Structure-based evolution of a promiscuous inhibitor to a selective stabilizer of protein–protein interactions

The systematic stabilization of protein–protein interactions (PPI) has great potential as innovative drug discovery strategy to target novel and hard-to-drug protein classes. The current lack of chemical starting points and focused screening opportunities limits the identification of small molecule stabilizers that engage two proteins simultaneously. Starting from our previously described virtual screening strategy to identify inhibitors of 14-3-3 proteins, we report a conceptual molecular docking approach providing concrete entries for discovery and rational optimization of stabilizers for the interaction of 14-3-3 with the carbohydrate-response element-binding protein (ChREBP). X-ray crystallography reveals a distinct difference in the binding modes between weak and general inhibitors of 14-3-3 complexes and a specific, potent stabilizer of the 14-3-3/ChREBP complex. Structure-guided stabilizer optimization results in selective, up to 26-fold enhancement of the 14-3-3/ChREBP interaction. This study demonstrates the potential of rational design approaches for the development of selective PPI stabilizers starting from weak, promiscuous PPI inhibitors.

P roteins interact with other proteins to exert their physiological functions in the context of complex spatiotemporally distributed protein-protein interaction (PPI) networks 1,2 . PPIs are attractive drug targets due to their essential regulation of nearly all cellular processes and, as such, PPI modulation has a vast therapeutic potential [3][4][5][6] . In fact, the inhibition of PPIs has rapidly evolved to the frontlines of modern drug discovery and has significantly extended the druggable genome 7,8 . However, the opposite strategy of PPI enhancement by small molecule stabilizers is underexplored, when in fact this strategy offers unique advantages due to the uncompetitive nature of stabilizers and specificity for a transient complex over the individual proteins [9][10][11][12] .
Whereas immunosuppressants rapamycin, cyclosporine, and FK506, and the antitumor drug paclitaxel have been long used in the clinic [8][9][10] , interest in PPI stabilization as a conceptual strategy has only recently surged, due to the success of synthetically engineered hetero-bifunctional probes (proteolysis-targeting chimera; PROTACs) 13,14 and the revelation of the molecular mechanism of lenalidomide and thalidomide (immunomodulatory drugs; IMiDs®) as PPI stabilizers 15,16 . Nevertheless, the majority of reported PPI stabilizers have been serendipitous discoveries and systematic design, screening, and technology platforms for PPI stabilizer discovery are largely lacking 17,18 . There is thus an urgent need for conceptual strategies for hit finding and rational optimization, empowering PPI stabilization.
Structure-based in silico approaches have proven their value in classical drug discovery [26][27][28][29][30] . Here, a structure-guided virtual screening and molecular docking cascade, employing a known PPI inhibitor class as starting point, is brought forward as a strategy for the identification of PPI stabilizers 3 . We report on a successful in silico screening strategy for stabilization of native PPIs via ligands with a molecular glue mode of action. The starting point for this approach is a screening methodology for the identification of inhibitory phosphonates/phosphates that bind to the phosphoserine/-threonine binding pocket in 14-3-3 and block 14-3-3 PPIs in a widespread manner 31 . Small-molecule stabilizers of the ChREBP/14-3-3 protein complex are identified that indeed engage a composite interface pocket constituted by both protein partners. Our structure-based optimization and two high-resolution X-ray crystal structures reveal a distinct difference in binding modes, enabling stabilatory and weak inhibitory activity of a common phosphonate scaffold to be entirely disconnected, resulting in up to 26-fold and selective PPI stabilization without significant PPI inhibition. These findings thus illustrate the power of this rational approach for future PPI stabilization-based drug discovery.

Results
Docking reveals small-molecule 14-3-3/ChREBP stabilizers. The importance of the phospho-group-both in 14-3-3-binding motifs and PPI inhibitors-directed the selection of chemical starting points to molecules that bind the phospho-accepting pocket of 14-3-3. Phosphate-and phosphonate-based inhibitors typically inhibit 14-3-3/client complexes in the low micro-molar (IC 50~1 -20 μM) range [31][32][33] . It has previously been shown that physiological levels of phosphate anions can furthermore affect the 14-3-3 phospho-interactome, via concentration-dependent dissociation of 14-3-3/client complexes 34 . Whereas phosphateand phosphonate-containing moieties thus generally compete with 14-3-3 target binding, for the 14-3-3/ChREBP complex the phospho-binding pocket is uniquely positioned at the rim of the interface, presenting an opportunity for phosphate-/phosphonate-based PPI stabilization. The two crystal structures of 14-3-3/ ChREBP in the Protein Data Bank (PDB; entries 4GNT and 5F74) served as entry points for the structure-based in silico screen. A phospho-binding pocket-centered receptor grid was generated for the structure of 14-3-3β bound to the α2 helix of ChREBP (Fig. 1a). The first step of the virtual screening procedure selected for a phosphate or phosphonate group by a substructure filter which yielded 869 virtual compounds (of the initial 5,993,085 in the public MolPort database) (Fig. 1b). After additional selection filters for drug-like properties, 471 compounds were subjected to molecular docking into the receptor grid using Glide 35,36 . Hits were additionally docked into the receptor using an induced fit docking protocol, taking conformational changes of amino acid side chains in the active site into account 37,38 . We selected 13 compounds for in vitro testing from the 200 top-ranked docking poses, based on visual inspection, divided among three distinct subclasses; AMP-like structures (class A); and non-AMP-like phosphates (B) and phosphonates (C ; Supplementary Tables 1-3 and Supplementary  Figs. 1-3). Class A including AMP itself did not show stabilization of the 14-3-3/ChREBP interaction in a fluorescence anisotropy assay ( Supplementary Fig. 4). We did observe stabilization by AMP based on ITC data ( Supplementary Fig. 5), validating it as a positive control and in line with literature. In both B-and C subclasses one hit was found to increase 14-3-3/ChREBP binding (1 and 2 with EC 50 values of 0.7 and 45 μM, and ligand efficiency (LE) of 0.28 and 0.32, respectively; Supplementary Fig. 4). The docking poses for 1 and 2 revealed that their phosphate or phosphonate groups were indeed ideally positioned in the basic cavity, constituted by the 14-3-3 Arg-Arg-Tyr phosphoaccepting triad, and interacting with the tryptophan side chain of ChREBP (Fig. 1c). 1 and 2 increased the binding affinity of 14-3-3β for ChREBP in a dose-dependent fashion up to 10-and 4fold, respectively (Fig. 1d).
Nearest neighbor analysis yields an improved stabilizer. The more attractive and synthetically more accessible phosphonatebased scaffold of 2, as compared with the reactive phosphate 1, prompted its chemical optimization to establish a structure activity relationship (SAR). An initial SAR-by-catalog study of eight compounds resulted in an increased PPI stabilization by 3 (EC 50 5.  Table 4), with a 14fold enhancement of the binding affinity of 14-3-3β for ChREBP (Fig. 2c). Shorter linkers to the second phenyl group, without an amide, were inactive (4-8), as was phenylphosphate (9) and a weak inhibitory effect was observed for phenylphosphonate (10) (Fig. 2b). Interestingly, phenylphosphonate-based scaffolds had also surfaced as hits for 14-3-3 in a virtual screen reported by us previously 31 . Whereas the focus of that work was on finding disruptors of the interaction between 14-3-3 and aminopeptidase N (APN), the target pocket appears identical. The fundamental differences between a PPI disruptor-that needs to tightly bind its target protein to compete with protein complex formation, and a PPI stabilizer-that binds a specific pocket at a PPI interface, lie at the basis of potential selectivity for stabilization, especially when targeting promiscuous PPI pockets. We hypothesized this also to be the case for the protein complex constituted of 14-3-3 and ChREBP and aimed to explore selective stabilization of this PPI by exploiting the phospho-pocket at its composite interface.
Crystal structure elucidates molecular glue mode of action. We thus set out to study the molecular mechanism and optimize the stabilizing activity of 3 by obtaining structural insights of its mode of action. The tertiary co-crystal structure of 14-3-3β bound to 3 and the ChREBP peptide was solved by X-ray crystallography ( Fig. 3 and Table 1). The overall complex resembles the previously reported crystal structures for a 14-3-3β dimer with the two antiparallel-binding ChREBP-α2 helices. 3 was indeed found clearly positioned in the phospho-accepting pocket of 14-3-3, interacting with R128 of ChREBP, and K51, R58, R129, and Y130 of 14-3-3 (Fig. 3c). A relevant, additional intramolecular polar interaction was observed for 3 between its amide nitrogen and a phosphonate oxygen, stabilizing its proteinbound state geometry. The phenyl of 3 on one side faces an ensemble of hydrophobic residues of both ChREBP (I120) and 14-3-3 (L218, I219, L174, and L222). R128 of ChREBP 'bridges'  Supplementary Fig. 6).
The crystal structure compared with the docking pose for 3 revealed a different orientation of its phosphonate, which is rotated around its tetrahedral geometry (by~109.5°) with the phenylphosphonate group pointing outward of the 14-3-3 central groove (crystal) versus into it (docking; Supplementary Fig. 7). This directs the orientation of the rest of the molecule in the crystal structure, resulting in optimal nestling in the 14-3-3/ ChREBP interface pocket (Fig. 3d, e), with the second phenyl beneficially engaging the hydrophobic roof of the groove.
Small library of analogs establishes crucial SAR. To study the stabilatory mechanism in more detail, a library around 3 was synthetized and analyzed for SAR. Linker length was found to indeed be essential for the stabilizing activity of 3, as demonstrated by the inactive derivatives with shorter linkers (11,12,(14)(15)(16) and lower EC 50 values for slightly longer linker variants (15 or 72 μM for 13 or 17, respectively) ( Table 2, Supplementary Figs. 8-10). A co-crystal structure was obtained for 14-3-3 bound by 12, one of the inactive short-linker analogs of 3 (Fig. 4a), revealing an identical binding pose to the previously described phosphonatebased inhibitors. Remarkably, with an intermediate linker length (n = 1 for (CH 2 ) n ; as noted in Table 1) for 12 compared with 3 (n = 2) and the reported inhibitors (n = 0), it not only appears to pinpoint the key-determining feature for the mode of action, but additionally hits the 'sweet spot' to turn the switch. A crystallographic overlay of 12 (binding to 14-3-3), with 3 (binding to the 14-3-3/ChREBP binary complex) shows two rotations of the molecules with respect to each other in their orientation in the binding pocket; around the phosphonate and around the central axis of the phenylphosphonate, which drags the side chain around (Fig. 4b), with this turn in binding orientation resembling a molecular switch between the two distinct modes. Further SAR revealed substitutions of the phenylphosphonic moiety were either not tolerated (Me, [18][19][20] or did not significantly enhance the activity (F, [21][22][23]. The second phenyl on the other hand, was hypothesized to provide an interesting opportunity for structure variations, for which substitutions on all positions might result in engaging the 14-3-3 side chains D215, K122, or N175 (Supplementary Fig. 9). Most substitutions analyzed resulted in similar or slightly improved stabilization (24)(25)(26)(27)(28)(29)(30)(31)(32)(33). However, the hydrophobic environment engaged by the phenyl does not tolerate a hydroxyl p-substitution (34 is inactive). Placing the hydroxyl group on the linker, however, was allowed, resulting in identical EC 50 values of 11.4 μM for both enantiomers (36 and 37), which can be explained by their most probable orientation toward the solvent-exposed side as can be observed from the crystal structure. Interestingly, whereas the methylated amide derivative (38) is inactive, removing the amide nitrogen (39) does not result in the same deleterious effect, suggesting its intramolecular hydrogen bond with a phosphonate oxygen is not essential for the molecule's conformation or stabilizing activity. Two derivatives, a p-F-substitution (26) and an o-OCH 2 Ph substitution (30) showed slightly improved stabilization activities, resulting in a cooperative enhancement of the 14-3-3/ChREBP binding affinity of 26-and 22-fold, respectively (Fig. 4c). Considering that characterization of complex stabilization in solution is dependent on the relative concentrations of the binding partners 39 , we collected 2D  titration data to investigate the extent of this effect (Supplementary Fig. 11). Titration of compounds 3, 26, and 30 to constant peptide (100 nM) and varying protein concentrations revealed a mild effect on EC 50 values observed (ranges between 6 and 15 μM; 3 and 4.4 μM, and 2.5 and 8 μM for 3, 26, and 30, respectively, for a protein concentration range of 1-50 μM), whereas no significant effect was observed for different fluorescent peptide concentrations (range 50-500 nM).
Selective stabilization of 14-3-3/ChREBP. The most potent stabilizers 3, 26, and 30, together with the inactive (i.e., no 14-3-3/ChREBP stabilization activity) short-linker analogs 11 and 12 were studied for their PPI modulatory mode of action and selectivity by titrations on 14-3-3β and five representative clientderived peptide motifs (Fig. 4d). These motifs were selected based on their distinct 14-3-3-binding sequences and included mode I/ II (TAZ), mode III (ERα), special mode (p53), and the only other reported non-phosphorylated motif, that of ExoS (Supplementary Table 5 and Supplementary Figs. 12 and 13). First, neither 11 nor 12 was found to influence the binding of p53 and the nonphosphorylated motifs of ChREBP and ExoS to 14-3-3, yet 11 showed PPI inhibition activity for 14-3-3/ERα and 14-3-3/TAZ, which was weaker for 12. The phenyl in 11 was directly coupled to the amide. Extending this distance with either a methyl (12) or ethyl (3) linker switched the activity from a weak 14-3-3 PPI inhibitor to a stronger and selective stabilizer for the 14-3-3/ ChREBP complex. This 'switch-on' effect is illustrated by the ChREBP-specific EC 50 of 9.3 μM observed for 3 and the simultaneous abrogation of inhibitory activity against any of the tested 14-3-3 PPIs (Fig. 4d). 26 and 30 were found to have EC 50 values in the same low μM range for enhancing the binding of ChREBP to 14-3-3 (6.4 and 7.0 μM, respectively), with only very weak inhibitory activity toward mainly TAZ and ERα, with merely hints of inhibition in the high μM range. The inhibitory power toward this set of representative 14-3-3 client motifs is thus absent or much weaker than their stabilizing activity. This demonstrates the highly selective nature of the activity of these compounds by addressing a unique pocket only present in the 14-3-3/ChREBP complex. Together, this data indeed confirms the molecular switch molecular mechanism as observed from the crystal structure (Fig. 4b), and this compound series further shows the evolution and ultimate uncoupling of promiscuous PPI inhibitory toward selective stabilatory activity, based on a highly similar scaffold. This can further be explained by the underlying mechanism of PPI inhibition, driven by the intrinsic affinity of the ligand for 14-3-3 alone to compete with complex formation, compared with cooperative enhancement by binding to a complementary, specific interaction surface, constituted by two protein partners that are engaged simultaneously by a stabilizer.
Here, constitution of a ternary complex results in strong stabilization if the small molecule has a low inherent affinity for one (or ideally both) protein partner(s) (low Kd) and high cooperativity (high alpha factor) as mathematically described in previous work 39,40 .

Discussion
We successfully employed an in silico structure-guided strategy to identify selective 14-3-3/ChREBP PPI stabilizers with a phosphonate-based chemotype. The strategy was steered by our previous success in virtual screening for 14-3-3 binders 31 . Structural analysis and SAR revealed the mode of action for stabilization activity as a cooperative molecular glue, by occupying a complementary PPI interface pocket, simultaneously engaging both protein partners. X-ray crystallography, together with biochemical binding data furthermore showed the evolution Table 2 Structure and activity for analogs of 3. of a compound series which revealed a molecular switch mechanism, dissecting weak and promiscuous PPI inhibition from strong stabilization of a specific interaction. Overall, this study demonstrates the principal interchangeability and relatedness of PPI inhibition and stabilization, serving as an inspiration for further efforts toward taking PPI inhibitors and rationally evolving these into stabilizers, potentially even beyond the 14-3-3 realm. Whereas it is especially relevant for hub proteins, such as 14-3-3, this notion in principle holds for many-if not all-globular peptide binding domains (PBDs) like PDZ, SH2, SH3, WW, WH1, PTB that interact with their partner proteins via short, linear (disordered) peptide motifs. Of these, around 1800 are known today 41 that potentially interact with 100,000 peptide motifs 42 . As such, PPI interfaces-and especially rim-of-the-interface regions-can be targeted by molecules with a potential inhibitor-stabilizer duality that by virtue of small chemical modifications can be directed in either of these two activities. The rational design approach validated here delineates a conceptual entry to the prospective discoveries of small molecules as stabilizers of native protein-protein interactions, which empowers future targeting of hard-to-drug proteins and pathways.

Methods
Compound discovery and synthesis. See Supplementary Methods for detailed descriptions of the virtual screening and molecular docking procedures, organic synthesis and characterization.
Purity and exact mass were determined ( Supplementary Fig. 14) using a highresolution liquid chromatography coupled with mass spectrometry (LC/MS) system comprised of an I-Class Acquity UPLC (Waters) with a Polaris C18A reverse-phase column 2.0 × 100 mm (Agilent), coupled to a Xevo G2 Quadrupole Time of Flight mass spectrometer (Waters). A flow rate of 0.3 mL min −1 was used with a gradient of acetonitrile + 0.1% formic acid (FA) in water + 0.1% FA (acetonitrile 15-75%). Deconvolution of the m/z spectra was done using the MaxENT I algorithm in the Masslynx v4.1 (SCN862) software.
Peptide synthesis. The ChREBP-derived α2 peptide (residues 117-142) was synthesized via Fmoc solid phase peptide synthesis on a TentaGel R Ram resin (Novobiochem; 0.20 mmol/g loading) using an Intavis MultiPep RSi peptide synthesizer. Briefly, Fmoc-protected amino acids (Novabiochem) were dissolved in Nmethyl-2-pyrrolidone (NMP, 4.2 eq., 0.5 M) and coupled sequentially to the resin    Protein crystallography, X-ray data collection, and refinement. 14-3-3σ ΔC/12. 14-3-3σ ΔC protein was dissolved in crystallization buffer (CB; 20 mM HEPES pH 7.5, 2 mM MgCl 2 , 2 mM βME) and mixed in a 1:2 molar stoichiometry with compound 12 (100 mM stock in DMSO) to a final protein concentration of 12.5 mg/mL. This was set up for sitting-drop crystallization in a 1:1 ratio in crystallization condition (CB; 0.095 M HEPES pH 7.1, 27% (v/v) PEG 400, 0.19 M CaCl 2 , 5% (v/v) glycerol). Crystals grew at 4°C within 4 days. Crystals were fished and flash-frozen in liquid N 2 . Diffraction data were collected on in-house X-ray diffraction system (equiped with Rigaku MicroMax-003 sealed tube X-ray source and Rigaku Dectris PILATUS3 R 200 K detector). Wavelength of data collection was 1.54187 Å, temperature 100 K. Data were indexed, integrated, and scaled using DIALS 43 . Phases were obtained by molecular replacement using PDB ID 4DHT as search model in Phaser 44 . Coot 45 and phenix.refine 46 were used in alternating cycles of model building and refinement. See Table 1 for data collection and refinement statistics. See Supplementary Fig. 18 for a portion of the electron density map. Ramachandran statistics for this dataset were obtained from the Ramachandran plot: favored/outlier residues 97.86/0.42%, respectively. The structural data for 14-3-3σ ΔC/12 were submitted to the PDB and obtained entry ID 6YE9. 14-3-3β ΔC/ChREBP-α2/3. 14-3-3β ΔC protein, acetylated ChREBP α2 peptide and compound 3 were dissolved in crystallization buffer (CB; 20 mM HEPES pH 7.5, 2 mM MgCl 2 , 2 mM βME) and mixed in a 1:2:2 molar stoichiometry with a final protein concentration of 12 mg/mL. This was set up for sitting-drop crystallization in a 1:1 ratio with the protein crystallization MPD Suite (Qiagen). Crystals were observed in condition #88 (0.1 M HEPES sodium salt pH 7.5, 30% (w/v) MPD, 5% (w/v) PEG4000) ( Supplementary Fig. 16A-C). After initial optimization of pH and PEG4000 concentration, needle-shaped crystals were subjected to the Additive Screen HT (HR2-138, Hampton Research) of which condition #12 resulted in crystals of an improved three-dimensional shape, even though still clustering around a single nucleation site (Supplementary Fig. 16D). The thus obtained crystallization-liquor constitution (0.1 M HEPES pH 7.1, 30% MDP, 1% PEG4000, 0.1 M Ni(II)Cl•6H 2 O) was subsequently homemade. Finally, the complex (prepared as described above) was set up for hanging-drop crystallization for crystal reproduction in a 1:1 ratio with the homemade crystallization-liquor. After 1 day of incubation at room temperature, rod-like clusters were observed (Supplementary Fig. 17A). These were crushed, resulting in small nucleation seeds ( Supplementary Fig. 17B) that were subsequently introduced into a fresh drop of pre-equilibrated protein complex in aforementioned crystallization-liquor using a cat whisker ( Supplementary  Fig. 17C). Single crystals were fished after 1-3 days of incubation at room temperature ( Supplementary Fig. 17D), and flash-frozen in liquid N 2 . Diffraction data were collected at the Deutsches Elektronen-Synchrotron (DESY Hamburg, Germany) PETRA-III beamline P11. Wavelength of data collection was 1.03320 Å, temperature 80 K. The dataset reported was obtained from a crystal that diffracted to a resolution of 2.09 Å. Data were indexed, integrated, and scaled using DIALS 43 . Phases were obtained by molecular replacement using PDB ID 5F74 as search model in Phaser 44 . Coot 45 and phenix.refine 46 were used in alternating cycles of model building and refinement. See Table 1 for data collection and refinement statistics. See Supplementary Fig. 18 for a portion of the electron density map. Ramachandran statistics for this dataset were obtained from the Ramachandran plot: favored/outlier residues 92.96/1.20%, respectively. The structural data for 14-3-3β ΔC/ChREBP−α2/3 were submitted to the PDB and obtained entry ID 6YGJ.