Data processing workflow with relion particle polishing

The dataset consisted of 468 micrographs, 41 (9%) of which were rejected after visual inspection, leaving 427 micrographs for processing. The super-resolution movies were first converted to MRC format with the newstack from IMOD (Mastronarde and Held, 2016), which increased their size to 5.1 GB/movie. Magnification anisotropy correction (measured to be 1.2%) and gain normalization were performed with mag_distortion_correct (Grant and Grigorieff, 2015). The movie frames were aligned with unblur (Campbell et al., 2012). Particles were picked by template matching in Gautomatch (http://www.mrc-lmb.cam.ac.uk/kzhang/). The CTFs were fitted with a homemade program in MATLAB (MathWorks, Natick, MA). The rest of the processing was performed in the GPU-accelerated beta version of Relion 2.0 (Kimanius et al., 2016). Particles were extracted from the aligned super-resolution micrographs and movies with 2x downsampling, resulting in a pixels size of 1.054 Å, matching the physical pixel size of the camera. The box size after downsampling was 180 pixels. The initial dataset of 145,870 particles was subjected to a 3D refinement with D7 symmetry which reached a resolution of 2.6 Å with a B-factor of 111 Ų. The output of the refinement was used to ‘polish’ the particles with the following parameters: running average three frames, stddev on translations one pixel, stddev on particle distance 100 pixels, B-factor highres-limit 2 Å, B-factor lowres-limit 7 Å. The polished dataset was 3D classified into five classes over several steps with finer and finer angular and translational sampling and E-step resolution limit of 12 Å. The best class, containing 93,596 particles (64% of the dataset), was 3D refined producing the final reconstruction with a resolution of 2.4 Å and a B-factor of 74 Ų. The other 3D classes refined to much lower resolutions with one of them showing some structure deformation in the form of separation between the top and bottom halves of the 20S. This was probably due to an incomplete assembly of some particles rather than being a valid functional state. To produce the resolution vs particle number data, subsets of random particles were extracted from the final dataset and 3D refined separately with a 60 Å low-pass filtered initial reference. The local resolution distributions were calculated with blocres from Bsoft (Heymann and Belnap, 2007).