Abstract

Introduction

The intense interest in developing antibodies as therapeutics continues to fuel advances in methodologies for discovering and evolving them (Nelson et al., 2010). In vitro methods such as phage and yeast surface display are particularly attractive because of the ease of using them for selecting lead candidates and evolving their binding properties, as well as the high degree of control they afford over antigen presentation (Sidhu, 2000; Gai and Wittrup, 2007). The compatibility of some of these methods with fluorescence-activated cell sorting (FACS) enables quantitative evaluation of antibody fusion proteins during selection, and prior to subcloning and assessment of the autonomous antibody variants (Boder and Wittrup, 1997; Feldhaus et al., 2003). These and other attractive attributes of in vitro display technologies have led to their widespread use for generating high-affinity antibodies (Bradbury et al., 2011).

Nevertheless, a key weakness of such in vitro selection methods is their increased susceptibility to yield antibodies and fragments thereof with modest specificity and/or stability relative to immunization (Bradbury et al., 2011). This is likely due to the reduced quality control (filtering) mechanisms that antibodies are subjected to during in vitro evolution relative to the analogous process in vivo. This has led to many creative approaches to improve the stringency of selection of stable and specific antibody variants in vitro (Jung et al., 1999; Boder and Wittrup, 2000; Orr et al., 2003; Brockmann et al., 2005; Chao et al., 2006; Demarest and Glaser, 2008; Arbabi-Ghahroudi et al., 2009a; Miller et al., 2010; Brockmann, 2012,Buchanan, 2012; Traxlmayr et al., 2012; Traxlmayr and Obinger, 2012; Fennell et al., 2013; Hasenhindl et al., 2013; Kim et al., 2014a; Turner et al., 2014). One widely used approach is to thermally stress antibody libraries displayed on phage or yeast, and then select variants that retain binding to antigen (Jung et al., 1999; Orr et al., 2003; Arbabi-Ghahroudi et al., 2009b; Mabry et al., 2010; Brockmann, 2012; Traxlmayr et al., 2012; Traxlmayr and Obinger, 2012; Fennell et al., 2013; Xu et al., 2013). Another common approach is to perform negative selections to eliminate antibody variants that bind to noncognate antigens (Bradbury et al., 2011; Miersch et al., 2015). A third approach is to use positive selections (typically in the absence of antigen) for binding to conformational ligands (such as Protein A and L) that recognize folded antibody domains and which bind to the frameworks of certain V_H/V_L subclasses (Bond et al., 2003; Jespers et al., 2004a,b; Barthelemy et al., 2008; Famm et al., 2008; Kim et al., 2014b).

We have developed an approach for generating lead antibody fragments (V_H domains) that recognize amyloid-forming polypeptides (such as the Alzheimer's Aβ peptide) by grafting hydrophobic self-recognition peptides into the third complementarity-determining region (CDR3; Perchiacca et al., 2012a). The resulting Grafted AMyloid-Motif AntiBODIES (gammabodies) typically recognize amyloid aggregates with modest (micromolar) affinity and the corresponding monomeric polypeptides with even weaker affinity (Perchiacca et al., 2012a,b, 2014). We reasoned that these lead antibody domains could be readily evolved for improved binding affinity using a directed evolution approach that involves generating mutant V_H libraries, displaying them on the surface of yeast, and selecting variants with improved binding via FACS. Therefore, we evaluated the feasibility of this approach for improving the binding of a V_H domain with Aβ residues 33–42 in CDR3 (Aβ33-42 V_H) to Aβ42 monomer. Here, we report the isolation of unstable and stable V_H variants with improved binding to Aβ as well as methods for improving the evolution of stable and specific antibody fragments on the surface of yeast.

Materials and methods

Results

Evolution of grafted amyloid-motif V_H domains using conventional yeast surface display leads to low stability and specificity

Toward our goal of evolving grafted V_H domains with amyloidogenic self-recognition peptides in CDR3 for improved binding affinity, we first fused a V_H domain [B1a (Barthelemy et al., 2008)] with Aβ residues 33–42 in CDR3 (herein referred to as wild type or WT; Fig. ​Fig.1)1) to the C-terminus of a yeast surface protein (Aga2). This fusion protein displayed well on the surface of yeast as judged by the presence of a myc tag at the C-terminus of the V_H domain (Fig. ​(Fig.2A,2A, top panel). However, it possessed little binding to Aβ42 monomer at micromolar concentrations. Therefore, we performed random mutagenesis throughout the entire domain and selected variants with improved affinity for Aβ42 monomer. The best V_H binder isolated after one round of mutagenesis and selection (D1) was subjected to a second round of mutagenesis and selection to isolate the most improved variant (D2).

Open in a separate window

Fig. 1

Sequence of the Aβ33-42 V_H antibody and structure of the parent V_H domain. (A) Amino acid sequence of the V_H domain (B1a) with Aβ residues 33–42 in CDR3. The CDRs are defined using Kabat numbering and underlined residues indicate the grafted CDR3 sequence. (B) Crystal structure of parent V_H domain (PDB: 3B9V) in which the original CDR3 is highlighted (without the grafted Aβ peptide).

Open in a separate window

Fig. 2

Flow cytometry analysis and secondary structure evaluation of WT and mutant versions of Aβ33-42 V_H domains that were evolved using the conventional display (anti-myc) antibody. The WT V_H domain was evolved for improved binding to biotinylated Aβ42 using four rounds of mutagenesis and selection to generate four variants (D1–D4). D1 is the first generation V_H variant and D4 is the fourth generation variant. (A) Aβ (500 nM) binding and V_H yeast display were evaluated using flow cytometry in PBS with 1% (v/v) Triton X-100 and 10 mg/ml BSA. The gated populations highlight the fraction of yeast cells displaying V_H domains. (B) Comparison of binding of displayed V_H domains to Aβ42 relative to a nonspecific hydrophobic peptide (IAPP) at 500 nM biotinylated peptide in PBS with 1% (v/v) Triton X-100 and 10 mg/ml BSA. The relative binding to each biotinylated peptide was normalized to V_H display levels. (C) Level of V_H display evaluated using an antibody that recognizes a myc tag at the C-terminus of the Aga2-V_H fusion proteins. The error bars represent standard deviations for two independent experiments. (D) Evaluation of the secondary structure of autonomous (nonfusion) WT and mutant Aβ33-42 V_H domains. The V_H domains were expressed in bacteria and characterized using CD spectroscopy.

This process was repeated two additional times and it yielded four unique V_H domains with increasing numbers of mutations. These variants (referred to as ‘D’ for display antibody) possessed 3 (D1), 6 (D2), 9 (D3) and 12 (D4) mutations (Supplementary Fig. S1), including one in the grafted amyloid segment in CDR3 (V100fM in D4, corresponding to V39M in Aβ42), one other in CDR3 (D100jG in D2 and later generations), one in CDR1 (D31G in D2 and later generations), and the rest in the framework regions. Notably, the single disulfide bond in the WT V_H that is known to contribute to folding stability (Glockshuber et al., 1992) was eliminated in the second round of mutagenesis and selection (D2 and later generations) due to the acquisition of a cysteine knockout mutation (C22R; Supplementary Fig. S1). Interestingly, even when the mutagenesis was limited to the CDRs (which do not include either cysteine), we still observed strong enrichment of cysteine knockout mutations that occurred at low frequency due to errors in the PCR amplification (data not shown).

The loss of the intramolecular disulfide bond suggested that the evolved V_H domains may be destabilized on the surface of yeast and bind nonspecifically to Aβ. To evaluate this possibility, we compared the binding of the WT and evolved V_H domains to another hydrophobic amyloid-forming polypeptide (IAPP) relative to Aβ42 (Fig. ​(Fig.2B2B and Supplementary Fig. S2). While the evolved domains displayed good specificity for Aβ42 in the stringent screening solution (10 mg/ml BSA and 1% (v/v) Triton X-100 in PBS; Fig. ​Fig.2B),2B), this specificity was reduced when a less stringent solution was used (1 mg/ml BSA in PBS; Supplementary Fig. S2). Moreover, V_H display levels were reduced more than 2-fold for the evolved mutants that carry the cysteine knockout mutation (D2, D3 and D4) relative to WT (Fig. ​(Fig.2C),2C), further suggesting that the evolved V_H domains were destabilized on the surface of yeast.

To test whether our observations were due to the intrinsic properties of the evolved V_H domains or specific to the Aga2-V_H fusion proteins, we isolated the V_H antibodies as autonomous domains to evaluate whether they were indeed destabilized. The V_H variants expressed in bacteria at moderate levels (~10–50 mg/l) that were lower than WT (~80 mg/l). Moreover, the V_H variants containing the cysteine knockout mutation (D2, D3 and D4) ran as multiple diffuse bands on an SDS–PAGE gel, while the WT and D1 V_H domains ran as single bands (Supplementary Fig. S3). We observed similar behavior via size-exclusion chromatography, as the D4 variant displayed multiple peaks (some of which are likely degradation products), while the WT was largely monomeric (Supplementary Fig. S3). CD revealed significant differences in secondary structure between the WT and D1 domains relative to the more evolved ones with the cysteine knockout mutation (Fig. ​(Fig.2D).2D). The loss of β-sheet structure for the D2, D3 and D4 variants strongly suggested that they were largely unfolded. Even the D1 variant that displayed β-sheet structure similar to WT was significantly destabilized, as its melting temperature was reduced from 75°C for the WT to 57°C for the D1 variant (Supplementary Fig. S4). Moreover, we find that the largely unfolded V_H variants (D2, D3 and D4) displayed significant nonspecific binding as autonomous domains and little binding to Aβ42 over background (Supplementary Fig. S5).

Protein A binding to V_H domains on yeast is a strong predictor of stability and specificity off yeast

Our unsuccessful efforts in evolving V_H domains for improved and specific Aβ binding led us to reevaluate our evolution approach. Previous studies have demonstrated that increased stability of antibody fragments and related proteins is correlated with increased yeast display of Aga2 fusion proteins (Shusta et al., 1999, 2000; Feldhaus and Siegel, 2004; Chao et al., 2006; Hackel et al., 2010). This observation is generally consistent with our findings that the most destabilized V_H domains (D2, D3 and D4; Fig. ​Fig.2D)2D) corresponded to those with the lowest yeast display levels (Fig. ​(Fig.2C).2C). We also evaluated the relationship between display level and V_H stability in more detail by generating all possible single and double V_H mutants derived from the D1 triple mutant (Fig. ​(Fig.3A3A and Supplementary Fig. S6). The yeast display levels of these single and double mutants (as well as the WT and D1 domains) show some correlation with their stabilities as autonomous domains (R² = 0.48). However, it is notable that the display levels changed by only ~1.6-fold for a relatively large change in stability (ΔT_m of ~18°C; Fig. ​Fig.3A).3A). This modest change in display levels for a significant change in stability may explain our difficulty in using display levels to guide the selection of sets of mutations that improve antigen binding without destabilizing the antibody domains.

Open in a separate window

Fig. 3

Comparison of the stability of autonomous Aβ33-42 V_H domains with their binding to sequence- and conformation-specific probes on the surface of yeast. The melting temperatures of isolated V_H domains were compared with (A) antibody binding to a myc tag at the C-terminus of the V_H domains, and (B) Protein A binding to a conformational epitope on the framework of the V_H domains. The flow cytometry measurements of anti-myc antibody and Protein A (50 nM) binding to displayed V_H domains were performed in PBS with 1 mg/ml BSA. The Aβ33-42 V_H variants are WT, the D1 mutant (D31G, E46K, Y91S), and single (D31G; E46K; Y91S) and double (D31G, E46K; D31G, Y91S; E46K, Y91S) mutants derived from the D1 variant. The melting temperatures of the V_H domains were obtained using CD spectroscopy.

We reasoned that the use of alternative affinity reagents for detecting V_H display that are more sensitive to antibody stability is necessary to improve the selection process. Therefore, we evaluated the use of a conformational ligand (Protein A) specific for a discontinuous epitope on the framework of folded V_H3 domains (such as the Aβ33-42 V_H) that is largely separated from the CDRs (Potter et al., 1996, 1997; Starovasnik et al., 1999; Graille et al., 2000). Previous studies have demonstrated that the relative Protein A binding to V_H3 domains is correlated with folding stability (Bond et al., 2003; Chang et al., 2014). Indeed, we also observe a strong correlation between Protein A binding to Aβ33-42 V_H mutants on the surface of yeast and their folding stability as autonomous domains (R² = 0.92; Fig. ​Fig.3B).3B). Notably, the dynamic range for Protein A binding to displayed V_H domains is more than an order of magnitude (12-fold change) for an ~18°C change in melting temperature for isolated V_H domains. Moreover, the strongly destabilized V_H antibodies that are largely unfolded as autonomous domains (D2, D3 and D4; Fig. ​Fig.2D)2D) show little binding to Protein A on the surface of yeast (Fig. ​(Fig.44).

Open in a separate window

Fig. 4

Flow cytometry analysis of binding of a conformational ligand (Protein A) to WT and mutant Aβ33-42 V_H domains on the surface of yeast. The mutant V_H domains (defined in Supplementary Fig. S1) were isolated from yeast display libraries using the conventional display (anti-myc) antibody. The experiments were conducted using 50 nM Protein A in PBS with 10 mg/ml BSA and 1% (v/v) Triton X-100. The gated populations represent yeast cells displaying V_H domains.

Use of Protein A for detecting V_H display enables co-selection of stable domains with improved Aβ binding

These encouraging results led us to investigate whether we could incorporate Protein A into our selections for sets of mutations that improve antigen binding without destabilizing the V_H domains. One conservative approach that we initially attempted was simply to use the conventional display antibody and Aβ to select for mutations that improve antigen binding in the initial library, and then the display antibody and Protein A to select for mutations that improve stability of the subsequent lead clone. However, this approach was not robust because selection with Protein A in the absence of Aβ resulted in significant enrichment in stabilizing mutations that reduced antigen-binding activity (and vice versa for selections with antigen only; data not shown). We were unable to optimize this methodology of alternating the type of selection (antigen or Protein A) to the point that we could reliably identify sets of mutations that improved affinity without reducing stability.

Therefore, we next asked whether Protein A could be used to replace the conventional display antibody to co-select for sets of mutations that improve Aβ binding without destabilizing the V_H domains. We reasoned that the Protein A binding site on the framework of V_H domains was separated enough from the CDRs such that Protein A and Aβ could bind simultaneously. Therefore, we performed four rounds of mutagenesis and selection—three rounds of random mutagenesis throughout the entire gene and one round focused on CDR3—to select four V_H domains with progressively improved Aβ binding. These variants (referred to as ‘P’ for Protein A) possessed 4 (P1), 7 (P2), 9 (P3) and 12 mutations (P4; Supplementary Fig. S7). These mutations include four substitutions in the grafted Aβ peptide in CDR3 (G100dR in P1 and later generations, corresponding to G37R in Aβ42; I100hT in P3 and later generations, corresponding to I41T in Aβ42; V100fI and V100gI in P4, corresponding to V39I and V40I in Aβ42), one other in CDR3 (E98K in P4), three in CDR2 (Y52H and S62R in P2 and later generations; S50R in P3 and later generations) and the rest in the framework regions. Importantly, these mutations do not eliminate the disulfide bond in the WT Aβ33-42 V_H domain, and we have not identified such cysteine knockout mutations for any V_H variants co-evolved for Protein A and antigen binding (including variants not reported here).

We first evaluated the relative binding of the evolved V_H domains to Aβ and Protein A (Figs ​(Figs55 and ​and6).6). Aβ binding progressively increased with each V_H generation, while Protein A binding was similar or modestly reduced. The specificity of the evolved V_H domains was evaluated using a nonspecific hydrophobic antigen (IAPP; Fig. ​Fig.6A).6A). As with the Aβ33-42 V_H variants evolved using the conventional display antibody, we observe good specificity for Aβ42 binding relative to IAPP in the stringent screening solution (10 mg/ml BSA, 1% (v/v) Triton X-100, PBS). However, this strong specificity was also maintained in a less stringent solution (1 mg/ml BSA, PBS; Supplementary Fig. S8), unlike the variants evolved with the display antibody (Supplementary Fig. S2). The yeast display levels (evaluated using the conventional display antibody; Fig. ​Fig.6B)6B) and Protein A binding (Figs ​(Figs5B5B and ​and6C)6C) for the V_H variants evolved using Protein A are higher than those evolved using the conventional display antibody, although the differences in display levels are modest.

Open in a separate window

Fig. 5

Flow cytometry analysis of WT and mutant Aβ33-42 V_H domains isolated from yeast libraries using co-selection with Protein A and Aβ42. Binding of the V_H domains (defined in Supplementary Fig. S7) to (A) Aβ42 and (B) Protein A was evaluated on the surface of yeast. The experiments were conducted using 500 nM Aβ42 or 50 nM Protein A in PBS with 10 mg/ml BSA and 1% (v/v) Triton X-100. The gated populations represent yeast cells displaying V_H domains. V_H display was detected using an antibody that recognizes a myc tag at the C-terminus of the Aga2-V_H fusion proteins.

Open in a separate window

Fig. 6

Flow cytometry analysis of the WT Aβ33-42 V_H domain and mutants thereof isolated from yeast libraries using either sequence-specific (anti-myc mAb) or conformation-specific (Protein A) probes. The WT V_H domain was evolved for improved binding to biotinylated Aβ42 and Protein A using four rounds of mutagenesis and selection. P1 is the first generation V_H variant and P4 is the fourth generation variant. (A) Aβ42 binding to WT and mutant Aβ33-42 V_H domains evolved using Protein A relative to nonspecific binding to a hydrophobic control peptide (IAPP). (B) Conventional display (anti-myc) antibody binding to V_H domains co-evolved using the anti-myc antibody (D series) or Protein A (P series). (C) Protein A binding to V_H domains co-evolved using the anti-myc antibody (D series) or Protein A (P series). Aβ42, IAPP and Protein A binding were normalized to V_H display levels. The flow cytometry analysis was conducted in PBS with 10 mg/ml BSA and 1% (v/v) Triton X-100 as well as 500 nM biotinylated antigen or 50 nM Protein A. Error bars represent standard deviations for two independent replicates.

We next evaluated the properties of the autonomous V_H domains co-evolved using Protein A. The four variants expressed at modestly reduced levels in bacteria (~40–60 mg/l) relative to WT (~80 mg/l), and possessed similar size and homogeneity as judged by SDS–PAGE (Supplementary Fig. S9). Importantly, the P4 domain co-evolved with Protein A that binds best to Aβ possesses similar β-sheet secondary structure as WT (Fig. ​(Fig.7A).7A). Moreover, the P4 variant has only modestly reduced stability relative to WT (T_m of 66°C relative to 75°C for WT; Fig. ​Fig.7B),7B), it is largely monomeric as judged by size-exclusion chromatography (Fig. ​(Fig.7C),7C), and it binds similarly to immobilized Protein A as WT (Supplementary Fig. S10), despite that P4 contains 12 mutations.

Open in a separate window

Fig. 7

Biophysical analysis of WT and mutant Aβ33-42 V_H domains. The D4 V_H domain is the fourth generation variant (12 mutations, defined in Supplementary Fig. S1) isolated using the conventional display (anti-myc) antibody, while the P4 domain is the fourth generation variant (12 mutations, defined in Supplementary Fig. S7) isolated using Protein A. (A and B) Characterization of V_H domains using CD spectroscopy in terms of their (A) far-UV spectra and (B) thermal unfolding transitions monitored at 235 nm. (C) Evaluation of the homogeneity of V_H domains using size-exclusion chromatography. In (B), the first and second thermal unfolding transitions are reported to evaluate reversibility. In (C), the chromatograms are shifted vertically for clarity.

The autonomous P4 V_H domain also displays significant improvements in specific binding to Aβ42 (Fig. ​(Fig.8).8). The IC₅₀ value for P4 binding to Aβ is 199 ± 12 nM, and it displays little nonspecific binding to IAPP. This improved Aβ binding was observed for antigen in solution (Fig. ​(Fig.8)8) and to a lesser extent for immobilized antigen (Supplementary Fig. S11). Interestingly, nonspecific binding to well plates blocked with milk was significantly lower for the P4 variant than for WT (Supplementary Fig. S12). The dissociation equilibrium constant (as judged by flow cytometry) for P4 binding to Aβ42 monomer was 167 ± 65 nM. Epitope mapping analysis revealed that the P4 domain—which we targeted to the C-terminus of Aβ via the self-recognition peptide in CDR3—retains binding specificity primarily for the C-terminus of Aβ42 (Supplementary Fig. S13).

Open in a separate window

Fig. 8

Evaluation of binding of the WT and P4 variants of the Aβ33-42 V_H domain to Aβ42 relative to a control hydrophobic peptide (IAPP). The binding analysis was performed in PBS with 50 mg/ml BSA and 0.1% (v/v) Triton X-100. The normalized binding signal was calculated by dividing background-subtracted signals by the background. Error bars represent standard deviations for two duplicate experiments.

Discussion

An unexpected finding of our work is the strong bias toward selection of destabilized V_H domains that recognize Aβ using conventional yeast surface display. This seems contradictory to several previous studies demonstrating that evolution of antibody fragments and related proteins on the surface of yeast yields improvements in stability in addition to affinity (Shusta et al., 1999, 2000; Hackel et al., 2010). Indeed, the latter findings are logical because mutations that increase folding stability generally increase expression and reduce degradation (such as intracellular proteolysis in the endoplasmic reticulum) during transit to the surface of yeast (Kowalski et al., 1998; Kieke et al., 1999; Shusta et al., 1999; Coughlan et al., 2004). Thus, it is expected that selection for increased display and antigen binding should favor selection of stable antibody domains with improved binding affinity.

However, our antibody fragments possess several unusual attributes that may explain the lack of strong correlation between V_H display and stability. One attribute that may be important is their high solubility, which is due in part to three negatively charged residues (DED) at each edge of the WT CDR3 loop. The WT Aβ33-42 V_H fails to aggregate even when boiled (Perchiacca et al., 2012b). The extreme solubility of the unfolded state may reduce the sensitivity of V_H display levels to destabilizing mutations given the lack of correlation between unfolding, aggregation and expression. Indeed, we find that strongly destabilized V_H domains (D2, D3 and D4) express in bacteria at moderate levels (~10–50 mg/l) and fail to aggregate (as judged by visual inspection) when stored at 4°C for weeks (data not shown). Moreover, we observe reversible unfolding for both the WT and D1 variant (Supplementary Fig. S4) despite that the latter is much less stable (T_m of 57°C for D1 relative to 75°C for WT).

Another contributing factor to the poor correlation between yeast display levels and V_H stability may be the small size of our domain antibodies. Indeed, other reports for some small, single-domain proteins also demonstrate reduced correlation between folding stability and yeast display levels (Park et al., 2006) or secretion levels as autonomous domains (Hagihara and Kim, 2002). This may be due to the simpler folding mechanisms of small, single-domain proteins that are less likely to be perturbed by destabilizing mutations. Nevertheless, the weak correlation between yeast display and stability has also been observed for some multidomain antibody fragments and related proteins (Esteban and Zhao, 2004; Kim et al., 2006; Traxlmayr et al., 2012; Xu et al., 2013), suggesting that the small size of our V_H domains does not fully explain this behavior.

It is also likely that the properties of Aga2 influence the degree of correlation between yeast display levels of Aga2-V_H fusion proteins and the stability of isolated V_H domains. Importantly, Aga2 (~10 kDa) is similar in size as our V_H domains (~15 kDa), and it is heavily glycosylated and expected to be highly soluble. Thus, it may be that display levels of fusion proteins are strongly influenced by the properties of Aga2 for fusion partners that are relatively small and soluble. Even the display levels of some Aga2-scFv fusions are weakly influenced by overexpression of proteins that improve scFv folding (such as protein disulfide isomerase) and which strongly enhance secretion of the isolated scFvs (Wentz and Shusta, 2007). This suggests that Aga2 can alter the secretory processing of its fusion partners and may reduce the ability of the secretory pathway to eliminate destabilized protein variants such as our unfolded V_H domains.

Our use of a conformational probe (Protein A) for detecting V_H display levels on the surface of yeast deserves further consideration. We find that it is relatively simple to select stabilizing mutations that improve Protein A binding to V_H domains in the absence of co-selection for antigen binding, as others have demonstrated using phage display (Bond et al., 2003; Jespers et al., 2004a,b; Barthelemy et al., 2008; Famm et al., 2008; Dudgeon et al., 2012a,b). This is consistent with the expectation that stabilizing mutations may occur at positions throughout the V_H domain and are common enough to be readily selected. We also find that it is relatively simple to identify destabilizing mutations that improve V_H binding to antigen in the absence of Protein A. This is likely due to the increased stickiness of destabilized V_H domains that mediates binding to hydrophobic antigens such as Aβ. In contrast, it was much more difficult to select sets of mutations that improve both Protein A and antigen binding, which is consistent with the expected rare occurrence of mutations (or sets of mutations) that improve specific antigen binding and which are not destabilizing. Thus, the use of Protein A during evolution of our V_H domains acts as a folding filter that greatly reduces the rate of false positives during our selections for specific Aβ binding.

Protein A has some notable advantages over other conformational ligands used previously for detecting proper folding on the surface of yeast (Shusta et al., 2000; Weber et al., 2005; Jones et al., 2006; Aggen et al., 2011; Traxlmayr et al., 2012; Traxlmayr and Obinger, 2012). We find that the level of Protein A binding to displayed V_H domains is directly correlated to the stability of autonomous V_H domains, even for variants that appear well folded on the surface of yeast and which display reversible unfolding transitions as isolated domains (Fig. ​(Fig.3B3B and Supplementary Fig. S6). This excellent sensitivity to resolve stability differences between folded antibody fragments is in contrast to other studies demonstrating that many conformational ligands fail to provide such resolution for related proteins (Traxlmayr et al., 2012; Traxlmayr and Obinger, 2012). Instead, most previous studies have used heat stress to unfold variants that are destabilized and conformational ligands to select stable variants that do not unfold during heat stress or are able to properly refold after such stress (Shusta et al., 2000; Weber et al., 2005; Jones et al., 2006; Aggen et al., 2011). Some of these studies require high enough temperatures that strongly reduce yeast viability and require the use of PCR after each sorting step to recover genes encoding stabilized antibody fragments (Traxlmayr et al., 2012), which limits the robustness and throughput of this approach. An additional limitation is the need for conformational ligands that recognize regions on the framework (which do not overlap with the binding loops) for co-selection of mutations that improve both affinity and specificity, which is problematic in some cases.

Nevertheless, an advantage of using heat stress and selection of stable variants guided by conformational probes other than Protein A is that the former approach is not limited to specific subclasses of variable domains. For V_H domains, Protein A is generally specific for the V_H3 subclass (Potter et al., 1996, 1997; Starovasnik et al., 1999; Graille et al., 2000; Bond et al., 2003). This limits the utility of our approach for evolving antibody fragments with non-V_H3 frameworks. Nevertheless, the significant improvements in selection of stable and specific antibody fragments afforded by conformational probes such as Protein A may warrant either transferring the Protein A binding site to other types of variable domains or generating new conformational ligands specific for folded variable domains. It is also notable that Protein L recognizes a conformational epitope on Vκ domains, including the Vκ I, III and IV subclasses (Nilson et al., 1992; Enokizono et al., 1997; Svensson et al., 2004), and key residues from this epitope have been grafted onto V_L domains from other subclasses to confer Protein L binding activity (Muzard et al., 2009). Moreover, given that the stability of V_H and V_L domains is strongly influenced by the V_H–V_L interface (Tan et al., 1998; Worn and Pluckthun, 1998; Jager and Pluckthun, 1999; Rothlisberger et al., 2005), it may be that evolving scFvs using even one conformational ligand (Protein A or Protein L) will be sufficient to stabilize both domains. If true, this would broaden the utility of our approach to evolving antibodies with V_H3 and/or Vκ I, III and IV domains without the need for mutating the framework.

Our findings also raise questions about the robustness of using negative selections during the isolation of antibodies specific for hydrophobic antigens. We find that the reliability of our selections for V_H domains with increased binding activity to Aβ was improved by including 1% (v/v) Triton X-100 (as well as 10 mg/ml BSA in PBS) in the binding and screening solutions. This stringent screening solution afforded reliable and consistent improvements in antigen binding per sort, likely by blocking weak nonspecific interactions. Indeed, we confirmed this by demonstrating that the stringent screening solution reduced binding to an extremely hydrophobic nonspecific peptide (IAPP) for destabilized V_H domains (such as D4) relative to a less stringent screening solution (1 mg/ml BSA in PBS; Supplementary Fig. S2). However, the destabilized D4 variant (which is unfolded as an autonomous domain and is weakly recognized by Protein A on yeast) only showed ~25% (1 mg/ml BSA in PBS) and ~5% (1% (v/v) Triton X-100 and 10 mg/ml BSA in PBS) of the binding signal for the nonspecific antigen (IAPP) relative to Aβ at high concentration (500 nM). It is unclear if these modest nonspecific signals are strong enough to enable effective negative selections that eliminate destabilized V_H domains. In contrast, we observed much larger (orders of magnitude) differences for Protein A binding to stable V_H domains relative to destabilized ones, which provides a robust methodology for eliminating destabilized V_H domains during selections for sets of mutations that improve specific antigen binding. We are currently evaluating the generality of our findings for designing and evolving a wide range of antibody fragments (including multidomain antibodies) that are specific for different hydrophobic antigens.

It is also important to understand the weaknesses of our findings. Our best evolved variant (P4) yielded maximum binding signals normalized to background of four to six in well plates, while conventional antibodies specific for Aβ yield maximum normalized signals at least 3-fold higher (data not shown). This may be due to the bivalency of IgGs and the resulting increased avidity as well as reduced nonspecific interactions relative to our V_H domains, and highlights the need for further optimization of our grafted V_H domains. It is also notable that the specificity of our autonomous V_H domains was lower than the corresponding Aga2-V_H fusions on the surface of yeast. For example, we observed significant binding specificity of the P4 variant to Aβ42 relative to IAPP on the surface of yeast (~130-fold higher Aβ binding) but this was reduced off yeast (~10-fold higher Aβ binding). This may be due to avidity effects on the surface of yeast and/or conformational differences between isolated V_H domains and Aga2-V_H fusion proteins. We also note that the parent V_H domain (B1a; Fig. ​Fig.1B)1B) as well as related V_H (including the WT and P4 variants) and unrelated V_L domains bind to a nonspecific, highly positively charged protein (lysozyme) as autonomous domains, but fail to do so on the surface of yeast (data not shown). These findings highlight challenges in relating binding specificity of isolated antibody fragments to those fused to Aga2 on the surface of yeast. These observations combined with the preference for our evolved antibody domains (such as the P4 variant) to bind Aβ in solution (Fig. ​(Fig.8)8) relative to immobilized Aβ (Supplementary Fig. S11) suggest that they are not optimal for immunoblotting or western blotting applications. Finally, the apparent conformational specificity of our evolved P4 variant was dependent on the assay format, as we observed preferential binding of the P4 variant to monomers relative to fibrils for Aβ immobilized in well plates and the opposite specificity for Aβ blotted on nitrocellulose membranes (data not shown). These findings may be due to changes in antigen conformation and/or avidity when immobilized in different formats.

Finally, it is notable that we selected multiple mutations (2–4) per round of evolution using Protein A to obtain the P4 variant (12 mutations) with improved affinity and moderate stability. We speculate that the multiple rounds of evolution resulted in selection of some mutations that improved antigen binding but reduced stability (given our findings for V_H domains evolved without Protein A co-selection) as well as other mutations that primarily contributed to stability. If correct, this process of accumulating different types of mutations that collectively result in improved affinity while maintaining thermodynamic stability in our directed evolution experiments may mimic the natural process of somatic hypermutation and B-cell clonal selection. Recent studies have demonstrated that multiple high-affinity antibodies accumulate somatic mutations with different functions, some of which contribute to affinity (and which are destabilizing) while others contribute primarily to stability (Sun et al., 2013; Wang et al., 2013). Analogous observations have also been made for enzymes (Wang et al., 2002; Chen et al., 2005) and other affinity proteins (Houlihan et al., 2015). We are currently evaluating whether our methodology for co-evolving V_H stability and affinity leads to accumulation of mutations outside the antigen-binding site that contribute primarily to stability and which compensate for destabilizing mutations within the antigen-binding site.

Supplementary data

Supplementary data are available at PEDS online.

Author contributions

M.C.J. and P.M.T. designed the research and wrote the paper. M.C.J., C.C.L., K.E.T., L.A.R., E.K.D. and A.J.S. performed the experiments.

Funding

This work was supported by the National Institutes of Health [R01GM104130 to P.M.T.], National Science Foundation [CBET 0954450; and 1159943 to P.M.T., Graduate Research Fellowships to M.C.J., K.E.T and L.A.R.], the Pew Charitable Trust [Pew Scholars Award in Biomedical Sciences to P.M.T.] and the Richard Baruch M.D. Chair (to P.M.T).

Conflict of interest: P.M.T. has received consulting fees and/or honorariums for presentations of this and/or related research findings at MedImmune, Eli Lilly, Bristol-Myers Squibb, Janssen, Merck, Genentech, Amgen, Pfizer, Adimab, Abbvie, Abbott, Roche, Boehringer Ingelheim, DuPont, Schrödinger and Novo Nordisk.

Supplementary Material

Acknowledgements

References