity of all particles in the model using physics-based equations for motion. These steps are repeated for a certain amount of simulation time to sample the configurations that molecules may adopt with applied physical and chemical constraints, such as bond length. Comparatively, MC methods involve iteratively evaluating random configurations and scoring the configurations based on the force field or the interatomic interactions. In MC simulations, if the proposed state’s score is statistically acceptable, it can be used as the next starting state and the probabilistic draw of a new configuration, and scoring is repeated until a certain number of rounds of sampling have occurred. Generally, the choice of MD or MC to carry out simulations depends on the complexity of the evaluated system. For complex systems such as fully atomistic models, MD methods might be more efficient for sampling different configurations in every iteration. By contrast, for less complex systems, such as the coarse-grained (CG) models discussed below, MC methods might more efficiently sample a broader ensemble of configurations than MD. A more in-depth discussion of different modeling methods and applications for high-concentration biotherapeutics is available elsewhere.²⁹

For antibodies in the biopharmaceutical context, it may be advantageous to simplify or CG the representation of the molecules from the atomistic level to a lower structural resolution (e.g., one bead per amino acid or as few as one bead per domain or molecule) (Figure 5A,B). With low-resolution, one bead per amino acid CG models (1bAA), each residue’s charge state at different pH levels can be represented with different protonation states on individual beads. These kinds of models enable simulation of weak PPIs, but not at high concentrations.²⁹ Alternatively, simplified CG models (Figure 5B,C) have been used to represent mAbs with fewer beads to reduce computational complexity.¹⁸⁹ Through decreasing the number of beads per molecule and sacrificing residue-level information, it becomes tractable to explicitly simulate high-concentration conditions with more than a thousand individual monomers in the simulation. These CG models thus allow for running simulations that can inform how different pH and ionic strength levels may affect PPIs and potentially viscosity. In a series of reports, Chowdhury and colleagues applied MD simulations of simplified CG models to better understand the effects of short- and long-range interactions on PPIs.^53,127,130 Their simulation approach is used in combination with experimental S(q) profiles, as described above, resulting in the calculation of the cluster size distribution (Figure 5C), which is subsequently used to predict solution viscosity (Figure 5D,E).¹²⁷ The viscosity calculations from combining simulations and experiments are accurate for their tested mAbs, illustrating the applicability of this type of modeling approach for developability screening. Furthermore, this approach can be used to understand the effects of excipients on cluster size, providing additional insight into the mechanism of excipient-mediated viscosity reduction.¹²⁷ For example, the data from Chowdhury et al. show that arginine reduces the number of both small and large clusters, potentially by disrupting SRAs (Figure 5D). While existing work supports the utility of CG models and simulation approaches, some questions, such as those related to protein-excipient interactions, protein-solvent interactions, or protein conformational sampling may require atomistic details. To bridge atomistic and CG models, multi-scale modeling approaches have been developed to improve computational efficiency while maintaining structural resolution.¹⁸⁷ The model developed by Hatch et al. performs well relative to the purely atomistic representations but can be improved with additional effort to include excipients to allow further insights relevant to formulation composition design.¹⁸⁷

Figure 5.

Computational models of antibody PPIs and their use in viscosity calculations. A) a model of the NISTmAb in which every atom is shown as a sphere. B) a representation of a 12-bead coarse-grained model of an antibody. C) a diagram illustrating how molecules are considered part of a cluster as used to calculate viscosity. D) the contribution of a cluster of a given size to calculated viscosity in the presence of the cosolutes arginine and sodium chloride. E) the relationship between the viscosity computed from molecular simulations and experimentally measured viscosity with different cosolutes as in D. A) rendered from the publicly available model of the NISTmAb.¹⁹⁰ B- E) reprinted with permission from “characterizing experimental monoclonal antibody interactions and clustering using a coarse-grained simulation library and a viscosity model” Amjad Chowdhury, Neha Manohar, Geetika Guruprasad, Amy T. Chen, Alfredo Lanzaro, Marco Blanco, Keith P. Johnston, and Thomas M. Truskett the journal of physical chemistry B 2023 127 (5), 1120–1137 DOI: https://doi.org/10.1021/acs.Jpcb.2c07616 copyright 2023 American chemical society.

One particularly useful application of simulation methods is to calculate parameters that can also be derived from experimental approaches. An example that has been extensively studied is the calculation of B₂₂ for protein and antibody models.^185,189,191 Since B₂₂ offers a thermodynamic measure of PPIs, computational approaches, such as Mayer Sampling MC simulations, offer a way to probe PPIs of a given molecule. Using an MC simulation approach, the strength of PPIs can be probed by simulating protein dynamics and calculating B₂₂ from the resulting ensemble of configurations. The pH and ionic strength conditions of implicit solvent CG models can be varied to identify conditions that yield values of B₂₂ that minimize PPIs.^185,186 Furthermore, these simulations might help suggest the structural regions involved in PPIs (Fab-Fab, Fab-Fc), which might provide guidance during preclinical candidate selection. Therefore, calculating B₂₂ at varied pH and ionic strengths could potentially suggest the range of solution condition where PPIs are minimized for a given molecule.^63,192 Yet, as discussed previously, since B₂₂ does not fully capture all the interactions a protein experiences at high concentrations, such as hydrodynamic and crowding effects, PPIs assessed at low concentrations may not be representative of PPIs at high protein concentrations. As a result, connecting computationally calculated B₂₂ to viscosity at high protein concentrations might be challenging.

All-atom MD simulations could be informative for understanding protein-excipient interactions for formulation composition design. MD relies on the specific forced field or set of parameters used to define the physical and chemical properties and the allowable dynamics (such as maximum bond lengths). The accuracy of the physics, and thus the results from MD simulations, are highly dependent on the force field used, total simulation time, and input configuration. It is thus important to consider these constraints when applying MD to study protein-excipient interactions. For example, atomistic force fields are primarily parameterized for protein and protein-like chemistries (e.g., carbon-oxygen bonds), but they may not accurately represent other chemical motifs present in excipients, such as inorganic salts and buffers. Protein-excipient interactions may be weak and existing force fields may not accurately capture these relevant but weak interactions. Including excipients as co-solutes in MD simulations and conducting simulations to improve drug development requires further force field parameterization. These improvements could allow the calculation of protein-excipient interaction potentials. For simulating viscosity-reducing excipients and other excipients, some tools such as the generalized Amber Force Field (GAFF) provide a way to compute the parameters that may be missing,¹⁹³ though GAFF builds off of the Amber force field and reduces flexibility in choosing other force fields for simulation. Comparatively, software from the open forcefield (OFF) initiative provides flexibility in running MD simulations with multiple solutes (protein, ligands) with up-to-date force fields.¹⁹⁴ Furthermore, the advent of ML approaches that can be applied to MD enable more physically accurate simulations.¹⁹⁵ In addition to questions regarding parameterization of force fields for MD simulations, the size and flexibility of biotherapeutic proteins, such as mAbs requires large computational resources. Simulation approaches, such as MD at atomistic resolution with explicit, implicit, or mixed solvents require large memory and computing resources, often making such simulations of antibodies intractable on the industrial scale.¹⁸⁷ On the other hand, most existing CG approaches rely on force fields that only implicitly account for protein-solvent interactions, but protein-excipient interactions may be more comparable to a mixed-solvent scenario. Thus, to discriminate between the effects of different buffer species or addition of excipients to the solution, the corresponding CG model would need to be trained against each set of solution conditions or potential solution components. To articulate the challenge with simulating antibodies in atomistic detail, Shahfar and colleagues estimated that simulation of a single antibody with sufficient solvent would require 10⁵ solvent atoms alone.¹⁸⁶ This size of system would take months to accumulate the nano to microseconds of cumulative simulation time likely required to accurately model the relevant protein-excipient interactions. Such timelines underscore the difficulty of applying existing MD approaches to high-concentration protein solutions and the need for novel and efficient approaches.

To conserve resources and time, it could be possible to simulate protein-excipient interactions in a low-concentration regime and quantify protein-excipient interactions via the preferential interaction parameters.^{88,179,196–207} In the low-concentration regime, it still may be necessary to deploy advanced methodologies to access a sufficiently broad configurational space to best approximate the protein-excipient interactions. Some alternative methods exist, including the SILCS method, which can be understood as simulating the interactions of organic probes with protein in a mixed solvent simulation and transferring those interaction potentials to excipients to understand the extent of protein-excipient interaction.^97,208 While SILCS describes the protein and excipient molecules separately, it does not explicitly assess or model protein-excipient interactions. Moreover, SILCS follows a methodology akin to ligand docking where only a single protein molecule is considered and therefore it cannot provide information in the context of a crowded, high-concentration solution. Yet there is some promise with other MD approaches as, over time, atomistic simulations increase in size and computational efficiency. In one example, Rickard and colleagues constructed a model of the E. coli cytoplasm with a variety of proteins and metabolites. Using a supercomputer, they collected an aggregate simulation dataset totaling five microseconds for 2x10⁵ atoms.²⁰⁹ The capacity of atomistic MD simulations to provide high structural resolution and increasingly accurate kinetic and thermodynamic simulations of molecular interactions suggests that MD may hold promise for future efforts to better understand protein–protein and protein-excipient interactions at high resolutions. Additionally, graphics processing unit (GPU)-powered simulations (including via cloud computing), along with ML-derived adaptive and enhanced sampling approaches, can further enable increased simulation efficiency.¹⁹⁵ Some new reports have attempted to model protein-excipient interactions with a variety of methods including simulations.^210–212

Analytical and statistical models of PPIs

Analytical or theoretical models, such as those from statistical thermodynamics and colloidal theory, have been used to predict formulation-relevant properties to help drive formulation composition design.^{121,136,145,191,213–215} Many fluid properties of antibodies can be modeled by colloidal theory, which often assumes simple spheres with limitations.¹²¹ These limitations include representing specific kinds of interactions, such as exclusively electrostatic interactions or Fab-Fab interactions, depending on the model.^{120,136,214,216} The short-range attraction and long-range repulsion (SALR) hard-sphere colloidal model developed by Virk and colleagues describes mAb viscosity and reflects the complex interplay between electrostatic and non-electrostatic interactions in the behavior of concentrated protein solutions.¹²¹ Virk and colleagues show that considering SALR can produce accurate predictions of B₂₂ and viscosity for a globular protein, and a mAb to a lesser extent, at different pH and ionic strengths. The SALR colloidal model provides a better description of concentration-dependent viscosity than a hard-sphere model with only electrostatic repulsions.¹²¹ However, the isotropic nature of this model fails to capture important protein molecular features for viscosity, such as surface charge heterogeneity or patches. In this regard, a ‘patchy’ hard sphere model, such as that from Kastelic et al. may be a more accurate representation for proteins, especially mAbs.⁷⁶ Kastelic and colleagues modeled antibodies with an interacting hard-sphere patchy model solved using Whertheim’s perturbation theory.⁷⁶ They showed that this model can predict the average cluster size distribution of a given molecule, which correlates with solution viscosity among other properties. In both the Virk and Kastelic models, the charge parameter on the spheres can be tuned based on the residue, domain, and/or overall protein charge, which is sensitive to pH and ionic strength. These models also only implicitly account for solution conditions through tuning different interaction parameters, and therefore they are limited to explicitly capture the effect of excipients on PPIs. For formulation composition design, it would be useful to conceive of a way to include the effect viscosity-reducing excipients may have on PPIs in the calculation of cluster size distribution and viscosity using these methods.

In addition to statistical mechanics-based models, there are also examples of ML models developed to ‘score’ candidate sequences by predicting whether a given sequence is likely to be above or below a viscosity threshold.^{64,151,177,179,217,218} Scoring of the molecule is in a fixed formulation. Such models lack some applicability to formulation composition design due to their lack of representation of non-protein solution components, including viscosity-reducing excipients. Yet these approaches have helped identify which in-silico derived protein descriptors or experimental data are helpful for predicting solution viscosities. Further studies should focus on expanding these models to include viscosity-reducing excipients and generate additional insights into different viscosity-reducing excipients’ mechanisms of viscosity reduction.

The development and training of ML models, such as data-rich deep-learning ML models, requires extensive computational resources and large, statistically diverse datasets.⁶⁴ These datasets are likely to combine experimental data, such as viscosity measurements, and in-silico data, such as molecular descriptors providing empirically guided predictions. The computational resource requirement for deep learning models, in particular, are much greater in comparison to descriptor calculations. Descriptor calculations can be done on a single computer in a matter of hours, while training even a simple neural network may take the equivalent of dozens of computers working together for several hours to days. This additionally does not account for constraints, such as whether a network or model needs retraining for new molecules or modalities, or the memory required to store the large amounts of data used for training and validation. Deep learning models require significant computational resources because models generally learn more efficiently from larger datasets. It may also be necessary to run model train/test computation cycles multiple times as part of a cross-validation approach to test for overfitting and generalizability (e.g., k-fold cross validation). The best way to build a generalizable, predictive model is to have a large amount of data (thousands to millions of data points) from which training and test datasets can be generated. The training dataset (~80% of data) is used to tune the model weights and parameters, while the test dataset (~20% of data) is used to assess if the model is well tuned for predicting the parameters of interest, such as viscosity. In the context of pharmaceutics, an ideal training set includes a diversity of protein sequences, CDR sequences and lengths, protein attributes (mixed charged and hydrophobic molecules), and solution conditions (diversity of excipients, pH, and ionic strength). However, the datasets used to train some published approaches lack sequence or protein attribute diversity and, as a result, the input data used for accurate predictions varies from report to report.

Although the consensus is that extended datasets are needed to ensure the accuracy and relevance of ML models, smaller datasets are still usable for developing simple neural networks. However, care must be taken to avoid overfitting. In one example (Figure 6A), Schmitt and colleagues applied a neural network to the task of concentration-dependent viscosity prediction.¹⁷⁷ In that work, they train a simple neural network with experimental data as well as sequence- and structure-derived in-silico data to predict either the slope (B) or intercept (A) of a linearized simple exponential function, ${\eta }_{\mathit{rel}}=\mathit{A}\ast e^{\mathit{Bc}},$ that describes the concentration (c) dependent viscosity (${\eta }_{\mathit{rel}})$.¹⁷⁷ The resulting model yields correlation coefficients greater than 0.9 for both the intercept and slope predictions. Although the predicted curve versus the experimental data shows great agreement across more than 25 antibodies, this approach assumes uniform solution conditions in all input data (pH 6.0). Expanding to more solution conditions could be labor intensive since the input data are from multiple types of experiments. Nonetheless, this work demonstrates that some of the necessary features for viscosity prediction may originate in both computational and experimental data types, emphasizing the need for both approaches to enable viscosity prediction and efficient formulation composition design. Notably, by training the model with high-concentration viscosity measurements, the data input into the ML model might implicitly capture important physical factors unique to high-concentration solutions, such as crowding or multibody effects. However, the model is not deployed in a way that allows for calculating the contribution of these factors from the model. By contrast, simpler models using only molecular descriptors often lack this information since descriptor calculations usually consider the protein as an isolated molecule not in a crowded context. Rai et al. constructed a convolutional neural net (Figure 6B) that uses the in-silico generated electrostatic surface potential map of the variable domain of a Fab to predict the viscosity of the full-length antibody at 150 mg/mL.⁶⁴ The convolutional neural network architecture is quite different from the simple neural network shown in Figure 6A, which enables the input to be a 3-dimensional protein surface with associated electrostatic surface potential. Comparatively, the inputs to the simple neural network are not representative of any three-dimensional structural information. The convolutional neural network approach performs well on the test antibodies, correctly classifying most tested molecules with receiver operator characteristic area under the curve (ROC-AUC) of 0.84, representing ~84% probability of correctly classifying a molecule as viscous (>20 cP at 150 mg/mL) or non-viscous. This work also showed the flexibility in developing viscosity predicting models since these two approaches comprise unique architectures and utilize different input data types to accurately predict viscosity. In both the Schmitt et al. and Rai et al. works, their models predict viscosity after being trained on datasets whose experimental viscosities were measured in a specific solution condition (Histidine pH 5.8–6). Given that the protein descriptors used to train both models only consider the individual protein and do not account for excipient-related effects, there is potential to further develop these tools for formulation composition design. Namely, the descriptor values could be expanded by recalculating them in different pH and ionic strength conditions to assess how solution conditions may modulate viscosity. Furthermore, analyses that suggest the driving factors behind a given molecule’s viscosity, such as done by Rai et al., can be leveraged to suggest whether a given excipient or category of excipients is likely to act as viscosity reducers. Rai et al. apply feature importance analysis to identify the molecular features that are correlated with high-viscosity predictions in their model. This kind of ‘interpretable’ ML can further suggest experimental approaches to validate the models’ predictions. For example, the high viscosity molecules could be re-engineered to remove the viscosity-promoting features, and their viscosity predicted as well as experimentally tested for validation. Ultimately, the addition of solution condition and excipient considerations, such as calculating the input computational data under varied solution conditions (pH, ionic strength), could improve the applicability of ML approaches to DP formulation development and viscosity reduction.

Figure 6.

Two example deep learning approaches for predicting solution viscosity from in-silico and experimental data. A) the architecture of a ‘simple’ neural network used to predict the concentration-dependent viscosity of input antibodies.¹⁷⁷ B) the architecture of a convolutional neural network that uses an input electrostatic surface potential (ESP) map from a mAb’s Fab to predict viscosity at 150 mg/mL.⁶⁴.

Practical considerations for formulation development of high-concentration biologics

Currently, there are theoretical limitations to our ability to rationally design viscosity-reducing formulation compositions. Primarily, there is limited mechanistic understanding of excipient-mediated viscosity reduction. Many hypotheses regarding the mechanism of viscosity reduction of common excipients, such as arginine, can be found in the literature, but none have been thoroughly tested, proven, or widely accepted. Thus, researchers continue to rely on empirical formulation screening.¹⁵² Both the pharmaceutical and academic communities also lack a robust mechanistic understanding of how specific protein properties, such as charge patterning, affect PPIs, and how those changes in PPIs, in turn, affect viscosity. Accordingly, even if a given excipient mechanism was well understood, the target protein ‘state’ for viscosity reduction may not be known. One source of this shortcoming is that common analytical techniques lack the structural resolution required for mechanistic elucidation of low-affinity interactions and the structural determinants of those interactions. Most available analytical techniques cannot capture the unique properties of a given molecule such as the surface charge patterning (in the protein’s native conformational ensemble) that could possibly underlie viscosity. Our current understanding has not fully connected sequence-structure properties to viscosity. As such, perhaps higher resolution structural techniques can aid in building the fundamental understanding required to connect protein properties to PPIs and viscosity. If it was possible to accurately measure properties such as side chain pKa, we could gain a deeper understanding of how properties, such as charge patterning, relate to viscosity and PPIs. Additionally, since the interactions between viscosity-reducing excipients and proteins may be transient and low-affinity (dissociation constants exceeding 1 mM¹⁷¹), measurement of the protein-excipient interactions may be quite difficult via common biophysical techniques excepting some structural approaches such as NMR.^171–173 In this regard, computational models offer great promise, but remain underutilized for formulation development.

While existing experimental techniques offer useful insights into the properties of the solution, sample requirements often limit how many formulations can be tested. Material availability, particularly at high concentrations, is often a constraint during developability assessments and early stages of DP development. For example, while the measurement of sample viscosity may only require volumes as low as 30 µl, those samples may be needed at concentrations exceeding 100 mg/mL per formulation for viscosity screening. Such amounts of protein may not be available for viscosity and formulation testing, especially considering the necessity of testing other attributes such as stability and potency. For example, grams of protein may be needed to generate samples for assessing stability, while sparing material for mechanistically probing viscosity may not be a high priority. One approach is to target smaller total volumes via high-throughput solution characterization approaches. However, high-throughput approaches face the difficulty of accurately dispensing small volumes of viscous material and ensuring good mixing with probes or excipients. Careful mixing is required to prevent bubble formation while ensuring homogeneity, which is more difficult with small volumes at high viscosities. The solution properties of high-concentration protein solutions thus limit the utility of common high-throughput techniques, including robotics. Practical limitations may become more evident if working with a concentrated process intermediate, such as post-UF/DF unformulated protein solutions that require further formulation or compounding steps to introduce additional excipients. For example, the concentrated process intermediate may be at significantly higher concentrations (and thus higher viscosity) to enable subsequent excipient addition and dilution to reach the target protein concentration. The concentrated process intermediate may also exhibit physical instabilities and problems with transfer steps given its potentially high viscosity, as well as the absence of stabilizing excipients such as surfactants. Additionally, concentration limits may prevent ‘spiking in’ excipients due to the consequent dilution of protein. Accordingly, protein concentration steps and/or buffer exchange are often required to assess various formulation compositions at high protein concentrations. The Gibbs-Donnan effect, however, often complicates buffer exchange and concentration operations since the proteins can cause unequal ion retention across the membrane, resulting in pH shifts.^19,34,85,101 New tools and instruments have been developed to increase the throughput of formulation screening by enabling medium throughput buffer exchange and protein concentration. These are instruments that can perform benchtop UF/DF on multiple samples at once, such as in multi-well plate formats using pressure to drive the fluid flow. For formulation screening studies, these instruments enable generation of more samples with somewhat less effort than existing approaches such as dialysis or spin column concentration and buffer exchange.²¹⁹ However, the starting viscosity of the solution may impact instrument performance, including process time and extent of buffer exchange (unpublished data). As with the comparable unit operation in manufacturing, lab-scale versions of operations, such as UF/DF, can suffer from the same problems with high-concentration DP, including filter fouling. Further development may be necessary to fully enable the use of these instruments for high-concentration formulation screening.

Like experimental approaches, computational and predictive models also face limitations in their application for formulation screening. First, predicting viscosity in a formulation-independent manner lacks applicability to identifying viscosity-reducing compositions for drug product development. Current formulation-independent approaches may be useful for developability assessments since they may help filter out protein sequences with PPI propensities or other unfavorable attributes (e.g., instability, oxidation). This may reduce the need to tune the formulation composition specifically for viscosity reduction, saving resources and time. For protein sequences where an ideal, low-viscosity sequence may not be possible, it may be important to have tools that can help assess the extent of viscosity modification possible from tuning the formulation composition. For instance, some models, such as DeepSCM,¹⁷⁸ can assess charge patterning based solely on protein sequence, but they lack information about how charge patterning relates to viscosity. More importantly, the calculation of these descriptors does not account for solution and formulation conditions, which play a very important role in the viscosity of a high-concentration drug product as we have discussed above. Nonetheless, formulation-independent models remain valuable during discovery stages for candidate selection or for identifying potential liabilities that formulation composition may address. Comparatively, the identification of new and complex protein descriptors in tandem with an understanding of these descriptors’ relationship to viscosity would be useful for the future applicability of computational methods to formulation composition design. New descriptors could leverage some findings from the literature that have been suggested to drive high viscosity. One of such examples is the calculation of the distance of a negatively charged carboxylate from neutralizing positive charges.⁶⁴ Although existing software allows for the calculation of viscosity-correlated descriptors, such as Fv charge, as a function of pH and ionic strength, they do not generally incorporate excipient-mediated viscosity effects in the descriptor calculations. Computing the effect of pH and ionic strength on viscosity-correlated descriptors could be a useful starting point for formulation composition design, enabling selection of solution conditions from which to begin testing viscosity-reducing excipients. Molecular modeling approaches, such as those that involve simulating PPIs or excipients, are partially limited due to challenges in representing small molecules (CG versus atomistic), the sampling (simulation time or configurations) required for low-affinity interactions, and the force fields for protein-excipient interactions.²²⁰ More efforts are needed to improve our understanding of protein-excipient interactions, which could not only provide a starting point to improve existing force fields but also enable improved modeling of physically relevant protein-excipient interactions. However, several open questions remain (1) Do protein-excipient interactions need to be more energetically favored than protein self-interactions for viscosity reduction? (2) Are the most effective viscosity reducers molecules that directly bind to or are weakly associated with the protein surface? and (3) How relevant are these excipients’ effects on solvent ordering and dynamics at the protein surface for viscosity reduction? Both modeling and experimental approaches will play a role in answering these questions. The recent work from Chowdhury et al. combining SAXS data with MC simulations of CG antibodies demonstrates a formulation-relevant approach that leverages the power of both experimental and computational approaches.^53,127,130 Other work is forthcoming and will highlight the utility of leveraging computational tools for assessing developability and PPIs.^221,222

Conclusions and future outlook

Worldwide, more than 200 antibody drugs and other protein-based therapeutics are approved, and many more are in clinical development, totaling nearly 1400 protein-based therapeutics.³ One of the drivers of the growth in protein therapeutics is their expansion into new therapeutic areas such as neuroscience, immunology, and cardio-metabolic disorders. For chronic conditions, patient convenience, flexible dosing options, and site-of-care shifts to at-home, self-administration are driving the growth in alternative delivery routes, such as SC delivery. Additionally, lifecycle management may drive the development of IV drugs for SC administration. Delivering the necessary doses of protein therapeutics via the SC route using standard PFS and AIs often requires low injection volumes and high protein concentrations. Such DPs may exhibit high viscosity, which poses manufacturing and injectability challenges, requiring the identification of viscosity-reducing formulation compositions. Large-volume injection devices, such as on-body injectors, can enable the use of low-concentration protein solutions, but are less mature and cost-effective compared to low-volume injection devices such as PFS.

Designing formulation compositions for high-concentration biotherapeutics currently relies on empirically screening for solution conditions and excipients that optimize the stability and viscosity of a given molecule of interest. Such an approach is time-consuming and poorly transferrable to other molecules and modalities. This empirical approach is used at least in part due to a lack of mechanistic understanding underlying the viscosity-reducing effect of common excipients, such as arginine. Furthermore, there is limited fundamental understanding of the relationship between the molecular properties of proteins and the interactions driving viscosity. While some studies have contributed to a better understanding of viscosity-reducing excipients, our understanding is insufficient to enable rational excipient selection based on prior knowledge of the protein molecular properties (e.g., surface heterogeneity, charge, hydrophobicity).

In comparison to experimental approaches, computational methods harbor the capacity to provide information across a range of length scales and contribute to building an integrated understanding of the origins of viscosity on a molecular level, as well as the mechanisms of excipient-mediated viscosity reduction. With some refining, computational approaches could integrate protein, excipients, and solution physics at multiple length scales from atomistic models to solution scale. The capability to measure protein-excipient, protein-protein, and protein-solvent interactions may be necessary to construct the resolution-spanning (sub-nanometer to micrometer) understanding needed for rational excipient selection. Currently, many computational methods exist for the resolution of protein-protein, and even protein-solvent interactions. While these methods do not typically include modeling of protein-excipient interactions, some approaches are evolving to provide resolution of these interactions. Properly including excipient molecules in computational methods may be critical since a primary gap is in the understanding of the viscosity-reducing excipient’s mechanism of action and the balance between protein-excipient and protein–protein interaction. For multi-scale modeling, this is a complex problem given the sizes of common viscosity-reducing excipients relative to the protein molecules and simulation structural resolution. For example, conducting atomistic simulations of an antibody complex with explicit solvent and multiple co-solutes may entail a simulation exceeding 10⁵ individual atoms and take months of computing time on a supercomputer to properly sample the low-affinity interactions that excipients may form with the protein molecules. Yet, such studies could be of great utility for understanding these interactions.

Combining information about protein structural properties, solution structure, formulation conditions, and viscosity may enable prediction of optimized formulation conditions. New ML models may be most useful if they predict viscosity in a formulation-dependent way, allowing for in-silico screening of formulation compositions. Some progress has been made in this regard, where ML models leverage electrostatic surface potential maps, which could change with the protein’s formulation condition and the dynamics of the CDRs.^64,180 Nonetheless, these initial reports do not recompute the electrostatic surface potential maps at different pH and ionic strengths. This may be because re-training the model with new in-silico data would require the corresponding experimental measurements at different pH and ionic strengths. With neural networks and similar ML models, it may be possible to connect in-silico protein properties, such as patch number or size, to viscosity through feature importance analysis. For example, ML models involve iteratively tuning the relative contribution (weight) of a given input property to the final predictive viscosity value. These final weights may be interpreted to understand how much a given property contributes to the specific model’s viscosity prediction. This kind of analysis may enable, for example, the identification of how many patches of a given size and character result in viscosity decreases. While this is only an approximation of the real-world contribution, Rai et al. conducted such an analysis and their findings echo previous hypotheses of the contribution of charged patches to viscosity.⁶⁴ Their model suggests that negatively charged and aromatic moieties on the protein surface are correlated with increases in solution viscosity. Experimental work, such as the mutational work from Apgar and colleagues, has also suggested that aromatic and negatively charged residues are important contributors to solution viscosity.⁶⁰ Although these deep learning models are relatively new, early studies suggest that their viscosity predictions are accurate.^64,177 However, mutational studies are needed particularly to evaluate whether mutations to residues that the Rai model predicts to drive viscosity would result in lower viscosity.⁶⁴

Toward the goal of transferability, existing computational methods, such as descriptor calculations or molecular simulations, require structural models but are not mAb specific. Therefore, some existing methods may be generalizable across various modalities. Existing mAb-centric viscosity prediction and estimation approaches rely on the construction of homology models for calculating 3-dimensional descriptors or running simulations that are used to ultimately estimate viscosity. To enable the use of these approaches for emerging modalities, such as fusion proteins, may still require homology models of the fusion protein. However, constructing these homology models may rely on as-of-yet unverified approaches and parameters, which limits the utility of the final structural model in existing viscosity prediction approaches. The current inefficiency of building homology models is partially addressed with the advent of deep learning models with accurate structure prediction capabilities, such as AlphaFold.²²³ These models are typically trained on existing public datasets, which have many examples of naturally derived molecules and mutants thereof. As a result, AlphaFold, and other structure-predicting ML models, may not perform well on therapeutic proteins. Biotherapeutics are highly engineered in both sequence and structure, and thus it may be difficult for modelers and engineers alike to adapt existing structure-prediction models for these engineered, protein modalities.

While structure-based predictions provide details known to be related to the high viscosity of mAb solutions, protein sequence ultimately dictates structure. While the viscosity-related features of the protein are embedded in the protein sequence, they may not be easily accessible. This suggests that with sufficient development, it may eventually be possible to develop deep learning models to predict viscosity and viscosity-reducing compositions, which are trained and parameterized on both sequence and structure data. This requires continued investigation of what quantitative changes to structural and surface features result in lower viscosity, as well as to develop methods to map these viscosity-driving features back into sequence to establish sequence–surface relationships. Then, these models could be constructed as to make predictions using sequence, akin to deepSCM and AlphaFold.^178,223 There are potential benefits of eventually developing sequence-based prediction of viscosity-reducing formulation compositions; namely, that they could be efficient and modality-agnostic. In comparison to structural model-based approaches, sequence-based methods could enable probing of possible formulation compositions for modalities that are more complex than mAbs. The construction of sequence-based computational approaches to viscosity-reducing formulation composition design would, however, rely on being able to train deep learning models with large, heterogenous datasets. Even for models that may be ready-to-apply, much concern exists for replacing and supplementing experimental approaches, especially to replace data used in regulatory filings or interactions.^224,225 This is, in part, due to the lack of regulatory consensus on what kinds of modeling approaches (mechanistic, quantitative structure activity relationship, empirical, ML/AI) are acceptable to use for supporting the characterization of the DP.^224,225 Moreover, there remain many data sharing limitations, including patents and intellectual property protections, which further restrict the dissemination of data, such as protein sequences, formulation compositions, stability, and viscosity that could be used for training deep learning models. The result is low sequence diversity in publicly available data, which may limit the generalizability of published findings to new modalities and prevent the development of tools that could help accelerate development timelines. More industry/academic partnerships can help to drive progress toward sequence-based predictions.²²⁴ Then, simple predictions from sequence would enable rapid screening of candidate sequences for low viscosity and viscosity-reducing formulation compositions. To develop these complex models, it is clear that understanding of the relationship between a protein’s sequence, structure, viscosity, and formulation is of paramount importance.

While instrument and sample constraints limit the quantity of available data, it is important to also consider the low modality, sequence diversity in published data. Most available literature has focused on the IgG1 mAb subclass. Although IgG2 and IgG4 mAbs are also developed as therapeutics, few studies that explore the differences in DP development between IgG subclasses, especially at high protein concentrations, are available.^1,58 Although some reports suggest IgG4 molecules may be more prone to higher viscosity and opalescence as compared to IgG1 molecules (at 125–150 mg/mL),^74,153 existing data in the literature suggests that even small sequence changes such as point mutations to a given antibody can result in significant changes in protein viscosity.^60,66 Thus, it is possible that differences observed between subclasses are largely a consequence of changing the protein sequence. Since many antibody framework sequences are highly similar, with sequence diversity arising largely in the CDRs, more sequence-diverse data can help researchers connect changes in protein properties to changes in viscosity and better establish the link between sequence and viscosity.

The next generation of computational and modeling approaches should include data that model the effects of formulation composition on solution and protein properties, including viscosity, rather than predicting viscosity at single, static conditions. Excipient properties can be represented via small molecule descriptors from available software packages.^226,227 Viscosity measurements from excipient screening can be included in training data by linking the protein descriptors (at the experimental pH and ionic strength) with descriptors for the excipients. With sufficiently diverse data, potential ML models could then be trained and designed to produce rank ordered lists of predicted viscosity-reducing conditions. The use of predictive methods could also accelerate candidate selection by identifying molecules that may be prone to PPIs and high viscosity before advancing to development stages. Furthermore, in development, the identification of potential viscosity-reducing excipients can help constrain the total number of test formulations, further improving efficiency. While use of computational viscosity prediction methods during developability assessments is not yet widespread, some caveats must be considered when using the existing tools. In particular, candidate molecules may be incorrectly classified as high viscosity when formulation composition is not considered, and this may result in the removal of developable molecules from consideration. Instead, the consideration of common formulation compositions in viscosity predictions could potentially prevent the filtering out of drug candidates in the discovery stage based on their formulation-agnostic predicted viscosity, since formulation conditions can still be optimized to decrease viscosity.

Physics-based methods will enable transferability to new modalities and afford industrywide innovation opportunities for viscosity modeling and formulation prediction. Training datasets need to be expanded to include examples of novel modalities; however, data on the development and formulation composition design of emerging modalities, such as fusion proteins and bispecifics, is less abundant owing to their relatively recent introduction to the industry. Given the industry’s experience in the development of mAbs, most available literature heavily features mAbs in training data for complex deep-learning and other kinds of statistical models. As the representation of emerging modalities increases, more data will be available for use in training models for connecting the differences in protein structure and chemistry to solution properties, including viscosity. Expanding available datasets is also important to increase the representation of ‘drug-like’ molecules given the extensive modifications made to therapeutic protein sequence and chemistry to optimize potency, stability, and other quality attributes. Meeting the challenges of developing complex biotherapeutics requires the generation and dissemination of expanded structural and experimental datasets for use in ML approaches. This effort must be collaborative, as academic research labs often focus on fewer and less representative molecules than those that pharmaceutical companies have in their pipelines. Academic and industry researchers can help push the development of deep-learning models forward through collaborations, as well as the publication of as much data as possible. While it may not be feasible to publish the sequence, concentration, exact solution composition, and experimental data for all molecules under pharmaceutical development, increased access to available data where possible could improve the development of high-concentration formulations. Companies should also invest in data mining from their own historical drug development efforts to drive understandings of formulation compositions and their effects on viscosity and stability as a function of protein sequence and structure.

Mechanistic understandings of the protein–protein and protein-excipient interactions that drive viscosity and viscosity reduction require continued investigation and will aid in building tools to rationally design viscosity-reducing formulation compositions for new protein molecules. Without physically grounded mechanisms for how a given excipient reduces viscosity, identification of the optimal formulation composition may remain subject to inefficient empirical optimization. The pharmaceutical landscape is changing, but efficient formulation development and robust manufacturing process and product development will continue to be high priorities for delivering necessary therapies to patients worldwide. The synergy between computational and experimental methods offers the opportunity for deeper understanding of high-viscosity solutions and mechanistic understandings underlying the effects of viscosity-reducing excipients. Leveraging established approaches along with novel predictive methods will improve formulation composition design and facilitate the development of needed patient-centric, high-concentration biotherapeutics.

Funding Statement

The authors acknowledge support from the Merck Postdoctoral Association, Rahway, NJ, USA.

Disclosure statement

All authors are employees of Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA, which provided funding to support M.A.C. as a postdoctoral fellow via the Merck Postdoctoral Association, Rahway, NJ, USA.