Helical assembly subunit refinement and classification (HASRC)

The helical-single-particle 3D reconstruction procedure generates density maps where all the asymmetric units in the helical assembly (one tubulin heterodimer with one-bound kinesin motor domain) are averaged. To independently determine the structure of the two motor domains of kinesin dimers bound to the microtubule (two-heads-bound), we implemented a procedure (HASRC) to isolate and classify individual subunits in a helical assembly. Different from a previous method⁴⁹, HASRC can separate two-bound kinesin dimer complex structures in fully decorated microtubules at near-atomic resolution. HASRC was implemented in Relion 3.1 adapting ideas and methods previously used to separate and classify the subunits contained in symmetrical multi-subunit assemblies^60,61. These ideas have also been used recently to refine the structure of microtubule protofilaments⁶². We implemented specific subunit location refinement and classification steps in HASRC that were essential to separate coexisting kinesin motor domain conformations at high resolution.

In addition to the ability to separate coexisting subunit structures, HASRC allowed to account for local lattice distortions, which resulted in up to 0.4-Å resolution improvement over the helically averaged maps. Thus, to produce maps at the highest possible resolution, we applied HASRC to all datasets, dimers in the two-heads-bound states as well as monomers and dimers in the one-head-bound states.

For each of the four KIF14 dimer two-heads-bound states datasets (K755 and K772, ANP and AAF states), the following procedure was used (Supplementary Fig. ²):

Relion helical refinement. The two independent Frealign helical refined half datasets were subjected to a single helical autorefinement in Relion 3.1 where each dataset was assigned to a distinct half-set and using as priors the Euler angle values determined in the helical-single-particle 3D reconstruction (initial resolution: 8 Å, sigma on Euler angles sigma_ang: 1.5, no helical parameter search).

Asymmetric refinement with partial signal subtraction. An atomic model of a KIF14 dimer two-heads-bound state was used to generate two soft masks (Supplementary Fig. ^{2b, c}) using EMAN pdb2mrc and relion_mask_create (low-pass filtration: 30 Å, initial threshold: 0.05, extension: 14 pixels, soft edge: 6 or 8 pixels). One mask (mask_full) was generated from a KIF14 dimer model bound to two tubulin dimers while the other mask (mask_kinesin) was generated with only the kinesin coordinates. The helical dataset alignment file was symmetry expanded using the 15R microtubule symmetry of the dataset. Partial signal subtraction was then performed using mask_full to retain the signal within that mask. During this procedure, images were re-centered on the projections of 3D coordinates of the center of mass of mask_full (C_M) using a box size as indicated in Supplementary Table ². The partially signal subtracted dataset was then used in a Relion 3D autorefinement procedure using as priors the Euler angle values determined form the Relion helical refinement and the symmetry expansion procedure (initial resolution: 8 Å, sigma_ang: 2, offset range corresponding to 3.5 Å, healpix_order and auto_local_healpix_order set to 5). The CTF of each particle was corrected to take into account their different position along the optical axis.

3D classification of the kinesin signal. A second partial signal subtraction procedure identical to the first one but using mask_kinesin and with particles re-centered on the projections of C_M was performed to subtract all but each pair of kinesin signals (Supplementary Fig. ²). The images obtained were resampled to 3.5 Å/pixel and the 3D refinement from step 2 was used to update the Euler angles and shifts of all particles. A 3D focused classification without images alignment and using a mask for the kinesin generated like mask_kinesin was then performed on the resampled dataset to separate the kinesin states (8 classes, tau2_fudge: 4, padding: 2, iterations: 175). Two of the resulting classes contained two well-resolved kinesin motor domains while the others had absent or not well-resolved kinesin densities (Supplementary Fig. ²f). The two classes with well-resolved kinesin densities differed in the location of the density connecting the two kinesin motor domains: one class with the connecting density at the center corresponding to a dimer with leading and trailing kinesin heads and the other class corresponding to the two unconnected kinesin heads of two distinct dimers. These two classes were equally populated as expected from the procedure used, which samples the microtubule axially at each tubulin heterodimer, rather than the two heterodimers span of the kinesin dimer. For the much smaller MT-K772-ANP dataset, another classification strategy had to be used: first a focused classification (3 classes, tau2_fudge: 6, padding: 2, iterations: 25) was used to eliminate the particles generating a low-resolution class average and an undecorated class average. A second focused classification (2 classes, tau2_fudge: 16, padding: 2, iterations: 175) enabled to separate the main class of the first classification into two similarly populated dimer configurations.

Subunit refinement. The subset of particles belonging to the class with a centered isolated dimer was further refined using a Relion 3D autorefinement with the same parameters used in step 2.

3D reconstructions with original images (not signal subtracted). To avoid potential artefacts introduced by the signal subtraction procedure, final 3D reconstructions were obtained using relion_reconstruct on the original image-particles without signal subtraction. Map filtration was then performed the same way as with the helically averaged maps, without symmetry imposition.

For the one-head-bound datasets (K735, K743, and K748 in all nucleotide states and K755 and K772 in the Apo and ADP states) the same procedure as described above was employed with the following modifications (Supplementary Fig. ³): in step 2, the mask_full was generated with a PDB containing 1 kinesin motor bound to 1 tubulin dimer and two longitudinally flanking tubulin subunits. The mask mask_kinesin was generated with a KIF14 motor domain model. All the datasets produced at least one class where the kinesin motor densities were well-resolved. In all cases, only a single motor domain configuration was found except for the K743 construct in the ANP and AAF states where two well-resolved classes with two different kinesin conformations (open and closed) were found (Supplementary Fig. ^3h–k). No further refinement was performed after classification (no step 4). The presence of two coexisting conformation in these datasets is consistent with the mixture of open and closed configuration densities observed in the helically averaged maps (Supplementary Fig. ¹⁰).