Previous work from our group used tyrosine kinases from the fibroblast growth factor receptor (FGFR) family of RTKs to identify four regions that form a long-range allosteric network which relays conformational changes between the A-loop to the kinase hinge located between the N- and C-lobes of the enzyme. These regions are: 1) the A-loop plug, 2) the αC tether, 3) the DFG latch, and 4) the molecular brake (Chen et al., 2017). Disruption of hydrogen bonds among N549, E565, and K641 at the molecular brake of FGFR2 kinase (FGFR2K) causes a conformational shift from an autoinhibited to an activated state (Fig. 1a) (Chen et al., 2007). To understand the mechanism driving this population shift (Chen et al., 2013; Chen et al., 2007; Chen et al., 2017; Huang et al., 2013; Nussinov et al., 2014), we used differential scanning calorimetry to assess the thermal stability of FGFR2K mutants harboring pathogenic mutations at each of the three molecular brake residues (i.e., N549K, E565A, K641R). Unphosphorylated and A-loop phosphorylated wild-type FGFR2K and the FGFR2K K659E mutant were used as controls. The latter mutation, which maps onto the kinase A-loop, introduces intramolecular hydrogen bonds that stabilize the kinase active state, thereby mimicking the action of A-loop tyrosine phosphorylation (Chen et al., 2013). Thermogram data showed that N549K, E565A, K641R mutants all had diminished melting temperatures, implying a mutation-induced destabilizing effect (Fig. 1b). Similar to N549K of FGFR2K, we also observed a reduction in the melting temperature for the analogous mutation (N535K) of FGFR4K (Fig. S1), which supports the conclusion that mutations at the molecular brake for other FGFRK isoforms induce a similar effect. In contrast to the reduced thermal stability for molecular brake mutants, the pathogenic K659E mutant and the A-loop phosphorylated FGFR2K displayed melting temperatures similar to that of the unphosphorylated kinase (Fig. S1). We conclude that in contrast to the K659E mutation, N549K, E565A, and K641R elevate the lowest energy conformation within the kinase free energy landscape, thereby facilitating the shift in equilibrium toward the active conformation (Fig. 1c). Importantly, the diminished thermal stability of the E565A enzyme explains why this mutation, in spite of imparting greater kinase activation compared to A-loop mutants, leads to a clinically milder form of skeletal disorder (Chen et al., 2017).

To delve deeper into the mechanism whereby molecular brake mutations destabilize the inhibited conformation, we used NMR spectroscopy, a technique capable of detecting site-specific protein dynamics under native conditions (Boehr et al., 2006; Kay, 2016; Kumar et al., 2018; Manley and Loria, 2012). Specifically, we carried out ¹H-¹H dipolar cross-correlated relaxation rate (η) experiments in order to calculate generalized order parameters (S²) for regions spanning throughout the kinase structure (Sun et al., 2011). The S² value ranges from zero to one and reflects fully disordered or ordered regions, respectively. To perform these measurements, we isotopically enriched wild-type FGFR2K and the most enzymatically active molecular brake mutant (i.e., E565A) with ¹³C at isoleucine, leucine, and valine in their methyl groups (ILV labeling (Rosen et al., 1996)). The S² and η values for both samples are presented in Fig. 2a and Table S2. Analysis of these values shows that these measurements could distinguish flexible loops from the rigid core of the kinase (Fig. S2). For example, the majority of N- and C-lobe residues involved in core hydrophobic packing such as V562 (in the β5 sheet), I696 (in the αF helix), and I760 (in the αI helix) displayed S² values above 0.6, indicative of order on a nanosecond timescale (Fig. 2a). On the other hand, loop residues such as I503 (in the β2-β3 loop), I541 (in the αC-β4 loop), and I651 and I654 (both in the A-loop) showed reduced order (i.e., S² values < 0.4), thus validating the suitability of our NMR approach for gauging differences in conformational dynamics induced upon mutation at the molecular brake.

To visualize changes in intrinsic dynamics between wild-type FGFR2K and the E565A mutant, differences in S² values (ΔS²) were mapped onto the inhibited FGFR1K crystal structure (PDB ID: 3KY2) (Fig. 2b, c). This plot revealed that the majority of residues in the C-lobe showed no significant differences, implying that the C-lobe is unaffected by the molecular brake mutation. However, the S² values for several sites within the N-lobe were significantly reduced in E565A relative to the wild-type enzyme. Specifically, we observed that residues in close vicinity to the E565A mutation had ΔS² values that were significantly reduced (e.g., V512, L551, V564), suggesting an increase in local dynamics on the nanosecond timescale. In addition, we also observed that two constituents surrounding the phenylalanine from the hydrophobic DFG latch, namely I541 and L617, incurred significant reductions in S² values (Fig. 2c). These residues, together with I547, are involved in locking the phenylalanine within the DFG motif in an inactive conformation (Chen et al., 2017). We conclude that the loss of hydrogen bonding seen in the E565A crystal structure leads to increased conformational flexibility around the DFG motif phenylalanine residue as well as a reduction in thermal stability observed using DSC (Fig. 1b).

To map the allosteric trajectory, we further investigated high-resolution crystal structures of active and inhibited FGFRKs (PDB IDs: 3KY2-inhibited and 2PVF-active). Our analysis revealed that I547 undergoes two major changes in its interaction patterns. First, hydrogen bonding distances between the carbonyl oxygen of I547 with both the side chains of N549 and K641 increase upon transition to the active conformation: N549 to I547 is 3.0 Å in the inhibited conformation and 6.9 Å in the active conformation, while K641 to I547 is 3.3 Å in the inhibited conformation and 4.6 Å in the active conformation (Fig. 3a). Second, the chi1 angle of the I547 side chain switches from a positive gauche conformation (59°) in the inhibited state to a negative gauche conformation (−68°) in the active form. These changes are accompanied by a difference in the F645 chi1 dihedral angle from 56° in the inhibited state to −69° in the activated state. Based on these coupled changes between molecular brake residues and I547 and the changes in I547 and F645 chi1 angles, we infer that I547 acts as a key intermediary of allosteric communication between the molecular brake and the DFG latch.

To gain further insight into the coupling mechanism between the molecular brake and the DFG latch, we utilized all-atom molecular dynamics simulations starting from an inhibited FGFR1K crystal structure bound to ATP and Mg²⁺ (PDB ID: 3KY2). Simulations were also carried out for E565A (Chen et al., 2017). We selected the FGFR1 isoform structure because we previously found that the A-loop conformation in the inhibited FGFR2K crystal structure was influenced by crystal packing, and thus did not represent an authentically inhibited kinase(Chen et al., 2017). Molecular dynamics were simulated for 200 nsec using three separate replicates, giving a total sampling time of 600 nsec for wild-type and the E565A mutant. The root mean squared deviation (RMSD) for heavy atoms (CA, C’, N, O) within structured regions of the kinase indicated stable trajectories for all replica runs (Fig. S3a). To correlate simulation data with the X-ray crystallographic and NMR data, we evaluated conformational fluctuations around the DFG latch by calculating the chi1 angle of the DFG phenylalanine throughout each trajectory (Fig. 3b and 3c). These plots show that E565A had a greater number of chi1 rotamer switches compared to the wild-type kinase. Specifically, E565A underwent dihedral angle changes between approximately 60° and −60°, in line with our analysis of crystal structures corresponding to inhibited and active states of positive gauche and negative gauche, respectively. These findings are consistent with our NMR spectroscopic data showing a decrease in order parameters around the DFG latch, thus reinforcing the existence of allosteric coupling between the molecular brake and DFG latch.

To further characterize the coupling mechanism between the molecular brake and the DFG latch regions, we quantified the number of hydrogen bonds formed by K641 and N549 with the carbonyl backbone of I547 and analyzed the chi1 angles of I547 in addition to those detailed above for F645. The simulation results for the wild-type enzyme revealed that the chi1 angles of F645 and I547 predominantly sampled the inhibited rotamer conformation (78.7%), whereas the active rotamers constituted only a minor population (3.7%). The remaining population (intermediates 1 and 2) showed mixed characteristics in which either the F645 rotamer resembled that of the inhibited state and I547 resembled that of the active state (15.3%), or F645 in the active state and I547 in the inhibited state (2.3%) (Fig. 3b). The extent of hydrogen-bonding indicated that the molecular brake was primarily intact, as evidenced by the high probability of two hydrogen bonds between the carbonyl oxygen of I547 and the side chains of N549 and K641 (Fig. 3b).

In stark contrast to the case of the wild-type enzyme, the corresponding populations of the E565A mutant for the inhibited and active forms were reversed: 7.8% populated the inhibited conformation and 49.8% populated the active state (Fig. 3c). Intermediate states were also present in which active F645 and inhibited I547 chi1 angles comprised 16.3%, while the opposite conformational states had a population of 26.1%. The higher percentage of intermediate rotamer states implies greater flexibility of the E565A mutant at the molecular brake region relative to the wild-type kinase. The E565A mutation also imparted larger perturbations on the hydrogen bond network of the molecular brake, as anticipated from loss of the salt-bridge involving this glutamic acid in the wild-type kinase (Fig. 3c). We conclude that E565A weakens hydrogen bonding interactions at the molecular brake, thereby lowering the free energy barrier for chi1 rotameric changes of I547 and F645 (Fig. 1c).

It is noteworthy that the overall RMSD of structured and unstructured regions in both wild-type and E565A simulations remained relatively small over the total simulation time of 600 nsec (Fig. S3a). While we observed local conformational fluctuations around the DFG latch and at the molecular brake (Fig. S3b), we found no evidence for transitioning of the A-loop to an active configuration (Fig. S3c). It has been reported that simulations on the millisecond timescale are required to observe A-loop conformational changes in EGFR kinase (Shan et al., 2013). For this reason, conformational changes of I547 and F645 were considered to be active-like transitions and may not represent a global transition to the active state. This is consistent with our previously reported activity data showing subtle cooperativity between the DFG-latch with the A-loop or the molecular brake with the A-loop (Chen et al., 2017). Namely, double mutations at the DFG-latch and A-loop or at the molecular brake and A-loop produced an additive effect of individual mutations (Chen et al., 2017). In contrast, combined mutations at the DFG-latch and molecular brake produce no greater activity than each mutation alone. Taken together, these observations support conformational coupling via a defined allosteric pathway involving hydrogen bonding of I547 to the molecular brake, which directly influences the orientation of the phenylalanine from the DFG latch.

To determine whether our allosteric coupling pathway between activated and inhibited kinase conformations in the FGFR family is applicable to other kinases, we analyzed crystal structures for tyrosine kinases containing isoleucine and phenylalanine residues at the analogous positions to I547 and F645 within FGFR2K. Because of the known influence of ATP-competitive inhibitors on the DFG conformation (DFG_in or DFG_out) (Muller et al., 2015), we focused on drug-free kinase structures. In addition, we included only those tyrosine kinases with glutamic acid and lysine residues corresponding to E565 and K641 of FGFR2K; note that these account for >85% of all tyrosine kinases (Fig. S4a). The calculated chi1 angle plots for phenylalanine and isoleucine residues showed clustering of structures with inhibited and activated side chain dihedral angles similar to those in the FGFRK family (Fig. 4). Specifically, out of 101 structures analyzed, 14 had chi1 angles that clustered in the inhibited conformation, while 80 were found within the active conformation. The remaining seven kinase structures had mixed characteristics in which the chi1 angle for the phenylalanine of the DFG motif was observed in an active conformation and the isoleucine was in an inhibited conformation. Notably, in these latter structures, the phenylalanine conformation may be biased because of the tightly bound sulfate ion by A-loop residues.

We also measured bond distances between the carbonyl oxygen of the analogous I547 residue and the side chain nitrogen at the analogous K641 position to further explore correspondence of the DFG latch orientation with hydrogen bonding at the molecular brake: longer bond distances would imply disrupted hydrogen bonding at the molecular brake, while shorter distances would be indicative of intact hydrogen bonding. Remarkably, we found that 12 of 14 inhibited structures had bond distances less than 3.5 Å, while 73 of 80 structures had bond distances greater than 3.5 Å. This dichotomy suggests a general activation mechanism that connects hydrogen bonding at the molecular brake to the rotamer orientation of the phenylalanine within the DFG motif. We note that some of the tyrosine kinases in this analysis do not share the asparagine position of the molecular brake conserved in FGFR, VEGFR, PDGFR, c-KIT, TIE, TEK, CSF1R, and FLT3. However, these kinases do have a hydrogen bonding pattern that resembles those interactions identified as part of the molecular brake (Chen et al., 2007). In particular, their structures show a conserved hydrogen bond from the backbone nitrogen of the analogous N549 position to the side chain carboxylate of E565 (Fig. S4b, c). This observation coupled with the high degree of conservation at the analogous E565 and K641 positions, implies that the molecular brake is a generalized regulatory element within the tyrosine kinase subfamily (Molina-Vila et al., 2014).

Our experimental and computational data lead us to conclude that mutations at the molecular brake impart local disorder, which is transmitted to the DFG latch via a change in the backbone hydrogen bonding distance to I547. The loss of hydrogen bonding alters the I547 side chain orientation which in turn directly influences the side chain conformation of the phenylalanine from the DFG motif. Thus, mutations at the molecular brake operate via an allosteric propagation mechanism in which there is a cooperative conformational transition from the site of mutation to the DFG latch. The timescale of A-loop motion is too slow to be captured using nanosecond MD simulations described in this work. Nevertheless, NMR chemical shift perturbations on A-loop residues I651 and I654 (Fig. S5) suggest that molecular brake mutants can sample the active state of the DFG latch component of the allosteric network implying that they can indeed populate the active state. This is evidenced by the remarkable consistency of active and inhibited A-loop conformations with the corresponding F645 and I547 chi1 angles observed in crystal structures where the A-loop has ample time to sample different conformations during the crystallization process.

Further information and requests for resources and reagents should be directed to and will be fulfilled by Nathaniel J. Traaseth (traaseth@nyu.edu)

cDNA fragments encoding residues P458 to E768 of human FGFR2c (Accession code: P21802-1) and residues L445 to E753 of human FGFR4 (Accession code: P22455-1) were amplified by PCR and sub-cloned into a pET bacterial expression vector with an N-terminal 6x-His-tag. Note that wild-type FGFR2 kinase was mutated at C491 to an alanine; this functional mutant was used for all subsequent site-directed mutants. All pathogenic mutations at the molecular brake were introduced using the QuikChange site-directed mutagenesis kit (Stratagene). The E. coli strain BL21(DE3) cells were transformed with the expression constructs, and kinase expression was induced with 0.1 mM (FGFR2K) or 1.0 mM (FGFR4K) isopropyl-L-thio-B-D-galactopyranoside overnight at 20 °C. The cells were lysed, and the soluble kinase proteins were purified using nickel affinity chromatography as previously described (Chen et al., 2007; Chen et al., 2017). Trace amounts of phosphorylation from the bacterial expression were removed by treating the kinases with FastAP™ Thermosensitive Alkaline Phosphatase (Thermo Scientific) and subsequently purified by anion exchange chromatography (Source 15Q, GE Healthcare Life Sciences) and gel filtration chromatography (Superdex 200 Increase 10/300GL, GE Healthcare Life Sciences).

¹H-¹H dipolar cross-correlated relaxation rates (η) were calculated from the intensities of the triple (forbidden) and single (allowed) quantum transition experiments as previously described (Sun et al., 2011). The relaxation delays (T) used for the triple and single quantum measurements were 1.25, 2.5, 5, 7.5, 10, 12, 15, and 18 msec. The forbidden experiment was acquired using 24 scans, while the allowed experiment was acquired using 16 scans. To correct for differences in peak heights due to unequal scans between the two experiments, the peak heights obtained from the forbidden experiment were multiplied by a factor of (16/24). Intensity ratios for each relaxation delay (I_forbidden/I_allowed) were fit to determine the ¹H-¹H dipolar cross-correlated relaxation rate using equation (3): (3)∣IforbidIallow∣=Cηtanh(η+δ2T)η+δ2−δtanh(η+δ2T C is 0.75 and δ compensates for relaxation effects arising from external protons; the latter was a floatable parameter in the fit. The acquisition parameters for 2D experiments utilized direct (¹H) and indirect (¹³C) spectral widths of 10000 and 3018 Hz and acquisition times of 60.0 msec and 19.6 msec, respectively. All measurements were performed in duplicate for FGFR2 kinase and E565A, which were 0.5 mM in concentration and prepared uniformly ¹⁵N labeled with –CH₃ groups labeled with ¹³C at Ile (C^δ1), Leu (C^δ1/δ2), and Val (C^γ1/γ2) in a perdeuterated background (referred to as ILV labeling). All experiments were carried out at 10 °C. Data were processed in NMRPipe (Delaglio et al., 1995) and analyzed in Sparky (Goddard and Kneller)

To calculate order parameters (S²) from the fitted η values in equation (3), we measured correlation times for wild-type and E565A kinases using the TRACT method (Lee et al., 2006). Datasets measuring the relaxation rates of the TROSY and anti-TROSY components were acquired in an interleaved fashion with relaxation delays from 0 to 200 msec in in 4 msec intervals. Relaxation curves were fit to an exponential decay to extract the relaxation rate of each TROSY component and the correlation time was determined as previously described (Lee et al., 2006). The correlation time (τ_C) was determined to be 27 nsec for both wild-type and mutant FGFR2Ks.

Order parameters (S²) were calculated using the following equation (Sun et al., 2011): (4)η=R2,HF−R2,HS2≈910(μ04π)2[P2(cosθaxis,HH)]2Saxis2γH4ħ2τCrHH6 where ℏ is the reduced Plank’s constant, r_HH is the distance between methyl group protons (1.813 Å), and P₂(cosθ_axis,HH) was set to −0.5 (Tugarinov et al., 2007). Errors in Figure 2 reflect non-linear least square fitting errors from datasets acquired in duplicate.

All measurements were performed on a NanoDSC (TA Instruments) at a scanning rate of 60 °C/hr from 10 °C to 70 °C. Kinase concentrations were 20 or 80 μM in 25 mM HEPES, 150 mM NaCl, pH 7.5. For AMP-PNP measurements, the concentrations of protein, AMPPNP, and magnesium chloride were 80 μM, 5 mM, and 10 mM, respectively. For phosphorylated samples, 1 mM EDTA was included in the DSC buffer to chelate Mg²⁺. Melting temperatures (T_m) were fit using the NanoAnalyze software (TA instruments).

Molecular dynamics simulations were performed using the AMBER 16.06 package with the ff14SB force field for the protein and TIP3P water model. Force fields developed for ATP(Meagher et al., 2003) and Mg²⁺ (Allner et al., 2012) were used in the nucleotide bound simulations. The initial structure for molecular dynamics simulations of wild-type enzyme was modeled based on the inhibited FGFR1 kinase structure 3KY2, with missing loops repaired using Modeller (Webb and Sali, 2017). The initial structure for the molecular brake mutant was obtained by mutating E565 into alanine. Wild-type and E565A mutant systems were neutralized using three and two Na⁺ ions, respectively, and were solvated using a cubic TIP3P water box with a 15 A buffer region between protein and boundary.

Each prepared simulation system was equilibrated using repeated steps of minimization and restrained dynamics. The solvated complex was initially minimized using a conjugate gradient minimization for 2000 steps followed by a 200 psec constant volume simulation at 300K with a 500 kcal/mol restraint on the crystal waters, cofactors, and the heavy atoms of the protein. Another round of conjugate gradient minimization for 2000 steps were performed using the same restraints followed by another round of minimization but with reduced restraints of 50 kcal/mol. A series of 200 psec constant volume simulations were performed with decreasing restraints from 50, 10, and 2 kcal/mol at 300 K. Lastly, two constant pressure simulations were run for 200 psec each at 300K with decreasing restraints of 2 and 0.5 kcal/mol. Finally, unrestrained production runs were initiated from the density equilibrated simulations with snapshots being collected every 10 psec. For all simulations, time steps were set at 2 fsec, with SHAKE constraints, particle mesh-Ewald (Salomon-Ferrer et al., 2013) for electrostatics and a 12 Å cut-off for non-bonded interactions.

All experimental DSC and NMR order parameter measurements were performed in duplicate. All listed and plotted points represent the average of two experiments; error in the DSC measurements reflect the standard deviation while errors associated with η and S² values represent the errors of the fit. The fits to equation (3) for the NMR order parameter measurements were performed in Prism 7 (Graphpad).