Reduction of monoclonal antibody viscosity using interpretable machine learning

ABSTRACT

Early identification of antibody candidates with drug-like properties is essential for simplifying the development of safe and effective antibody therapeutics. For subcutaneous administration, it is important to identify candidates with low self-association to enable their formulation at high concentration while maintaining low viscosity, opalescence, and aggregation. Here, we report an interpretable machine learning model for predicting antibody (IgG1) variants with low viscosity using only the sequences of their variable (Fv) regions. Our model was trained on antibody viscosity data (>100 mg/mL mAb concentration) obtained at a common formulation pH (pH 5.2), and it identifies three key Fv features of antibodies linked to viscosity, namely their isoelectric points, hydrophobic patch sizes, and numbers of negatively charged patches. Of the three features, most predicted antibodies at risk for high viscosity, including antibodies with diverse antibody germlines in our study (79 mAbs) as well as clinical-stage IgG1s (94 mAbs), are those with low Fv isoelectric points (Fv pIs < 6.3). Our model identifies viscous antibodies with relatively high accuracy not only in our training and test sets, but also for previously reported data. Importantly, we show that the interpretable nature of the model enables the design of mutations that significantly reduce antibody viscosity, which we confirmed experimentally. We expect that this approach can be readily integrated into the drug development process to reduce the need for experimental viscosity screening and improve the identification of antibody candidates with drug-like properties.

KEYWORDS:

• Antibody engineering
• charge
• computation
• developability
• formulation
• Fv
• hydrophobicity
• in silico
• isoelectric point
• mutation

Introduction

Antibody therapeutics are widely used for treating many human diseases given their attractive combinations of activities, developability properties, and safety profiles. These natural biomolecules commonly have high affinity and specificity for their target antigens, which limits their potential off-target effects, and their breakdown products are amino acids, which limits their potential toxicities. They also typically have attractive pharmacokinetic properties and effector functions due to their Fc regions, the latter of which can be manipulated in numerous ways to achieve a wide range of behaviors.^1–3 Equally important, antibodies have some of the most attractive biophysical properties of any class of proteins, including high folding stability, solubility, and low aggregation propensity. These and other attractive attributes have resulted in >100 approved antibody drugs to date and hundreds more currently in clinical trials.⁴

Nevertheless, there is intense interest in simplifying the administration of therapeutic antibodies using subcutaneous delivery, which has led to the need for concentrated antibody formulations. It is commonplace for antibodies, which normally have excellent biophysical properties at dilute concentrations (e.g., <10 mg/mL), to display highly variable viscoelastic properties when concentrated to much higher levels (e.g., >100 mg/mL).^5–7 It remains difficult to reliably predict which antibodies or mutants thereof will display suboptimal viscoelastic properties, and solving this problem would be significant due to the slow and costly nature of preparing concentrated antibody formulations for viscosity measurements.^8–10 Moreover, by the point in the development process that antibody viscosity can be measured, it is typically too late to change the antibody sequence to address viscosity problems.^8,11–14

The development of models for predicting antibody viscosities or classification of antibody viscosity levels (e.g., high or low viscosity) has typically suffered from at least one or more of the following four problems. First, there is relatively little viscosity data available for concentrated antibody formulations that can be used for model development, and this prevents rigorous model development and testing. Second, some of the previously reported models are not accessible to most investigators due to the need to either license them or perform complex calculations that are impractical for general and routine use.^15–17 Third, some of the existing models are difficult to interpret, as judged by the difficulty in using them in a simple way to predict mutations that reduce viscosity. Fourth, most models have not been validated for predicting new mutations that reduce antibody viscosity rather than simply predicting antibodies that were held out during model training.

Here, we sought to address each of these limitations and develop a classification model for predicting the level of antibody viscosity in a relatively simple and widely accessible manner (Figure 1). First, we have used one of the largest sets of self-consistent viscosity measurements collected at high antibody concentrations (>100 mg/mL), including 62 mAbs used for model training and 17 mAbs held out for testing.¹⁸ Second, we have developed a classification model that only requires the Fv amino acid sequence and generation of homology models in a widely accessible computational package (Molecular Operating Environment, MOE). Third, our model is a simple decision tree based on physical antibody properties, such as Fv isoelectric point, which is simple to interpret and use for redesigning antibodies to reduce viscosity. Fourth, we experimentally confirm that the model accurately predicts new mutations that reduce viscosity of suboptimal antibodies.

Figure 1. Overview of approach for training and testing a decision tree model for predicting the level of antibody viscosity for IgG1s and identifying mutations that reduce viscosity.

Results

Identification of antibody molecular features that mediate viscous behavior

The goal of this study was to develop a model for classifying antibodies with either high or low viscosity, which we define herein as those with viscosities >20 centipoise (cP) (high) and < 20 cP (low) for concentrated antibody formulations (>100 mg/mL mAb) at a common pH (pH 5.2). We generated a panel of 79 IgG1 human antibodies that possessed diverse physicochemical properties and germlines (Figure 2), as well as diverse heavy chain CDR3s (Table S1). Notably, the properties of the 79 IgG1s in this study were similar to those of a larger panel of 94 clinical-stage IgG1s, the identities of which are summarized in Dataset S1. Antibody viscosities for the 79 IgG1a were measured at 149 ± 13 mg/mL (pH 5.2), and the actual antibody concentrations for each viscosity measurement have been reported previously.¹⁸

Figure 2. Physicochemical properties and germlines for the panels of 79 IgG1s in this study and 94 clinical-stage IgG1s. (a-d) Distributions of Fv (a) net charge (pH 5.2), (b) isoelectric point, (c) charge symmetry (pH 5.2) and (d) hydrophobicity for IgGs in this study (blue) and clinical-stage antibodies (yellow). (e) Distribution of germline families for IgGs in this study (blue) and clinical-stage antibodies (yellow).

Next, we first split the 79 antibodies into two sets, one that contained 62 antibodies used for feature analysis and model training, and another with 17 antibodies only used for model testing, in a manner that resulted in similar fractions of viscous antibodies in each set (27–35%). We generated five homology models of the Fv regions of each antibody using a standard modeling program (MOE). The models were energy minimized, titrated to pH 5.2, and 35 features from the Fv region were extracted and averaged for the five homology models per antibody.

Notably, several single molecular features were correlated with antibody viscosity (Figure 3a). For example, the total hydrophobic patch area and the percentage of hydrophobic patch area in Fv were positively correlated with antibody viscosity (Spearman’s ρ of 0.43-0.45 and p-value of 5 × 10⁻⁴-2×10⁻⁴). Likewise, the amount of negatively charged patch area, the number of negatively charged patches, and the % of negatively charged patch area in Fv were negatively correlated with antibody viscosity (Spearman’s ρ of 0.33–0.39 and p-value of 8 × 10⁻³-2×10⁻³). Similar findings for both hydrophobicity and negative charge were also observed for the complementarity-determining regions (CDRs). Moreover, we observed that several Fv charge features, such as Fv net charge and isoelectric point (calculated using either sequence or structure), were negatively correlated with viscosity (Spearman’s ρ of 0.43–0.47 and p-value of 5 × 10⁻⁴-1×10⁻⁴). Three of the correlations are shown in greater detail, including for Fv isoelectric point based on sequence (Figure 3b), largest hydrophobic patch area in Fv (Figure 3c), and number of negatively charged patches in Fv (Figure 3d).

Figure 3. Analysis of the antibody Fv molecular features linked to high antibody viscosity. (a) Spearman’s $\mathit{\rho }$ values for the correlations between Fv and CDR features and antibody viscosity. (b-d) Specific examples of correlations between antibody Fv properties and antibody viscosity, including (b) Fv isoelectric point, (c) largest hydrophobic patch area in Fv, and (d) # of negative patches in Fv.

Given the substantial time and resources needed to evaluate antibody viscosity at high concentrations (>100 mg/mL) for the numbers of antibodies in this study, we also evaluated the self-association for the same antibodies using affinity-capture self-interaction nanoparticle spectroscopy (AC-SINS) at much lower antibody concentrations (0.05 mg/mL; Fig. S1).^19–21 Notably, we observed a relatively strong correlation between AC-SINS and viscosity measurements (ρ of 0.75 and p-value of 4 × 10⁻¹²), and similar (albeit weaker) correlations between Fv molecular features and AC-SINS measurements relative to the corresponding correlations with the viscosity measurements.

Decision tree model identifies viscous antibodies using three molecular features

We next sought to develop a model that accurately classified antibodies with high or low viscosity while maintaining simplicity and interpretability. Machine learning models are often ‘black boxes’, learning complex relationships, but they are difficult to interpret and use to guide rational redesign of problematic antibodies. Therefore, we sought to develop decision tree classifiers as our model architecture, which are simple to interpret. We used a standard decision tree classifier algorithm (see Methods for details) to classify antibodies with high (>20 cP) or low (<20 cP) viscosity. To prevent overfitting, decision tree fitting was limited to a maximum depth of three with a minimum of four antibodies per leaf and a minimum impurity decrease of 0.05. Leave-one-out cross-validation was performed to evaluate overfitting, identifying a final model with the highest overall accuracy and least overfitting (Figure 4a). The model performed relatively well with an overall accuracy of 89% (confusion matrix provided in Fig. S2) and an average validation accuracy of 68%. A summary of the antibody viscosities and properties is provided in Supplementary Dataset 2.

Figure 4. Decision tree model for predicting antibodies with low (<20 cP) or high (>20 cP) viscosity at pH 5.2. (a) The model was trained using a set of 62 mAbs, which includes 45 mAbs with high viscosity and 17 mAbs with low viscosity. mAbs with Fv isoelectric points <6.3 were predicted to display high viscosity (Type III antibodies). Moreover, for those with isoelectric points >6.3, mAbs with the largest hydrophobic patch in Fv <261 Ų were predicted to display low viscosity (Type I antibodies). Finally, for those with the largest hydrophobic patch in Fv >261 Å,² mAbs with <5.8 negative patches in Fv were predicted to display high viscosity (Type IV antibodies) and low viscosity for those with >5.8 negative patches in Fv (Type II antibodies). (b-c) The four types of antibodies are separated by their three properties (and the corresponding cutoffs) from the decision tree in (a), which are shown for pairs of Fv properties, namely (b) isoelectric point versus largest hydrophobic patch and (c) isoelectric point versus number of negative patches.

This model identified three physicochemical features in the Fv region for classifying antibodies with different levels of viscosity (Figure 4a). We further grouped antibodies into four types to describe their viscous behaviors based on the terminating decisions given by the model. Antibodies with Fv isoelectric points <6.3 (as calculated based on Fv sequence) were predicted to display high viscosity and categorized as Type III antibodies, which corresponded to 14 of the 62 antibodies in the training set. Ten of the 14 were correctly predicted, while four were false positives for high viscosity.

Next, antibodies with Fv isoelectric points >6.3, which included 48 of the 62 antibodies in the training set, were evaluated in terms of their largest hydrophobic patch in Fv (Figure 4a). The largest patches for most (38 of 48) antibodies were <261 Å,² which resulted in a prediction of low viscosity. These 38 antibodies have Type I behavior (Fv isoelectric points >6.3 and largest hydrophobic patches <261 Ų), and only two of them were incorrectly predicted as low viscosity.

Finally, the remaining 10 antibodies with relatively large hydrophobic patches, as well as Fv isoelectric points >6.3, were next evaluated in terms of their number of negatively charged patches in Fv (Figure 4a). Four antibodies with >5.8 negatively charged patches in Fv were predicted to display low viscosity, which resulted in all of them being correctly scored. Although the number of negatively charged patches is positively correlated with viscosity for the 62 antibodies (Figure 3d), we observe the opposite behavior for the 10 antibodies with relatively high Fv isoelectric points (>6.3) and large hydrophobic patches (>261 Å^2; Figure 4b,c). A relatively high number of negatively charged patches (>5.8) in the Fv region appears to counter the increased hydrophobicity due to the relatively large hydrophobic patches (>261 Ų), which results in a correct prediction of low viscosity for the four antibodies (Figure 4a). These antibodies, which display Type II behavior, possess Fv isoelectric points >6.3, largest hydrophobic patches >261 Å,² and number of negatively charged patches >5.8. Conversely, the other six antibodies, which possess similar sizes of largest hydrophobic Fv patches as Type II antibodies (Figure 4b), possess lower numbers of negatively charged patches (Figure 4c) and were predicted to display high viscosity (Figure 4a). These antibodies displayed Type IV behavior with Fv isoelectric points >6.3, largest hydrophobic patches >261 Ų, and number of negatively charged patches <5.8. Of the six Type IV antibodies, five were correctly predicted. Overall, most of the incorrectly predicted antibodies display one or more unusual physicochemical properties, which we highlight in Table S2.

We next evaluated our model using a hold-out set of 17 antibodies (Figure 4b). This set of antibodies, which includes 11 mAbs with low viscosity and 6 with high viscosity, was unseen during training. Encouragingly, the model correctly predicts the level of viscosity for 15 of 17 antibodies with an overall accuracy of 88% (confusion matrix in Fig. S2), which is similar to the overall training accuracy of 89% for the 62 antibodies. Two of the incorrectly predicted antibodies in the hold-out set were experimentally determined to be viscous but predicted to display low viscosity. Overall, these findings demonstrate the strong prediction ability of our decision tree model to differentiate between antibodies with high and low viscosities.

We also compared the performance of our three-feature model to models based only on each individual feature in our model and two previous models for the entire set of 79 antibodies (Figure 6).^22,23 Our model displayed relatively high AUC (0.92), accuracy (0.89), precision (0.79), recall (0.83) and F1 score (0.81; Figure 6a). In contrast, single feature models displayed reduced performance, even though we selected the best cutoff that maximized model (AUC) performance (Figure 6b–d). For example, an Fv pI cutoff of >6.3 achieved a similar AUC (0.87) as our three-feature model (AUC of 0.89), but the accuracy and other performance metrics of the single feature model were lower (Figure 6b). We also used a previously reported model for predicting viscosity at pH 5.5 and 180 mg/mL based on the Fv charge, charge symmetry (also known as cross charge) and hydrophobicity index.²² For this model, we identified an optimized cutoff value for our dataset, and observed modest performance relative to our model (Figure 6e). Moreover, we evaluated a second model that is based on the Hamaker constants and charges of the variable regions, identified an optimized cutoff for our dataset, and observed modest performance relative to our model (Figure 6f). Finally, we tested a decision tree model based on the antibody net charge and the normalized number of hydrophilic minus hydrophobic residues in the Fv regions based on a previously reported model,¹⁷ but found that this model predicted all 79 antibodies in this study (Dataset S2) to be low viscosity. A summary of the performance metrics for each model corresponding to the overall (79 mAb), training (62 mAbs) and hold-out (17 mAbs) sets of antibodies is provided in Table S3.

We also tested the ability of our model to classify a set of 38 antibody mutants whose viscosities have been reported at pH 5.8 and 150 mg/mL,²⁴ and also observed strong model performance (AUC of 0.98; Fig. S3). Finally, we also tested our model using an independent set of 21 antibodies with viscosity data at pH 6 (10 mM histidine),¹⁷ and also observed strong performance (AUC of 0.85). This suggests that our model could be used to rank antibodies for their relative risk for high viscosity, even at different formulation conditions and for antibodies with diverse sequences.

Model-guided design of mutations that reduce antibody viscosity

The strong model performance observed in Figures 4–6 motivated us to test if we could use this model to design mutations that reduce the viscosity of antibodies with unacceptably high viscosities. Therefore, we selected two antibodies from our set, namely A24 and A54, that display high viscosity (219 cP for A24 and 23 cP for A54), and designed mutations that are predicted to reduce their viscosities (Figures 7a,b). The Fv isoelectric points for both wild-type antibodies were below the cutoff of 6.3, including values of isoelectric point values of 5.7 for A24 and 6.2 for A54, which resulted in the correct prediction of high viscosity for both antibodies. Therefore, we sought to introduce mutations into the Fv region that increased the isoelectric point of the Fv region. Moreover, for designed antibody variants with isoelectric points >6.3, we also sought to introduce mutations that would result in the largest hydrophobic patches in Fv to be below the model cutoff (<261 Ų). For A54, this was especially important because the wild-type antibody had a hydrophobic patch above this cutoff (292 Ų), while the other wild-type antibody (A24) did not (120 Ų).

Figure 5. Evaluation of decision tree model for predicting antibodies with low or high viscosity using a hold-out set of antibodies. The model in Figure 4a was tested with a hold-out set of 17 mAbs, which includes 11 mAbs with low viscosity and 6 mAbs with high viscosity. The hold-out set of antibodies was not used during model training.

Figure 6. Evaluation of decision tree model performance relative to the performance of single features and previously reported models. (a-d) The performance of the (a) model in this work for the 79 mAbs was compared to single feature models, including (b) Fv pI, (c) largest Fv hydrophobic patch, and (d) # of negative patches in Fv. In addition, the performance of two additional previously reported models is shown, including (e) model #2 (Sharma, Patapoff et al., PNAS, 2014) and (f) model #3 (Lai, Swan et al., mAbs, 2021).

We chose to mutate the heavy chain CDRs of the antibodies, and only introduce mutations at sites in which the wild-type residue is not highly conserved and the mutant residue is relatively common in human repertoires (see Methods for more detail). This led to the selection of 1–2 mutations per variant for A24 and 2–3 mutations per variant for A54. The six variants of A24 had higher isoelectric points (pIs of 6.6–7.2) than the wild-type antibody (pI of 5.7), and comparable hydrophobic patches (120–126 Ų) than the wild-type antibody (120 Ų). Likewise, the two variants of A54 had higher isoelectric points (pIs of 6.6–6.9) than the wild-type antibody (pI of 6.2), and smaller hydrophobic patches (212–244 Ų) than the wild-type antibody (292 Ų).

Next, we produced the eight antibody variants and evaluated their high concentration viscosities (Figures 7c,d). Notably, we observe that the six variants of A24 display large decreases in antibody viscosity and all are below 20 cP (Figure 7c). Similarity, the two variants of A54 also displayed reduced viscosity (Figure 7d). Given that the viscosity of the wild-type antibody for A54 (23 cP) was much lower than that for A24 (219 cP), the reductions in viscosity for A54 were smaller, but the viscosities of both variants were below the 20 cP cutoff. The variant with a single mutation [I(98)K in V_H] displayed intermediate reduction in viscosity (13.1 cP), while the double mutant [Y(58)K and G(97)K in V_H] resulted in a further reduction in viscosity (8.5 cP). Overall, these findings demonstrate how an interpretable model can be used to design mutations that reduce antibody viscosity in a simple and predictable manner.

Figure 7. Design of mutations that reduce antibody viscosity. (a-b) Mutations were designed using the decision tree model for two antibodies (A24 and A54), and (c-d) tested experimentally at 150 mg/mL. In (a), the feature values are shown for the wild-type antibodies as orange (A24) and green (A54) solid lines.

Model identifies clinical-stage antibodies with predicted high viscosity

The ability of our model to predict antibodies with high viscosity led us to next evaluate predictions for clinical-stage IgG1s. Using the same panel of 94 clinical-stage IgG1s reported in Figure 2 (Dataset S1), we classified the antibodies by their predicted viscosity behavior (Figure 8a). Notably, 18% of antibodies were predicted to display high viscosity in the set of clinical-stage antibodies (total for Type III and IV antibodies) relative to 30% of antibodies in the panel of antibodies in this study. Moreover, most of the predicted high viscosity antibodies in the clinical-stage set were Type III antibodies, as observed in our antibody set, which correspond to antibodies with low Fv pIs (pI < 6.3). Conversely, most of the predicted low viscosity antibodies are Type I antibodies in both set of antibodies, which correspond to antibodies with moderate-to-high Fv pIs (pI>6.3) and largest hydrophobic patches <261 A². These findings are shown more directly in terms of the Fv pI (Figure 8b) and net charge at pH 5.2 (Figure 8c) for the two sets of antibodies, which reveals that low pI and low net charge (pH 5.2) are the primary predicted risk factors for antibodies with high viscosity for the formulation studied in this work.

Figure 8. Evaluation of viscosity model predictions and antibody properties for clinical-stage mAbs relative to the mAbs in this study. (a) The percentage of mAbs with each type of predicted viscosity behavior, as defined in Fig. 4. (b-c) Distribution of (b) Fv pI and (c) Fv net charge at pH 5.2 for different levels of predicted viscosity. In (a), the number of mAbs is shown on top of each bar.

Discussion

One interesting aspect of our work is the identification of molecular features that are strongly linked to either increased or decreased antibody viscosity. Our finding that Fv isoelectric point is the most strongly negatively correlated feature with antibody viscosity (Spearman’s $\mathit{\rho }$ of −0.48 and p-value of 9×10⁻⁵) is consistent with several previous studies.^{16,17,22–28} Likewise, our finding that the % of hydrophobic patches in the Fv region is the most strongly positively correlated feature with antibody viscosity (Spearman’s $\mathit{\rho }$ of 0.45 and p-value of 0.0002) is also generally consistent with the findings from several previous studies that also identified the importance of hydrophobicity in mediating antibody viscosity.^16,17,22 It is interesting that antibody electrostatic and hydrophobic properties displayed similar levels of importance in our study, which has been observed in some previous studies^16,22 while charge was observed to be much more important than hydrophobicity in other studies.^10,25 While the molecular origins of the differences in the relative importance of antibody charge and hydrophobicity in different studies is unclear, it is notable that our study is based on the largest set of antibodies with high concentration viscosity measurements reported to date and may provide unique insights into the determinants of antibody viscosity.

It is also interesting that our model reveals four types of antibodies with unique mechanisms of mediating high viscosity. Type III antibodies, which are the most common type of viscous antibodies (Figure 8a), display low Fv isoelectric points and likely mediate high viscosity due to attractive electrostatic interactions. In contrast, Type IV antibodies appear to be viscous due to hydrophobic interactions, as they have large hydrophobic patches and low numbers of negative patches. For the non-viscous antibodies, Type I antibodies are the most common and appear to minimize viscosity due to minimizing attractive electrostatic and hydrophobic interactions via their combination of relatively high isoelectric points and small hydrophobic patches. Interestingly, Type III antibodies, which are less common and possess large hydrophobic patches, appear to minimize viscosity by reducing hydrophobic interactions via their relatively high numbers of negative patches. The latter phenomena of reducing hydrophobic interactions using negatively charged patches has been observed in previous studies of antibody viscosity²⁹ and aggregation.^30–34

Our classification model for predicting the level of antibody viscosity also deserves further consideration. It is notable that our model does not require proprietary methods and can be immediately implemented by others. The only requirement is the use of the MOE software package to generate homology Fv structures and extract the two structure-based features (largest hydrophobic patch and number of negative patches in Fv) in addition to the single sequence-based feature (Fv isoelectric point). This is significant because several reported models are not widely used because they are proprietary and/or overly complex.^15,16,23,27 Another notable advantage of our model is its interpretability, as it is simple to evaluate the predicted origins of antibodies with either high or low viscosity. We demonstrated that the interpretability of our model could be readily used to design mutations to reduce antibody viscosity. While some interpretable viscosity models have been reported,^17,22 none have been reported based on as large of antibody datasets as those reported in our study nor have they been used to design viscosity-reducing mutations.

It is also important to consider that the viscosity measurements were performed at pH 5.2 (10 mM sodium acetate and 9% sucrose), and the model predictions may be limited in their ability to accurately predict the level of viscosity at other solution conditions. For example, we demonstrated that our model strongly differentiated between a panel of antibody mutants with a wide range of viscosities (Fig. S3),²⁴ but antibodies with predicted low viscosities were generally those with viscosities ~10–100 cP instead of ~5–20 cP in our study. The difference is likely due to the fact that the previous study was performed in a higher pH formulation (pH 5.8), which would be expected to result in higher viscosities relative to the formulations at lower pH (pH 5.2) in our study based on general trends observed previously.^35,36 While it would be simple to retrain our model at other formulation conditions as such viscosity data becomes available, we caution users to consider this issue when extrapolating our model to other formulation conditions. Moreover, the model was trained on human IgG1 antibodies, and extrapolation to other human antibody subclasses (e.g., human IgG4s) may be problematic due to their unique charge and hydrophobicity properties.^37–39 Finally, our proof-of-concept study to design antibody mutations that reduce viscosity did not address the effects of these mutations on binding affinity, and future work is needed to evaluate the feasibility of designing viscosity-reducing mutations, including those in the CDRs, that also maintain high binding affinity.

Our findings also suggest opportunities for several exciting future studies and applications. First, our model can be readily used early in the discovery process to identify antibody candidates with low predicted levels of viscosity. This is important because our model only requires the Fv amino acid sequence and could reduce the risk of developing antibodies that will display unacceptably high viscosities later in the development process when experimental viscosity characterization is feasible. Second, our model can be readily used to design mutations that reduce viscosity. While we demonstrated this direction in this study, we envision that our model could be used in concert with other experimental or computational approaches for co-optimizing antibody properties that also strongly depend on the same variable regions, such as affinity, solubility, aggregation, and nonspecific binding.^9,34,40–46 Given that common antibody properties such as isoelectric impact antibody properties in opposite ways,^34,47 we expect our model will help constrain the selection of antibody mutations that co-optimize other key antibody properties while simultaneously minimizing viscosity. These and other applications of our model are expected to improve the predictable generation of drug-like therapeutic antibodies that are well suited for use in high concentration antibody formulations necessary for subcutaneous delivery.

Materials and methods

Antibodies and their properties

The 79 mAbs in this study were human IgG1s and were produced as reported previously.¹⁸ The antibody constant regions were from the human IgG1z SEFL2.2 scaffold,⁴⁸ which has been engineered to lack effector function through removal of the C_H2 glycosylation site (N297G) and for improved stability through introduction of a disulfide bond (R292C, V302C). For the 94 clinical-stage IgG1s, the amino acid sequences of the variable (V_H and V_L) regions were obtained from two previous publications.^49,50 The physicochemical properties of the antibodies, including Fv isoelectric point (pI), sequenced-based Fv net charge at pH 5.2, Fv charge asymmetry at pH 5.2 (the product of the V_H and V_L net charges), and hydrophobicity, were calculated using multiple methods. The Fv pI was extracted from homology models generated using MOE software. The sequence-based net charges at pH 5.2 in different regions (Fv, V_H and V_L) were calculated using Biopython (version 1.78). The hydrophobicity was calculated based upon the Eisenberg scale for each amino acid.⁵¹ The germline family of each antibody was determined by ANARCI.⁵²

Affinity-capture self-interaction nanoparticle spectroscopy

For AC-SINS measurements, capture antibody (0.4 mg/mL, Jackson ImmunoResearch, 109-005-008) was buffer exchanged (pH 4.3, 20 mM sodium acetate), mixed with gold nanoparticles (used as is), and incubated for 1 h at 4°C. Thiolated PEG (Sigma, P4338-1 KG) was added (final concentration of 0.1 mM) and incubated overnight at 4°C. The gold-capture conjugate solution was passed through a 0.22 μm PVDF membrane (Millex-GV, 13 mm, Millipore) with particles retained on the membrane. Phosphate-buffered saline (PBS) at 1/40 of the starting volume was used to elute concentrated conjugated particles. Dilute antibody solutions (50 μg/mL, 45 μL; pH 5.2, 10 mM sodium acetate, 150 mM NaCl) were incubated with 5 μL of gold conjugates for 1 h at room temperature. Absorbance spectra were measured on a Biotek Synergy Neo plate reader (Biotek, Winooski, VT) in 1 nm increments between 450 and 650 nm. The wavelength of the inflection point of a quadratic equation fit to describe the 40 data points surrounding the maximum measured absorbance was calculated to determine the plasmon wavelength. Plasmon wavelengths were normalized between high (CNTO-607) and low control AC-SINS antibodies.

Antibody structural modeling

The antibody homology models were generated using MOE software with the Amber10:EHT forcefield and a dielectric constant value of 4. The MOE antibody modeler was used to build an initial homology model based on searching templates from a Fab/antibody structure database, including antibody structures in Protein Data Bank (PDB), for V_H, V_L and individual CDR loops. Finally, the initial antibody model was energy minimized with a minimum gradient setting of 0.00001 RMS kcal/mol/A.² Protein properties that were descriptive guides for selection of amino acid substitutions were extracted from final homology models. Models were then exported to PyMOL for visualization.

Model development

Machine learning analysis was implemented via sklearn package (version 1.0.2) in Python (version 3.8). The DecisionTreeClassifier algorithm was trained on the feature set using gini impurity, a max depth of 3, a minimum of 4 samples per leaf, and a minimum impurity decrease of 0.05 per split. Different models were generated using the GridSearchCV algorithm. The best model was selected to maximize both training and testing performance. Leave-one-out cross validation was used to evaluate model robustness.

Antibody mutant design

Two antibodies, A24 and A54, were selected for optimization, capturing a wide range of suboptimal viscosity behavior. Mutations that altered feature values in a desirable manner were tested in a virtual screen. The primary CDR sites that were considered for mutation were those whose wild-type residues occur in human antibody repertoires at frequencies of <50%, and the mutations that were primarily considered at each CDR site are those that occur in human repertoires at frequencies of >10%. The frequencies do not consider specific germlines, as the frequency values are averaged over all human germlines. The human antibody repertoire frequencies for each CDR site were accessed at abYsis (species: homo sapiens). Mutations that satisfied these criteria were screened individually by creating homology models of their respective antibodies with each single mutation. Single mutations that improved the value of at least one model feature without negatively impacting the others were further considered. These mutations were tested as multi-mutation variants.

The mutation design strategies for both antibodies involved an increase in the isoelectric point of the Fv region, suggesting that the removal of negative charge and the addition of positive charge would be beneficial for reducing viscosity. In addition to this, the model predicted that A54 exhibited too large of a hydrophobic patch in the Fv region, which would need to be remediated if the antibody would be classified as optimal. Therefore, mutations that disrupted this hydrophobic patch, in HCDR3, were also considered. Successful variants that achieved optimal predictions were selected for experimental characterization.

Production of engineered variants

The engineered variants of A24 and A54 were produced for viscosity measurement at Amgen (Thousand Oaks, CA). The heavy and light chains were cloned into a proprietary monocistronic Amgen mammalian expression vector with either puromycin or hygromycin resistance cassettes, respectively, and transfected into CHO-K1 cells. Dual antibiotic selection media was added on day 3 post-transfection, and cell recovery was followed using a Vi-CELL BLU cell viability counter (Beckman Coulter). Once cells had reached >90% viability, batch productions were seeded at 1E6 viable cells/mL. After 7 days, cell viability and viable cell density were measured, cells were removed by centrifugation, and CM was filtered using 0.2 µm cellulose acetate filters. CM titers were measured by ForteBio analysis using Protein A sensor tips (Molecular Devices, Fremont, CA).

For purification of the A24 variants, clarified supernatants were affinity captured by MabSelect SuRe chromatography (Cytia, Piscataway, NJ) using 25 mM Tris, 100 mM NaCl, pH 7.4 as the first wash buffer and 25 mM Tris-HCl, 500 mM L-Arg-HCl, pH 7.5 as the second wash buffer, followed by a third wash repeating the 25 mM Tris, 100 mM NaCl, pH 7.4 buffer. The elution buffer was 100 mM acetic acid, pH 3.6. Peak fractionation used to collect the Protein A elution was initiated when the absorbance at 280 nm was above 50 mAU and stopped when the absorbance was less than 100 mAU. The elution pools were diluted in line 1:1 with 50 mM MES, pH 6.5 and filtered through a 0.22 μm cellulose acetate filter.

For purification of A54 variants, clarified supernatants were purified by MabSelect SuRe chromatography followed by SP Sepharose high-performance anion exchange chromatography (Cytia, Piscataway, NJ) using an ÄKTA instrument configured for automated tandem purification of the 5 mL column as described elsewhere.⁵³ The wash buffer for the affinity column was 25 mM Tris pH 7.4, 100 mM NaCl. The elution buffer was 100 mM acetic acid, pH 3.6, and the Protein A elution was initiated when the absorbance at 280 nm was above 50 mAU and stopped when the absorbance was less than 50 mAU. The Protein A eluate was in-line diluted 1:4 with 50 mM sodium acetate, pH 5.0 and directly loaded on a SP Sepharose high-performance column and then washed with 5 column volumes of 20 mM sodium acetate, pH 5.0 followed by elution using a 40-column volume gradient to 60% SP-Buffer B (20 mM sodium acetate, 1 M NaCl, pH 5.0). Peak fractionation was used to collect 1.25 mL fractions, starting when the absorbance at 280 nm reached 50 mAU and ending when the absorbance dropped below that value. Pools of the SP fractions were made based on the main peak of the chromatogram.

Concentration of IgGs

For parental antibodies (A24 and A54) as well as the A24 variants, the final elution pools from chromatography were dialyzed against approximately 30 volumes of 10 mM sodium acetate, 9% sucrose, pH 5.2 using Slide-A-Lyzer dialysis cassettes with a 10 kDa cutoff membrane (Thermo Scientific, Waltham, MA) and further concentrated to 150 mg/mL (within 10%) using Vivaspin-20 centrifugal concentrators with a 30 kDa cutoff membrane (Sartorius Stedim Biotech, Goettingen, Germany). For A54 variants, the pools were first reduced in volume, buffer exchanged with 10 mM sodium acetate, 9% sucrose, pH 5.2 and brought to the final concentration using the Big Tuna buffer exchange device (Unchained Labs, Pleasanton CA).

For all antibodies, concentrated material was filtered through a 0.2 μm cellulose acetate filter, and the concentration was determined by the absorbance at 280 nm using the calculated extinction coefficient. Sample purity was determined by LabChip GXII analysis under non-reducing (with 25 mM iodoacetamide) conditions. Analytical SEC was carried out using a BEH200 column (Waters, Milford, MA) with an isocratic elution 100 mM NaPO₄, 250 mM NaCl, pH 6.9.

Viscosity measurements

Viscosities of A24 and A54 antibodies were measured using an Anton Paar MCR Rheometer. A flow sweep procedure was applied from 10 to 1000 1/s using a steel 20 mm Peltier plate with 1.988° Cone. Viscosity was measured in Pa-s, where 1 mPa-s = 1 cP at 1000 1/s. An aliquot of 80 µL was loaded onto the plate for each measurement. Viscosity was measured for each molecule at approximately 150 mg/mL with 0.01% Tween 80 surfactant added for a final formulation buffer of 10 mM sodium acetate, 9% sucrose, 0.01% Tween 80, pH 5.2.

Viscosities of optimized variants were measured using a Viscometer-Rheometer-on-a-Chip (VROC) Initium One Plus instrument (RheoSense, San Ramon, CA). The VROC is a rheometer/viscometer that integrates both a microfluidic channel and a microelectromechanical system (MEMS) pressure sensor array. The VROC chip calculates the viscosity based on the Hagen-Poiseuille equation by measuring the pressure drop along the rectangular microfluidic channel at a given flow rate. The flow rate was automatically selected for each sample to ensure optimal signal on the chip pressure sensors (>5% full scale). The VROC initium includes a robotic autosampler and a 100 µL test syringe that injects the sample to the VROC chip. The protocol was programmed to inject 50 µL of the sample to the chip. The E02 measuring chip was selected for all sample measurements and measurements were repeated 10 times at 25°C. Viscosity was reported as the average of the 10 repeated measurements. The E02 chip and test syringe were both temperature-controlled at 25°C ± 0.1°C with a 2 min soak time before viscosity measurement.

Abbreviations

AC-SINS	=	affinity-capture self-interaction nanoparticle spectroscopy
CDR	=	complementarity-determining region
CM	=	conditioned media
Fv	=	fragment variable
IgG	=	immunoglobulin
mAb	=	monoclonal antibody
PBS	=	phosphate-buffered saline
pl	=	isoelectric point
VH	=	variable heavy
VL	=	variable light

Acknowledgments

We thank members of the Tessier lab, Kenneth Walker, Yax Sun, and Iain Campuzano for their helpful suggestions. We thank Kevin Graham, Katie Douglas, Anne Abrajano, Ling Liu, Ning Sun, Cai Guo, Eric Gislason, Mai Tran, Julian Reed, Vivian Li, Noi Nuanmanee, Qian (Ken) Chen and Carl Kolvenbach for contributions to antibody cloning, expression, purification, and analytics. The authors declare the following competing financial interest(s): M.M. and E.M.P-O are full-time employees, and M.M., E.M., D.W., and E.M.P-O are shareholders of Amgen Inc.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2024.2303781

Additional information

Funding

The work was supported by the National Institutes of Health [1T32GM140223-01]; National Science Foundation [CBET 1804313,CBET 1803497]; University of Michigan. [Rackham Graduate Fellowship and Albert M. Mattocks Chair].