“DFG-Flip” in the Insulin Receptor Kinase Is Facilitated by a Helical Intermediate State of the Activation Loop

Abstract

We have characterized a large-scale inactive-to-active conformational change in the activation-loop of the insulin receptor kinase domain at the atomistic level via untargeted temperature-accelerated molecular dynamics (TAMD) and free-energy calculations using the string method. TAMD simulations consistently show folding of the A-loop into a helical conformation followed by unfolding to an active conformation, causing the highly conserved DFG-motif (Asp¹¹⁵⁰, Phe¹¹⁵¹, and Gly¹¹⁵²) to switch from the inactive “D-out/F-in” to the nucleotide-binding-competent “D-in/F-out” conformation. The minimum free-energy path computed from the string method preserves these helical intermediates along the inactive-to-active path, and the thermodynamic free-energy differences are consistent with previous work on various other kinases. The mechanisms revealed by TAMD also suggest that the regulatory spine can be dynamically assembled/disassembled either by DFG-flip or by movement of the αC-helix. Together, these findings both broaden our understanding of kinase activation and point to intermediates as specific therapeutic targets.

Introduction

Receptor tyrosine kinases are tightly regulated ligand-activated transmembrane glycoproteins that catalyze the phosphorylation of specific tyrosines on protein substrates. Activation of the insulin receptor (IR), a constitutively dimeric receptor tyrosine kinase, consists of trans-autophosphorylation of three activation loop (A-loop) tyrosines (1158, 1162, and 1163) located in each cytoplasmic kinase domain of IR (IRKD) (1,2). The inactive (3) and active (4) crystal structures of bilobal IRKD reveal that 1), the A-loop is displaced by ∼20 Å, and 2), the αC-helix in the N-lobe rotates toward the C-lobe by ∼30°. The highly conserved DFG motif at the N-terminus of the A-loop (D¹¹⁵⁰, F¹¹⁵¹, and G¹¹⁵²) flips from a “D-out/F-in” to “D-in/F-out” conformation during activation (Fig. 1 a). Flipping back and forth of the DFG-motif has been suggested to facilitate nucleotide binding and release (5). For a productive inactive-to-active transition, flipping of the phenylalanine of the DFG motif out of the nucleotide-binding pocket places its side chain into a network of side chains known as the regulatory (R) spine, which forms a continuous hydrophobic core linking the N-lobe with the base of the C-lobe. A similar chain of contacts forms upon binding of the nucleotide and is known as the catalytic (C) spine; formation of both the R- and C-spines is the hallmark of activation of all known tyrosine kinases (6–8). Evidently, conformational changes in the A-loop and αC-helix can dynamically assemble and disassemble the R- and C-spines as part of the activation/deactivation cycle (8). Such changes are hinted at by comparison of inactive and active crystal structures of IRKD. However, the exact mechanism of conformational change in the A-loop of IRKD involving the DFG-flip and R/C-spine formation remains elusive, chiefly due to the lack of structural data on intermediate IRKD conformations (9). Approaches using physical all-atom modeling can therefore in principle provide some needed insight.

Figure 1.

(Color online) (a) Key structural features are highlighted in the inactive (1IRK.pdb) and active (1IR3.pdb) crystal structures of the insulin receptor kinase domains. (Transparent cartoons) Protein backbone; (red) A-loop and αC-helix; (cyan and blue) side chains of Asp¹¹⁵⁰ and Phe¹¹⁵¹ of the DFG-motif, respectively. The side chains of residues in R-spine (orange sticks) and C-spine (green sticks), along with their transparent surface-renderings, respectively. (b) Schematic representation of the simulation domain (73 Å × 76 Å × 76 Å) as viewed along the z axis: (blue cartoon) N-lobe and (cyan cartoon) C-lobe; (red) crystallographic water molecules; (yellow spheres) ions; and (wireframe) solvent. The system contains 40,044 atoms in all.

Observing large-scale conformational changes in biomolecules using unbiased molecular dynamics (MD) simulations is particularly challenging because of difficulties in surmounting associated free energy barriers in reasonable computational time. Hence, a combination of targeted molecular dynamics (10), and/or the string method with swarms-of-trajectories (11), or metadynamics (12) have proven useful in studying similar transitions in other kinases (13–15). Long MD simulations of a mutant Abl kinase (a nonreceptor TK) captured the DFG-flip and suggest an important role for the protonation state of the DFG-aspartate (16). Abrams and Vanden-Eijnden have recently demonstrated (17) the usefulness of a new technique, temperature-accelerated molecular dynamics (TAMD) (18,19), that requires no-target bias to study large-scale conformational changes in proteins. In this contribution, we have applied TAMD to the A-loop (1150–1168) of inactive IRKD. We observe DFG-flip in four independent trajectories, and we delineate mechanistic details governing this conformational change. Using the string method (20), we further refine the TAMD-generated trajectories to a minimum free-energy path (MFEP) for activation, which remarkably preserves the existence of helical conformations of the A-loop. We briefly discuss the significance of the mechanism behind DFG-flip in designing novel therapeutics targeting kinases.

Methods

Molecular dynamics simulations: system setup

We generated all MD trajectories using NAMDv2.8 (21) and the CHARMM force field (22) with CMAP correction (23). VMDv1.9 was used for system creation and protein rendering (24). The initial coordinates for the inactive state of IRKD were taken from the crystal structure of Hubbard et al. (3) (PDB:1IRK). The crystallographic water molecules were retained but the ethylmercuric phosphate molecule was deleted. The crystal structure provided coordinates for the residues 981–1283, and the missing residues of the N-terminus at positions 978–980 were modeled. We solvated this structure using explicit (TIP3P) water and included all hydrogen atoms. Based on simple pK_a estimates, we elected not to protonate Asp¹¹⁵⁰ (see Discussion). Charge neutrality was maintained by adding counterions at a salt concentration of 150 mM. The system contains a total of 40,044 atoms (Fig. 1 b). Similar protocols were followed for all simulations: 500 cycles of energy-minimization via conjugate-gradient optimization and a constant temperature (300 K) via the Langevin thermostat with damping coefficient of 5 ps⁻¹.

The equilibration phases were carried out initially for ∼2–3 ns in the NPT ensemble to adjust the box volume, and thereafter continued for ∼5 ns in the NVT ensemble using 1-fs time step with all bonds flexible, including those to hydrogens (i.e., no holonomic restraints), in all simulations. Periodic boundary conditions were used throughout. Nonbonded interactions were cut off beyond 10 Å with smooth switching taking effect at 8.0 Å. Long-range electrostatic interactions were handled using the particle-mesh Ewald method. We carried out five ∼6–7-ns-long independent MD equilibrations of the inactive IRKD, which were used to sample initial conditions for eight TAMD runs (see below).

Temperature-accelerated molecular dynamics

The theoretical basis of the temperature-accelerated molecular dynamics was originally presented by Maragliano and Vanden-Eijnden (18,19), and has been reviewed in detail (25). More recently, the TAMD was further developed and applied by Abrams and Vanden-Eijnden (17) as a unique method to study the large-scale conformational changes in proteins. For this reason, we will simply describe the underlying equations here. The coupled system of equations describing TAMD are

$$\begin{array}{l}{m}_i{\ddot{x}}_i=-\frac{\partial V\left(x\right)}{\partial x_i}-\kappa \sum _{j=1}^m\left[{\theta }_j^{\ast }\left(x\right)-{\theta }_j\right]\frac{\partial {\theta }_j^{\ast }\left(x\right)}{\partial x_i}-\gamma m_i{\dot{x}}_i+{\eta }_i\left(t;\beta \right),\hfill \\ \overline{\gamma }{\overline{m}}_j{\dot{\theta }}_j=\kappa \left[{\theta }_j^{\ast }\left(x\right)-{\theta }_j\right]+{\xi }_j\left(t;\overline{\beta }\right),\hfill \end{array}$$ (1)

where θ^∗ (x) = (θ^∗₁ (x), θ^∗₂ (x),…….., θ^∗_m (x)) are collective variables that are functions of the atomic Cartesian coordinates; m_i values are the masses of x_i; V(x) is the interatomic MD potential; κ is the coupling spring-constant; γ is the Langevin friction coefficient; η is the white noise satisfying fluctuation-dissipation theorem at physical temperature β⁻¹; $\overline{\gamma }$ and ${\overline{m}}_j$, respectively, are fictitious friction and masses of the variables θ_j; and ξ is the thermal noise at artificial temperature ${\overline{\beta }}^{-1}$.

The aforementioned set of equations describe the motion of x(t) and θ (t) over the extended potential

$$\begin{array}{c}{U}_{\kappa }\left(x,\theta \right)=V\left(x\right)+\frac{\kappa }{2}\sum _{j=1}^m{\left[{\theta }_j^{\ast }\left(x\right)-{\theta }_j\right]}^2\end{array}.$$ (2)

As shown elsewhere (18), by choosing κ so that θ^∗ (x(t)) ≈ θ (t), and fictitious friction coefficient $\overline{\gamma }$ so that θ moves slower than x, we can generate a trajectory θ (t) that moves at artificial temperature ${\overline{\beta }}^{-1}$ on the free energy landscape computed at the physical temperature β⁻¹. In this work, we have chosen a TAMD friction $\overline{\gamma }$ of 50 ps⁻¹ and a spring constant κ of 100 kcal/mol·Å². As collective variables (CVs), we choose the Cartesian coordinates of centers of mass of spatially contiguous groups of residues. Particularly, the A-loop residues 1150–1168 were divided into four subgroups (three groups of five residues each, and one group of four residues) and hence 12 CVs. Therefore, the conformation sampling of only the A-loop was accelerated via TAMD and the remaining atoms in the system evolved under standard Langevin dynamics.

We did not apply TAMD to the whole structure because alignment of the inactive (3) to active (4) crystal structure reveals that major contribution to the backbone C_α root mean-square deviation (RMSD) comes solely from the A-loop. TAMD was hence applied to the A-loop of inactive IRKD at a fictitious thermal energy ${\overline{\beta }}^{-1}=5$ kcal/mol, where $\overline{\beta }=1/k_{\text{B}}\overline{T}$, in which k_B is Boltzmann's constant and $\overline{T}$ is the fictitious temperature. We carried out a total of eight ∼40-ns-long TAMD simulations starting from initial conditions sampled using five independent MD equilibration (see above) trajectories of the inactive IRKD crystal structure (3). Details of these runs are summarized in Table 1. One TAMD trajectory (run No. 1) was successful in generating the entire conformational change, whereas three other simulations (runs Nos. 2–4) were partially successful, and remaining four (runs Nos. 5–8) failed to generate the conformational change on 40-ns timescale (see Table 1). The results for TAMD run No. 1 are described in the main article, while the results for additional partially successful or unsuccessful runs (runs Nos. 2–8) are described in the Supporting Material.

Table 1.

Details of all TAMD simulations (∼40 ns each)

Run No.	RMSD (Å) closest/end	Flip∗ ↻/↺	Helix	R-spine
1	8.27/8.27	√ (↺)	√	√
2	18.30/26.48	√ (↺)	√	√
3	10.37/27.82	√ (↻/↺)	√	×
4	16.15/21.52	√∗(↺)	√	×
5	20.20/23.82	×	×	×
6	19.03/26.10	×	×	×
7	19.70/28.33	×	×	×
8	15.81/24.15	×	×	×

Results from run No. 1 are discussed in the main article and for runs Nos. 2–8 in the Supporting Material.

^∗Residues are only semiflipped.

Free energy calculation of the activation pathway via string method

The string method (20,26–30) is a technique to compute the minimum free energy path (MFEP) in a large but finite set of CVs. The MFEP is defined as the curve whose tangent is always parallel to the gradient of the free energy (the mean force) times a metric factor that, for the case of linear CVs used here, is constant and diagonal. The algorithm works by iteratively refining (see below) an initial string, i.e., a collection of discrete configurations of the system known as “images”. Diverse problems have been recently studied using the string method such as the hydrophobic collapse of a hydrated chain (31) and a pore opening/closing transition of an ion channel (32). The string method with swarms-of-trajectories (11), a variant of the original string method (20), was applied to the inactive-to-active transition of the catalytic domain of human c-Src kinase (14), which is a transition similar to what we describe in this work. In the following, we present the details of the string method as it has been used in this work.

Initial string

The images in the initial string in our work were chosen from the ∼40-ns-long TAMD run No. 1 (see above) of the inactive-to-active transition of IRKD. Because TAMD trajectories can spend significant time exploring the phase-space locally (as seen in failed TAMD runs), not every configuration in a TAMD trajectory contributes to the reactive trajectory. Hence, we carefully pruned the TAMD trajectory such that only those configurations that always take the system away from the initial (inactive) state, yet toward the final (active) state, are chosen. The configurations chosen this way, although unique and discrete, are not equidistant in CV-space, which is a requirement for the iterative string evolution. Hence, using the reparameterization procedure described in the original string method work (20), we created 24 (0, 1, 2, ⋯, 23) equidistant images that form our initial string.

Because the initial string is merely a rough estimate of MFEP, the iterative refinement of this initial guess toward MFEP is required. We note that although the iterative procedure of the string method can, in principle, evolve any guessed/interpolated path connecting the initial and final state to MFEP, a carefully generated initial path (as demonstrated in this work) can save significant computational cost. We suggest that TAMD is ideally suited for generating such initial paths, because the conformational transitions in TAMD are obtained from the exploration of the free-energy landscape defined at the physical temperature β⁻¹ (17,19,20).

String evolution

Application of the string method algorithm requires the calculation of the mean force at each image along the string at each iteration. The mean force can be obtained from restrained MD simulations, i.e., using the potential defined in Eq. 2 but with θ_j fixed, and using the estimator in Maragliano et al. (20). In this article, we used 100-ps-long restrained simulations with force constant k = 5 kcal/mol·Å², after having equilibrated the system at each image using a higher k. The 100-ps-long trajectories were used for the mean force estimation at each image (see below). However, the mean forces in the final converged string were estimated using longer trajectories for each image. Weak harmonic restraints were also used on the protein backbone to avoid net translation and rotation of IRKD during restrained runs, because CVs in our case are not internal coordinates. The images in the string were updated by measuring the mean restraining force (or the negative gradient of the free energy F) for each CV (θ^∗_j) using the following equation:

$${\overrightarrow{f}}_{{\theta }_j^{\ast }}^{mean}=-\frac{\partial F\left({\theta }^{\ast }\left(x\right)\right)}{\partial {\theta }_j^{\ast }\left(x\right)}\approx k\left(\langle {\theta }_j^{\ast }\left(x\right)\rangle -{\theta }_j\left(x\right)\right).$$ (3)

We also note that the string images were updated using the full force computed at each image and not its orthogonal component, because it has been shown (29) that using the full force simplifies and improves the original implementation. Hence, it was not necessary to smooth the string in this work (a smoothing step was used in the projected-force evolution of the string to improve the tangent calculation). All string calculations were carried out with explicitly solvated and ionized images in NAMD simulation package using the CHARMM force field (see above).

Potential of mean force (free-energy) calculations

The free energy profile at each string iteration can be computed by integrating the mean force along the string (20) using

$$\text{PMF}=-\underset{0}{\overset{\alpha }{\int }}\left[\sum _{j=1}^N{\overrightarrow{f}}_{{\theta }_j^{\ast }}^{mean}\times t_{{\theta }_j^{\ast }}\right]d\alpha ,$$ (4)

where N is the total number of CVs (12 here), ${\overrightarrow{f}}_{{\theta }_j^{\ast }}^{mean}$ can be computed using Eq. 3, and $t_{{\theta }_j^{\ast }}$ is the tangent to the path for that CV. The tangent calculation was performed using the forward difference formula for the first image, the backward difference formula for the last image, and the central difference formula for all other images. The integral was computed using the trapezoidal rule. The potential of mean force along the path is then a sum over the product of mean force at each image multiplied by the tangent to the path at that point. We point out that the free-energy profile computed by the string method is in the space of CVs, which is nonetheless multidimensional (12-dimensional here corresponding to 12 CVs).

String convergence

To assess the convergence to the MFEP, we monitored three properties during successive string iterations: a), the free energy profiles; b), the RMSD of the string in CV-space; and c), the average string RMSD over different iterations. We present these results in Fig. S1 in the Supporting Material, while the snapshots of converged images in MFEP are presented in Fig. S2. Twenty-five iterations of the string, in total, were carried out until convergence. The error bars were estimated by measuring the statistical fluctuations of each image in the converged string using independent simulations. The last image in the string was also allowed to evolve.

Results

TAMD-driven conformational change in the A-loop

In Fig. 2 a, we show time-evolution of the root mean-squared deviation (RMSD) of the A-loop, R-spine, C-spine, and F¹¹⁵¹ of the DFG-motif with reference to the active crystal structure (4) during two different simulation stages: the first ∼7 ns of initial MD-equilibration (shaded region) followed by ∼40 ns of TAMD simulation (run No. 1; see Table 1). The inactive conformation of the A-loop displays an RMSD relative to the active conformation of ∼20 Å. We observe no appreciable change in RMSD of the A-loop during the initial ∼7 ns MD-equilibration, which indicates that the A-loop is highly stable and remains close to its inactive conformation. During the next ∼40 ns of TAMD, we observe that the A-loop departs from its inactive conformation at ∼25 ns and evolves toward an activelike state as indicated by a significant decrease in its RMSD from ∼25 to ∼8 Å. We also observe that along with this conformational change in the A-loop, the initially broken R-spine (see Fig. 1 b for spine residues) experiences a sudden transition (at ∼25 ns) in RMSD from ∼8 to ∼3 Å, the majority of which is contributed by a decrease in RMSD of F¹¹⁵¹ from ∼9 to ∼1.5 Å. During the entire simulation, the C-spine initially displays an increase in RMSD, only to heal by ∼1 Å during last 25 ns.

Figure 2.

(Color online) (a) RMSD versus simulation time (ns) for the activation-loop (the A-loop), R-spine, C-spine, and Phe¹¹⁵¹ with respect to the active crystal structure for TAMD-run1 (see Table 1). (Gray background in the plots) The first ∼7-ns of MD equilibration, which is followed by ∼40 ns of TAMD. (b) Representative snapshots of IRKD from TAMD simulation are shown at various time-points with the A-loop (red) and side chains of Asp¹¹⁵⁰ and Phe¹¹⁵¹ (cyan and blue), respectively. (Black cartoon) Conformation of the A-loop in the active crystal structure. See Fig. 1b for the labels and coloring scheme of R- and C-spine residues. (Large panel, center) Conformations of IRKD with highlighted structural motifs: αC-helix, nucleotide-binding loop, and the activation loop from TAMD simulation at t = 7.59 (red), 17.09, 22.89, 40.09, and 47.00 ns (blue). (Arrow) Directional guide along the increasing simulation time.

At 12 different time-points in Fig. 2 b, we also show the snapshots of IRKD highlighting the conformational change in the A-loop from inactive (red cartoon) to its activelike conformation (black cartoon) along with residues of R-spine, C-spine, and the DFG-motif. We observe that at ∼16.95 ns, the A-loop (residues 1152–1165) begins to fold into a helical conformation that it maintains until ∼25 ns before unfolding into an extended activelike conformation during the rest of the simulation. A key implication of helical folding/unfolding transition in the A-loop is that it leads to change in backbone dihedral angles of the DFG-motif residues D¹¹⁵⁰ and F¹¹⁵¹, such that they flip from a D-out/F-in to D-in/F-out conformation. Specifically, the folding of the A-loop leads to an anticlockwise twist in the side chains of D¹¹⁵⁰ and F¹¹⁵¹ (snapshots at 20.31, 20.63, and 22.39 ns), whereas spontaneous unfolding of the A-loop places the side chains of D¹¹⁵⁰ and F¹¹⁵¹ in the ATP-binding pocket, and underneath αC-helix, respectively.

Although DFG-flip is complete after ∼25 ns, an intact R-spine (continuous orange surface connecting the two lobes shown in Fig. 1 b) does not form until ∼39 ns (snapshots at 26.57, 32.69, and 39.03 ns). This is because the N-lobe moves outward during the first-half of simulation (also indicated by the increase in the RMSD of C-spine, Fig. 2 a), and begins to transition toward the C-lobe after the A-loop is out of its way post-DFG-flip. Because two residues of the hydrophobic R-spine reside in the αC-helix (which is part of the N-lobe and is still oriented away from the C-lobe), the intact R-spine is only observed after the αC-helix rotates toward the C-lobe at ∼40 ns. The coupling of the movement of αC-helix, Gly-rich nucleotide-binding-loop, and the A-loop is also highlighted in the snapshots of IRKD at four different simulation frames (central panel in Fig. 2 b), where the αC-helix first moves outward along with Gly-rich loop (snapshots at 7.59, 17.09, and 22.89 ns) before rotating inward (snapshots at 40.09, and 47.00 ns) during the transition of the A-loop to the activelike conformation.

Key salt-bridging interactions and the tyrosine solvent-exposure

An important characteristic of the active IRKD is two salt bridges: a kinase-conserved K¹⁰³⁰-E¹⁰⁴⁷ salt bridge, and an the A-loop stabilizing R¹¹⁵⁵-Y¹¹⁶³ salt bridge formed after the trans-autophosphorylation of Y¹¹⁶³ (4). However, we observe that R¹¹⁵⁵ and K¹⁰³⁰ can participate in other transient ion-pair interactions that couple residues of the A-loop with the residues of αC-helix and the N-lobe during kinase activation. In Fig. 3 a, we show interatomic distance traces of the following salt-bridging interactions: R¹¹⁵⁵ with E¹⁰⁴⁷, Y¹¹⁶³, and D¹¹⁵⁰, and K¹⁰³⁰ with E¹⁰⁴⁷ and D¹¹⁵⁰. In Fig. 3 b, we also highlight the position of the A-loop, and of residues participating in the salt-bridging interactions at selected simulation times. Distance traces in Fig. 3 a show that none of the salt bridges exists in the beginning of simulation. However, during the conformational change in the A-loop, we observe an intermittent R¹¹⁵⁵-E¹⁰⁴⁷ salt bridge (snapshot at 18.37 ns in Fig. 3 b), which breaks and remains broken for the next 8 ns only to reform between ∼26 to 36 ns (Fig. 3 a and snapshot at 35.87 ns in Fig. 3 b) beyond which the salt bridge is broken again due to the movement of the A-loop.

Figure 3.

(Color online) (a) Interatomic distance-traces for various salt-bridging interactions formed by Lys¹⁰³⁰ and Arg¹¹⁵⁵. (Gray background) The first ∼7 ns of MD equilibration, which is followed by ∼40 ns of TAMD. (Horizontal lines) Distances in the active crystal structure. (b) Snapshots of IRKD highlighting the side chains of residues involved in salt-bridging interactions with the A-loop (shown in red). (c) Snapshots of IRKD highlighting the sequence of solvent-exposure for three the A-loop tyrosines (Tyr^{1158,1162,1163}). (Black cartoon) Conformation of the A-loop in the active crystal structure; (last two panels) the three tyrosine residues.

While participating in this salt bridge, R¹¹⁵⁵ additionally serves as a chaperone for the formation of the K¹⁰³⁰-E¹⁰⁴⁷ salt bridge by pulling E¹⁰⁴⁷ (and in turn the αC-helix) toward K¹⁰³⁰ (snapshot at 40.97 and 45.60 ns in Fig. 3 b). R¹¹⁵⁵ also intermittently interacts with D¹¹⁵⁰ (snapshot at 19.83 ns in Fig. 3 b). We do not observe a R¹¹⁵⁵-Y¹¹⁶³ salt bridge except a transient proximity of both residues (snapshot at 20.95 ns in Fig. 3 b) because this interaction requires phosphorylated Y¹¹⁶³, which is not the case in our simulations. We also observe that K¹⁰³⁰ can engage both E¹⁰⁴⁷ and D¹¹⁵⁰ (after DFG-flip) simultaneously in salt-bridging interactions (snapshot at 40.97 ns in Fig. 3 b). But in the final state, we only observe a stable K¹⁰³⁰-E¹⁰⁴⁷ salt bridge and not a K¹⁰³⁰-D¹¹⁵⁰ salt bridge (Fig. 3 a, and snapshot at 45.60 ns in Fig. 3 b).

The crystal structures suggest that Y¹¹⁵⁸ and Y¹¹⁶² are significantly exposed to the solvent on phosphorylation, whereas Y¹¹⁶³ is buried. The A-loop transition studied in this work also provides insights into the solvent exposure of these catalytic tyrosines. In Fig. 3 c, we show the snapshots of IRKD at different simulation times highlighting the side chain of three tyrosine residues (1158, 1162, and 1163). We observe that during TAMD, Y¹¹⁶² is the first residue to leave the active site and expose itself to the solvent, immediately followed by Y¹¹⁵⁸ (snapshots at 20.53 and 23.43 ns). The last tyrosine to expose itself to the solvent is Y¹¹⁶³ (snapshot at 34.69 ns), which becomes further buried due to conformational change in the A-loop, leaving Y¹¹⁵⁸ and Y¹¹⁶² solvent-exposed as seen in the active crystal structure (snapshots at 41.03 and 43.90 ns). These results suggest that Y¹¹⁶², Y¹¹⁵⁸, and Y¹¹⁶³ likely are the first, second, and third tyrosines respectively to undergo trans-autophosphorylation by the other kinase domain, which is consistent with earlier experiments (33). Also, Hubbard et al. (3) suggest that Y¹¹⁶² is likely the first residue to disengage from the active site because it unblocks the nucleotide-binding site of IRKD.

Free-energy calculation of the activation pathway via String method

The conformation of the A-loop in the equilibrated inactive IRKD is devoid of a defined secondary structure. However, we observe in our independent TAMD trajectories (see Fig. S6, Fig. S7, and Fig. S8) that the A-loop adopts a helical conformation, before unfolding and subsequently adopting an extended activelike conformation (Fig. 2 b and see Fig. S6, Fig. S7, and Fig. S8). The fact that TAMD induces exploration of the physical free-energy surface implies that states observed under TAMD with moderate fictitious temperatures (here, 5 kcal/mol) are in fact statistically significant, suggesting the folding/unfolding transition that drives DFG-flip is an important part of the activation mechanism. However, to validate this implication, and thereby elucidate the detailed transition mechanism, we used the string method (20,26–30) to refine the activation pathway generated by TAMD (run No. 1) toward the minimum free-energy path (MFEP; see Methods). The free energy profiles computed along both the initial and converged pathways appear in Fig. 4 a. The initial pathway generated by TAMD converges to the MFEP in 25 iterations during which the free-energy of the end image drops from ∼115 kcal/mol to ∼20 kcal/mol (Fig. 4 a and see Fig. S1).

Figure 4.

(Color online) (a) Free energy profiles along the initial (red) and final converged (black) activation pathways. (Profiles for all string method iterations appear in Fig. S1 in the Supporting Material.) (b) Representative conformations of IRKD from MFEP images 1, 4, 6, 12, 13, 14, 18, and 23, with the A-loop (red), and the side chains of Asp¹¹⁵⁰ and Phe¹¹⁵¹ (cyan and blue, respectively).

The α-helical conformations of residues Gly¹¹⁵² to Asp¹¹⁶¹ are observed in MFEP images 2–6, 8, 9, 11, and 12 (Fig. 4 b and see Fig. S2). We observe that images 11 and 12 belong to a metastable region with ΔF ≈ 18.24/18.71 ± 1.84 kcal/mol (with reference to image 0), respectively. Structurally, these two images show a half-way anticlockwise rotation of the side chains of Asp¹¹⁵⁰ and Phe¹¹⁵¹, where the conformation of these residues is neither as seen in the inactive nor the active crystal structures (Fig. 1 b). The flip completes between images 12 and 14, where we observe image 13 at a transition state with a forward free-energy barrier (12→13) of ∼0.79 kcal/mol and a backward free-energy barrier (13←14) of ∼1.07 kcal/mol. The important distinguishing feature of the conformation at image 13 is the loss of the helical character in A-loop, which further places the side chain of Phe¹¹⁵¹ underneath the αC-helix and Asp¹¹⁵⁰ in the ATP-binding pocket (see image 14 in Fig. 4 b). Final images beyond the 16th image also belong to a metastable region in the free-energy profile, with image 18 having the lowest ΔF ≈ 15.10 kcal/mol (with reference to image 0), where the intact R-spine is observed. Phe¹¹⁵¹ at image 18 has an RMSD of 0.51 Å with reference of its conformation in the active crystal structure. The free-energy profile also shows that the relative free-energy difference between images 9 and 16 is ∼1.5–2.5 kcal/mol, which likely suggests a dynamic equilibrium among these conformations. A positive ΔF suggests that IRKD prefers to remain in an inactive conformation in the absence of ATP or phosphorylated A-loop tyrosines.

Discussion

In this work, we have studied the large-scale conformational change in the A-loop of the insulin receptor kinase domain with a combination of TAMD simulations and the string method in collective variables. A major finding of this study is the observation of a spontaneous helical folding/unfolding transition in the N-terminal residues of the A-loop during TAMD that drives the flip of the kinase-conserved DFG-motif from a D-out/F-in to D-in/F-out conformation. In fact, the possibility that the A-loops of many other kinases adopt helical conformations is apparent from crystal structures (see Fig. S9). We find that the DFG-flip happens during the early part of the simulation when the A-loop is still relatively far away (in RMSD) from its conformation in the active state, whereas the inward rotation of the αC-helix is not observed until the A-loop has completely moved out to facilitate this functional motion (Fig. 2). Although the flip of F¹¹⁵¹ positions all residues of the hydrophobic R-spine on the same side, a contiguous surface (spanning both the N- and C-lobe) indicating an intact R-spine only forms post-αC-helix rotation. These results suggest that the R-spine can be dynamically disassembled via two different mechanisms: a), D-out/F-in conformation of the DFG-motif as seen in the inactive IRKD structure leads to a broken R-spine, and b), without the movement of the αC-helix as seen in TAMD simulations, the R-spine remains broken despite correct F¹¹⁵¹ placement. Consistent with our results, a careful structural analysis of various kinases by Taylor and Kornev (8) has suggested that IRK and AKT likely utilize the DFG-flip mechanism for the disassembly of the R-spine, whereas other kinases such as inactive Src and cyclin-dependent kinase 2, where the DFG-motif is already in a D-in/F-out configuration, likely disassemble R-spine via motion of the αC-helix. Our simulations, however, also suggest that not only the former but the latter mechanism as well is working for the R-spine assembly/disassembly in IRKD. Indeed, a mutant unphosphorylated inactive structure of IRKD (Asp¹¹⁶¹→Ala) with an ATP-analog has been crystallized by Till et al. (9) in which F¹¹⁵¹ is flipped, but the αC-helix is still in an outward inactivelike conformation. Integrity of the R-spine and the inward rotation of the αC-helix have been suggested to be the hallmarks of an active kinase, both of which are present in the TAMD-generated activelike state. The region in the vicinity of the R-spine has been suggested to be strategically of great importance (8) as many kinase inhibitors that out-compete ATP for binding to the active site exploit the flip of DFG-motif, F¹¹⁵¹ of which belongs to the R-spine. His¹¹³⁰ from the R-spine and Ala¹⁰²⁸ from the C-spine belong to other kinase-conserved motifs such as the HRD-motif (5) and the VAxK-motif (8), respectively. We find that the conformations of other residues in these conserved motifs in our simulations are consistent with the known active crystal structure (see Fig. S3).

In our simulations, we also observe a complex network of salt-bridging interactions formed by K¹⁰³⁰ (with E¹⁰⁴⁷ and D¹¹⁵⁰), and R¹¹⁵⁵ (with D¹¹⁵⁰ and E¹⁰⁴⁷). A key feature of many active kinases (8), the K¹⁰³⁰-E¹⁰⁴⁷ salt bridge, is observed, and an intermittent R¹¹⁵⁵-E¹⁰⁴⁷ salt bridge is also seen. The observation of the latter salt bridge is consistent with many other simulation studies on different kinases where this intermittent or stable salt bridge has been observed (13,15,16,34,35). Because the K¹⁰³⁰-E¹⁰⁴⁷ salt bridge helps anchor the αC-helix in the active state as seen in the present and previous simulation studies of other kinases, and K¹⁰³⁰ also interacts with nucleotide when bound (4), mutation of K¹⁰³⁰ are expected to abolish the activity of IRKD as was demonstrated by Ebina et al. (36).

The MFEP calculation via the string method reveals the existence of metastable intermediate states: at images 11 and 12 the A-loop has departed away from its inactivelike conformation and the αC-helix is still in an inactive conformation, whereas in images 18 and 19, the A-loop and αC-helix are in more activelike conformations. Similar intermediate states have been seen along the activation pathways studied in previous simulations of CDK5 (15), Hck (34,35), c-Src (14), and Abl (16). We also note that the magnitude of the free-energy barriers for this transition suggested in our work are consistent with similar studies on other kinases such as CDK5, where Berteotti et al. (15) computed a physical activation barrier of ∼16–20 kcal/mol, and Gan et al.'s (14) c-Src free-energy profiles indicate activation barriers of ∼20 kcal/mol. How the presence of an ATP molecule can help lower this activation barrier remains an important object of future study, because at physiological concentrations of 2–8 mM ATP (in most insulin-responsive tissues) (37), at least one out of the two IRK domains in each IR likely has bound ATP. Similarly, how the activation pathway is affected by inter-kinase contacts remains unclear at this point, as we have only simulated a single IRKD.

DFG-flip may not require transient helix-folding in the A-loop for all kinases. Indeed, in a study of the nonreceptor tyrosine kinase Abl, Shan et al. (16) combined ultra-long MD simulations and experiments to demonstrate pH-dependence of a drug (imatinib) binding to Abl kinase, which is consistent with the claim that the protonation state of the aspartate of the DFG motif, already likely having a significantly higher pK_a than free aspartate in Abl, plays a role in mediating the DFG conformational change. In that work, long unbiased MD simulations also spontaneously resulted in DFG-flip, but it must be emphasized that this evidently required mutation of Met²⁹⁰ to Ala, which unblocks access to a site for the phenylalanine side chain.

We stress that no such similar mutation was made in our study of IRKD. Furthermore, whereas the pK_a of the DFG-aspartate in active Abl kinase was determined to be 11.7 (16) via a Poisson-Boltzmann calculation using the H++ server (biophysics.cs.vt.edu/H++), meaning it is likely protonated, we find using the same method that the pK_a of the DFG-aspartate in IRKD is 5.3 in the inactive form (PDB:1IRK), and −0.3 in the active form (PDB:1IR3). There is, therefore, no basis to support a protonation-dependent mechanism for DFG-flip in IRKD. However, we cannot entirely rule out the possibility that the DFG aspartate in IRKD is protonated at physiological pH, as establishing this fact would require careful quantitative experiments as well as simulations of the key residues of IRKD in different protonation states, which are difficult to infer based only upon the crystal structures. It is probably worth remembering, therefore, that because Abl and IRKD hail from two wholly different families of kinases, it is not necessarily unlikely that they are regulated via different microscopic mechanisms.

As a final point of discussion, we find from this work that, as a tool for exploring CV space in an all-atom simulation of at least this system, TAMD alone can reveal important features of an activation mechanism. This is noteworthy because TAMD is not target-biased; it merely enables rapid exploration of the physical free-energy landscape for enhanced conformational sampling (17–19) relative to standard MD. Such exploration is by construction stochastic, and not every 40-ns TAMD trajectory generated a successful transition (Table 1). Nevertheless, we show that the successful trajectories harvested from TAMD trials reveal mechanistic features of the minimal free-energy transition pathway through a chosen CV-space, which should encourage the use of TAMD for conformational exploration in many other systems. We also suggest that TAMD exploration offers some possible advantages over other means of generating initial pathways for refinement using the string method. The pathway generated via TAMD was refined to an MFEP in ∼25 iterations; in contrast, application of the string method with swarms-of-trajectories to a targeted-MD-generated initial pathway for c-Src required over 55 iterations for convergence (14) (though it should be noted that this latter work considered a larger number of CVs than we do here).

Conclusions

Our simulations have provided detailed atomistic insights into the mechanisms governing the activation of the IRKD and associated DFG-flip. TAMD simulations combined with the string method reveal that flip of the DFG-motif in IRKD is made possible by a helical intermediate of the A-loop that spontaneously unfolds into an activelike state in the end. A free-energy difference (ΔF) of ∼15 kcal/mol (with an initial activation barrier of ∼17–20 kcal/mol) between the inactive and the activated-state is observed, suggesting that the kinase prefers to stay inactive in the absence of ATP or phosphorylation consistent with what has been seen in previous work (13,15,16,34,35). The observation of metastable states along the activation pathway also corroborates earlier modeling and simulation work on various kinases.