Machine Learning Applied to Determine the Molecular Descriptors Responsible for the Viscosity Behavior of Concentrated Therapeutic Antibodies.

Predicting the solution viscosity of monoclonal antibody (mAb) drug products remains as one of the main challenges in antibody drug design, manufacturing, and delivery. In this work, the concentration-dependent solution viscosity of 27 FDA-approved mAbs was measured at pH 6.0 in 10 mM histidine-HCl. Six mAbs exhibited high viscosity (>30 cP) in solutions at 150 mg/mL mAb concentration. Combining molecular modeling and machine learning feature selection, we found that the net charge in the mAbs and the amino acid composition in the Fv region are key features which govern the viscosity behavior. For mAbs whose behavior was not dominated by charge effects, we observed that high viscosity is correlated with more hydrophilic and fewer hydrophobic residues in the Fv region. A predictive model based on the net charges of mAbs and a high viscosity index is presented as a fast screening tool for classifying low- and high-viscosity mAbs.

Computational Screening for mAb Colloidal Stability with Coarse-Grained, Molecular-Scale Simulations.

Monoclonal antibodies (mAbs) are an important modality of protein therapeutics with broad applications for numerous diseases. However, colloidal instabilities occurring at high protein concentrations can limit the ability to develop stable, high-concentration liquid dosage forms that are required for patient-centric, device-mediated products. Therefore, it is advantageous to identify colloidally stable mAbs early in the discovery process to ensure that they are selected for development. Experimental screening for colloidal stability can be time- and resource-consuming and is most feasible at the later stages of drug development due to material requirements. Alternatively, computational approaches have emerging potential to provide efficient screening and focus developmental efforts on mAbs with the greatest developability potential, while providing mechanistic relationships for colloidal instability. In this work, coarse-grained, molecular-scale models were fine-tuned to screen for colloidal stability at amino-acid resolution. This model parameterization provides a framework to screen for mAb self-interactions and extrapolate to bulk solution behavior. This approach was applied to a wide array of mAbs under multiple buffer conditions, demonstrating the utility of the presented computational approach to augment early candidate screening and later formulation strategies for protein therapeutics.

A Systematic Approach to Evaluating Closed System Drug-Transfer Devices During Drug Product Development.

The use of Closed System Drug-Transfer Devices (CSTDs) has increased significantly in recent years due to NIOSH and USP recommendations to use them during preparation of hazardous drugs. Mechanistic and material differences between CSTDs and traditional in-use components warrant an assessment of their impact on product quality and dosing accuracy. Using a combination of prevalent CSTDs with biologic molecules, we performed an extensive assessment of the effect of using CSTDs for dose preparation and observed no negative impact on product quality attributes. Additionally, we found that the CSTD hold-up volume is 2 to 4-fold higher than conventional in-use components and exhibited a strong dependence on the CSTD brand used. We also found that the CSTD brand and dosing volume have a major influence on dosing accuracy with suboptimal protein recovery at very low dosing volumes. We identified entrapment of product in the CSTD spike as the root cause for this sub-optimal recovery and found that flushing the CSTD spike with a brand-new syringe and not the dosing syringe aided in complete protein recovery. Taken together we present a systematic approach to evaluate the risks and impact of CSTD to drug product quality, dose preparation, and dosing accuracy.

Machine Learning Feature Selection for Predicting High Concentration Therapeutic Antibody Aggregation.

Protein aggregation can hinder the development, safety and efficacy of therapeutic antibody-based drugs. Developing a predictive model that evaluates aggregation behaviors during early stage development is therefore desirable. Machine learning is a widely used tool to train models that predict data with different attributes. However, most machine learning techniques require more data than is typically available in antibody development. In this work, we describe a rational feature selection framework to develop accurate models with a small number of features. We applied this framework to predict aggregation behaviors of 21 approved monospecific monoclonal antibodies at high concentration (150 mg/mL), yielding a correlation coefficient of 0.71 on validation tests with only two features using a linear model. The nearest neighbors and support vector regression models further improved the performance, which have correlation coefficients of 0.86 and 0.80, respectively. This framework can be extended to train other models that predict different physical properties.

Characterizing aggregate growth and morphology of alanine-rich polypeptides as a function of sequence chemistry and solution temperature from scattering, spectroscopy, and microscopy.

The aggregation behavior and stability of a series of alanine-rich peptides, which are included as components of peptide-polymer conjugates, were characterized using a combination of biophysical techniques. Light scattering techniques were used to monitor changes in peptide morphology and size distributions as a function of time and temperature. The results show large particles immediately upon dissolution in buffer. At room temperature, these particles relaxed to reach a mostly monomeric peptide state, while at higher temperatures, they grew to form aggregates. Circular dichroism spectroscopy (CD) was used to monitor temperature- and time-dependent conformational changes as a function of peptide sequence and incubation time. CD measurements reveal that all of the sequences are helical at low temperatures with transitions to non-helical conformation with increased temperature. Samples incubated at room temperature were able to recover their original helicity. At increased temperature, the shorter and longer peptide sequences showed notable changes in conformation, and were not able to recover their original helicity after 72 h. After incubation for up to one week, ß-sheet conformations were observed in these two cases, while only α-helical conformation loss was observed for the peptide of intermediate molecular weight. Transmission electron microscopy measurements reveal the formation of fibrils after 72 h of incubation at 60 °C for all samples, in agreement with the scattering measurements. Additional quenching experiments show that peptide aggregation can be stalled when solutions are cooled to room temperature.

A single molecular descriptor to predict solution behavior of therapeutic antibodies.

Despite the therapeutic success of monoclonal antibodies (mAbs), early identification of developable mAb drug candidates with optimal manufacturability, stability, and delivery attributes remains elusive. Poor solution behavior, which manifests as high solution viscosity or opalescence, profoundly affects the developability of mAb drugs. Using a diverse dataset of 59 mAbs, including 43 approved products, and an array of molecular descriptors spanning colloidal, conformational, charge-based, hydrodynamic, and hydrophobic properties, we show that poor solution behavior is prevalent (>30%) in mAbs and is singularly predicted (>90%) by the diffusion interaction parameter (k D), a dilute-solution measure of colloidal self-interaction. No other descriptor, individually or in combination, was found to be as effective as k D. We also show that well-behaved mAbs, a substantial subset of which bear high positive charge and pI, present no disadvantages with respect to pharmacokinetics in humans. Here, we provide a systematic framework with quantitative thresholds for selecting well-behaved therapeutic mAbs during drug discovery.

Predicting High-Concentration Interactions of Monoclonal Antibody Solutions: Comparison of Theoretical Approaches for Strongly Attractive Versus Repulsive Conditions.

Nonspecific protein-protein interactions of a monoclonal antibody were quantified experimentally using light scattering from low to high protein concentrations (c2) and compared with prior work for a different antibody that yielded qualitatively different behavior. The c2 dependence of the excess Rayleigh ratio (Rex) provided the osmotic second virial coefficient (B22) at low c2 and the static structure factor (Sq=0) at high c2, as a function of solution pH, total ionic strength (TIS), and sucrose concentration. Net repulsive interactions were observed at pH 5, with weaker repulsions at higher TIS. Conversely, attractive electrostatic interactions were observed at pH 6.5, with weaker attractions at higher TIS. Refined coarse-grained models were used to fit model parameters using experimental B22 versus TIS data. The parameters were used to predict high-c2 Rex values via Monte Carlo simulations and separately with Mayer-sampling calculations of higher-order virial coefficients. For both methods, predictions for repulsive to mildly attractive conditions were quantitatively accurate. However, only qualitatively accurate predictions were practical for strongly attractive conditions. An alternative, higher resolution model was used to show semiquantitatively and quantitatively accurate predictions of strong electrostatic attractions at low c2 and low ionic strength.

Light Scattering to Quantify Protein-Protein Interactions at High Protein Concentrations.

Static and dynamic (laser) light scattering (SLS and DLS, respectively) can be used to measure the so-called weak or colloidal protein-protein interactions in solution from low to high protein concentrations (c2). This chapter describes a methodology to measure protein-protein self-interactions using SLS and DLS, with illustrative examples for monoclonal antibody solutions from low to high protein concentrations (c2 ~ 1-102 g/L).

Electrostatically Mediated Protein-Protein Interactions for Monoclonal Antibodies: A Combined Experimental and Coarse-Grained Molecular Modeling Approach.

Electrostatically mediated protein-protein interactions (PPI) can influence key product properties such as solubility, solution viscosity, and aggregation rates. Predictive models would allow for candidates/formulations to be screened with little or no protein material. Three monoclonal antibodies that display qualitatively different experimental PPI were evaluated at a range of pH and ionic strength conditions that are typical of product formulations. PPI parameters (kD, B22, and G22) were obtained from static and dynamic light scattering measurements and spanned from strongly repulsive to strongly attractive net interactions. Coarse-grained (CG) molecular simulations of PPI (specifically, B22) were compared against experimental PPI parameters across multiple pH and salt conditions, using a CG model that treats each amino acid explicitly. Predicted B22 values with default model parameters matched experimental B22 values semiquantitatively for some cases; others required parameter tuning to account for effects such as ion binding. Experimental PPI values were also analyzed for each monoclonal antibody within the context of single-protein properties such as net charge, and domain-based and global dipole moments. The results show that PPI predicted qualitatively and semiquantitatively by CG molecular modeling of B22 can be an effective computational tool for molecule and formulation assessment.

Predicting Protein-Protein Interactions of Concentrated Antibody Solutions Using Dilute Solution Data and Coarse-Grained Molecular Models.

Protein-protein interactions for solutions of an IgG1 molecule were quantified using static light scattering (SLS) measurements from low to high protein concentrations (c2). SLS was used to determine second osmotic virial coefficients (B22) at low c2, and excess Rayleigh profiles (Rex/K vs. c2) and zero-q structure factors (Sq=0) as a function of c2 at higher c2 for a series of conditions (pH, sucrose concentration, and total ionic strength [TIS]). Repulsive (attractive) interactions were observed at low TIS (high TIS) for pH 5 and 6.5, with increasing repulsions when 5% w/w sucrose was also present. Previously developed and refined coarse-grained antibody models were used to fit model parameters from B22 versus TIS data. The resulting parameters from low-c2 conditions were used as the sole input to multiprotein Monte Carlo simulations to predict high-c2Rex/K and Sq=0 behavior up to 150 g/L. Experimental results at high-c2 conditions were quantitatively predicted by the simulations for the coarse-grained models that treated antibody molecules as either 6 or 12 (sub) domains, which preserved the basic shape of a monoclonal antibody. Finally, preferential accumulation of sucrose around the protein surface was identified via high-precision density measurements, which self-consistently explained the simulation and experimental SLS results.

Biophysical characterization and molecular simulation of electrostatically driven self-association of a single-chain antibody.

Colloidal protein-protein interactions (PPI) are often expected to impact key behaviors of proteins in solution, such as aggregation rates and mechanisms, aggregate structure, protein solubility, and solution viscosity. PPI of an anti-fluorescein single chain antibody variable fragment (scFv) were characterized experimentally at low to intermediate ionic strength using a combination of static light scattering and sedimentation equilibrium ultracentrifugation. Surprisingly, the results indicated that interactions were strongly net-attractive and electrostatics promoted self-association. Only repulsive interactions were expected based on prior work and calculations based a homology model of a related scFv crystal structure. However, the crystal structure lacks the charged, net-neutral linker sequence. PyRosetta was used to generate a set of scFv structures with different linker conformations, and coarse-grained Monte Carlo simulations were used to evaluate the effect of different linker configurations via second osmotic virial coefficient (B22 ) simulations. The results show that the configuration of the linker has a significant effect on the calculated B22 values, and can result in strong electrostatic attractions between oppositely charged residues on the protein surface. This is particularly relevant for development of non-natural antibody products, where charged linkers and other loop regions may be prevalent. The results also provide a preliminary computational framework to evaluate the effect of unstructured linkers on experimental protein-protein interaction parameters such as B22 .

Protein Partial Molar Volumes in Multicomponent Solutions from the Perspective of Inverse Kirkwood-Buff Theory.

Inverse Kirkwood-Buff (KB) solution theory can be used to relate macroscopic quantities with molecular scale interactions and correlation functions, in the form of KB integrals. Protein partial specific volumes ([Formula: see text]) from high-precision density measurements can be used to quantify solvent-solute and solute-solute KB integrals. Currently, general expressions for [Formula: see text] as a function of cosolute concentration (c3) have been provided for only binary and ternary solutions. We derive a general multicomponent expression for [Formula: see text] in terms of the relevant KB integrals for the case of low (infinite dilution) protein concentration but arbitrary cosolute concentrations. To test the utility of treating a quaternary system with a pseudoternary approximation, α-Chymotrypsinogen (aCgn) solutions with a series of solutes (NaCl, sucrose, and trehalose) were compared as a function of solute concentration with and without buffer present. Comparison between those ternary and quaternary solutions shows equivalent results within experimental uncertainty and suggests the pseudoternary approximation may be reasonable. In the case of aCgn, doing so also revealed that the preferential interactions can depend on pH. Analysis of steric contributions also provides an example that illustrates how KB integrals allow one to interpret [Formula: see text] in terms of contributions from molecular volume, excluded volume, and hydration/solvation effects.

Predicting Protein Interactions of Concentrated Globular Protein Solutions Using Colloidal Models.

Protein interactions of α-chymotrypsinogen A (aCgn) were quantified using light scattering from low to high protein concentrations. Static light scattering (SLS) was used to determine the excess Rayleigh ratio (Rex) and osmotic second virial coefficients (B22) as a function of pH and total ionic strength (TIS). Repulsive (attractive) protein-protein interactions (PPI) were observed at pH 5 (pH 7), with decreasing repulsions (attractions) upon increasing TIS. Simple colloidal potential of mean force models (PMF) that account for short-range nonelectrostatic attractions and screened electrostatic interactions were used to fit model parameters from data for B22 vs TIS at both pH values. The parameters and PMF models from low-concentration conditions were used as the sole input to transition matrix Monte Carlo simulations to predict high concentration Rex behavior. At conditions where PPI are repulsive to slightly attractive, experimental Rex data at high concentrations could be predicted quantitatively by the simulations. However, accurate predictions were challenging when PPI were strongly attractive due to strong sensitivity to changes in PMF parameter values. Additional simulations with higher-resolution coarse-grained molecular models suggest an approach to qualitatively predict cases when anisotropic surface charge distributions will lead to overall attractive PPI at low ionic strength, without assumptions regarding electrostatic "patches" or multipole expansions.

Coarse-Grained Antibody Models for "Weak" Protein-Protein Interactions from Low to High Concentrations.

So-called "weak" protein-protein interactions are important for the control of solution properties and stability at elevated protein concentrations (c2) but are not practical to capture in atomistic simulations. This report focuses on a series of coarse-grained models for predicting second osmotic virial coefficients (B22) and high-concentration Rayleigh scattering (osmotic compressibility) as a function of c2 for monoclonal antibodies (MAbs) that are of interest in biotechnology. B22 and molecular volume along with c2-dependent osmotic compressibility were calculated for a series of models with increasing structural detail. Models were refined to include contributions from sterics, short-ranged van der Waals and hydrophobic attractions, screened electrostatics, and the flexibility of the mAb hinge region. The results highlight shortcomings for spherical models of MAbs and a useful balance between numerical accuracy and computational burden offered by models based on 6 or 12 spherical, partly overlapping domains. The results provide bounds for realistic values of effective charges on variable domains in order for MAbs to be stable in solution and more generally illustrate semiquantitative bounds for the space of model parameters that can reproduce experimental behavior and provide a basis for future development of computationally efficient and accurate CG mAb models to predict both low- and high-c2 behavior.

Relating Protein-Protein Interactions and Aggregation Rates From Low to High Concentrations.

At low protein concentrations (c2), non-native protein aggregation rates are known to be sensitive to changes in conformational stability and "weak" or "colloidal" protein-protein interactions. Protein-protein interactions are also known to be strong functions of c2. In the present work, protein-protein interactions and rates of aggregation were quantified systematically for a monoclonal antibody (MAb) across a broad range of c2 at pH 5.1 and 6.5, with or without 5 wt/wt % sucrose or 100 mM NaCl present. Aggregation rates were determined from initial-rate analysis with size-exclusion chromatography, and interactions were quantified with static and dynamic laser light scattering. A number of hypotheses were tested regarding whether changes in protein-protein interactions can be predictive of changes in aggregation rates versus c2. Hypotheses were based on (i) changes in thermodynamic activity; (ii) statistical mechanical fluctuation theory; and (iii) surface-contact probabilities. Arguments based on (i) and (ii) were qualitatively inconsistent with experimental rates and scattering. Hypothesis (iii) was reasonably successful and resulted in a semiquantitative correlation between rates and protein-protein interactions across almost 2 orders of magnitude in c2. However, (iii) requires one to assume that the concentration-dependent protein-protein Kirkwood-Buff integral is a reasonable surrogate for contact probabilities.

Predicting unfolding thermodynamics and stable intermediates for alanine-rich helical peptides with the aid of coarse-grained molecular simulation.

This report focuses on the molecular-level processes and thermodynamics of unfolding of a series of helical peptides using a coarse-grained (CG) molecular model. The CG model was refined to capture thermodynamics and structural changes as a function of temperature for a set of published peptide sequences. Circular dichroism spectroscopy (CD) was used to experimentally monitor the temperature-dependent conformational changes and stability of published peptides and new sequences introduced here. The model predictions were quantitatively or semi-quantitatively accurate in all cases. The simulations and CD results showed that, as expected, in most cases the unfolding of helical peptides is well described by a simply 2-state model, and conformational stability increased with increased length of the helices. A notable exception in a 19-residue helix was when two Ala residues were each replaced with Phe. This stabilized a partly unfolded intermediate state via hydrophobic contacts, and also promoted aggregates at higher peptide concentrations.