High-resolution structures of kinesin on microtubules provide a basis for nucleotide-gated force-generation.

Protein expression and purification

The wild-type, monomeric human K349 construct was bacterially expressed and purified as described ( Kull et al., 1996 ), and 15% (wt/vol) sucrose was added before snap freezing in liquid nitrogen and storing at −80°C. A plasmid for the mutant N255K construct was generated from the wild-type plasmid using the QuikChange site-directed mutagenesis kit (Agilent Technologies; Santa Clara, CA). Thawed K349 (either wild-type or N255K) was exchanged into EM buffer (25 mM PIPES, 25 mM KCl, 1 mM EGTA, 1 mM DTT) using three rounds of dilution and concentration in a Microcon ultracentrifugal filter (EMD Millipore; Billerica, MA). Microtubule batches were grown from 250 µg of lyophilized bovine brain tubulin (Cytoskeleton; Denver, CO), resuspended in 25 µl EM buffer and clarified (Beckman TLA 120.2, 100K RPM, 4°C) prior to incubation at 37°C. Taxol (2 mM in DMSO) was added to equimolar levels with tubulin after 10 min of incubation. Following ∼45 min of polymerization, microtubules were brought to room temperature and a ∼twofold excess of K349 was added prior to pelleting through a glycerol cushion (50 µl of EM buffer + 60% glycerol wt/vol + 200 µM taxol) in order to remove unbound motor and unpolymerized tubulin (20 min, Beckman TLA 120.2, 50K RPM, 24°C). The motor–microtubule complex was resuspended in ∼10 µl of EM buffer plus 200 µM taxol.

Mant-ADP release measurements

Stopped-flow measurements were performed at 298 K using an SF-300X (Indiana University) stopped-flow apparatus (KinTek Corp.; Austin, TX) equipped with a Xenon arc lamp (Hamamatsu; Japan). Kinetics of the interaction of mant-ADP with kinesin was measured by equilibrating a kinesin·mant-ADP (1:1) complex followed by rapid mixing with a high concentration of MgATP or varying MT concentrations plus MgATP to chase the mant-ADP from the active site as described ( Sadhu and Taylor, 1992 ). Mant fluorescence was monitored over time, I ex,max = 356 nm, I em,max = 448 nm (400 nm long-pass filter).

Cryo-EM sample preparation

Samples were plunge-frozen in liquid ethane, using Quantifoil holey carbon grids with 1 µm hole diameter, 1.5 µm spacing (Quantifoil Micro Tools GmbH; Germany). In order to optimize motor decoration on the microtubules, glow discharge was not applied to the grids and samples were diluted ∼5–10× into distilled water (0.35 µl sample plus 3 µl d(H 2 O)) on a piece of Parafilm prior to grid application, in order to compensate for evaporation that occurs prior to contact with liquid ethane ( Sindelar and Downing, 2007 ). For the ADP, AMPPNP, and ADP•Al•F x conditions, the grid droplet mixture was supplemented by 2 mM nucleotide (2 mM ATP + 2 mM AlCl 3 + 10 mM NaF for the latter condition). Following our previously published protocol ( Sindelar and Downing, 2007 ), most of the initially applied buffer on the grids was wicked away by touching the grid edgewise with a piece of filter paper (Whatman). The grids were then blotted completely, and after a 0.5–1 s delay, immersed into liquid ethane using a home-built plunge-freezing apparatus. Consistent with our prior experience ( Sindelar and Downing, 2007 ), failure to dilute the sample buffer ∼5–10× prior to plunge freezing led to poor and/or inconsistent kinesin decoration of the microtubules.

Data collection and image processing

Micrographs of the microtubules decorated by the wild-type K349 construct were collected using the SerialEM package to collect data semi-automatically on 300 kV FEG-equipped electron microscopes (FEI F30 for the no-nucleotide data set, FEI Titan for the ADP•Al•Fx data set), using K2 direct electron detecting cameras (Gatan; Pleasanton, CA) in video mode. For bare microtubules, and microtubules decorated by the N255K construct, micrographs were collected using a TF-20 FEG-equipped instrument and a Gatan US4000 CCD detector. In all cases, the defocus was systematically varied from approximately 1 to 2.5 µM through the course of the data collection. The total dose was ∼15 electrons/Angstrom squared, distributed over 15 or 16 video frames. The no-nucleotide data set was collected at ∼13K magnification using the camera's super-resolution mode and subsequently binned 2×, yielding 4K by 4K image dimensions with an effective pixel size of 1.99 Å. The ADP•Al•Fx data set was collected at approximately the same magnification as the no-nucleotide data set but in regular counting mode, resulting in 4K by 4K images with an effective pixel size of 2.097 Å. After performing drift analysis ( Li et al., 2013 ), whole-image video frames were aligned and averaged (K2 data only) before manually selecting overlapping boxed segments corresponding to individual microtubules. Defocuses and astigmatism parameters were estimated using the CTFFIND3 program ( Mindell and Grigorieff, 2003 ). Single-particle analysis and 3D reconstruction was then performed as described ( Sindelar and Downing, 2010 ), with some modifications. Initial reference models were generated by applying a parametric model of the microtubule ( Chrétien and Wade, 1991 ) to PDB models of the no-nucleotide kinesin-microtubule complex ( Sindelar and Downing, 2007 ), converting the atomic coordinates to EM density using the SPIDER package ( Frank et al., 1996 ), and applying a low-pass filter with a 20 Å cutoff frequency. Multi-reference analysis performed using a range of different microtubule assemblies (12–15 protofilaments) established that 14-protofilament microtubules outnumbered 13-protofilament by approximately 3:1, with only marginal populations of other symmetry forms. Based on this determination, the 13- and 14-protofilament sub-populations were separately selected for further analysis. Initial estimates of the in-plane rotation of individual segments were made using the Radon transform, as described ( Li et al., 2002 ). Automated scripts using the SPIDER package ( Frank et al., 1996 ) were used for semi-exhaustive searching of XY shifts and Euler angles (out-of-plane tilt range: +/−15°). Reference alignment was used to determine the position of the microtubule seam on a per-microtubule basis as described ( Sindelar and Downing, 2007 ), as was the filament polarity, and then every segment was subjected to local refinement using the established seam orientation and polarity. Evidently due to disorder in the N255K kinesin head orientations, the correlation scores used to identify the seam orientation were far noisier for this mutant. Following initial refinement of the Euler angles and shifts for each box, the estimated position of each 8 nm repeat of the microtubule was mapped back onto the micrograph and a new stack of boxes was extracted for a final round of SPIDER refinement. Thus, the final data set included exactly one box for each 8 nm repeat identified in the selected microtubules. Subsequent structure refinement and 3D reconstruction was performed using the FREALIGN package ( Grigorieff, 2007 ) with specific modifications for helical image processing ( Alushin et al., 2010 ). For the FREALIGN refinements, four rounds of refinement and reconstruction were initially performed, using successively higher resolution cutoffs for the refinement target (20 Å, 15 Å, 12 Å, 10 Å). For each reconstruction, 13- or 14-fold pseudo-helical symmetry was applied in order to transform all imaged asymmetric subunits onto a single ‘good’ protofilament, as previously described ( Sindelar and Downing, 2007 ). Helical parameters were derived from the corresponding, canonical microtubule form ( Chrétien and Wade, 1991 ) but adapted for the measured axial repeat distance in our data sets. The reconstructed ‘good’ protofilament (along with the remainder of the reconstruction) was subsequently replicated and transformed 13 or 14 times using the same symmetry parameters in order to generate protofilament models for the entire microtubule. The transformed ‘good’ protofilaments were subsequently selected by complementary wedge-shaped masks and summed in order to obtain the final asymmetric microtubule model. To reduce the influence of solvent noise in the refinement, a tight mask around the protein density was generated from a thresholded low-resolution version of the current map, and then smoothed with a soft Gaussian filter (∼10 Å half-width). This mask was then applied to the reconstructed volume prior to the next round of refinement. The same mask was employed for FSC calculations. Once the FREALIGN steps were completed, the resulting 3D map was then filtered to 20 Å and fed back into the SPIDER/FREALIGN pipeline, in order to reduce errors related to seam identification, polarity, and local searches. This additional step was omitted from the refinement of the N255K mutants. Following the second cycle of SPIDER/FREALIGN analysis, FSC calculations indicated that the resolution of the wild-type reconstructions was approximately 6 Å (0.143 criterion; Figure 1—figure supplement 1A,B ), and no signs of higher-resolution features were apparent. We then subdivided the aligned video frames corresponding to each micrograph into summed groups of three (5 frames each), and re-extracted a new FREALIGN image stack from these sub-frames, resulting in a threefold larger number of stacked images. We then performed four more rounds of FREALIGN analysis with the new stack, using Euler angles and shifts determined previously as starting parameters. The resulting reconstructions resolved individual beta strands and alpha helical pitch in some portions of the map ( Figure 1—figure supplement 1F,G ), although as noted in the main text the kinesin density was more poorly resolved and there was evidence of anisotropic blurring in the tubulin density (results not shown). The final resolution, obtained through FSC comparison of reconstructed half-data set volumes after applying a soft mask, was ∼5 Å (0.143 criterion) for both nucleotide states ( Figure 1—figure supplement 1A,B ). For the final no-nucleotide reconstruction, a total of 10,029 8 nm repeats were used, for a total of ∼140000 asymmetric units (14 protofilaments). A total of 2394 8 nm repeats, for a total of ∼33,600 asymmetric units, were used for the final ADP•Al•F x reconstruction (14 protofilaments). Molecular dynamics flexible fitting calculations In order to derive an atomic model for the microtubule complex of ATP-bound kinesin, we subjected the coordinates of tubulin and kinesin from the 4HNA structure to a hybrid all-atom molecular dynamics method (MDFF) in which cryo-EM density contributes a force field term (steering potential) designed to guide the simulation toward the experimentally observed structure ( Trabuco et al., 2008 ). Version 1.91 of NAMD was used for all molecular dynamics calculations ( Phillips et al., 2005 ). We used an explicit solvation model (TIP3P), because conformational instabilities were observed in trial simulations that used a less expensive continuum solvent model. All simulations were run using periodic boundary conditions, but no attempt was made to model the longitudinal or lateral interfaces of tubulin. While this omission likely produced artifacts near the boundaries between the simulated tubulin heterodimer and neighboring tubulin subunits, these boundaries are remote from the motor–microtubule interface. Moreover, the EM steering potential used in these simulations conformation strongly clamps the conformation of tubulin. Thus, our simulation setup is expected to minimize or eliminate the propagation of any boundary artifacts toward kinesin and its interface with tubulin, which is the focus of the current study. To complete the starting model, twenty water molecules were manually placed within the kinesin nucleotide pocket at sites corresponding to crystallographic waters identified in a high-resolution (1.7 Å) X-ray structure of KIF4 kinesin co-complexed with AMPPNP ( Chang et al., 2013 ). Hydrogen atoms were added to protein atoms and crystallographic water molecules using the ‘guesscoord’ command from the NAMD package ( Phillips et al., 2005 ). The system was then placed in a TIP3P water box of dimensions 125 × 125 × 95 Å, and sodium and chloride ions were randomly added to this water box to a concentration of 50 μM, at a ratio that neutralized the charge of the system. After energy minimization (600 steps), the system was subjected to a three-phase equilibration, in which protein atoms were initially subjected to positional restraints that were then progressively released. The equilibration simulations were performed in the NPT ensemble, using a Nosé–Hoover Langevin piston with a target pressure of 1 atm, a decay period of 200 fs and a time constant of 100 fs. The temperature was maintained at 310 K using a Langevin temperature bath with a time constant of 5 ps −1 . In the first equilibration phase, which ran for 50 ps, all protein and ligand atoms were restrained by a 20 kcal/mol/Å harmonic potential; the second equilibration step restrained protein backbone atoms only and the nucleotide ribose rings (50 ps); and the final step removed all harmonic constraints and was run for 600 ps. For all simulations, the CHARMM27 force field was used, with long-distance electrostatic interactions computed by the particle-mesh Ewald method and a non-bonded cutoff distance of 10 Å. Force field terms for GDP and GTP were adapted from ADP and ATP, respectively. Following equilibration, we performed MDFF simulations, but with substantial modifications in order to reduce the possibility of overfitting, which has been shown to be a significant concern with this method ( Bai et al., 2013 ). Importantly, we limited the influence of the EM steering potential to include only protein backbone atoms and nucleotide ligands. Thus, side chain atoms in our MDFF simulations are subjected exclusively to physics-based force field terms, which is reasonable because side chains were not visualized in our maps. Related to this point, as noted in the main text, a low-pass filter was used to remove signal beyond 6 Å resolution in the cryo-EM map, in order to reduce the level of contaminating noise prior to generating the EM steering potential. Limiting the EM steering potential to backbone atoms tends to moderate the contribution of the EM term in the simulated energy function, thus compensating for the limited number of experimental constraints available from our cryo-EM maps. Convergence of the fitted structures within the MDFF trajectories In conventional applications of MDFF, the EM steering potential is introduced as a step function after an initial equilibration period, and the relative strength of the potential (compared with the physics-based terms in the molecular dynamics force field) is described by an adjustable parameter, ξ, whose magnitude must be empirically determined. If the value of ξ is too low, it will not force a conformational transition within an achievable amount of simulation time; on the other hand, if the value is too high the molecular conformation becomes distorted (‘overfitting’). However, a quantitative method for choosing an appropriate value for the ξ parameter has not been presented. To address this problem, we developed a modified MDFF protocol in which the value of ξ follows a linear ramp function, starting at zero and slowly increasing over the course of 10 nanoseconds. We then monitored the system for convergence by calculating the root mean squared deviation between the coordinates of the fitted structure and the starting model ( Figure 2—figure supplement 1 ). The resulting simulation trajectory exhibits a rapid conformational transition that occurs within the first 2 nanoseconds of the simulation, corresponding to relatively low values of the GSCALE parameter (GSCALE