Design of rigid protein–protein interaction inhibitors enables targeting of undruggable Mcl-1

The Impossible Mcl-1 PPI Site.

Despite the pivotal role of Mcl-1 in leukemia, Mcl-1 inhibitor development has lagged behind its antiapoptotic family members Bcl-2 and Bcl-xL by almost 20 y. Compared with Bcl-2 and Bcl-xL, the PPI pocket of Mcl-1 is shallower and less flexible, which substantially exacerbates the development of potent oral small-molecule inhibitors (10). We analyzed the PPI pocket of Mcl-1 in its apo and Bim peptide-bound crystal structures in detail, using the two scoring methods, number of probe points (PP) (11, 12), and protein-ligand binding (PLB) index (13) (Fig. 1A). The PP method scores protein crevices by volume and small-molecule drug-like shape, whereas the PLB index method additionally weighs such PP scores by the abundance of pocket amino acids frequently found in drug-like sites, such as Trp, His, Phe, and Met, thereby ruling out high scores for highly hydrophilic pockets. Apo Mcl-1 and Bim-bound Mcl-1 possess PP scores of 26 and 10, respectively, which are far below average druggable pocket PP scores of 30 and above. Similarly, the PLB indices for apo and Bim-bound Mcl-1 show low values of 0.0 and 0.1, respectively, while drug-like pockets often show PLB indices of 1.0 and above.

Nevertheless, we and researchers at Amgen and Servier took on this challenge and delivered small-molecule Mcl-1 PPI inhibitors, which advanced to clinical Phase 1 or 2 oncology studies and were previously thought impossible (Fig. 1B). Three additional Mcl-1 inhibitors, namely ABBV-467, PRT1419, and GS-9716, by teams from AbbVie, Prelude Therapeutics, and Gilead Sciences, respectively, have entered the clinic, but have not been disclosed in chemical structure detail yet (4, 14). Thereof, the clinical trials of AMG-397 and AZD5991 have been terminated by now (15, 16). Nevertheless, this impressive number of clinical small molecules gives evidence of the crossorganizational breakthrough made in Mcl-1 drug development.

Clinical Small-Molecule Inhibitors Induce Druggable Pocket.

We set out investigating the drug–protein interactions in atomic detail by determining the 3D crystal structure of the clinical Phase 1 inhibitor AMG-176 (1) bound to Mcl-1 (SI Appendix, Fig. S1). Our structure reveals that 1 occupies an overlapping pocket to the BH3 cognate peptides, here exemplified by the pan-selective Bim peptide bound to Mcl-1 (Fig. 2A) (17). The superimposition of the Mcl-1-bound crystal structures of 1 and Bim illustrates that Bim binds as an a-helix to an extended surface groove between helices a3 and a4 of Mcl-1, whereas the synthetic inhibitor 1 cuts deeply into the hydrophobic core of Mcl-1. The important salt bridge interaction between Arg263 of Mcl-1 and the conserved Asp15 of the BH3 peptides (residue numbering as in PDB-code = 2PQK) is formed by the acidic acylsulfonamide of 1. However, the intraprotein interactions of Arg263 differ markedly. In Bim-bound Mcl-1, the guanidinium of Arg263 is salt bridged to the carboxylate of Asp256 at the C terminus of a4, forcing the NH of Asp256 out of the a-helical H-bond network. By contrast, in Mcl-1 bound to 1, the guanidinium of Arg263 forms H-bonds with the carbonyls of V253 and F254. In effect, the entire helix a4 is shifted toward its N terminus by 3 Å compared with Bim-bound Mcl-1. Hydrophobic core-forming residues on a4, such as Leu246, Val249, and Met250, are shifted along, opening up the sizeable crevice utilized by the 6-chlorotetralin of 1.

This exceptional move of a4 in a folded a-helical bundle protein is made possible by the NH of Val258 losing its H-bond to the carboxylate of Asp256, and by Gly257 losing its a-helical H-bond in favor of an extended conformation. The side chain of Leu267 in helix a5 rearranges in concert to fill the void left by the shifted a4 hydrophobic core residues. As a consequence, a PPI binding site shallow and not drug like in both apo and cognate peptide-bound Mcl-1 becomes unexpectedly voluminous and druggable in the presence of 1.

Mcl-1 Plasticity upon Inhibitor Binding.

We extended our binding site analysis to additional potent, crystal structure-defined, synthetic inhibitors of Mcl-1, of which two, S64315 (2) (18) and AZD5991 (3) (19), have entered clinical trials, and the discovery tool compound 4, derived from a DNA-encoded library screening hit (20). One key mutually conserved interaction of 1, 2, 3, 4, and Bim is with Arg263, which forms salt bridges with the carboxylates of 2 and 3 and Asp15 of BH3 peptides. The synthetic inhibitor 4 adopts a cation-π interaction between its para-chlorobenzyl and Arg263. The remaining interactions differ markedly.

1 and 3 show large binding site volume increases, as quantified by more than a doubling of PP and higher PLB indices compared with apo and Bim-bound Mcl-1 (Fig. 1A), consistent with both inhibitors 1 and 3 inserting their 6-chlorotetralin and naphthalin moieties, respectively, deeply into the Mcl-1 core. By contrast, 2 and 4 show merely moderate PP and PLB index increases with respect to apo and Bim-bound Mcl-1. Molecular analysis of their pocket interaction reveals that 2 fails to pry open the Mcl-1 core between helices a3 and a4 and does not engage with Met250 and Phe270, thereby explaining the low probe point and PLB scores. Inhibitor 4 does penetrate the Mcl-1 core with its flat chloro-(2-phenylethyl)benzene substituent, but opens up much less pocket volume than 1 and 3. Since 4 still sandwiches in between the high-scoring pocket residues Met250 and Phe270, it achieves a PLB index elevation to 0.57.

Allosteric Mcl-1 Plasticity by Crystallography and in Solution.

Next, we looked for Mcl-1 plasticity beyond its Bim-binding site. Comparison of the Ca backbone rmsd values between the whole protein and residues within 4 Å of the pocket of 1 demonstrates that conformational rearrangements of around 1 Å occur throughout the protein and between all ligated complexes investigated here (Fig. 3). Prominent pocket movements of helices a3 and a4 as well as the a4–a5 loop of Mcl-1 trigger plasticity throughout the entire protein (Fig. 3). Previous exploration of Mcl-1 peptide ligand–induced conformational changes identified systematic differences in protein backbone rmsd values of approximately 0.7 Å between crystal and NMR-derived solution structures (17). Alerted by these potential crystal-packing effects, we investigated the binding of inhibitors 1, 2, 3, 5, and Bim by ¹⁵N-HSQC NMR spectroscopy in solution. Inhibitor 5 is the 15-fold more potent methylsulfon analog of 4 (Fig. 1) and expected to bind to the same site as 4 (20).

Protein conformational changes produce chemical shift perturbations in backbone amide ¹H-¹⁵N chemical shifts that can be quantified by the ECHOS-NMR (easy comparability of higher order structure by NMR) analysis of two-dimensional ¹⁵N-HSQC spectra. This method yields spectral correlation coefficient (R) values of 1.0 for no change at all and R values as low as 0.0 for maximum changes, such as protein unfolding (21). We validated this statistical quantification method by studying different batches of ¹⁵N-labeled Mcl-1 unbound and bound to the low-affinity binder 1-methylindole-2-carboxylate (6) (22). We obtained R values of >0.99 and 0.97 for unbound and Mcl-1 bound to 6, respectively (SI Appendix, Fig. S2). By contrast, inhibitors 1, 2, 3, 5, and Bim produced R values of 0.57, 0.48, 0.56, 0.53, and 0.59, respectively, indicating substantial Mcl-1 rearrangement upon ligand binding in solution. Hence, Bim and inhibitors 1 to 5 induce Mcl-1 plasticity of similar magnitude both in solution and in single crystals, but deep druggable pocket inductions are unique to these inhibitors.

Preorganization Inhibitor Design Is Key to Success.

While ligand-induced protein rearrangements are universal in biology, the creation of unprecedented drug-like pockets is certainly not. We were intrigued to further investigate the landmark differences between BH3 peptide ligands and inhibitors 1 to 5. Affinity and potency measurements in vitro showed similar interaction strengths for Bim and 1 to 5, thereby not explaining the stark pocket volume changes unique to 1 to 4 (Table 1).

Rather, we asked whether the macrocyclic structures of 1, 3, and 5 could be a root cause. We studied 1 to 3, and 5 by free-in-solution conformational analysis. This method utilizes the experimental NMR restraint NOE distances, scalar couplings, and ¹H chemical shifts to calculate the 3D conformers that a small molecule predominantly populates free in solution (23). We found that the free inhibitor 1 adopts a 3D conformation identical to the crystal-defined conformer of 1 bound to Mcl-1 within the accuracy of the forcefield calculations (Fig. 4A). This stunning preorganization of 1 free in solution is not trivial, as the computational molecular dynamics (MD) simulation of the theoretical conformer space produces several important diversions (SI Appendix, Fig. S3). The methoxy substituent on the macrocycle of 1 can fold inward to H bond with the NH of the sulfonamide, and the 6-chlorotetralin can adopt many poses. However, our NMR experimentation demonstrates that 1 free in solution is entirely preorganized to induce the exact binding pocket observed by crystallography.

Analogous free ligand analysis of 2 again reveals an unexpected degree of preorganization free in solution (Fig. 4B). The 2-chloro-3-methyl-phenol and the 2-hydroxy-3-phenyl-propanoate of 2 adopt identical 3D poses free in solution as in the bound crystal structure. In the absence of a macrocyclic backbone, this lack of rotational space sampling is a well-designed conformational property of 2. By NMR experimentation, we found that the ortho hydrogen of the 2-hydroxy-3-phenyl-propanoate packs within 2.8 Å against the 2-chloro-3-methyl-phenol ring, as evidenced by intramolecular NOE and a characteristic upfield chemical shift for this hydrogen, which is 0.9 ppm smaller than predicted by the 3D-insensitive HOSE method (24) (Fig. 5). In effect, this edge-to-face aromatic ring alignment generates a tiny hydrophobic core that stabilizes 2 in a largely bioactive conformation. Only its methyl-piperazine and anisole substituents are unstructured in solution, both of which form merely solvent-exposed surface interactions with Mcl-1 in the crystal structure.

Free-in-solution conformer analysis of 3 confirmed the findings of our previous chemical shift study (19). 3 is fully preorganized to bind Mcl-1 in the very same pose as observed in the bound crystal structure (Fig. 4C). Similar to 2, the 3D conformer of 3 shows two aromatic rings packed edge to face, as determined by intramolecular NOE and a chemical shift of the involved pyrazole hydrogen shifted by 1.1 ppm upfield as compared with the HOSE prediction method (Fig. 5).

Finally, inhibitor 5, despite being macrocyclic, shows much less preorganization than 1 to 3 (Fig. 4D). The methylsulfonylbenzene of 5, forming the important interaction with the guanidinium of Arg263, shows complete rotational freedom in the unbound state. Likewise, the 1-chloro-2-(2-phenylethyl)benzene, inserting deeply into the Mcl-1 core to contact Met250 in the crystal structure of 4, is unstructured free in solution. Merely its methylene connection to the macrocyclic backbone is rotationally restricted through a face-to-face aromatic ring packing, as evidenced by intramolecular NOE. The 1.3 ppm upfield chemical shift of the 9-hydrogen of the tetrahydro-1-benzazepin-2-one as compared with its HOSE-predicted chemical shift supports this aromatic ring packing (Fig. 5). The cis/trans amide bond configurations are preserved between unbound 5 and bound 4, but the macrocycle backbone still shows larger deviation from its bioactive conformer than observed for the macrocycles 1 and 3. Taken together, this lack of preorganization goes along with an approximately 100-fold lower affinity for Mcl-1 than 1 to 3 in our surface plasmon resonance assay, and a harsh 1000-fold drop-off in cellular potency in our MOLP-8 cell assay (Table 1).