Dynamics of human protein kinases linked to drug selectivity

Protein kinases are major drug targets, but the development of highly-selective inhibitors has been challenging due to the similarity of their active sites. The observation of distinct structural states of the fully-conserved Asp-Phe-Gly (DFG) loop has put the concept of conformational selection for the DFG-state at the center of kinase drug discovery. Recently, it was shown that Gleevec selectivity for the Tyr-kinases Abl was instead rooted in conformational changes after drug binding. Here, we investigate whether protein dynamics after binding is a more general paradigm for drug selectivity by characterizing the binding of several approved drugs to the Ser/Thr-kinase Aurora A. Using a combination of biophysical techniques, we propose a universal drug-binding mechanism, that rationalizes selectivity, affinity and long on-target residence time for kinase inhibitors. These new concepts, where protein dynamics in the drug-bound state plays the crucial role, can be applied to inhibitor design of targets outside the kinome. eLife digest The Ser/Thr kinase Aurora A is an important target for the development of new anticancer therapies. A longstanding question is how to specifically and effectively inhibit only this kinase in a background of over 550 protein kinases with very similar structures. To this end, understanding the inhibition mechanism of Aurora A by different drugs is essential. Here, we characterize the kinetic mechanism of three distinct kinase drugs, Gleevec (Imatinib), Danusertib (PHA739358) and AT9283 (Pyrazol-4-yl Urea) for Aurora A. We show that inhibitor affinities do not rely exclusively on the recognition of a specific conformation of the Asp-Phe-Gly loop of the kinase. Our quantitative kinetics data put forward an opposing mechanism in which a slow conformational change after drug binding (i.e., induced-fit step) dictates drug affinity.

6 evidence for such conformational selection, but identifies an induced-fit step after drug binding as 172 the overwhelming contribution for Gleevec selectivity towards Abl compared to Src (Agafonov et 173 al., 2014). Here, we ask the obvious question if this mechanism of Gleevec binding to Abl might 174 exemplify a more general mechanism for kinase inhibitors.

175
To assess which kinetic steps control drug affinity and selectivity, we first studied the 176 binding kinetics for Gleevec to Aurora A by stopped-flow spectroscopy using intrinsic tryptophan

192
In order to more rigorously analyze the data and test the model, all time courses of the 193 fluorescence changes were globally fit using the microscopic rate constants determined above as 194 starting values (Figure 4) to the model in Figure 3G, where also the resulting microscopic rate 195 constants are given. The lack of a conformational transition after drug binding (i.e., induced-fit 196 step) in Aurora A should dramatically decrease drug affinity in comparison to Abl. Indeed, Gleevec 197 binds to Aurora A with a , of 24 ± 7 µM ( Figure 3F) compared to the low nM affinity to Abl 198 (Agafonov et al., 2014). Two pieces of independent evidence establish that there is indeed no 199 induced-fit step in Gleevec binding to Aurora A: (i) the calculated KD from the kinetic scheme is in 200 agreement with the macroscopically measured , (c.f. Figure 3G and 3F), and (ii) the observed 201 #--from the dilution experiment ( Figure 3E) coincides with the physical dissociation rate (i.e., 202 intercept of the binding plot, 31 ± 2 s -1 , in Figure 3C). In summary, the lack of an induced-fit step 203 7 for Gleevec binding to Aurora A is the major reason for Gleevec's weak binding, and not the DFG-204 loop conformation.

206
Kinetics of Danusertib binding to Aurora A: three-step kinetics with conformational 207 selection and an induced-fit step 208 Next, we wanted to shed light on why Danusertib, unlike Gleevec, binds very tightly to Aurora A.

209
A high-resolution X-ray structure shows Danusertib bound to Aurora A's active site with its DFG-210 loop in the out conformation ( Figure 5A) (Fancelli et al., 2006), and to rationalize Danusertib's 211 high affinity we measured the kinetics of Danusertib binding to Aurora A directly by stopped-flow 212 experiments at 25 °C. An increase in fluorescence intensity was observed at all Danusertib 213 concentrations and showed double-exponential behavior ( Figure 5B). The dependence of the two 214 observed rates constants on drug concentration is linear for one of them ( Figure 5C) and non-215 linear for the other with an apparent plateau at approximately 16 ± 2 s -1 ( Figure 5D). The step with 216 linear inhibitor concentration dependence corresponds to the second-order binding step, whereas 217 a non-linear concentration dependency hints at protein conformational transitions. For a 218 hyperbolic increase of the observed rate with substrate concentrations, one cannot a priori 219 differentiate between a conformational selection and an induced fit mechanism. However, 220 conformational selection happens before drug binding, and the intrinsic slow DFG-in to DFG-out 221 interconversion in Aurora A revealed by Gleevec binding ( Figure 3A) must, therefore, be 222 unaltered. Since the apparent rate of 16 ± 2 s -1 ( Figure 5D) is two orders of magnitude faster, it 223 can only reflect an induced-fit step (i.e., #$% = / + '/ ).

224
So, what happened to the conformational selection step? We hypothesize that the lack of 225 this step in our kinetic traces is due to a too small amplitude of this phase, or not observable 226 because of photobleaching having a bigger effect at the longer measurement times. To lessen 227 potential photobleaching, we reduced the enzyme concentration and increased the temperature

231
While these experiments clearly establish the three-step binding mechanism, it does not 232 provide accurate rate constants for the conformational selection step and it cannot be observed  reliable rate constants ( ( = 0.09 ± 0.01 s -1 and '( = 0.06 ± 0.005 s -1 ) for the conformational 236 selection step in Aurora A, which will be used as "knowns" in what follows. We hypothesize that 237 8 the conformational selection step reflects the interconversion between inactive/active 238 conformations and is correlated with the DFG-out and -in position (Figure 1). The following 239 observations support our hypothesis: (i) two crystal structures for the apo-protein show Trp277 in 240 very different environments ( Figure 1E), (ii) Danusertib has been proposed to selectively bind to 241 the DFG-out conformation based on a co-crystal structure ( Figure 5A) (Fancelli et al., 2006), and 242 (iii) the same slow step is observed for binding of both Gleevec and Danusertib.

243
Next, the dissociation kinetics for Danusertib was measured by fluorescence and 244 appeared to be extremely slow with an observed slow-off rate of (3.2 ± 0.3) ´ 10 -4 s -1 ( Figure 5E).

245
Rationalization of complex binding kinetics cannot be done anymore by visual inspection and 246 kinetic intuition, which can, in fact, be misleading. In order to elucidate the correct binding 247 mechanism and obtain accurate kinetic parameters, all kinetic traces were globally fit ( Figure 6) 248 to the three-step binding scheme ( Figure 5I

275
Our results illuminate trivial but profound principles of binding affinity and lifetime of 276 drug/target complexes: a conformational selection mechanism always weakens the overall 277 inhibitor affinity, while an induced-fit step tightens the affinity depending on how far-shifted the 278 equilibrium in the enzyme/drug complex is (Equations 2, 3 and 4). For DFG-out binders (e.g.,

279
Danusertib and Gleevec), the DFG-in and -out equilibrium weakens the overall affinity 1.6-fold; 280 however, the conformational change after drug binding results in a four orders of magnitude 281 tighter binding for Danusertib and is the sole reason for its high affinity to Aurora A compared to 282 Gleevec. The dissociation constants for the bimolecular binding step & is very similar for both 283 inhibitors. Finally, the lifetime of Danusertib on the target is very long because of the very slow 284 conformational dynamics within the Aurora A/Danusertib complex ( '/ = (7.1 ± 0.5) ´ 10 -4 s -1 ).

287
We chose AT9283 as a third inhibitor to characterize the binding mechanism because it has been 288 described as a DFG-in binder based on a crystal structure of AT9283 bound to Aurora A (PDB 289 2W1G, (Howard et al., 2009)). We, therefore, anticipated that in its binding kinetics one can now

314
Finally, the 10-fold difference between the '/ from the global fit ( Figure 7H) and the 315 experimentally observed slow off-rate can be reconciled by kinetic partitioning as shown in Figure   316 7-figure supplement 1A.

319
In an effort to structurally verify our model we solved a crystal structure of Aurora A with AT9283 320 bound and indeed observed the DFG-out conformation (PDB 6CPG, Figure 9B and     Figure 10A) that describe the physical binding step (i.e., linear dependence on mant-ATP 338 concentration; Figure 10B) and the induced-fit step ( Figure 10C). The observed rate constant 339 11 approaches a maximum value defined by the sum of / + '/ ( Figure 10C) and the intercept can 340 be estimated to be '/ and is consistent with the value obtained from the #--experiment ( Figure   341 10D). We find that mant-ATP can bind to both the DFG-in or -out conformations, consistent with 342 our nucleotide-bound crystal structures ( Figure 1A-D) and recent single-molecule fluorescence 343 spectroscopy data that indicates that nucleotide binding does not significantly affects this 344 equilibrium (Gilburt et al., 2017). To confirm the model, the kinetic data were globally fit to a two-345 step binding mechanism ( Figure 10H, G). The calculated , from the corresponding microscopic 346 rate constants ( Figure 10H) is comparable with experimental macroscopic , obtained from a 347 titration experiment ( Figure 10E, F).

348
The presence of an induced-fit step for the natural substrate ATP suggests that such 349 conformational change after ligand binding is a built-in property of the enzyme. In other words, 350 inhibitors take advantage of the inherent plasticity of the enzyme that is required for its activity 351 and regulation. The main difference between ATP and inhibitor binding is the rate constant for the 352 reverse induced-fit step ( '/ ). In the case of ATP, this rate is much faster and, therefore, does     The "use" of a far-shifted induced-fit step for a promising drug is logical for the following

394
Our results exemplify why rational drug design is so challenging. The characterization of 395 the complete free-energy landscape of drug binding is needed, which will require more 396 sophisticated computational approaches guided by experimental data such as provided in our

504
An interscan delay of 1.5 s was used with 5,000 scans per transients, giving rise to a total 505 acquisition time of 2.5 h per spectrum. To remove background signal from the probe and avoid 506 baseline distortions, data acquisition was started after a ~100 µs delay (using the "delacq" macro) 507 and appropriate shifting of the data followed by backward linear prediction was performed. The 508 16 data were apodized with an exponential filter (2.5 Hz line broadening) and zero-filled before 509 Fourier transform. To improve the signal-to-noise ratio several data sets were recorded 510 consecutively and, provided that the sample remained stable, added together after processing 511 (two for apo Aurora A, four for Aurora A + AMPPCP, and five for W277L + AMPPCP, respectively).

548
Under the rapid equilibrium approximation, the binding and dissociation steps of Gleevec 549 to Aurora A are fast compared to conformational selection, therefore the value of ( and '( can    K D = (K 1 + 1)*K 2 *K 3 *K 4 K 1 *K 2 *K 3 + K 3 *K 4 + K 4 20 10 µM/10 µM Aurora A/mant-ATP complex was diluted with buffer (ratio 1:10). A significant     Gleevec (labeled as G) binding scheme to Aurora A corresponds to a two-step binding mechanism: conformational selection followed by the physical binding step. The corresponding microscopic rate constants obtained from the global fit and calculated overall equilibrium and dissociation constants are shown. Fluorescence traces are the average of at least five replicate measurements (n > 5), and error bars and uncertainties given in C-G denote the (propagated) standard deviation in the fitted parameter.   Figure  5). Fluorescence traces are the average of at least five replicate measurements (n > 5). Global fitting was performed using the KinTek Explorer software using the model shown in Figure 5I.   Figure 8. Global fits of AT9283 binding and dissociation kinetics to Aurora A at 25 °C. Binding kinetics was monitored by stopped-flow fluorescence at different concentrations of AT9283 (indicated) to 0.5 μM Aurora A. Dissociation kinetics were obtained for fully equilibrated drug/kinase complex (k obs,off ) or for the initial encounter complex (k off,djump ) by using a 1 hour or a short 2 s incubation of the kinase with AT9283, respectively, before inducing dissociation by a buffer wash using Creoptix WAVE waveguide interferometry. Global fitting was performed with KinTek Explorer software using the model in Figure 7H (reduced χ 2 = 3.2). Fluorescence traces are the average of at least five replicate measurements (n > 5). Alternative binding models of AT9283 to Aurora A cannot explain the experimental data. (A) Our initial three-state binding scheme, where AT9283 binds only the DFG in state of Aurora A and is followed by an induced-fit step, is incorrect. The best global fit (shown in red) did not describe the data as can be seen by visual inspection and from the reduced χ 2 value of 36. (B) An alternative model, where AT9283 can bind to Aurora A irrespective of the state of the DFG-loop, and binding is followed by an induced-fit step did not result in adequate fits (data not shown) and a reduced χ 2 value of 52. In both cases the values for the interconversion between AurA out and AurA in were taken from the Gleevec experiment ( Figure 5-figure supplement 2). Fluorescence traces are the average of at least five replicate measurements (n > 5).