Gatekeeper mutations activate FGF receptor tyrosine kinases by destabilizing the autoinhibited state

Significance

Anticancer drugs inhibiting the kinase domain of receptor tyrosine kinases block downstream phosphorylation necessary for cellular signaling and growth. However, a major impediment to successful cancer therapy is the emergence of drug resistance mutations at a residue situated at the kinase hinge region between the N- and C-lobes, which provides fitness advantages for tumor growth. In this work, we used kinase assays, NMR spectroscopy, and MD simulations to determine the mechanism of kinase activation from drug-resistant gatekeeper mutations. We found that these mutations destabilize the autoinhibited conformation of the kinase by weakening the hydrophobic spine and modulating the conformation of the conserved DFG motif, thereby leading to the conclusion that gatekeeper mutations increase in the population of the active state.

Abstract

Many types of human cancers are being treated with small molecule ATP-competitive inhibitors targeting the kinase domain of receptor tyrosine kinases. Despite initial successful remission, long-term treatment almost inevitably leads to the emergence of drug resistance mutations at the gatekeeper residue hindering the access of the inhibitor to a hydrophobic pocket at the back of the ATP-binding cleft. In addition to reducing drug efficacy, gatekeeper mutations elevate the intrinsic activity of the tyrosine kinase domain leading to more aggressive types of cancer. However, the mechanism of gain-of-function by gatekeeper mutations is poorly understood. Here, we characterized fibroblast growth factor receptor (FGFR) tyrosine kinases harboring two distinct gatekeeper mutations using kinase activity assays, NMR spectroscopy, bioinformatic analyses, and MD simulations. Our data show that gatekeeper mutations destabilize the autoinhibitory conformation of the DFG motif locally and of the kinase globally, suggesting they impart gain-of-function by facilitating the kinase's ability to populate the active state.

Intracellular and intercellular signaling by receptor tyrosine kinases (RTK) play essential roles in the regulation of cell growth, differentiation, metabolism, apoptosis, and senescence (1). Not surprisingly then, aberrant RTK signaling arising from amplifications, mutations, or their misexpression is implicated in a wide array of human developmental, metabolic and immune disorders, and cancer (2–4). Underscoring their importance in cancer, a recent study analyzing over 9,000 tumor samples from 33 cancer types found that alterations in genes encoding RTKs and constituents of the RAS signaling pathway dominate the landscape of altered genes in cancer (4). RTK gain-of-function mutations typically cluster within regulatory regions of the tyrosine kinase domain, including the juxtamembrane, activation loop (A-loop), and kinase hinge (5, 6). Epidermal growth factor receptor (EGFR) and fibroblast growth factor receptor (FGFR) are two of the most frequently mutated RTK families in human cancers (4). These mutations drive oncogenicity typically by subverting physiological mechanisms of ligand-induced kinase domain activation leading to uncontrolled cell growth and promotion of angiogenesis (1), key hallmarks of cancer. Accordingly, inhibition of the tyrosine kinase domain of RTKs via ATP-competitive small molecule inhibitors is the mainstay therapy for a multitude of solid and hematological cancers (6, 7).

Following the first successful kinase inhibitor Imatinib (Gleevac) (8, 9), there are now 71 FDA-approved kinase drugs with 39 targeting RTKs involved in cancer (10). A major impediment to successful cancer therapy is the emergence of drug-resistant mutations at a residue within the kinase hinge region (11–14), which provides fitness advantages for tumor growth and hinders long-term remission (15). This location is often occupied by a threonine or valine, and in most instances, it becomes mutated to a larger hydrophobic residue most frequently isoleucine, methionine, leucine, and phenylalanine (Fig. 1) (16). The bulkier side chains sterically block access of the inhibitor to the hydrophobic pocket at the back of the ATP-binding cleft leading to poorer binding affinity and drug efficacy (14, 17–20). As a result, this residue is referred to as the gatekeeper in kinase jargon.

Fig. 1.

Distribution of gatekeeper mutations within tyrosine kinases from the kinase mutations and drug response database (16). Heatmap of gatekeeper mutations, where the native gatekeeper residue is displayed on the Y-axis and the residue it is mutated to is shown on the X-axis. The greatest occurrence of mutations is colored in red and stems from the same mutation observed in different tyrosine kinases and in multiple tumor samples; white boxes correspond to no observed mutations. Amino acid residues are ordered from hydrophobic to hydrophilic in the Left to Right and Bottom to Top directions.

An underappreciated fact regarding gatekeeper mutations is that they elevate the kinase activity of the affected RTK. Understanding the molecular basis for this gain-of-function effect is important because the enhanced kinase activity of gatekeeper mutants correlates with greater signaling and worse patient outcomes (14, 21–23). It has been proposed that gatekeeper mutations encourage the active state of the kinase by strengthening the hydrophobic regulatory spine (R-spine) (24–26), a network of four hydrophobic residues connecting the N- and C-lobes of the kinase (Fig. 2A) (27). However, a couple of observations are inconsistent with this hypothesis. First, for many kinases, an intact hydrophobic R-spine occurs in both the autoinhibited and active states (28, 29). Second, certain gatekeeper mutations introduce hydrophilic residues, including in ALK (L1196Q), EPHA2 (T692N), FGFR3 (V555E), FGFR4 (V550E), and KIT (T670E) (16, 30) (Fig. 1). Although infrequent, such hydrophilic substitutions are difficult to reconcile with the hydrophobic R-spine hypothesis. Therefore, investigation of both hydrophobic and hydrophilic gatekeeper mutations would offer a comparative approach for elucidating mechanisms of activation.

Fig. 2.

Gatekeeper mutants display increased phosphorylation activity in vitro. (A) Cartoon representation of FGFR2K structure (PDB ID: 2PVF) highlighting the catalytic and regulatory spines (C- and R-spines). Left: Whole view of the kinase structure with the nonhydrolyzable nucleotide AMP-PCP (gray), gatekeeper residue (blue), and hydrophobic R- and C-spine residues (gray) represented in sticks. The surface of gatekeeper and hydrophobic spine residues are also displayed in blue and gray, respectively. Mg²⁺ atoms are shown in green spheres. Right: expanded view of the gatekeeper and hydrophobic spine residues represented as in the Left panel. (B) Phosphorylation assays of wild-type FGFR2K and its gatekeeper mutants evaluated using native gel electrophoresis (Top) and IB analyses with an antibody specific for A-loop phosphorylated FGFR at both Y656 and Y657 (Bottom). (C) Quantification of phosphorylation at A-loop tyrosines (Y656/Y657) from the IB data in panel (B) for FGFR2K and gatekeeper mutants. Intensities for each respective kinase sample are normalized to a value of 1 for the 20-min time point.

Here, we used a multipronged approach consisting of kinase assays, NMR spectroscopy, and molecular dynamics (MD) simulations to determine the mechanism of RTK activation by drug-resistant gatekeeper mutations using FGFR2 kinase (FGFR2K) as the model RTK system. Our findings indicate that gatekeeper mutations destabilize the autoinhibited state of the kinase and modulate the conformation of the conserved DFG motif, thereby leading to the proposed model where mutations shift the equilibrium to the active state of the kinase.

Results

Gatekeeper Mutations Elevate Trans-Autophosphorylation Activity of FGFR2K.

To assess the effects of gatekeeper mutations on FGFR2K activity, we compared kinase trans-autophosphorylation activities of V564I and V564E gatekeeper mutants with wild-type FGFR2K in vitro. V564I is an authentic FGFR2K mutation causing drug resistance (31) whereas the V564E mutation is analogous to the V550E mutation in the FGFR4 kinase domain detected in a childhood cancer (30). Each kinase sample was incubated with ATP and Mg²⁺ and analyzed for phosphorylation as a function of time using kinase mobility shifts in native gel electrophoresis and immunoblots (IBs) with an antibody specific for A-loop phosphotyrosines (pY656/pY657; “2P” state). Note that kinase autophosphorylation occurs in an intermolecular manner (1), as depicted in our asymmetric dimer structure of FGFR kinase showing A-loop trans-autophosphorylation of Y656 (32). In phosphorylation assays, both V564E and V564I gatekeeper mutants displayed an increased rate of intrinsic kinase activity as evident by earlier shifts in protein bands in native gels and earlier appearance of phosphorylated bands in IBs relative to wild-type FGFR2K (Fig. 2B). Quantification of the IBs revealed that gatekeeper mutants phosphorylated A-loop tyrosines ~threefold to fourfold faster than wild-type FGFR2K (compare 3-min time point in Fig. 2C). These results indicate that V564E and V564I gatekeeper mutants confer gain-of-function to wild-type FGFR2K in vitro.

Gatekeeper Mutations Enhance the Conformational Dynamics of FGFR2K When Bound to ATP.

Our kinase trans-autophosphorylation data imply that the gatekeeper mutations encourage the active state conformation of the kinase. To test this hypothesis, we applied Carr–Purcell–Meiboom–Gill (CPMG) relaxation dispersion (33, 34) to characterize the conformational dynamics of the kinases in the absence of ATP and Mg²⁺ cofactor (i.e., apo state) on the μsec–msec timescale. Wild-type FGFR2K and its gatekeeper mutants V564E and V564I were isotopically enriched with ¹³C at methyl groups of isoleucine, leucine, and valine residues (i.e., ILV labeling). Surprisingly, CPMG results showed no significant relaxation dispersions for either wild-type FGFR2K or the gatekeeper mutants (SI Appendix, Fig. S1), indicating the absence of detectable higher energy conformations that convert with the lowest energy state on the μs–ms timescale.

We suspected that differences in conformational dynamics between wild-type FGFR2K and gatekeeper mutants may only be captured in the ATP/Mg²⁺-bound state. Indeed, the nucleotide is essential for kinase function, as the triphosphate moiety serves as the energy source for phosphoryl transfer and the adenine base of ATP interacts with hydrophobic residues from the N- and C-lobes of the kinase to encourage formation of the catalytic spine (C-spine) necessary for kinase activation (35, 36). Hence, we repeated the CPMG experiments in the presence of ATP and Mg²⁺. However, such an experiment was technically challenging since the kinase undergoes tyrosine trans-autophosphorylation in the presence of ATP. To overcome this problem, we engineered a nonphosphorylatable version of FGFR2K by mutating all five phosphorylatable tyrosines (Y466F/Y586L/Y588P/Y656F/Y657F; “0Y” nomenclature or WT^0Y). A ¹H/¹³C HMQC spectral comparison of WT^0Y with wild-type FGFR2K showed only minor chemical shift perturbations near the mutation sites (SI Appendix, Fig. S2), which commonly arise from local changes in electrostatics introduced by the mutation. Since there were no large chemical shift perturbations in residues distant to the mutation sites, we inferred that the tyrosine mutations had a minimal effect on the kinase structure. Accordingly, we prepared “0Y” versions of both gatekeeper mutants (V564E^0Y and V564I^0Y) for subsequent measurements.

Next, we performed CPMG relaxation dispersion experiments on WT^0Y, V564E^0Y and V564I^0Y mutants in the presence of saturating concentrations of ATP (20 mM) and Mg²⁺ (40 mM) as in the kinase assay (Fig. 3 A and B and SI Appendix, Fig. S3). Due to the occurrence of ATP hydrolysis at longer incubation times (SI Appendix, Fig. S4), we reduced the total experimental time by collecting CPMG relaxation dispersion data at 50-Hz and 1,000-Hz frequencies and at two magnetic field strengths. We found that gatekeeper mutants displayed significantly larger ΔR₂ values for residues near the gatekeeper residue (L550, L551), within or proximal to the αC helix (L528, L531, V562), the A-loop (I654), near the substrate binding site (L692, I707, L716), and within or proximal to the hydrophobic R-spine (L550, V562, I563, I623, I642) (Fig. 3B). Each of these regions are known to undergo conformational changes important in kinase activation, including the A-loop, which blocks substrate access to the enzyme’s active site in the autoinhibited conformation (37).

Fig. 3.

ATP binding increases conformational heterogeneity of gatekeeper mutants. (A) ¹H/¹³C HMQC spectra of WT^0Y, V564I^0Y, and V564E^0Y in the absence (apo state; black) and presence of ATP (20 mM) and Mg²⁺ (40 mM; purple). Residues highlighted with a box indicate example residues with notable differences between WT^0Y and the gatekeeper mutants. (B) Results from CPMG relaxation dispersion experiments and intensity retention calculations plotted on the homology model of autoinhibited FGFR2K. The difference between CPMG relaxation dispersion experiments performed at two frequencies, ΔR₂, was calculated as in Eq. 1 and plotted on the homology model of autoinhibited FGFR2K. Significant ΔR₂ values are shown from 20 s⁻¹ (blue) to 2 s⁻¹ (light blue). ΔR₂ values less than 2 s⁻¹ are shown in white. Peaks with intensity retention (I_R) values less than 0.25 from panel (A) are displayed in purple. Hydrophobic R-spine residues are displayed in green; residue labels in bold correspond to the gatekeeper residue and hydrophobic R-spine residues. Note that some residues displayed missing peaks in the CPMG experiment for WT^0Y (L647, L675) and V564E^0Y (L675) and could not be analyzed; these residues are not displayed in the respective plots. (C) I_R values for HMQC spectra from panel (A) plotted on the homology model of autoinhibited FGFR2K. I_R values were calculated as the intensities of ATP and Mg²⁺-bound states divided by the intensities of the apo states. I_R values plotted range from 0 (purple) to 1 (white). Hydrophobic R-spine residues are displayed in green; residue labels in bold correspond to the gatekeeper residue and hydrophobic R-spine residues.

The increased activity and ΔR₂ values for V564E^0Y and V564I^0Y mutants relative to WT^0Y suggested a population shift away from the major state of WT^0Y (see Methods section CPMG Data Interpretation). Thus, it became imperative to define this major conformation of WT^0Y when bound to nucleotide and Mg²⁺ (i.e., autoinhibited or active state). To do so, we designed a paramagnetic relaxation enhancement (PRE) experiment to measure distances between an MTSL spin label attached to a cysteine in the C-lobe (I707C) and tyrosine residues within the A-loop. Based on FGFR kinase crystal structures (PDB IDs: 3KY2 and 2PVF), Y656 and Y657 are within ~20 Å from the MTSL spin label when in the autoinhibited kinase conformation and >30 Å when in the active kinase conformation. Hence, the sensitivity of the PRE experiment (≤25 Å) would be informative for deciphering the major state. ¹H/¹⁵N HSQC spectra of ¹⁵N-Tyr- and MTSL-labeled kinase were collected for the oxidized and reduced states in the presence of nonhydrolyzable nucleotide AMP-PCP and Mg²⁺ (SI Appendix, Fig. S5). Note that the rather long data collection times precluded the use of native ATP due to the occurrence of ATP hydrolysis. HSQC spectra for Y656 and Y657 within the A-loop displayed decreased intensities in the oxidized spectrum relative to the reduced dataset and corresponded to distances of ~20 Å from the spin label. This finding was consistent with the autoinhibited state and provided evidence that WT^0Y in complex with nucleotide is primarily in the autoinhibited conformation.

These PRE findings taken together with results from CPMG relaxation dispersion experiments and activity assays support a model where gatekeeper mutations induce a population shift toward the active state of the kinase. Although the two-point CPMG approach used in this work does not directly identify the minor state as the kinase active conformation, these results do indicate a population shift away from the predominant autoinhibited conformation characterized in the PRE experiment. Thus, the correlation of increased CPMG relaxation dispersion with increased activity for gatekeeper mutants relative to wild-type FGFR2K supports our activation hypothesis where gatekeeper mutants induce an equilibrium shift toward the active conformation. This activation model is also consistent with FGFR kinase crystal structures, which show the enzyme primarily crystallizes in two states, autoinhibited and active conformations (29).

ATP Binding Increases the Conformational Heterogeneity of Gatekeeper Mutants.

Notwithstanding the insights gained from CPMG relaxation dispersion data, some residues near the ATP-binding pocket could not be analyzed due to low signal intensities induced by the addition of ATP and Mg²⁺ (see purple spheres in Fig. 3B). These reductions in peak intensities further supported the CPMG results that ATP binding induced conformational dynamics on the μs–ms timescale (38). Notably, these effects were independent of the kinase concentration (SI Appendix, Fig. S6), indicating that the reduced signal intensities were not due to kinase dimerization, as we previously reported for a gain-of-function mutation in the P + 1 pocket of FGFR kinase (32). Thus, to complement CPMG data, we quantitatively compared ¹H/¹³C HMQC peak intensities of V564I^0Y and V564E^0Y with WT^0Y in the absence and presence of saturating concentrations of ATP (20 mM) (Fig. 3 A–C).

Mapping peak intensity changes onto the crystal structure revealed significantly greater reductions for V564E^0Y and V564I^0Y relative to WT^0Y in key regulatory regions of the kinase, including the R- and C-spines, DFG motif, and A-loop. Peaks corresponding to I547, I548, and L647 surrounding the DFG motif were completely absent for V564E^0Y and V564I^0Y but were only reduced in intensity for WT^0Y. C-lobe residues of the C-spine also displayed larger intensity reductions for the gatekeeper mutants than for WT^0Y (V632, L633, V634, L692, L696) (39). This observation is consistent with the role of ATP in promoting the hydrophobic C-spine and priming the active site for substrate binding and catalysis (40, 41). We also observed reductions in intensity for L550 in V564I^0Y, one of the residues within the R-spine, as well as other residues nearby the spine for V564I^0Y and V564E^0Y (L551, V562, I563, I564 of V564I only, I623, I642). In addition, peak intensity for I651 located in the A-loop and 20 Å away from the ATP binding site, decreased by ~80% for gatekeeper mutants (83.6% for V564I^0Y and 76.8% for V564E^0Y) compared with 60% for WT^0Y. These reductions/ablations in peak intensities for V564E^0Y and V564I^0Y observed in ¹H/¹³C HMQC experiments indicated that the gatekeeper mutants are conformationally more heterogenous than WT^0Y when bound to ATP. From these data, we conclude that gatekeeper mutations induce perturbations within the hydrophobic spine and A-loop, regions both involved in the conformational change from the autoinhibited to the active state. This transition involves the C-terminal end of the A-loop moving away from its autoinhibitory position in the substrate-binding pocket to enable the N-terminal portion of the A-loop to form new contacts with the αC helix (αC tether) and the catalytic loop (29).

Gatekeeper Mutants Bind ATP Tighter Than FGFR2K.

We also performed NMR titration experiments to quantify binding affinities of the resulting WT^0Y, V564I^0Y, and V564E^0Y kinases using the native ATP nucleotide and Mg²⁺ cofactor (Fig. 4A and SI Appendix, Fig. S7). Fitting chemical shift perturbation data revealed K_d values of 450 ± 70 μM for WT^0Y, 280 ± 30 μM for V564I^0Y, and 140 ± 20 μM for V564E^0Y (Fig. 4B). The 1.6-fold and 3.2-fold improvements in ATP-binding affinities of V564I^0Y and V564E^0Y relative to WT^0Y imply that conformations of gatekeeper mutants are differentially affected upon binding to ATP/Mg²⁺ relative to unphosphorylated FGFR2K. Notably, the 20 mM ATP concentration used for phosphorylation assays is at least 44-fold above the K_d values; therefore, the elevated kinase activity of the gatekeeper mutants reported in Fig. 2 cannot be attributed to the improved ATP-binding affinities of the gatekeeper mutants. Furthermore, the ATP concentration in cells is typically 1 mM and above (i.e., higher than the K_d values) (42), so ATP is expected to be bound to wild type and mutants. Finally, we also measured the binding constant of FGFR2K to the nonhydrolyzable nucleotide AMP-PCP in the presence of Mg²⁺ and observed weaker binding (K_d = 2.9 ± 0.1 mM; SI Appendix, Fig. S8), which has been reported with AMP-PNP (43). This divergence from our K_d value determined for ATP as well as reported K_m values for ATP (70 to 840 μM) (24, 44, 45) corroborated our experimental approach to assess FGFR2K conformational dynamics using the native ATP nucleotide.

Fig. 4.

Gatekeeper mutations of FGFR2K enhance the ATP-binding affinity. (A) ¹H/¹³C methyl HMQC spectra of L483 (β1 strand) at three concentrations of ATP using samples of WT^0Y (Left), V564I^0Y (Middle), and V564E^0Y (Right). Full titrations for all peaks are displayed in the supporting information. Note that the L617 peak (22.2 ¹³C ppm and 0.25 ¹H ppm) was removed from the spectra for clarity. (B). Chemical shift perturbation as a function of ATP concentrations (absolute value shown). ¹³C chemical shifts were scaled by a factor of 0.25 to account for the ¹H chemical shift range. Binding affinity was quantified using a global nonlinear least-square fit to assigned residues experiencing chemical shift perturbations in fast exchange. Individual residues correspond to the symbols as represented below. WT^0Y: ○I548^Cδ1, ▲V516^Hγ, △I548^Hδ1, ●L483^Hδ, ●V516^Hγ, △L572^Hδ, ▲ L560^Hδ, ○ L483^Cδ; V564I^0Y: ○L496^Hδ, ▲ L483^Hδ, △ V516^Hγ, ●V516^Cγ, ●V516^Hγ, △L560^Hδ, ▲L572^Hγ; and V564E^0Y: ○V516^Hγ, ▲L483^Cδ, △I623^Cδ1, ●L483^Cδ, ●L560^Hδ, △L603^Cδ, ▲L603^Hδ.

MD Simulations Reveal Gatekeeper Mutants Destabilize the Autoinhibited State.

To complement the NMR experiments, we next performed MD simulations on the ATP/Mg²⁺-bound states of wild-type FGFR2K (unphosphorylated and phosphorylated forms) and the V564I and V564E gatekeeper mutants for a total of 5 μs (five replicates) or 3 μs (three replicates) starting from the autoinhibited or active conformation, respectively. We first focused on the hydrophobic R-spine due to its proposed role in kinase activation (27, 28) and our NMR observations on the gatekeeper mutants for this region. We quantified whether the hydrophobic R-spine was intact as defined as side chain contact distances less than 4.5 Å among the four constituents of the spine (M538, L550, H624, F645). The fraction of intact R-spine was calculated every 5 psec of the MD simulations and plotted as a function of each replicate (Fig. 5). MD simulations starting from the active conformation showed a similar extent of intact R-spine formed for wild-type FGFR2Ks and gatekeeper mutants (Fig. 5 A and B). Next, we performed the same analysis on simulations starting from the autoinhibited conformation. The unphosphorylated FGFR2K possessed an intact R-spine for all replicates (Fig. 5 C and D). In contrast, the V564I and V564E gatekeeper mutants showed a broken R-spine due to the disruption of contacts between M538 and F645 in the DFG motif (Fig. 5D). Likewise, we observed disruption of the hydrophobic R-spine in the phosphorylated FGFR2K simulations, albeit through disruption of contacts between H624 and F645. Notably, our simulations displayed a comparable effect for hydrophilic V564E or hydrophobic V564I gatekeeper mutations on the hydrophobic R-spine. We also analyzed whether the hydrophobic C-spine was intact in active and autoinhibited simulations, as defined as contact distances less than 4.5 Å among the adenine base of ATP and the eight residues encompassing the spine (V495, A515, L572, V632, L633, V634, L692, I696). We found the C-spine to be intact in both wild-type FGFR2K and gatekeeper mutant simulations starting from the active state but slightly destabilized for V564E and phosphorylated wild type in simulations starting from the autoinhibited state (SI Appendix, Fig. S9). From these data, we conclude that the gatekeeper mutations confer gain-of-function by destabilizing the autoinhibited state, most significantly at the hydrophobic R-spine. Notably, this conclusion stands at odds with the accepted view in the field that gatekeeper mutations act by strengthening the hydrophobic R-spine purported to be a unique hallmark of the active conformation (24–26).

Fig. 5.

MD simulations reveal disruption of the hydrophobic R-spine within the autoinhibited state. (A and C) The percentage of R-spine formed calculated from MD stimulations every 5 psec starting from the active (A) and autoinhibited states (C) for unphosphorylated wild-type FGFR2K (WT; gray), V564I (blue), V564I (pink), and phosphorylated wild-type FGFR2K (pWT; green). Each bar represents one replicate of 1-μs sampling time. (B and D) Screenshots of representative structures for the gatekeeper (position 564) and hydrophobic R-spine residues (M538, L550, H624, F645) sampled during MD simulations starting from the active (B) and autoinhibited states (D). In each panel, the hydrophobic R-spine residues correspond to the order displayed in the Right panel. Gatekeeper and R-spine residues are shown in sticks and a surface representation.

Mechanism of Autoinhibited State Destabilization.

To globally analyze MD simulations of wild-type FGFR2K and gatekeeper mutants, we derived a computational approach for classifying autoinhibited and active conformations through the analysis of 52 crystal structures from all four FGFR kinase isoforms. A matrix of the minimum residue distance for 192 conserved residues among the isoforms was calculated and subjected to principal component analysis (PCA) (46). Two distinct clusters emerged following transformation of the distance matrix onto the first and second principal components (PCs) (Fig. 6A). Structures within these two clusters correlated with autoinhibited and active states of FGFR kinases, as concluded from phosphorylation assays (29, 47) (SI Appendix, Table S1). For instance, unphosphorylated FGFR kinases (e.g., 3KY2 and 1FGK) defined one cluster, while A-loop phosphorylated FGFR kinases (e.g., 2PVF) and gain-of-function mutants in the A-loop and/or hinge region (e.g., 4J97 and 5UI0) comprised a second cluster. Two out of the 52 PDBs analyzed did not fall in either cluster: 6PNX and 4TYG. The former structure is an asymmetric dimer complex of the A-loop transphosphorylation complex where the monomers correspond to the substrate-acting (6PNX^A) and enzyme-acting (6PNX^B) kinases (32). The enzyme-acting kinase was in close vicinity to the active state cluster, in agreement with its role as the enzyme in A-loop phosphorylation, while the substrate-acting kinase significantly differed from both clusters. The 4TYG structure of FGFR4 kinase also significantly differed from both clusters, which likely stemmed from the inserted position of the A-loop into its own ATP-binding pocket (48). Such an insertion has not been observed in other FGFR kinase structures. Based on this analysis, we conclude that 4TYG and the substrate-acting kinase of 6PNX (6PNX^A) represent unique conformations that do not resemble either the autoinhibited or active conformations.

Fig. 6.

Classification of PDB structures and contacts distinguishing active and autoinhibited conformations. (A) Left: Approach to classify FGFR kinase structures using a PCA analysis of the minimum residue distance matrix for conserved residues among the four FGFR isoforms. Right: Crystal structures are plotted on the first two PCs showing two distinct clusters highlighted as autoinhibited and active states. Representative structures are highlighted from each cluster where 3KY2 and 1FGK represent unphosphorylated wild-type FGFR kinases and 3GQI and 2PVF represent phosphorylated wild-type FGFR kinases. 6PNX^A and 6PNX^B correspond to the substrate-acting and enzyme-acting kinases, respectively, in the A-loop transphosphorylation asymmetric dimer structure of FGFR3 kinase (32). A complete list of PDBs and where they cluster is displayed in SI Appendix, Table S1. (B) Left: Scaled t-value between active and inactive clusters of each distance–distance pair is plotted against the minimum residue distance. Red lines represent the cutoff region used to extract significant autoinhibited and active contacts. Right: Autoinhibited and active contacts are plotted on their representative structure and contacts are highlighted in yellow dotted lines. Key regions in the kinase are colored as follows: αC helix (blue), A-loop (red), and catalytic loop (green).

Using the classification of autoinhibited and active clusters, we elucidated a set of distances that differed among kinase conformations and offered an unbiased and quantitative view of residue pairs experiencing distance changes upon activation. Namely, we found 20 active contacts formed in the active state but disrupted in the autoinhibited state and 22 autoinhibited contacts formed in the autoinhibited state but disrupted in the active state (Fig. 6B and SI Appendix, Table S2; see Methods for details). MD simulations were analyzed by constructing a heat map for each residue contact, where red corresponds to an intact contact (<4.5 Å) and blue corresponds to a disrupted contact (>4.5 Å) (Fig. 7A and SI Appendix, Fig. S10). The heat map representation indicates that our simulations do not capture the complete interconversion between autoinhibited and active conformations defined by the cluster analysis. Despite this time sampling limitation, several residue contacts displayed significant differences between wild-type and gatekeeper mutants. Simulations of the gatekeeper mutants starting from the autoinhibited conformation displayed a greater probability of disrupting autoinhibited contacts and forming active contacts (Fig. 7 B–D). These contacts involve key residues in the β3 strand (K517), αC helix (E530), DFG motif (D644, F645), A-loop (L647, A648), and catalytic loop (I623, H624, R625). For example, in the gatekeeper mutant simulations, the DFG+1 and DFG+2 residues (L647, A648) had a higher probability of forming active contacts with residues in the catalytic loop (I623, H624, R625) (Fig. 7 B and D). Likewise, autoinhibited distances were disrupted for gatekeeper mutants, which remained intact for the wild-type kinase, including the DFG+1 residue (L647) with the aspartate of the DFG motif (D644) and the conserved lysine involved in ATP coordination (K517) (Fig. 7C). The DFG+2 residue (A648) also displayed a reduced probability to form an autoinhibited contact with the αC helix (E530).

Fig. 7.

Gatekeeper mutants destabilize the autoinhibited state around the DFG motif and DFG+1 and DFG+2 residues. (A) Heatmap of active or autoinhibited contacts that are formed (red) and broken (blue) during MD simulations. The two heatmap columns refer to simulations starting from the autoinhibited state (Left) or active state (Right). Residue pairs displayed for autoinhibited and active contacts represent the three greatest differences between wild-type and gatekeeper mutants from simulations initiated from the autoinhibited state structure (PDB ID: 3KY2). (B) Representative snapshots of MD simulations of wild-type FGFR2K (gray), V564I (light blue), and V564E (pink) that display contacts formed and disrupted starting from the autoinhibited state. Select residues are displayed that correspond to those identified in panel (A) and those of the DFG motif and gatekeeper position. Black dotted lines represent distances 4.5 Å or less for autoinhibited contacts for the wild-type FGFR2 snapshot and active contacts for the gatekeeper mutation snapshots. (C and D). Density plots corresponding to the minimum residue distance of the indicated autoinhibited (C) or active contact (D) for MD stimulations starting from the autoinhibited conformation. Distances were calculated for each 5 psec of the MD stimulation. Colors of the density plots are displayed within the Right panel of (C).

In contrast to simulations starting from the autoinhibited conformation, we observed more similarities among wild-type and gatekeeper mutations for simulations starting from the active conformation (Fig. 7A and SI Appendix, Fig. S10). The heat map showed some disrupted active contacts with the greatest differences at residue pairs near the phosphorylated tyrosines within the C-lobe (R664 and R678) that forms interactions with the A-loop (L647, K659, and N662). It is likely the active state differences arise from the reduced stability of the A-loop conformation for unphosphorylated tyrosines in the A-loop relative to the phosphorylated forms. Nevertheless, these differences were relatively minor compared with those observed in simulations starting from the autoinhibited conformation. Overall, the analysis of MD simulations implies that gatekeeper mutations destabilize the autoinhibited conformation through structural changes primarily impacting the DFG, DFG+1, and DFG+2 residues without directly interacting with the phenylalanine residue from the DFG motif (SI Appendix, Fig. S11). These results are consistent with our hydrophobic spine analysis indicating a destabilization of the autoinhibited conformation. We propose that gatekeeper mutations disrupt interactions within the autoinhibited conformation, which induces a population shift to the active state such that the kinase occupies a catalytically competent conformation a greater fraction of time and leads to gain-of-function.

Discussion

Stabilization of the hydrophobic R-spine in the active conformation has been proposed to underlie the gain-of-function of gatekeeper mutations in the Abl- and Src-family tyrosine kinases (25, 26). This mechanism rests on the idea that an intact hydrophobic R-spine is a unique feature of the active state of tyrosine kinases. Contrary to this model, we discovered that hydrophobic and hydrophilic gatekeeper mutations of FGFR kinase manifest their gain-of-function by destabilizing the autoinhibited state. Unlike gatekeeper mutants, unphosphorylated FGFR2K displayed an intact hydrophobic spine in both the autoinhibited and active states. These findings lead us to propose that the observed gain-of-function for gatekeeper mutations occurs from a population shift away from the autoinhibited state toward the active state.

We hypothesize the mechanistic difference between Abl- and Src-family kinases and FGFR-like kinases stems from the ability of the former kinases to sample the DFG^out conformation (26, 28), which is an inactive state typically observed with type II inhibitors and one that possesses a broken hydrophobic R-spine. The DFG^out conformation has not been observed for FGFR kinase and other RTKs in the absence of drugs (28). Instead, these kinases possess two distinct DFGⁱⁿ conformations, where one conformer corresponds to an active state (DFG^{in, active}) and the other to an autoinhibited state (DFG^{in, inhibited}) (28, 29). Notably, we found that the chi1 angle of the DFG motif phenylalanine residue strongly correlated with the active and autoinhibited states identified from our cluster analysis (Fig. 8A) (29, 37). Simulations of the autoinhibited conformation displayed a greater probability for gatekeeper mutations to sample the DFG^{in, active} conformation relative to wild-type FGFR kinase (Fig. 8B). Hence, we propose that the mechanism of gatekeeper mutations is to create steric repulsion forces with the hydrophobic R-spine at M538 in the autoinhibited state that influence the conformation of the DFG motif phenylalanine. Based on this hypothesis, any mutation at the gatekeeper position, hydrophobic or hydrophilic, that destabilizes the hydrophobic R-spine may induce activation through a similar mechanism as the isoleucine and glutamate mutations studied in this work. Notably, mutation of the gatekeeper position to phenylalanine, somewhat common in FGFR kinases (13, 49) and RTKs in general (Fig. 1), is likely to increase steric interference with M538 and lead to activation. These steric interactions likely weaken the hydrophobic R-spine in the autoinhibited conformation by introducing conformational heterogeneity and enable the DFG phenylalanine to sample the active state conformation more easily (i.e., DFG^{in, active}) (Fig. 8C). The DFG^{in, active} conformation is more compatible with rotation of the N-lobe, a key step for transitioning to the active state (35).

Fig. 8.

Mechanism of gatekeeper mutation activation occurring through the hydrophobic R-spine and DFG motif. (A) F645 chi1 dihedral angle calculated for the autoinhibited (DFG^{in, inactive}) and active (DFG^{in, active}) clusters displayed in Fig. 6A. Two PDBs (4UXQ, 4QRC) within the autoinhibited cluster were excluded from the plot, since they were bound by type II inhibitors and in the DFG^out conformation. (B) The percentage of F645 in the autoinhibited conformation (i.e., chi1 angle > 0°) from MD stimulations starting from the autoinhibited state for unphosphorylated wild-type FGFR2K (WT; gray), V564I (blue), V564E (pink), and phosphorylated wild-type FGFR2K (pWT; green). The dihedral angle was calculated every 5 psec of the MD simulation. Each bar in the graph corresponds to one replicate of 1-μs sampling time. (C) Model depicting mutations at the gatekeeper residue (in cyan) disrupting the hydrophobic R-spine (in yellow) of the autoinhibited conformation and influencing the conformation of the DFG phenylalanine (F645). The αC-helix is depicted in gray, the substrate binding site is depicted as a purple oval, and the N- and C-lobes of the kinase are shown as gray circles. The A-loop text in the substrate-binding site forms an autoinhibited interaction, while the A-loop text outside depicts the active conformation of the loop.

We also found that mutations at the gatekeeper position influence the conformation of the DFG+1 and DFG+2 residues within the autoinhibited state. Namely, contacts between DFG+1 and DFG+2 residues and the αC-helix were seen in simulations for gatekeeper mutations but not those with wild-type. We envision these interactions encourage N-lobe rotation downward and the transition to the active state. Such interactions between the N-terminal portion of the A-loop and the αC-helix are reminiscent of the αC-tether we previously reported as a determinant of activated kinases (29). Our findings also concur with the importance of the DFG+1 residue in substrate specificity and allostery by modulating selectivity between serine- and threonine-containing substrates in Ser/Thr kinases (50). Furthermore, this residue is responsible for the most common oncogenic mutation (L858R) in EGFR that induces a ~50-fold increase in catalytic efficiency and is predicted to destabilize the inactive state by disrupting autoinhibitory interactions (51) and encourage the formation of kinase dimers (52, 53).

We surmise that the conformational changes induced by gatekeeper mutations at the R-spine and DFG motif occur on a relatively fast timescale (ns–μs), while the global transition from the autoinhibited state to the active state occurs on a slower timescale (μs–ms) (Fig. 8C). Destabilization of the autoinhibited state and subsequent conformational changes at the DFG motif phenylalanine likely serves as the underlying source of fast motions that promotes N-lobe rotation and A-loop movement to the active conformation. This hypothesis is corroborated by our MD simulations capturing local conformational changes induced by gatekeeper mutations on the ns–μs timescale around the DFG motif and NMR experiments sensing dynamics at the DFG motif and A-loop on the μs–ms timescale. Based on the similarity of FGFR kinase with other RTKs (SI Appendix, Fig. S12), such as PDGF, VEGF, KIT, and c-Met receptors, the mechanism discovered for FGFR kinase may be applicable to gatekeeper mutations in other tyrosine kinases to subvert autoinhibitory mechanisms, leading to enhanced phosphorylation activity and increased clinical severity for patients possessing such mutations.

Methods

Protein Expression and Purification.

FGFR2K domain constructs, corresponding to P458 to E768 and with the C491A mutation, were expressed with an N-terminal His₆-tag to aid in protein purification. The C491A mutation of FGFR2K is referred to as wild type throughout the text. The “0Y” construct of FGFR2K was prepared by mutating all five phosphorylatable tyrosines in the protein (i.e., Y466F/Y586L/Y588P/Y656F/Y657F). Note that Y586L and Y588P mutations mimic native positions in the FGFR4 isoform. Kinases were expressed in Escherichia coli BL21(DE3) cells and induced with 0.1 mM IPTG overnight at 20 °C. Protein was purified using Ni-NTA resin and size exclusion chromatography. The N-terminal His₆-tag was cleaved using TEV protease and phosphorylated states were removed using FastAP alkaline phosphatase (Thermo Scientific). Anion exchange chromatography was performed to ensure purification of the unphosphorylated state. For NMR experiments, protein was expressed in minimal media with ¹⁵NH₄Cl, deuterated glucose, 99% D₂O, and ILV precursors, as previously described (29).

Activity Assay of Gatekeeper Mutants.

The activity assay reaction mixture consisted of 50 μM FGFR2K, 20 mM ATP, 40 mM MgCl₂, 20 mM Tris pH 7.5, and 150 mM NaCl. The addition of ATP initiated transphosphorylation on five tyrosine residues within FGFR2K, including the tyrosine residues within the A-loop. Reactions were quenched at 0, 0.5, 1, 3, 5, 10, and 20 min by adding EDTA to a final concentration of 50 mM. The transphosphorylation reaction was observed with native gel electrophoresis using a 7.5% acrylamide gel with a tris-glycine running buffer (25 mM Tris, 200 mM glycine, pH 8.3). The same reactions were analyzed using IB analysis by running samples on an SDS–PAGE gel followed by transfer to a nitrocellulose membrane. The membrane was blocked with 5% nonfat dry milk in PBS buffer at pH 7.4 for 1 h at 37 °C and subsequently for 30 min at room temperature. Phosphorylation was then detected using an antiphosphorylation-FGFR antibody for phosphorylation of A-loop tyrosines (Cell Signaling Technologies, part number 3471) by incubating the antibody overnight at 4 °C. Following incubation, the membrane was washed and incubated with antirabbit secondary antibody for 2 h at 37 °C. Membranes were imaged for chemiluminescence by addition of horseradish peroxidase (HRP) enzyme.

ATPase Activity Assay of Gatekeeper Mutants.

ATP hydrolysis rates were derived from ³¹P one-dimensional NMR spectra. Each experiment was performed using a 15-μs excitation pulse, 300 ms for acquisition (spectral width of 32467.533 Hz), and a 2.5-s recycle delay. The total length of a single experiment was 5 min (160 scans). The kinetics of ATP hydrolysis was captured over a total time of 11.75 h by collecting a series of experiments one after another. Each sample of WT^0Y, V564I^0Y, and V564E^0Y consisted of 600 μM kinase in the presence of 10 mM ATP and 20 mM MgCl₂. Quantification of the time series was performed by integrating the γ-phosphate peak of ATP at each time point.

CPMG Relaxation Dispersion Experiments.

Methyl ¹H/¹³C CPMG experiments (54) were performed using Bruker AVANCE III and AVANCE NEO NMR spectrometers operating at ¹H frequencies of 600 MHz and 800 MHz, respectively, each equipped with a TCI cryogenic probe. Data were acquired using a constant time delay of 40 ms with frequencies for the 180° pulses of 50 and 1,000 Hz. ΔR₂ values were calculated using Eq. 1, where I_1,000 is the peak intensity at 1,000 Hz, I₅₀ is the peak intensity at 50 Hz, and T is the constant time delay.

$$\mathrm{\Delta }{R}_2=\ln \left(\frac{I_{1,000}}{I_{50}}\right)\frac{1}{T}$$ [1]

ΔR₂ values and errors reported reflect the average and SD between two replicate experiments. CPMG experiments on the apo form of FGFR2K were performed at 10 °C with a protein concentration of 500 μM in 25 mM HEPES pH 7.5, 150 mM NaCl, and 10% D₂O. For ATP-bound experiments, 20 mM ATP and 40 mM MgCl₂ were used to saturate the kinase. In these experiments, ATP hydrolysis was observed and therefore a relatively low concentration of kinase was used (50 μM) to slow the reaction while maintaining sufficient signal-to-noise for data interpretation. Experiments with ATP were performed at 25 °C with 16 scans and 60 complex points in indirect dimension. We quantified the amount of ATP hydrolysis that occurred before and after the experiment using ³¹P NMR. The highest amount of hydrolysis occurred with gatekeeper mutants, corresponding to 68% of the starting ATP concentration for V564E^0Y, 64 to 71% for V564I^0Y, and 84 to 89% for WT^0Y.

CPMG Data Interpretation.

CPMG relaxation dispersion data are described by the following equation for two-site chemical exchange (states “A” and “B”) (55, 56):

$$\begin{array}{c}{R}_{2eff}=\frac{R_2^A + R_2^B + k_{ex}}{2} - \frac{1}{2{\tau }_{cp}}{\cosh }^{-1}[D_{+}\cosh {\eta }_{+}-D_{-}\cosh {\eta }_{-}]\\ D_{\pm }=\frac{1}{2}\left[\pm 1+\frac{\psi +2\text{Δ}{\omega }^2}{\sqrt{{\psi }^2+{\zeta }^2}}\right] {\eta }_{\pm }=\frac{{\tau }_{cp}}{\sqrt{2}}{\left[\pm \psi +\sqrt{{\psi }^2+{\zeta }^2}\right]}^{1/2}\\ \psi =k_{ex}^2-\text{Δ}{\omega }^2 \zeta =-2\text{Δ}\omega k_{ex}\left(p_A-p_B\right)\end{array}$$ [2]

$p_A$ and $p_B$ are the populations of A and B, $k_{\mathit{ex}}$ is the exchange rate, which is equal to the forward ($k_1$) and reverse rate ($k_{-1}$) constants describing the equilibrium between states A and B, $\mathrm{\Delta }\omega$ is the difference in chemical shifts between the two states, $R_2^A$ and $R_2^B$ are the transverse relaxation rates for states A and B, and ${\tau }_{\mathit{cp}}$ is the time between successive 180° refocusing pulses. This equation assumes the difference between $R_2^A$ and $R_2^B$ is negligible relative to the difference between $k_1$ and $k_{-1}$.

Gatekeeper mutants displayed larger ΔR₂ values calculated using Eq. 1 relative to the WT^0Y sample. There are three reasons (not mutually exclusive) why these ΔR₂ differences could arise according to Eq. 2: 1) change in Δω, 2) change in k_ex, or 3) changes in p_A and p_B. For the first case, we anticipate that Δω values will be the same for wild-type and gatekeeper mutants for residues not near the site of mutation. Indeed, many of the key changes we observed are greater than 10 Å from the gatekeeper position. We consider this a somewhat safe assumption considering FGFR kinases almost exclusively crystallize in either the autoinhibited or active conformation (47). Indeed, our quantitative and unbiased clustering analysis in Fig. 6 displays two clusters representing 52 FGFR kinase crystal structures of wild-type and mutated variants. In other words, the active state (or autoinhibited state) is essentially identical whether a mutation is present or not and therefore is expected to have the same Δω for wild-type or mutants. The second possible reason for a change in ΔR₂ values is a change in k_ex. However, since k_ex is a sum of k₁ and k₋₁, the only way k_ex can change without changing populations is if both k₁ and k₋₁ change by exactly the same factor. This narrow condition is very unlikely and would not explain the increased activity of gatekeeper mutants relative to wild-type displayed in Fig. 2. In the more likely case that k_ex changes, which stems from a change in k₁ and/or k_-1, this would result in a change in populations. Therefore, based on these arguments and the increased activity of gatekeeper mutants, we conclude that the increased ΔR₂ values for gatekeeper mutants correspond to a change in the population relative to wild-type.

Intensity Retention NMR Experiments.

Intensity retention experiments were performed on ILV-labeled kinases using a Bruker AVANCE NEO NMR spectrometer operating at a ¹H frequency of 800 MHz equipped with a TCI cryogenic probe. ¹H/¹³C HMQC spectra were collected at 25 °C on 50 μM and 250 μM samples of WT^0Y, V564I^0Y, and V564E^0Y in 25 mM HEPES pH 7.5, 150 mM NaCl, and 40 mM MgCl₂ in the presence and absence of 20 mM ATP. Intensity retention values were calculated as the peak intensity in the ATP/Mg²⁺-bound state divided by the intensity in the absence of ATP. ¹H/¹³C HMQC spectra for 50 μM samples in the absence of ATP were collected with eight scans, while spectra in the presence of 20 mM ATP were collected with 12 scans to account for the dilution upon nucleotide addition (50 μL nucleotide into 200 μL kinase sample). ¹H/¹³C HMQC spectra for 250 μM samples in the presence and absence of ATP were each collected with four scans. To compensate for the dilution factor and/or scan difference in HMQC spectra, peak heights of kinase samples at 250 μM in the presence of ATP were multiplied by 1.25, while peak heights of kinase samples at 50 μM in the absence of ATP were multiplied by 1.2. The peak height multiplications were done in an identical manner for WT^0Y, V564I^0Y, and V564E^0Y.

PRE Experiment.

A ¹⁵N tyrosine selectively labeled sample was prepared using the I707C mutant of FGFR2K, as described previously (29). Prior to labeling, the kinase was treated with 20 mM DTT for 15 min in 20 mM Tris pH 7.5 and 150 mM NaCl. DTT was buffer exchanged, and the pH was adjusted to 6.5. MTSL ((1-oxyl-2,2,5,5-tetramethyl-3-pyrroline-3-methyl)-methanethiosulfonate) was added in half-molar equivalents relative to the kinase concentration (32 μM) and incubated on ice for 10 min at each of two total additions. The progress of the reaction was monitored using a Bruker ultrafleXtreme^TM MALDI-TOF. The sample was buffer exchanged to 25 mM HEPES pH 7.5 and 150 mM NaCl for NMR experiments. The kinase sample concentration for NMR experiments was 106 μM and included the addition of 10 mM AMP-PCP, 20 mM MgCl₂, and 10% D₂O.

¹H/¹⁵N HSQC experiments were collected for MTSL-labeled I707C FGFR2K in the oxidized and reduced forms using a Bruker AVANCE NEO NMR spectrometer operating at a ¹H frequency of 800 MHz equipped with a TCI cryogenic probe. The oxidized sample was obtained by treating the reduced sample with 10 molar equivalents of ascorbic acid relative to the kinase concentration. ¹H/¹⁵N HSQC experiments were performed with 704 scans, spectral widths of 12,500 Hz (¹H) and 1,469.734 Hz (¹⁵N), acquisition time of 81.92 ms, 24 complex points in the indirect dimension, and a recycle delay of 4 s. ¹H/¹⁵N HSQC intensity retentions were calculated by dividing peak heights of the oxidized HSQC spectrum over that of the reduced spectrum. Distances were calculated as previously described using a correlation time of 22.2 ns (57, 58).

NMR Nucleotide Titrations.

ATP titrations were performed at 25 °C on WT^0Y, V564I^0Y, and V564E^0Y using a Bruker AVANCE NEO NMR spectrometer operating at a ¹H frequency of 800 MHz equipped with a TCI cryogenic probe. ¹H/¹³C HMQC experiments were acquired on ILV-labeled kinases with 4 scans and 60 complex points in the indirect dimension to minimize the experiment time due to ATP hydrolysis. The entire titration was acquired in ~3 h. Titrations were carried out with 50 μM kinase in 25 mM HEPES pH 7.5, 150 mM NaCl, and 40 mM MgCl₂. ATP was titrated at several concentrations ranging from low μM to mM, as displayed in the figure legends. The AMP-PCP titration was performed at 10 °C on FGFR2K using a Bruker AVANCE III NMR spectrometer operating at a ¹H frequency of 600 MHz equipped with a TCI cryogenic probe. ¹H/¹³C HMQC experiments were performed on 533 μM ILV-labeled kinases with two scans and 100 complex points in the indirect dimension. Binding constants were fit using a nonlinear least-squares regression as shown in Eq. 3.

$$\begin{array}{c}\text{Δ}{\delta }_{obs} = \text{Δ}{\delta }_{\max }\frac{\left({[P]}_t+{[L]}_t+K_d\right) - \sqrt{{\left({[P]}_t+{[L]}_t+K_d\right)}^2 - 4{\left[P\right]}_t{\left[L\right]}_t}}{2{\left[P\right]}_t}\end{array}$$ [3]

FGFR Kinase Structural Analysis.

Fifty-two crystal structures that correspond to each of the four FGFR isoforms were analyzed: 32 FGFR1, 9 FGFR2, 3 FGFR3, and 8 FGFR4. PDBs with unresolved A-loops were removed from the analysis, as the A-loop was found to be an important part of the classification. Of the 208 conserved residues among all FGFR isoforms of the kinase domain, 192 of these residues were used in the analysis since some regions were consistently not resolved in crystal structures (e.g., P-loop and kinase insert). Using the 192 residues, a distance matrix of all PDBs was calculated using the minimum residue distance (MRD). PCA was then used on MRD of the 52 high-resolution crystal structures and plotted according to the PC1 and PC2. The first two PCs accounted for 66% and 7% of the total variance. From there, PDBs were clustered by the PC1 and PC2 using DBSCAN clustering algorithm (59) using Scikit-learn (60).

To determine the distances that account for the variance between clusters, a t-score was calculated as shown in Eq. 4, where $d_{\mathit{ij}}^i$ and $s_{\mathit{ij}}^i$ represent the average distance pair and the standard deviation of the distance within the cluster, respectively.

$$\mathit{t-score} = \frac{d_{\mathit{ij}}^1 - d_{\mathit{ij}}^2}{\sqrt{s_{\mathit{ij}}^1 + s_{\mathit{ij}}^2}}$$ [4]

The t-score was then scaled (scaled t-score) by the minimum of the distance pair ($d_{\mathit{ij}}$) to enhance the t-score of smaller distances.

$$\mathit{scaled} \mathit{t-score} = \frac{\mathit{t-score}}{\mathit{min}\left(d_{\mathit{ij}}\right)}$$ [5]

All distance pairs with a scaled t-score greater than 3.5 Å in magnitude and a minimum distance less than 4.5 Å were extracted. This yielded 43 distances formed in the active state, and 45 distances formed in the autoinhibited state. Next, distance pairs were filtered to represent unique contacts formed in the active and autoinhibited states. The first filtering was performed to remove long-range distances by removing distances that have an average PDB distance greater than 5 Å. The next filtering removed distances shared in active and autoinhibited states by removing average distances between the clusters that varied by 1 Å or less. Last, distances were filtered based on the stability of the contact in MD simulation. The stability of active and autoinhibited contacts was determined by the MD simulations starting from 2PVF and the FGFR2K homology model of 3KY2, respectively. Distances were removed if the percent contact formed was less than 25%, where a contact is defined as formed if the minimum residue heavy atom distance is less than 4.5 Å. Active distances removed include L647-L665, L647-P666, R625-L665, and K658-D677. Autoinhibited distances removed include T660-L665, R664-S702, R630-T660, R664-E695, R573-N662, and R573-R664. This yielded 20 active contacts formed in the active state but disrupted in the autoinhibited state, and 22 autoinhibited contacts formed in the autoinhibited state but disrupted in the active state (SI Appendix, Table S2). These contacts represented ~0.5% of the total number of conserved residue pairs

MD Simulations.

Molecular dynamic simulations were performed starting from autoinhibited and active states. The active state was represented by an FGFR2K structure with A-loop tyrosines phosphorylated with ATP analog and substrate peptide [PDB ID: 2PVF (45)]. The inactive state was modeled from a homology model of FGFR1 kinase [PDB ID: 3KY2 (61)] using Rosetta comparative homology modeling (62, 63), since there are no FGFR2K crystal structures in the authentically autoinhibited conformation. Simulations were performed on wild-type, V564I, V564E, and phosphorylated state (3P) with ATP and Mg²⁺ bound and by removing the substrate peptide from PDB ID 2PVF. The phosphorylated state was simulated with pY656, pY657, and pY586 to mimic the phosphorylation state in the crystal structure (PDB ID: 2PVF). The protein was solvated with TIP3P water with a 15 Å buffer region and neutralized with Na⁺ (e.g., eight ions for wild-type, nine ions for V564E, and 14 ions for 3P). For ATP-bound simulations, parameters for ATP and Mg²⁺ were implemented from a previously parameterized system and coordinates were modeled from PDB ID 2PVF (45, 64, 65). Simulations were built in either Amber 16.06 or Amber 20.11 and ran in either Gromacs 5.1.4 and Gromacs 2020.4 for 2PVF and 3KY2, respectively. Parameter files were converted from Amber to Gromacs using ACPYPE (66). All simulations were performed using the amber14sb forcefield (67). Hydrogen bonds to heavy atoms were constrained using the LINCs algorithm with a 2-fs timestep (68). The van der Waals interactions were switched off from 1.0 to 1.2 nm. Electrostatic interactions were treated with particle-mesh Ewald with a cutoff of 1.2 nm (69). A steepest descent minimization was completed until the maximum force was below 1,000 kJ mol⁻¹ nm⁻² with heavy atoms constrained. The system was equilibrated for 1 ns under constant volume and temperature with the heavy atoms constrained. Following this equilibration, a constant pressure and temperature simulation was performed, where the maximum force on heavy atoms of the protein was reduced from 1,000 kJ mol⁻¹ nm⁻² to 0 kJ mol⁻¹ nm⁻² over 3 ns. From there, 30 ns unconstrained MD simulations were performed prior to the production runs. For simulations starting from the autoinhibited state, heavy atoms of ATP were released prior to the release of force constants of protein because ATP was modeled into the structure. Temperature was maintained at 300 K using the Nose–Hoover thermostat for protein and solvent separately (70). Pressure was maintained at 1 atm using the Parrinello–Rahmen coupling (71). Simulations were repeated three or five times for 1 μs each saving coordinates every 5 ps for the analyses. Mutant structures were obtained using the Dunbrack rotamer library implemented in Chimera (72). Loops missing in PDB ID 2PVF were modeled in using Modeler (73).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

MD simulation data are publicly available in the link: https://yzhang.hpc.nyu.edu/MD_data_FGFR. All study data are included in the article and/or SI Appendix.