2.4. Cryo-EM data processing

Cryo-EM data processing was carried out using a combination of Sphire 1.3 (Moriya et al., 2017), Relion 3.0.8 (Scheres, 2012), and Topaz 0.2.3 (Bepler et al., 2019) and is summarized in Supplementary Figures 1 and 2. For each individual data set, individual frames of dose-fractionated exposure were aligned and summed using MotionCor2, and the defocus of the summed frames were estimated using CTER (Penczek et al., 2014; Zheng et al., 2017). The outputs of the MotionCor2 and CTER were used to perform CTF and drift assessments within the SPHIRE software suite (Moriya et al., 2017). The defocus, astigmatism frequency, and drift cutoff values, along with the resulting average drift and final number of micrographs, for each data set are indicated in Supplementary Table 5. An initial set of 500–1000 particles were manually picked for each data set from a subset of aligned movies and were used to train a neural-net automated particle picker using Topaz software. A Topaz cutoff score was chosen to yield 95% recall of true positives (as determined by generation of a precision-recall curve) (Bepler et al., 2019).

Frame alignment and CTF estimation were then rerun on each truncated set of movies using Relion’s implementation of MotionCor2 and CTFFind4, respectively (Mindell and Grigorieff, 2003; Zheng et al., 2017). The coordinate set of all particles in each data set were used to re-extract particles (box size 250 pix) and perform two rounds (25 iterations each) of initial reference-free 2D classification (mask diameter 260 Å) to generate 200 2D class averages. After each round of 2D classification, only intact, ring-shaped classes, which are easy to identify by eye, were selected to continue in the refinement process (see Supplementary Table 6 for the fraction of particles remaining during each step of the refinement process). The particles selected after the second round of 2D classification were used to generate a de novo initial reference-free 3D model using no imposed symmetry. Following imposition of D2 symmetry, this model was used for unsupervised 3D classification (6 classes, 100 iterations, no symmetry imposed), and only classes containing intact rings were selected for further refinement in each data set. Following selection of unbroken particles from 3D classes, 7492 particles were randomly selected from each data set for one round of 3D autorefinement. Following refinement, combination of the two half-maps along with B-factor adjustment (“postprocessing”) was performed using a mask generated with a low pass filter of 15 Å, initial binarization threshold 0.0119, binary map extended by 3 pix, with a soft edge of 3 pix. Analysis of 3D FSC results for each data set was completed using the Remote 3DFSC Processing Server found at https://3dfsc.salk.edu/ and these results were analyzed using the two half maps generated from 3D autorefine and the postprocessed map as inputs (Tan et al., 2017).

For processing of the full-size (untruncated) data sets, following selection of particles from 3D classes, all particles in each data set were used during the aforementioned 3D autorefinement and postprocessing steps. Following postprocessing, CTF refinement was performed with CTF parameter fitting and per-particle defocus fitting flagged, and then particle polishing was performed for each data set using the Bayesian method of particle motion estimation in Relion. The mask generated for the chameleon-plunged full size data set was generated with a low pass filter of 15 Å, initial binarization threshold 0.0529, binary map extended by 4 pix, with a soft edge of 3 pix. Similarly, the mask generated for the Vitrobot-plunged full size data set was generated with a low pass filter of 15 Å, initial binarization threshold 0.0089, binary map extended by 3 pix, with a soft edge of 3 pix. The refined particle sets were used to rerun high-resolution 3D refinement and postprocessing with the aforementioned masks to generate the final maps. The final resolution at FSC = 0.143 was 4.3 Å for the chameleon-plunged full size data set and 5.6 Å for the Vitrobot-plunged full size data set. Local resolution was estimated for both data sets using the Relion LocalRes executable within the Relion 3.0.8 GUI.