Computationally designed cyclic peptides derived from an antibody loop increase breadth of binding for influenza variants

Summary

The influenza hemagglutinin (HA) glycoprotein is the target of many broadly neutralizing antibodies. However, influenza viruses can rapidly escape antibody recognition by mutation of hypervariable regions of HA that overlap with the binding epitope. We hypothesized that by designing peptides to mimic antibody loops, we could enhance breadth of binding to HA antigenic variants by reducing contact with hypervariable residues on HA that mediate escape. We designed cyclic peptides that mimic the heavy chain complementarity-determining region 3 (CDRH3) of anti-influenza broadly neutralizing antibody C05 and show that these peptides bound to HA molecules with < 100 nM affinity, comparable to that of the full-length parental C05 IgG. In addition, these peptides exhibited increased breadth of recognition to influenza H4 and H7 subtypes by eliminating clashes between the hypervariable antigenic regions and the antibody CDRH1 loop. This approach can be used to generate antibody-derived peptides against a wide variety of targets.

Graphical Abstract

In Brief

Sevy et al. have developed a method for creating cyclic peptides from antibody loops using computational design. These peptides have broad binding activity against seasonal influenza variants and represents a new strategy for peptide discovery.

Introduction

Since the mid-1990s, the role of monoclonal antibodies (mAbs) as therapeutic agents has been growing rapidly. MAbs have many favorable properties as therapeutics, as they can be generated quickly by a number of different discovery strategies, and it is possible to isolate mAbs that bind to virtually any target of interest. To date, over 60 antibody-based drugs have been approved for therapeutic use, and over 550 antibodies are currently being developed as therapeutics (Carter and Lazar, 2018). Antibodies are attractive as drugs because of their high affinity and remarkable specificity for the target of interest.

However, antibody therapeutics also face major limitations. As large (~ 150 kDa) biological molecules, they can be difficult to produce and store compared to small molecules. MAbs are susceptible to degradation and modification over time, and the cost of production is high (Elgundi et al., 2017). Antibodies cannot cross cell membranes efficiently, and therefore are only effective against targets accessible on the surface of microorganisms or cells (Carter and Lazar, 2018). In addition, in cases where antibodies are needed to recognize multiple related proteins with sequence polymorphisms, such as viral surface proteins, the specificity that is a cardinal feature of the antibody-antigen interaction can be viewed as a limitation.

To address these limitations, we developed a method of computational design of cyclic peptides based on the structure of an antibody hypervariable loop to generate mini-antibodies. Antibodies bind their targets with a surface comprising six complementarity-determining regions (CDRs), the most diverse of which is the third loop on the heavy chain (CDRH3) (Murphy et al., 2012). Potent and cross-reactive antiviral antibodies commonly have long CDRH3 loops that mediate antigen recognition (Corti and Lanzavecchia, 2013), a trend that is also evident for other non-viral targets (Bonsignori et al., 2014; Shih et al., 2012; Thomas, 1993). Long CDRH3 loops typically contact a conserved portion of the antigen and can mediate broad recognition. However, the large overall footprint of such antibodies often includes contacts made by CDRH1 and CDRH2 and by light chain CDRs. If antigen residues in the periphery of the contact area are hypervariable (i.e., vary among viral field strains), minor structural variations in those strains result in loss of binding and thus prevent broad reactivity for that mAb. We hypothesized that we could enhance the breadth of activity of a neutralizing antibody against a hypervariable viral protein by designing a peptide that adopts the conformation of the antibody CDRH3 loop but eliminates interactions with variable antigenic residues. As a proof-of-principle exercise, we used computational modeling to design a peptide derived from anti-influenza human mAb C05 whose CDRH3 binds to the highly conserved receptor-binding site on hemagglutinin (HA) (Ekiert et al., 2012). We show that this peptide binds with high affinity to influenza HA and exhibits increased breadth, as it binds to viral subtypes not recognized by the parental antibody, due to the highly conserved footprint of the peptide.

Results

Experimental workflow

To design antibody loop-based peptides, we performed computational modeling simulations using the Rosetta software suite (Alford et al., 2017) (Fig. 1). We focused our efforts on anti-influenza mAb C05, as this mAb binds its antigen primarily with a long (26 amino acid) CDRH3 loop that has a nearly identical conformation in the bound and unbound states. We first removed the CDRH3 loop from the structure of C05 in the unbound state (PDB ID 4fnl) and added cysteines to the N-and C-termini in silico (Fig. 1A). We reasoned that a disulfide-stabilized cyclic peptide would have reduced conformational entropy compared to a linear peptide and would more readily adopt the active conformation (Bogdanowich-Knipp et al., 1999). In addition to a cyclic peptide spanning the full CDRH3 we also created a 16-residue truncated version spanning residues 97 – 100L (Kabat numbering). Truncation sites were determined by visual inspection of the CDRH3 conformation. We used Rosetta to fold both the full-length peptide (C05 WT) and the truncated peptide (C05 truncated WT) over 1,000 simulations and analyzed the folding energy landscape of the peptides, where a peptide likely to fold into its active conformation had low scoring models with low Cα RMSD (Fig. 1B). After folding the wild-type peptide sequence, we used RosettaDesign to optimize the peptide sequence to increase stability in the active conformation. We then refolded the sequence-optimized peptides using molecular dynamics simulation to assess the likelihood that these spontaneously fold into the bioactive conformation. Based on the computational analysis we selected eight redesigned peptides for synthesis and experimental characterization (Fig. 1C).

Figure 1. Experimental workflow of designing CDRH3-derived cyclic peptides.

A. Influenza antibody C05 was chosen due to its long CDRH3 loop involved in antigen recognition. The CDRH3 loop was removed from the antibody and cyclized with a disulfide bond in silico. B. Folding simulations of the full-length and truncated C05 cyclic peptide were performed using Rosetta loop modeling. The wild-type (WT) CDRH3 sequence (blue) was redesigned to stabilize the peptide and improve the energy landscape (red). Favorable energy landscapes have native-like models (low Cα RMSD) with low Rosetta score (lower left quadrant). Score is expressed in Rosetta energy units (REU). C. WT (blue) and redesigned (red) cyclic peptides were synthesized and characterized for binding kinetics to recombinant influenza HA using biolayer interferometry (BLI).

Peptide folding simulations

Starting with the full-length (C05 WT) and truncated CDRH3 (C05 truncated WT) peptides, we folded these peptides in silico to predict their low energy conformations (Fig. 1B). The peptides failed to converge on the active conformation, suggesting that alternative low energy conformations exist that would presumably not function. Next, we redesigned the group of peptide models within 2 Å of the native conformation (36 full-length and 49 truncated models fulfilled this criteria). To redesign the peptide models, we used Rosetta fixed backbone design allowing for all amino acids except cysteine to optimize the amino acid sequence for low energy in the near-native (< 2 Å) conformation, creating 50 designs per model. Although the theoretical amino acid space for redesign is exceedingly large (20²⁶ or 20¹⁶ for full-length or truncated peptides, respectively), the sequences converged after 50 independent design trajectories, each sampling Thousands of energetically favorable sequences. A successful redesign would result in sequences that improved folding into the active conformation – therefore we selected the population of designs improving energy by at least 2 REU from the native sequence (10 and 23 designs for full-length and truncated peptides, respectively) and refolded the redesigned sequence. Out of these models, three full-length and five truncated peptides displayed an improved folding profile compared to the wild-type sequence (Fig. S1). These variants featured a decreased overall Rosetta score compared to the wild-type peptides, suggesting a stabilization of the active conformation. The variants also exhibited a characteristic “funnel” shape in the Rosetta score plots, where a decrease in energy is correlated with low RMSD models, suggesting that they primarily fold into the active conformations. We used a decoy discrimination metric from (Conway, et al. 2014) to quantify the fitness of an energy landscape, where more negative values represent peptides more likely to fold in the desired conformation. In the paper introducing this metric, differences in the discrimination metric as small as 0.02 were considered significant, as this metric is highly stable between simulations (Conway, et al. 2014). We observed variations of 0.04 in the metric between simulations of the same peptide (data not shown). Therefore, we consider changes of the magnitude observed after design (most ranging from −0.09 to −0.4) to be meaningful.

In addition, we modeled these peptide variants in the context of the antibody-antigen interaction. Since the fixed backbone design step was performed in the absence of antigen, several of the mutations that improve conformational stability also affect antigen binding. Therefore, we reverted these mutations one by one, measuring the peptide stability (total score) and binding affinity to antigen (interface score, predicted ΔΔG). Further, we refolded the reverted variants to identify a variant that both exhibited an improved folding profile and retained antigenic contact residues (Fig. S2). This analysis prioritized eight redesigned peptides, one derived from each parent sequence, for further characterization. For brevity, to refer to these antigen-optimized sequences we remove the variant number and use only the design number herein (i.e. d1v1 becomes d1, d2v3 becomes d2, etc.).

Molecular dynamics simulations

To further assess the expected stability of our designed peptides in solution, we performed a total of 85.5 μs of molecular dynamics (MD) simulation on the eight redesigned peptides reverted to maintain antigen binding (listed in Fig. S2), plus the native sequence of full-length and truncated peptides (Table S1). While Rosetta frequently can identify native energy minima de novo, it does not explicitly sample kinetic intermediates within or between energy wells (Baker, 2019; Kaufmann et al., 2010). Because MD simulations are well-suited to characterize local fluctuations, we performed 500 ns simulations of the eight best redesigned peptides in explicit solvent and measured the per-residue heavy atom root-mean-square fluctuations (RMSF) relative to the corresponding full-length or truncated C05 CDRH3 peptide (Fig. S3). Of the three full-length and five truncated peptide designs previously optimized by Rosetta modeling, one full-length (d1; Fig. S3A) and one truncated (d4; Fig. S3B) displayed increased local stability (i.e. reduced RMSF) around the antigen interface site relative to the corresponding C05 CDRH3 peptide. Indeed, comparison of these unbound peptide simulations with 250 ns simulations of the same peptide designs in complex with HA demonstrate near ideal overlap of the peptide antigen binding motif, despite conformational fluctuations in non-interacting N-and C-terminal regions (Fig. S3E – F). To evaluate the global stability of our re-designed peptides, we performed two additional 4.0 μs MD simulations for each unbound peptide and estimated conformational Markov models. Consistent with our RMSF calculations, the probability-weighted per-residue RMSDs of our d1 and d4 peptides to their designed structures demonstrate the most increased stability in the interface region relative to their respective C05 peptides. Interestingly, several other designs display a global stability improvement relative to their respective C05 peptides, in agreement with their Rosetta energy scores (Fig. S3C – D). Taken together, these results predict that designs d1 and d4 are the most conformationally stable peptides and benefit from a reduced entropic cost to binding.

Experimental characterization of redesigned peptides

Based on the in silico modeling of these peptides by two complementary approaches, we synthesized and characterized eight redesigned and two wild-type peptides, four of the full-length and six of the truncated form, cyclized by a disulfide linkage between the N-and C-termini. We then tested binding of these peptides to recombinant influenza HA by biolayer interferometry (BLI). We found that while binding of the wild-type peptides to HA was not detected, two redesigned peptides (one full-length (C05 d1) and one truncated (C05 truncated d4)) displayed high-affinity binding (Fig. 2). These peptides bound to HA proteins of both the H1 and H3 antigenic subtype with an affinity below 100 nM. The remaining six redesigned peptides did not show observable binding (data not shown). To verify that the observed binding was not an artifact of the BLI system, we repeated the binding assay using ELISA on streptavidin-coated plates (Fig. S4A). We observed a clear binding signal for C05 d1 on this system, albeit at lower apparent affinity. C05 truncated d4 showed low but observable binding in ELISA, however the EC₅₀ could not be calculated due to lack of saturation. We attribute this finding to the fact that the truncated peptide lacks the torso of the CDRH3 loop, reducing the total linker distance between the biotin tag and functional portion of the peptide. Therefore, we conclude that peptides C05 d1 and C05 truncated d4 bind reproducibly to influenza HA.

Figure 2. Redesigned cyclic peptides bind with high affinity to group 1 and 2 HAs.

Redesigned cyclic peptides C05 d1 (upper panels) and C05 truncated d4 (lower panels) bind to H1 (B, F) or H3 (C, G) HAs with high (<100 nM) affinity. The wild-type CDRH3 sequence does not bind to H1 HA in either full-length (A) or truncated (E) formats. To identify the peptide epitope, peptides were loaded onto a biosensor, which then was treated with recombinant HA from H1 A/Solomon Islands/03/2006 virus followed by baseline at 90 sec and either a receptor-binding site (C05), stem (CR6261), or irrelevant (MPXV-26) antibody at 120 sec (D, H).

To verify specificity of binding to the receptor-binding site on influenza HA, we performed competition binding on a biosensor using BLI. We first immobilized peptides to the biosensor, then bound recombinant HA followed by either a receptor-binding site antibody (mAb C05), stem binding antibody (mAb CR6261), or irrelevant antibody (MPXV-26, Fig. 2D and H). We observed that, while CR6261 bound the HA tethered to peptide, binding of C05 was not detected, indicating that peptides bind specifically to the receptor-binding site as predicted. The irrelevant antibody showed no association with HA.

To directly compare the affinity of the redesigned peptides to the affinity of C05 IgG, we measured binding to a monomeric head domain of HA from the H1 influenza virus strain A/Solomon Islands/03/2006. We observed an avidity effect from using a trimeric HA that depended on the density of the immobilized molecule (either peptide or IgG), suggesting monomer binding is the most unbiased method to directly compare affinities. The peptides bound monomeric HA head domain with an affinity comparable to that of C05 IgG (Fig. S4B). C05 IgG bound with an affinity of 88 nM, while the peptide affinities were 124 nM or 506 nM for C05 d1 or C05 truncated d4 peptides, respectively. The C05 IgG bound with high on and off rates, consistent with previous work on monovalent binding of C05 (Ekiert et al., 2012), whereas the peptides bound with slower on and off rates.

We repeated binding assays with a linear peptide of the same sequence to test our hypothesis that cyclization increases affinity by stabilization of proper conformation. We reduced the disulfide bond in the peptides before coupling to the biosensor and repeated binding to HA. Although we could still observe low levels of binding with the linear peptides, the affinity was reduced significantly (Fig. S4C), supporting our hypothesis that cyclization does improve peptide activity.

To investigate the biological activity of these peptides, we performed hemagglutination inhibition (HAI) assays with influenza H1 virus. Neither of the peptides showed HAI activity when tested in concentrations up to 50 μM (data not shown). In an effort to improve the avidity of the peptides we preloaded peptides onto streptavidin tetramers and repeated HAI, which also did not show activity when tested in concentrations up to 2.5 μM (data not shown).

Peptides have increased breadth compared to IgG

C05 IgG is known to have unusual breadth for an antibody to the influenza receptor-binding site, binding to both group 1 and 2 HA molecules (Ekiert et al., 2012). After confirming the binding activity of two redesigned peptides, we tested binding to a diverse panel of influenza HA, including strains known to bind or not bind C05 IgG. Remarkably, we found that the redesigned peptides not only maintained the breadth of C05 IgG, but they exhibited increased breadth (Table 1). The peptides recognized new strains within the H1 subtype, increasing breadth to A/Puerto Rico/8/1934 for peptides d1 and d4 and A/California/04/2009 for peptide d4. The binding to A/Puerto Rico/8/1934 is consistent with previous work showing that a computationally redesigned C05 mutant antibody gains binding to this strain (Sevy et al., 2019). In addition, HAs from two new influenza virus A subtypes, H4 and H7, were bound by peptides but not by IgG. Although breadth was increased for several strains, this came at the cost of binding for other subtypes, namely H1 A/mallard/Alberta/35/1976, H3 A/Panama/2007/1999, and H3 A/Perth/16/2009 (for C05 d4). In addition, the affinity was reduced for several strains for which breadth was maintained.

Table 1.

C05-based cyclic peptides have increased breadth of recognition of diverse influenza HA molecules compared to the parental IgG molecule.

Group	Subtype	Strain	C05 d1	C05 d4	C05 IgG
1	H1N1	A/Solomon Islands/03/2006	+++	+++	++++
		A/Solomon Islands/03/2006 head domain	++	++	+++
		A/Brevig Mission/1/1918	-	-	-
		A/Tottori/YKO 12/2011	-	-	-
		A/mallard/Alberta/35/1976	-	-	++
		A/Puerto Rico/8/1934	+++	+++	-
		A/Texas/36/1991	-	-	-
		A/New Caledonia/20/1999	+++	+++	++
		A/California/04/2009	-	++++	-
	H2N2	A/Japan/305/1957	+++	++	++++
		A/Singapore/1/1957	++++	+++	++++
	H5N1	A/Vietnam/1203/2005	-	-	-
		A/lndonesia/5/2005	-	-	-
	H9N2	A/turkey/Wisconsin/1/1966	++++	+++	++
	H16N3	A/black-headedgull/Sweden/4/1999	-	-	-
2	H3N2	A/Hong	+++	+++	++++
		Kong/1/68	-	-	-
		A/Brisbane/10/2007	+++	+++	++++
		A/Perth/16/2009	+++	-	++++
		A/Panama/2007/1999	-	-	++++
		A/Bangkok/1/1979	-	-	-
	H4N6	A/duck/Czechoslovakia/1956	+++	+++
	H7N9	A/Shanghai/02/2013	+++	+++
		A/Netherlands/219/2003	-	-	-
	H15N8	A/shearwater/Western Australia/2576/1979	-	-	-

⁺⁺⁺⁺<10 nM

⁺⁺⁺10–100 nM

⁺⁺100–1,000 nM

^-Specific binding not detected

As a control we performed binding to an irrelevant antigen from HIV (Fig. S4D). We did observe a low level of binding to the irrelevant antigen that was above background signal, indicating that there is a weak nonspecific component to the peptide interaction. However, the signal from specific binding to HA was clearly distinct from the nonspecific signal (Fig. S4D). Therefore, we only considered binding to variant HAs if the signal was at least twice as strong as that to the irrelevant antigen. All K_D, k_on, and k_off values are reported in Table S2.

Structural analysis of peptides

Peptides that showed binding activity, d1 and d4, were highly mutated compared to the wild-type CDRH3 sequence (Fig. 3A), with 11/28 or 10/18 amino acids mutated, respectively. Interestingly, the designed mutations in these peptides converged on the same amino acid at three positions, converging on hydrophobic residues (Fig. 3A). A structural analysis of the peptide models suggests that the Rosetta-designed mutations function by creating hydrophobic patches on the peptide to induce proper folding. Mutations in d1 are predicted to create a patch between Y9, P11, and Y16, and another patch involving L8, Y19, V20, and I21, which both cross opposing strands of the loop and encourage proper loop closure (Fig. 3B). In addition, mutation Q3 is predicted to create a hydrogen bond with the neighboring main chain, and mutations to E24 and R26 are predicted to create an electrostatic interaction in the torso region of the loop (Fig. 3B). In peptide d4, two hydrophobic patches are predicted to form, consisting of residues Y5 and L7, and residues L4, Y12, L15, P16, and L17 (Fig. 3C). Both of these peptides are predicted to fold into conformations similar to that of the C05 CDRH3 loop, with Cα RMSDs of 1.6 and 1.8 Å for d1 and d4, respectively.

Figure 3. Structural analysis of redesigned cyclic peptides.

A. Sequence alignment of redesigned peptides compared to the wild-type C05 CDRH3 sequence. Residues with the same identity as wild-type are shown as dashes. The total number of mutations in each peptide is also shown. B-C. Models of redesigned cyclic peptides (tan) are compared to the C05 CDRH3 structure (gray), and Cα RMSD over all residues in the peptide is shown below. The mutations introduced into the peptide sequence in their structural contexts are highlighted.

Evasion of HA hypervariable elements by peptides

To compare the predicted binding poses of peptides d1 and d4 to that of C05 IgG, we used ROSETTADOCK to model the bound conformation of the peptides in the receptor-binding site of the HA from nine different subtypes (Fig. S5). The peptides docked to H3 A/Hong Kong/1/1968 are predicted to adopt very similar conformations to C05 IgG (Fig. 4A). Peptides d1 and d4 have a Cα RMSD of 1.7 and 2.4 Å, respectively, when aligning the HA component and calculating RMSD of the peptide to the C05 CDRH3 loop. These docked poses therefore agree well with the experimental binding data and suggest that the redesigned peptides mimic C05 in their recognition of HA.

Figure 4. Cyclic peptides contact a minimal epitope on the surface of influenza HA.

A. Models of peptides d1 and d4 (tan) were docked into the receptor-binding site of influenza HA (blue) from H3 A/Hong Kong/1/1968 (PDB ID 4fnk) and compared to the co-crystal structure of C05 IgG (PDB ID 4fp8). Cα RMSD was calculated by superimposing the HA and measuring RMSD over all residues on the peptide compared to the IgG co-crystal structure. B, C. The binding footprint (orange) of either C05 IgG or peptides d1 and d4 was calculated for two antigens for which the peptides have increased breadth, H7 A/Shanghai/02/2013 (B) and H4 A/duck/Czechoslovakia/1956 (C). Buried surface area on the HA surface was calculated and is shown below each structure.

To explain the enhanced breadth of the peptides, we compared the conformation of hypervariable antigenic elements in strains H4 A/duck/Czechoslovakia/1956 and H7 A/Shanghai/02/2013 to the binding pose of C05. We identified two antigenic elements, the loop at position 150 of the HA (150-loop) and the helix at position 190 (190-helix), which are known to influence binding of receptor-binding site antibodies (Lee et al., 2014; Wu et al., 2017; 2018). H7 A/Shanghai/02/2013 has an insertion in the 150-loop compared to H3 A/Hong Kong/1/1968 that directly clashes with the CDRH1 of C05 (Fig. S6A). In H4 A/duck/Czechoslovakia/1956, the 150-loop has amino acid substitutions that clash with the CDRH3 of C05, and the 190-helix has substitutions clashing with the CDRH1 (Fig. S6B). By removing the CDRH1 and introducing mutations into the CDRH3 of the redesigned peptides, we reduced the binding footprint to avoid these antigenic elements.

Although the peptides exhibited increased breadth against subtypes H4 and H7, they also failed to bind to H3 strains A/Panama/2007/1999 and (for peptide d4) A/Perth/16/2009, which are bound by C05 IgG with high affinity. To explain this loss of binding, we analyzed the hypervariable antigenic elements present in these strains. In contrast to the previously discussed H4 and H7 strains, the 150-loop and 190-helix make favorable contacts with the CDRH1 of C05 in these two H3 strains (Fig. S6C). In fact, deletion of the five-residue insertion in the C05 CDRH1 has been previously shown to abolish binding to A/Perth/16/2009 (Ekiert et al., 2012) due to loss of favorable van der Waals interactions between aromatic residues. Therefore, although the reduced binding footprint of CDRH3-mimic peptide can enable binding to novel subtypes, this comes at the cost of binding energy to several existing targets.

Based on our hypothesis that the peptides achieve increased breadth by reducing the binding footprint on the HA surface, we compared the footprint of C05 IgG and peptides d1 and d4 docked to nine HAs of the H1, H2, H3, H4, and H7 subtypes (Table 2). The IgG tended to contact a larger surface area than the peptides as predicted, with average buried surface areas of 730, 627, or 618 Ų for IgG, d1, or d4, respectively. We then compared the buried surface area of peptides docked into the subtypes with increased binding breadth, H4 and H7 (Fig. 4B, C). In the docked models the peptides have a greatly reduced footprint compared to IgG, primarily due to reduced contacts on the 150-loop and 190-helix. The binding footprint of these peptides in the docked conformations therefore represents the minimal binding epitope of HA that is conserved across H1, H2, H3, H4, and H7 subtypes.

Table 2.

Buried surface area on the HA of various subtypes.

	Buried surface area on the HA (A2)
Strain	C05 IgG	C05 d1	C05 truncated d4
H1 A/Solomon Islands/03/2006	639	638	586
H1 A/California/04/2009	777	580	569
H1 A/Puerto Rico/8/1934	734	720	593
H2A/Japan/305/1957	753	596	800
H2A/Singapore/1/1957	731	666	634
H3A/Perth/16/2009	780	686	614
H3 A/Hong Kong/1/1968	651	568	609
H4 A/duck/Czechoslovakia/1956	760	597	628
H7A/Shanghai/02/2013	742	588	534
Average	730	627	618

Mock co-complexes of C05 IgG with each subtype were created by aligning the HA structure to antigen in the co-crystal structure of C05 (PDB ID 4fp8). Peptides C05 d1 and C05 truncated d4 were docked into the receptor-binding site of each HA and the buried surface area of the lowest energy model was calculated.

Although the binding affinity of the peptides to monomeric HA from H1 A/Solomon Islands/03/2006 was comparable to IgG, the peptides associated and dissociated from the antigen much more quickly (Fig. S4B and Table S2). We hypothesized that this difference in association and dissociation rates can be explained by an induced fit mechanism of binding. In this model, the peptides establish new interactions with the antigen through a subtle shift in conformation inaccessible to the more rigid CDRH3 of the IgG molecule. To support this hypothesis, we measured the binding energy and binding density of the antibody-antigen and peptide-antigen interfaces of the docked models (Table S3). We defined binding density as the total binding energy divided by buried surface area. We found that, although the peptides have a reduced buried surface area compared to IgG, they increased the density of binding energy over that surface. This suggests that the peptides can access conformational space unavailable to the IgG molecule to establish new antigen interactions at the binding epitope.

Discussion

Summary of results

In this paper, we show that computational design can be used to engineer antibody CDRH3-based peptides with high affinity and enhanced breadth of recognition for antigenic variants. We designed variants of the full-length and truncated CDRH3 loop of anti-influenza antibody C05 and identified two variants with potent activity. These peptides bound specifically to the influenza receptor-binding site and gained recognition for two HA subtypes, H4 and H7, not bound by the IgG molecule. Models of these two peptides suggest that they bind in a similar orientation to the C05 CDRH3 loop and achieve increased breadth of recognition by avoiding contact with the HA 150-loop and 190-helix hypervariable antigenic elements.

Although the peptides exhibited high-affinity binding on BLI they were not able to inhibit viral hemagglutination. We attribute this lack of activity to several factors. First, BLI may overestimate affinity compared to other kinetic platforms (Yang et al., 2016), therefore the peptides may have lower affinity than reported here. Second, antibodies targeting the influenza receptor-binding site are well known to rely on avidity for their neutralization activity, as many potently neutralizing antibodies show weak binding as monovalent antibody fragments (Fab) (Ekiert et al., 2012; Lee et al., 2014; Schmidt et al., 2013; Whittle et al., 2011). The peptides in this study were monomeric and therefore suffer from the same lack of avidity as a Fab. We repeated HAI assays with peptide loaded onto streptavidin tetramers in an attempt to increase avidity, which also failed to show activity, presumably due to incorrect geometry compared to the HA trimer. In future work, we plan to optimize the geometry of peptide presentation using scaffold proteins, an approach which has been shown to improve neutralizing activity of computationally designed proteins (Strauch et al., 2017).

Cyclic peptide implications

Cyclic peptides have long been pursued as inhibitors of protein-protein interactions (Crook et al., 2017; Owens et al., 2017), modulators of antibody activity (van Rosmalen et al., 2017), mimics of viral antigenic loops (Bird et al., 2014), and antibody mimics (Casset et al., 2003; Kadam et al., 2017; Levi et al., 1993). However, cyclic peptide design has been met with many challenges. Most protein loops do not readily assume their active conformations as peptides, severely limiting the scope of which antibodies can be mimicked by peptides. The work in this study surpasses this limitation using structure-based computational design, by showing that the C05 CDRH3 peptide has no activity when using the wild-type sequence and only functions after introducing designed variations. Instead of limiting the mimicry of antibodies by use of the naturally occurring CDRH3 peptide, we introduce a more systematic approach to create stable peptides from loops based on structure-based design principles. This technology has the potential to be applied to a wide variety of projects involving antibody therapeutics.

Minimal epitope for influenza HA receptor-binding site recognition

We identified a minimal epitope that is conserved across many influenza A subtypes, including H1, H2, H3, H4, H7, and H9. Influenza antibodies, especially those that target the highly conserved receptor-binding site, typically have restricted breadth to a single subtype (Lee et al., 2014; Schmidt et al., 2015; Whittle et al., 2011; Winarski et al., 2015) or to a subset of strains within a subtype (Ekiert et al., 2012; Krause et al., 2012; Lee et al., 2012). This restriction of breadth is due to steric hindrance between antibody residues and hypervariable antigenic elements. The peptides designed in this study are of significant interest since they avoid contact with these variable antigenic elements and target a minimal epitope capable of achieving high-affinity binding against a variety of diverse HA antigens. Identification of this highly conserved epitope can be applied to the design of highly potent small molecule and protein inhibitors (Kadam and Wilson, 2018). The computational methods used to engineer these peptides also can be applied to other systems to identify minimal epitopes required for broad and potent activity.

STAR Methods

Resource Availability

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, James Crowe (james.crowe@vanderbilt.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

The authors declare that the data supporting the findings of this study are available within the paper and its supplementary information files. The code used to perform computational modeling and analyze results during this study are available at https://github.com/sevya/cyclic_peptide_protocol_capture (doi: 10.5281/zenodo.3544793).

Experimental Model and Subject Details

Cells

HA proteins and C05 IgG were expressed by transient transfection of Expi293F human embryonic kidney cells in serum-free medium (ThermoFisher Scientific, cat no. A14527; sex=female).

Shuffle T7 E.coli competent cells were grown in Luria broth supplemented with ampicillin at 30 °C and induced by the addition of 1 mM IPTG overnight at 18 °C. A/Solomon Islands/3/2006 H1N1 (Influenza Reagent Resource, FR-331) were propagated and tittered in monolayer cultures of MDCK cells (ATCC, CCL-34; sex=female).

Method Details

Structure preparation

To generate templates for peptide modeling, we first extracted the CDRH3 loop from the crystal structure of antibody C05 from the Protein Data Bank (PDB, code 4fnl) using PyMol (Schrodinger, LLC, 2015). The loop was renumbered starting from residue 1. For the truncated form, the CDRH3 was truncated based on visual inspection. Cysteines were added to the N and C termini using the PeptideStubMover functionality in Rosetta (see Protocol Capture for details).

Cyclic peptide closure and redesign

Once peptides were processed, they were subjected to loop closure simulations using the Rosetta Generalized Kinematic Closure (GeneralizedKIC) protocol (Bhardwaj et al., 2016; Stein and Kortemme, 2013). Peptides were closed using a residue at the tip of the CDRH3 as the anchor point and perturbed using random perturbations of φ and ψ angles along the loop. Full details of the loop modeling protocol can be found in the Protocol Capture accompanying this manuscript. 1,000 models were generated for each peptide, and the score and Cα RMSD compared to the wild-type loop were calculated.

After modeling wild-type peptides, we redesigned the sequence to increase convergence on the active conformation. We identified all folded loops from the previous simulation that were within 2 Å of the native loop and subjected the top ten by score to Rosetta fixed backbone design (Leaver-Fay et al., 2013). We then simulated folding of the redesigned peptides and calculated score and Cα RMSD compared to the wild-type loop. To identify sequences that converged more readily on the active conformation, we used the funnel discrimination metric described in (Conway et al., 2014).

Binding energy calculations

To ensure the redesign of peptides did not introduce residues that would clash with the antigen, we modeled the designed sequence in the context of the antibody-antigen interface. We modeled the isolated CDRH3 loop (PDB ID 4fnl) in complex with H1 A/Solomon Islands/03/2006 (PDB ID 4hkx).

We created the mock complex of these proteins by aligning to the co-crystal structure of C05 in complex with an H3 antigen (PDB ID 4fp8). We then threaded the mutated residues and refined the structure using Rosetta relax with a 1.0 Å backbone constraint to the starting coordinates. Binding energy (ΔΔG) was calculated as below:

$$\Delta \Delta \text{G }={\mathrm{E}}_{\text{complex}}–\left({\mathrm{E}}_{\text{Ab}}+{\mathrm{E}}_{\text{Ag}}\right)$$

where E_Ab and E_Ag are the energies of the antibody and antigen alone, respectively. When mutated residues affected the binding energy of the complex, we reverted these mutants individually and determined which contributed most to the increase in binding energy. We then modeled revertants with the same loop modeling protocol, and mutants that were favorable in both convergence on the active conformation and binding energy were selected for experimental characterization.

Molecular dynamics simulations

To test whether the cyclic peptides would remain stable in solution, we subjected them to molecular dynamics simulation. As input structures, we used the lowest energy models from the Rosetta loop closure simulations. System parameterization was completed using Leap from AmberTools18. Simulations were performed with Amber18 (Case et al., 2018) in the ff14SB forcefield (Maier et al., 2015). Structures were solvated in a rectangular box of TIP4PEW explicit solvent neutralized with Joung-Cheatham monovalent ions (Joung and Cheatham, 2008; Vega et al., 2006). Peptide was buffered on all sides with 12 Å solvent. Solvent and ions were minimized with 500 cycles of steepest descent followed by 1000 steps of conjugate gradient descent (CGD) while protein atoms were restrained with a force constant of 10.0 kcal/mol/A². The protein then underwent 500 cycles of steepest descent followed by 1000 steps CGD minimization in buffer restrained with a force constant of 5.0 kcal/mol/A². Finally, restraints were removed from the system for 500 steps steepest descent followed by 1000 steps of CGD minimization.

Post-minimization, SHAKE was implemented to constrain covalent bonds to hydrogen atoms. Systems were slowly heated in NVT ensemble to 100K over 50 ps with a 1 fs timestep. Subsequently, systems were heated in NPT ensemble at 1 bar with isotropic position scaling from 100K to 300K over 500 ps and 1 fs timestep. Equilibration/production simulations were run in the NPT ensemble at 300K for either 500 – 4000 ns (unbound peptides) or 250 ns (peptide-HA complexes) with a Monte Carlo barostat and integrated every 2 fs. Temperature was controlled using Langevin dynamics with a collision frequency of 1 ps⁻¹. Periodic boundary conditions were imposed on the system throughout heating and equilibration/production. Electrostatics were evaluated using the Particle Mesh Ewald (PME) method and a distance cutoff of 8.0 Å. RMSF trajectory analysis was performed on the individual 500 ns trajectories using CPPTRAJ (Roe and Cheatham, 2013).

Markov model analysis of molecular dynamics simulations

Markov models were generated with the PyEMMA Python Library version 2.5.7 (Scherer et al., 2015). For each peptide, we combined the two 4.0 μs simulation trajectories, featurized the trajectories by backbone and dihedral angles, and performed time-lagged independent component analysis (TICA) of the featurized trajectories with a lag time of 10 ps. For each peptide, we then constructed a 10-state Markov chain model, resampled 1,000 structures probabilistically from each metastable state, computed the stationary distribution for each metastable state, and finally computed the per-residue RMSD from the target design structure for each metastable state. Per-reside RMSD calculations were performed with CPPTRAJ (Roe and Cheatham, 2013). The per-residue RMSDs were re-weighted by their macrostate’s respective stationary distribution probability, and the final per-residue RMSD is given as the sum of the re-weighted per-residue RMSDs across all 10 metastable states in the Markov model.

Docking

To generate models of the bound peptides the lowest energy peptide models were aligned to the CDRH3 loop position of C05 in PDB ID 4fp8 and saved in complex with the HA from nine different subtypes (PDB IDs 4hkx, 3ubq, 1rvx, 3ku3, 2wr7, 4fnk, 4kvn, 5xl3, and 4ln3). 1,000 models were generated of the two peptides docked into the nine antigens using RosettaDock (Gray et al., 2003).

The protein-peptide interface then was refined using the Rosetta relax protocol (Combs et al., 2013). The ROSETTA score and Cα RMSD was calculated for each model compared to the wild-type C05 CDRH3 loop. Buried surface area calculations were performed using PyMol (Schrodinger, LLC, 2015).

Experimental characterization

Eight redesigned peptide candidates were selected for characterization based on the previously described criteria, along with two control peptides with the wild-type C05 sequence. Peptides were synthesized by Genscript with a disulfide linkage between residues at the N and C termini, and a C-terminal polyethylene glycol (PEG) 6 linker connected to a lysine-linked biotin group. Peptides were synthesized with a purity of >90%.

Recombinant protein expression

Sequences encoding the HA genes of interest were optimized for expression in human cells and synthesized (Genscript). Genes were constructed as soluble trimer constructs by replacing the transmembrane and cytoplasmic domain sequences with a GCN4 trimerization domain and a 6x-His tag at the C-terminus. Synthesized genes were cloned into the pcDNA3.1(+) mammalian expression vector (Invitrogen). HA protein was expressed by transient transfection of Expi293F cells (ThermoFisher Scientific). Supernatants were harvested after 7 days, filter-sterilized with a 0.2-μm filter, and purified using affinity chromatography with a 5 mL HisTrap excel column (GE Healthcare). HA head domain was synthesized as a maltose-binding protein (MBP) fusion in pMAL-c5x vector (New England BioLabs). Head domain was expressed in SHuffle T7 Express competent E. coli (New England BioLabs) to enable disulfide formation in the cytoplasm, induced by the addition of 1 mM IPTG overnight at 18 °C, and purified using amylose resin (New England BioLabs).

Genes encoding C05 heavy and light chains were synthesized (Synthetic Genomics) and cloned into an Ig expression vector (McLean et al., 2000) using the Gibson Assembly Master Mix reagent (New England BioLabs). C05 IgG was expressed by transient transfection of Expi293F human embryonic kidney cells in serum-free medium (ThermoFisher Scientific). The supernatants were harvested after 7 days, filter-sterilized with a 0.2-μm filter, and purified using a 5 mL HiTrap MabSelectSure protein A column (GE Healthcare).

Biolayer interferometry assay

Binding kinetics were determined using biolayer interferometry (BLI) with an Octet Red instrument (FortéBio, Menlo Park, CA). Peptides were loaded onto streptavidin biosensors at 5 μM in kinetics buffer (PBS + 1% BSA, 0.05% Tween 20). The binding experiments were performed with the following steps: 1) baseline in kinetics buffer for 60 s, 2) loading of peptide for 30 s, 3) baseline for 60 s, 4) association of HA for 120 s, and 5) dissociation of HA into kinetics buffer for 300 s. A reference well was run in all experiments, where peptide was loaded onto the biosensor, but antigen was not present, and was subtracted from all sample wells to correct for drift and buffer evaporation. Trimeric HAs were diluted two-fold starting from a concentration of 1.25 μM, and monomeric HA was diluted two-fold from a starting concentration of 20 μM. At least three to four dilutions of HA were used to fit kinetic curves. To eliminate nonspecific effects, the binding curves were compared to binding to an irrelevant antigen (HIV gp120), and binding was only considered significant if the signal was >2x as strong as the irrelevant signal. To test the effect of cyclization on peptide affinity, the peptides were reduced using 2.5 mM TCEP before loading onto the biosensor. Curves were fit to a 1:1 or 2:1 binding model using the FortéBio software. Curve fits were accepted only if they fulfilled an R² of > 0.9. Competition binding was performed with the following steps: 1) baseline for 30s, 2) loading of 5 μM peptide to streptavidin tips for 60 s, 3) baseline for 30 s, 4) addition of HA (H1 A/Solomon Islands/03/2006) at 50 μg/mL for 90 s, 5) baseline for 30s, 6) addition of competitor mAb at 50 μg/mL for 120 s. Small variations in binding response of HA to peptide between different competitor mAb samples result from variation in peptide loading density.

ELISA binding assay

Biotin-labeled cyclic peptides were bound to a pre-coated streptavidin ELISA plate (Streptavidin Coated High Capacity Plates, ThermoFisher Scientific) at 1 μM and incubated for 1 hour at 37 °C. The plates were then blocked with 10% goat serum (Gibco) in PBS for 1 hour at 37 °C. HA monomeric head domain protein was diluted serially 2-fold in blocking buffer at a starting concentration of 14 μM. To detect binding, plates were incubated with a mouse anti-His mAb coupled to HRP (ThermoFisher Scientific). Binding was detected by addition of 100 μL of TMB substrate (ThermoFisher Scientific) and incubated for 5–10 min before quenching the reaction with 100 μL of 1 N HCl. Plates were read at 450 nm using a BioTek plate reader. After plate coating and primary and secondary antibody incubation, plates were washed 3x with wash buffer (PBS +0.05% Tween 20, Cell Signaling Technologies). EC₅₀ values were calculated in GraphPad Prism using robust nonlinear regression. All ELISAs were performed in triplicate. In addition to the ELISA experiments described above, we also attempted an alternate format where HA was coated onto an ELISA plate and peptide binding was detected by a streptavidin-HRP secondary. However, due to peptides binding nonspecifically to the ELISA plate, this format produced high background noise and was unreliable for measuring binding affinity (data not shown).

Viruses and hemagglutination assay

Influenza virus strain A/Solomon Island/3/2006 H1N1 strain was provided by Influenza Reagent Resource of US CDC. The working stocks used for hemagglutination inhibition assay (HAI) were made in MDCK cell culture. For HAI, 25 μL of four hemagglutination units of virus were incubated for 1 hour at room temperature with 25 μL two-fold serial dilutions of peptides starting at 50 μM in PBS. The 50 μL of antibody-virus mixture was incubated for 45 minutes at 4 °C with 50 μL of turkey red blood cells (Rockland) diluted in PBS. The HAI titer was defined as the highest dilution of antibody that inhibited hemagglutination of red blood cells.

Quantification and Statistical Analysis

GraphPad Prism software version 6 was used to determine average values, standard errors, and standard deviations for Fig. S4. ELISA binding curves were fit to the log(agonist) vs. response (3 parameters) equation. Software supplied by the manufacturer (ForteBio) was used to fit BLI curves and determine binding constants and error.

Key Resources Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
C05 IgG	Ekiert et al.2012	NA
6x-His Tag Monoclonal Antibody (HIS.H8), HRP	ThermoFisher	MA1–21315-HRP
Bacterial and Virus Strains
Influenza virus strain A/Solomon Island/3/2006	Influenza Reagent Resource of US CDC	FR-331
SHuffle T7 Express competent E. coli	New England BioLabs	C3026J
Biological Samples
Turkey Red Blood Cells	Rockland Inc.	R408–0050
Chemicals, Peptides, and Recombinant Proteins
Rosetta-designed cyclic peptides	This paper	NA
Recombinant HA protein	This paper	NA
HI Vgp120	This paper	NA
TMB substrate	Thermo Fisher	N301
Critical commercial assays
Gibson Assembly Master Mix	New England BioLabs	E2611S
Deposited Data
Crystal structure of apo C05	Ekiert et al. 2012	PDB ID 4fnl
Crystal structure of H1 A/Solomon Islands/03/2006	Schmidt et al. 2013	PDB ID 4hkx
Co-crystal structure of C05 with H3 A/Hong Kong/1/1968	Ekiert et al. 2012	PDB ID 4fp8
Crystal structure of C05 with H3 A/Hong Kong/1/1968	Ekiert et al. 2012	PDB ID 4fnk
Crystal structure of H3 A/Perth/16/2009	Nakamura et al. 2013	PDB ID 4kvn
Crystal structure of H4 A/duck/Czechoslovakia/1956	Song et al. 2017	PDB ID 5xl3
Crystal structure of H7 A/Shanghai/02/2013	Yang et al. 2013	PDB ID 4ln3
Crystal structure of H1 A/California/04/2009	Xu et al. 2011	PDB ID 3ubq
Crystal structure of H1 A/Puerto Rico/8/1934	Gamblin et al. 2004	PDB ID 1rvx
Crystal structure of H2 A/Japan/305/1957	Xu et al. 2010	PDB ID 3ku3
Crystal structure of H2 A/Singapore/1/1957	Liu et al. 2009	PDB ID 2wr7
Experimental Models: Cell Lines
Expi293F cells	Thermo Fisher Scientific	A14527
MDCK cells	ATCC	CCL-34
Recombinant DNA
pcDNA3.1(+)	Invitrogen	V79020
pMAL-c5x	New England BioLabs	N8108S
McLean Ig expression vector	McLean et al. 2000	N/A
Software and Algorithms
Rosetta	Alford et al.2017	https://www.rosettacommons.org/
Amber18	Case et al. 2018	https://ambermd.org
PyMol	Schrodinger, LLC, 2015	http://pymol.org
GraphPad Prism	GraphPad	https://www.graphpad.com/
Code for data analysis	This study	https://github.com/sevya/cyclic_peptide_protocol_capture
PyEMMA Python library version 2.5.7	Scherer et al., 2015	http://emma-proiect.orq/latest/
CPPTraj	Roe and Cheatham, 2013	https://amber-md.aithub.io/coDtrai/CPPTRAJ.xhtml
ForteBio software	N/A	https://www.fortebio.com/products/octet-svstems-software
Other
HisTrap Excel column (5 mL)	GE Healthcare	17371206
HiTrap MabSelectSure column (5 mL)	GE Healthcare	11003494
Pre-coated streptavidin ELISA plate	Thermo Fisher Scientific	15500

Highlights.

• Cyclic peptides based on antibody loops were modeled and optimized using Rosetta

• Computationally designed peptides bind with high affinity to diverse influenza strains

• Peptides have broad binding activity due to their small footprint on influenza surface