(A) Processing of the core part of the complex. 2186 and 6580 movies from two different data sets were preprocessed in parallel and used for initial particle sorting by 2D and 3D classifications using the box size of 128 pixels with 4.85 Å per pixel. After initial classifications, approximately 625,000 particles were selected and combined for further processing. A round of ab-initio reconstruction followed by hetero-refinement with five classes was performed to separate the best particles for further refinements. About 407,000 particles from three classes, which reached 9.9 Å resolution (Nyquist for the binned data), were extracted using the box size of 320 pixels with 1.94 Å per pixel, and another round of 2D classification was performed. Afterward, a new round of ab-initio reconstruction and hetero-refinement with six classes was conducted. Particles from two classes that reached the best resolutions (7 and 7.1 Å) were further processed separately. The first of the selected two classes was refined by NU-refinement to 4.3 Å resolution, while the second class was refined to 4.9 Å. Subsequently, particles were extracted in full box size (672 pixels with 0.924 Å per pixel) and used for further homogeneous and NU-refinement, which resulted in two consensus maps, better resolving either upper or lower part of HOPS (reached 4.2 and 4.4 Å resolution, respectively). FSC curves generated in cryoSPARC are shown for both consensus maps. Afterward, local refinements were applied to improve each of the consensus maps. Local refinements of the first map resulted in two maps covering the SNARE-binding module (3.6 Å resolution) and the upper part for the core of HOPS (4 Å resolution). Local refinements of the second consensus map provided maps of the bottom of HOPS core better covering either Vps18 (4.4 Å resolution) or Vps39 (5 Å resolution). The subset of particles used for generation of both consensus maps was in parallel probed for flexibility using 3D classification (Figure 2—figure supplement 3A,B) and 3D variability analyses (Figure 2—figure supplement 3D) (dashed arrows; see Materials and methods for details). (B) Processing of the distal parts of HOPS. Movies used for core reconstruction (see (A)) were combined with movies from two additional data sets (2841 and 8338 movies, respectively) and preprocessed. After template picking, about 3.9 million particles were selected and subjected to several rounds of heterogeneous, homogeneous, and NU-refinements to select classes best resolving upper (Vps41) and lower (Vps39) distal parts of HOPS. Afterward, selected particles were used for a round of NU-refinement with a box size of 448 px with 1.82 Å per pixel. The particles from this NU-refinement were then subjected to 3D variability analysis with masks covering either the Vps41 or Vps39 volume. This was followed by 3D variability display in the cluster mode to further select particles. Finally, particles from best clusters were subjected to local refinements at a full box size of 882 pixels with 0.924 Å per pixel resulting in the maps resolving N-terminal regions of Vps39 (7.5 Å resolution) or Vps41 (6.9 Å resolution).