High concentration formulation developability approaches and considerations

ABSTRACT

The growing need for biologics to be administered subcutaneously and ocularly, coupled with certain indications requiring high doses, has resulted in an increase in drug substance (DS) and drug product (DP) protein concentrations. With this increase, more emphasis must be placed on identifying critical physico-chemical liabilities during drug development, including protein aggregation, precipitation, opalescence, particle formation, and high viscosity. Depending on the molecule, liabilities, and administration route, different formulation strategies can be used to overcome these challenges. However, due to the high material requirements, identifying optimal conditions can be slow, costly, and often prevent therapeutics from moving rapidly into the clinic/market. In order to accelerate and derisk development, new experimental and in-silico methods have emerged that can predict high concentration liabilities. Here, we review the challenges in developing high concentration formulations, the advances that have been made in establishing low mass and high-throughput predictive analytics, and advances in in-silico tools and algorithms aimed at identifying risks and understanding high concentration protein behavior.

KEYWORDS:

• Aggregation
• developability
• high concentration formulations
• in-silico
• physical instability
• subcutaneous administration
• viscosity

Introduction

Monoclonal antibodies (mAb), and other protein-based therapeutics, have become a huge success, as exemplified by the more than 100 different antibodies approved in the United States.^1,2 The growing need for biologics to be administered subcutaneously and ocularly, which limits the total volume of drug that can be delivered, coupled with certain indications requiring high doses, has resulted in a growing need for high concentration (>50 mg/mL) formulations. While chemical liabilities such as deamidation and isomerization are independent of the protein concentration, physical instabilities such as aggregation are protein concentration dependent. Due to this, high concentration mAb formulations can pose substantial manufacturing, stability, and delivery challenges.³ As protein concentrations have increased, so has the occurrence of physical instability (e.g., opalescence, aggregation, particles) and viscosity challenges.⁴ In addition, as molecule formats have become more complex, high concentration instabilities have proven to be more challenging.

Two general approaches are used to mitigate physical instabilities and viscosity issues. The first is to optimize the protein formulation during development by finding conditions that minimize potential liabilities (for example by changing the formulation pH and/or testing various excipients), thereby ensuring acceptable drug product manufacturing, administration conditions, and shelf life. However, this approach does not allow the use of platform formulations, as each molecule would require its own unique formulation, nor does it help with challenges during manufacturing where the protein may not be in its final formulation. The process to optimize the formulation is often slow and may result in a formulation that does not meet the intended shelf life, especially in the case where there is a need to balance multiple competing degradation pathways. Furthermore, formulation development can only improve high concentration behavior of biomolecules to a certain extent. While some liabilities can be addressed efficiently (e.g., agitation-induced aggregation, particle formation), others such as aggregation due to thermal unfolding and viscosity require the selection of suitable molecules for the intended purpose before technical development is started. As such, the second approach is to use a developability assessment program during the candidate selection phase to identify potential liabilities.^4–6 Early identification of such liabilities provides the opportunity to select a different variant or to attempt to fix the liability using protein engineering approaches. However, screening high concentration liabilities requires large amounts of protein and therefore limits the number of candidates that can be screened within a reasonable time period and with limited resources.⁷ In addition, while there has been success at identifying and removing chemical liabilities (such as deamidation and isomerization), mitigating physical instabilities is more complex. In these cases, there is generally no single amino acid that needs to be replaced, as physical instabilities are the result of protein-protein surface interactions. As such, the mechanisms behind aggregation and viscosity and approaches to predict, and therefore mitigate, such behavior have been the subject of much interest. Here, we review some of the recent work aimed at understanding, measuring, predicting, and mitigating high concentration instabilities and viscosity challenges. For a comprehensive summary of the other aspects of developability, we refer you to a recent review.⁶

Aggregation

Antibody aggregation is a substantial concern, as aggregates are generally viewed as immunogenic, may be hyper-potent, and conversely can reduce overall efficacy.⁸ Aggregation observed on stability may limit product shelf-life, which complicates supply chains. Aggregation occurs when a protein forms high molecular weight species either through weak nonspecific interactions (self-association) or via covalent bonds (covalent-aggregation).^9,10 In either case, it can lead to the formation of soluble (e.g., dimers, trimers) or insoluble aggregates that can manifest as particulates. Protein aggregation can be influenced by a number of factors, such as molecular properties; formulation composition, including active pharmaceutical ingredient (API) concentration; agitation; temperature; light; and contact with certain materials.¹¹ While controlling processing conditions, storage, shipment, and handling may help reduce aggregation, it may not fully mitigate aggregation. In those cases, other factors such as molecular properties and formulation composition are important tools in reducing stability risks.

Optimization of formulations to prevent protein aggregation is typically performed by leveraging prior experience combined with trial and error. In a typical study, the protein is formulated in a variety of buffers, surfactants, excipients, and pH levels and stressed via elevated temperature, agitation, and freeze/thaw conditions to identify the optimal formulation with minimum aggregation propensity. However, the propensity of a molecule to aggregate at elevated temperatures may not be predictive of the propensity to aggregate at the intended storage temperature due to differences in degradation pathways. Recently, Bunc et al. were able to predict the aggregate fractions after three years of storage at 5°C using accelerated temperature degradation.¹² The authors used a branched kinetic model that breaks the stability behavior into a low temperature and a high temperature pathway. Using this model, the authors were able to predict the aggregation of six antibodies and resolved the curvature that is typically observed when applying an Arrhenius model. The data also suggested that at low temperature, aggregation is triggered by chemical modifications, while unfolding is the main driver at high temperatures.

Aggregation during frozen storage

The challenges of protein aggregation are not limited to handling at drug product (DP) storage temperature of 2–8°C and/or handling during typical production (e.g., 18–25°C). Frozen storage (</= −20°C) is a standard practice for protein drug substance (DS), and the storage temperature and storage material must be carefully selected. Various stress factors that contribute to protein instability during DS processes have recently been reviewed by Das et al.¹³ and will not be reviewed herein. However, while the general assumption is that frozen storage increases the stability profile, high concentration mAb formulations have been found to be susceptible to instabilities such as increased aggregation during frozen storage. Lazar et al.¹⁴ have shown that mAbs have a potential to undergo cold denaturation at storage temperatures near −20°C at pH 6.3. Singh et al.¹⁵ reported increased aggregation of an IgG2-trehalose formulation when stored at −20°C. The glass transition temperature (Tg’) of the matrix was determined to be −29°C. Interestingly, aggregation in this formulation increased with time at −20°C but not at −40°C or −10°C. This is most likely because of the crystallization of freeze-concentrated trehalose from the frozen solute at temperatures higher than the Tg’ of the matrix. Lack of aggregation at −10°C was attributed to lower cryoconcentration of the protein at these temperatures, as well as higher mobility of the formulation that allows for any unfolded proteins to refold back into native state. Connolly and others¹⁶ expanded on the frozen instability studies and found that faster cooling rates (>100°C/min) result in trehalose crystallization and protein aggregation. They also reported that at lower cooling rates of≤1°C/min, trehalose remains predominantly amorphous and there is no effect on protein stability. Interestingly, the authors found that phase distribution of amorphous and crystalline trehalose dihydrate in frozen solutions depends on the ratio of trehalose to mAb. Based on these observations, the authors found that an optimum ratio of 0.2 to 2.4 trehalose-mAb (w/w) ratio is ideal for frozen storage.

Bluemel et al.¹⁷ report an interesting cryoconcentration effect on mAb instability. Samples with lower mAb concentration showed increased formation of high molecular weight species (i.e., aggregates). In contrast, higher concentrated samples led to more subvisible particles (SVPs) when the samples were stored above Tg’, that is, at −20°C or at −10°C. The authors went on to show that the protein and small excipients freeze-concentrate differently, resulting in different local protein to stabilizer ratios within a container, ultimately leading to mAb instability under these storage conditions.¹⁸ Interestingly, Bluemel et al.¹⁹ have recently reported on the use of computational fluid dynamic (CFD) simulations of mAb solutions. They conclude that CFD simulations are a promising tool to describe large-scale freezing and thawing of mAb solutions and could potentially save time and resources. In addition to these instability pathways, a relatively new hypothesis on the interactions of proteins with air bubbles leading to aggregation has been proposed by Salnikova et al.²⁰ Authelin et al. recently published a detailed perspective covering these topics.²¹ In summary, higher protein concentrations in the DS can result in frozen instabilities and have to be carefully evaluated.

Viscosity

Self-association of mAbs increases with concentration and can result in high viscosity, which challenges the capability of the ultrafiltration diafiltration (UFDF) unit operation during manufacturing, resulting in slow operation and/or an unacceptable pressure in the system. If the viscosity is only slightly elevated, it is possible to mitigate it using single pass tangential flow filtration (TFF) or increasing the recirculation temperature to decrease the viscosity. However, increasing the recirculation temperature can also result in an increase in aggregation and chemical degradation. High viscosity can also affect the filling operation, including the accuracy and fill rate. If the fill rate is slowed enough, or stopped due to an interruption, drying at the filling needles could occur, which can lead to aggregates/particulates in the final drug product or yield loss due to the need to flush the filling lines. During administration, high viscosity can prevent the use of a syringe and needle, in particular for subcutaneous administration with a pre-filled syringe. Above a certain viscosity, the injection force required to expel the solution might be too high for certain populations (e.g., geriatric populations) and/or dose volumes. It is important that instructions for use and other training material clearly highlight any product needs for warming the DP to room temperature prior to injection, if needed to reduce viscosity to a level that is convenient for administration. Although the use of autoinjectors or other devices can overcome some limitations imposed by high viscosity, they can add complexity, lengthen development time if an autoinjector must be available during pivotal clinical trials, and increase the cost of goods. Therefore, substantial effort has gone into the development of high concentration protein formulations with low viscosity. One approach is to use excipients that can help lower the viscosity. The other approach is to use experimental and computational approaches to measure and predict viscosity in order to help select molecules with low viscosity.

Viscosity reducing formulations

Viscosity is the result of noncovalent intermolecular interactions that result in the formation of a molecule network. As a result, any molecule or excipient that has the ability to disrupt the network can decrease the viscosity of the sample. Charge-based interactions tend to have the biggest impact, as they affect molecules over longer ranges compared to hydrophobic interactions. Sodium chloride can lower the viscosity of a formulation by shielding protein charge and therefore decreasing electrostatic driven protein-protein interactions (PPI). In contrast, it has been hypothesized that arginine works by direct binding to the protein surface.²² This can result in both charge shielding, as well as a reduction in hydrophobic forces by interacting with aromatic residues. While typically salts such as sodium chloride decrease the viscosity, there have been reports of it increasing the viscosity.^23,24 As such, each salt has to be evaluated on a case-by-case basis, making evaluation a very slow and tedious process. Caffeine has also shown promise at reducing viscosity, and in some cases performed better than both sodium chloride and arginine.²⁴ Results obtained by Zeng et al.²⁴ suggested that formulating with caffeine reduces the attractive PPIs, leading to lower viscosity. Using rigid body docking simulations, the authors found that caffeine may preferentially bind to positively charged groups and aromatic groups on the protein surface.

Other excipients have also been evaluated alone, or in combination, to address viscosity and stability issues. Recently, Banik et al.²⁵ reported that the use of excipients such as carnitine HCl, (S)-(+)-camphorsulfonic acid, and ornithine HCl resulted in varying effects to viscosity and self-interaction parameters that were also pH dependent. These findings again highlight how the charge distribution on molecules can have a substantial impact on what excipients are most effective in reducing viscosity. Proline is an example of another excipient that has been shown to reduce viscosity and is used in the 200 mg/mL polyclonal human IgG drug product Hizentra.²⁶ Other amino acids and excipients that have hydrophobic and charged properties have also been shown to lower viscosity in certain cases, furthering the understanding that disrupting API network formation is key to lowering viscosity.^27,28

Other viscosity-lowering approaches use the formation of nanoclusters or microparticles that restrict the viscosity effects to interparticle interactions, rather than intermolecular interactions. This approach creates two separate dominant forms of molecular interactions, those within the particles, which do not have a major influence on viscosity, and between the surfaces of the particles, which is less than that of soluble monomers. In one example, nanoclusters were formed by lyophilizing a mAb formulation containing trehalose as a crowding agent and reconstituting it in a buffer close to the mAb’s isoelectric point using a lower volume than the pre-lyophilization volume. This induced macromolecular crowding and led to the formation of nanoclusters that revert back to monomer upon injection.^29,30 Other particle-based technologies, using techniques such as electrospraying and microglassification offered by companies like Elektrofi and Lindy Biosciences, claim concentrations >400 mg/mL may be reached. However, these microparticle approaches may require substantial process development and require the use of non-aqueous solvents for the DP to be administered.

Experimental measurement and prediction of viscosity, solubility, and aggregation

High concentration behavior is influenced by a complex dynamic of different molecular interactions (prominently hydrophobic and electrostatic). Proximal energy effects caused by molecular crowding in high concentrated protein solutions further complicate the prediction of thhe behavior based on single parameters.³¹ One of the major liabilities for molecules intended for high concentration formulation is molecular self-interaction. Depending on the exact nature of those interactions and the underlying cause, a strong tendency to self-interact can lead to high viscosity, opalescence, aggregation, and/or particle formation. Colloidal interactions in a native, chemically modified, or (partially) unfolded state can have distinct consequences for a protein in solution. Furthermore, molecules are exposed to different unfavorable conditions during manufacturing and administration that can trigger degradation via distinct pathways. Given the complexity of underlying molecular mechanisms of self-interaction and the variety of external stressors, it is clear that there is no simple one-size-fits-all solution for the prediction of high concentration behavior in biopharmaceutical development. Even small-scale experimental evaluation of viscosity, solubility, and aggregation under specific conditions requires large mass amounts of protein due to the need to formulate at high concentration and the volume required by the analytics. For example, the accepted method for evaluating the viscosity behavior of a liquid dosage form at high concentration involves formulating the protein at multiple concentrations and generating a viscosity curve using a cone and plate rheometer. Due to the high mass requirements (~25–75 mg), screening multiple variants can be slow and/or cost prohibitive.⁴ As such, there is much interest in methods that can increase throughput with lower mass requirements (Table 1).

Table 1. Overview of small-scale assays to prediction solution behavior of proteins in high concentration formulations.

Method	Parameter	Predictive for	Material need	Throughput	References
AUC	ks	viscosity	high	low	^Citation33
DLS/SLS	Kd/A2 (self-interaction)	Viscosity, aggregation/ particle formation	medium	low/medium	^Citation25,Citation33,Citation34,Citation98
SINS	Wavelength shift (self-interaction)	Viscosity, opalescence	low	high	^Citation36,Citation39
SIC/CIC	Retention time (self-interaction)		medium/high	low/medium	^Citation42–45,Citation99
SI-BLI	Binding response rate (self-interaction)	Viscosity, opalescence	low	high	^Citation40,Citation41
Scissor	turbidity	Precipitation upon injection	high	low	^Citation100,Citation101
PEG/ammonium sulfate precipitation	Equilibrium protein concentration as function of precipitant concentration (solubility)		low/medium	medium/high	^Citation48,Citation49
Small scale stability (F/T, temperature, mechanical stress)		Aggregation, particle formation	medium/high (*)	low

(*) relevant concentration of proteins has to be used, long term high concentration exposure under physiological conditions becomes more relevant in e.g., Ophthalmology.

Molecular self-interaction in the formulation is determined by three major factors: 1) the inherent physicochemical characteristics of the protein, 2) the concentration of the protein in solution, and 3) the presence of other components in the solution (e.g., excipients). While the desired final concentration is mostly determined by the required clinical dose and the route of administration, both molecule selection and formulation development (or early evaluation of formulation suitability) can be guided using assays predicting high concentration behavior. In the following section, we discuss small scale (low mass) options of these assays and give an overview of available tools for the characterization of observed instabilities prominent in these solutions.

Viscosity and self-interaction

The first set of methods for assessing self-interaction propensity of proteins at low concentrations is based on assessing the second virial coefficient (termed A2 or B₂₂), which describes the weak pairwise interaction between two solute molecules in a solvent. A small-scale study of three mAbs demonstrated that A2 can be an effective low concentration (<20 mg/ml) predictor of protein aggregation and viscosity at high concentrations.³² A less tedious alternative to measuring A2, depending on the available instruments, is to measure its component, the diffusion interaction parameter, kD, which is amenable to high-throughput assessment using dynamic light scattering (DLS).^33,34 Using a panel of 29 mAbs (28 IgG1 and 1 IgG4) Connolly et al.³³ demonstrated that kD is a useful predictor of viscosity in buffer systems of both high and low ionic strengths. Kingsbury et al.³⁴ expanded on this study by measuring kD for 59 mAbs (44 IgG1, 4 IgG2, and 11 IgG4) and trying to correlate the outcome with their high concentration opalescence and viscosity behavior. The authors found that kD measured under dilute concentrations in a low ionic strength His-HCl buffer could very well predict high concentration opalescence and viscosity liabilities. They could also set a kD cutoff of 20 mL/g as a threshold below which most of the antibodies in their panel tended to show high concentration opalescence or viscosity.

A second family of methods that have proven useful for measuring the self-interaction between protein molecules using extremely low amounts of protein fall under the umbrella of self-interaction nanoparticle spectroscopy (SINS)-based techniques. These approaches take advantage of the extreme sensitivity of the localized surface plasmon resonance spectrum of gold nanoparticles (AuNP) to the inter-nanoparticle distance to report on the self-interaction of test proteins that are immobilized on the AuNP-surface.³⁵ The extremely low protein concentrations (1–100 μg/ml) and mass (μg scale) needed, the applicability with unpurified samples, and amenability to high-throughput analysis make these techniques suitable for application at very early project phases, including during binder screening. One of the first reported variants of a SINS method used for the characterization of antibodies was affinity-capture SINS (AC-SINS).^36,37 In a seminal work, Lie et al.³⁷ optimized the experimental steps of AC-SINS and used it to measure the self-interaction propensity of more than 400 mAbs. Using the known high concentration properties of a small subset of these mAbs, the authors claimed to predict the developability of the entire panel. However, no attempt was made to correlate these predictions with actual analysis at high protein concentration for the entire panel. The biggest drawback of AC-SINS is that it is only applicable under physiological buffer conditions (PBS, pH 7.4), which prohibits testing of proteins in formulation buffers at pH 5–6. Attempts to overcome this limitation have focused on engineering the surface properties of the AuNPs by functionalizing them with polymers that sterically or electrostatically stabilize the nanoparticles. A recently developed method in this area, termed PEG-stabilized SINS (PS-SINS), uses polyethylene glycol (PEG) functionalization to extend the applicability of the method to a wider pH range.³⁸ A complementary method, which uses poly-L-lysine (PLL) functionalization (termed charge stabilized SINS, CS-SINS) to achieve a similar extension in applicability, was used to analyze the behavior of a panel of 56 antibodies composed of molecules with IgG1, IgG2, and IgG4 frameworks.³⁹ The authors showed that the behavior of IgG1 and IgG2 mAbs in the CS-SINS assay correlated very well with their high concentration viscosity and opalescence properties. However, this correlation was much weaker for IgG4 mAbs, which shows that different molecular pathways contribute to high concentration liabilities even for closely related molecule types.

An alternative method to study protein self-interaction in high throughput is based on biolayer interferometry (BLI). Initially described by Sun et al.,⁴⁰ Domnowski et al.⁴¹ implemented the method on an Octet HTX system, thereby substantially increasing the throughput. This allows the ranking of molecules in a standard buffer system, as well as the early evaluation of different formulations and their impact on self-interaction. The broad buffer compatibility of the self-interaction (SI-) BLI technology and the commercial availability of sensor tips with different capture functionalities are advantages of this approach. However, only indirect correlations to high concentration properties have been established.

One early approach for the determination of self-interaction parameters, such as the second virial coefficients, is self-interaction chromatography (SIC). Using a chromatographic resin functionalized with a protein of interest, the retention time of the protein can be correlated to its self-interaction propensity.^42,43 While both SINS and SI-BLI are assays designed with high-throughput applications in mind, SIC can be used to characterize the self-interaction of a protein in more detail, detecting heterogeneous behavior and even sampling fractions of broad peaks to better understand the heterogeneity. Furthermore, an in-depth characterization of the influence of solution components is possible to guide formulation or process development for a low number of molecules.⁴⁴ However, the preparation of a specifically functionalized column resin for SIC is both skill and labor intensive and requires relatively large amounts of protein. Therefore, SIC is less suitable for ranking many molecule candidates during developability assessment.

In contrast, cross-interaction chromatography (CIC) circumvents the need for a specific column preparation for each individual mAb.^44,45 Attachment of a polyclonal IgG mixture or unrelated mAbs to the resin allows a higher throughput of mAb candidates to be evaluated. The relatively large amount of sample needed to specifically functionalize a column resin for SIC is also omitted, making CIC a more suitable option for early molecule assessment. Nevertheless, a good inverse correlation between retention time on the column and protein solubility was shown for a series of mAbs evaluated by Jacobs et al.⁴⁵ Potential association of test molecules with the general structural features of IgGs caused by Fab-Fc or Fc-Fc interactions can explain these observations. However, it must be considered that a specific case of self interaction, caused by mutual association of features within the complementarity-determining regions (CDRs), will be missed using CIC. This disadvantage can become more relevant when evaluating new formats, especially when the constant part of the molecules compared to the variable part comprises a smaller percent of the overall construct (e.g., Fab fragments). While SIC is a powerful tool to characterize one or few final candidates with regards to the exact influence of the solvent, CIC can be used during early selection of candidates in molecule discovery.

Solubility

Solubility, or propensity to precipitate, is an important factor in the formulation of stability and API behavior upon administration to patients. Determining a molecule’s solubility in a low volume assay can be a useful tool for increasing the number of molecules that can be screened in early project stages. This can be accomplished in a high-throughput manner by inducing precipitation at low concentration by the addition of ammonium sulfate (AS) or polyethylene glycol (PEG). While PEG mostly acts via the excluded volume effect (“molecular crowding”), AS potentially alters the interaction behavior of molecules.⁴⁶ Therefore, PEG is considered to be the preferred precipitant in such experiments. The general principle is to add different concentrations of the precipitant and measure the free soluble protein concentration after separation of soluble and insoluble fractions.⁴⁷ The resulting midpoint concentration of PEG is a good indicator for protein solubility and correlates well with other predictors such as kD.⁴⁸ An alternative study describes using microscopy to qualitatively assess precipitation and other effects such as phase separation or using turbidimetry as a direct readout for precipitation without the need for prior separation of insoluble fractions.⁴⁹ Advantages of this method include the low protein concentration required (in the range of 1 mg/mL for IgGs) and the potential for high-throughput application with an optimized small-scale assay and a suitable readout.⁴⁹ Scannell et al.⁴⁸ also stated that PEG precipitation might be superior to DLS-based kD determination when working with ionic excipients. Therefore, early formulation screening for challenging molecules could benefit from this approach. In contrast to most of the above mentioned self-interaction assays, and similar to light scattering-based kD determination, PEG/AS precipitation does not require the immobilization of the interacting molecules and thus avoids potential artifacts arising from this.

Aggregation and characterization of observed species

Low volume and/or low concentration stability studies are a standard part of the developability assessment for protein therapeutics. Aggregation liabilities can be identified using a combination of freeze-thaw stress, mechanical stress, and high-temperature stress conditions using DP storage and physiological conditions. Interestingly, in a study of 137 mAbs from different stages of clinical development, aggregation was observed at 40°C (stress stability) only weakly correlated to other parameters such as self-interaction.⁵⁰ Therefore, the specific evaluation of stability in small-scale experiments is still necessary to cover this important aspect of developability assessment. The design of such studies requires careful consideration since, at high concentration, molecular crowding (or molecular proximity) effects lead to a shift in prominence of different forces that influence molecular interactions. Short-range forces such as van der Waals and hydrophobic interaction become more important factors of protein-protein interactions at high protein concentrations. In contrast, protein-protein interactions at lower concentrations (and therefore longer molecular distances) are dominated by longer range charge or dipole effects.³¹ As such, choosing suitable conditions and a relevant concentration for accelerated stability assessment is crucial to avoid missing molecule liabilities that only become apparent at higher concentrations (>100 mg/mL). Furthermore, specific routes of administration can result in a high local concentration under physiological conditions, and efforts should be made to assess the potential aggregation risk under physiological conditions. Given the complexity and diversity of the physiological environment to which a protein therapeutic can be exposed, small-scale stability conditions tailored to reflect this could become a prerequisite for developability assessments in the future.⁵¹ The relevance of conformational stability (routinely determined by thermal unfolding and the calculation of the denaturation midpoint temperature, T_m) as compared to colloidal stability for protein aggregation, is controversial. Jain et al.⁵⁰ did not report any correlation between T_m and accelerated stability results, indicating that under common small-scale stress conditions (and in the range of T_m values represented in the mAb test-set), this parameter has no significant impact on aggregation.

Regardless of the study design, detailed characterization of the observed impurities/degradation products (prominently high molecular weight species and/or particles) is required. This endeavor can be challenging and requires the use of a specialized toolkit of analytical methods. Apart from the low-resolution biophysical methods used for investigation of protein oligomerization and complex formation, such as DLS or more time-consuming experimental techniques like analytical ultracentrifugation, the most established and representative method for assessment of species of different molecular weights in the biopharmaceutical industry is size exclusion chromatography (SEC). While SEC alone does not provide an accurate size of the observed species, coupling it to a multi-angle light scattering (MALS) detector enables determination of the molar mass and concentration, which are proportional to the light scattered by the particles.⁵² Despite the wide and validated applicability of the SEC method for release and stability monitoring, further investigation of aggregation phenomena may require orthogonal techniques to SEC to gain additional or more precise data that is typically not necessary for release and stability monitoring.

Such orthogonal techniques include native mass spectrometry (MS) coupled to liquid chromatography for separation. Modern ultra-high-mass-range (UHMR) mass spectrometers allow acquisition of high-quality mass spectra in order to determine molar masses of up to several MDa with superior mass resolution and accuracy.⁵³ Nevertheless, native MS is a gas-phase technique, which requires volatile buffers and has a low tolerance for salts and detergents.⁵²

Another orthogonal method, mass photometry (MP), can be used for the analysis of proteins and protein complexes under native buffer conditions using very low sample quantities. This is achieved by the use of interferometric scattering microscopy, which enables detection and quantification of light scattered by single particles.⁵² The free choice of buffers for investigation of observed impurities, as well as its high sensitivity are advantages of this technology. MP provides quantitative, label-free detection and mass determination of biomolecules. In addition, MP makes studying stoichiometry, energetics and dynamics of protein complexes possible.⁵⁴ While the low sample consumption of MP is an advantage, MP is only applicable at nanomolar or lower concentrations and is thus unable to detect unstable high molecular weight species that dissociate at such low concentrations. Furthermore, the resolution of MP is lower than that of MS and, while sufficient to detect aggregation/multimerization, can fail to resolve species originating from degradation of the API.

To complement these methods and cover the full size range of possible aggregates, techniques for investigation of visible and subvisible particles such as optical microscopy, light obscuration, membrane microscopy, flow imaging, conductivity-based particle counter, and fluorescence microscopy, are available.⁵⁵ Additionally, laser diffraction, as well as DLS, nanoparticle tracking analysis, or turbidimetry/nephelometry can be used to assess the size of particles. Den Engelsman et al.⁵⁵ have made a comprehensive comparison of these methodologies regarding their power of observations, their advantages, and disadvantages.

Current state of the experimental toolbox for high concentration behavior prediction

While high-throughput predictive assays with impressive correlations to unwanted high concentration solution behavior of proteins have been developed, it is currently impossible for a single assay to predict which liability will ultimately be manifested.^38,39,41,49 Furthermore, most assays have been developed for standard IgG formats. Due to the complexity of the underlying molecular mechanism of self-association, the limitation of these assays becomes relevant for alternative molecule formats (e.g., Fabs, DutaFabs, fusion proteins, Contorsbodies).^56,57 In fact, Starr et al.³⁹ observed differences in the correlation between CS-SINS score and diffusion interaction parameter [kD] when evaluating different IgG subclasses. It would not be surprising if similar differences are observed for more diverse formats. Similarly, Domnowski et al.,⁴¹ while showing a good correlation between response rate in SI-BLI to kD by DLS, observed exceptions that indicate that the exact method setup (e.g., molecules immobilized vs free in solution) could bias the obtained results. While for early molecule selection, these exceptions are acceptable, a late developability assessment of a limited number of candidates could benefit from a confirmation using a second orthogonal method. Additional studies are needed to determine which assays have broader general applicability or can be quickly adjusted to format needs and which approaches might be too specific to keep pace with a rapidly evolving drug discovery pipeline.

In-silico approaches for viscosity and aggregation prediction

In addition to the development of high-throughput analytics to measure and understand high concentration liabilities, there have been many advances in the use of computational approaches. Studies have suggested that electrostatic and hydrophobic interactions both play a role in viscosity, solubility, and aggregation of proteins.^58–61 In particular, it has been shown that negative charge on the variable fragment (Fv) of a IgG1 mAb correlates with high viscosity, presumably due to intermolecular interactions with corresponding positive charge patches on Fc regions.⁶² This has led to the development of various methods that quantify the charge and hydrophobicity of proteins and then attempt to correlate those results to viscosity and aggregation. The first generation of methods relied on protein sequence to calculate properties of interest, such as the isoelectric point or hydrophobic content, and predict high concentration behavior. One such approach correlated the viscosity of 14 antibodies with the net charge at the formulation pH, charge symmetry, and the hydrophobicity index.⁶² In that study the authors showed that hydrophobicity and charge distribution affect the solution viscosity, while increased charge decreases it. Similar sequence-based approaches, including methods such as AGGRESCAN, PASTA, TANGO, and Zyggregator, have been used to predict protein aggregation.^63–67 However, these methods are not antibody specific, tend to have high levels of false positives for mAbs, and ignore the contribution of neighboring residues. This led to the creation of structure-based prediction methods that take into account the three-dimensional (3D) protein structure to gain a better representation of surface patches that could result in protein-protein interactions. One such method is AGGRESCAN3D (A3D), which combines the aggregation propensity of amino acids obtained from AGGRESCAN with structural information from the 3D protein structure to identify aggregation prone regions.^68,69 By taking into account only residues that are surface exposed and within a spherical region around the residue’s central carbon, A3D is able to lower the false positive rate, as it ensures only residues that are able to interact are taken into account. Another method known as the spatial charge map (SCM) uses homology modeling to build the 3D structure of the Fv region and calculates the spatial summation of charge of the surface exposed residues.⁷⁰ While SCM has been shown to differentiate between high and low viscosity molecules using three different datasets, the cutoff between the two populations varied each time. This could result from differences in experimental conditions between datasets or the method picking up differences in molecule frameworks. In addition, SCM does not account for the effect of buffer on viscosity, and thus this method cannot be used to determine if the high viscosity can be fixed via formulation changes.

Similar approaches can be taken to calculate hydrophobicity of an antibody with one common method being the spatial aggregation propensity (SAP). SAP measures the hydrophobicity of a region by taking into account the surface exposure and hydrophobic nature of particular amino acids and those proximal to that residue.⁷ Compared with purely sequence-based approaches, SAP is able to determine hydrophobic patches created by multiple hydrophobic residues that are near each other in the folded state. To demonstrate the applicability of SAP, Chennamsetty et al. used two antibodies with high hydrophobic scores, identified the regions responsible for the hydrophobicity, and introduced single or multiple point mutations to lower the hydrophobic nature.⁷ After incubating the mutants at elevated temperatures, they demonstrated that the new variants had higher thermal stability as measured by the percent monomer still present after thermal stress. In addition, functional analysis demonstrated that they preserved the antigen-binding activity when mutations were not in the CDR. Such a technique can prove useful for selecting as well as optimizing therapeutic candidates.⁷¹ Another method known as Aggscore takes into account the distribution of hydrophobic and electrostatic patches on the protein surface.⁷² As opposed to SAP, Aggscore has a scoring function that takes into account the intensity and relative orientation of the respective surface patches. However, Aggscore was not trained on antibodies and while the authors correlated Aggscore to hydrophobic interaction chromatography (HIC), CIC, and standup mono-layer adsorption chromatography (SMAC), the method was not evaluated against high concentration properties such as viscosity, aggregation, or opalescence.

Building on SAP, Lauer et al. devised the Developability Index (DI), which incorporates the competing effects of electrostatic and hydrophobic interactions.⁷³ Using a dataset of 12 antibodies of different subtypes, the authors were able to generally predict the aggregation at elevated temperatures and across a range of pH. While the method provides a general trend of aggregation propensity, the validation dataset is limited, and it does not give insight into other high concentration liabilities such as viscosity or opalescence. Similarly, Raybould et al. developed the Therapeutic Antibody Profiler (TAP) to provide developability rules similar to the Lipinski rules for small-molecule development.⁷⁴ To develop TAP, Raybould et al.⁷⁴ modeled the variable domain of clinical stage antibodies⁵⁰ and used them to determine five metrics that were thought to be important in developability. The metrics include the length of the CDRs, the extent and magnitude of surface hydrophobicity, positive and negative CDR charge and charge asymmetry. The authors then demonstrated that TAP was able to flag both an affinity-matured molecule that exhibited high levels of aggregation due to a large surface hydrophobic patch, as well as another antibody that exhibited poor expression. Using such tools, residues of interest could be identified and used to introduce mutations that either add or remove charge or change a hydrophobic patch and thereby improve a molecule’s developability.

Beyond calculating the charge and hydrophobicity of a molecule, it has become increasingly common to calculate additional descriptors that can correlate with high concentration liabilities. For example, Li and coworkers modeled the Fv regions and calculated molecular descriptors to understand and predict the viscosity of antibodies.⁷⁵ The authors used a panel of 11 antibodies, all formulated under-standardized conditions and found that the charge, zeta-potential, and pI of the Fv region correlated with the viscosity of highly concentrated solutions. In addition, at 150 mg/mL, molecules with high viscosity had negatively charged patches on the Fv region, which could interact with the positively charged IgG1 Fc regions. This was similar to Vikas et al.’s results showing Fv charge distribution affected viscosity.⁶² However, using only the Fv to understand high concentration liabilities disregards the effect of the constant regions, which have been shown to engage in transient intermolecular networks.^76,77 It is also possible that the introduction of Fc mutations (for example for half-life extension) can destabilize the Fc and alter protein-protein interactions.⁷⁸ To account for this, there has been a shift to modeling the whole antibody. In one example, Tomar et al. predicted the concentration-dependent viscosity curve from the combination of experimental data and homology-based structural models.⁷⁹ Using molecular descriptors calculated on 16 mAbs, the authors found that the hydrophobic surface area of the full-length antibody, as well as the charges on the VH, VL, and hinge region could be used to predict the viscosity curves. Similar to other works, the robustness of the method still needs to be determined since the results are limited to a small dataset and a single formulation condition.

While homology modeling the whole antibody further improves model accuracy, molecules are inherently dynamic and that motion can result in different descriptor values when compared to those from a single static structure. To account for this, it is becoming common to use molecular dynamics (MD) to capture the side chain fluctuations and understand the impact of conformational dynamics on antibody properties. In addition, MD can be used to identify key residues that may be involved in protein-protein interactions and propose mutations to disrupt such interactions.⁸⁰ Furthermore, MD can also identify additional types of interactions such as cation-$\pi$ and/or $\pi$-$\pi$, which may contribute to high concentration behavior.⁸⁰ However, doing full atom molecular dynamics on a complete antibody is computationally intensive and hence limits the speed and throughput as a screening tool. With the introduction of even larger complex structures, the limitation becomes even more pronounced. One way to overcome this limitation is to coarse grain (CG) model the molecule, which allows the simplification of an antibody to a reduced system composed of 8–16 beads, with each bead containing a single charge representing the summation of all partial charges in that domain.^{76,77,81–83} Not only does this allow the modeling of the whole antibody, but also the study of interactions between multiple copies of a protein or even multiple proteins. In one study, a 12 and 26 bead model was constructed to model two antibodies that had different viscosity behavior at high protein concentrations.⁷⁷ Using such an approach, the authors studied the impact of domain-level charge-charge interactions on the two antibodies and visualized the type of interaction (Fab-Fab or Fab-Fc). Interestingly, the antibody that had high viscosity at high concentration formed dense clusters, whereas the antibody with low viscosity at high concentration did not. To further understand the mechanism behind network formation, the authors investigated the impact of charge swap mutants on the viscosity of the two antibodies and found differences in the number of Fab-Fab interactions across the variants.⁸² Similar results were found by Buck et al. when they studied four different antibodies using CG models.^{76,77,81–83} In this case, the authors used an asymmetrical model that is more representative of the structure in solution and found that molecules with higher viscosity also had a higher number of pairwise interactions.⁷⁶ While these studies highlight the importance of electrostatic interactions in network formation, hydrophobic interactions will likely play an important role when the intermolecular distance decreases. Izadi et al. expanded the previous work and developed a CG model that not only accounts for higher-order electrostatic multipole moments, but also for hydrophobicity.⁸³ After tuning the model with data from three antibodies, the authors predicted the viscosity behavior of 12 different antibodies and demonstrated that the model could be useful in predicting the effect of ionic strength on the viscosity of solutions. Such an approach could be useful to determine the likelihood that a high concentration formulation could be developed in the case where no low viscosity variants exist.

In addition to the use of MD and coarse grain modeling, there has been an increased interest in using machine learning (ML) to predict high concentration liabilities as well as formulation fixes for those liabilities.^84–91 After removing high correlated descriptors, Lai et al. used a dataset of 21 approved antibodies to develop an aggregation model.⁸⁷ In that model, the positive SCM values and the solvent-accessible surface area (SASA) of the CDR-H2/H3 region correlated with the aggregation rate. The SAP value of the CDR-H3 was also identified as being important, which agreed with previous studies that charge and hydrophobicity are key factors. Comparison of the two-factor linear model with the dynamic mode of A3D showed that the new model performed much better for this limited dataset (prediction coefficient of 0.71 vs 0.28). In a subsequent study, Lai et al. measured the viscosity of 27 commercially approved mAbs and used the experimental data, along with computationally derived descriptors, to develop a decision tree to classify antibodies into high and low viscosity.⁸⁵ The authors found that antibody net charge and Fv amino acid composition drive viscosity. In particular, they found that the number of hydrophilic and hydrophobic residues in the Fv region are important. Expanding the dataset to predict the aggregation propensity of 20 pre-clinical and clinical stage antibodies gave mixed results.⁸⁸ In the case of aggregation, the authors found that positive SCM values and the solvent-accessible surface area of hydrophobic residues on the variable fragment were the most important, which agreed with their previous results. However, for viscosity, the authors found that the model developed with commercial antibodies did not adequately predict the viscosity of the clinical-stage antibodies.⁸⁸ A potential explanation is that the original model was trained on too small of a dataset, resulting in feature selection that is not representative of a large population of antibodies. The smaller the dataset, the higher the risk of overfitting ML models, resulting in models that are not predictive when applied to a new dataset. Since datasets of high concentration molecules are limited, this becomes a recurring problem in predicting high concentration liabilities. In addition, given the expense of producing such data, it is likely that this will remain a bottleneck for some time. To try to overcome this, the authors combined both commercial and clinical-stage antibodies to produce a model trained on 47 molecules. In that model, the two most important parameters are the number of hydrophobic residues and net charges on the light chain variable region. More data will be required to determine if the new model accurately predicts high concentration behavior. Another model developed by Cloutier et al. used an elastic net (EN) and support vector machine (SVM) model to predict antibody-excipient interactions for various excipients including sucrose, arginine hydrochloride, and sodium chloride.⁹¹ In that study, the aggregation and viscosity behavior of antibodies was correlated with the relative strength and weakness of the protein-excipient interactions. The authors proposed that such an approach can be used in combination with SAP and SCM to determine when to mutate the antibody to correct such liabilities.

In summary, many advances in predicting molecule aggregation and viscosity have been made. Although the approach behind each method varies, charge and hydrophobicity are typically found to play a role in these high concentration liabilities. In addition, the limited datasets in combination with the small number of problematic molecules in those datasets, have resulted in models that fail to have wide applicability. However, computational models are expected to improve and eventually help identify high risk molecules before they enter development. In the cases where no viable alternative exists, computational approaches can provide insight into possible engineering solutions. Two recent reviews provide an in-depth discussion of some of the recent advances in computational assessments for developability.^92,93

Conclusion and future outlook

With the focus on higher protein concentrations and the introduction of more complex molecule formats, high concentration instabilities have been observed more frequently. If predictive assessments, whether experimentally or computationally, are to be applied, additional data to build and validate models will be required. However, due to the high protein masses and the labor-intensive process needed to generate high concentration data, most studies conducted to date have been carried out at low concentration. For example, the extensive survey conducted by Jain et al.⁵⁰ looked at biophysical properties found in clinical stage and commercial antibodies with the assumption that antibodies that reached that stage of clinical trials have characteristics amenable to therapeutic development. However, most of those molecules were not developed at high concentration and thus the high concentration data do not exist. It is possible that, as the protein concentration of those molecules increases, additional factors relevant to developability will be uncovered while documented thresholds for each property will need recalibrating.

Building a deeper understanding of high concentration liabilities will require systematic data of molecules at high concentration. While automation can help decrease the experimental burden, many limitations exist, including the high protein masses required by the standard analytics. Identification of low volume high-throughput analytics could decrease this burden, but additional work is required to establish correlations between low mass, high-throughput assay results, and high concentration aggregation. As such, the need for large amounts of protein remains. In addition, automation of the sample preparation becomes difficult due to the increase in viscosity that is often observed at higher protein concentrations, and sample-to-sample variation in viscosity adds to the complexity. The move to novel formats also adds complexity to the process of developing high concentration formulations. Not only is the number of potential format combinations increasing,⁹⁴ but in-silico modeling of such formats still remains relatively unexplored.

Building on the theme of more data, it is becoming increasingly common to use multi-attribute monitoring (MAM) methods for the residue level quantification of antibodies.^95–97 The use of such methods provides deeper characterization of product variants and could be useful in furthering our understanding of protein aggregation, in particular during long-term and accelerated storage. When coupled with automation and low volume analytics, the resulting data could be used to build various models that filter molecules with unfavorable properties from entering development. Eventually, the use of computational tools could increase the number of variants that are evaluated, further increasing the probability of technical success. Ultimately, a fundamental understanding of high concentration liabilities will be required, especially as molecule formats become more complex.

Acronyms

A3D	=	AGGRESCAN3D
AC-SINS	=	affinity-capture SINS
API	=	active pharmaceutical ingredient
AS	=	ammonium sulfate
AUC	=	analytical ultracentrifugation
AuNP	=	gold nanoparticles
BLI	=	biolayer interferometry
CDR	=	Complementarity-determining regions
CFD	=	computational fluid dynamics
CG	=	coarse grained
CIC	=	cross-interaction chromatography
CS-SINS	=	charge stabilized SINS
DI	=	developability index
DLS	=	dynamic light scattering
DS	=	drug substance
DP	=	drug product
EN	=	elastic network
Fv	=	Fragment variable
kD	=	diffusion interaction parameter
MALS	=	multi-angle light scattering
MAM	=	multi-attribute monitoring
MD	=	molecular dynamics
ML	=	machine learning
MP	=	mass photometry
MS	=	mass spectrometry
PEG	=	polyethylene glycol
PEG-SINS	=	PEG-stabilized SINS
PLL	=	poly-L-lysine
PPI	=	Protein-protein interactions
QSAR	=	quantitative structure-activity relationship
SAP	=	spatial aggregation propensity
SASA	=	solvent accessible surface area
SCM	=	spatial charge map
SEC	=	size exclusion chromatography
SI-BLI	=	self-interaction biolayer interferometry
SIC	=	self-interaction chromatography
SINS	=	self-interaction nanoparticle spectroscopy
SLS	=	static light scattering
SMAC	=	standup mono-layer adsorption chromatography
SVM	=	support vector machine
SvP	=	subvisible particle
TAP	=	therapeutic antibody profiler
TFF	=	tangential flow filtration
Tg’	=	glass transition temperature
UFDF	=	ultrafiltration diafiltration
UHMR	=	ultra-high-mass-range

Acknowledgments

The authors would like to thank Jasper Lin and Trevor Swartz for their careful review and comments.

Disclosure statement

No potential conflict of interest was reported by the authors.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

The author(s) reported that there is no funding associated with the work featured in this article.