Multiple Steps to Activate FAK’s Kinase Domain: Adaptation to Confined Environments?

Abstract

Protein kinases regulate cell signaling by phosphorylating their substrates in response to environment-specific stimuli. Using molecular dynamics, we studied the catalytically active and inactive conformations of the kinase domain of the focal adhesion kinase (FAK), which are distinguished by displaying a structured or unstructured activation loop, respectively. Upon removal of an ATP analog, we show that the nucleotide-binding pocket in the catalytically active conformation is structurally unstable and fluctuates between an open and closed configuration. In contrast, the pocket remains open in the catalytically inactive form upon removal of an inhibitor from the pocket. Because temporal pocket closures will slow the ATP on-rate, these simulations suggest a multistep process in which the kinase domain is more likely to bind ATP in the catalytically inactive than in the active form. Transient closures of the ATP-binding pocket might allow FAK to slow down its catalytic cycle. These short cat naps could be adaptions to crowded or confined environments by giving the substrate sufficient time to diffuse away. The simulations show further how either the phosphorylation of the activation loop or the activating mutations of the so-called SuperFAK influence the electrostatic switch that controls kinase activity.

Introduction

Phosphorylation of proteins is recognized as one of the most important posttranslational modifications, because it is involved in almost every signaling pathway (1–3), and has been the target of various drug developments (4,5). Therefore, understanding the mechanism of the regulation of kinases and phosphatases is of major interest in health and disease. Crystallographic data of active and inactive kinases have shed light on the different conformations of kinases (6) and kinetic measurements provide information about the underlying mechanisms (7), but the dynamics of kinases and their cycle of substrate binding, reaction, and product release are still not completely understood at an atomistic level.

The focal adhesion kinase (FAK) is one of the most upstream tyrosine kinases in the formation of focal adhesions, at the crossroad of extra- and intracellular signaling (8–10). FAK has both scaffolding and signaling functions, and is involved in many cell processes, such as adhesion, spreading, and migration (8,11). FAK signaling is involved in numerous diseases including rheumatoid arthritis (12), and it has been found that the activity of FAK is altered in various types of cancer (13,14).

FAK consists of three domains: the FERM, kinase (KD), and focal adhesion targeting (FAT) domains (Fig. 1 A). The FERM domain can bind to the KD and is thought to autoinhibit kinase activity (Fig. 1 A) (15). Release of the autoinhibitory state is the first step in kinase activation. The KD of FAK corresponds to the well-conserved fold of protein kinases (16). It consists of two lobes, a small N-terminal lobe (N-lobe) and a larger C-terminal lobe (C-lobe) (Fig. 1 B). The catalytic core and the ATP-binding site are located in the groove formed by these two lobes (see Fig. 1 D). The structures of the active, phosphorylated kinase complexed with a nonhydrolysable ATP analog (Fig. 1 B) are known, as well as the nonphosphorylated, inactive kinase domain with an inhibitor (Fig. 1 C) (15). The most striking difference between the active and inactive structure of FAK is the conformation of the activation loop (residues 564–592). The loop can be doubly phosphorylated on Y576 and Y577 and form a β-hairpin-like conformation in the active conformation, whereas it is unstructured in the nonphosphorylated and inactive form (15). Phosphorylation of the two tyrosine residues Y576 and Y577 by Src increases the catalytic activity of FAK >20-fold (15). Distinct structural differences in the conformation of the activation loop between the active and inactive state is a common feature of protein kinases (17). It was thought that the conformation of the activation loop is regulated by phosphorylation of the activation loop and that the activation loop directly controls substrate binding and catalysis (18). Although this is the case for some kinases, there are kinases where the activation loop modulates phosphoryl transfer, but not substrate access (19).

Figure 1.

Crystal structures of FAK. (A) The autoinhibited structure of the KD with the FERM domain (PDB 2J0J) together with the FAT domain (PDB 1K40). (B) Crystal structure of the catalytically active kinase domain with ATP analog AMP-PNP (PDB 2J0L). The activation loop (red) forms two small beta strands (denoted by KD⁼). (C) The crystal structure with staurosporin as inhibitor (PDB 2J0J) shows an unstructured activation loop (KD^≈), which was partly modeled (orange). (D) Interactions between KD⁼ and ATP-Mg.

Several computational studies tried to bridge the gap between the catalytically active and inactive structures obtained by crystallography. For instance, the conformational change of the activation loop of Src, cyclin-dependent kinase 5 (CDK5) and vascular endothelial growth factor receptor 2 (VEGFR2) was investigated by coarse-grained (20) or biased molecular dynamics (MD), string method swarms of trajectories derived from targeted MD, metadynamics, and accelerated MD (21–24). All but the coarse-grained study are based on biased models with a priori knowledge of the end-state, because the expected timescale of milliseconds to seconds for the conformational change of the activation loop (20,21) exceeds the timeframe that can be simulated by all-atom MD. Moreover, several studies investigated the role of phosphorylation of the activation loop in CDK2, protein kinase Cθ (PKCθ), and protein kinase A (PKA) (25–27). Computer simulations deepened the understanding of an electrostatic network that regulates kinase activity in many kinases (28). The catalytically important salt bridge between a conserved glutamate residue of the αC helix (residue E471 in FAK) and a lysine in the N-lobe (K454 in FAK) can be switched on and off through a network of charged residues in Src and Abl (21,29). Further simulations showed that the energy barrier between the active and inactive state of the Lyn and CDK2 kinase is determined by a displacement of the αC helix (30).

The enzymatic kinetics of FAK’s KD were studied experimentally (31), but the underlying structural mechanisms remained unknown. Here, we use MD simulations to study the fast conformational changes of FAK’s KD in the range of 100 ns. We show that the ATP-binding pocket is structurally unstable and closes transiently upon nucleotide removal from the active conformation that displays a structured activation loop. In contrast, the ATP-binding pocket remains open when the kinase is in the inactive conformation displaying an unstructured activation loop. Furthermore, the role of ATP-Mg on the activation loop conformation and interlobular dynamics is analyzed, as well as how these dynamics are affected by the activating mutations K578E and K581E of the so-called superFAK (32).

Methods

Initial structures

The kinase domain of FAK (residues 411–686) was simulated starting from two different crystal structures: the phosphorylated active conformation in complex with the nonhydrolyzable ATP analog Adenylyl Imidodiphosphate (AMP-PNP) and Mg (Protein Data Bank (PDB) code 2J0L, resolution 2.30 Å (15)) and the nonphosphorylated inactive conformation with the inhibitor staurosporine compound AFN941, which binds to the ATP-binding pocket (PDB code 2J0J, resolution 2.8 Å (15)). All nonprotein atoms were removed except in the simulation with ATP-Mg, where ANP was mutated to ATP and the Mg atom was retained. In the inactive conformation, residues 574 to 583 are unresolved and were modeled with MODELLER (33). The residues Y576 and Y577 were simulated in both the nonphosphorylated and phosphorylated state. For the SuperFAK simulations, the K578E and K581E mutations were introduced using the Mutate plugin in VMD (34). In these simulations, Y576 and Y577 were not phosphorylated. The setup of all the simulations is listed in Table 1.

Table 1.

Summary of simulations

Name	Initial structure	Activation Loop	Phosphorylation	Ligand	Mutation	Simulation time [ns]
KD=	2J0L	structured	–	–	–	3 × 130
KD= (a)	2J0L	structured	–	–	–	3 × 130
p-KD=	2J0L	structured	pY576, pY577	–	–	3 × 130
p-KD= (a)	2J0L	structured	pY576, pY577	–	–	3 × 130
holo p-KD=	2J0L	structured	pY576, pY577	ATP-Mg	–	100
KD≈	2J0Jb	unstructured	–	–	–	3 × 130
KD≈	2J0Jc	unstructured	–	–	–	3 × 130
holo KD≈	2J0Jb	unstructured	–	ATP-Mg	–	100
holo p-KD≈	2J0Jb	unstructured	pY576, pY577	ATP-Mg	–	100
p-KD≈	2J0Jb	unstructured	pY576, pY577	–	–	3 × 130
SuperFAK=	2J0L	structured	–	–	K578E, K581E	100
holo-SuperFAK=	2J0L	structured	–	ATP-Mg	K578E, K581E	100
SuperFAK≈	2J0Jb	unstructured	–	–	K578E, K581E	100
					Total	3330 ns

^aWater was first equilibrated for 2 ns with protein atoms fixed.

^bUnresolved residues of the activation loop were modeled with MODELER (1 model).

^cUnresolved residues of the activation loop were modeled with MODELER (3 different models).

MD simulations

MD simulations were performed with NAMD (35) using the CHARMM27 (36–38) force field. The starting structures were immersed in a box of TIP3P water molecules (39) with a padding of 15 Å in all three dimensions using the Solvate plugin in VMD. To neutralize the system, Na⁺ and Cl⁻ ions were added at a physiological concentration of 150 mM. First, the system was energy minimized for 2000 steps with the protein coordinates fixed and additional 2000 steps without any restrains. The system was then heated up gradually from 0 to 310 K, increasing the temperature every 100 steps by 3.1 K. MD simulations were performed under constant pressure and temperature (NPT ensemble) using Langevin dynamics with a damping factor of 1 ps⁻¹ and Nose-Hoover Langevin piston method with a damping time constant of 50 fs and decay period of 100 fs. Long-range electrostatic forces were simulated using the particle mesh Ewald summation with a grid size smaller than 1 Å. Full electrostatic interactions were calculated every 4th step. Van der Waals interactions were calculated using a switching function starting at 10 Å with a cutoff at 12 Å. Rigid bond lengths and angles of water molecules were used and an integration step of 1 fs was chosen. The repetitions of simulations were done with identical starting structures with a different random number seed for the initial velocity set at 310°K. After 10 ns, the root mean-square deviation (RMSD) of the protein with respect to the crystal structure remained essentially constant (Fig. S1 in the Supporting Material). Where specified, protein atoms were fixed for the first 2 ns to allow for an optimal hydration of the ATP-binding pocket.

Analysis

The trajectories were analyzed and visualized with VMD version 1.9. For the RMSD and root mean-square fluctuation (RMSF) analysis, the backbone atoms of the protein were aligned to the starting structure. Hydrogen bonds were assigned with distance and angle cutoff at 3.5 Å and 30°. The analysis was performed using Java scripts developed and tailored to our needs. The average distances d1 and d2 were tested with an unpaired student’s t-test. The covariance matrix of the carbon alpha atoms was calculated from the trajectories using Carma (40).

Computation

All simulations were performed at the Swiss National Supercomputing Centre CSCS on a Cray XE6. The systems contained around 45,000 atoms. In total, >3 μs were simulated.

Results

This study only focuses on KD of FAK, without the FERM domain, and thus considers events after the release of autoinhibition. Hereafter, we will furthermore differentiate the two distinct conformations of the activation loop, whereby the structured activation loop will be denoted with KD⁼ and represents the catalytically active form of KD, which allows substrate access to the active site (41), whereas we use KD^≈ for the catalytically inactive conformation commonly associated with the unstructured activation loop. In the crystal structure of KD^≈, residues 574 to 583 of the activation loop were unresolved and thus modeled by an unstructured loop (Fig. 1 C).

ATP-binding pocket dynamics

Following the phosphorylation of the substrate and to proceed through the catalytic cycle, ADP has to be released from the nucleotide-binding pocket of FAK’s KD. The ATP-binding pocket resides between the N- and C-lobe of the kinase, formed by the G-loop and some residues of the C-lobe (Fig. 1 D). Both, in the catalytically active (PDB 2J0L) and inactive (PDB 2J0J) crystal structures, a ligand is bound to the nucleotide-binding pocket. In the active structure, the nonhydrolysable ATP analog AMP-PNP is bound (Fig. 1 B), whereas the kinase inhibitor staurosporine AFN941 stabilized the inactive conformation (Fig. 1 C). To gain insights into the structural changes of the KD that happen upon release of the nucleotide, we removed the ligand and replaced it with ∼7–9 water molecules using the Solvate plugin in VMD. Both structures were thus in an open configuration of the ATP-binding pocket at time zero (Fig. 2 A). All simulations of unliganded KD were repeated three times.

Figure 2.

MD simulations of FAK’s unliganded KD. (A) Surface representation of initial structures without ligand and open ATP pocket. (B) Although the pocket is unstable and can close when KD displays a structured activation loop (KD⁼) (t = 80 ns), the pocket remains open in the conformation with the unstructured activation loop (KD^≈) (t =130 ns) (see also Movies S1–S9). (C) Superposition of the initial (transparent) and final (solid) structures. (D) The residues forming the binding pocket, both in the open (t = 0 ns) and closed (t = 80 ns) state. In the closed state, the residues framing the pocket form hydrogen bonds (shown in red). (E) The average distance d1 and d2 between the ATP pocket residues (residues I428–E506 and Q432–D546) are lower in the six KD⁼ simulations (blue) compared to six KD^≈ simulations (green). (F) Distances d1 and d2 plotted over time for one representative nonphosphorylated KD⁼ and KD^≈ simulation. (G) The number of hydrogen bonds between the upper and lower residues that frame the ATP pocket is shown over time for the three repetitions of FAK with the unstructured (KD^≈) and structured activation loop (KD⁼), respectively, in both the nonphosphorylated (np) and phosphorylated pY576 and pY577 (p) form. (H) Occurrence of hydrogen bonds between residues that frame the ATP binding pocket. ^**P < 0.01, Student’s t-test.

Although the nucleotide pocket remained open throughout the simulations in the KD^≈ simulations with the unstructured activation loop (Fig. 2 B, Movie S1, Movie S2), the pocket was structurally unstable and transiently closed in the KD⁼ configuration (Fig. 2 B, Movie S3, Movie S4). The conformational change from an open to a closed state is shown in Fig. 2, C and D. The distance between the pocket-forming residues in both initial structures is around 10 Å and there are no hydrogen bonds between the upper (residues 426–432) and lower (residues 503–506, 546, 550) lobe of the pocket. When the groove between the N- and C-lobes closed, the residues at the edge of the pocket formed several backbone and side-chain hydrogen bonds (Fig. 2 D). For the KD⁼ and KD^≈ simulations in both the phosphorylated or nonphosphorylated form that were repeated three times, the number of hydrogen bonds between the upper and lower lobe of the pocket is shown over time (see Fig. 2 G) and the relative occurrence (see Fig. 2 H). In five out of six KD^≈ simulations, the pocket remained open. In one, it transiently closed for 24.3 ns. In contrast, the pocket is unstable in five out of six simulations of KD⁼ within the simulated timeframe of 130 ns. The dynamics of a nonphosphorylated KD⁼ simulation, where the pocket closes within the first nanosecond, and of one nonphosphorylated KD^≈ simulation are shown in Movie S5 and Movie S6, and the other simulations are shown in Movies S1, S2, S3, and S4.

Strikingly, the pocket dynamics depend on the configuration of the activation loop. To further characterize the state of the ATP-binding pocket, two distances d1 and d2 between the pocket-forming residues (Fig. 2 D) were quantified here as function of time. The average distances d1 and d2 of the KD⁼ simulations were significantly lower than in the KD^≈ simulations (Fig. 2 E). The distances d1 and d2 are shown over time for a KD⁼ and KD^≈ simulation with a nonphosphorylated activation loop (np-KD⁼ and np-KD^≈) in Fig. 2 F and Movie S5 and Movie S6. Repeatedly, we observed a transient closing of the pocket in KD⁼ and the stably open pocket in KD^≈, irrespective of whether the activation loop was phosphorylated on Y576 and Y577 or not. To exclude that the initial conformation of the modeled part of the activation loop influences this result, we ran additional three simulations of nonphosphorylated KD^≈ (np-KD^≈) for 130 ns with different initial configurations of the modeled loop. As expected, only a few and transient hydrogen bonds formed between the upper and lower lobes of the pocket (Fig. S2, Movie S7). To exclude that the pocket closure resulted from insufficient initial solvation of the pocket, we repeated the KD⁼ simulations, where the protein atoms were fixed for the first two nanoseconds (Fig. S3, Movie S8, Movie S9). During this period, around 22 water molecules filled the cavity. After release of the constraints, the pocket showed again unstable pocket dynamics. The pocket closed in several simulations and similar hydrogen bonding between the pocket-forming residues were observed as in the previous KD⁼ simulations. In contrast to the simulations with the first set of constrains (Fig. 2 G), the preequilibration of the water lead to a more transient closure of the ATP-binding pocket. The KD⁼ pocket can close and reopen within the simulated timeframe, which was not observed in the KD^≈ simulations. Consistently, we find that the ATP-binding pocket is rendered into a structurally instable state upon removal of an ATP analog, and this finding is independent of how the cavity was filled with water molecules.

Effect of activation loop conformation on the nucleotide-binding pocket dynamics

In the simulations of KD⁼ that displayed the structured activation loop, the αC helix and the G-loop moved rapidly toward the catalytic core (Fig. 2 D), which closed the ATP-binding pocket transiently. Both hydrophobic and electrostatic interactions of residues that frame the catalytic core are regulating the ATP-binding pocket accessibility. The motion of the αC helix is highly correlated with the motion of the G-loop (see covariance between the αC helix and the G-loop in Fig. 4 E) through a network of hydrophobic residues including F533. In KD⁼, L567 of the activation loop stably pointed toward the C-lobe, away from the αC helix (see Fig. 3 B). The αC helix and the G-loop could swing down to form a more compact, closed conformation. In this conformation, Q432 of the G-loop formed a relatively stable hydrogen bond with D546 of the catalytic loop, which lies deeper in the C-lobe (see Fig. 3 B). These interactions were often seen when the ATP-binding pocket was in a closed configuration.

Figure 4.

Effect of ATP-Mg on the conformation of the activation loop. (A) The RMSF of the Cα atoms calculated for the time range from 10 to 100 ns is shown for the three nonphosphorylated apo (np), the three phosphorylated apo (p), and the ATP-Mg simulation of KD⁼ with the structured activation loop (AL, red). When ATP is included, the RMSF of the second part of the activation loop is reduced (see inset). (B) Corresponding B factors from the crystal structure (PDB 2J0L). (C) ATP forms a salt bridge with K583 of the activation loop. (D) The distance d between the gamma phosphate of ATP and the amine nitrogen of K583 is plotted over time. (E and F) Covariance matrix calculated from the motion of Cα atoms of one KD⁼ simulation without and with ATP-Mg.

Figure 3.

Mechanism of the transient closures of the nucleotide-binding pocket. (A) In KD^≈ with the unstructured activation loop (red and orange), L567 is mainly pointing toward the αC helix (violet) and Q432 can form several hydrogen bonds with residues in the activation loop. Note that the hydrogen bond partners (residues 566–572) are resolved in the crystal structure. (B) In KD⁼, the activation loop (red) is structured. L567 is less flexible and points toward the catalytic loop (green). Q432 forms a stable side-chain hydrogen bond with D546, which is part of the catalytic loop.

In contrast and when simulating how KD^≈ with an unstructured activation loop equilibrated after removal of the inhibitor, the αC helix did not swing down during our simulations (Fig. 2 C). Instead, interactions between the αC helix and the activation loop kept the nucleotide-binding pocket open. For instance, Q432 of the G-loop had several hydrogen bond partners with the unstructured activation loop in KD^≈ (L567, S568, E572, Fig. 3 A). Note that these hydrogen bond partners are resolved in the crystal structure and were not part of the modeled loop. Additionally, the hydrophobic residue L567, which is part of the activation loop, pointed mostly toward the αC helix in the unstructured conformation. Compared to KD⁼, the flexibility of these residues was greatly enhanced. In one out of six simulations of the KD^≈, the nucleotide-binding pocket closed transiently for 24.3 ns (Fig. 2 E). During this period, D430 from the C-lobe formed two hydrogen bonds with R550 of the N-lobe. In contrast to the KD⁼ simulations, the αC helix did not swing down. It was kept at its initial position through interactions with the activation loop. The hydrogen bonds between the C- and N-lobe were not stable in KD^≈, the pocket opened again and remained open for the rest of the simulation. In summary, KD⁼ and KD^≈ promote opposite relaxation processes within the nucleotide-binding site of FAK.

Stabilization of the structured activation loop by ATP-Mg

Upon insertion of ATP together with an Mg ion into the nucleotide-binding pocket of KD⁼ by mutating ANP of the crystal structure to ATP, the fluctuations between the N- and C-lobe were lowered. The presence of ATP thus stabilized the protein and increased the correlation of the motions of the two lobes (see Fig. 4, E and F). Most importantly and in addition to this general stabilizing effect, ATP directly interacted with the structured conformation of the activation loop. The local RMSF of the carbon alpha atoms of the activation loop decreased when ATP was bound (Fig. 4 A). The γ-phosphate formed a salt bridge with the amine of K583 in the activation loop (Fig. 4 C). The distance between the two atoms is plotted over time in Fig. 4 D. During the simulated 100 ns, the salt bridge was present 72% of the time. This interaction lowered the RMSF of the activation loop between the second short β-strand and the APE motif (residues 578–592, Fig. 4 A, inset).

Electrostatic switch that connects the activation loop to the catalytic site

Because the catalytic activity of FAK was reported to be much higher when the activation loop is phosphorylated at Y576 and Y577 (15,42), we asked how the phosphorylated residues of the activation loop would interact with the catalytic core. An electrostatic network of conserved charged residues, including the phosphorylated tyrosines, connects the activation loop to residues of the catalytic site, which is thought to regulate catalytic activity (28). The phosphorylated tyrosine residue 577 interacted in KD⁼ with R545 and R569 (Fig. 5 A). In this conformation, the catalytically important residue D546 and the ATP positioning K454 and E471 were in a position favorable for catalysis, and did not interact with the two arginine residues R545 and R569. It was observed in one simulation of KD^≈ that this electrostatic network can be switched off. While Y577 was far away from the active site, E471 in the activation loop flipped and interacted with R545 and R569 (Fig. 5 B). This inactive, switched-off state was reached rapidly in one simulation and remained stable over the simulated 130 ns. Although we did not observe differences of the nanosecond ATP-pocket dynamics upon phosphorylation of the activation loop (Fig. 2 G), we show that the phosphorylated residues Y576 and Y577 of the activation loop interact with catalytically important residues through a network of charged residues.

Figure 5.

Electrostatic switch that controls catalytic activity. (A) In the on-state of the switch, the phosphotyrosine pY577 interacts with R569 and R545 and the characteristic salt bridge of an active kinase between K454 and E471 is formed. Note that the activation loop is structured (KD⁼) (B) In the off-state, E471 does not interact with K454, but with the two arginine residues R569 and R545. The activation loop (red) is unstructured (KD^≈) and pY577 is distant to R545 and R569. (C) The two mutated residues E581 and E578 of SuperFAK⁼ interact electrostatically with R569 and R545.

Finally, it is known that point mutations in the activation loop (K578E, K581E) can increase the catalytic activity of FAK’s KD in vitro, which leads to the name superFAK (32). Upon removal of either the ATP analog, or of the inhibitor staurosporine, and by mutating residues K578 and K581, the simulations of the SuperFAK showed the same behavior as seen previously for the wild-type: the ATP-binding pocket was structurally unstable when the activation loop was structured, but remained open when the activation loop was unstructured. The activating mutations K578E and K581E of the SuperFAK both interacted via salt bridges with R569 and R545 (Fig. 5 C). In one simulation, E578, and in another, E581 interacted electrostatically with R569, very similar to pY577 in the phosphorylated-KD⁼ simulations. Most commonly, only one of the two glutamate residues interacted with R569, and the other pointed toward the solvent, like pY576 in the KD⁼ simulations. This position is directed toward the autoinhibitory FERM domain in FAK. In Fig. 5 C, an intermediate state is shown, where the two mutated residues interacted simultaneously with R569.

Discussion

Two different conformations of FAK’s KD are known, one of the phosphorylated kinase in complex with a nonhydrolysable ATP analog that is catalytically active (Fig. 1 B), as well as one of the inactive KD in the presence of an inhibitor (Fig. 1 C) (15). We could therefore study the dynamics of the relaxation processes within KD upon either removal of the ATP analog (Fig. 2 A; KD⁼) or of the inhibitor (Fig. 2 A; KD^≈). The most significant finding of our MD simulations is that the ATP-binding pocket is structurally unstable upon removal of the ATP analog from FAK’s KD⁼, which displays the structured activation loop (Fig. 2 B and Movies S3, S4, S5, S8, and S9). In contrast, the pocket remained open after removal of the inhibitor from KD^≈ in its catalytically inactive configuration that displays the unstructured loop (Fig. 2 B, Movies S1, S2, S6, and S7). This outcome was perhaps unexpected, because one could have expected that the active KD⁼ has a stably opened ATP-binding pocket, as is the case for PKA (43). Based on these findings, we would like to suggest that the transient closure of the ATP-binding pocket of KD⁼ reduces the on-rate of ATP, as the probability of a collision leading to a binding event is decreased when the pocket is closed. Our simulations thus suggest a multistep-structural model in which ATP might preferentially bind the inactive conformation of FAK’s KD that displays the unstructured activation loop, and not as previously expected the active conformation with the structured activation loop. MD simulations allowed us for the first time, to our knowledge, to capture that the ATP-binding pocket undergoes major structural fluctuations after removal of the ATP analog, dynamics that cannot easily be captured by experimental techniques. Approaches that could verify the proposed mechanisms include conformational sensitive fluorescence resonance energy transfer sensors or NMR spectroscopy as it was used for example to study the dynamics of cAPK (44) or PKA (45).

Positional isotope exchange experiments with [γ-¹⁸O]ATP and viscogen experiments show a slow, viscosity depending step upon nucleotide binding and after phosphoryl transfer but before product release (31). Schneck et al. proposed that the slow step (k_conf ≈ 0.1 s⁻¹) corresponds to the conformational change of the activation loop. In agreement with this, our simulations suggest that ATP might preferentially bind KD^≈ displaying the unstructured activation loop, rather than KD⁼. Additionally, we observed a stabilizing function of ATP on interlobal dynamics and fluctuations of the structured activation loop (see Fig. 3 F), suggesting that ATP-binding might subsequently facilitate the conformational transition from an unstructured to a structured display of the activation loop.

Toward the functional role of phosphorylating the activation loop of KD, it is known that the nonphosphorylated FAK can be catalytically active, but the catalytic activity of FAK in solution is much higher when the activation loop is phosphorylated at Y576 and Y577 (15,42). The phosphorylated activation loop can thereby interact with catalytically important residues through electrostatic interactions (28). Both the active and inactive crystal structures of FAK showed the characteristic salt bridge between E471 and K454. Upon equilibration of the crystal structures in explicit water, we observed here the inactive state characterized by a flip of E471 in the αC helix (Fig. 5 B), similar to crystallographic structures of other inactive kinases, for example Hck (46). Y576 and Y577 were far away from the network in the unstructured loop, leaving R569 free to interact with E471. Phosphorylation of the activation loop allows Y577 to interact favorably with catalytically important residues in the structured conformation of the loop, which might explain the increase in catalytic activity of FAK upon phosphorylation (15,42). Similarly, we found here that mutated E581 and E578 interacted with the electrostatic network (Fig. 5 C).

In other kinases, a displacement of the αC helix has been seen between the active and inactive structures and a swing of the αC helix away from the active site was related to their inactivation (16). Similar to the pocket closure here, a domain closure between the N- and C-lobe was observed in a MD simulation study of ERK2 (47). The authors of the ERK2 study considered the closure as the first step of kinase activation. Simulations by others might further suggest that the mechanism of nucleotide-binding pocket accessibility might work differently in other kinases. For instance, in CDK5, the opposite behavior to FAK was observed (23). The atomistic, metadynamic simulations show that ATP cannot bind the closed, i.e., the conformation with the unstructured activation loop. In PKCθ, MD simulations show that phosphorylation instead of activation loop conformation regulates ATP-binding pocket accessibility (26). It is plausible that different kinases show different mechanisms of activation and action, especially as the inactive conformations of kinases are diverse (17,48,49). Other kinase like PKA also show a conformational change in their catalytic cycle, but their rate constant k_conf is two to three orders of magnitude higher (50) and might be related to a different mechanism. Indeed, the conformational changes of PKA upon ligand binding are not related to the activation loop, but to small changes in an allosteric network around the ligand binding pocket and at distal sites of the kinase (45). The different mechanism might have been evolved due to the specific requirements of different signaling pathways and subcellular localization of the two kinases in their catalytically active or inactive states.

What might be the physiological significance of the results reported here for FAK? We would like to propose that the structural instability of the ATP-binding pocket of FAK’s KD slows down its catalytic rate because transient closures of the ATP-binding pocket will reduce the ATP on-rate. Taking these short cat naps before starting a next catalytic cycle might give the phosphorylated substrate sufficient time to diffuse away from FAK’s KD, rather than to rebind. Adjusting the ATP on-rate kinetics of a kinase in the catalytic cycle might be particularly important in environments where the diffusion rate of the substrate is slowed, for example by spatial confinement or by crowded conditions. FAK is one of the most upstream tyrosine kinases orchestrating the formation of focal adhesions (8–10). FAK is therefore recruited to the plasma membrane where it is subsequently catalytically active in crowded protein environments (51,52). In the crowded environments of focal adhesions, it has been shown already that the diffusion rate of various proteins is slowed compared to that in the cytosol (53). Tuning catalytic cycle rates of a kinase to the substrate diffusion rates by transient closures of the ATP-binding pocket might decrease the risk of a blocking association of a phosphorylated substrate. One could also imagine that binding of a regulatory protein to the kinase domain of FAK could perhaps alter the pocket dynamics and consequently influence the catalytic cycle.

Finally, FAK overexpression is observed in many cancer cells and often associated with poor prognosis (14). Several ATP-competitive (54–57) and allosteric (58) inhibitors of FAK were developed recently, some of which are in clinical trials (12). The insights of this study on the mechanism of FAK might furthermore provide hints for a rational, structure-based design of inhibitors. Most kinase inhibitors bind competitively to the ATP-binding pocket. As the pocket-forming residues are highly conserved, selectivity is a major issue (59). Given that the kinase has to undergo conformational changes before and after phosphoryl transfer, small molecules that block the kinase in any state of the catalytic cycle could inhibit its activity. For instance, one could target the ATP-bound kinase and stabilize this conformation. Candidate target sites include regions that are less conserved than the ATP-binding pocket, which could lead to inhibitors with improved kinase specificity.