Computationally-guided optimization of small-molecule inhibitors of the Aurora A kinase–TPX2 protein–protein interaction

Computational binding free energy predictions were validated against experiment and used to design new inhibitors of an important protein–protein interaction.


COMPUTATIONAL METHODS
Free energy perturbation (FEP) theory is a computational technique, which is used in conjunction with molecular dynamics or Monte Carlo sampling to determine protein-ligand free energies of binding [1][2][3]. In principle, FEP includes all components of the binding free energy, including changes in ligand and receptor conformation, hydration penalties and changes in entropy. The predictive power of FEP is limited only by the accuracy of the underlying molecular mechanics force field and the finite sampling of conformational degrees of freedom. Both of these limitations are undergoing continuing improvement, and the current manuscript addresses the latter to some extent by enhancing the sampling of ligand and receptor degrees of freedom.
The rigor of FEP calculations does come with disadvantages. Compared to other computational tools that are more common in fragment-based drug design, such as fragment-based active site mapping, virtual screening of fragment libraries, and de novo fragment-based growing/linking (reviewed in Refs. [4,5]), the computational expense and required user expertise of FEP is higher. However, with the increasing use of graphical processing units (GPUs) to reduce simulation times and automated set-up of calculations to reduce user errors [6][7][8], we envisage FEP becoming even more widely used alongside traditional computational methods and experiment for fragment-based drug design [9].
In this communication, FEP simulations followed standard protocols [10][11][12][13]. Initial coordinates were constructed from the X-ray crystal structure of compound 2 bound to Aurora A using the MCPRO software [1]. ATP, counter-ions and crystal water molecules were removed from the structure. The model comprised the 113 residues closest to the ligand. Residues within a 10 Å radius sphere were allowed to move during the simulations and backbone moves were controlled by the concerted rotation algorithm [15]. The unbound ligand and complexes were solvated in a 25 Å water cap. The latter used the JAWS algorithm [16] with standard parameter settings to place water molecules optimally around the binding site. The protein energetics were described using the OPLS/AA force field, the ligands with OPLS/CM1A [17] and water with TIP4P. FEP calculations used 11 λ windows of simple overlap sampling [18]. Each window consisted of between 10 and 30 million (M) configurations of equilibration and 40M/50M configurations of averaging for the bound/unbound simulations.
At each λ window, four replicas were run in parallel with replica exchange with solute tempering (REST) enhanced sampling applied to the ligand and, as described in the main text, the protein residue L178 [12,13]. The REST method is implemented in MCPRO as described in Ref. [12] with the following modifications:  Intramolecular angle bending force field terms are no longer re-scaled. That is, the REST scaling factor is applied only to dihedral and non-bonded terms. This modification significantly increases replica exchange rates without adversely affecting conformational sampling.
 Any protein residue may now be selected for enhanced sampling, as well as the ligand. That is, the dihedral rotation and non-bonded force field terms of the selected residue are now effectively re-scaled to reduce potential energy barriers to rotation in high temperature replicas of the system [12]. Since the number of atoms in the REST region is now higher, care must be taken to ensure adequate acceptance rates for replica exchange.
REST scaling factors were chosen to be exponentially distributed (25, 86, 160, 250 o C). With these replica temperatures, replica exchange was in the range 15-30% for all λ windows. Exchange attempts between pairs of neighboring replicas were attempted every 10 000 MC steps. The REST protocol was used alongside the 'flip' algorithm in which selected dihedral angles undergo attempted jumps that are much larger than typical MC moves [12,19]. The flip algorithm was applied to the side chain χ2 dihedral angle in L178 and the angle between the phenyl ring and quinoline core of the ligand (χ and φ in Figure 1 of the main text). The jumps were of random size in the range 60 o to 300 o .

CHEMICAL SYNTHESIS OF AURORA A KINASE-TPX2 INHIBITOR ANALOGS
Synthesis of compounds 1-8 proceeded as described in Ref [20]. Synthesis of 9 and 10 is described below.

FLUORESCENCE POLARIZATION ASSAY
The fluorescence polarization assay as well as the procedure to obtain the protein (Hisx6-Aurora A123-403), were described in detail and validated in Ref [20]. To summarize, the fluorescence polarisation (FP) measurements were carried out in a 384-well, low-volume, black, flat bottom polystyrene NBS microplate (Corning 3820) using a PHERAstar Plus plate reader (BMGLabtech). The polarization values are reported in arbitrary millipolarization units (mP) and were measured at an excitation wavelength of 540 nm and an emission wavelength of 590 nm. The optimized assay used His-Aurora A123-403 (300 nM), TAMRA-TPX21-43 (12 nM) in the assay buffer (PBS, 0.3 mM DTT). The Z-factor for the TAMRA-TPX21-43 assay was calculated to be 0.66-0.71. The total volume per well was 30 μL. Compounds were subjected to assays at decreasing concentrations by serial dilutions and two (HTS) to four replicates were measured. The binding curves were analyzed using GraphPad Prism 6.0 software. Nonlinear regression curve fit with variable slope was used to yield IC50 (competitive binding) and Kd values (direct binding, no receptor depletion). Lower limit constraint was imposed on the analysis to match the mP of the TAMRA-TPX2 peptide only control. Ki values were calculated from IC50 using model equations as defined by Nikolovska-Coleska et al. [21]. The displacement of TAMRA-TPX21-43 peptide from Aurora A protein upon the titration of the compounds 6 (described previously, [20]), 9 and 10 is shown in Supplementary Figure 1

CRYSTALLOGRAPHIC DATA
Production of Aurora A for crystallography, crystallization ligand soaking and structure determination was performed as described previously [20]. Structure determination statistics are shown in Supplementary Table 1.
The coordinates and associated structure factors have been submitted to the Protein Data Bank under accession numbers 5OBJ (Aurora A in complex with compound 2) and 5OBR (Aurora A in complex with compound 5).