Targeting the Side-Chain Convergence of Hydrophobic α-Helical Hot Spots To Design Small-Molecule Mimetics: Key Binding Features for i, i + 3, and i + 7

Abstract

The conformational convergence of hydrophobic α-helical hot spots was revealed by analyzing α-helix-mediated protein–protein interaction (PPI) complex structures. The pharmacophore models were derived for hydrophobic α-helical hot spots at positions i, i + 3, and i + 7. These provide the foundation for designing generalizable scaffolds that can directly mimic the binding mode of the side chains of α-helical hot spots, offering a new class of small-molecule α-helix mimetics. For the first time, the protocol was developed to identify the PPI targets that have similar binding pockets, allowing evaluation of inhibitor selectivities between α-helix-mediated PPIs. The mimicry efficiency of the previously designed scaffold 1 was disclosed. The close positioning of this small molecule to the additional α-helical hot spots suggests that the decoration of this series of generalizable scaffolds can conveniently reach the binding pockets of additional α-helical hot spots to produce potent small-molecule inhibitors for α-helix-mediated PPIs.

Graphical Abstract

INTRODUCTION

α-Helix accounts for the largest class of protein secondary structures and plays a key role in mediating protein–protein interactions (PPIs).¹ α-Helix mimetics offer a powerful tool to unravel complex signaling networks by modulating specific PPIs and have emerged as an attractive therapeutic strategy to mitigate aberrant PPIs. The peptide truncates of α-helix domains are far less organized in solution and sensitive to proteolysis. Mixed α/β-peptide foldamers can faithfully mimic natural α-helices with superior protease resistance, disrupt α-helix-mediated PPIs, and modulate specific signaling pathways.^2–5 Covalent tethering, including side chain hydrocarbon stapling^6,7 and backbone H-bond mimicry,^8,9 has been developed to constrain peptide conformations and promote α-helicity, thereby enhancing protease resistance and cell penetrance. Hydrocarbon-stapled peptides were shown to disrupt the desired PPIs and exhibit therapeutic effects in vitro and in vivo.^10–14 H-bond surrogates were also reported as PPI antagonists in cell-based and in vivo studies.^15–17

Small-molecule α-helix mimetics are developed to mimic key binding elements of α-helices.^18–20 The advantage of using small-molecule mimetics is that these compounds are metabolically stable, can easily permeate the cell membrane, and are readily absorbed. For many α-helix-mediated PPIs, the residues at positions i, i + 3/4, and i + 7 of α-helices are the projecting hot spots. Hamilton and co-workers reported 3,2′,2″-trisubstituted terphenyls as the first small-molecule scaffold of which the staggered conformation projects three substituents to mimic β-carbon (Cβ) atoms of α-helical side chains at positions i, i + 3/4, and i + 7.²¹ The latter studies revealed the power of this generalizable scaffold, and small-molecule inhibitors were developed for calmodulin/smMLCK,²¹ gp41 oligomeric,²² Bcl-xL/Bak,^23,24 and MDM2/p53 PPIs.^25,26 This class of compounds has been called small-molecule α-helix proteomimetics. Inspired by the success of this scaffold, other important scaffolds have been reported to improve physicochemical properties and synthetic accessibility, such as terpyridines,²⁷ benzoylureas,^28,29 aromatic oligoamides,^30–34 benzamides,³⁵ oligooxopiperazines,^36,37 pyrrolopyrimidines,³⁸ and triazines.³⁹ Some of these small-molecule α-helix proteomimetics have been derivatized to show selectivities for the biological targets,^{25,37,40–42} regulate specific signaling pathways in cell-based studies,^41,43–45 and exhibit in vivo activities in animal studies.⁴⁶ All of these small-molecule α-helix proteomimetics use preorganized scaffolds to overlap with the spiral backbone of α-helices and project three substituents to the areas where the side chains are located.

For PPIs the projecting hot spot of the ligand protein packs against the hot spot pocket of the target protein.⁴⁷ We postulate that the hydrophobic α-helical hot spot needs to stretch out to interact with the hot spot pocket of the target protein. In our earlier work, we observed the limited spatial arrangements of the side chains of hydrophobic α-helical hot spots at positions i, i + 3, and i + 7 when forming PPI complex structures.⁴⁸ On the basis of this observation, a scaffold, 4′-fluoro-N-phenyl-[1,1′-biphenyl]-3-carboxamide (1) in Figure 1, that itself can mimic the binding mode of hydrophobic side chains of three α-helical hot spots at positions i, i + 3, and i + 7 was designed. As the first case study, this scaffold was applied to mimic BCL9 α-helical hot spots Leu366 (i), Ile369 (i + 3), and Leu373 (i + 7) and derivatized into 2 as a β-catenin/BCL9 PPI inhibitor with a K_i of 2.1 μM.⁴⁸ Further optimization led to 3 with a K_i of 0.47 μM.⁴⁹ These early encouraging results have motivated us to develop a series of techniques in this study to analyze the interplay between α-helical hot spots and their corresponding binding pockets, reveal conformation convergence, and identify important binding features at the hydrophobic PPI interface. This knowledge is critical for developing novel small-molecule α-helix mimetics to directly mimic the binding mode of the side chains of α-helical hot spots.

Figure 1.

Our strategy to design small-molecule α-helix mimetics.

RESULTS

Characterization of α-Helical Hot Spots.

We retrieved all PPI complex structures in HippDB^50–52 with hydrophobic α-helical hot spots leucine (Leu), isoleucine (Ile), methionine (Met), valine (Val), phenylalanine (Phe), and tryptophan (Trp) at positions i, i + 3, and i + 7 (Table S1). These PPIs were superimposed based on the backbone non-hydrogen atoms of α-helical hot spots at positions i, i + 3, and i + 7. To characterize the spatial arrangements of these hydrophobic α-helical hot spots, the vectors starting from α-carbon (Cα) atoms to the side-chain centroids (the geometrical center of side-chain non-hydrogen atoms) were devised to represent the orientations of the side chains (Figure 2A). This side-chain centroid vector offers visual inspection and quantification of the spatial arrangement of the side chains of α-helical hot spots by hierarchical clustering (Figure 2C). The superimposition of α-helices showed two side-chain groups (orientations I and II) for α-helical hot spots Leu, Met, Phe, and Trp and one side-chain group for α-helical hot spots Ile and Val, as shown in Figure 2B,C and Figures S1–S3. The side-chain centroid vectors can clearly differentiate these side-chain groups of Leu, Met, Phe, and Trp. Further, the side-chain centroid vectors identify that few α-helical hot spots exhibit the other side-chain orientations (termed off-rotamers), which will be analyzed below. For the side chains of hot spots Leu, Met, Phe, and Trp, the hierarchical clustering results based on side-chain shape and side-chain centroid vectors share an overall 90% similarity (Figure S4), suggesting that the side-chain centroid vectors can represent the overall shape or conformation for hierarchical clustering.

Figure 2.

(A) Side-chain centroid vectors of hydrophobic α-helical hot spots. The geometrical centroids of the side chains were shown as red dots. (B) Orientations of α-helical hot spots Leu at position i. (C) Scatter plot of the side-chain centroid vectors of Leu at position i. (D) Percentages of α-helical hot spots whose side-chain orientations were successfully predicted by the mean dihedral angles of the Dunbrack rotamer library. (E) Independence of the side-chain orientations between α-helical hot spots Leu at positions i, i + 3, and i + 7. α-Helical hot spots were colored in red for those with orientation I and blue for those with orientation II at positions i (left), i + 3 (middle), and i + 7 (right).

The side-chain conformations of residues are known to be backbone-dependent, and the protein side-chain rotamer library has been reported.^53–56 To disclose to what extent the presence of α-helix backbone would affect the side-chain conformations of α-helical hot spots, the dihedral angles of the side chains of α-helical hot spots were calculated and compared with the mean dihedral angle values in the Dunbrack rotamer library.⁵⁶ As shown in Figure 2D, about 90% side chains of α-helical hot spots adopt similar conformations as those of non-hot-spot residues. The most populated side-chain dihedral angles of each α-helical hot spot were close to the mean values of the Dunbrack rotamer library (Figures S5–S8).

The side-chain groups (i.e., orientations I and II) of α-helical hot spots are largely determined by dihedral angle χ₁, while dihedral angles χ₂ and χ₃ affect the side-chain convergence in each group. Dihedral angle χ₁ is restricted by helical backbone. The dihedral angle χ₂ changed along with χ₁ to avoid steric clashes, while the restricted dihedral angle χ₂ in turn influences the value of dihedral angle χ₃.⁵⁵ The restriction brought directly or indirectly from the backbone structures leads to limited numbers of hot-spot side chain groups, with the side chains in each group convergent. The single bond between Cβ and Cγ of α-helical hot spots Leu, Met, Phe, and Trp had two orientations with χ₁ around 180° and 300° (Figure S9A,C,E,F), respectively. The additional methyl groups on Cβ of α-helical hot spot Ile and Val brought in more restrictions of the side-chain conformations by their backbones,⁵⁵ rendering one preferred orientation with χ₁ around 300° for Ile and 180° for Val (Figure S9B,D). This resulted in the dihedral angles between the additional methyl Cγ, Cβ, Cα, and backbone N atoms being around 180° for Ile and 300° for Val, respectively. Therefore, only one side-chain group was observed for hot spots Ile and Val. There were about four dihedral angle groups for Phe (Figure S9E) and Trp (Figure S9F), but owing to the symmetry of the phenyl ring of Phe, the side-chain conformation of χ₂ was identical with that of χ₂ + 180°. Correspondingly, four clusters merged into two. Similarly, two groups of dihedral angle χ₂ of Trp had very close side-chain centroid vectors, and four clusters were convergent into two. In addition, a small group of α-helical hot spots (about 10%) was observed to adopt strained conformations (i.e., off-rotamers), which might be caused by the sufficient binding energy compensation from the hot spot pockets.⁵⁷

The convergent spatial arrangements of α-helical hot spots inspired us to investigate the interdependence of the side-chain orientations between hot spots at positions i, i + 3, and i + 7. The Leu···Leu···Leu subset was used as an example to illustrate the result. As shown in Figure 2E, the orientation of hot spot Leu at position i is independent of that at positions i + 3 or i + 7 and vice versa. We postulated that this independence would lead to the comparable numbers of PPIs adopting either side-chain orientation I or II at positions i, i + 3, and i + 7. This was confirmed when we categorized PPI structures based on the side chain groups of α-helical hot spots (Figure S10).

Analysis of the Binding Pockets of Target Proteins.

The distributions of hydrophobic and hydrophilic atoms in the binding pockets around α-helical hot spots were analyzed to characterize the hydrophobic feature of the binding sites of PPI target proteins. As shown in the left panel of Figure 3A, the average number of hydrophobic atoms was rapidly increased and remained unchanged while the average number of hydrophilic atoms kept increasing, indicating the strong hydrophobic feature of the close contacting regions 2.0–5.0 Å away from α-helical hot spots. A more obvious trend can be observed in the right panel of Figure 3A. The average percentage (i.e., density) of hydrophobic atoms was ≥2-fold higher than that of the hydrophilic atoms within 4.5 Å. The density of hydrophobic atoms began decreasing and was lower than that of hydrophilic atoms in the regions ≥5.5 Å from α-helical hot spots. It should be noted that even though the hydrophobic atoms in the binding pockets were much enhanced within 3.0–4.5 Å from α-helical hot spots, there were still some polar atoms, which are mainly from backbone N and O atoms.

Figure 3.

Distribution of hydrophobic and hydrophilic atoms in the binding pockets of target proteins around (A) all α-helical hot spots, (B) α-helical hot spots with different residue types, and (C) α-helical hot spots Leu in different side-chain orientation groups.

The atom distributions of α-helical hot spots at the individual positions i, i + 3, and i + 7 are shown in Figure S11. As expected, pockets around α-helical hot spots of different positions showed identical atom distribution. The hydrophobic and hydrophilic atoms in the binding pockets were categorized to assess the effect of the individual residue types of α-helical hot spots. Interestingly, the distributions of the average atom density were similar among the binding pockets (Figure 3B, right panel). However, the numbers of the involved atoms are slightly different: the α-helical hot spot with a large side chain, such as Trp, has more binding pocket atoms (Figure 3B, left panel). The atoms in the binding pockets were analyzed to examine whether the distribution of these atoms would be affected by the difference of the side-chain orientations of α-helical hot spots. The distributions of hydrophobic and hydrophilic atoms of the binding pockets have the same patterns between conformational groups I and II (Figure 3C).

The atom distributions of the binding pockets to accommodate α-helical residues at positions i − 1 to i + 8 were also analyzed as the control groups. These residues were categorized based on whether they are hot spots and whether they are hydrophobic residues. As shown in Figure S12, the pockets that accommodate hydrophobic α-helical hot spots at the other positions displayed the same distribution of the binding pocket atoms as that for positions i, i + 3, and i + 7. In contrast, the pockets that accommodate hydrophilic α-helical hot spots exhibited unique enhanced hydrophilic atoms around region 2.5–3.0 Å and enhanced hydrophobic atoms around region 3.0–4.5 Å away from α-helical hot spots (Figure S13). The presence of hydrophobic region 3.0–4.5 Å away from hydrophilic α-helical hot spots keeps dry the hydrophilic atoms within 2.5–3.0 Å and allows strong polar attractive interactions, such as H-bonding, with hydrophilic α-helical hot spots, which explains the high binding affinity brought by the hydrophilic hot spots. The binding regions around α-helical non-hot-spots consist of fewer pocket atoms (Figures S14 and S15). For this group, the differences of the distribution of hydrophobic (3.0–4.5 Å away from α-helical residues) and hydrophilic (2.5–3.0 Å away from α-helical residues) atoms are not obvious.

Apart from hydrophobic interactions, aromatic–aromatic interactions with α-helical hot spots Phe and Trp and H-bond interactions with α-helical hot spot Trp were analyzed. About 81% Phe and 79% Trp bound with their binding pockets without the assistance of aromatic–aromatic interactions (Figure S16A,B, left panels). This percentage is comparable with that of the entire protein structures (Figure S16A,B, middle panels). Similarly, H-bond interactions between α-helical hot spot Trp and the binding pockets are as frequent as in the entire protein structures (Figure S16C, left and middle panels). Further, there is no preference of the directions of these strong directional interactions at the hydrophobic PPI interface (Figure S16, right panels). It was noted that a small fraction of α-helical hot spot Trp (15 out of 95 entries) formed the hydrogen bonds with structural water molecules. However, the directions of these hydrogen bonds were also random. Very few interactions involving charges (e.g., cation–π interaction) were observed in the hydrophobic binding pockets of target proteins.

Pharmacophore Models of the α-Helical Hot Spots at Positions i, i + 3, and i + 7.

As discussed above, the dihedral angles of the singe bonds between Cβ and Cγ atoms of hydrophobic α-helical hot spots all are around 180° and 300° (illustrated in Figure 4A). The side-chain atoms of all hydrophobic α-helical hot spots at positions i, i + 3, and i + 7 are convergent to two spatial areas, allowing deriving the pharmacophore models with acceptable tolerance radii (Figure 4B). The protocol of pharmacophore generation is disclosed and described in the Experimental Procedures. Eight pharmacophore models were then derived to cover two major side-chain orientations at each α-helical position, as shown in Figure 4C. To validate the derived pharmacophore hypotheses toward mimicking the side chains of α-helical hot spots, the percentage of mimicry was calculated. The pharmacophore models have >98% mimicry efficiency when RMSD is set to 1.2 Å (Figure S17).

Figure 4.

(A) Diagrams showing two side-chain groups of α-helical hot spots. The side-chain atoms attaching to the bonds of the same color (red or green) are convergent to the same group. (B) Hydrophobic features of α-helical hot spots at positions i, i + 3, and i + 7. (C) Eight pharmacophore models with geometric parameters labeled. The tolerance radii of the hydrophobic features were set to 1.7 Å.

Identification of α-Helix-Mediated PPIs with Similar Binding Pockets.

Given that the conformational preferences of α-helical side chains determine PPI complex formation,⁵⁷ α-helical hot spots were used as the ligand references. The binding features between α-helical hot spots and its target proteins were detected by IChem triplet interaction fingerprints (TIFPs).^58,59 GRIM⁵⁹ was employed to align the interaction pseudoatoms by the graph matching algorithm and derive the graph-alignment GRIMscore of the clique. IChem TIFPs/GRIM has successfully identified all α-helix-mediated PPIs that have similar binding pockets in our data set. One example is shown in Figure 5. While Mcl-1/Noxa PPI (Figure 5A) and vinculin Vh1/α-actinin VBS PPI (Figure 5B) belong to two different classes of proteins, the binding pockets of the three α-helical hot spots are quite similar. The superimposition of the structures indicated that the PPI contacting atoms are basically overlapped with each other (Figure 5C). The other PPI pairs with the similar local binding sites were listed in Table S2.

Figure 5.

(A) Binding site of Mcl-1/Noxa PPI around α-helical hot spots Leu27, Ile30, and Val34 (colored green). (B) Binding site of vinculin Vh1/α-actinin VBS PPI around α-helical hot spots Leu747, Ile750, and Ile754 (colored green). Interaction pseudoatoms are shown as gray ball-and-stick presentation. (C) Superimposed binding pockets of Mcl-1 (red) and vinculin Vh1 (blue). Key residues are shown as wire representation, and the contacting atoms are shown as ball-and-stick presentation. (D) Probabilities of the presence of different numbers of additional of α-helical hot spot at position i − 1 to position i + 8. (E) Probabilities of the presence of one additional α-helical hot spot at the different positions. (F) Illustration of the substitution patterns of scaffold 1 to reach the additional α-helical hot spots at different positions.

Mimicry Efficiency of the Designed Scaffold 1.

The mimicry efficiency of scaffold 1 was evaluated. As shown in Table S1, 635 α-helices (78%) were successfully mimicked by this scaffold (RMSD < 1.2 Å). For those failed, (1) 86% α-helices contain hot spots Phe and/or Trp and (2) the two pharmacophore models in which the distance between i + 3 and i + 7 is 8.8 Å (as shown in Figure 3C) cannot be mimicked well by 1. An extensive conformational search of 1 indicated that the distance between the centroids of the middle and bottom phenyl rings of 1 cannot reach 8.8 Å.

α-Helical hot spots at positions i, i + 3, and i + 7 are often accompanied by the presence of hot spots at the other positions for PPIs. A survey study on the occurrence of additional hot spots at positions i − 1 to i + 8 in HippDB indicated that 29% PPIs displayed one additional hot spot and 42% PPIs had more than one additional hot spot (Figure 5D). For the PPIs with one additional hot spot at positions i − 1 to i + 8, most of them (83%) had the additional hot spot at positions i − 1, i + 4, i + 6, and i + 8 (Figure 5E). Scaffold 1 was superimposed with α-helical hot spots at positions i, i + 3, and i + 7 to offer possible substitution strategies to mimic the additional hot spots. One example was chosen for the additional hot spot at each position, as shown in Figures S18 and S19 and summarized in Figure 5F. The detailed analyses were described in Notes S2 and S3 in Supporting Information. Scaffold 1 directly occupies the positions of the side chains of α-helical hot spots at positions i, i + 3, and i + 7, which facilitates introduction of substituents to reach additional hot spots at positions i − 1 to i + 4 and positions i + 6 to i + 8. For example, the bottom phenyl rings of the 4-fluorophenyl moiety of 2 and of 3-tetrazolylphenyl moiety of 3 were introduced to mimic the binding mode of BCL9 Phe374 (i + 8) and increase the inhibitory potency.

DISCUSSION AND CONCLUSIONS

Small-molecule α-helix mimetics have the potential to serve as an important strategy to disrupt α-helix-mediated PPIs for drug discovery. The existing small-molecule α-helix proteomimetics are designed to mimic the spatial arrangement of the spiral backbones of α-helices and rely on the substituents to occupy the areas where the side chains of α-helical hot spots are located. While the substituents on these generalizable scaffolds are critical as they mimic α-helical hot spots, the scaffolds themselves are designed to provide skeletal support and are not expected involved in binding. Aiming at providing knowledge for the design of novel small-molecule α-helix mimetics that directly mimic the binding mode of the side chains of α-helical hot spots, we have developed a series of new techniques to analyze the interplay between α-helical hot spots and their corresponding binding pockets and discovered the side-chain convergence of α-helical hot spots at positions i, i + 3, and i + 7 in the PPI complex structures (Figures 2B,C, S1–S3, S9).

The comparison of the dihedral angles of the side chain of α-helical hot spots with that in the Dunbrack rotamer library indicated that the presence of hot spot interactions was not as anticipated the driving force of the side-chain convergence of α-helical hot spots. Instead, the side chains of α-helical hot spots adopt their low-energy conformations in the PPI complex structures, and the binding with the hot spot pockets only accounts for less than 10% discrepancies between the dihedral angle values of α-helical hot spots and the mean values of the Dunbrack rotamer library (Figures 2D). Hence, the high backbone dependence converges the side-chain of α-helices to the limited spatial areas and allows the design of scaffolds to directly mimic the side chains of hydrophobic α-helical hot spots. Further, the comparable numbers of PPIs adopting either side-chain orientation I or II at positions i, i + 3, and i + 7 highlighted the importance of the flexibility of a versatile side-chain α-helix mimetic. Flexible bonds need to be introduced into this series of scaffolds to mimic the side chains with different orientations.

The atomic-level analysis of the binding pockets suggested that hydrophobic atoms were statistically enhanced in the contacting area (Figure 3A), which would account for the strong binding affinity of hydrophobic hot spots. This hydrophobic contacting region is common between the binding pockets for hydrophobic α-helical hot spots at different positions (i, i + 3, or i + 7, Figure S11), for hydrophobic α-helical hot spots with different residue types (Leu, Ile, Met, Val, Phe or Trp, Figure 3B), and for hydrophobic α-helical hot spots with different side-chain conformations (orientation I or II, Figure 3C). It should be noted, however, that the atom-based binding pocket analysis in Figure 3 offers the common hydrophobic binding features but does not disclose the shape and concavity of the binding pockets, as well as the electronic properties. The binding of a specific hydrophobic α-helical hot spot relies on not only common hydrophobic binding features but also the degree of match with the concavity and electronic properties of the binding pockets. Hence, this method does not offer the information on the binding preference of a specific α-helical hot spot. The analysis of aromatic–aromatic and hydrogen bonding interactions indicated that the frequency and directions of these strong directional interactions were not enhanced at the hydrophobic PPI interfaces, allowing the design of generalizable α-helix mimetics to mimic the binding mode of hydrophobic side chains of α-helical hot spots.

The side-chain convergence of hydrophobic α-helical hot spots in PPI complex structures made it possible to superimpose α-helical hot spots of different PPI targets. The protocol to derive the pharmacophore models of the side chains of α-helical hot spots was developed, and the pharmacophore models for α-helical hot spots at positions i, i + 3, and i + 7 were derived (Figure 4). These pharmacophore models provide intuitive insights in the spatial arrangements of the side chains of α-helical hot spots and enlighten the design and assessment of novel small-molecule α-helix mimetics.

Many α-helix-mediated PPIs use the same hydrophobic α-helical hot spots (e.g., Leu···Leu···Leu) to interact with their target proteins. It is important to evaluate the selectivity of small-molecule inhibitors between them. The current practice is the random selection of α-helix-mediated PPIs, sometimes with the guide of biological rationale, to evaluate inhibitor selectivity.^{25,37,40–42} The identification of α-helix-mediated PPIs that have similar binding pockets is of great significance. After assessing different available methods, we noticed that the side chains of α-helical hot spots could serve as the alignment reference to superimpose the local binding sites to compare the similarity of PPI target structures using IChem TIFP/GRIM, and the PPI targets with the similar binding pockets can be identified (Table S2). Generalizable scaffolds, such as 1 in Figure 1, can be used to mimic the key, common binding features of the side chains of α-helical hot spots. The introduction of R_A–R_C groups to this scaffold could trigger inhibitor selectivity by identifying adjacent pockets that only exist in the desired PPI and inducing steric clashes with the unwanted PPIs. The occurrence of polar atoms and aromatic side chains, as described in the analysis of the binding pockets, allows introduction of directional interactions (e.g., hydrogen bonding, π–π interaction, cation–π interaction) between the hot spot pockets of target proteins and the substituents on small-molecule α-helix mimetics (e.g., R_A–R_C in Figure 1) to trigger selectivity for the desired PPI target protein over the unwanted target proteins. The identification of similar binding pockets also facilitates the design of small-molecule inhibitors to modulate a group of desired α-helix-mediated PPIs while leaving the unwanted PPIs unaffected to minimize off-targets.

The conformational search indicated that 78% α-helices were successfully mimicked by this scaffold 1 (RMSD < 1.2 Å). However, α-helical hot spots with large side chains (e.g., Phe and Trp) cannot be readily mimicked, and some α-helices whose hot-spot side chains exhibit long distance between positions i + 3 and i + 7 cannot be reached by scaffold 1. Some of these problems can be solved by the selective substitution of the scaffold, while the mimicry of others would require new scaffolds with more suitable geometries. The mimicry efficiency of scaffold 1 offered the directions for the design of new small-molecule α-helix mimetics.

The presence of the additional hot spots beyond positions i, i + 3/i + 4, and i + 7 often makes the mimicry of α-helices by small-molecule α-helix mimetics more challenging as these hot spots locate on more than one faces of the α-helices.¹ Unlike α-helix proteomimetic scaffolds that mimic the positioning of α-helical spiral backbone structures, scaffold 1 directly occupies the positions of the side chains of the α-helical hot spots at positions i, i + 3, and i + 7. The distance between 1 and the side chains of many additional α-helical hot spots is much closer, which facilitates the introduction of substituents to mimic additional hot spots at positions i − 1 to i + 4 and positions i + 6 to i + 8. Hot spot at position i + 5, despite that it only accounts for about 6% in the presence of additional hot spots, is not easy for 1 to mimic. Multifaced α-helix proteomimetics, on the other hand, have been reported to conveniently mimic this additional hot spot.^60–63

In conclusion, new techniques were developed to analyze the interplay between α-helical hot spots and their corresponding binding pockets. The side-chain centroid vector was devised to represent the orientations of α-helices for hierarchical clustering. The dihedral angles of α-helical hot spots in PPI complex structures were calculated and compared with those in the Dunbrack rotamer library. The side chains of all hydrophobic α-helical hot spots are maximally convergent into two clusters at each position, which is primarily determined by helical backbone restriction. The counting of hydrophobic/hydrophilic atoms in the binding pockets can disclose the common hydrophobic binding features in the binding pockets. The hydrophobic atoms are significantly accumulated while the functional groups for strong directional interactions are not. These results allow designing generalizable scaffolds to directly mimic the binding mode of the side chains of α-helical hot spots, which would offer the development of drug-like small-molecule α-helix mimetics. For the first time the protocol of pharmacophore generation was disclosed for α-helix mediated PPIs, and the pharmacophore models were derived for the PPIs with α-helical hot spots at positions i, i + 3, and i + 7. The use of α-helical hot spots as the reference ligands enables IChem TIFP/GRIM to identify the similar binding pockets between different PPIs. These results will facilitate the design of new small-molecule α-helix mimetics that target the side-chain convergence of α-helical hot spots. The introduction of substituents to these generalizable scaffolds, such as R_A–R_C in Figure 1, can trigger inhibitor potency for example through mimicking additional α-helical hot spots, or trigger inhibitor selectivity by identifying the binding features around the binding pocket that are unique for the desired PPI.

EXPERIMENTAL PROCEDURES

Data Set Preparation.

The HippDB database^50–52 for helical PPI complex structures (https://www.nyu.edu/projects/arora/hippdb/helix.php) was interrogated (89 682 PPI entries with α-helical hot spots on Dec 14, 2018) to retrieve all of the structures that have (1) hydrophobic hot spots at positions i, i + 3, and i + 7 of α-helices (data were obtained from HippDB database); (2) concave pocket(s) in target proteins, where a concave pocket was defined as the presence of at least two pocket residues 4.5 Å around α-helical hot spots; (3) no missing side chain or backbone atoms of target proteins within 6 Å around α-helical hot spots; and (4) ≤ 3 Å resolution of X-ray crystal structures (data were obtained from the PDB). The analyzed hydrophobic residues of α-helices were leucine (Leu), isoleucine (Ile), methionine (Met), valine (Val), phenylalanine (Phe), and tryptophan (Trp). The PPIs whose α-helical hot spots and hot spot pockets contain selenomethionine were included in the data set and joined in the methionine residue group. The repeated α-helical hot spot pattern in one PPI crystal structure was only counted once (e.g., the hot spot group, Leu3, Leu6 and Leu10, in a protein dimer structure was only counted once). Out of all 89 682 PPI structures, 1478 structures have hydrophobic α-helical hot spots at positions i, i + 3, and i + 7. The filtration of the repeated α-helical hot spot pattern in the oligomeric structures afforded 956 PPI complex structures. The application of the concave pocket criteria gave 925. The application of the criteria, no missing side chain or backbone atoms of target proteins within 6 Å around α-helical hot spots, still offered 925 PPI complex structures. The filtration of X-ray crystal structures with resolution if >3 Å resulted in 814 PPI complex structures that meet all of the above requirements.

Protein Structure Preparation.

Crystallographic coordinates of the 814 retrieved PPI complex structures were obtained from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (https://www.rcsb.org/). Preparation of the crystal structures was achieved with Protein Preparation Wizard⁶⁴ in Schrodinger Maestro⁶⁵ and Tripos SYBYL X2.0. Protein structure chains 6 Å away from α-helical hot spots at positions i, i + 3, and i + 7 were removed. If another protein chain was connected with the chain that is involved α-helix-mediated PPIs by the metal atom, all these chains were retained. Small-molecule ligands, water molecules, and metals beyond 4 Å from these retained protein chains were removed. Any missing atoms that are 6 Å away from α-helical hot spots were added by the computer modeling programs: the missing backbone atoms were added using SYBYL X2.0 and the missing side chains were added by Protein Preparation Wizard. The protonation states of the residues were set to pH 7.0 when hydrogens were added. The water molecules beyond 4.00 Å from het groups were deleted. The orientations of the remaining water molecules were sampled at pH of 7.0. Hydrogens were minimized using the OPLS3 force field.^66–69 These PPIs were superimposed based on the backbone non-hydrogen atoms of α-helical hot spots at positions i, i + 3, and i + 7 to minimize the root-mean-square deviation (RMSD).

Hierarchical Clustering of the Side Chains of α-Helical Hot Spots Using Side-Chain Centroid Vectors.

The side-chain centroid vector was defined as the vector starting from α-carbon (Cα) of α-helical hot spots to the geometrical center of the side-chain heavy atoms (non-hydrogen atoms). The coordinates (x, y, z) of each side-chain centroid were used for hierarchical agglomerative clustering in MatLab R2016a.⁷⁰ Hierarchical agglomerative clustering methods build clusters with a hierarchy based on the similarity between objects and generally contain three steps. Step 1: The pairwise distances between all pairs of objects were calculated, and matrix X was generated. Step 2: The linkage algorithm returned a matrix Z that encoded a tree of hierarchical clusters of the rows of matrix X. Step 3: Clusters were generated from the hierarchical agglomerative cluster tree. The Euclidean distance was used as the pairwise distance scale of atom centroids, and the average-linkage algorithm was chosen as the agglomerative technique to construct clusters. The maximum of clusters was determined by visual inspection (Figure 2B,C and Figures S1–S3). The side-chain centroid vectors of α-helical hot spots Leu, Met, Phe, and Trp were classified as two groups, and the side chains of α-helical hot spots Ile and Val were clustered in one group. Hierarchical clustering methods were only performed for the side chain of hot spots Leu, Met, Phe, and Trp.

Because the presence of side-chain centroid outliers usually affects the clustering results, we removed the outliers before clustering. The removal of outliers was based on the algorithm reported by Pais and co-workers,⁷¹ in which the outliers were detected based on the comparison of the average connectivity of objects with the connectivity of the analyzed objects when a distance parameter is given. The objects would be defined as the outliers when its connectivity is lower than one-third of the average connectivity value. The algorithm was stopped when the number of outliers reaches 10% of the whole data set to avoid deleting too many data points. The single linkage method is sensitive to the presence of outliers. Since our data set contains some data points that would still be detected as outliers and affect the clustering result, we decided to use the average linkage method instead of the single linkage method recommended by the outlier detection algorithm.⁷¹

Comparison of Side-Chain Shape and Side-Chain Centroid.

To assess the effectiveness of side-chain centroid vectors for hierarchical clustering, the side chains of α-helical hot spots Leu, Met, Phe, and Trp were hierarchically clustered based on the side-chain shape. The shape parameters were calculated by Phase Shape Screening⁷² in Schrodinger.⁷³ The similarity score S between the side-chain shape of every two residues was calculated. The Shape similarity score S ranged from 0 to 1, with the analyzed hot spot side chain itself having a similarity score of 1. To meet the requirements of hierarchical clustering methods where the most similar data pair should have the closest distance, 1 − S was used to construct matrix X. Then, the distance matrix was clustered using the average linkage algorithm of the hierarchical clustering method in MatLab R2016a. The clustering results were compared with the results clustered by side-chain centroid vectors. It should be noted that 10% of data points were removed as the outliers in the previous hierarchical clustering studies using side-chain centroid vectors. The same data points were excluded in this analysis to ensure the total number of data points in side-chain shape clustering and side-chain centroid clustering is equivalent. The percentage of similarity P was calculated as

$$P(\%)=\frac{N_{\text{sc}1}+N_{\text{sc}2}}{N_{\text{tc}}}\times 100$$

where N_sc1 is the number of data points shared by side-chain shape cluster 1 and side-chain centroid cluster 1; N_sc2 is the number of data points shared by side-chain shape cluster 2 and side-chain centroid cluster 2; N_tc is the total number of data points in side-chain shape cluster 1 plus the total number of data points in side-chain shape cluster 2 (or the total number of data points in side-chain centroid cluster 1 plus the total number of data in side-chain centroid cluster 2). The results are shown in Figure S4.

Comparison of Side-Chain Dihedral Angles.

The dihedral angles χ₁ (for Leu, Ile, Met, Val, Phe, and Trp), χ₂ (for Leu, Ile, Met, Phe, and Trp), and χ₃ (for Met) of the side chains of α-helical hot spots were calculated. The mean dihedral angles provided by the Dunbrack backbone-dependent rotamer library (Simple Mode, 5% step down)⁵⁶ were used as the standard for comparison (http://dunbrack.fccc.edu/bbdep2010/license/index.html). The Dunbrack backbone-dependent rotamer library consists of rotamer frequencies, mean dihedral angles (χ₁₀, χ₂₀, …), and variances (σ₁, σ₂, …) as a function of the backbone dihedral angles (ϕ₀, ψ₀). The successful prediction of the observed α-helical hot spot in our data set by the Dunbrack rotamer library requires that the observed side-chain dihedral angles (χ₁, χ₂, …) of hot spots fall in the range of the mean dihedral angle value (χ₁₀, χ₂₀, …) ± 3 times of the standard deviation (σ₁, σ₂, …), shown as (χ₁₀ ± 3σ₁, χ₂₀ ± 3σ₂, …), when the observed backbone dihedral angles (ϕ, ψ) are close to the calculated dihedral angles (ϕ₀, ψ₀). For example, the observed side-chain dihedral angle of α-helical hot spot Leu is (−70°, 180°) and its backbone dihedral angle is (−63°, − 39°); all rotamers of Leu with backbone dihedral angle (−60°, −40°) in the Dunbrack rotamer library will be searched. Since (−70°, 180°) falls in the range of (−69.2° ± 3 × 6.4°, 172.7° ± 3 × 7.6°), this α-helical hot spot will be considered to be successfully predicted by the Dunbrack rotamer library. The percentage of the predicted α-helical hot spots P_x (x = Leu, Ile, Met, Val, Phe, and Trp) was calculated as

$$P_{\text{x}}(\%)=\frac{N_{\text{ph}}}{N_{\text{th}}}\times 100$$

where N_ph is the number of the predicted hot spots of residue type x in the data set, of which the side-chain dihedral angles (χ₁, χ₂, …) fall into those predicted by the Dunbrack rotamer library. N_th is the total number of α-helical hot spots of residue type x in the data set with the side-chain dihedral angles (χ₁, χ₂, …). The results are shown in Figure 2D.

One potential problem is that even though the side-chain conformations were successfully predicted by the Dunbrack rotamer library, the percentage of each rotamer group might not be close to the predicted rotamer frequency. To address this problem, the percentage of the successfully predicted α-helical hot spots in each rotamer group was compared with the rotamer frequency provided by the Dunbrack rotamer library. For a given backbone dihedral angle (ϕ₀, ψ₀), all rotamers whose backbone dihedral angles (ϕ, ψ) fall near (ϕ₀, ψ₀) were included in the analyzed subset. For these data of each subset, the percentages of all hot spots with side-chain dihedral angles (χ₁, χ₂, …) falling in the range (χ₁₀ ± 3σ₁, χ₂₀ ± 3σ₂, …) were counted, and the percentage of rotamer r in rotamers $P_{\text{r}\left({\varphi }_0,{\psi }_0\right)}$ was calculated as

$$P_{\text{r}}\left({\varphi }_0,{\psi }_0\right)(\%)=\frac{N_r}{{\sum }_i^j{N}_i}\times 100$$

where j is the total group number of rotamers with the backbone dihedral angle (ϕ₀, ψ₀), N_r is the total number of hot spots in rotamer group r, and N_i is the total number of hot spots in rotamer group i.

For example, for residue Leu with a backbone dihedral angle (ϕ, ψ) near (−60°, −40°), the predicted rotamers (χ₁, χ₂) were (−69.2°, 172.7°), (−178.6°, 60.2°), (−176.1°, 153.6°), (−89.2°, 57.3°), (−89.2°, −60.6°), (−174.8°, −78.8°), (70.9°, 84.5°), (71.9°, 165.8°), and (70.3°, −63.0°), and the numbers of hot spots in each rotamer group were 145, 94, 4, 2, 4, 1, 0, 0, and 0, respectively. Hence, the group number was 9, and the percentage of rotamer (−69.2°, 172.7°) in rotamers with backbone dihedral angle (−60°, −40°) was 58.0%, which was compared to the predicted rotamer frequency 64.0%. The results are shown in Figures S5–S8. Because of the restriction of the size of our data set, we did not filter our data points with those requirements used by the Dunbrack rotamer library: (1) the resolution of the crystal structure is better than or equal to 1.8 Å; (2) the R factor of the crystal structure is lower than 0.22; (3) the mutual sequence identity of the chains is 50% or less.⁵⁶ Nonetheless, we still obtained the satisfactory P_x and the similar rotamer distribution (P_r). Further filtration, if permitted by the size of the future data set, is expected to yield the result that is more consistent with Dunbrack’s prediction.

To better understand the side-chain conformations of α-helical hot spots, dihedral angles χ₁ (for Leu, Ile, Met, Val, Phe, and Trp) and χ₂ (for Leu, Ile, Met, Phe, and Trp) were plotted and described in Figure S9.

The interdependence between hot-spot side chains at positions i, i + 3, and i + 7 was analyzed by examining the PPI complex structures with the same α-helical hot spot patterns (e.g., Leu···Leu···Leu as shown in Figure 2E). The PPI complex structures, in which the side chains of α-helical hot spots at positions i, i + 3, and i + 7 were classified in group I or II by hierarchical clustering, were counted. It should be noted that the side chains of α-helical hot spots Ile and Val were counted for both orientation groups. The results are shown in Figure S10.

Hydrophobic Binding Feature Detection.

To analyze hydrophobic features of the binding pockets, we first counted different types of residues in the binding pockets of target proteins. For the contacting residues 4.5 Å around hydrophobic α-helical hot spots,⁷⁴ the frequency of the presence of a specific hydrophobic residue was not found to be affected by the residue type of α-helical hot spots, despite that hydrophobic residues are more frequently observed in this region.

The hydrophobic and hydrophilic atoms in the binding pockets were then analyzed. The binding region of the pocket of the target protein was defined as 6.0 Å away from α-helical hot spots. The atoms in this region were defined as either hydrophobic or hydrophilic based on their partial charge magnitudes reported by Rossky and co-workers.⁷⁵ Only the heavy atoms were analyzed to avoid over-estimation of the hydrophobic or hydrophilic effect brought by the hydrogen atoms. The binding region was further divided into eight subregions based on the distance from α-helical hot spots: 2.0–2.5, 2.5–3.0, 3.0–3.5, 3.5–4.0, 4.0–4.5, 4.5–5.0, and 5.5–6.0 Å. It is worth noting that the region 0–2.0 Å is too close and no pocket atom exists in this region to avoid steric clashes. For 814 PPIs, hydrophobic or hydrophilic atoms in each subregion were counted. The percentages of these atoms were calculated by the equation that the atom numbers of one type (i.e., hydrophobic or hydrophilic) were divided by the total atom number in the analyzed region. For each subregion, the average atom numbers and percentages were calculated and analyzed. The average atom percentages of the whole target protein were calculated. The results are shown in Figure 3A.

α-Helical hot spots at positions i, i + 3, and i + 7 were classified based on their positions (i, i + 3, or i + 7), residue types (Leu, Ile, Met, Val, Phe, or Trp) and side-chain groups (orientation I or II). The classification of α-helical hot spots based on side-chain groups was achieved by hierarchical clustering methods described above. The results are shown in Figure S11 and Figure 3B,C.

To analyze the atom distributions of residues at positions i − 1 to i + 8, residues Gly, Ala, Val, Leu, Ile, Met, Trp, Phe, and Pro were defined as hydrophobic residues, and residues Ser, Thr, Cys, Tyr, Asn, Gln, Asp, Glu, Lys, Arg, and His were treated as hydrophilic residues. The HippDB database was used to determine whether these residues are predicted to be hot spots. These residues were classified as (1) hydrophobic hot spots, (2) hydrophilic hot spots, (3) hydrophobic non hot spots, (4) hydrophilic non hot spots. The results are shown in Figures S12–S15.

Analysis on Aromatic–Aromatic Interactions and Hydrogen Bonding Interactions.

Aromatic–aromatic interactions (Figure 6A,B) and hydrogen bonding interactions (Figure 6C) were counted based on the default criteria of Schrodinger Maestro.⁷⁶ For the edge-to-face aromatic–aromatic interaction (Figure 6A) to occur, the distance between the centroids of two aromatic rings d_cen must be lower than 5.5 Å, and the angle between the ring planes, θ, must be greater than 60.0°. For the face-to-face aromatic–aromatic interaction (Figure 6B), the distance between the centroids of the rings d_cen was defined to be lower than 4.4 Å, and the angle between the ring planes θ must be lower than 30.0°. For the hydrogen bonding interaction (Figure 6C), the distance between the H atom to the acceptor (A) atom, d_DA, was defined to be lower than 2.8 Å, the N–H···A angle θ_D should be greater than 120.0°, and the H···A–B angle θ_A must be greater than 90.0°.

Figure 6.

Geometries of (A) edge-to-face and (B) face-to-face aromatic–aromatic interactions. The normal vector of the aromatic ring was drawn as the red arrow, starting from the centroid of the aromatic ring. The distance between the centroids of two aromatic rings is shown as d_cen, and the angle between the aromatic rings is shown as θ. (C) Geometry of the hydrogen bonding interaction. The acceptor atom is shown as A, and the atom connecting with the acceptor atom is shown as B. The distance from the H atom to the acceptor atom is shown as d_DA. The N–H···A angle is shown as θ_D, which should be greater than 120.0°. The H···A–B angle is shown as θ_A, which should be greater than 90°.

The number of aromatic–aromatic interactions per residue (Phe and Trp) and the number of hydrogen bonding per residue (Trp) were counted for each α-helical hot spot. The number of these interactions per residue was counted for this residue throughout the whole pocket protein and used as the control group. The number of α-helical hot spots that are engaged in none, one, and more than one aromatic–aromatic or hydrogen bonding interaction was calculated. The results are shown in Figure S16.

Generation of the Pharmacophore Models.

The pharmacophore model is to map the most populated area of two groups of α-helical hot spot side chains. To avoid the effects of off-rotamers toward deriving the pharmacophore models, 10% α-helical hot spots in each position (i, i + 3, and i + 7) were treated as outliers and removed from the data set using the outlier detection algorithm described above.⁷¹ Once one side-chain centroid vector of α-helical hot spot at one position (e.g., position i + 3) was detected as the outlier, the residues at the other two positions (e.g., i and i + 7) from the same α-helical structure were also removed.

Out of 814 PPIs, 631 structures were retained to drive the pharmacophore models. The side chains of α-helical hot spots Leu, Met, Phe, and Trp have two groups of orientations based on the locations of their side-chain centroids. Only one group of the side chains was observed for the Cβ-branching α-helical hot spots, Ile and Val. The dihedral angle analysis indicated that the dihedral angles χ₁ of these two residues were both near 180° and 300°, and the Cγ atoms of Ile and Val were located in the same areas as that of α-helical hot spots Leu, Met, Phe, Trp in groups I and II, respectively. Therefore, the side chains of α-helical hot spots Leu, Met, Phe, Trp were classified into two groups based on their side-chain centroids, and the side-chain atoms of Ile and Val were also classified into two groups based on the spatial arrangements of their Cγ atoms (e.g., CG1 atom as group I and CG2 atom as group II).

(1). Determine the location of the pharmacophore features at positions i, i + 3, and i + 7.

To derive the pharmacophore model, the vector starting from atom A with the coordinates (x_A, y_A, z_A) to atom B with the coordinate (x_B, y_B, z_B) is written as V_A→B:

$$V_{\text{A}\to \text{B}}=\left(x_{\text{B}}-x_{\text{A}},y_{\text{B}}-y_{\text{A}},z_{\text{B}}-z_{\text{A}}\right)$$

The distance between atoms A and B can be calculated as d_A→B:

$$d_{\text{A}\to \text{B}}=\left|V_{\text{A}\to \text{B}}\right|=\sqrt{{\left(x_{\text{B}}-x_{\text{A}}\right)}^2+{\left(y_{\text{B}}-y_{\text{A}}\right)}^2+{\left(z_{\text{B}}-z_{\text{A}}\right)}^2}$$

For the α-helical hot spot at position p (p = i, i + 3, or i + 7) and its side chain in the side-chain group g (g = I or II), the side-chain atom vector V_{Cα,p,g→A,p,g} was defined as the vector starting from the Cα atom with the coordinates (x_Cα,p,g, y_Cα,p,g, z_Cα,p,g) toward each side-chain heavy atom A with the coordinates (x_A,p,g, y_A,p,g, z_A,p,g).

The geometrical center vector of all side-chain atoms V_Smean,p,g was used to denote the center of the hydrophobic features of α-helical hot spots at position p in side-chain group g:

$$V_{\text{Smean, }p,g}=\frac{1}{n}\sum _j^n{V}_{\text{C}\alpha ,p,g\to S_j,p,g}$$

where n is the total number of the side-chain atoms of α-helical hot spot at position p in side-chain group g, S_j is the jth side-chain atom. In total, six geometrical center vectors were obtained as V_Smean,i,I, V_Smean,i,II, V_Smean,i+3,I, V_Smean,i+3,II, V_Smean,i+7,I, and V_Smean,i+7,II.

(2). Calculate the Distances between the Pharmacophore Feature Centers at Positions i, i + 3, and i + 7.

For α-helical hot spot at position p in side-chain group g, the Cα atom vector of hot spot at position p was defined as the vector starting from origin O with the coordinates (0, 0, 0) to the Cα atom of hot spot with the coordinates (x_Cα,p,g, y_Cα,p,g, z_Cα,p,g):

$$V_{\text{O}\to \text{C}\alpha ,p,g}=\left(x_{\text{C}\alpha ,p,g},y_{\text{C}\alpha ,p,g},z_{\text{C}\alpha ,p,g}\right)$$

To take into account the variations of the Cα positioning between different α-helix-mediated PPIs, the geometrical center vector of the Cα atom vectors was calculated to identify the center of the Cα atom of hot spots at position p in side-chain group g:

$$V_{\text{C}\alpha \text{mean},p,g}=\frac{1}{n}\sum _j^n{V}_{\text{O}\to \text{C}\alpha ,p,g}$$

where n is the total number of Cα atoms of α-helical hot spots at position p in side-chain group g, S_j is the jth Cα atom. In total, six geometrical center vectors of the Cα vectors were obtained: V_Cαmean,i,I, V_Cαmean,i,II, V_{Cαmean,i+3,I}, V_{Cαmean,i+3,II}, V_{Cαmean,i+7,I}, and V_{Cαmean,i+7,II}.

The pharmacophore center vector, V_pharm,p,g, starting from the origin O toward the center of the hydrophobic features of α-helical hot spots at position p in side-chain group g, can be calculated as

$$V_{\text{pharm},p,g}=V_{\text{Cαmean},p,g}+V_{\text{Smean, }p,g}$$

In total, six pharmacophore center vectors were obtained as V_pharm,i,I, V_pharm,i,II, V_pharm,i+3,I, V_pharm,i+3,II, V_pharm,i+7,I and V_pharm,i+7,II.

The distance between the center of the hydrophobic features at position p₁ in group g₁ and that at position p₂ (p₂ ≠ p₁) in group g₂ is $d_{p_1,g_1\to p_2,g_2}$ as shown in the following formula, which can be calculated as mentioned.

$$d_{p_1,g_1\to p_2,g_2}=\left|V_{p_1,g_1\to p_2,g_2}\right|=\left|V_{\text{pharm },p_2,g_2}-V_{\text{pharm },p_1,g_1}\right|$$

For example, the distance between the center of hydrophobic features at position i in group I and that at position i + 3 in group II is

$$d_{i,\text{I}\to i+3,\text{II}}=\left|V_{i,\text{I}\to i+3,\text{II}}\right|=\left|V_{\text{pharm},i+3,\text{II}}-V_{\text{pharm},i,\text{I}}\right|$$

(3). Define the Tolerance Radius of the Hydrophobic Feature at Each Position.

Since we only included the most populated areas of two groups of side chains and assumed that the larger side chain can be mimicked by adding hydrophobic substituents onto the scaffolds, the tolerance radius of the hydrophobic features was defined as at least 1.7 Å to ensure the inclusion of all Cβ atoms. The results are shown in Figure 4.

(4). Validate the Pharmacophore Model.

The centers of the hydrophobic features in eight pharmacophore models were superimposed with the side-chain centroids of α-helical hot spots. The successful mimicry of the side chains of hot spots in each PPI requires that the RMSD after superimposition of at least one pharmacophore model is lower than the defined RMSD threshold. To give a general idea of the mimicry and avoid overmimicry, the RMSD was set increasing from 0 to 1.2 Å instead of a certain RMSD threshold. The results are shown in Figure S17.

Comparison of the Binding Site Similarity.

The comparison of the binding sites of different PPIs was achieved using software package IChem 5.2.9⁵⁸ (http://bioinfo-pharma.u-strasbg.fr/labwebsite/download.html), which provides the program triplet interaction fingerprints (TIFPs) and the graph matching algorithm GRIM.⁵⁹ TIFPs encode the binding features (hydrophobic, aromatic–aromatic, H-bond donor, H-bond acceptor, positive ionizable, negative ionizable, and metal chelation) between target proteins and their ligands into protein–ligand interacting pseudoatoms (IPAs). IPA can be the interacting ligand atom (the InterLig mode), the interacting protein atom (the InterPro mode), the geometric center of interacting atoms (the Centered mode), or the MERG mode that includes the interacting ligand and protein atoms, and the geometric center of interacting atoms. The IPAs of the reference binding site and the query binding site were aligned by GRIM,⁵⁹ in which all possible matches of the graphs (represented by IPAs as well as their links/distances with other IPAs) of the reference and query binding sites were evaluated. The match is accepted only when the IPAs have the same pharmacophoric features (e.g., H-bond acceptor, etc.) and the same identity (e.g., ligand, protein pseudoatoms, and/or the geometric centers). The largest clique, which refers to the one with the maximal set of possible graph matches, was detected. The graph-alignment GRIMscores (GrSc) of the clique was then derived by the empirical scoring function:

$$\text{GrSc}=0.5006+0.0151N_{\text{lig}}+0.0039N_{\text{Center}}+0.0143N_{\text{Prot}}+0.2098{\text{Sum}}_{\text{Cl}}-0.0720\text{RMSD}-0.0003\text{ DiffI }$$

where N_lig, N_Prot, N_Center are the numbers of the matched IPAs generated based on ligand atoms, protein atoms, and their geometric centers; RMSD refers to the root-mean square deviation of the matched clique; and DiffI is the absolute value of the difference in IPA numbers between reference and query binding sites. The term Sum_Cl is calculated by

$${\text{Sum}}_{\text{Cl}}=\raisebox{1ex}{$\sum \text{ pair weights in clique }$}\!\left/ \!\raisebox{-1ex}{$\sum \text{ all possible pair weights }$}\right.$$

Detailed description of the algorithm can be found the original literature.⁵⁹

In this study, the “ligand” was defined as α-helical hot spots at positions i, i + 3, and i + 7, and the binding sites were defined as the regions 6.0 Å around α-helical hot spots. TIFPs were used to encode the interactions between α-helical hot spots and their corresponding binding pockets into IPAs. Both ligand and pocket were considered in the comparison of the binding sites. Hence, the MERG mode was used to generate three pseudoatoms for each interaction site: the IPAs at the position of the interacting ligand atom, the IPAs at the position of the interacting binding pocket atom, and the IPAs in the middle of the interacting ligand and binding pocket atoms. The default was used for the rest of parameters. The generated pseudoatoms between different PPIs were matched using GRIM at the default setting. Each of the 814 PPIs in the data set was compared with each other. Therefore, 814 × 814 GrSc were generated and sorted, with the most similar binding site pair ranked as the highest. The similar binding sites (GrSc > 0.59 as validated⁵⁹), aligned by GRIM based on the interaction pseudoatoms, were carefully examined to identify the PPIs that have the similar binding pockets. Examples are shown in Figure 5A–C and Table S2.

Assessment of the Mimicry Efficiency of the Previously Designed Scaffold 1.

All possible conformations of scaffold 1 were generated by systemically rotating each rotatable single bond in 1 from 0° to 360° (5° increment). The generated conformers were checked to ensure that the relative energy between the conformation and the global minimum conformation of scaffold 1 (calculated by OPLS3 force field^66–69) is lower than 5 kcal/mol. The filtered conformers were sorted based on their relative energy with the most stable conformations to be assessed first.

The definition of three centroids of scaffold 1 was shown in Figure 1, where the first centroid (mimicking hot-spot side chains at position i) is the geometrical center of fluorophenyl group. The second and third centroids (mimicking hot-spot side chains at positions i + 3 and i + 7) are the geometrical centers of two phenyl groups connected by the amide group. These centroids of the scaffold were superimposed with the side-chain centroids of α-helical hot spots to minimize the RMSD. If the RMSD after superimposition is lower than 1.2 Å, these three α-helical hot spots are considered to be successfully mimicked by the scaffold.

The HippDB database has the information of whether the residues at positions i − 1 to i + 8 are predicted to be α-helical hot spots. The numbers of α-helix structures were counted when the numbers of additional hot spots are 0, 1, 2, 3, and >3, respectively. For α-helix structures with one additional hot spot, the numbers of α-helix structures were counted when the additional hot spot was located at positions i − 1, i + 1, i + 2, i + 4, i + 5, i + 6, and i + 8, respectively. Their percentages of occurrence over the total number of α-helix structures were calculated, as shown in Figure 5D and Figure 5E.

The examples of the mimicry of the additional α-helical hot spot were chosen based on the following requirements: (1) the relative energy between the conformation and the global minimum conformation of scaffold 1 was lower than 5 kcal/mol (calculated by the OPLS3 force field^66–69); (2) the RMSD between the three centroids of scaffold 1 and hot-spot side chains was lower than 1.2 Å; (3) the superimposed scaffold 1 had no steric clash with the corresponding binding pockets of the target proteins. The steric clash was defined by Schrodinger Maestro,⁷⁷ where a clash was considered when the distance between two atoms is lower than 0.75 times the sum of their van der Waals radii. The results are shown in Figures S18 and S19.

ABBREVIATIONS USED

PPI

protein–protein interaction

TIFP

triplet interaction fingerprint

GrSc

GRIMscores

IPA

protein–lignad interacting pseudoatom