Image processing

A diagram of the procedures used in data processing is presented in Supplementary Fig. ². Approximately 3000 particles were manually picked and used to generate 2D classes for templates for auto-picking. A total of 1,730,910 particles were auto-picked from 4100 micrographs with RELION 2.0⁵⁴. After 2D classification, ten good 2D classes were used to generate an initial model using e2initialmodel.py⁵⁵, and a total of 1,001,249 good particles were then selected and subjected to 3D auto-refinement. The particles were further subjected to several cycles of 3D classification with six classes and a local angular search step of 3.75° with the output from different global angular search iterations of the 3D auto-refinement as input. The class with fully intact particles was considered as a good class, which contains useful high-resolution information and usually has the smallest value of the accuracy of rotation and translation. A total of non-duplicated 655,998 particles were selected from the good classes of local angular search 3D classification. These particles were subjected to local angular search 3D auto-refinement with a soft mask applied, resulting in a 4.5-Å resolution map. The particles were classified into four classes using multi-reference, and the best classes were selected and combined. The final particle number for the 3D auto-refinement is 105,118, thereby resulting in a 4.1-Å resolution map after post-processing. The resolution was estimated with the gold-standard Fourier shell correlation 0.143 criterion⁵⁶ with the high-resolution noise substitution method⁵⁷.