Nonspecificity fingerprints for clinical-stage antibodies in solution

Significance

Monoclonal antibody drugs successfully target cancer as well as autoimmune and infectious diseases. However, the optimization of target affinity often comes at the cost of increased nonspecific interactions, which can compromise successful application. Here, we address this barrier by using microfluidic technologies to quantify the physicochemical properties and nonspecific interactions of clinical-stage antibodies. Our results show how electrostatics drive binding to proteins and polymers that are found in the bloodstream, resulting in particle formation between charge-complementary antibodies and DNA. We determine the binding affinities for off-target interactions, and we demonstrate that target avidity can increase the apparent affinity by two orders of magnitude. Based on our findings, we propose a quantitative nonspecificity score, which can be employed across medical and biosciences.

Abstract

Monoclonal antibodies (mAbs) have successfully been developed for the treatment of a wide range of diseases. The clinical success of mAbs does not solely rely on optimal potency and safety but also require good biophysical properties to ensure a high developability potential. In particular, nonspecific interactions are a key developability parameter to monitor during discovery and development. Despite an increased focus on the detection of nonspecific interactions, their underlying physicochemical origins remain poorly understood. Here, we employ solution-based microfluidic technologies to characterize a set of clinical-stage mAbs and their interactions with commonly used nonspecificity ligands to generate nonspecificity fingerprints, providing quantitative data on the underlying physical chemistry. Furthermore, the solution-based analysis enables us to measure binding affinities directly, and we evaluate the contribution of avidity in nonspecific binding by mAbs. We find that avidity can increase the apparent affinity by two orders of magnitude. Notably, we find that a subset of these highly developed mAbs show nonspecific electrostatic interactions, even at physiological pH and ionic strength, and that they can form microscale particles with charge-complementary polymers. The group of mAb constructs flagged here for nonspecificity are among the worst performers in independent reports of surface and column-based screens. The solution measurements improve on the state-of-the-art by providing a stand-alone result for individual mAbs without the need to benchmark against cohort data. Based on our findings, we propose a quantitative solution-based nonspecificity score, which can be integrated in the development workflow for biological therapeutics and more widely in protein engineering.

Monoclonal antibodies are prominent among the current best-selling pharmaceuticals, and the 100th unique mAb was approved recently (1). These versatile biological drugs have been optimized to target a wide range of conditions including autoimmune, cancer, and infectious diseases. Features such as high target binding affinity and specificity, biocompatibility, immune effector functions, and desirable pharmacokinetic properties are factors leading to antibodies becoming one of the preferred therapeutic modalities. However, the development process is costly and often unsuccessful in delivering a commercial product (2, 3). Antibody hits identified during discovery are often subjected to affinity maturation campaigns using display-based platforms. Recent studies have demonstrated that these maturation campaigns can compromise mAb development through a trade-off between affinity and specificity (4–7). Off-target interactions are typically driven by suboptimal physicochemical properties, e.g., charged or hydrophobic patches on the mAb surface, which cause nonspecific binding, often defined as weak reversible interactions with molecules and interfaces other than the intended target (8). A propensity for nonspecific interactions can therefore be detected by challenging the mAbs with charged and/or hydrophobic nonspecificity probes (5, 9, 10). The propensity for nonspecific interactions emerges as a critical determinant of developability, i.e., the suitability of antibody drug candidates to succeed as medicines (Fig. 1) (5, 9, 10). Consequently, screens for nonspecificity are rapidly becoming industry standard and are implemented in parallel with functionality measures during the discovery and optimization phases (5, 9–14).

Fig. 1.

Illustration of the challenges faced by mAbs. (Left) High mAb concentration in low complexity solutions yields a high risk of self-association and high viscosity. (Right) Low mAb concentration under crowded conditions in a complex solution. Risk of off-target binding to isolated nonspecificity partners or cell surfaces which can result in limited bioavailability, unwanted side effects, or fast clearance.

Remarkably, in spite of the current focus on screening for nonspecificity as a key developability parameter, the underlying physical chemistry remains largely unexplored (5, 8, 9). Direct measurements of nonspecific binding events in solution have not been within the scope of mAb development campaigns, and traditional assays are poorly suited to address these questions, as they rely on indirect reporters, such as light scattering, and potentially disruptive processes such as surface attachment or for samples to pass through a matrix (5, 12, 15). Here, we address the need for quantitative solution-based nonspecificity assays by applying microfluidic techniques, and we generate nonspecificity fingerprints for a panel of clinical-stage antibodies by challenging them with a range of physiologically relevant ligands (Table 1 and SI Appendix, Table S1) (16–18). Crucially, by evaluating changes to physicochemical parameters, such as the size and charge of antibodies, we detect interactions in a quantitative manner without surface attachment, and we are able to evaluate the binding affinity of these interactions directly (17–21).

Table 1.

mAb panel

Antibody sequence	Name	Status	PSR	CIC	ELISA	Flags
Briakinumab Fv-SIA	Br	Discontinued	X	X	X	8
Atezolizumab Fv-SIA	Az	Approved		X		7
Bococizumab Fv-SIA	Bo	Discontinued	X	X	X	7
Figitumumab Fv-SIA	Fi	Discontinued	X	X	X	7
Lenzilumab Fv-SIA	Le	In clinical trials	X	X	X	7
Sirukumab Fv-SIA	Si	Discontinued	X	X	X	7
Gantenerumab Fv-SIA	Ga	Discontinued	X		X	4
Denosumab Fv-SIA	De	Approved			X	3
Dupilumab Fv-SIA	Du	Approved			X	2
Brentuximab Fv-SIA	Bx	Approved			X	1
Adalimumab Fv-SIA	Ad	Approved				0
Tovetumab Fv-SIA	To	Discontinued				0

Construct information: isotype-matched IgG1 scaffold with Fragment variable–sequence identical analogues (Fv-SIA) for 12 mAbs; abbreviation; reported flags by poly-specificity reagent (PSR); cross-interaction chromatography (CIC); enzyme-linked immunosorbent assay (ELISA); and the total number of flags reported from ten assays (5).

The bivalent antibody format permits ligand binding avidity, which could exacerbate nonspecific interactions in surface-based assays against an immobilized ligand (CIC, ELISA, and similar) (12, 15). The ability to tune the avidity potential is key for developing an accurate understanding of the nonspecificity profile for mAbs and gain a better understanding of the role of avidity in determining mAb affinity and how this effect may translate to in vivo situations, e.g., as a driver of binding to interfaces such as the endothelial cell surface (12). We find that avidity is a key promoter of nonspecific binding, and we evaluate the contribution of avidity potential to the apparent binding affinity ($K_d$) by exposing a mAb to DNA oligomers of decreasing length.

We propose that nonspecificity issues can be identified through changes to the ligand size or charge; this approach yields a score, which can be evaluated as a stand-alone measurement without the need for a cohort dataset as benchmark. Taken together, we establish microfluidics as a platform for the quantitative, in-solution characterization of biologics. Furthermore, we provide crucial insights into the impact of avidity on nonspecific binding and rationalize the physicochemical drivers that promote these interactions, enabling these features to be addressed during the development of targeted proteins and biological drugs.

Results and Discussion

Physicochemical Properties of Monoclonal Antibodies.

A dozen mAbs are selected based on their reported biophysical properties in the study by Jain et al. (5) (Table 1). In particular, we have chosen those with a high number of developability flags in past studies (Br, Az, Bo, Fi, Le, and Si). Only a few flags were raised for Bx, De, Du, and Ga, mainly in surface-based nonspecificity screens, and we were interested in finding whether these nonspecific interactions translated to solution-based assays. Additionally, we include Ad and To as references with good reported biophysical properties, i.e., we expect a low propensity for nonspecific binding in these antibodies (5). In addition, the mAb selection is designed to include sequence charges ranging from the near-neutral Bx to the highly positively charged Le (5).

Microfluidic technology enables us to interrogate key solution properties of mAbs, and we screen the diffusion coefficient ($D$), the corresponding hydrodynamic radius ($R_H$), the electrophoretic mobility (${\mu }_e$), and the effective solution charge ($z_e$) (SI Appendix, Table S2) (16–19, 21). Notably, we screen unlabeled antibodies by imaging intrinsic protein fluorescence (SI Appendix, Fig. S1) (21). This feature allows us to not only measure the nonspecific interactions but also to perform quality control on our antibody and ligand stocks.

Antibody self-association is a common issue that can potentially compromise the reliability of interaction studies, measurements of the mAb $R_H$ can detect this process (13, 22, 23). We screen for nonspecific interactions at a relatively low mAb concentration (1 mg/ml) in order to limit the risk of mAb self-association. No significant increase in mAb size is observed under native solution conditions at physiological salt levels (hs, 150 mM NaCl, 20 mM HEPES-NaOH pH 7.4) or low salt (ls, 15 mM NaCl, 2 mM HEPES-NaOH pH 7.4) (SI Appendix, Table S2). All samples and buffer solutions contain 0.01% v/v Tween20 to prevent surface adhesion within the microfluidic channels.

Avidity Enhances Nonspecific Interactions.

Screens for nonspecific interactions are an integral part of the drug development pipeline (5, 9–11, 13, 14). When administered as drugs, mAbs encounter insulin and cell-free DNA fragments, which can be single or double stranded, in circulation (24–26). A propensity for nonspecific interactions with either biomolecule could thus affect the function of mAbs through competition from off-target interactions (5, 7, 9–11, 13, 14). DNA and insulin are thus two frequently employed nonspecificity probes, and antibodies that interact strongly with these negatively charged biomolecules have been reported to have an increased risk of poor in vivo half-lives (11). Therapeutic mAbs thus need to achieve a balance between typically positive sequence charges to facilitate interactions when targeting membrane proteins and avoidance of nonspecific interactions, e.g., due to charge patches on the mAb surface (10).

We use microfluidic diffusional sizing to record the $R_H$ for mixtures of mAbs and monomeric insulin-CF488 (1 $\mathrm{\mu }$M, icf) or single-stranded DNA oligomers labeled with Cy3 (20-mer, 1 $\mathrm{\mu }$M) at physiological salt levels. Fluorophore-labeling of the nonspecificity ligands enables parallel size measurements for both the unlabeled mAbs and the targets (SI Appendix, Fig. S2). The analyte concentrations in solution are well defined and can be tuned to investigate specific binding affinities, which are not accessible to, e.g., ELISA, where the analyte concentrations are challenging to quantify. In addition, microfluidic technologies are compatible with high-viscosity solutions, and nonspecificity assays could also be conducted at elevated protein concentrations, in serum, or cell lysate (17).

Insulin and DNA provide an opportunity to compare nonspecific binding to nonavidity and avidity targets as insulin is small and monomeric while DNA is a homopolymer, allowing for multiple interactions (Fig. 2 and SI Appendix, Fig. S2). No significant interactions between the antibodies and insulin are observed in solution, whereas a subset of the mAbs bind to DNA oligomers, even at physiological pH and ionic strength. In particular, the observed DNA $R_H$ increases in the presence of Bo, Br, Ga, and Le indicating binding (Table 1 and Fig. 2 A and B). None of the mAbs flagged in the solution-based measurements have received regulatory approval yet.

Fig. 2.

(A) Nonspecific binding assayed in free solution under nonavidity conditions in the high salt buffer between unlabeled mAb (1 mg/ml, 6.7 $\mathrm{\mu }$M) and 1 $\mathrm{\mu }$M monomeric insulin-CF488 (icf, blue) against the reported ELISA signal for each mAb against immobilized insulin (5). Approved mAbs in shades of green, in clinical trials in brown, and discontinued in pink. (B) The observed DNA $R_H$ against the reported ELISA signal for each mAb against ssDNA 20-mer (colors as in A) (5). (C) Diffusional sizing of 100-nM Cy3-labeled DNA oligomers as a function of the concentration of the mAb Le at physiological salt concentrations. Three oligomer lengths 20 (20-mer, red), ten (10-mer, orange), and six bases (6-mer, yellow) were investigated to evaluate the polymer size requirement for avidity effects in nonspecific mAb-DNA interactions. The 20-mer forms higher order complexes with the mAb ($R_{\mathit{complex}}=10.7\mathrm{nm}>R_{\mathit{mAb}}$) with a $K_{dapp}$ of 3.4 $\mathrm{\mu }$M per antigen binding site (assuming two per antibody). When target avidity is removed by reducing the oligomer length to ten or six bases, we observe a dramatic decrease in binding affinity, by approximately two orders of magnitude ($K_{dapp}\approx$ 154 and 239 $\mathrm{\mu }$M, respectively). The antibody $R_H$ was measured in parallel for all sample compositions with Le $\ge$ 640 nM and found to be constant within error (purple shaded areas show average $\pm$ SD for the three datasets). (Right) Schematic illustrations and $K_{dapp}$. Error bars show mean $\pm$ SD for three independent repeats. (D) Screen for particle formation (demixing) as a function of Le:20-mer ratio and NaCl concentration (27). The concentrations of DNA (1.5 $\mathrm{\mu }$M) and buffer (20 mM HEPES pH 7.4) were kept constant. Droplet contents were classed as mixed or demixed by image analysis of the Cy3-20-mer fluorescence.

The microfluidic diffusional sizing results are in good agreement with the ELISA data by Jain et al. (5) where these four mAbs are among the five with the highest ELISA signal against ssDNA (Fig. 2B). The literature ELISA screens were performed against immobilized ligands (5) and avidity effects would result in low effective dissociation rates, thus increasing the apparent interaction affinity and giving rise to false positive readings for otherwise soluble targets. Indeed, in the surface-based assay, most of the mAbs appear to bind insulin and DNA to some extent as seen from the range of ELISA values (Fig. 2 A and B). These intermediate values are challenging to interpret, and this group includes approved mAbs. The standard approach is to flag only, e.g., the highest 10% of readings (5). This grouping will depend on the cohort and not just the individual mAb. In contrast, the solution-based measurement provides a clearly defined yes/no result depending on the value of the $R_H$.

Three of the approved mAbs, Br, De, and Du, were selected for this study because they were flagged as poorly behaved in surface-based assays (ELISA and baculovirus particle) (Table 1) (5). We were interested in whether the reported interactions would also be observed without surface-attachment. In solution, neither ligand showed a significant $R_H$ increase in high salt buffer with these mAbs (Fig. 2 A and B and SI Appendix, Fig. S2).

Le results in the largest DNA size increase (SI Appendix, Fig. S2D). We therefore set out to determine the role of binding avidity in promoting this interaction (Fig. 2C). To this end, we measure the $R_H$ of 100 nM Cy3-labeled DNA oligomers (20-mer, 10-mer, and 6-mer) as a function of Le concentration. Condensation to form micron-scale particles is observed for a larger oligomer (100-mer) (SI Appendix, Fig. S4B). Here, we consider the interactions as Fragment antigen binding (Fab)-DNA binding, as this fragment is varied between the antibodies in our screen, whereas they share an IgG1 Fc. Le has micromolar affinity for the 20-mer ($K_{dapp}$ = 3.4 $\mathrm{\mu }$M per antigen binding site—two per mAb), and clusters were formed ($R_{\mathit{complex}}$ = 10.7 nm $>$$R_{\mathit{Le}}$). The concentration of DNA used here is comparable to those of cell-free dsDNA in plasma (1 to 10 ng/ml in healthy individuals, rising to 100s of ng/ml in cancer patients) (26, 28) the interaction would thus be relevant at a typical mAb $c_{\mathit{max}}$ of 0.1 mg/ml (0.67 $\mathrm{\mu }$M) in plasma (29). This type of nonspecific interaction could thus be highly significant for the function of therapeutic antibodies.

For mAb concentrations of 640 nM ($\approx$0.1 mg/ml) and above, the antibody in the sample mixture is sized via intrinsic protein fluorescence (Fig. 2C). The antibody is in excess at the plateau values, and the complexes are therefore likely to involve multiple mAbs cross-linked by the DNA oligomer. However, the constant $R_{\mathit{mAb}}$ in these samples indicates that DNA binding does not induce aggregation of the whole Le population.

Remarkably, when the polymers are shortened to prevent ligand avidity, the apparent Le affinity is reduced by approximately two orders of magnitude ($K_{\mathit{dapp}}\approx$ 154 and 239 $\mathrm{\mu }$M respectively when $R_{\mathit{complex}}$ was fixed to $R_{\mathit{mAb}}$) (Fig. 2C). As we also observe for monomeric insulin, charge complementarity alone is not sufficient for high-affinity nonspecific interactions at physiological salt concentrations. Avidity therefore emerges as a key factor promoting off-target mAb interactions.

Screening for Nonspecific Interactions.

The microfluidic platform for quantitative biophysical characterization of mAbs provides a versatile measure of nonspecific interactions under a wide range of solution conditions. We create a nonspecificity fingerprint for the mAbs by monitoring the change in $R_H$ for mAbs and ligands, when we challenge the mAbs with molecules that may be encountered in vivo at high salt (150 mM NaCl) and low salt (15 mM NaCl) (Fig. 3).

Fig. 3.

(A) Nonspecificity fingerprints for clinical-stage mAbs in the presence of nonspecificity probes (hs, 150 mM NaCl): negatively charged polymers (1 $\mathrm{\mu }$M ssDNA oligomers and 1 mg/ml heparin), a negatively charged peptide hormone (1 $\mathrm{\mu }$M monomeric insulin), and human serum albumin (2 mg/ml HSA), which carries both a negative charge and has exposed hydrophobic patches (30, 31). Heatmap of the relative mAb $R_H$, ligand labels are shown in purple for data acquired via intrinsic fluorescence and gray for labeled ligands (green heatmap data). Only HSA has significant levels of intrinsic fluorescence at the concentrations used here; the observed $R_H$ is therefore a weighted average of the mAb and HSA contributions ($<$R${}_{\mathit{HmAb}}$). The antibodies are ordered according to the number of reported flags (5) as in Fig. 1B, and labels are color coded by approval status (green = approved, brown = in clinical trials, pink = discontinued). (B) Nonspecificity fingerprints in low salt buffer (ls, 15 mM), same color scheme.

Based on their advanced development stage, the mAbs investigated here would not be expected to show nonspecificity at high salt. We detect nonspecific binding to negatively charged polymers, where the interaction is enhanced by avidity, even in the high salt buffer (Figs. 2 and 3A). A relative $R_H$ of 1 means that no nonspecific binding was detected. The typical error on the diffusional sizing experiment is 10%; for individual nonspecificity ligands, we would therefore consider a relative $R_H\ge 1.21$ between the samples with and without nonspecificity ligand as a sign of interaction (Fig. 3).

When we lower the buffer ionic strength, electrostatic interactions are promoted and nonspecific binding is increased (Fig. 3B). We observe extensive interactions with the negatively charged polymers, e.g., with heparin (Fig. 3B). For DNA, precipitation is observed upon dilution of the buffer, even for mAbs with otherwise good biophysical properties, such as Ad (SI Appendix, Figs. S4 and S5) (5). Le forms particles with the DNA 100-mer even in the high salt buffer and at concentrations considerably below formulation conditions (100 mg/ml) (SI Appendix, Fig. S4B). Intriguingly, we have recently reported liquid–liquid phase separation for nontherapeutic antibodies (32). We therefore expanded our analysis of the Le construct to investigate the propensity for particle formation between Le and the ssDNA 20-mer using an automated microfluidic platform to detect solution demixing as a function of the salt concentration and Le:DNA ratio (Fig. 2D and SI Appendix, Fig. S6) (27). This analysis showed that the Le-DNA mixtures condensate to form particles in a salt-dependent manner (SI Appendix, Fig. S6) and that droplets were demixed even at high salt concentrations. Furthermore, the boundary between mixed and demixed solutions shifted as a function of the Le:DNA ratio, with the samples being less prone to particle formation when the mAb was in a large excess (Fig. 2D). At low Le:DNA ratios, under avidity conditions, networks of mAb-polymer interactions can form, resulting in demixing. In contrast, at high Le:DNA ratios, the formation of soluble complexes is favored. Our findings indicate that liquid–liquid phase separation may occur in clinical antibodies, and this form could be relevant for bioavailability and pharmacokinetics in vivo where charge-complementary polymers such as DNA are encountered in the bloodstream. As the large mAb-DNA particles lead to precipitation, we probed for interactions with DNA at physiological salt levels only when generating the nonspecificity fingerprints (Fig. 3).

At 35 to 50 mg/ml, HSA is the most abundant protein in plasma, and it interacts with a wide range of drug molecules (30–32). Indeed, binding to HSA is promoted, e.g., by drug lipidation as a strategy to extend the half-life of medicines (31). Interestingly, nonspecific mAb binding to HSA increases when the ionic strength is lowered (compare Fig. 3 A and B); this dependence on charge screening by the solvent ions indicates that the interaction is electrostatic, and not exclusively due to the hydrophobic patches on the surface of HSA.

Denosumab is marketed as an IgG2 rather than IgG1 antibody; SI Appendix (33). The different isotype used in this study (De; Table 1) may be less stable than the antibody drug, resulting in poorer biophysical properties and nonspecificity behavior. This antibody is flagged for interactions with nonspecificity ligands in the ls buffer. In addition, $R_{\mathit{Hrel}}$ = 1.2 for the mAb against insulin, this value is not matched by an increase in ligand size ($R_{\mathit{Hrel}}$ = 1.1), and we therefore attribute the higher $R_{\mathit{mAb}}$ to antibody-only interactions (Fig. 3A).

We note that four of the mAbs were consistently flagged as prone to nonspecific interactions (Bo, Br, Ga, and Le) (Fig. 3). Notably, none of the approved mAbs are flagged; we approximate these as false positives in our nonspecificity assay, although even approved mAbs could have compromised pharmacokinetics (SI Appendix, Fig. S6). In order to identify accurately mAbs with a propensity for nonspecific interactions, we therefore propose the use of the average relative $R_H$ for two or more ligands as a measure of nonspecificity (SI Appendix, Fig. S7B).

Charge Property Evaluation.

The charge properties of mAbs are important for target recognition, potency, pharmacokinetics, formulation, and as a quality control measure, e.g., for posttranslational modifications (34, 35). Current strategies for assessing these use column chromatography to fractionate samples followed by mass spectrometry or capillary electrophoresis to assess the sample charge (34, 35). Mass spectrometry in particular has limited buffer system compatibility, and column separation techniques may require nonphysiological conditions, such as low pH, to operate optimally (35). The microfluidic platform employed here is compatible with common biological buffer systems and can be applied to determine $z_e$ under native solution conditions (Fig. 4A) (19). While the free-flow electrophoresis step is preferentially performed in low ionic strength buffers for optimum separation, microfluidics is a highly modular technology, and a desalting element can be included upstream of the electrophoresis channel (36).

Fig. 4.

(A) Microfluidic free-flow electrophoresis design. Solid electrodes (gray) are incorporated along the main channel as previously described (16). An electric field is applied perpendicularly to the direction of flow, and sample molecules migrate according to their electrophoretic mobility (${\mu }_e$). (B) Heat map of ${\mu }_e$ for the mAbs (1 mg/ml) in isolation and in the presence of nonspecificity targets heparin (1 mg/ml), monomeric insulin-CF488 (icf, 1 $\mathrm{\mu }$M), and human serum albumin (HSA, 2 mg/ml), all in low salt buffer (red for increasing negative charge, and blue for positive ${\mu }_e$). Ligand labels for data acquired via intrinsic fluorescence are in purple, and data for the labeled ligand are in gray.

We have selected the antibodies in this study to have a wide distribution of sequence charges, from the near-neutral Br to the highly positively charged Le (5). The ${\mu }_e$ reports on the sample size to charge ratio and is therefore a useful parameter for detecting interactions where the complex $M_w$ is similar to that of the observed protein and the ligand is unlabeled, for instance mAb versus mAb (150 kDa) $+$ heparin (8 to 25 kDa). The mAbs bind insulin in low salt buffer, and it is therefore a useful nonspecificity probe under these conditions (Figs. 3B and 4B). We measure ${\mu }_e$ for individual mAbs and mixtures with nonspecificity ligands in low salt buffer (SI Appendix, Table S2 and Fig. S8).

While there is little interaction with HSA in hs buffer, both electrophoresis and diffusional sizing show binding between HSA and the majority of mAbs under ls conditions (Figs. 3B and 4B). The mAbs interact extensively with heparin at low ionic strengths. The change in $M_w$ would be small for mAb binding to a single polymer; however, $R_H$ measurements reveal that clusters are formed (Figs. 3B and 4B). The increase in charge complementary interactions at low ionic strength indicates that these are favored by electrostatics.

We calculate $z_e$ based on $R_H$ and ${\mu }_e$ and find that while $z_e$ increases with sequence charge, the absolute value is lower (SI Appendix, Fig. S7) (19). Charge-screening through interactions with solvent molecules and within the protein structure can contribute to the lower solution charge (37). Additionally, charged residues may be buried within the immunoglobulin folds, form salt bridges, or have suppressed pK${}_{\mathrm{A}}$ values, e.g., due to clustering of charged residues. The antibodies are all expressed on the same IgG1 scaffold, so differences in overall charge originate mainly in the variable regions and the choice of $\kappa$ or $\lambda$ light chains.

Four of the five mAbs with the highest $z_e$ and sequence charges are flagged as prone to nonspecific interactions (Fig. 3 and SI Appendix, Fig. S7). High levels of positive charge have been associated with faster clearance rates in vivo; there is thus a trade-off between antibody potency and pharmacokinetics (38). It may therefore be desirable to avoid or modulate excessive positive charges during the development process (SI Appendix, Fig. S7 and Fig. 5). However, while the most charged mAbs (Bo, To, Ga) have not been successful in clinic, mAbs with lower $z_e$ have also failed (e.g., Si). As may be expected, a good charge profile is required, but not sufficient for success in therapeutic antibodies. Notably, antibodies with low (Br) and moderately high (Ad, De, Du) $z_e$ have received approval as drugs. This observation indicates that not just the overall charge, but also the distribution in the antibody structure can play a key role in mAb developability. Indeed, the presence of charged patches on the mAb surface has been highlighted as a driver of nonspecificity (10).

Fig. 5.

(A) Matrix of Pearson correlation coefficients for the microfluidic screening assays employed in this study. The results are grouped by the observed species (mAb in purple or ligand in gray), the ionic strength (high salt and low salt, ranges shown below), and the measured parameter ($R_H$ or ${\mu }_e$ of the total sample). The ligand type (polymer or isolated protein) is also shown. The physicochemical parameter interrogated for each sample is also shown. (B) A plot of $R_H$ for a ligand with avidity at high salt (ssDNA 20-mer) against a nonavidity ligand at low salt (monomeric insulin). Four mAbs emerge as nonspecific binders against both ligands (Bo, Br, Ga, Le). (C, Left) Ranking of the mAbs by an average of the relative $R_H$ of two nonspecificity probes in this study (20-mer, hs and insulin, ls); a score of 1 means that no interaction was detected. Right column, Number of flags from results in the worst 10% of the panel of 137 mAbs in the literature (5).

The effective charge and the $\zeta$-potential can be used to assess colloidal stability and overall physicochemical suitability, with highly charged mAbs are less likely to be successfully developed (Fig. 5A). However, for systems, such as proteins with a heterogeneous surface charge distribution, a more detailed analysis of the charge distribution is needed to locate potentially problematic charged patches and to assess the solution stability. We note that the mAbs investigated here have $\zeta$-potentials in the range of 0 to 17 mV, which would render them unstable or only marginally stable in solution, if the charges were evenly distributed on the surface (SI Appendix, Fig. 8B).

Improving Screening Campaigns.

Our investigations of the physicochemical properties of antibodies and their nonspecific interactions under native solution conditions highlight the need to amend current screening approaches with tools such as the microfluidic platform presented here to obtain accurate and quantitative readouts. We plot the Pearson correlation coefficients for the data collected in this study (Fig. 5A). The antibody ${\mu }_e$ and $z_e$ are highly correlated due to the similar $R_H$ of the mAbs in this study. The mAb charge correlates with $\mathrm{\Delta }{\mu }_e$ for heparin, HSA, and insulin interactions with the mAbs (Fig. 4B).

Monitoring the properties of the nonspecificity probe directly increases the sensitivity of the assay, as the mAb $R_H$/${\mu }_e$ dominate in most mAb-ligand complexes. A subset of the mAbs (Bo, Br, Ga, Le) emerge as prone to nonspecific interactions (Figs. 4 and 5B). We therefore suggest verifying nonspecific binders across two or more targets in solution as a standalone measure of mAb nonspecificity, e.g., by recording the average relative ligand ratio ($R_{\mathit{Hrelavg}}$) (Fig. 5C). Clinical status is often used as an approximation of true/false positives when evaluating developability screening strategies (5). A good nonspecificity profile is one of the key requirements for successful therapeutics. However, several other factors influence the decision on whether to proceed with a drug, and antibodies with perfectly acceptable physicochemical profiles may not progress, e.g., due to the lack of a business case. Discontinuation is therefore not a comprehensive marker poor biophysical properties. Approved mAbs would not be expected to show significant off-target interactions, and we therefore use approved drugs as an approximation for false positives in our assessment of nonspecificity. We evaluate the $R_{\mathit{Hrelavg}}$ threshold required to avoid false positives, and our analysis shows that none of the approved mAbs are flagged when $R_{\mathit{Hrelavg}}\ge$1.2 (SI Appendix, Fig. S7B). We therefore apply this value as a threshold, which can be further optimized as more data become available.

Potentially problematic biophysical properties are often flagged as poor performance relative to a cohort, requiring large datasets to define the threshold values (Table 1 and Fig. 5C) (5, 39). Quantitative solution-based assays enable us to analyze the nonspecificity behavior of individual antibodies in absolute terms, without using a cohort dataset as benchmark for relative performance. Two of the mAbs that were flagged for surface-based interactions (Fi and Si) did not show nonspecific binding in our assay (Table 1 and Fig. 5). When we examine the reported ELISA and BVP values, we find that these were only slightly above the worst 10% threshold (2.89 for Fi, 2.17 for Si, threshold = 1.9 for ELISA; 5.65 and 9.66 respectively with threshold = 4.3 for BVP) (5). Flagging a fixed percentage of a cohort can be useful when comparing data across a panel of assays (5, 39). However, this approach groups mAbs without weighting how far above the threshold they were, and the result can be biased by the cohort composition, which will influence the outcome for individual molecules.

We compare our approach to literature reports, and we find that the mAbs we flag also have poor results in surface- and column-based screens (Fig. 2 A and B and SI Appendix, Fig. S3) (5, 12). In a further validation of our approach, we compare our data with the results reported by Kraft et al. for their column-based heparin interaction assay (SI Appendix, Fig. S3) (12). We find that they measured the highest relative heparin retention times for the same four mAbs, which we highlight here as prone to nonspecific interactions. In a hybrid approach, ELISA or another high-throughput screening method could therefore be used to prescreen mAbs to identify constructs for detailed investigation. Az, which is used in cancer treatment, but was flagged as problematic in the literature, performed well in the solution-based measurements (5, 23). The ability to evaluate individual mAbs can be particularly useful when assessing the effect of variations in an otherwise similar group of antibodies, e.g., variant selection following mutagenesis. The microfluidic platform allows different molecular formats to be accessed (e.g., Fabs, nanobodies) without the need for further development.

Off-target nonspecific interactions are one of the potential barriers to mAb developability. This factor is considered together with criteria such as efficacy, cost-benefit, self-association, and specific off-target binding, when deciding whether to advance or discontinue a drug. Critically, we did not expect all discontinued mAbs to have poor nonspecificity profiles, and indeed, three discontinued mAbs gave good results in our nonspecificity screens but could have been discontinued due to other factors (40, 41).

Conclusions

Microfluidic solution-based measurements are used to generate a multiparameter, multicomponent nonspecificity fingerprint for a panel of clinical-stage antibodies, paving the way for standardized comprehensive physicochemical characterization of proteins and antibodies. This approach can identify the physicochemical parameters that drive nonspecific interactions, providing a route toward addressing undesirable physicochemical properties, for instance, before committing to the resource-intensive later stages of drug development.

The microfluidic strategy outlined here provides a well-defined readout and baseline conditions, reducing the risk of false positives compared to surface-based assays such as ELISA. We demonstrate the importance of avidity in enhancing nonspecific interactions as an inherent property of the target (e.g., polymers) and experimental setup (e.g., surface immobilization of the probe). By recording a quantitative score with a physical significance based on direct interaction measurements, we find that a subset of the clinical-stage mAbs investigated here are prone to nonspecific interactions.

The solution properties of monoclonal antibodies are key to their success as therapeutics, and their propensity for nonspecific interactions is emerging as a central determinant of their developability potential. Quantitative assessments of these parameters provide a valuable addition, not only to the protein drug development pipeline, but also to the wider protein engineering community, and add to our basic knowledge of the physical chemistry governing protein interaction specificity. We envision that the approach outlined here can provide a valuable addition to the state-of-the-art for biophysical chemistry screening workflows.

Materials and Methods

Microfluidic Measurements.

Microfluidic devices were prepared using standard soft lithography methods (42). Microfluidic diffusional sizing, electrophoresis, and PhaseScan measurements were performed as described previously (16, 17, 19, 20, 27, 43). See SI Appendix for extended methods for microfluidic device preparation, diffusional sizing, free-flow electrophoresis, and PhaseScan.

Sample Preparation.

HSA and heparin were purchased (A3782 Sigma Aldrich and A3004,001 from Applichem). DNA oligomers were purchased with a 5’ Cy3 modification (100, 20 mer: Merck, Germany, HPLC purified and lyophilized; 10, 6-mer: Biomers, Germany, HPLC purified and lyophilized); see SI Appendix for sequences. Monomeric insulin was labeled in-house; see SI Appendix. Buffer components were purchased from Sigma.

The primary sequence of the clinical-stage antibodies in IgG1 subclass format where taken from the paper by Jain et al. (5) The mAbs were expressed and purified according to standard protocols; see SI Appendix.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Data, Materials, and Software Availability

Python code for diffusional sizing data fits available at https://zenodo.org/record/3881940#.YRGzCSV4UlQ (43). All other data are included in the article and/or SI Appendix.