Visualization of translation and protein biogenesis at the ER membrane – Nature

CRISPR–Cas9 knockout of CCDC47

FreeStyle 293-F cells (Thermo Fisher Scientific, R79007) were transfected with the plasmid pSpCas9(BB)-2A-Puro (PX459) V2.0 from the F. Zhang laboratory (Addgene plasmid 62988) containing the 20-bp single guide RNA (sgRNA) target sequence 5′-CACCGGTACACGGTGAACTCGTGCG-3′, PAM: AGG or 5′-CACCGGGAGGAAGCGGGCGAGGTGC-3′, PAM:GGG. Transfection was performed using Lipofectamine 2000 (Thermo Fisher Scientifiic, 11668019) according to the manufacturer’s instructions using 1 μg DNA per ml of culture at a cell density of 1 × 10⁶ cells per ml. Cells were cultured for 48 h in FreeStyle 293 expression medium (Thermo Fisher Scientific, 12338018) on an orbital shaker (120 RPM) at 37 °C and supplemented with 5% CO₂. Two days after transfection, cells were collected and resuspended in complete Dulbecco’s Modified Eagle Medium (DMEM) (Thermo Fisher Scientific, 11966025) (supplemented with 10% fetal bovine serum (Thermo Fisher Scientific, 10100147) and GlutaMAX-I (Thermo Fisher Scientific, 35050061)) with 0.5 μg ml⁻¹ Puromycin (InvivoGen, ant-pr-1). Subsequently, cells were plated in T175 flasks (Thermo Fisher Scientific, 159910) and grown for 7 days in complete DMEM with 0.5 μg ml⁻¹ Puromycin with periodical medium exchange or sub-culturing when confluency was reached.

After 7 days of Puromycin selection, surviving cells were dislodged, collected, and resuspended at 5 cells per ml in conditioned complete DMEM. One-hundred and fifty microlitres per well of cell suspension was plated into sterile 96-well plates and cultured for 14 days. Cell colonies derived from single cells were used for further cell expansion. After 14 days in culture, conditioned complete DMEM was exchanged for FreeStyle medium and cell colonies transferred into 24-well plates. Subsequently, cells were grown to confluency and further expanded into 6-well plates and 10-cm dishes before analysis.

Cell culture

HeLa and U2OS cells (from ATCC, CVCL_0042 and CVCL_0030 in Cellosaurus.org, respectively) were grown in standard tissue culture conditions (37°, 5% CO2) in DMEM Glutamax (Gibco). HEK 293F cells (Thermo Fisher Scientific, R79007) were grown in suspension in FreeStyle medium with 120 rpm agitation. Cell lines were not authenticated and were tested for negative mycoplasma.

ER-vesicle preparation

HEK 293F wild-type or CCDC47 knockout cells (0.5–1 × 10⁶ cells per ml, 50 ml) were collected and washed (3 times with PBS, at 300g, 5 min, 4 °C). HEK 293F cells used for ER stress studies were treated with 10 mM DTT for 2 h before collection. Cells were resuspended in lysis buffer (2–4 ml, 10 mM HEPES-NaOH pH 7.4, 250 mM sucrose, 2 mM MgCl₂, 0.5 mM DTT, protease inhibitor cocktail (Roche)) and lysed using a Isobiotec cell cracker (5–10 passes, 14 μm clearance, on ice). The lysate was cleared (1,500g, 2–3 × 5 min, 4 °C, in 2 ml tubes) using a cooled tabletop centrifuge. Vesicles were pelleted (10,000g, 10 min, 4 °C), and washed with resuspension buffer (10 mM HEPES, 250 mM sucrose, 1 mM MgCl₂, 0.5 mM DTT). The pellet was resuspended at a concentration of ~50 mg ml⁻¹ determined by A₂₈₀, frozen in liquid nitrogen and stored at −80 °C until further use. The supernatant was used for proteomics as control.

Twenty micrograms of microsomes were used for SDS–PAGE followed by immunoblotting using antibodies against SEC61α (Abcam, ab15575; 1:1,000), TRAPγ (Sigma Aldrich, hpa014906; 1:1,000) and CCDC47 (Abcam, ab241608; 1:1,000).

Mass spectrometry data acquisition

Approximately 100 μg of the isolated ER-microsome and cytosolic fraction (supernatant) were digested using an S-Trap micro-MS column (protifi) according to the vendor’s protocol.

Proteins were solubilized in lysis buffer (10% SDS, 100 mM Tris, pH 8), reduced (100 mM TCEP), alkylated (400 mM CAA in isopropanol) and denatured (27.5% phosphoric acid). For protein trapping, samples were flown over an S-Trap micro spin column, (10,000g, 30 s) and further washed with binding buffer (100 mM triethylammonium bicarbonate (TEAB) buffer, in 90% methanol). Protein digestion was achieved with an overnight incubation at 37 °C using a water bath (Grant Instruments, JB Academy) after the addition of digestion buffer (10% trypsin, 2% lysine, 50 mM Tris). Protein peptides were retrieved by washing with elution buffer (50 mM Tris), using a tabletop centrifuge (10,000g, 1 min).

Eluted peptides were lyophilized and dissolved in 2% formic acid prior to liquid chromatography–mass spectrometry (LC–MS/MS) data acquisition. MS data were acquired using an Ultimate 3000 RSLC nano system (Thermo Scientific) coupled to an Exploris 480 (Thermo Scientific). Three technical replicates of each sample were measured. Peptides were first trapped in a pre-column (Dr. Maisch Reprosil C18, 3 μm, 2 cm × 100 μm) prior to separation on the analytical column packed in-house (Poroshell EC-C18, 2.7 μm, 50 cm × 75 μm), both columns were kept at 40 °C in the built-in oven. Trapping was performed for 10 min in solvent A (0.1% v/v formic acid in water), and the elution gradient profile was as follows: 0–10% solvent B (0.1% v/v formic acid in 80% v/v acetonitrile) over 5 min, 13–44% solvent B over 37 min, 44–100% solvent B over 4 min, and finally 100% B for 4 min before re-equilibration in 100% A for 8 min. The mass spectrometer was operated in a data-dependent mode. Full-scan mass spectra were collected in a mass range of m/z 350–1,300 Thomson (Th) in the Orbitrap at a resolution of 60,000 after accumulation to an AGC target value of 10⁶ with a maximum injection time of 50 ms. In-source fragmentation was activated and set to 15 eV. The cycle time for the acquisition of MS/MS fragmentation scans was set to 1 s. Dynamic exclusion properties were set to n = 1 and to an exclusion duration of 10 s. HCD fragmentation (MS/MS) was performed with a fixed normalized collision energy of 27% and the mass spectra acquired in the Orbitrap at a resolution of 30,000 after accumulation to an AGC target value of 10⁵ with an isolation window of m/z = 1.4 Th.

Raw data were processed using the MaxQuant software⁵¹ version 2.0.1.0 with standard settings applied. In brief, the extracted peak lists were searched against the reviewed Human UniProtKB database (date 15 July 2021; 20,353 entries), with an allowed precursor mass deviation of 4.5 ppm and an allowed fragment mass deviation of 20 ppm. Cysteine carbamidomethylation was set as static modification, and methionine oxidation, N-terminal acetylation as variable modifications (maximum 5 modifications per peptide allowed). Both LFQ quantification and ‘match between runs’ were enabled. The iBAQ values in Supplementary Fig. 4b are approximate absolute abundances of the identified proteins derived by the normalization of the summed peptide intensities by the number of theoretically observable peptides for a given protein.Raw data were processed using the MaxQuant software⁵¹ version 2.0.1.0 with standard settings applied. In brief, the extracted peak lists were searched against the reviewed Human UniProtKB database (date 15 July 2021; 20,353 entries), with an allowed precursor mass deviation of 4.5 ppm and an allowed fragment mass deviation of 20 ppm. Cysteine carbamidomethylation was set as static modification, and methionine oxidation, N-terminal acetylation as variable modifications (maximum five modifications per peptide allowed). Both LFQ quantification and ‘match between runs’ were enabled. The iBAQ values in Supplementary Fig. 4b are approximate absolute abundances of the identified proteins derived by the normalization of the summed peptide intensities by the number of theoretically observable peptides for a given protein.

Grid preparation

ER vesicles were diluted in resuspension buffer to a concentration of 2–3 mg ml⁻¹ and 2 μl were applied onto a glow-discharged lacey carbon grid (Quantifoil). Four m,icrolitres of BSA-conjugated gold beads (10 nm, UMC Utrecht) diluted in resuspension buffer without sucrose were added and mixed with the sample on grid. Grids were immediately blotted from the backside for 5–6 s and plunged into a mix of liquid ethane and propane using a manual plunger.

For the adherent cell lines (Hela and U2OS), cells were seeded on R2/2 holey carbon on gold grids (Quantifoil) coated with fibronectin in a Mattek dish and incubated for 24 h. The suspension HEK 293F cells were grown to mid-log phase, and the cells were then directly pipetted onto glow-discharged R2/1 Carbon on Copper grids (Quantifoil). Grids were immediately blotted from the back for 10 s and plunged into liquid ethane propane mix using a manual plunger.

Lamella preparation

Lamellae were prepared using an Aquilos FIB-SEM system (Thermo Fisher Scientific). Grids were sputtered with an initial platinum coat (10 s) followed by a 10 s gas injection system (GIS) to add an extra protective layer of organometallic platinum. Samples were tilted to an angle of 15° to 22° and 12 μm wide lamellae were prepared. The milling process was performed with an ion beam of 30 kV energy in 3 steps : (1) 500 pA, gap 3 μm with expansion joints, (2) 300 pA, gap 1 μm, (3) 100 pA, gap 500 nm. Lamellae were finally polished at 30–50 pA with a gap of 200 nm.

Data acquisition

We acquired 869 tilt series on a Talos Arctica (Thermo Fisher Scientific) operated at an acceleration voltage of 200 kV and equipped with a K2 summit direct electron detector and energy filter (Gatan). Images were recorded in movies of 7–8 frames at a target defocus of 3 μm and an object pixel size of 1.72 Å. Tilt series were acquired in SerialEM (3.8)⁵² using a grouped dose-symmetric tilt scheme⁵³ covering a range of ±54° with an angular increment of 3°. The cumulative dose of a series did not exceed 80 e⁻ Å⁻².

Lamella data used in this analysis has been collected in one session on a pool of grids of human cell lines. Twenty-seven tilt series were acquired on six different lamellae on a Talos Arctica (same instrument as above). Images were recorded in movies of 5–8 frames at a target defocus of 4 μm and an object pixel size of 2.17 Å. Tilt series were acquired in SerialEM using a grouped dose-symmetric tilt scheme covering a range of ±60° with a pre tilt of ±10° and an angular increment of 3°. The cumulative dose of a series did not exceed 70 e⁻ Å⁻².

Reconstruction and particle localization

Video files of individual projection images were motion-corrected in Warp (1.0.9)⁵⁴ and combined into stacks of tilt series with the determined contrast transfer function (CTF) parameters. The combined stacks were aligned using the gold fiducials in IMOD (4.10.25)⁵⁵. Per-tilt CTF estimation for entire tilt series was performed in Warp and full deconvoluted tomograms were reconstructed by weighted back projection at a pixel size of 20 Å. Ice thickness was determined manually for a subset of 50 tomograms and results in an average thickness of 156 nm. Particle coordinates were determined by template matching against a reconstruction of a human 80S ribosome filtered to 40 Å and downsampled to match the tomogram pixel size (20 Å) using pyTOM (0.994)⁵⁶. Most false-positive hits were manually removed in pyTOM. The determined positions of ribosomes were used to extract subtomograms and their corresponding CTF volumes at a pixel size of 3.45 Å (2× binned) in Warp. Video files of individual projection images were motion-corrected in Warp⁵⁴ and combined into stacks of tilt series with the determined CTF parameters. The combined stacks were aligned using the gold fiducials in IMOD⁵⁵. Per-tilt CTF estimation for entire tilt series was performed in Warp and full deconvoluted tomograms were reconstructed by weighted back projection at a pixel size of 20 Å. Ice thickness was determined manually for a subset of 50 tomograms and results in an average thickness of 156 nm. Particle coordinates were determined by template matching against a reconstruction of a human 80S ribosome filtered to 40 Å and downsampled to match the tomogram pixel size (20 Å) using pyTOM⁵⁶. Most false-positive hits were manually removed in pyTOM. The determined positions of ribosomes were used to extract subtomograms and their corresponding CTF volumes at a pixel size of 3.45 Å (2× binned) in Warp.

Lamellae data were processed as above with slight variations. Video files of individual projection images were motion- and CTF-corrected in Warp and combined into stacks of tilt series. The combined stacks were aligned using patch tracking in IMOD. CTF estimation for entire tilt series was performed in Warp and full tomograms were reconstructed by weighted back projection at a pixel size of 17.36 Å. Ice thickness was determined manually and was found to be <200 nm for all lamellae. Particle coordinates were determined by template matching against a reconstruction of a human 80S ribosome filtered to 40 Å using downsampled to match the tomogram pixel size (17.36 Å) pyTOM. The determined positions of ribosomes were used to extract subtomograms and corresponding CTF volumes at a pixel size of 8.68 Å (4× binned) in Warp.

Subtomogram analysis

The extracted subtomograms were aligned in RELION (3.1.1)⁵⁷ using a spherical mask with a diameter of 300 Å against a reference of an 80S ribosome obtained from a subset of the same data. The extracted subtomograms were aligned in RELION (3.1.1)⁵⁷ using a spherical mask with a diameter of 300 Å against a reference of an 80S ribosome obtained from a subset of the same data. The aligned particles were refined in M (1.0.9)¹⁷ using the reconstructions of the two half maps as a reference and a tight soft mask focused on the LSU at a pixel size of 3.45 Å. Particles were subjected to 2–3 rounds of refining image warp grid, particle poses, stage angles, volume warp grid, defocus and pixel size. After refinements, new subtomograms and their corresponding CTF volumes were extracted at a pixel size of 6.9 Å (4× binned) and subjected to 3D classification (without mask, without reference, T = 4 and classes = 50) to sort out remaining false positives, poorly aligned particles, and lone LSUs. The remaining 134,350 particles were used for subsequent focused classification steps to dissect ribosomal intermediate states or translocon variants.

Classification of ER ribosome populations

All 134,350 particles were subjected to 3D classification (without reference, with soft mask, T = 4, classes = 20) in RELION, focused on the area at the ribosomal tunnel exit including the membrane and translocon. Particles were sorted into SEC61–TRAP-bound, SEC61–TRAP–OST-bound, SEC61-multipass-bound and EBP1-bound ribosomes and a combined class of ribosomes with ambiguous densities. Ribosomes with ambiguous densities were subjected to two further classification rounds and sorted the respective class from above until no further separation could be achieved. Ribosomes that associated with the EBP1 were designated ‘soluble’, ribosomes associated with translocon variants were designated ‘membrane-bound’ and ribosomes associated with ambiguous densities were designated ‘unidentified’.

Subtomograms of the multipass translocon were recentered by 17 nm from the centre of the ribosome towards SEC61 and extracted in M at a voxel size of 6.9 Å. Subsequently, subtomograms were classified focused on the luminal domains of TRAP and NCLN (with reference of all multipass translocons, with soft mask, T = 4, classes = 3) or focused on the cytosolic domain of CCDC47 (with reference, with mask, T = 3, classes = 2). The TRAP-multipass translocon was further refined using local angular searches in RELION or, to obtain ribosome-centred reconstructions of the multipass translocon populations, subtomograms were recentered again by 17 nm towards the centre of the ribosome in M and subjected to another round of refinement.

Refinement of the OST translocon

The 42,215 best-correlating particles (5,554 particles were poorly aligned) of the OST-bound ribosome were used for refinement focused on the LSU in M using the same parameters as above at a pixel size of 1.72 Å (unbinned), which resulted in a reconstruction at an overall resolution of ~4 Å. However, densities of OST or TRAP in the ER lumen were poorly resolved. To improve local resolution of the translocon components, the reconstruction was recentered by 19.5 nm from the centre of the ribosome towards the OST translocon and subtomograms were extracted in M at a pixel size of 3.45 Å. The particles were aligned in RELION using the average of the recentered reconstruction of the OST translocon as reference and a tight soft mask focused on SEC61, TRAP and OST. Subsequently, the aligned particles were refined in M as above at a pixel size of 1.72 Å resulting in a reconstruction at an overall resolution of 8 Å. Local resolutions estimated using M¹⁷ ranged from 6–7 Å for the OST and 8–9 Å for TRAP and the N-terminal domain of RPN2, indicating flexibility. Local refinement focused on the TRAP complex did not improve its resolution, presumably because the protein complex was too small to provide sufficient signal for reliable refinement.

After refinement in M, translocon-centred OST-particles were extracted at a pixel size of 6.9 Å and subjected to classification in RELION (without reference, with mask, T = 10, classes = 4) focused on the chaperone binding site. The resulting classes were refined in M as above using masks focusing on SEC61, TRAP, OST and chaperone.

Classification of ribosomal intermediates

Ribosomal intermediate states were obtained by hierarchical classification focused on the rotation of the SSU and on the tRNA and elongation factor binding sites. First, all 134,350 particles were classified into classes of ribosomes with non-rotated and rotated SSU (with reference, with soft tight mask focused on SSU, T = 4, classes = 2). Subsequently, non-rotated and rotated particles were each subjected to two rounds of classification (with reference, with mask focused on tRNA and elongation factor binding site, T = 10–20, classes = 10–20). Classes with fragmented densities, such as pre/pre+, rotated−1/rotated−1+, non-rotated idle/translocation, were separated in the second round of classification (with reference, with mask focused on tRNA and elongation factor binding site, T = 10–20, classes = 2–4).

Classification of intermediate states was first performed for individual populations of ER translocon-bound or soluble ribosomes, which revealed similar results for each population. However, to improve performance of classification, especially for translocon-associated populations with a low number of particles, we pooled all translocon and soluble populations and performed classification of intermediates on the entire dataset. Subsequently, particle sets of individual intermediate states were dissected according to the translocon-associated and soluble ribosome populations.

The classification workflow was repeated four times to assess the technical uncertainties of 3D classification, which was determined at 5% to 15% and correlates inversely with class size. To assess experimental reproducibility, we combined two smaller datasets of ER-derived vesicles (31 tomograms, 6,101 particles; 58 tomograms, 3,836 particles) with the large dataset (869 tomograms, 134,350 particles) and processed them as described above. After obtaining classes of intermediate states, particle numbers were determined for each dataset and class.

The classification workflow was applied to in situ data with slight variations: extracted subtomograms were used for 3D classification with image alignment against a low pass filtered 80S ribosome map as reference in RELION to exclude false positive. The remaining 5,818 ribosome subtomograms were refined in RELION and re-extracted in Warp at a pixel size of 4.34 Å (2× binned). Two times-binned subtomograms were refined in RELION with a mask on the LSU prior to a first round of 3D classification without image alignment with a mask on the SSU to separate rotated from non-rotated ribosomes. A second round of classification was performed using a mask positioned on the tRNA and elongation factors sites, optimizing the mask extension and class number to this data in order to yield stable classes despite limited resolution and particle number. The different classes were finally subjected to iterative refinement in M.

Refinement of intermediate states

Classes of ribosomal intermediate states were simultaneously refined in M at a pixel size of 1.72 Å (unbinned) using tight masks focused on the entire 80S ribosome, tRNAs and elongation factors, which were individually generated for each intermediate. Refinement of image warp grid, particle poses, stage angles, volume warp grid, defocus and pixel size were performed iteratively (2–3 iterations). Globally or locally filtered and sharpened maps were generated by M and used for visualization or model building.

Model building

Initial models for each chain of SEC61 and the OST were downloaded from the Alphafold database⁵⁸. A polyalanine helical stretch was manually built to account for the plug density. The OSTA chains were manually docked into the higher-resolution OSTA SPA map EMD-10110, followed by refinement through an iterative cycling between phenix (1.20.1) refine⁵⁹, isolde (1.0b5)⁶⁰ and Coot (0.9.8.2)⁶¹. The initial model for TRAP was built using AlphaFold Colab³⁷ and Coot⁶¹. The initial model for TRAP was built using AlphaFold Colab for multimeric complexes⁶² and was divided into the transmembrane part and the luminal part. Each model was manually fitted into our subtomogram average (STA) density in UCSF Chimera (1.14.0)⁶³, followed by normal-mode guided refinement using iMODFIT (1.51)⁶⁴. Long flexible loops not visible in our density were manually removed from the models. SEC61, OSTA and luminal TRAP domains were fitted and refined into a STA centred on the OST, while the TRAP transmembrane helices were fitted and refined into the original ribosome-centred STA, in which they were better defined. Each model was refined using iterative cycling between phenix refine, Isolde and Coot. Models were then combined for one last round of refinement together in the OST centred STA. Validation was performed using Molprobity (4.5.1)⁶⁵. UCSF ChimeraX (1.3.0)⁶³ was used for visualization of all models and reconstructions.

Single-particle analysis

Suspension HEK 293F cells were grown to mid-log phase (0.5–1 × 10⁶ cells per ml, 50 ml). Cells were pelleted at 500 g for 5 min and washed twice in ice cold PBS and resuspended in 10 mM Hepes KOH, pH 7.5, 250 mM sucrose, 2 mM magnesium acetate, 0.5 mM DTT, 0.5 mM PMSF, protease inhibitor tablets). Cells were lysed with 30 passages through a 21-gauge needle. The lysate was cleared by centrifugation steps at 1,000g for 10 min, 1,500g for 15 min and 20,000g for 20 min. The final supernatant was loaded onto a 1 M sucrose cushion and spun at 300,000g for 1 h. The final ribosomal pellet was resuspended in lysis buffer and snap frozen in liquid nitrogen. For grid preparation, 3.5 μl of the ribosome preparation was pipetted onto glow-discharged R 3.5/1 2 nm C holey grids (Quantifoil) and blotted for 2.5 s at force 0 using a Vitrobot (Thermo Fisher Scientific) before subsequent plunging into liquid ethane.

Single-particle cryo-EM data were acquired on a Titan Krios (Thermo Fisher Scientific) equipped with a cold FEG, Falcon 4i detector and Selectris X energy filter 10 eV slit at a pixel size of 0.729 Å per pixel. A total of 17,000 movies was acquired with EPU 3 (Thermo Fisher Scientific) in EER format. A cumulative dose of 40 e⁻ Å⁻² was used.

The data was processed in Relion 3.1.1. Movies were motion-corrected and CTF was estimated. Particles were picked with the logpicker and reconstructed at a pixel size of 6 Å per pixel for subsequent 2D classification, followed by 3D classification with image alignment to exclude false-positive and low-quality particles. A total of 66,000 particles was then subjected to 3D classification without image alignment using a mask on the A tRNA site and the GTPase centre. 19,000 particles were selected in a class corresponding to the classical pre+ state, refined, re-extracted at 1.0 Å per pixel and refined again. CtfRefine was performed followed by another round of refinement. Masks on the A-site tRNA site and elongation factor, as well as on the peptidyl transferase centre were used for particle subtraction and focused refinements to improve the quality of the maps in these regions.

For model building, a previous crystallographic structure of eEF1A in the extended GDP bound conformation (PDB 4C0S) was used as starting model and was first briefly refined in real space in the higher-resolution crystallographic electron density map using Isolde and phenix refine, in order to improve the starting geometry of the model. The resulting model was then refined in our map through iterative cycling between phenix refine⁵⁹, Isolde⁶⁰ and Coot⁶¹. The model was validated using Coot⁶¹ and Molprobity⁶⁵.

Sequence conservation

The degree of sequence conservation was determined using the ConSurf server⁶⁶ using 150 homologous sequences with a sequence identity ranging from 35%–95%. The conservation score was plotted onto the surface of the respective protein model in UCSF Chimera.

Polysome analysis

For the neighbourhood analysis, ribosome positions and orientations were read from the RELION star files resulting from subtomogram alignment in a python script (Python 3.8.11, Numpy 1.20.3, Scipy 1.7.1). For each ribosome we determined distance vectors between itself and its n closest neighbours (n = 4), excluding neighbours further than 100 Å. The vectors were rotated with the inverse orientation of the respective ribosome, resulting in the coordinates of neighbours in the coordinate system of an ER-bound ribosome with the xy plane corresponding to the ER membrane. These vectors were sampled on a 3D-histogram with voxels corresponding to 15³ ų and divided by the total number of analysed neighbours to indicate the probability of finding a neighbouring ribosome particle in each voxel. The plots were projected on the xy plane to visualize the density of neighbours surrounding ER-bound and soluble ribosomes.

A threshold was chosen to identify clusters for trailing and leading neighbours. For ER-bound neighbours a binary mask was created in the 3D-histogram above a probability of P = 0.0005, while for soluble ribosomes the threshold was put at P = 0.0003. Both masks were dilated by two voxels. The soluble and ER-bound trailing masks were combined in a trailing mask for the whole dataset, and the same procedure was performed for the leading mask. The masks were used to annotate associations of ribosome pairs in a polysome. A trailing–leading connection was confirmed if the neighbour localized in the trailing–leading mask area and the analysed ribosome also positioned in the leading–trailing area of the respective neighbour (that is, the inverse calculation).

The trailing/leading states of neighbours were used in R to fit a multinomial mixed-effects logistic regression model (mclogit 0.9.4.2⁶⁷ in R 3.6.1). The ribosome’s state was used to predict probabilities of leading and trailing states, where the tomogram index was used as a random effect to account for sample and imaging variation. We used the same model to predict probabilities of translation states in polysome chains. For visualization, the probabilities were extracted with their 95% confidence interval, representing the region of 95% certainty that the modelled mean is the population mean. Variation between tomograms was shown by calculating the frequency of certain events per tomogram—for example, the 42nd tomogram might have 7 pre+ ribosomes of which 6 are associated in polysomes resulting in a frequency of 0.86. Random association probability was calculated by fractional abundance of each state in the dataset. For the plots showing the fold increase, the modelled mean and confidence interval lower and upper bounds were divided by the random association probability and displayed with logarithmic y-axis. Statistical significance for the fitted logistic parameters was determined with a two-sided Wald-test (as reported by mclogit) and used to annotate plots. P values were adjusted for multiple comparisons with the Hochberg method as implemented in R with p.adjust (method=‘hochberg’).

Previously published data

We made use of previously published atomic models from the PDB (accession codes 5AJO, 4CXG, 4UJE, 6Y0G, 6Y57, 6GZ5, 6Z6L, 6Z6M, 5LZS, 4C0S, 5LZT, 5IZK, 6O85, 5LZZ, 6GZ3, 6GZ4, 6GZ5, 6SXO, 1BN5, 6W6L, 6ENY, 6S7O, 3JC2) and the AlphaFold Protein Structure Database (AF-O00178, AF-P30101). Moreover, we used the following EM densities from the EMDB for analyses: EMDB-2904, EMDB-2908.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.