De novo design of D-peptide ligands: Application to influenza virus hemagglutinin

Significance

Although D-peptides offer superior metabolic stability and immunogenicity properties compared to L-peptides, their discovery has been limited by current screening methods. We introduce a computational strategy for the de novo design of D-peptides that precisely targets specific protein epitopes, bypassing the need for synthesizing D-enantiomeric proteins. This approach successfully identified multiple D-peptide binders for influenza A hemagglutinin. Our methodology provides a robust alternative for designing stable, nonimmunogenic peptide therapeutics, potentially accelerating drug development against a wide range of targets.

Abstract

D-peptides hold great promise as therapeutics by alleviating the challenges of metabolic stability and immunogenicity in L-peptides. However, current D-peptide discovery methods are severely limited by specific size, structure, and the chemical synthesizability of their protein targets. Here, we describe a computational method for de novo design of D-peptides that bind to an epitope of interest on the target protein using Rosetta’s hotspot-centric approach. The approach comprises identifying hotspot sidechains in a functional protein–protein interaction and grafting these side chains onto much smaller structured peptide scaffolds of opposite chirality. The approach enables more facile design of D-peptides and its applicability is demonstrated by design of D-peptidic binders of influenza A virus hemagglutinin, resulting in identification of multiple D-peptide lead series. The X-ray structure of one of the leads at 2.38 Å resolution verifies the validity of the approach. This method should be generally applicable to targets with detailed structural information, independent of molecular size, and accelerate development of stable, peptide-based therapeutics.

Peptides constitute a very appealing class of drugs that are recognized for their attractive pharmacological properties. According to GlobalData (www.globaldata.com, accessed December 9, 2024), 72 synthetic peptide-based medications are currently marketed in the United States, Europe, and/or Japan, including impactful drugs like Mounjaro/Zepbound for type 2 diabetes/obesity, Pluvicto for prostate cancer, and Goserelin for breast cancer. Moreover, 146 synthetic peptide drugs are in active clinical development, of which 24 are in phase 3 studies (1). Due to their high binding specificity, peptides can attain target selectivity levels that translate into safe and efficacious drugs (2, 3). However, these benefits come at a cost of more complex development that is related to inherently limited oral bioavailability and poor physiologically relevant stability (2–4). The peptide stability issue can be alleviated by using D-stereoisomers of amino acids to construct D-peptides, which cannot be easily degraded or posttranslationally modified. Thus, D-peptides are more likely to succeed downstream in the drug development process with relatively low attrition rates (2, 5–7). D-peptides are less immunogenic compared to L-peptides (8), further increasing their potential as prospective drugs. Given these inherent benefits, peptide discovery programs often seek to identify leads containing D-amino acids; however, existing D-peptide design methodologies have substantial limitations with only 30% of currently approved peptide drugs possessing a D-amino acid component and none being entirely a D-peptide (3).

In the D-scanning method, D-amino acids are often selectively inserted into L-peptides, but the success of such an approach strongly depends on the specific mode of binding, and often only a small number of amino acids can be replaced by D-amino acids without loss of activity (9, 10). Similarly, in the structure-based retro-inverso methodology, inverting the peptide backbone chirality and amino acid sequence may maintain the activity of the D-peptide ligand, but the approach is limited to the design of short linear or β-hairpin D-peptide binders (11–13). Strategies for design of helical D-peptides have been demonstrated and applied to inhibitor design (14–16). Nonproteinogenic amino acids can be incorporated, albeit to a limited extent into phage (17, 18) or mRNA display libraries (19, 20). Furthermore, in the mirror image phage display approach (21, 22) (Fig. 1A), the all-D enantiomer of the target protein is chemically synthesized, folded in vitro, and used for phage display selection. The approach has been very successful in a number of cases (23–27), but the requirement for synthesis of an all-D target is still a substantial limitation (26). Considerable improvements have recently been achieved in chemical synthesis of proteins. Direct manufacturing of peptide chains up to 164 amino acids was demonstrated using a fast-flow automated chemistry platform (28, 29) or using an enzyme-cleavable solubilizing-tag to facilitates the chemical synthesis of mirror-Image proteins (30). Even though certain proteins up to 300 amino acids can be synthesized using native chemical ligation, folding of such long polypeptides often requires a cofactor or a chaperone (31, 32). Physics-based de novo D-peptide design models have been recently developed, focusing on D-peptide scaffold matching followed by subsequent sequence refinement (33). To overcome the challenges faced by current D-peptide discovery methodologies, we report a practical de novo computational strategy for design and screening of D-peptidic ligands.

Fig. 1.

De novo structure-based design of D-peptides. (A) Schematic depiction of the mirror-image phage display methodology (21). [a → b] A target protein is synthesized with D-amino acids and folded. [b → c] Bacterial phages (gray) presenting a library of L-peptides or proteins (orange) are used to identify binders. [c → d] When synthesized with D-amino acids, the binders bind the original target L-protein. (B) Computational de novo D-peptide design methodology. [a] Structure of the target L-protein (blue) in complex with a ligand protein (yellow). [b] The target protein structure is mirror-inverted in silico. A library of L-polypeptide ligands (orange) is designed using different scaffolds to present the original ligand interaction hotspots (yellow). [c] The best ligands are selected based on a scoring function. [d] Corresponding D-peptides are synthesized and tested for binding to the L-target. (C) Blueprint of D-peptide ligand design against influenza A hemagglutinin. bnAb FI6v3 Fab (yellow) is shown in complex with HA (blue) from H1N1 A/California/04/2009 (H1/Cal) (PDB ID 3ZTN) (34). L- and D-enantiomeric disembodied hotspot residues Phe100D (L-Fh and D-Fh) and Trp100F (L-Wh and D-Wh) from FI6v3 HCDR3 are shown in space-filling view (C/O/N in yellow, red, and blue colors) and the peptide scaffold (brown) is in cartoon representation.

Results

De Novo Design, Stability, and Immunogenicity Data of Designed D-Peptides.

In the present approach, the target protein is mirror-inverted in silico into its D-enantiomer, and libraries of L-peptides are designed against the desired epitopes (Fig. 1 B and C). D-enantiomers of the best scoring designs are subsequently synthesized and tested for binding and specificity. The approach is analogous to mirror image phage display (Fig. 1A), but overcomes the need for actual synthesis of D-enantiomeric targets. The applicability of our approach is demonstrated here by computational design of D-peptide ligands against an ~ 220 kDa protein, influenza A hemagglutinin (HA) (Fig. 1C). In order to incorporate ease of synthesis and in vitro folding into our D-peptide designs, we focused on peptide scaffolds with ≤35 amino acids (35).

To validate that such peptide scaffolds indeed exhibit favorable properties of stability and nonimmunogenicity, we compared L- versus D-enantiomers of a panel of peptides [TC10B (Protein Data Bank (PDB) ID 1L2Y), HP35 (PDB ID 3TRW), FSD1(PDB ID 1FSD), Tripzip4 (PDB ID 1LE3), cyc-bCgA^350–362 (PDB ID 1N2Y), and PSP1 (PDB ID 1B1V)] (36) (Fig. 2A and SI Appendix, Table S1). These selected peptides were synthesized in both enantiomeric forms and subjected to a cocktail of proteases in concentrated MDCK cell lysates (Fig. 2B). All D-peptides were entirely stable for 24 h under the experimental conditions, while the corresponding L-peptides were degraded within 2 to 6 h (Fig. 2B). Cyclic peptide cyc-bCgA^350–362 was the only exception where both L- and D-peptide variants were stable for more than 23 h. We then tested the immunogenicity of four peptides (HP35, FSD1, Tripzip4, cyc-bCgA^350–362) in BALB/c mice (Fig. 2C). D-peptides FSD1, Tripzip4, and cys-bCgA formulated with or without alum adjuvant induced virtually no detectable IgG1. Low IgG1 titers were detectable in some of the mice immunized with D-peptide HP35 using CFA/IFA adjuvant, but these titers were considerably lower compared to the corresponding L-peptide (Fig. 2 D and E). Thus, our data confirm that D-peptides indeed are stable and with low to no immunogenicity and, therefore, suitable starting points for design of hemagglutinin targeting therapeutics.

Fig. 2.

Enzymatic stability and immunogenicity of D-peptides. (A) Panel of peptides used for the enzymatic stability study. L- and D-enantiomers of the peptides were subjected to MDCK cell lysate and half-life calculated from degradation profiles (B). The values, reported in hours, are an average of two independent experiments. (B) Degradation profiles of L- and D-peptides in MDCK cell lysate. Colored lines correspond to the D-peptides while gray lines with the same symbols correspond to L-peptides. (C) Schema of the in vivo immunogenicity experiment. (D) IgG1 responses in mice immunized with L- or D-peptides in two different adjuvants (Alum, CFA/IFA). Phosphate-buffered saline (PBS) was used as solvent control. ELISA of the sera of individual mice (n = 5/group) was tested 41 d after the first injection in 1/50 dilution in duplicate. The peptide used for immunization was coated onto 96-well microtiter plates and the bound antibody was detected with peroxidase-conjugated goat anti-mouse IgG1. Mean values per mouse are shown. Animals immunized with L-peptides are grouped on the Left side and animals treated with corresponding D-peptides on the Right side. Sera that scored positive, i.e. the measured OD₄₅₀ value was above a predetermined cut-point (ranging from 0.163 to 0.265), determined with preimmune sera for each peptide separately (SI Appendix, Table S7), are shown in blue. Samples below the cut-point (negative) are shown in gray. (E) Dose–response curves for peptides (FSD1 and HP35) that showed the strongest responses in (panel D). Day 41 sera of mice immunized with FSD1 or HP35 peptides were serially twofold diluted, starting from a 1/50 dilution, and tested in single experiments with the same ELISA as described for panel D. Each colored titration curve corresponds to an individual mouse immunized with the indicated peptide.

D-Peptide Design Against Influenza Hemagglutinin.

To design D-peptide binders that target influenza A HA, we based our approach on the cocrystal structures of complexes of two influenza broadly neutralizing antibodies (bnAbs): CR9114 fragment antigen-binding (Fab) in complex with HA from H5N1 A/Vietnam/1203/2004 (H5/Viet) (PDB ID 4FQI) (37), and FI6v3 Fab with HA from H1N1 A/California/04/2009 (H1/Cal) (PDB ID 3ZTN) (34). In the Fab complex structures, Fab amino acids having more than 80% of the side-chain heavy atoms (including Cα) within 4 Å of the epitope, and having at least 5 such heavy atoms, were identified to select tightly bound hotspot amino acids with high ligand efficiency. The procedure yielded Phe⁵⁴, Phe⁷⁴, and Tyr⁹⁸ from CR9114 and Phe^100D and Trp^100F from FI6v3. The remaining amino acids from the Fabs were subsequently removed and the coordinates of HA and hotspot amino acid were mirror-inverted in silico (Fig. 1C) (n.b. for ease of understanding, only the FI6v3 hotspot residues are shown here as the example of D-peptide design). Since our approach entailed design of L-peptides against a D-protein, chirality of the hotspot’s backbone was locally inverted with mirror-image transformation through a plane spanning the Cα-Cβ-H atoms and only involving the backbone atoms, thereby retaining the bound conformation and interaction of the hotspot side chains with the D-HA protein.

Next, the preselected set of 315 peptide scaffolds (SI Appendix, Table S1) and their multiple conformations were docked on the mirror-inverted HA (D-HA) to select scaffolds complementary in shape with both the target protein and the hotspots (38). To increase the likelihood of a match with the docked scaffolds, multiple conformations of the backbone atoms of the hotspots were generated by using redocking and inverse rotamer generation in Rosetta (38). Scaffolds that could be docked onto the D-HA and able to recapitulate any combination of two hotspots, so-called hits, were subject to a recursive protocol (SI Appendix, Table S2) designed to generate peptide libraries by using RosettaDesign (39) and RosettaDock (40) tools. Within the protocol, peptides were mutated with canonical and noncanonical amino acids (SI Appendix, Table S3) to optimize binding energy and binding efficiency, which was defined as binding free energy per mutation. This latter parameter was introduced to increase the likelihood of correct folding of the designed peptides by biasing toward designs with fewer mutations. For each of the hit scaffold classes, combinations of mutations discovered by the library design protocol were added to the library, whereas designs in which the hotspot amino acids were not reproduced were removed.

D-Peptide Synthesis, Binding, and Specificity.

The resulting library contained 75 D-peptide designs belonging to 10 different peptide scaffold classes and 10 negative controls (wild-type D-peptides) (Fig. 3 A and B and SI Appendix, Table S4). Depending on the scaffold class, 20-80% of the designed peptides were successfully synthesized and folded in vitro (41). and tested for binding against influenza A HA (Fig. 3 B–E and SI Appendix, Figs. S1 and S2 and Tables S4 and S5). Competition ELISAs were conducted against the HA from the 2009 pandemic strain H1N1 A/California/07/2009 (H1/Cal) (Fig. 3 B–D and SI Appendix, Fig. S2 and Table S4) (42). Median inhibitory concentrations (IC₅₀) for the successful designs ranged from 10 to 88 µM with corresponding inhibitory constants (K_i) of 2 to 21 µM (Fig. 3 B and C and SI Appendix, Table S4). Two peptide classes did not result in measurable binding competition. Specificity values were defined and calculated as the ratio K_i^2D1/K_i^HB80.4, where K_i^2D1 is the approximated K_i in competition with HA head binding antibody 2D1 (43) and K_i^HB80.4 with the stem-binding small protein HB80.4 (44) (Fig. 3 B–D and SI Appendix, Table S4). A total of 32 peptides from six scaffold classes, with K_i values below 30 µM and specificity above 10, were defined as hits (Fig. 3C and SI Appendix, Table S4). Analysis of the hits from the top 6 scaffold classes (Fig. 3 B–E and SI Appendix, Fig. S3 and Table S4) showed that 4 scaffolds (1ACWΔC, 2YEN, 1FVY, and 2FQC) were based on the hotspots from antibody CR9114, one 2LJS had hotspots from FI6v3, and scaffold class 1ALG acquired hotspots from both antibodies (Fig. 3B).

Fig. 3.

Design, synthesis, and testing of D-peptide libraries against HA from group 1 influenza A/California/07/2009 (H1N1) virus. (A) Structures of scaffolds for the 10 identified peptide classes are shown in cartoon representation. 1ACWΔC is derived from 1ACW by introducing two mutations (C6A and C10A) and an 8 amino acid C-terminal truncation. (B) Binding competition IC₅₀ and calculated K_i values for selected peptides from the 10 scaffold classes. Different combinations of FI6v3 and CR9114 hotspots were used and between 3-14 peptide variants were synthesized for each scaffold class. (C) Affinity and specificity of all synthesized peptides with different symbols corresponding to different peptide scaffolds. Hit area with affinity (K_i) <30 µM and specificity >10 is colored in yellow. K_i values were estimated from the pIC₅₀’s and based on the concentration and affinity of the small protein HB80.4 that binds to the HA stem (44) using the Cheng–Prusoff equation. Specificity values were defined and calculated as the ratio K_i^2D1/K_i^HB80.4, where K_i^2D1 is the approximated K_i in competition with head binding antibody 2D1 and K_i^HB80.4 with stem binder HB80.4. (D) AlphaLISA curves for DP93 and DP99 in binding competition with HB80.4 and the negative control, HA head binding antibody 2D1. (E) Sequence of selected peptides from six different scaffold classes. Letters represented as uppercase are L-amino acids and lowercase are D-amino acids. Noncanonical amino acids are shown beside the table. H– and –OH in the peptide sequence indicate that N- and C-termini are free and uncapped, whereas H- and -NH₂ in the peptide sequences indicate that N and C termini are uncapped and amidated, respectively.

Crystal Structure of a Designed D-Peptide with Influenza Hemagglutinin.

Based on structural diversity, binding specificity, and hotspot representation from the two different antibodies, peptide DP93 (scaffold 2LJS) was selected for structural characterization by X-ray crystallography (Fig. 3 B and D). Cocrystals that diffracted to 2.38 Å resolution were obtained in complex with H1N1 A/Puerto Rico/8/1934 (H1/PR8) HA (Fig. 4 and SI Appendix, Fig. S4 and Table S6). DP93 binds with a stoichiometry of three per trimer in the highly conserved binding site at the interface of HA1 and HA2 in the HA stem region (Fig. 4 A–C). The binding epitope and orientation of DP93 in the crystal structure are similar to the designed DP93 model and the stem-targeting bnAb FI6v3 (Fig. 4 E–G). Molecular recognition of DP93 involved noncovalent interactions with the HA protein surface (Fig. 4H) and N-linked glycans, which surround the binding site (Fig. 4A and SI Appendix, Fig. S5). N-terminal tag residues D-pyroglutamic acid (dPCA) and D-propargylglycine (π) from DP93 (Fig. 3E) make direct and water-mediated hydrogen bond (H-bond) interactions with Asn²¹ and IIe¹⁸ from HA1 and HA2, respectively (Fig. 4H). Peptide residues dSer⁵ and dTyr²⁷ make similar water-mediated and direct H-bonds with HA1 Ser³⁹ and Thr³¹⁸, and Ser²⁹¹, respectively (Fig. 4H). Further, dIIe⁶ and non-natural amino acid D-homoleucine (b⁷) recognize the HA2 helix-A via water-mediated H-bond interactions with Thr⁴⁹ and Asn⁵³. Along with the intermolecular interactions with HA, DP93 makes series of intramolecular interactions that are mediated via a potassium ion and water molecules (Fig. 4H). Thus, the crystal structure of DP93 in complex with HA verifies that the computational de novo design approach developed here was successful and provides a robust way to precisely and accurately design and efficiently screen for D-peptide ligands against a particular target.

Fig. 4.

Crystal structure of D-peptide DP93 with influenza H1/PR8 HA. (A) The crystal structure of DP93 in complex with influenza group-1 H1 HA from H1N1 A/Puerto Rico/8/1934 (H1/PR8) strain. The HA trimer is represented as a molecular surface with one protomer colored (HA1; pink and HA2; blue) and the other two protomers in whitish gray. DP93 is shown in a tube backbone representation (red) and glycans on the HA surface are in cyan sticks. (B and C) A zoomed-in view of one of three DP93 binding sites in the HA trimer is shown, with the atoms of DP93 in red and S in yellow, respectively. (D) 2D-representation of DP93. The amino acid sequence is represented in lowercase single letters for D amino acids and uppercase for L amino acids. Disulfide bonds are represented in black solid lines and dPCA-π as an N-terminal tag. (E) Computationally predicted binding mode of DP93. All C, and side chain O and S atoms of DP93 are in purple, red, and yellow, respectively. The overall Calpha RMSD between the model and the crystal structure of DP93 is 4.9 Å, mainly as a result of higher RMSDs for peptide regions that do not contact the HA. (F) Superimposition of the interacting side chains of the DP93 computational model (purple) and X-ray crystal structure (red) with HA. The hydrophobic interacting sidechains of the model and crystal structure are generally in close agreement. Sidechains of HA interacting residues of DP93 are shown and highlighted with dotted ellipses. (G) Superimposition of the hotspot phenylalanine residue from FI6v3 Fab of the HA–Fab complex (PDB 3ZTN) and DP93 crystal structure. Hotspot residues dPhe² and ф¹⁹ from DP93 occupy the same conserved hydrophobic pocket on HA as Phe^100D and Trp^100F from HCDR3 of Fab FI6v3. (H) Molecular interactions of DP93 in complex with H1/PR8 HA. Polar interactions are depicted in black dotted lines and measured in Å. HA1 (pink ribbon), HA2 (blue cartoon), water (green sphere), and potassium (orange sphere). Polar interactions were also incorporated into our design workflow, although the designed hydrophobic interactions in (F) dominated as expected for binding to the HA hydrophobic stem region. Abbreviations for D-amino acid residues are as follows: dAla (a), D-alanine; dCys (c), D-cysteine; dAsp (d), D-aspartic acid; dGlu (e), D-glutamic acid; dPhe (f), D-phenylalanine; dHis (h), D-histidine; dIle (i), D-isoleucine; dLys (k), D-lysine; dLeu (l), D-leucine; dMet (m), D-methionine; dAsn (n), D-asparagine; dPro (p), D-proline; dGln (q), D-glutamine; dArg (r), D-arginine; dSer (s), D-serine; dThr (t), D-threonine; dVal (v), D-valine; dTrp (w), D-tryptophan, and dTyr (t), D-tyrosine. Noncanonical D-amino acids abbreviated as dPCA, D-pyroglutamic acid; π, D-propargylglycine; b, D-homoleucine; j, D-aminobutyric acid; ф, D-homo-phenylalanine.

Discussion

We have extended the applicability of the hotspot-centric approach (38, 45) to computational de novo design of D-peptidic ligands. Using this methodology, D-peptidic ligands can be developed that have potential to combine the high specificity of biologics with the stability and nonimmunogenicity of small molecules. The de novo design approach was highly effective in predicting D-peptide hits with promising affinities and selectivities against influenza A HA. 32 peptides (42%) in the designed D-peptide library against H1/Cal HA had affinities below 30 µM and showed specific competition with the stem epitope binder. Six out of ten D-peptide scaffold classes that were successfully synthesized and folded contained at least a single promising D-peptide design, and one D-peptide was cocrystallized with HA to validate the design. The obtained hits are deemed useful starting points for further structure-based maturation (46) to high-affinity HA binders, including current avian/bovine H5N1 HAs (47). Thus, the hotspot-centric de novo design methodology can be used as a reliable alternative to other screening approaches (9–11, 21, 22) to develop D-peptides. Since design of each peptide can be performed in parallel, computation is easily scalable on CPU clusters. Our de novo D-peptide design strategy is generally applicable to targets with high-resolution structural information, independent of their size, and without the need to chemically synthesize the target protein in the D-enantiomeric form. In the case of influenza hemagglutinin, synthesis of its D-enantiomeric form is indeed deemed unfeasible by today’s standards. Thus, this approach can serve as a promising alternative strategy to develop peptide therapeutics especially when the stability or immunogenicity of L-peptide leads are a debilitating issue.

Materials and Methods

Peptide Library Design.

Peptide scaffolds with sequence length of ≤35 amino acids were selected from the Protein Data Bank (PDB) using the following criteria. Transmembrane and posttranslational modified peptides were excluded to avoid solubility, folding, and synthesis-related issues. Peptides without clearly defined secondary structures, such as short linear unstructured polypeptides were also excluded. Additionally, peptides in protein–peptide complexes were also excluded because their bound structures may not correspond to their unbound structures. The final resulting scaffold set contained 315 PDB files (SI Appendix, Table S1), most of which originated from NMR based structure determination and thus contain multiple models. In total, 6749 scaffold models were used for computational peptide library design.

D-peptide Design Against Influenza A Hemagglutinin (HA).

Two broadly neutralizing antibodies FI6v3 Fab in complex with H1N1 A/California/04/2009 (H1/Cal) HA (PDB ID 3ZTN) (34) and CR9114 Fab in complex with H5N1 A/Vietnam/1203/2004 (H5/Viet) HA PDB ID 4FQI) (37) were used for the D-peptide design against HA. In both complexes, amino acids with more than 80% of their side-chain heavy atoms (including Cα) within 4 Å of the epitope and containing at least 5 such heavy atoms were identified as hotspot amino acids, yielding Phe^100D and Trp^100F from FI6v3, and Phe⁵⁴, Phe⁷⁴ and Tyr⁹⁸ from CR9114 (n.b. hotspot residues that could not be independently redocked with Rosetta were also excluded). The design and screening of the computational peptide library were performed against the HA structure from H1N1 A/South Carolina/1/1918 (H1/SC) (PDB ID 3GBN). The 3ZTN and 4FQI structures were superimposed on an H1/SC HA monomer and the antibody Fab structures were removed except for the hotspot residues including backbone. The hotspot residues were redocked on the H1/SC HA structure using RosettaDock (40). The HA hotspot complexes were then mirror inverted. Subsequently, each hotspot backbone was mirror inverted through the plane defined by their respective Cα-Cβ-H atoms. To generate alternate conformations and remove steric clashes, hotspots were redocked using RosettaDock (40), generating hundreds of alternate hotspot poses while maintaining the original hotspot amino acid side-chain – HA interactions. The hotspots from both bNAbs were incorporated into the peptide scaffolds using the Rosetta hotspot-centric approach described by Fleishman et al. (38), where scaffolds are docked onto the epitope and simultaneously matched with the best combination of hotspot residues (38) (script S1and S2). 315 scaffolds were used, each having multiple NMR models. In total, 6,749 models were used for the docking and matching step. We used the MM standard scoring function to incorporate noncanonical amino acids and, for the matching step, we removed the Dunbrack rotamer energy term. A total of 79 scaffolds were docked onto the epitope and were able to connect the hotspots, resulting in 195 unique designs. These in silico hits constituted the initial pool of peptides for the follow-up designs. The hits had an average Rosetta score of -6.7 RU, buried on average of 1,017 Ų of solvent accessible surface area (SASA) in the complex, and had on average five mutations including hotspots.

We designed a six-step algorithm to expand these initial hits into libraries by applying sequence transformations and filters (SI Appendix, Table S2). For each of the steps, the Rosetta score (∆G), represented in Rosetta Units (RU), was the key indicator of improvement and was calculated with a Perl routine (SI Appendix, Table S2). Change of the Rosetta score upon mutation (∆∆G) was calculated as ∆∆G = ∆G_mut − ∆G_parent and used to judge improvements of designs after introducing mutations. In step 1, for each peptide from the design pool, each mutated amino acid was reversed back to parent if the predicted ∆∆G was larger than -0.5 RU. Resulting designs with Rosetta score ≤−7 and not more than 10 mutations were added to the design pool. Step 1 was performed on the whole design pool until the algorithm did not produce any new mutants to be accepted. In step 2, a fixed backbone design (39) with natural and non-natural amino acids (SI Appendix, Table S3) was performed, excluding cysteine, glycine, or proline positions or the termini. All contact residues were allowed to vary and only mutations to cysteine were excluded. Single mutations were accepted and added to the design pool if ∆∆G was smaller than −0.5 RU, solvent accessible surface area, as calculated by GROMACS (48), was larger than 1,100 Ų, and the total number of mutations (N_mut) was not larger than 10. Step 2 was performed on the whole design pool until the algorithm did not produce any new mutants to be accepted (SI Appendix, Table S2). In step 3, for all peptides stored in the design pool, all one, two, and three amino acid combinations were allowed to mutate simultaneously into both canonical and noncanonical amino acids. Cysteine, proline, and glycine positions as well as termini were excluded as before. New designs were only accepted and added to the design pool if ∆∆G was smaller than −1.5 RU, SASA was larger than 1,100 Ų, and the total number of mutations was not larger than 10. In step 4, for all peptides stored in the design pool, the mutation reversal step was performed as in step 1, with the same acceptance criteria. All accepted peptides were added to the design pool. In step 5, for all peptides stored in the design pool, single mutation was performed as in step 2. New designs were only accepted and added to the designs pool if ∆∆G was smaller than -1.0 RU, SASA was larger than 1,100 Ų, and the total number of mutations was not larger than 10. In step 6, all designs with ∆G/ N_mut < −1 RU were removed from the design pool.

During the library design process, all mutations were followed by a redocking step. Only the redocked structures were saved in the pool and used for calculating parameters. After step 6, the library, consisting of 1,681 designs, was analyzed by visual inspection. Scaffold families were aligned and inspected one by one to judge the resulting designs. The main criterion used was whether the designs preserved the introduced hotspot interactions. For some scaffold classes, we did indeed observe mutations at the hotspot sites that did not preserve hotspot interactions. These groups of designs were excluded. Scaffold families not resulting in a single binding pose were rejected. The final peptide library contained 75 designs based on 10 scaffold classes (SI Appendix, Table S4).

Synthesis of D-Peptides.

De novo designed D-peptides were synthesized using Fmoc-based solid-phase peptide synthesis (SPPS) strategies at Pepscan Inc. (www.pepscan.com). After trifluoroacetic acid mediated cleavage and deprotection, peptides were purified by reversed-phase high-performance liquid chromatography (RP-HPLC) and analyzed by liquid chromatography–mass spectrometry (LC–MS). Following purification, peptides were lyophilized and stored at −20 °C prior to use. Analytical data for sets of D-peptides are reported in SI Appendix, Fig. S1. In one family (PDB 2LJS) of D-peptides (e.g., DP93), additional noncanonical amino acids D-pyroglutamic acid (dPCA) and D-propargyl glycine (π) were introduced at the N terminus for ease of synthesis and purification (SI Appendix, Fig. S1).

In Vitro Folding of D-Peptides.

Two different methods were used for in vitro folding of the chemically synthesized D-peptides. Dimethyl sulfoxide (DMSO) based folding was performed by adding a mixture of 100% DMSO (100 μL) and 20 mM Tris buffer pH 8.0 (1.4 mL) to 1 mg of the linear D-peptide. Glutathione (GSH) based folding was performed by adding a mixture of 150 μL glutathione folding buffer (8 mM oxidized glutathione and 10 mM reduced glutathione dissolved in 40 mL 20 mM Tris buffer pH 8.0) and 1.35 mL 20 mM Tris buffer pH 8.0 to 1.0 mg linear D-peptide. The following steps were common for both folding methods. The folding reaction mixtures described above were shaken for 24 h at room temperature, followed by addition of 50 μL of 10% TFA in MQ water, and overnight evaporation in a SpeedVac apparatus at 40 °C. The crude reaction mixtures were resuspended in a mixture of 0.1% TFA in MQ (150 μL) and DMSO (50 μL). Individual SepPak C18 cartridges (100 mg stationary phase) were treated with 0.1% TFA in acetonitrile (1.4 mL) followed by 0.1% TFA in MQ (1.4 mL) through gravity assistance. The resuspended crude reaction mixtures were loaded onto the prepared C18 cartridges, a fresh one for each reaction mixture, and absorbed through gravity assistance. A vacuum-assisted (10 cm Hg) washing step was performed by addition of 0.1% TFA in MQ (1.4 mL). The flow-through, which should not contain peptidic material, was collected for each peptide. A vacuum-assisted (10 cm Hg) elution step was performed by addition of 0.1% TFA in acetonitrile (1.4 mL) and the eluate containing was collected for each peptide. Samples from each elute (20 μL) were diluted with 0.1% TFA in MQ (180 μL) in UV transparent microtiter plates and the absorptions at 215 nm and 280 nm were measured to assess the concentrations of the folded D-peptides. Samples from each elute (25 μL) were diluted with 0.1% TFA in MQ (75 μL) and subjected to HPLC analysis (Waters X-bridge C18 column, linear gradient from 0.1% TFA in MQ to 0.1% TFA in acetonitrile over 25 min, column temperature 35 °C, detection at 214 nm and 280 nm). The remaining elute fractions were evaporated to dryness in a SpeedVac, overnight at 40 °C. From the concentration assessment data, a minimum volume (V_min) was determined to be added to the resulting pellets, which would lead to 10 mM stock concentrations. The volume of DMSO required to achieve a 10 mM stock solution or V_min, whichever was the largest, was added to each individual pellet leading to clear solutions throughout. An amount equal V_min from each individual sample was then transferred to a well in the (leftmost) column 1 of a microtiter plate and diluted with phosphate-buffered saline (PBS) (19× V_min). The remaining columns were filled with 5% DMSO in PBS (10x V_min). A volume equal to 10x V_min was then transferred from column 1 to column 2 followed by mixing, after which this process was repeated from column 2 to 3, up to column 10. A volume equal to 10x V_min was then discarded from column 10 leading to a uniform concentration plate. Insolubility was assessed by measurement of absorption at 650 nm, where a signal above background (5% DMSO in PBS) was judged indicative of scattering through precipitation. In such rare cases where significant precipitate was observed, the plate was spun (3,000 rpm, for 10 min) and the clear supernatants transferred to a fresh plate.

Preparation of Madin–Darby Canine Kidney (MDCK) Cell Lysate.

MDCK cell lysate was prepared by first seeding a 95% confluent culture of MDCK cells in culture medium: DMEM (Gibco) supplemented with 10% fetal bovine serum (Gibco) and L-glutamine) in 16 T175 flasks. The cells were cultured overnight at 37 °C, 10% CO₂. The medium was decanted and cells were washed with PBS (Gibco) after which 2 mL of TrypLE Select (Gibco) was added to each flask. After trypsinization, cells were harvested in culture medium, washed twice with PBS, and cells were resuspended in 20 mL PBS. The collected MDCK cells split into 20 x 1 mL Lysin Matrix M tubes (MP Biomedicals), and the cells were homogenized in an MP FastPrep-24 (4.0 m/s, 5 × 20 s). The cell lysates were three times subjected to snap-freezing (−80 °C) and slow thawing, centrifuged to remove debris (10 min at 13,000 rpm), and the cell lysate was decanted into Eppendorf vials. The lysates were stored at -80 °C until use.

Stability of L- vs D-Peptides Against Proteases in MDCK Cell Lysate.

1.0 mg aliquots of the peptides of interest were dissolved individually in 100 μL PBS pH 7.4 (Gibco) after which 100 μL of MDCK lysate was added. All samples were incubated at 37 °C and 25 μL samples taken at 0, 0.5, 2, 4, 6, and 23-h time points. To all samples, 80 μL of a 4:1 mixture of buffer A (5% acetonitrile, 95% MQ, and 0.1% trifluoroacetic acid) and buffer B (95% acetonitrile, 5% MQ, and 0.1% trifluoroacetic acid) was added immediately. All samples were stored at -20 °C until analysis. The samples were run on the RP-HPLC system equipped with a C18 X-bridge column (Waters) kept at 35 °C employing a 25-min gradient from 100 % buffer A to 100% buffer B at 0.3 mL/min. Absorption at 214 nm was monitored. Chromatograms obtained for the t = 0 h samples were employed to determine the retention time of the uncleaved peptide and to normalize its peak across all subsequent timepoints. The degradation profiles were used to calculate half-life by nonlinear regression (Prism GraphPad 7) (Fig. 2 A and B).

Immunogenicity of D-Peptides.

Immunogenicity studies were performed by AGRO-BIO (La Ferté Saint Aubin, France). A study protocol was approved by an ethical committee at AGRO-BIO and AGRO-BIO Institutional Animal Care Utilization Committee (document AGRO0005). Female BALB/cByJ mice (8 wk old) were used for the immunization and preimmune sera were collected at day −7. Peptides were administered via the intraperitoneal route in 100 µL at 50 µg at day 0 and at 20 µg at day 14 and day 28. Peptides were administered in either PBS, Alum (adjuvant 1; Imject Aluminium, Thermo Scientific) or complete/incomplete Freund’s adjuvants (adjuvant 2; CFA/IFA, Sigma), n = 5 mice per peptide per condition. Serum samples were collected at predose (preimmune) and on days 12, 26, and 41.

LCMS Analysis of Pharmacokinetics/Pharmacodynamics (PK/PD) Data of Peptides.

Peptide samples were prepared in a 4:1 v/v mixture of buffer A (95% MQ, 5% acetonitrile, 0.1% TFA) and buffer B (5% MQ, 95% acetonitrile, 0.1% TFA) and run on a Waters Aquity UPLC/MS system equipped with an XBridge C18 column (130Å, 5 µm, 2.1 mm × 100 mm) at 35 °C. A 10 min gradient of 100% buffer A to 100% buffer B was used with a flow of 0.3 mL/min. Detection was at 214 and 280 nm with an inline Waters ESI-MS detector in positive ionization mode.

ELISA.

ELISAs were performed at Eurofins ADME Bioanalyses SAS (Vergèze, France). Peptides 100 µL (1 µg/mL in PBS pH 7.4) were coated 12 to 72 h at 4 °C in 96-well Nunc-Immuno MaxiSorp plates. Plates were washed three times with PBS, 0.05% Tween-20 (PT), blocked with PBS, 2% bovine serum albumin (BSA) for >1 h at room temperature, washed again three times with PT, and incubated for 2 h at room temperature with diluted mouse serum (1:50 or higher in PBS, 1% BSA). After washing the plates three times with PT, bound IgG1 was subsequently detected with peroxidase-conjugated goat anti-mouse IgG1 (Jackson Immuno Research, 115-35-205) followed by staining with tetramethylbenzidine. To determine the cut-point for negative/positive screening of the serum samples of immunized mice, predose serum samples of individual mice (1:50 diluted in PBS, 1% BSA, n = 5 per peptide) were analyzed to assess the background signal. The cut-point was defined as mean OD + (1.645*SD). Screening of serum samples from day 41 was performed in duplicate in 1:50 dilution in PBS, 1% BSA. Samples were considered positive if the mean optical density (OD) was above the cut-point. Day 41 samples that scored positive in the screening ELISA were subsequently titrated in singlicate and results are reported in Fig. 2 D-E.

Expression and Purification of Influenza A Hemagglutinin (HA).

The HA ectodomain was expressed using a baculovirus expression system as previously described (49, 50). Briefly, HA was fused with gp67 signal peptide at the N terminus and to a biotinylation site, thrombin cleavage site, foldon trimerization domain, and His-tag at the C-terminus. Expressed HA was purified using metal affinity chromatography using Ni- NTA resin. For crystallization studies, the HA was digested with trypsin (New England Biolabs, 5 mU trypsin per mg HA, 16 h at 4 °C) to produce uniformly cleaved HA1/HA2, and to remove the trimerization domain and His-tag. The digested material was purified by gel filtration (GE Healthcare).

Competition ELISA.

To assess binding of the D-peptides to the stem epitope of influenza HA, the D-peptide libraries were screened in an ELISA competition assay. To verify binding specificity, competition with a negative control, HA head binding antibody 2D1, was performed. 96-well half-area high binding plates were coated overnight at 4 °C with 50 µl of 0.5 µg/ml H1/Cal HA (Protein Sciences Corporation). Only PBS was added to column 12. The next day, the plates were washed (3× 150 µl with PBS—0.05% Tween-20) followed by a blocking step for 1 h at room temperature (block buffer: PBS—1% BSA, 0.1% Tween-20). A concentration range (~500 to 1 µM) of the purified folded D-peptides was added to the plates (twofold dilution, 10-steps, 25 µl/well) followed by 25 µl of stem binder HB80.4-flag (2 nM) or head binder 2D1 (0.5 nM). Plates were incubated for 1 h while shaking. Column 11 contained no inhibitor and served as high control (upper limit). Column 12 contained no hemagglutinin and no inhibitor and served as low control (lower limit). Plates were then washed again and secondary antibody (50 µl/well) was added to the plates [anti-flag M2 peroxidase (Sigma, cat A8592-1MG, or anti-human Fc peroxidase (Jackson, Cat 209-035-098)]. Plates were incubated for 1 h while shaking, followed by a plate wash. A volume of 25 µl BM Chemiluminescence ELISA Substrate (POD, Roche 11582950001) was added to all wells followed by an incubation of 3 min while shaking in the dark after which the luminescent signal was read out. Data were analyzed and fitted with a standard four-parameter logistic nonlinear regression model using SPSS statistical software package.

Crystallization and Structure Determination of the D-Peptide DP93–HA Complex.

Gel filtration fractions containing H1N1 A/Puerto Rico/8/1934 (H1/PR8) HA were concentrated to ~10 mg/mL in 20 mM Tris, pH 8.0, and 150 mM NaCl. D-protein DP93 at ~10× molar excess was incubated with H1/PR8 HA for ~1 h at room temperature and centrifuged at 14,000 g for ~2 to 3 min before screening for crystallization. Crystallization screens were set up using the sitting drop vapor diffusion method with our automated CrystalMation robotic system (Rigaku) at TSRI. Within 3 to 7 d, diffraction-quality crystals had grown in crystallization condition 0.2 M tri-potassium citrate, 20% w/v PEG3350 at 4 °C. The resulting crystals were cryoprotected with 5 to 15% ethylene glycol, flash cooled, and stored in liquid nitrogen until data collection. Diffraction data were collected at 100 K on the General Medicine and Cancer Institute’s Collaborative Access Team (GM/CA-CAT) 23IDD beamline at the Advanced Photon Source (APS) at Argonne National Laboratory. The diffraction data were processed with HKL-2000 (51). Initial phases were determined by molecular replacement using Phaser (52) with an HA model from PDB ID 1RU7. Refinement was carried out in Phenix (53), alternating with manual rebuilding and adjustment in COOT (54). The final coordinates were validated using MolProbity (55). Data collection and refinement statistics are summarized in SI Appendix, Table S6. In the D-peptide DP93–HA complex, an extra density blob was assigned to a potassium ion based on the following criteria: coordination number and presence of potassium ion in the crystallization condition (0.2 M tri-potassium citrate, 20% w/v PEG3350 at 4 °C).

Structural Analyses.

Surface areas buried on the HA upon binding of DP93 were calculated with the Protein Interfaces, Surfaces and Assemblies (PISA) server at the European Bioinformatics Institute (56). MacPyMol (DeLano Scientific) was used to render structure figures.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

X-ray structure data have been deposited in PDB (9DXX) (57). All study data are included in the article and/or SI Appendix.