Surface patches induce nonspecific binding and phase separation of antibodies

Significance

The susceptibility of antibodies to bind unwanted off-targets is a key issue in the development of therapeutic antibodies. Although the mechanisms have yet to be resolved, such unwanted interactions are linked to aberrant assembly processes, which can impact storage and administration as well as the potency of antibodies. In our work, we quantify these nonspecific interactions to correlate them with the protein surface properties, and link nonspecific off-target interactions to the propensity of antibodies to undergo heteromolecular phase separation. We show that both phenomena are governed by the nature and size of surface patches and demonstrate that modulations in surface patches can vastly change nonspecific binding as well as macroscopic behavior as manifested by phase separation.

Abstract

Nonspecific interactions are a key challenge in the successful development of therapeutic antibodies. The tendency for nonspecific binding of antibodies is often difficult to reduce by rational design, and instead, it is necessary to rely on comprehensive screening campaigns. To address this issue, we performed a systematic analysis of the impact of surface patch properties on antibody nonspecificity using a designer antibody library as a model system and single-stranded DNA as a nonspecificity ligand. Using an in-solution microfluidic approach, we find that the antibodies tested bind to single-stranded DNA with affinities as high as K_D = 1 µM. We show that DNA binding is driven primarily by a hydrophobic patch in the complementarity-determining regions. By quantifying the surface patches across the library, the nonspecific binding affinity is shown to correlate with a trade-off between the hydrophobic and total charged patch areas. Moreover, we show that a change in formulation conditions at low ionic strengths leads to DNA-induced antibody phase separation as a manifestation of nonspecific binding at low micromolar antibody concentrations. We highlight that phase separation is driven by a cooperative electrostatic network assembly mechanism of antibodies with DNA, which correlates with a balance between positive and negative charged patches. Importantly, our study demonstrates that both nonspecific binding and phase separation are controlled by the size of the surface patches. Taken together, these findings highlight the importance of surface patches and their role in conferring antibody nonspecificity and its macroscopic manifestation in phase separation.

Antibodies have emerged as highly potent drug molecules, particularly due to their high specificity, long in vivo half-life, and ability to induce immunogenic effector functions (1, 2). As of 2021, more than one hundred therapeutic antibodies for the treatment of various human diseases, including cancer (3–5), Alzheimer’s disease (6), osteoporosis (7, 8), and HIV (9), have been approved by the US Food and Drug Administration (10). Despite the enormous success of antibody-based pharmaceuticals, only 21% of candidates admitted to clinical trials have been approved (11). This adds to significant developmental costs of 1.4 billion dollars per approved compound (12). The primary reasons for these failures are likely due to unsatisfactory clinical readouts and commercial considerations. Interestingly, however, it has been reported that antibodies in the market possess a lower tendency for nonspecificity compared to candidates that fail in clinical trial Phase 2 or 3 (13), suggesting that unsuccessful clinical trials can be associated with undesired nonspecific binding.

Indeed, target specificity is often hard to achieve in the development of therapeutic antibodies, and as a result, antibodies that exhibit nonspecific off-target binding are common outcomes (Fig. 1)(14). This is because target specificity is not an inherent property of antibodies in particular (15–21) or of proteins in general. A tendency for nonspecific binding is, in fact, known to limit the size of the proteome (22–24), govern functionality of protein interaction networks (25), and shape the intricacies of forming immune responses (26, 27). In vivo, this is resolved by negative selection processes, which enable simultaneous optimization of specificity with other factors such as affinity (25–28). Replication of the in vivo maturation and negative selection process in vitro is often difficult to achieve with similar precision due to a trade-off between affinity and specificity (18, 29). Nonetheless, obtaining highly specific antibodies is crucial, not least because several recent reports have demonstrated how insufficient specificity often manifests in critical development issues (Fig. 1B) such as decreased therapeutic effects due to suboptimal pharmacokinetics, fast in vivo clearance rates (30–36), or even in vivo toxicity (37). Moreover, nonspecificity has also commonly been linked to physical stability issues related to bioprocessing and formulation such as high solution viscosity, low solubility, and phase separation (36, 38–40). Hence, specificity is not only the strength of antibodies, but in many ways, it can be their weakness and a key hindrance to their developability (i.e., the overall suitability for successful development into a therapeutic).

Fig. 1.

Quantification of nonspecific binding in a rationally designed antibody library using an in-solution microfluidic approach. (A) Nonspecificity is defined as the binding of antibodies to unwanted target binding partners. (B) This commonly leads to problematic manifestations in important development issues such as self-association and undesired in vivo pharmacokinetics. Individual case studies have highlighted surface patch properties (i.e., the extent of groupings of amino acids with similar physicochemical properties on the protein surface) as key contributors to nonspecificity. (C) In our study, we perform a systematic assessment of the impact of surface patch properties on nonspecific binding using the humanized anti-trinitrophenyl (HzATNP) library. The molecular model shown displays the residues which are mutated within the library, which was generated using CamSol to present changes in the surface patch properties, which we quantify here using the Molecular Operating Environment (MOE). This allows calculation of surface patch areas of hydrophobicity (Φ_H, green), positive charge (Φ_P, blue), and negative charge (Φ_N, red). (D) Nonspecific binding is probed against single-stranded DNA oligos and monitored using microfluidic diffusional sizing (MDS). MDS measures the diffusion profiles of labeled molecular species in a microfluidic chip to determine their hydrodynamic radii. Complex formation as a result of nonspecific binding yields an increase in the observed hydrodynamic radius (R_h) allowing for quantitative solution phase study of nonspecific antibody–DNA binding.

Despite being recognized as a vital developmental issue for the generation of therapeutic antibodies (13, 41), the physicochemical origins of nonspecificity have not been addressed sufficiently. As such, it has remained difficult to reduce the tendency for nonspecific binding in antibodies as well as other proteins. Furthermore, it is unclear to date how formulation and environmental conditions impact nonspecificity behavior and its problematic manifestations. In recent years, case studies investigating the importance of surface patches (14, 33, 34, 36, 39 (i.e., groupings of amino acids with similar physicochemical properties on the protein surface) have shown high potential in addressing nonspecificity behavior. One example is the work of Dobson et al. (36) who studied the role of a hydrophobic amino acid patch on the protein surface of the MEDI1912 antibody. This surface patch leads to extensive nonspecific binding which translates into critical development issues like high in vivo clearance rates and high solution viscosity, all of which could be removed by mutational disruption of a surface patch. Indeed, the importance of modulating protein surface properties is reported as a powerful strategy not only in the context of antibodies but also for other proteins, for example, in controlling their cellular motion (43). However, detailed studies on how the surface patch properties of antibodies translate into nonspecific binding and other macroscopic manifestations are yet to be investigated more systematically using antibody libraries.

Moreover, insights into the molecular origins of nonspecificity have been hampered by methodological shortcomings. To date, the study of nonspecific binding usually relies on determining interactions through surface immobilization techniques or indirect measurements of physicochemical properties. These approaches, however, fail to accurately represent binding events under native conditions (i.e., directly in solution without surface and avidity effects), which—for weak, low affinity interactions characteristic of nonspecificity—can yield substantial changes in the perceived properties (44). In addition, conventional approaches do not allow for an accurate quantification of the interactions and determination of binding affinities. This makes it difficult to correlate nonspecific interactions with observations of in vivo behavior and to use nonspecific binding readouts for prediction assays based on protein physicochemical features. Hence, approaches able to probe nonspecificity in solution while providing accurate quantification of interaction affinities are highly desirable.

Here, we provide insights into the molecular origins of antibody nonspecificity by examining nonspecific binding in a rationally designed antibody library using an in-solution microfluidic approach. Specifically, we study the humanized anti-trinitrophenyl (HzATNP) antibody library, designed with CamSol, to represent systematic variations of the physicochemical surface patch properties (38, 45) (Fig. 1C) and quantify the patch sizes using the molecular dynamics sampling tool within the Molecular Operating Environment (MOE). We assess the impact of the changes between variants on the nonspecific binding to single-stranded DNA, a common nonspecificity ligand which is a valuable indicator for the overall propensity of nonspecific interactions also in vivo (35, 41, 46). By establishing microfluidic diffusional sizing (MDS) (47, 48) (Fig. 1D) as a tool for probing antibody nonspecificity, we determine nonspecific binding affinities as tight as K_D = 1 µM. The systematic surface property assessment highlights that nonspecific binding of HzATNP antibodies to DNA at physiological salt conditions is driven primarily by a well-defined hydrophobic patch in the complementarity-determining regions (CDRs) around residue V99 in the light chain (LC). Reduction of the hydrophobic patch area by charged mutations reduces the propensity for nonspecific binding by decreasing DNA affinities by multiple orders of magnitude. Furthermore, investigations under lowered salt conditions lead to the observation of DNA-induced phase separation and nanocluster formation. We further characterized the transition from the individual molecule binding to the phase-separated state by mapping the phase space using PhaseScan, a high-throughput droplet microfluidic-based approach (49) and combining it with detailed analysis of binding events. Finally, we link the observed nonspecific binding and DNA-induced assembly across the library to the surface patch sizes to identify generic property profiles of protein mutants that display a tendency for adverse nonspecific binding and assembly.

Results

Probing the Influence of Surface Patch Properties on Nonspecific Binding in the HzATNP Antibody Library by MDS.

The HzATNP library consists of a total of 17 antibodies designed around the humanized anti-trinitrophenyl wild-type (WT) antibody to represent a variety of protein solubilities. The library was generated by Wolf-Perez et al. (38, 45) using a computational design algorithm in CamSol, which predicts solubility based on the extent and the frequency of physicochemical surface property hot spots. The library consists of variants with charged and/or hydrophobic mutations that modulate the protein surface patch properties by disrupting or enhancing aforementioned hot spots. In our study, we investigated a subset of the library consisting of the WT antibody as well as variants 1, 2, 4, 6, 10, 13, 15, and 16 (Table 1). This subset of nine antibodies presents a total of 16 different surface patch property mutations distributed across the fragment antigen-binding (Fab) domain of the antibody and therefore effectively covers the physicochemical space of the original library of 16 variants with only one position not being sampled in this smaller subset. Furthermore, a primary objective of our investigations was to identify the impact and role of the individual mutations that are causative for the changes in behavior of the individual protein variants.

Table 1.

Summary of mutations and patch areas presented in the HzATNP antibody library

	Mutations		Φ H		Φ P		Φ N
Variant	HC	LC	Fab	CDR	Fab	CDR	Fab	CDR
1	T68G, Q108E	T5D, V99R	657	272	1,227	238	877	105
2	T68G, Q108E	V99R	665	278	1,213	250	864	109
4	T68A	V99K	678	304	1,244	250	773	84
6	S70E	V99D	699	308	1,162	194	843	146
WT	–	–	720	336	1,177	199	751	92
10	–	K193V	749	350	1,149	197	785	93
13	E16V, K120V	–	792	338	1,159	202	90	90
15	E16V, K120F	K193Y	886	340	1,126	196	756	90
16	E16V, D72F, K120W	K193F	1,003	334	1,133	192	728	93

Mutations in the HzATNP library around the wild-type (WT) antibody in the heavy-chain (HC) and light-chain (LC) protein domains and resulting hydrophobic (Φ_H), positive charged (Φ_P), and negative charged (Φ_N) patch surface areas (all in Ų) for the entire Fab and only complementarity-determining regions (CDR)-associated residues as determined using the MOE (46).

To complement analysis by CamSol, we further quantified changes by different mutations on the surface patch properties using the molecular dynamics sampling tool within the MOE, which allows for the calculation of hydrophobic (Φ_H), positive charged (Φ_P), and negative charged (Φ_N) surface patch areas across the entire Fab or the CDRs (50, 51). Thereby, total charged patch areas (Φ_ΣPN = Φ_P + Φ_N) and charged patch area difference (Φ_ΔPN = Φ_P–Φ_N) can be quantified. ΔΦ denotes differences in surface patch properties of protein variants compared to WT, unless explicitly stated otherwise. The surface area values for each variant are listed in Table 1. Here, variants 1, 2, 4, and 6 display an increase in the total charged patch area in the Fab region as compared to the WT antibody (ΔΦ_ΣPN^{Fab, v.X} > 0), hence further referred to as charged variants, whereas variants 10, 13, 15, and 16 show an increase in the hydrophobic patch area (ΔΦ_H^{Fab, v.X} > 0), hence further referred to as hydrophobic variants. Taken together, the antibody library represents rationally selected, systematic changes in the surface patch properties, making it an ideal model system to study nonspecific binding.

As a nonspecificity ligand, we used single-stranded DNA, which is commonly seen as a powerful proxy for relevant nonspecific interactions in vivo (35, 41). Accordingly, DNA can be encountered in vivo at high concentrations, especially in several disease contexts such as cancer (52). Furthermore, the nature of antibodies as bivalent constructs allows for a potential increase in the observed binding affinity for ligands that present multiple binding sites. This is particularly important because avidity-driven interactions to negatively charged macromolecular polymers are a potential key driver of increased clearance rates in vivo (14, 53, 54).

To measure the interactions between antibodies and DNA, we adopted the in-solution microfluidic-based approach MDS (47) to acquire diffusion profiles of molecular species within a microfluidic channel. From the broadening of the diffused species in space and time, the diffusion coefficient D can be extracted, and the corresponding hydrodynamic radii R_h can be calculated using considerations of the advection–diffusion process and the Stokes–Einstein relation (47, 48). Binding events can be observed upon complexation of binding partner to the labeled species as it leads to an apparent increase in the observed hydrodynamic radius. Using MDS, we measured the hydrodynamic radii of Cy3-labeled single-stranded DNA (50-mer of random sequence, see the Materials and Methods section in SI Appendix) at a constant concentration of 1 µM in the presence of an excess of the individual antibody variant (6.7 µM) (Fig. 2). The excess in antibody concentration was chosen to ensure binding saturation or, at least, to observe a significant size increase for the DNA oligomer. Importantly, therapeutic antibodies are commonly administered to patients at concentrations which are orders of magnitude higher than the ones used here (e.g., 150 mg/mL corresponds to approximately 1 mM), making these interactions highly physiologically relevant and a source of potential problematic behavior even in vivo. Interactions were first probed under physiological salt conditions [20 mM HEPES buffer, pH = 7.4, 150 mM NaCl supplemented with 0.01% (v/v) Tween]. We also performed experiments, in which we varied the antibody concentration for selected variants over several orders of magnitude to obtain binding curves and determine the apparent nonspecific binding affinity of the antibody–DNA interaction (Fig. 3).

Fig. 2.

Hydrodynamic radius changes of labeled DNA oligos in complexation with HzATNP library antibodies. (A) Single-stranded DNA 50-mer labeled with Cy3 is subjected to binding with the various HzATNP antibody library variants. (B) DNA hydrodynamic radius (R_h) recorded via microfluidic diffusional sizing (MDS) concentration alone and in the presence of HzATNP antibody library variants under physiological salt [purple bars, 20 mM HEPES buffer, pH = 7.4; 150 mM NaCl, with 0.01% (v/v) Tween] and low salt conditions [blue bars, 2 mM HEPES buffer, pH = 7.4, 15 mM NaCl, with 0.01% (v/v) Tween]. DNA and antibody concentrations were 1 µM and 6.7 µM, respectively. Variants are categorized into hydrophobic (ΔΦ_H^Fab > 0) and charged variants (ΔΦ_ΣPN^Fab > 0).

Fig. 3.

Nonspecificity affinities of HzATNP antibody variants against DNA and investigation of the hydrophobic surface patch properties. (A) Changes in the normalized hydrodynamic radius of DNA at 1 µM with varying WT antibody concentrations at physiological salt conditions [20 mM HEPES buffer, pH = 7.4, 150 mM NaCl supplemented with 0.01% (v/v) Tween]. Nonnormalized R_h values range from ≈ 3.5 to 7 nm. Errors represent SDs from triplicate measurements. At high excesses of antibody concentration, binding site saturation on the DNA is observed, hence plateauing of the hydrodynamic radii. (B) Comparison of the dissociation constant K_D of the interaction of different variants to DNA. (C) Molecular model of the protein surface showing a large hydrophobic patch in the complementarity-determining regions (CDRs) (Φ_H^CDR = 336 Ų) centered around the V99 residue common to variants that display a decreased affinity. (D) Affinity comparison of the protein variants containing only the single point mutations present in v. 6 (S70E, V99D) and v. 4 (T68A, V99K) confirming that the V99 mutations are responsible for the decrease in affinity. (E) Close-up of the hydrophobic patch for both V99D (Left) and V99K (Right) highlighting a disruption and decrease in the patch. (F) Mapping the CDR hydrophobic versus charged patch sizes for all variants. Points represent individual protein variants which are colored for their nonspecific DNA binding affinity (blue ~ 1 µM, yellow ~ 100 µM). High-affinity variants display large excesses of hydrophobic patches, whereas affinities decrease with mutations to charged patches.

CDR-Located Hydrophobic Patch Controls the Nonspecific Binding Affinity to Single-Stranded DNA under Physiological Salt Conditions.

First, we probed potential complexation of the HzATNP WT antibody with DNA, and upon mixing, an increase in the DNA hydrodynamic radius from R_h^DNA ≈ 3.6 nm for free DNA to R_h^DNA+WT ≈ 7 nm was observed. We further examined the binding affinity of the nonspecific interactions between the WT antibody and DNA (Fig. 3A) by recording a binding curve (Fig. 3B) and obtained a K_D value in the micromolar regime (K_D^v.WT = 1.8 ± 0.7 µM, effective stoichiometry 2.7:1 antibodies per DNA molecule; see SI Appendix for stoichiometry calculation and SI Appendix, Fig. S1).

Next, we probed the DNA binding of the charged variants (i.e., variants 1, 2, 4, and 6). These exhibited a reduced propensity to form complexes with DNA compared to the WT antibody (Fig. 2B). At 1 µM DNA and 6.7 µM antibody concentration, only variant 6 exhibited binding to DNA (R_h^DNA+v.6 = 5.0 ± 0.5 nm); however, this is a smaller complex size compared to the WT antibody, which corresponds to an ∼60% DNA binding site saturation. Variants 1, 2, and 4 displayed no binding. The decrease in binding appears to be independent of the overall net charge change introduced by the mutations (z_net^v.1 = –1, z_net^v.2 = 0, z_net^v.4 = +1, and z_net^v.6 = –2). Hence, the observed nonspecificity is very unlikely to be primarily electrostatically driven, which is not unexpected given the high concentrations of salt ions present under physiological conditions because of increased Debye screening. Analysis of the binding affinities highlights that variants which carry mutations at the V99 position display a decrease in affinity by at least an order of magnitude compared to the WT antibody (Fig. 3B). Further investigations of the molecular structure using the MOE highlight that V99 sits in the center of a large hydrophobic patch located in the CDRs of the antibody, which is decreased for all the charged variants tested (Table 1). This indicates that the binding to the single-stranded DNA is primarily driven by hydrophobic interactions to exposed nucleobases presented by the largely unstructured oligo.

To further confirm that the hydrophobic patch is key in establishing the observed nonspecific binding to DNA, we created single point mutation variants from variants 4 and 6 (Table 2). The S70E (K_D^v.S70E = 1.9 ± 0.8 µM) and T68A (K_D^v.T68A = 1.1 ± 0.5 µM) mutants display an almost identical binding affinity compared to the WT antibody (K_D^v.WT = 1.8 ± 0.7 µM), as shown in Fig. 3D. In comparison, both the V99D (K_D^v.V99D = 14.5 ± 7 µM) and V99K (K_D^v.V99K = 140 ± 28 µM) single-mutant variants show a largely decreased binding affinity. This aligns with a significant decrease in hydrophobic patch size compared to the WT (ΔΦ_H^{CDR, V99D} = −50 Ų, ΔΦ_H^{CDR, V99K} = −39 Ų).

Table 2.

Impact of single point mutations on surface patches and DNA binding affinity

Mutations	Variant	KD/µM	Φ H CDR	Φ P CDR	Φ N CDR
–	WT	1.8 ± 0.7	336	199	92
T68A	v. 4, HC	1.1 ± 0.5	357	187	88
V99K	v. 4, LC	140 ± 28	297	255	90
T68A, V99K	v. 4	190 ± 30	304	250	84
S70E	v. 6, HC	1.9 ± 0.8	348	205	109
V99D	v. 6, LC	15 ± 7	286	202	125
S70E, V99D	v. 6	15 ± 6	308	194	146

Investigated single point mutation protein variants and measured DNA nonspecific binding affinity K_D (in µM) and CDR-located surface patches (Φ_X^CDRall in Ų).

Interestingly, V99K displays a larger decrease in the binding affinity by about an order of magnitude despite the introduction of a positive charge complementary to the overall negatively charged DNA. Molecular models of the hydrophobic patch, however, highlight that the longer aliphatic side chain of the lysine residue appears to nestle into the hydrophobic residues along the patch. This allows for V99K to cut through the patch (Fig. 3 E, Right), whereas V99D only decreases the patch size unilaterally (Fig. 3 E, Left). Furthermore, there is a lysine residue in close proximity to V99 which, upon mutation of V99 to another positive charged residue, gives a small local positive charged patch (ΔΦ_P^{CDR, V99K} = 56 Ų). This small “counter patch” could likely allow for more effective solvation and shielding of the hydrophobic region making the water exclusion in the interface much less favorable and, thereby, decreases the affinity to DNA.

Next, we probed the interaction of DNA with variants 10, 13, 15, and 16, all of which display a larger hydrophobic patch area in the entire Fab region compared to the WT antibody (ΔΦ_H^Fab > 0, up to +280 Ų). These variants, however, did not exhibit a significant change in behavior compared to the WT antibody as they all displayed saturation at 6.7 µM antibody concentration (Fig. 2B). We further quantified the binding affinities (Fig. 3B) and found that there is no apparent difference in DNA nonspecific binding between the WT antibody and other hydrophobic variants. Interestingly the V99 containing hydrophobic patch is largely associated with CDR residues. Indeed, when evaluating the hydrophobic surface patch properties only across CDR-associated residues, the hydrophobic variants show minimal differences compared to the WT antibody (ΔΦ_H^CDR < 15 Ų). This suggests that although the overall protein surface properties are amended in these mutations (E16V, K120V/F/W, and K193V/F/W), they do not affect the nonspecific binding paratope. The molecular model of the WT antibody as derived by the MOE further shows that all of the additional mutations are not in proximity to the V99-associated hydrophobic patch (Fig. 1C) and, hence, do not increase the possible interaction surface. These observations highlight that specific surface patches can drive nonspecific binding and that the impact of surface patches can largely only be modulated by mutations in close proximity to or part of the patch, or potentially by allosteric effects. This is further supported by the fact that the V99K mutant protein displays a similar decrease in the binding affinity (~10² increase in K_D) to library variants that have multiple other charged mutations distributed across the Fab (Q108E, S70E, and T5D).

Given the high predictive power of CDR-associated surface patch properties for the nonspecific binding of the individual variants, we next set out to correlate the behavior across the entire library. Indeed, when plotting Φ_H^CDR versus Φ_ΣPN^CDR for each variant, high-affinity nonspecific binders (K_D ~ 1 µM) are completely segregated from variants less prone to nonspecific binding (K_D > 10 µM) (Fig. 3F). More specifically, high-affinity binders exclusively display Φ_H^CDR > 330 Ų, whereas low-affinity binders only display Φ_H^CDR < 310 Ų. Similarly, high-affinity binders exclusively display Φ_ΣPN^CDR < 325 Ų, whereas low-affinity binders only have Φ_ΣPN^CDR > 325 Ų. As such, there appears to be a balance between hydrophobic and charged patches as low-affinity binders display an excess charged patch area (Φ_ΣPN^CDR > Φ_H^CDR). These observations suggest that quantification of surface patch areas can be directly related to observed nonspecific binding affinities across an antibody library.

Reduction of Ionic Strength Reveals DNA-Induced Phase Separation of Antibodies.

Having established an understanding of the binding behavior under physiological salt conditions, we performed MDS measurements at lowered ionic strengths to assess the role of electrostatic interactions in particular. Electrostatic contributions are expected to become increasingly dominant at lower ionic strengths, given the accompanying decrease in Debye screening. Such electrostatic interactions are expected to be highly important also in vivo as antibodies commonly target proteins, which are integrated into charged membranes and encounter a plethora of charged macromolecules in the blood stream with the endothelial glycocalyx (53). Moreover, varying ionic strength is a common strategy in protein and antibody development to investigate the impact of fundamental physicochemical forces underlying association processes and to optimize formulation strategies (55–57). To this end, we studied the interactions between HzATNP library antibodies with DNA in buffer with reduced salt content [2 mM HEPES buffer at pH = 7.4, 15 mM NaCl, with 0.01% (v/v) Tween]. Similar to the physiological conditions above, we first probed interactions at 1 µM DNA and 6.7 µM antibody for all variants. As shown in Fig. 2B, we find a size increase for pure DNA in low salt conditions as compared to physiological salt conditions (R_h^{DNA, LS} = 4.5 ± 0.2 nm versus R_h^{DNA, HS} = 3.6 ± 0.3 nm). The size increase at low salt can be rationalized by an increase in monomer repulsion within the polymer chain (58).

For the interaction between the antibodies and the DNA, we observed limited changes to R_h^{DNA, LS} at lowered salt conditions for the WT antibody, the hydrophobic variants (variants 10, 13, 15, and 16), and variant 6 (Fig. 2B). Significant hydrodynamic radius increases, however, were obtained for the other variants of the charged mutation series (variants 1, 2, and 4; see Fig. 2B). Variants 1, 2, and 4, all of which have an increased positive charge (particularly V99K/R), form nanoclusters up to ∼9 to 10 nm in the hydrodynamic radius at low salt concentrations. Notably, both DNA and antibody individually do not form such large complexes (SI Appendix, Fig. S2). This indicates a change in the binding mode toward a nonstoichiometrically constrained electrostatic network assembly process as such complexes are expected to have roughly twice the molecular weight (59) compared to saturated complexes under physiological conditions. This was further confirmed by binding curve titrations in which variant 2 (Fig. 4A) and variant 4 (SI Appendix, Fig. S3) did not show binding saturation with increasing antibody concentrations as the hydrodynamic radius was found to continuously increase. This suggests that binding of additional antibodies can allow for further recruitment of DNA strands based on additive electrostatic interactions to form clusters, as opposed to saturation of defined binding sites. This is likely associated with the introduction of an additional positive charge by the V99K/R mutations common to these variants. Indeed, an increase in the CDR charged patch area is observed for variant 1, 2, and 4 compared to the WT (ΔΦ_P^CDR ~ 40 to 50 Ų). Furthermore, a large excess of positive charged surface patch area compared to the negative charged surface patch area in the CDRs is observed (Φ_ΔPN^CDR = Φ_P^CDR–Φ_N^CDR = 166 Ų).

Fig. 4.

Changes in binding regime at lower ionic strengths reveal DNA-induced phase separation of antibodies. (A) Binding curve titration of selected HzATNP antibody variants against DNA at 1 µM and lowered ionic strengths [2 mM HEPES buffer, pH = 7.4, 15 mM NaCl supplemented with 0.01% (v/v) Tween, 4 °C]. The lack of binding site saturation observed for both WT and variant 2 compared to variant 6 indicates a change in binding regime toward electrostatic network assembly into clusters and phase separation. Points beyond the phase boundary were recorded in the supernatant postcentrifugation. Errors represent SDs from triplicate measurements. (B) Bright-field microscopy images of the observed macroscopic phase behavior recorded at 40 µM antibody and 5 µM DNA. Variant 2 compared to the WT antibody displayed a significantly larger mean condensate size (d) immediately after mixing. (C) Molecular models of the surface charge potential to represent the surface properties under decreased ionic strength using the Adaptive Poisson-Boltzmann Solver (ABPS) Electrostatics PyMOL2 plug-in, including indication of mutation sites. Resulting CDR-associated positive charged patch (Φ_P^CDR, blue box), negative charged patch (Φ_N^CDR, red box), and charge patch difference (Φ_ΔPN^CDR, purple box) surface area for the individual variants.

To investigate this electrostatic assembly further, we imaged mixtures of variants 2 and 4 with DNA using fluorescence microscopy (5 µM DNA and 40 µM antibody) (Fig. 4B) and observed spherically shaped assemblies that readily fused and merged. This indicates that these assemblies are phase-separated condensates, resulting from a demixing of antibody and DNA. Control experiments with both DNA and antibody at the same conditions confirm that neither DNA nor antibody displays this phase separation behavior on their own (SI Appendix, Fig. S4). Further testing with 20- and 100-mer DNA as well as PolyA RNA (700 to 3,500 kDa) also yielded phase-separated condensates, suggesting that the behavior is not based on an interaction specific to a particular secondary DNA structure or length (SI Appendix, Figs. S5 and S6). Finally, we identified that the demixed phase undergoes complete dissolution upon disruption of electrostatics via addition of high salt content but leads to aggregate formation upon addition of the hydrophobic disruptor in 1,6-hexanediol (SI Appendix, Fig. S7). This corroborates the initial hypothesis that the observed behavior represents phase separation of antibodies with DNA based on additive electrostatic interactions.

Despite the decrease in the positive charged patch area of the WT antibody compared to variant 2 (ΔΦ_P^CDR = −51 Ų), the WT antibody also displays no DNA binding site saturation with increasing antibody concentrations (Fig. 4A), despite forming much smaller clusters in hydrodynamic radius. The cluster assembly, however, coincides with phase separation at higher antibody concentrations. Hence, the V99K/R mutations only further promote the electrostatic assembly as the WT antibody already displays a large excess of positive charged over the negative charged surface patch area (Φ_ΔPN^CDR = 107 Ų). When imaging directly after mixing, variant 2 also tends to form larger condensates on average compared to the WT under the same conditions (droplet diameter: d_v.2 = 5.3 ± 1.7 µm versus d_v.WT = 1.3 ± 0.4 µm both at 40 µM antibody and 5 µM DNA) (Fig. 4B). This more effective recruitment into condensates is an indication of less electrostatic repulsion within antibody–DNA complexes for variant 2 compared to the WT. Finally, variant 6 does not show any of the observed assembly behavior besides DNA interactions reminiscent of those observed under physiological conditions. This finding is in agreement with an increase in the negative charged patch area compared to the WT and variant 2 by ΔΦ_P^CDR = +56 Ų and Δ_v.2Φ_P^CDR = 37 Ų, respectively. Hence, the additional negatively charged mutations in variant 6 particularly V99D generate additional repulsion capable of disrupting cluster formation and phase separation. Accordingly, the positive and negative charged patch areas appear to be much more balanced for variant 6 with Φ_ΔPN^CDR = 44 Ų. This trend of the charged patch balance is further supported by mapping the individual variants for both their Φ_P^CDR and Φ_N^CDR (SI Appendix, Fig. S8). Here, the cluster forming and phase-separating variants 1, 2, and 4 are characterized by much higher Φ_P^CDR compared to the rest of the library. The hydrophobic variant cluster around the WT with decreased Φ_P^CDR but similar Φ_N^CDR was compared to variants 1, 2, and 4. To confirm their phase separation behavior, variant 16 was further subjected to microscopy imaging and also shown to phase separate similar to the WT antibody in the presence of DNA (SI Appendix, Fig. S9). Interestingly, variant 6 does not further display a decrease in Φ_P^CDR but is separated from the rest of the library by its much larger Φ_N^CDR.

To further characterize the phase behavior of antibody–DNA condensates and map the chemical phase space in more detail, we applied PhaseScan (49), a droplet microfluidic-based approach for generating high-resolution phase diagrams. We mapped out the phase space for the WT antibody and variant 2 over a broad range of conditions as shown in Fig. 5A. Overall, the phase boundary is shifted to higher DNA concentrations for variant 2. Hence, a higher concentration of DNA is necessary to allow for phase separation of variant 2 compared to the WT, while phase separation is more favorable with increasing DNA concentrations for variant 2. As such, our data suggest that the additional positive surface charge of variant 2 leads to more effective isolation of individual DNA strands at large excess of antibody, while higher contents of the negatively charged DNA can be stabilized in condensates (SI Appendix, Fig. S8). Below c_DNA^sat (∼4 µM), the DNA–antibody complexes show a continuous increase in size and formation of high–molecular weight clusters (HMWCs) with increasing antibody concentrations without undergoing phase separation (c_DNA = 0.1 µM, Fig. 5 B, Lower). Hence, the electrostatic network assembly is an inherent property of the system, as opposed to an induced behavior in phase-separating conditions. Above c_DNA^sat, cluster formation and size increase with higher antibody concentrations is similarly observed in the dilute phase (c_DNA = 5 µM, see Fig. 5 B, Higher). DNA acts as a limiting cross-linker between antibody molecules, which are required to undergo large-scale network formation and condensation from isolated clusters. Hence, negatively charged macromolecules can act as triggers for antibody phase separation even at low concentrations, which is supported by a stoichiometric ratio analysis in condensates, highlighting a large excess of antibodies compared to DNA molecules (antibody:DNA ratio > 100:1, see SI Appendix, Fig. S11)(60).

Fig. 5.

Mechanistic characterization of DNA-induced antibody phase separation. (A) Phase diagrams of variant 2 (Left) and WT (Right) at room temperature as obtained by PhaseScan with gray points indicating a mixed phase and red points phase-separated conditions. Individual points in the phase diagram represent a condition probed by a water-in-oil droplet generated with combinatorial microfluidics, as shown by images of representative droplets. (Scale bar, 40 µm.) In each phase diagram, the probability of phase separation is assessed via a Random Forest Classification. This projects a probability of phase separation and is represented by the red (probability of phase separation = 1) to gray (probability of phase separation = 0) gradient that is drawn behind the individual points. This phase separation probability is indicated by the gradient bar next to the variant 2 phase diagram. (B) Binding curve titrations across horizontal cross-sections (const. DNA, varying antibody concentrations) of the phase diagram below (c_DNA = 0.1 µM) and above (c_DNA = 5 µM) the minimal DNA content necessary to undergo phase separation. Phase-separated species were removed by centrifugation to study cluster formation in the dilute phase. Errors represent SDs from triplicate measurements. In both instances, a continuous increase in the complex size is observed. This indicates that network formation based on additive electrostatic interactions is an inherent property of the system, not just of phase-separating conditions. Hence, the DNA saturation concentration functions as a critical concentration of available linkers between DNA–antibody clusters.

Although the decrease in ionic strength is not directly comparable to physiological conditions, phase separation was observed at antibody concentrations at least an order of magnitude lower than used in common formulations. The in vivo environment can further present high abundances of macromolecular species which in addition to high antibody concentrations post injection can generate crowding effects. Both increases in protein concentration and macromolecular crowding are well known to drastically increase the propensity of protein phase separation. Indeed, additional experiments under physiological salt concentrations but in the presence of polyethylene glycol (PEG) as a crowding agent show that a mixture of WT antibody and DNA (40 µM WT, 5 µM DNA; same concentrations as probed at lowered ionic strength) undergoes phase separation at moderate concentrations of the crowder PEG (at 5% PEG 10 k MW). Notably, both the antibody and DNA individually do not phase separate (SI Appendix, Fig. S9). As such, this observation of DNA-induced antibody phase separation could even be utilized as a screening tool to assess antibodies for their heteromolecular assembly propensity.

Discussion

In this work, we have presented a systematic study of the impact of surface patch properties on antibody nonspecificity using the HzATNP antibody library as a model system and DNA as a nonspecificity ligand. In utilizing and establishing MDS to probe nonspecific interactions in the solution phase, we quantified nonspecific binding to observe changes of multiple orders of magnitude in K_D via only few mutations. We have identified a hydrophobic patch in the CDRs as the key determinant of nonspecific binding with K_D values up to 1 µM in affinity. Here, only mutations disrupting the patch decrease the observed binding affinity with the individual point mutation V99K displaying the most pronounced effect (K_D > 100 µM). We further quantified the changes in the hydrophobic, positive charged, and negative charged patch surface areas throughout the library using the computational protein evaluation tool MOE. This allowed us to segregate high-affinity binders (K_D ~ 1 µM) from low-affinity binders (K_D > 10 µM) by considering their hydrophobic versus total charged patch areas (Fig. 6, Top). Furthermore, DNA-induced phase seperation was discovered based on a change in condition to lowered ionic strengths. The decrease in Debye screening causes the charged surface patch properties to become the dominating factor distinguishing between behaviors (Fig. 6, Bottom). More specifically, our experiments revealed a continuous increase in DNA–antibody complex sizes with increasing antibody concentrations based on additive electrostatic interactions for a subset of the library, which displays positively charged mutations in the V99 position. In tracing binding behaviors as well as the phase space for selected variants, we rationalized how the balance between positive and negative charged patches can translate into a macroscopic behavior. As such, variant 2 does phase separate, forming large condensates and showing no binding saturation, whereas variant 6 does not phase separate and displays binding site saturation with increasing antibody concentration. Taken together, we quantitatively connect nonspecific binding and assembly to the surface patch properties across an antibody library to discern guidelines for optimizing protein specificity (Fig. 6).

Fig. 6.

Balance between surface patches governs nonspecific binding and assembly of HzATNP antibodies with DNA. At physiological salt conditions, a hydrophobic patch drives nonspecific binding of the WT antibody to DNA, whereas at lowered ionic strengths, the surface charge properties become dominant. Mapping the hydrophobic (Φ_H^CDR) versus total charged patch sizes (Φ_ΣPN^CDR) for all variants shows clear segregation of low-affinity binders from high-affinity binders at physiological salt. Points represent individual protein variants which are colored for their nonspecific DNA binding affinity (blue ~ 1 µM, yellow ~ 100 µM). At lowered ionic strengths, the balance between positive (Φ_P^CDR) and negative (Φ_N^CDR) charged patches governs high–molecular weight cluster (HMWC) formation and phase separation. Points represent individual protein variants which are colored for the hydrodynamic radius increase induced to DNA postaddition of antibody under lowered ionic strength at 1 µM DNA and 6.7 µM antibody (blue ~ 5 nm, yellow ~ 2 nm). These selection criteria highlight how surface patch properties could be applied to guide disruption of the surface patches to improve nonspecific binding and assembly.

Methodological shortcomings have so far hindered quantifying nonspecific interactions, with common tools relying on disruptive preparation steps such as surface immobilization or indirect readouts such as chromatography retention times (41). Determining accurate readouts for nonspecificity, however, is crucial to reveal the potential impact of nonspecificity in vivo and for optimizing protein design and engineering approaches. By applying MDS, we provide quantitative characterisation under native solution conditions and determine DNA nonspecific binding affinities with K_D values of up to 1 µM. Our approach further reveals that an individual V99K mutation in the center of a hydrophobic patch reduced the nonspecific binding affinity by two orders of magnitude. Hence, these measurements show great promise for optimizing drug candidates as the changes in behavior can be assessed more accurately and extensively. Furthermore, the physicochemical driving forces of the nonspecific binding can be assessed by subjecting interactions to changes in ionic strength. Identification of electrostatically or hydrophobically driven nonspecificity in this manner could be a valuable addition to the screening workflow without requiring much prior information on the interacting systems. Hence, application of MDS to the nonspecificity of antibody libraries against a wide range of potential targets could be a critical step toward improving antibody specificity.

An important element of the study of nonspecific binding is its correlation to unwanted molecular assembly events and macroscopic manifestations in vitro and in vivo (14). Antibody phase separation has been described previously for single-component systems, where strong surface charge asymmetries generate regions with complementary charges on an individual antibody (61–63). It has, however, remained unclear on a more fundamental level how nonspecific binding events can trigger or lead to these processes in heteromolecular systems (i.e., systems composed of the antibody itself and other components). The observation that some antibodies within our library undergo phase separation in the presence of DNA, while others do not, provides a critical opportunity to investigate and rationalize this assembly event triggered by a nonspecificity ligand (i.e., DNA). Furthermore, the fact that heteromolecular phase separation is also observed in the presence of moderate crowder and at much lower protein concentrations than in typical formulation conditions could suggest potential relevance also in vivo. This is particularly important as our binding analysis highlights that DNA at low concentrations acts as a cross-linker between antibodies. Moreover, given the important role of negatively charged biopolymers in conferring phase separation, other molecules such as highly abundant carbohydrates in the endothelial glycocalyx after administration may induce phase separation. Hence, this tendency for increased heteromolecular assembly might suggest functional inferiority of variants that form nanoclusters and phase separate. Here, methods like MDS, dynamic light scattering or turbidity measurements could be applied to detect nanoclusters and condensates at lower ionic strengths as potential screening tools.

Nonspecificity has been an elusive topic to grasp, with the responsible molecular features not having been addressed sufficiently. Hence, it is necessary to rely on extensive screening campaigns rather than utilizing rational design approaches to reduce nonspecificity. Our data, however, highlights that only mutations which directly disrupt the causative surface patch decrease the nonspecific affinity. Investigations on the molecular origins of antibody nonspecificity, however, commonly rely on sequence-based net properties (64–66) or abundances of individual amino acids (67–71). These approaches are unlikely to directly impact the nonspecific binding paratopes and, hence, are largely limited to excluding candidates with extreme properties (e.g., very high sequence charge compared to common distributions). Here, we establish rational design guidelines across the library based on the antibody surface patch areas from which nonspecific binding and assembly can be predicted and optimized. This indicates the potential to assess problematic nonspecificity by identifying suitable surface patch property profiles and targeting them for mutagenesis. Finally, we note that the nonspecificity paratope likely stems from the binding site raised against the intended target in the small-molecule trinitrophenyl. Interestingly, surface patches are commonly formed in affinity maturation processes as a consequence of generating or increasing affinity for the intended target (14). Here, introducing counter patches or limiting patch sizes might be a key to eliminate the potential formation of problematic surface patches during maturation via a priori negative selection and design.

Materials and Methods

HzATNP Antibody Variants.

Design, characterization, and expression of the HzATNP antibodies were performed as described in detail previously (38, 72). Briefly, mutations were introduced to the WT heavy-chain and light-chain vectors (pNNC340 and pNNC341) using a mutagenesis kit, after which PCR was applied to generate linearized vectors. Transformation into Escherichia coli DH5alpha competent cells (Thermo Fisher Scientific) was then performed, and the protein-encoding sequences were then identified after plasmid harvesting from overnight cultures via a sequencing service. Expi293F™ cells (Thermo Fisher Scientific) were then transfected with both heavy- and light-chain vectors and harvested after 5 d of incubation. Purification of antibody variants was performed by protein A resin affinity chromatography and gel filtration, after which antibody samples were stored in aliquots at roughly 5 to 10 mg/mL at −80 °C. Protein stocks were thawed on ice and shock frozen using liquid nitrogen, up to a maximum of three freeze–thaw cycles.

DNA and RNA Oligos.

DNA oligos were purchased from Merck (Darmstadt, Germany; HPLC purified and dry), with strands including a 5′ modification with Cy3. The 100-mer, 50-mer, and 20-mer sequences used were C​TCA​CCC​ACA​ACC​ACA​AAC​AAT​TTA​AAT​AAT​ATT​AAA​TAA​TAT​TAA​TAT​ATT​ATC​GAT​TAA​ATA​ATA​ATT​AAT​TAA​TATTGGTTGGATGGTAGATGGTGA, CTCACCCACAACCACAAACAATTTAAATAATATTAAATAATATTAATATA, and CTCACCCACAACCACAAACA, respectively. PolyA RNA lyophilized was purchased from Merck (Darmstadt, Germany). PolyT RNA 20-mer, including a 5′ modification with Cy5, was purchased from Biomers (Ulm, Germany)

Fabrication of MDS Devices.

Fabrication of MDS devices was performed according to previous descriptions (47). After designing devices using AutoCAD software (Autodesk, San Rafael, CA, USA), printed acetate masks of these designs (Micro Lithography Services, Essex, UK) were used to generate SU-8 3025 molds. This was done utilizing photolithographic methods, spin coating, and development via PGMEA washing after which approximately 25-µm device heights, as determined by a profilometer (Dektak, Bruker, USA) were obtained. The master was then placed in a Petri dish which was filled with polydimethylsiloxan (PDMS) (two-component system, Momentive, Techsil, Bidford-on-Avon, UK) supplemented with carbon black nanoparticles, and after baking, the PDMS mold was removed from the silicon waver, cut into individual chips, and punched to yield inlet and outlets. Afterward, devices were bonded to glass coverslips by simple contacting postsurface radicalization in an oxygen plasma oven (Femtro from Electronic Diener, Ebhausen, Germany).

MDS of DNA and DNA–HzATNP Antibody Complexes.

Operation of the MDS devices was performed according to previous descriptions (47, 48). In brief, microfluidic devices were prefilled with experiment buffer [20 mM HEPES buffer, pH = 7.4, 150 mM NaCl, with 0.01% (v/v) Tween or 2 mM HEPES buffer, pH = 7.4, 15 mM NaCl, with 0.01 % (v/v) Tween], and negative pressure was applied to the outlet via a glass syringe (Hamilton, Bonaduz, Switzerland) connected to a syringe pump (neMESYS, Cetoni GmbH, Korbussen, Germany) allowing for flow control. Sample and buffer loading was then performed through reservoirs connected to the respective inlets, and images were recorded through a custom-built inverted epifluorescence microscope which was equipped with a CMOS camera (Prime 95B, Photometrics, Tucson, AZ, USA), bright-field LED light sources (Thorlabs, Newton, NJ, USA), and a Cy3-laserline filter cube (Laser2000, Huntingdon, UK). Extraction of the diffusion profiles was performed via analysis of the fluorescence images with a custom written analysis software (73) that fits the recorded diffusion profiles with numerical model simulations, which are solutions to the diffusion–advection equations for mass transport under flow (48).

Diffusional Sizing of HzATNP Antibodies via Intrinsic Fluorescence.

Intrinsic fluorescence was used to detect HzATNP antibody diffusion profiles and was only applied in high excesses of antibody, where the intrinsic fluorescence signal of DNA could be neglected. For UV measurements, quartz instead of glass coverslips was used to mitigate the absorption of glass itself. Here, a custom-built inverted microscope which was equipped with a 280-nm LED light source (M280L3, Thorlabs, Newton, NJ, USA), a charge-coupled device camera (Prime 95B, Photometrics, Tucson, AZ), and filter set Semrock TRP-A-000 (Laser2000, Huntingdon, UK) was utilized as described previously (73, 74).

DNA–HzATNP Antibody Library Phase Separation.

Phase separation was exclusively induced in vitro, and specific conditions are reported for each specific dataset. Generally, phase separation was induced by gently mixing the individual components in 500-µL centrifuge tubes. Buffer [2 mM HEPES buffer, pH = 7.4, 15 mM NaCl, with 0.01% (v/v) Tween] was added first, followed by DNA and then antibody. Usually 10 µL of 40 µM antibody and 5 µM DNA 50-mer were prepared, and images were obtained by application of 1 to 2 µL of sample onto a coverslip, followed by immediate imaging in sealed chambers. For phase separation experiments of the HzATNP library with PolyA, PolyA was spiked with 5% (w/v) of PolyT Cy5 labeled, to give a stock solution of 15 mg/mL in 50 mM HEPES buffer pH = 7.2, which was diluted in the sample to yield buffer conditions of 2 mM HEPES (pH 7.4) and 15 mM NaCl, 0.01% (v/v) Tween. Specific conditions for the usage of 20-mer, 100-mer, and PolyA are described in detail where shown, where fluorescence was observed using an inverted microscope (OpenFrame, Cairn Research, Faversham, UK) equipped with a 20× air objective (Nikon, Surbiton, UK) and appropriate filters (Laser2000, Huntingdon, UK) and a high-sensitivity camera (Prime BSI Express sCMOS, Photometrics, Tucson, AZ, USA).

Dissolution Experiments.

For the dissolution experiments, DNA–HzATNP antibody phase-separating samples at 50 µM variant 4 and 6.25 µM DNA 100-mer were prepared as discussed previously. Then, 20% (v/v) of 1M NaCl or 50% (w/v) 1,6-hexanediol was added to give sample conditions of 40 µM variant 4 and 5 µM DNA 50-mer plus 212 mM sodium chloride or 10% (w/v) 1,6-hexanediol, respectively. Imaging was performed as reported above.

Mapping Chemical Phase Space Using PhaseScan and Tie-Line Determination.

PhaseScan was operated according to previously described procedures (49). Briefly, droplets were generated on the microfluidic device using automated syringe pumps, allowing for combination and variation of the aqueous droplet components prior to the droplet-generating junction, where FC-40 oil [containing 1% (w/v) fluorosurfactant, RAN Biotechnologies] was flown into the chip at a constant flow rate to generate droplets. Droplets were then imaged subsequently with an inverted microscope (OpenFrame, Cairn Research, Faversham, UK) using a 20× objective and collected through a camera (Kinetix, Photometrics, Tucson, AZ, USA). Based on the fluorescence intensity in each channel, every droplet could be mapped to a unique point in the phase space and assessed for phase separation. Dilute DNA concentrations were then determined by evaluating the background fluorescence intensity in each droplet from which tie-line gradients were then evaluated by segmenting droplets by the dilute phase concentration and fitting to the points (60).

Antibody Homology Model Generation.

Homology models of the Fab regions of the antibodies were created using the Antibody Modeler application in MOE 2022.09 (50). Using the method described by Maier and Labute (75), a homology search of the antibodies in the Protein Data Bank was performed to identify the best matching framework and CDR templates for both chains. These were grafted together to produce high-quality chimeric Fv templates for homology modeling. Because there are CH1 mutations in the library, Fab sequences were input to the Antibody Modeler, which was run in the seq mode, guaranteeing that the sequence of the model contains the entire Fab, including capping groups that neutralize the Fv C termini. For this, the C1 domains are modeled from the PDB:1N8Z Fab structure template.

Property Calculations on 3D Models.

The patch property descriptors Φ were calculated using the Protein Properties application of MOE 2022.02, which produced 100 conformations of each model where the framework is restrained, and side chains are free to move using LowModeMD, and alternate protonation states are sampled from pH 6.4 to 8.4 (centered at pH = 7.4) using the Protonate three dimensional (3D) method (76, 77). Hydrophobic patches consist of regions where a hydrophobic potential equal to or greater than that of a methyl group persists over a surface area greater than 50 Ų, where the hydrophobic potential is determined using the SLogP method for each atom and mapping the result onto the surface. Charged patches are formed where there is excess forcefield charge (Amber10:EHT) sustained over a surface area of 40 Ų, and all descriptors Φ_H^CDR, Φ_P^CDR and Φ_N^CDR are similarly calculated by the sum of the surface areas of the hydrophobic patches involving at least one CDR atom averaged over the 100 samples using a Boltzmann weight centered at the target pH of the calculation (pH = 7.4). Patch descriptors carry errors but were evaluated as small compared to the changes induced by relevant surface patch mutations.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Data, Materials, and Software Availability

All study data are included in the article and/or SI Appendix.