Cryo-EM and image processing

The translation mixture was incubated at 10°C for 20 min and purified by gel filtration chromatography on a BioSuite 450 HR 8-μm column (Waters) at 4°C. Fractions that contained ribosomes and radioactivity (corresponding to f[³H]Met-tRNA^fMet) were collected and frozen in liquid nitrogen. The ribosome samples were applied to Quantifoil R 2/2 holey carbon grids coated with a thin layer of carbon. Grids vitrified in a FEI Vitrobot were then transferred to a Titan Krios (FEI) electron microscope operated at 300 kV. Images were acquired using a K2 detector at a magnification of 47,170 (yielding a pixel size of 1.06 Å). Initial beam-induced motion correction was performed at the level of entire micrographs (42). Contrast transfer function parameters were estimated using CTFFIND3 (43). Data were processed using the Relion software package (44).

An initial data set of 150,569 particles was manually picked and subjected to a round of two-dimensional (2D) class averaging. The selected 140,985 particles gave rise to a refined 3.2 Å map after particle movement correction (that is, “polishing”). The final accumulative electron dose was ~26 e⁻/Ų. These particles were used for treatment of structural heterogeneity (fig. S1): First, 3D classification with exhaustive angular search (0.9°) into five classes yielded three distinct recognizable classes—classical state–like (12.2%), hybrid state–like (19.1%), and the state with the P-site tRNA only (60%). The two smaller classes gave suboptimal reconstructions, and the particles from these classes were discarded.

To identify the complexes in a better-defined conformation, we performed additional rounds of classification on the isolated subpopulations. This approach allowed us to further clean the classical-state subset (the selected particles yielded a 5.1 Å map); however, despite the fragmented appearance of the A/P tRNA, they were unsuccessful in the attempt to find different subpopulations in the hybrid-state subset (these particles yielded a 4.65 Å map).

The visual inspection of the reconstruction for blurry parts of density indicated the existence of remaining structural heterogeneity in the main subset. The heterogeneity was particularly well seen in the mRNA structure. To improve the density, we took a new approach: subtraction of part of the signal from the particles, creating two new data sets of experimental images, each of them without one of the subunits. We then performed classification on the new set of particles with no signal for the 50S subunit while keeping all orientations fixed at the values determined in the refinement of the consensus 30S subunit reconstruction. Sorting into six groups yielded two major classes, one of which showed an improved density for the mRNA, particularly for the A-site secondary structure. The remaining classes were ignored. To obtain meaningful subclasses, it was necessary to limit the resolution (using the“-strict_highres_exp” Relion1.4 command) to 10 Å. Because the information about the interface between the subunits was not well defined in the separately masked reconstructions, the entire ribosome was refined to generate the final map at 3.6 Å resolution. Local resolution variability was estimated using ResMap (Fig. 1E and fig. S3) (45).

The take-off complexes represented approximately 60% of the total ribosome population in the sample, which matched the maximal percentage of bypassing of the in vitro system. The other two subpopulations were ribosomes that resembled hybrid-state (A/P, P/E, and rotated 30S) and classical-state forms (less than 20 and 10%, respectively). The quality/resolution of these 3D maps was low, and there were no structural details that might support the assignment of the peptidyl-tRNA or the identity of the tRNAs in A and P sites. At this point, we cannot specify whether these complexes were by-products of the sample preparation/purification or ribosomes stalled at previous translocation stages.

Note that multiple additional runs (not shown in fig. S1) with varying numbers of classes were performed at different stages of sorting. During these runs, orientation searches were performed exhaustively, that is, every 0.9° (restricted to local searches) or kept fixed at the orientations from the consensus refinements. Separate 3D autorefinements were also performed for each of the corresponding subsets after each round of classification. The visual inspection of the resulting 3D reconstructions, in conjunction with the local resolution assessment, helped in planning the best strategy to deal with the structural heterogeneity of the sample.