Molecular Architecture of the 40S⋅eIF1⋅eIF3 Translation Initiation Complex

1Protein synthesis is catalyzed by the ribosome in a process that consists of initiation, elongation, and termination. In bacteria, three initiation factors (IF1, IF2, IF3) and the Shine-Dalgarno sequence are sufficient to accurately pair the AUG start codon in messenger RNA (mRNA) with the anticodon loop of methionyl initiator transfer RNA (Met-tRNAi). Eukaryotes, in contrast, require at least 25 different polypeptides assembled into eight eukaryotic initiation factors (eIFs) to initiate protein synthesis and employ a complex scanning mechanism to probe the 5′ leader sequences of mRNA for the correct start site. Initiation is targeted by a number of regulatory pathways linked to cellular processes such as cell growth, differentiation, and environmental stress responses, and the functional disruption or decoupling of these regulatory interactions has been observed in a number of cancers.

2Eukaryotic translation initiation begins with the cooperative assembly of the 43S preinitiation complex (PIC), composed of the eIF2/GTP/Met-tRNAi ternary complex (TC), eIF1, eIF1A, eIF5, and eIF3 on the 40S ribosomal subunit. In canonical eukaryotic translation, the 43S PIC recruits mRNAs by engaging the eIF4F cap-binding complex to form the 48S PIC. Within the 48S PIC, eIF1, eIF1A, eIF3, and eIF4G promote the accurate scanning of the mRNA leader region and ensure the proper recognition and pairing of the start codon with Met-tRNAi.

3EIF3 is a large and structurally complex molecular assembly that, in the majority of eukaryotes, consists of 11–13 subunits (eIF3a-eIF3m) with a molecular weight of 600–800 kDa. Six of the subunits (eIF3a, eIF3c, eIF3e, eIF3k, eIF3l, and eIF3m) contain PCI (proteosome, COP9/signalosome, eIF3) modules, and two (eIF3f and eIF3h) contain MPN (Mpr1-Pad1-N-terminal) domains. The PCI⋅MPN core is structurally conserved in the 26S proteasome lid, the COP9 signalosome, and eIF3, forming a distinctive, multilobed structure. PCI modules are characterized by an N-terminal helical domain (HD) and a C-terminal winged helix domain (WHD), which mediates PCI dimerization. In the subnanometer EM structures of the proteasome lid, six WHDs oligomerize to form a horseshoe-shaped arc with their HDs radiating outward, an organization also observed in the EM reconstructions of eIF3, the 43S, and 43S⋅IRES complexes and the COP9 signalosome. S. cerevisiae and related yeasts lack six components of the PCI⋅MPN core, retaining only two PCI proteins among the six universally conserved eIF3 core subunits (eIF3a/Tif32, eIF3b/Prt1, eIF3c/Nip1, eIF3i/Tif34, eIF3g/Tif35, and eIF3j/Hcr1) (Figure 1A). The conserved core displays a modular architecture, with the interaction between eIF3b and the C-terminal portion of eIF3a connecting the PCI subunits to the eIF3g/eIF3i subcomplex and to eIF3j.

4Recent EM reconstructions as well as in vivo and in vitro experiments indicate that the N-terminal ends of the helical subdomains of the eIF3a/eIF3c PCI modules form two intermolecular bridges with the 40S subunit in the vicinity of ribosomal proteins rpS13/uS15 and rpS27/eS27 as well as rpS26/eS26, rpS1/eS1, and rpS0/uS2. In addition, footprinting and hydroxyl-radical probing experiments implicate helix 16 of the 40S subunit in eIF3 binding. Other studies have placed eIF3j, which interacts with the eIF3b-RRM (RNA recognition motif) and the eIF3a-CTD (C-terminal domain), near the decoding center of the 40S subunit. Nevertheless, information about the molecular architecture of eIF3 is incomplete due to the lack of atomic structures of eIF3 subunits and the dynamic nature of the 40S⋅eIF3 interaction. In this study, we present X-ray structures of all major structured domains of S. cerevisiae eIF3 and utilize a combination of negative-stain electron microscopy (EM), chemical crosslinking followed by mass spectrometry (CX-MS) and integrative structural modeling to derive a detailed molecular architecture of the yeast 40S⋅eIF1⋅eIF3 complex.

5As a first step in our hybrid approach (Figure 1B), we obtained a number of soluble eIF3 fragments and subcomplexes that were amenable to crystallographic analysis (Table S1 available online). The structure of the full-length PCI domain of eIF3a (S. cer. residues 1–496) was solved by single-wavelength anomalous dispersion (SAD) with selenomethionine (Se-Met)-labeled protein to a resolution of 3.3 Å (Table S2). As observed in a truncated structure of eIF3a and in previously characterized PCI modules, the WHD is structurally similar to other PCI proteins, whereas the N-terminal domain has a series of distinct, flatly arranged helical repeats (Figure 2A). The positioning of these helices is reminiscent of the concertina-like arrangement of tetratricopeptide repeats (TPR), although, as in other PCI proteins, TPR-like sequence elements are not detectable. Unlike other PCI HDs, which typically have a significant right-handed superhelical twist, eIF3a has two distinct segments to its repeat. Whereas helices 10–14 have a right-handed pitch, helices 1–9 are arranged along a flat plane that is distinct from other PCI proteins with extended helical repeats (Figure 2C).

6Using the full eIF3a PCI-domain structure as a guide, N-terminal eIF3a truncations were generated to obtain crystals of a dimeric eIF3a/eIF3c assembly, encompassing the full predicted PCI domain of eIF3c (S. cer. residues 251–812) and residues 228–496 of eIF3a. The structure of the dimeric eIF3a/eIF3c complex was solved at 3.5 Å resolution using Se-Met SAD (Figure 2B and Table S2). The superposition of the common eIF3a WHD component in these structures provides a detailed atomic model for the full eIF3a/eIF3c PCI heterodimer. As expected, the PCI domain of eIF3c shares the basic architecture of the fold (Figure 2B) but with a more elaborate helical domain that extends over 450 residues (Figure 2C). The helical domain can be subdivided into two regions: helical-element 1 (HE-1) is composed of helices 1–7, whereas helical element 2 (HE-2) contains helices 9–13. HE-1 and HE-2 form TPR-like right-handed superhelices similar to those found in other PCI proteins, with the relative orientations of HE-1 and HE-2 defined by helix 8, which disrupts the helical repeat at the HE-1/HE-2 interface, inducing an almost 90° bend in the domain (Figures 2B and 2C). An extensive, conserved protein segment formed by a 40 amino acid insertion between helices 13 and 17 meanders back into the HE-1/HE-2 interface to further stabilize the kink (Figure 2C).

7The dimerization interface between eIF3a and eIF3c involves two distinct interaction regions. Helix 23 in the WHD of eIF3a engages the WHD of eIF3c (Figures 2B and 2D) in an arrangement similar to the Thp1-Sac3 PCI-dimer. Because the WHD domains are the least well-ordered regions in our structure, the details of the interaction cannot be unambiguously established but involve a number of conserved hydrophobic side chains on eIF3a (W463, L470, I480, and I482) and eIF3c (F732, Y733, F736, L744, and F745) as well as a number of conserved polar and charged residues (Q445, E466, and H484 on eIF3a and S737 on eIF3c) (Figure 2D). A comparison between eIF3a⋅eIF3c and Thp1⋅Sac3 shows that, although the overall interaction mode is conserved between these two PCI complexes, the relative positions of the WHDs differ by up to 6 Å between these two heterodimers (Figure S1). The second interaction interface between eIF3a and eIF3c involves a loop between helices 18 and 19 of eIF3c that engages a cleft formed by helices 11, 12, and 15 of eIF3a (Figures 2B and 2D), allowing a stacking interaction between residue F664 of eIF3c and Y257 of eIF3a. This interaction induces a rearrangement of up to 3 Å at the end of eIF3a helix 15 to accommodate the eIF3c loop.

8Yeast eIF3 has a significantly reduced repertoire of eIF3 components, but the conservation of eIF3a and eIF3c in all eukaryotes is high. Docking of the eIF3a/eIF3c PCI heterodimer into the EM density of the mammalian 43S reveals an excellent fit into the two structural elements that mediate 40S binding by the PCI-MPN core of eIF3 (Figure 3A). In this placement, eIF3c engages the Zn-binding knuckle of rpS27/eS27 through a conserved pocket within HE-1, whereas eIF3a primarily contacts rpS1/eS1 (Figure 3B). The two PCI modules flank ES7, the only expansion segment within the central RNA domain that defines the 40S platform (Figure 3B).

9Many viral internal ribosome entry site (IRES)-containing RNA elements require the presence of eIF3. The recent structure of the 40S-bound CSFV IRES in complex with eIF3 revealed the structural basis of this interaction, with the IRES RNA occupying the eIF3 binding site at the 40S platform and creating a secondary binding site for the PCI⋅MPN core. Because IRES interactions are also mediated by eIF3a/eIF3c, we were able to model the interaction by fitting the eIF3a/eIF3c heterodimer into the 43S⋅IRES EM density map without any adjustments (Figure 3C). This fit of the eIF3a/eIF3c heterodimer places two loop elements (between helices 2 and 3 of eIF3a and helices 1 and 2 of eIF3c) important for IRES and 43S binding in close contact to two extrahelical bulges conserved within domain IIIB of the CSFV and HCV IRESs (Figure 3D). Consistent with our model, these RNA elements are protected from RNase digestion when bound by eIF3.

10Based on the convincing fit of the eIF3a/eIF3c structure into the EM density, we attempted to dock the recently solved paralogs of eIF3e (CSN1) and the eIF3h-eIF3f heterodimer (Rpn8/Rpn11) into the eIF3⋅IRES EM density along with older models for eIF3k, eIF3m, and eIF3l (Figure S2A). Using the previously identified positions of the various eIF3 subunits as a guide, we fitted the models to agree with features of the EM density (Figure 3E). The proposed arrangement defines the molecular interactions within the core subunits, with the HCV⋅IRES, and with the 40S subunit. Remarkably, even the helical bundle structure recently proposed for the 26S proteasome lid fits the remaining density after placement of the PCI and MPN domains (Figure 3E), suggesting that it is a conserved feature of the PCI⋅MPN core of these functionally diverse complexes. The final model also fits the lower-resolution mammalian 43S EM maps (Figure S2B).

11The eIF3b subunit plays a critical anchoring role within the 5 subunit core of eIF3. Though structural information is available for the N-terminal RRM domain and a segment of the C-terminal tail, great uncertainty remained about the central region of the protein, predicted to form either a single or a double β-propeller structure. We purified and crystallized the 57 kDa central domain of eIF3b (S. cer. residues 132–626) and solved its structure using Se-Met SAD to a resolution of 2.2 Å (Table S3). The middle domain of eIF3b forms a β propeller with nine blades, a configuration not previously observed in a native polypeptide (Figure 4A). In our crystals, two β propellers are present in the asymmetric unit, with a domain swap encompassing segments of blades 1 and 9 (Figures S3A and S3B). For clarity, our other figures depict a monomeric, unswapped eIF3b model.

12The middle domain of eIF3b begins with a short helical element, followed by the first blade of the β propeller. The A strand of this blade is provided by the most C-terminal strand in a 3+1-type velcro closure. The blades of eIF3b are fairly symmetric, with extensive loops decorating blades 3, 4, 5, 7, and 8 (Figure 4A). The central channel of eIF3b is marked by a conserved structural loop motif that constricts the “bottom” entrance of the central cavity, creating a funnel-like shape (Figure S3C). The structure and underlying sequences of the loops are conserved among the various blades of S. cerevisiae eIF3b and in the blades of eIF3b β propellers from other organisms (Figures S3C and S3E). This particular loop orientation is also seen in the six-bladed TolB β propeller, which has a similar consensus sequence (the AxSPD motif) linking strands A and B (Figure S3D). The nine-bladed eIF3b β propeller fills the gap between previously characterized eight- and ten-bladed β propellers (Figure 4B). A search for other nine-bladed β propellers in the S. cerevisiae genome revealed that the conserved sequence element and its associated nine blades are also present in eIF2A (Figure S3E), a repressor of IRES-mediated translation initiation that is downregulated during stress events in S. cerevisiae. It remains to be investigated whether these shared structural features reflect functional commonalities with eIF3b.

13The C-terminal helical domain of eIF3b mediates the association of eIF3b with the dimeric eIF3i/eIF3g complex. Crystals were obtained for a complex containing the eIF3b-CTD (S. cer. residues 655–698), full-length eIF3i, and the eIF3g-NTD (N-terminal domain) (S. cer. residues 1–135). The heterotrimeric structure was solved to a resolution of 2 Å using molecular replacement with the eIF3b-CTD/eIF3i complex as the search model (Table S3). The arrangement of eIF3i and the eIF3b fragment was nearly identical to the previously solved dimeric complex, with a root-mean-square deviation (rmsd) of 0.4 Å. Of the 135 residues in the crystallized eIF3g fragment, clear density was observed for the first 96 residues. The first 46 residues of eIF3g form a β hairpin that makes an important crystal contact but does not directly interact with eIF3i or eIF3b. Indeed, no residue of the eIF3g-NTD makes a direct contact with the eIF3b-CTD (Figure 4C). Residues 47–90 of eIF3g are responsible for the tight association with eIF3i, meandering along one-third of the outside surface of the eIF3i β propeller and making extensive contacts with blades 1, 6, and 7 of eIF3i (Figures 4C, S4A, and S4B). The molecular interaction between the eIF3g-NTD and eIF3i can be divided into three areas. The first is defined by the insertion of three eIF3g residues (R52, W55, and Y58) between blades 1 and 2 of eIF3i and involves both hydrophobic and polar interactions (Figure S4A). The second is characterized by extensive polar contacts along the outside of blades 1 and 7, including a short β strand (residues E77–V79) extending the sheet of blade 7 of eIF3i (Figure S4B). Finally, another set of exclusively hydrophobic interactions (residues L81, L83, and W87 of eIF3g) occurs between blades 6 and 7 of eIF3i (Figure S4B). The most prominent elements in the interaction surface are highly conserved among all organisms (Figures S4C and S4D), suggesting that the interaction mode between eIF3i and eIF3g is universally conserved.

14Electron microscopy studies of initiation events have been hampered by the labile nature of 40S-eIF1 interactions during cryoelectron microscopy (cryo-EM) sample preparation as described for both yeast and mammalian 40S⋅eIF1 complexes. Because eIF1 is critical for eIF3 binding to 40S, we similarly observe no consistent eIF3 occupancy in 40S samples prepared for cryo-EM, even with samples crosslinked by GraFix. In contrast, identical samples prepared by negative stain show clear density for eIF3, suggesting that this preparation method is more compatible with a functional 40S⋅eIF1⋅eIF3 assembly. The 28 Å single-particle reconstruction of the L. kluyveri 40S⋅eIF1⋅eIF3 complex (Figure 5B) reveals two large areas of extra density on the solvent-exposed side of 40S compared to complexes without eIF3 (Figure 5A). Similar to the eIF3 density in the mammalian 43S structure, one extra density is localized below the platform, matching the eIF3a/eIF3c PCI heterodimer in size and shape, whereas the other region of extra density is on the solvent-exposed side of the 40S subunit, halfway between the platform and the beak. In our hybrid structural approach, we use the EM envelope as a boundary to validate the results of our integrative structure modeling outlined below.

15We carried out CX-MS experiments to identify contact points between eIF3 subunits and 40S⋅eIF1 and to guide our placement of eIF3 subunits on 40S. Because of the propensity of S. cerevisiae eIF3 samples to form dimers due to the blade swapping within the eIF3b subunit (Figure S3A), we carried out the CX-MS experiments on eIF3 samples from the budding yeasts Lachancea kluyveri and Debaryomyces hansenii (see the Extended Experimental Procedures for details). These organisms, like S. cerevisiae, only possess the universally conserved, six-component eIF3.

16To define the position of eIF3 on 40S, we prepared and crosslinked samples of the yeast eIF3 core bound to 40S⋅eIF1. Samples were either crosslinked in high salt to prevent nonspecific ribosome binding and eIF3 aggregation or, alternatively, in low salt but with an additional sucrose gradient ultracentrifugation step to remove dimeric 40S/40S⋅eIF3 species and unbound eIF3. Samples were subjected to LC-MS/MS analysis, yielding 965 interlinks and intralinks (i.e., crosslinks between different proteins or within the same protein, respectively) (Table S4). Specifically, we were able to detect 155 interlinks between eIF3 subunits and 40S and 461 interlinks between 40S ribosomal proteins.

17The extensive number of crosslinks connecting 40S ribosomal proteins permitted a detailed analysis of the characteristics of our data set. When mapped onto the X-ray structure of the S. cerevisiae 40S subunit, 86% of these crosslinks fall within 35 Å (Figure 5C and Table S5), the maximal lysine Cα-Cα distance that our crosslinker can bridge. Within this crosslink test set, the distribution of all mapped distances versus the Id score (Figure 5D) or versus the FDR rate (Figure S5A) shows the expected correlation, as 96% of crosslinks with an Id score of >36 and 94% crosslinks with an FDR of <0.05 are satisfied (Table S5). However, this analysis also showed that a majority of crosslinks in the lower confidence score ranges are also satisfied, suggesting that useful information is lost if standard data cutoffs are applied. Another characteristic observed in our test data set is that certain unique crosslinks are detected more frequently across independent samples than others (Figure 5C and Table S4). These highly redundant crosslinks were more likely to be satisfied than those detected once or only a few times (Figure 5E), suggesting that the number of occurrences of a specific crosslink should be an important consideration when analyzing our data set.

18The observed characteristics of the 40S crosslink test set influenced our strategy for the integrative structural modeling of the 40S⋅eIF1⋅eIF3 complex (Figure S6). In addition to promoting crosslinks with high redundancy, we decided to include all crosslinks with an Id score greater than 28 (FDR <0.3) and used the information gained from the test set to split the crosslink data set into three confidence classes (high, medium, and low; Figures 5D, S5, and Table S5). In order to fully explore the contributions of high-, medium-, and low-confidence crosslinks, we opted for a Bayesian framework to allow us to qualify every piece of information based on the test set. The goal was to lessen the impact of inconsistent crosslinks while benefitting from accurate crosslinks exclusively present within the low and medium confidence ranges (Figure 6B and Experimental Procedures). The 965 crosslinks from the 40S⋅eIF1⋅eIF3 data set were used to model the positions and orientations of the six eIF3 subunits on the 40S⋅eIF1 complex, using crystallographic structures and comparative models of the 40S subunit, eIF1 protein, eIF3 domains, and eIF3 interfaces (Figures 6A, S6, and S7, and Extended Experimental Procedures). 40S, eIF1, and eIF3 domains were represented by sets of beads arranged into either rigid bodies or flexible strings (Figures 6A and S7). None of the available EM maps were integrated in the calculation. Candidate models were ranked by a scoring function (described in the Extended Experimental Procedures) that reflected how well the models satisfied crosslinking data as well as excluded volume and sequence connectivity restraints.

19We computed 90,000 structural models by sampling the positions and orientations of the rigid bodies as well as the positions of the remaining beads, guided by the scoring function. The 500 best-scoring models were grouped into two clusters, using the rmsd as a structural similarity criterion (Figure S8 and Extended Experimental Procedures). The cluster of models with the higher population and lower average score was chosen as the final solution set.

20These solutions satisfied the excluded volume of the beads and their sequence connectivity restraints. As expected from our experimental design, most of the satisfied crosslinks within our data set of 126 unique eIF3-eIF3 and 40S⋅eIF1-eIF3 crosslinking restraints were redundant and of high confidence (Figures 6B and S5 and Table S5). Importantly, the eIF3-eIF3 and 40S⋅eIF1-eIF3 data set had crosslink confidence and frequency distributions comparable to the 40S⋅eIF1 crosslink data set (Table S5 and Figure S5).

21The average pairwise rmsd of the solutions in the favored cluster is 36 Å, or 30 Å if the poorly determined eIF3g-RRM domain is omitted from the rmsd calculation. This precision allowed us to determine the positions and orientations of the eIF3 domains (Table S6). We represented the cluster of solutions by individual localization densities (Figures 6C and S8), defined as the probability of observing a specific eIF3 domain at a given point in space. Our modeling results, as described by these localization densities, place eIF3 at the back of 40S, arranged into a continuous structure that encircles the 40S⋅eIF1 complex, comprising two large modules and three linker regions (Figure 6C). These results are fully validated by the EM reconstructions, as our localization densities are in remarkable agreement with our difference density EM map (Figure 6D). In fact, the extensive overlap between the localization densities and the EM map allowed us to assign specific eIF3 domains to regions of the EM difference density (Figure 6D).

22The localizations of the eIF3b β propeller, the eIF3b-RRM, and the PCI modules of eIF3a and eIF3c have a precision of <30 Å and are the best-defined components within our modeling solutions (Table S6 and Figure 6C). The high precision of the eIF3a/eIF3c PCI heterodimer localization stems, in part, from three unique crosslinks between our structural model of the eIF3a/eIF3c PCI heterodimer and 40S (Figures 6B, 7A, and 7B). However, an extensive network of crosslinks involving the peripheral, non-PCI regions of eIF3a (S. cer. residues 496–964) and eIF3c (S. cer. residues 1–251) provide critical additional restraints (Figure 6B). These linking elements connect the PCI modules with the eIF3b/eIF3i/eIF3g module and with eIF1, eIF2, and eIF5, respectively. Because no structural information is available for these regions, they are represented by beads comprising 20 residue segments in our modeling (Figure S7). Despite this coarse representation, eIF3c-NTD and eIF3a-CTD form tightly clustered interaction networks containing both inter- and intralinks that provide additional restraints on either side of the eIF3a/eIF3c PCI heterodimer-binding site (Figures 7A and 7B). The localization densities for the linker regions match regions of weak difference density in our EM reconstructions (Figures 6D, 7A, and 7B), further suggesting that portions of the eIF3c-NTD and eIF3a-CTD form small ordered regions. In particular, the extensive crosslink network involving the first part of the eIF3a-CTD and the C-terminal helix of eIF3c hints at the formation of a structure analogous to the helical bundle present in the larger PCI⋅MPN core (Figures 7B, 7E, and 7F). Biochemical evidence suggests that the second half of the eIF3a-CTD extends near the entrance of the mRNA channel, interacting with h16/h18, rpS2/uS5, rpS3/uS3, and the eIF3b-RRM. Although we do not observe any specific crosslinks that support this orientation, an extended orientation for this region of eIF3a is consistent with our overall model.

23In our solutions, the eIF3b β propeller localizes between RNA helices ES6A and h16 and above rpS9/uS4, rpS4/eS4, and rpS24/eS24 with a high precision of 15 Å (Table S6). The placement is guided by ten unique crosslinks between four eIF3b residues and seven rpS9 residues (Figures 6B and 7C) and is consistent with previous RNase protection experiments that mapped an interaction between mammalian eIF3 and h16. To obtain a representative set of structures from our solutions, we further clustered the eIF3b β propeller conformations on the 40S subunit using an rmsd cutoff of 3.5 Å (Extended Experimental Procedures). This analysis allowed us to select ten structures from our solutions that represent the most frequent positions of the eIF3b β propeller. The overlay of the selected eIF3b β-propeller structures reveals a tight cluster of structures that share a common interaction surface with 40S (Figure 7C). The placements of the eIF3b-CTD/eIF3i/eIF3g-NTD and of the eIF3b-RRM are less precise, relying exclusively on crosslinks and sequence connectivity constraints to the eIF3b β-propeller domain, as there are no satisfied crosslinks with 40S (Figure 6B). Indeed, the position of the eIF3b-CTD/eIF3i/eIF3g-NTD complex is the feature that defines the two clusters within our 500 best-scoring solutions. The position of our best-solution set relies heavily on a redundant, medium-confidence crosslink between eIF3i and the eIF3b β propeller, whereas the position in the less-populated cluster is dominated by connectivity restraints that are less constrained in this orientation, leading to the lower precision observed for this cluster (Figure S8 and Table S6). In our final solution set, the position of the eIF3i β propeller within the eIF3b-CTD/eIF3i/eIF3g-NTD subcomplex is consistently oriented orthogonally to the eIF3b β propeller (Figure 6C), with the “bottom” face of the eIF3i β propeller facing the “top” surface of the eIF3b β propeller (Figures 7E and 7F). This arrangement places the C-terminal helices of eIF3b in an elongated conformation along the “bottom” and “top” surfaces of the eIF3b and eIF3i β propellers, respectively, and the eIF3g-NTD is localized between eIF3i and 40S near the interface of the β propellers, consistent with evidence that eIF3g stabilizes the eIF3b/eIF3i interaction and interacts with ribosomal proteins rpS3/uS3 and rpS20/uS10. At the N-terminal end of eIF3b, the localization density of the eIF3b-RRM is adjacent to the eIF3b β propeller near the mRNA entrance channel, linking eIF3b to eIF3j, which our solutions place in its expected position near the mRNA channel entrance.

24Consistent with our localization densities, the EM envelope of the mammalian 43S complex, despite the presence of the DHX29 helicase, shows features that agree with our model. Of the density elements assigned to eIF3, a toroid-shaped element between the platform and the beak of 40S closely matches our position for the eIF3b β propeller (Figure 7D). Previously, this density was tentatively attributed to eIF3i. Furthermore, the localization densities of the eIF3b-CTD/eIF3i/eIF3g-NTD complex and the eIF3b-RRM match regions of weak additional density in the mammalian 43S structure (Figure 7D), supporting the localization and orientation of the eIF3b/eIF3i/eIF3g module in our solution set. This extensive agreement between the EM densities and our modeling solutions gives us confidence that our models for the yeast and mammalian 40S⋅eIF1⋅eIF3 complexes reveal the regions of the 40S subunit involved in eIF3 interactions as well as the positions and relative orientations of nearly all eIF3 domains and the likely paths of all major linker domains (Figures 7E and 7F).

25The eIF3 complex plays a prominent role in regulating the intricate molecular events of translation initiation. Its large size and modular structure have made molecular studies of this assembly challenging, and a comprehensive molecular understanding of the 40S⋅eIF3 interaction has remained elusive.

26Our hybrid approach allowed for a large number of low-resolution restraints to be effectively integrated and optimized, resulting in a sophisticated and surprisingly detailed interaction map of eIF3 on the ribosome (Figures 7E and 7F). This arrangement enables eIF3 components to completely encircle the 40S subunit and to engage and coordinate the binding of other components of the translation initiation machinery at both ends of the mRNA channel. EIF3 can therefore act as a scaffold, ensuring that factors required at both the mRNA channel entry and exit are assembled simultaneously, allowing eIF3 to act as the “orchestra conductor” during initiation, coordinating the input from a number of “sections” to ensure that mRNA recruiting and scanning at one end occurs only when elements critical for AUG recognition and scanning modulation are assembled at the opposite end. Interestingly, because eIF1 and eIF3j have high 40S-binding affinities, interactions at the ends of the two eIF3 “arms” may enable the cooperative assembly of the full 43S PIC.

27This model is consistent with the vast majority of biochemical and genetic data that has accumulated over the years. Mutations and disruptions in factors mapped to the mRNA channel entrance result in a leaky scanning phenotype and defects in rescanning downstream of short uORF sequences, suggesting that eIF3b, eIF3g, eIF3i, and eIF3j prevent AUG codon bypass and modulate rescanning downstream of uORFs by engaging the mRNA in the vicinity of the channel entrance. At the other end of the eIF3 “clamp,” the eIF3c-NTD extends across the subunit interface, connecting the PCI⋅MPN core of eIF3 with the mRNA channel and making critical contacts with eIF1, eIF5, and TC and, in mammals, with eIF4 to create an assembly that is highly proficient at codon differentiation during 5′ UTR scanning. In contrast to mutations in the eIF3b/eIF3i/eIF3g/eIF3j module, mutations in the N-terminal tail of eIF3c result in defects in TC recruitment and scanning fidelity, suggesting distinct roles for the two “arms” of eIF3 that extend into the mRNA channel.

28The interactions mediated by the six universally conserved eIF3 subunits probably represent the core functions of eIF3 during translation initiation. The remaining six PCI⋅MPN eIF3 components not found in budding yeasts are likely to act as “gatekeepers,” engaging elements within mRNPs (including the cap-binding complex) to regulate their access to the ribosome. In S. cerevisiae, the loss of eIF3 PCI⋅MPN core components correlates with the simplified 5′ UTR structures of most transcripts and the low intron density found in these organisms. Our detailed molecular description of the interaction between eIF3 and the 40S ribosomal subunit maps individual functionalities of eIF3 to distinct regions of the 43S PIC and provides an intricate structural framework to integrate biochemical, biophysical, and genetic studies of eukaryotic translation initiation.

29A full list of constructs used in this study is given in Table S1. The purification protocols for eIF1, various eIF3 subunits and complexes, and the 40S subunit are described in Extended Experimental Procedures.

30All proteins (15–20 mg/ml) were crystallized by sitting-drop vapor diffusion. Native and Se-Met SAD data were collected at beamline PX of the Swiss Light Source and processed in XDS. Structures were solved using AutoSol or Phaser MR in Phenix and models built using Phenix AutoBuild and COOT. Model refinement was performed with Phenix. More details are provided in the Extended Experimental Procedures and Tables S2 and S3.

31EIF3 samples were assembled from purified, recombinant subunits and were subsequently mixed with 40S subunits as described in the Extended Experimental Procedures. Final 40S⋅eIF1⋅eIF3 samples were assembled under various buffer conditions and crosslinked. Some samples were purified on sucrose gradients before RNase treatment, enzymatic digestion, and enrichment of crosslinked peptides was carried out. LC-MS/MS analysis was performed on an Orbitrap Elite mass spectrometer, and the data were searched using the xQuest/xProphet software pipeline. Further details are provided in the Extended Experimental Procedures.

32The Lachancea kluyveri 40S⋅eIF1⋅eIF3 complex was prepared and crosslinked as described for CX-MS. The sample was then subjected to gradient fixation using GraFix, grids were prepared by negative staining using uranyl acetate, and images were acquired in a Tecnai F20 electron microscope under low-dose conditions. Single-particle images were split into initiation factor occupied and unoccupied ribosomal complexes based on an initial reconstruction of the whole data set using supervised classification (Imagic Spider RRR). The resolution of the final structure of the initiation factor occupied and unoccupied ribosomal complex was estimated as 28.1 and 24.6 Å, respectively (Fourier shell = 0.5 criterion, semi-independent data half sets).

33The integrative modeling approach proceeded through four steps, as described in Extended Experimental Procedures. To benefit from the entire crosslink data set, the Bayesian scoring function separately modeled the uncertainty of high-, medium-, and low-confidence crosslinks (as defined by Id score and FDR), thereby modulating their relative weights. The 275 eIF3-eIF3 and 40S⋅eIF1-eIF3 crosslinks restrained the positions and orientations of eIF3 domains relative to 40S⋅eIF1, whereas the remaining 690 40S⋅eIF1 crosslinks were included as a training set to increase the accuracy of the uncertainty estimation. Because the frequency of crosslink observation across eight independent experiments correlates with their accuracy in our test set (Figure 5E), the unique crosslinks were weighted proportionally to their frequency. An extensive description of the methodology is provided in the Extended Experimental Procedures, and the modeling scripts and models are available at http://salilab.org/40S-eIF1-eIF3.

34Proteins were cloned into pET-derived vectors using ligation-independent-cloning and expressed in E. coli BL21-DE3 pRIL cells in 2xYT media overnight at 18°C. Cells were lysed in 50 mM HEPES pH 7.5, 800 mM KCl, 40 mM imidazole, 0.5 mM Tris-(2-Carboxyethyl) phosphine hydrochloride (TCEP) and 10% Glycerol. S. cer. or L. klu. His-MBP-tagged eIF3c and His-tagged eIF3a were independently purified over Ni-affinity columns, then combined and purified over a Heparin column. D. han. His-eIF3b/eIF3i/eIF3g was expressed from a polycistronic construct and purified over Ni-Sepharose and Heparin columns. Bacterial cells expressing S. cer. or L. klu. eIF3i were mixed with cells expressing S. cer. or L. klu His-MBP-tagged eIF3g before lysis and purification over Ni-Sepharose and HiTrap Q. L. klu. His-eIF3b, His-MBP-eIF3j and His-eIF1 were expressed separately and purified over Ni-Sepharose and Heparin columns. Samples for crystallography were assembled and digested with TEV protease overnight to remove all tags before final purification over a Superdex 200 gel-filtration column. The 6-subunit eIF3 core complex was assembled from purified L. klu. eIF3a/eIF3c, eIF3b, eIF3i/eIF3g (or eIF3b/eIF3i/eIF3g in the case of D. han.) and eIF3j before overnight TEV cleavage to remove all tags. After orthogonal purification to remove uncleaved proteins, tags and TEV protease, the sample was applied onto a Superdex 200 column to yield the full eIF3 complex. Se-Met-derivatized proteins were expressed in minimal media supplemented with an amino acid cocktail including Se-Met.40S ribosomal subunits were purified from Lachancea kluyveri (DSMZ strain 70517). Cells were grown to an OD of 10 in a 50-l fermenter (INFORS) at 24°C, harvested by centrifugation, washed with 1x PBS, resuspended in lysis buffer and snap frozen in liquid nitrogen. Yeast cells were lysed using a cell cracker at 40 kpsi, the lysate cleared by centrifugation (Sorvall SS-34) and then applied onto a high-density sucrose cushion (60% w/v) in low salt for a 24 hr ultracentrifugation at 50k rpm in a TI-70 rotor. Pelleted ribosomal particles were resuspended in high-salt buffer and the subunits separated on 10%–47% w/v sucrose gradients by ultracentrifugation. 40S bands were carefully removed and pooled, then buffer exchanged into 50 mM HEPES pH 7.5, 10 mM MgCl2, 75 mM KCl and 0.5 mM TCEP.Because of the propensity of S. cer. eIF3b to form dimers, we focused on two related budding yeasts for our EM and CX-MS experiments. The full 6-component eIF3 complex, lacking only the disordered C terminus of eIF3a (equivalent to S. cer. residues 870-964) was assembled from individual components. While we were able to assemble the full eIF3 complex from Lachancea kluyveri, we were unable to purify eIF3a and eIF3c from Debaromyces hansenii. Because our polycistronic construct for D. han. eIF3b/eIF3i/eIF3g expressed large quantities of the trimeric complex, and because of the high degree of conservation among these related budding yeasts, we decided to test whether we could obtain a hybrid complex of L. klu. and D. han. to increase the variability of the samples for CX-MS. Hybrid eIF3 complexes behaved similarly to the nonhybrid L. klu. complex and showed the same areas of extra density in negative-stain single particle reconstructions (data not shown). 40S⋅eIF1⋅eIF3 (10 μM) was assembled with a 1.2x molar excess of initiation factors in 25 mM HEPES, 15 mM MgCl2, 125 mM KCl and 2 mM TCEP prior to crosslinking. For each round of experiments, several different salt concentrations (40 mM to 175 mM KCl) as well as sample concentrations (2 μM to 10 μM) were explored to find optimal conditions for stable complex formation prior to crosslinking.All proteins were exchanged into crystallization buffer (20 mM HEPES, 75 mM KCl and 0.5 mM TCEP) prior to crystallization. The eIF3a-PCI domain (18 mg/ml) was crystallized by sitting-drop vapor diffusion against a solution of 14%–18% PEG 3350, 400-600 mM KSCN in a broad pH range (6.8-8.8) using bis-tris-propane, HEPES or bicine as buffers. Crystals belonging to space group C2 grew over 2-5 days and were cryo-stabilized with the addition of 25% ethylene glycol. The eIF3a/eIF3c dimer (20 mg/ml) was crystallized by sitting-drop vapor diffusion against a solution of bis-tris-propane pH 6.0-6.3, 9%–11% PEG 6000 and 400-500 mM KSCN. Crystals belonging to space group P3221 appeared overnight and achieved maximum size after 3 days. Crystals were stabilized and dehydrated over a 2 hr period by the addition of 25% ethylene glycol and 12% PEG 20K over 4 equal concentration steps. This dehydration treatment was essential for diffraction beyond 4.5Å and to promote isomorphicity along the crystallographic c-axis between native and Se-Met crystals. The eIF3b β propeller (20 mg/ml) was crystallized in space group P21 after 2-5 days in 10%–12% PEG MME 5k, 100-200 mM KSCN or KH2PO4 and pH 8 – 9.5 using Tris-Acetate or CHES as buffers. Crystals were cryo-stabilized with the addition of 25% Glycerol and flash frozen in liquid nitrogen. The trimeric eIF3b-CTD/eIF3i/eIF3g-NTD complex (15 mg/ml) was crystallized in 5%–7% PEG 20K, 25%–28% PEG MME 550 at pH 6.5 with Imidazole or MES buffers and belonged to space group C2221. The structures of eIF3a, eIF3a/eIF3c and eIF3b were solved by single-wavelength anomalous dispersion techniques using Se-Met substituted proteins. eIF3b-CTD/eIF3i/eIF3g-NTD was solved using the yeast eIF3b/eIF3i structure as a molecular replacement model.The eIF3a, eIF3b β propeller and eIF3b-CTD/eIF3i/eIF3g-NTD structures were solved and built using standard methodologies as described in Experimental Procedures. The structure of the eIF3a/eIF3c heterodimer presented more of a challenge due to its lower resolution, prevalence of split spots and unusually high overall thermal factors (Wilson B = 170 Å2; Average refined B factor = 175 Å2). Initial models were built into the 3.7 Å experimental SAD maps obtained from AutoSol, as they were more informative than the phase-extended 3.5 Å maps. After extensive rounds of manual rebuilding in Coot and refinement in Phenix, eventually incorporating the higher-resolution data, a high-confidence model was built, although the presence of split lattices and the high B-factors prevented further improvement of the R-factors.H. sapiens eIF3 subunits were threaded onto the model templates (Figure S2) using Phyre2. Models are represented as poly-alanine but labeled according to the predicted human residue number, except for the helical bundle, which is entirely poly-alanine. The model was fitted into the 43S⋅IRES EM map, using automated fitting with manual adjustments in Chimera.Throughout the paper, we use S. cer. residue numbering for all data. All crosslink data are therefore mapped onto the S. cer. 40S and eIF3 sequences based on multiple sequence alignments. S. cer. ribosomal proteins are identified by their historical names as well as the newly proposed naming system.EIF3 samples from L. klu. or L. klu. and D. han. were assembled from purified, recombinant subunits as described above. 60 μl of purified 40S⋅eIF1⋅eIF3 sample was cross-linked with 1 mM disuccinimidyl suberate d0/d12 (DSS, Creative Molecules Inc.) at room temperature (RT) for 30 min and quenched for 10 min at RT with ammonium bicarbonate (50 mM final concentration). Samples assembled in high salt, where no significant ribosome oligomerization occurred, were processed immediately, while low salt samples, which showed a significant level of ribosome oligomerization, were applied to 10%–47% sucrose gradients to separate monomeric 40S fractions from oligomeric assemblies and unbound eIF3 before further processing. There are strength and weaknesses in each of these approaches, as the extra purification steps lower the final peptide yields, while the nonpurified samples retain a background of eIF3 interlinks from the unbound eIF3 fraction of the sample. To ensure a large and unbiased set of identified crosslinked peptides, we included crosslinks from both data types in the final data set, although we did exclude eIF3⋅eIF3 interlinks from our nonpurified samples, as there is no unbiased way to determine whether the crosslink was generated when eIF3 was bound to 40S or free in solution.After 1 hr treatment with 10 μg/ml RNase A (Roche), the samples were reduced with 2.5 mM TCEP in 8 M urea at 37°C for 30 min and alkylated with 5 mM iodoacetamide (Sigma-Aldrich) for 30 min at RT in the dark. Before digestion, the samples were diluted with ammonium bicarbonate to a 1 M final concentration of urea. 2% w/w trypsin (Promega) was added overnight at 37°C and inactivated by acidification with 2% (w/v) trifluoroacetic acid (TFA). Peptides were purified with Sep-Pak C18 MicroSpin columns (Waters, Milford, MA), according to the manufacturer’s protocol, followed by enrichment of crosslinked peptides using size exclusion chromatography. LC-MS/MS analysis was carried out on an Orbitrap Elite mass spectrometer (Thermo Electron, San Jose, CA) and samples were shot in duplicates. Data were searched using xQuest in iontag mode against a database containing the protein sequences of all eIF3 subunits, eIF1 and all 40S ribosomal proteins with a precursor mass tolerance of 10 ppm. For matching of fragment ions tolerances of 0.2 Da for common-ions and 0.3 Da for crosslink ions were used and only peptides with a minimal length of 4 amino acids were considered. False discovery rates (FDR) of crosslinked peptides were assigned using xProphet. Crosslinked peptides were identified with a delta score of < 0.95 and a linear discriminant (ld) score of > 28 as described in the text. Mass spectra of the most abundant peaks were additionally analyzed by visual inspection in order to ensure good matches of ion series for both cross-linked peptides.The integrative approach to determine the S. cerevisiae 40S⋅eIF1⋅eIF3 complex architecture proceeds through four stages: (1) gathering of data, (2) representation of subunits and translation of the crosslinking data and the prior knowledge into a Bayesian scoring function, (3) configurational sampling to produce an ensemble of models that minimize the Bayesian scoring function, and (4) analysis of the ensemble (Figure S6). All residue ranges described below refer to the S. cerevisiae sequence.We used the crystallographic structures of the 40S ribosome subunit (PDB 3U5F and 3U5G) as well as the eIF3 subunit domains described in the main text (the eIF3a/eIF3c PCI heterodimer, the eIF3b β propeller, and the eIF3b-CTD/eIF3i/eIF3g-NTD complex). We removed chain h (the Stm1 protein) from the 40S structure, and rigid-body adjusted the position of helix ES6A to better fit the 43S EM density. We increased the structural coverage of the complex by adding comparative models for the eIF3b-RRM (residues 37-626, human homolog PDB 2KRB) in complex with a short eIF3j fragment (residues 30-41), eIF3g-RRM (residues 183-282, human homolog, PDB 2CQ0), eIF3j-CTD helical domain (residues 146-210, human homolog, PDB 3BPJ), and eIF1 (human homolog, PDB 2IF1). Comparative models were built with MODELER 9.12. The S. cerevisiae eIF1⋅40S interface was modeled using the Tetrahymena thermophila 40S⋅eIF1 complex structure. Sequence segments without a template structure included eIF3a-CTD (residues 496-964), eIF3c-NTD (residues 1-251), eIF3g linker (residues 97-183), eIF3j disordered regions (41-145) as well as other missing loops or terminal domains (Figure S7). The 965 crosslink observations (Table S4) were identified by mass spectrometry from eight independent experiments. We used the crosslink Id-Score and FDR to define three crosslink confidence classes, allowing us to construct a noise model for the crosslinking data (Figures 5D, S5 and Table S5).The domains of the 40S⋅eIF1⋅eIF3 subunits were represented by beads, arranged into either a rigid body or a flexible string, based on the available crystallographic structures and comparative models (Figure S7). To balance the thoroughness of configurational sampling and precision of model representation, we represented the structures in a multi-scale fashion. For the crosslink and excluded volume scoring terms, the crystallographic structures and comparative models were coarse-grained by representing each consecutive segments of 1 and 10 residues by a bead, centered on the center of mass, respectively (Figure S7). Sequence segments missing in the crystal structures were substituted by a single or multiple beads of the corresponding size (Figure S7). For each domain and interface with an atomic model, the beads representing a structured region were kept rigid with respect to each other during configurational sampling (i.e., rigid bodies). The rigid bodies include: the 40S⋅eIF1 complex, the eIF3a/eIF3c PCI heterodimer, eIF3b(655-699)/eIF3i/eIF3g(43-86) complex, eIF3b-RRM/eIF3j(30-41), eIF3g-RRM, eIF3b short helix (631-655), eIF3g(8-38) N-terminal double-stranded β sheet, eIF3g(86-97) short helix, and the C-terminal helices eIF3j(146-167), eIF3j(168-187), and eIF3j(187-210). Segments without a crystallographic structure or comparative model (i.e., segments predicted to be disordered or structured without a known homolog) were represented by a flexible string of large beads corresponding to a maximum of 30 residues each (Figure S7).The Bayesian approach estimates the probability of a model, given information available about the system, including both prior knowledge and newly acquired experimental data. The model M ≡ (X, {αi}) includes the structure coordinates X and additional parameters {αi} (below). Using Bayes’ theorem, the posterior probability p(M|D,I), given data D and prior knowledge I, iswhere the likelihood function p(D|M,I) is the probability of observing data D, given I and M; and the prior is the probability of model M, given I. To define the likelihood function, one needs a forward model that predicts the data point (i.e., the presence of a crosslink between two given residues) given the structure, and a noise model that specifies the distribution of the deviation between the observed and predicted data points. The Bayesian scoring function is the negative logarithm of p(D|M,I)p(M|I), which ranks the models identically to the posterior probability. All information gathered in Stage 1 is encoded into the Bayesian scoring function as a likelihood function (crosslinking data), as well as priors (sequence connectivity restraints to enforce proximity between beads representing consecutive sequence, and the excluded volume between all pairs of beads).For a given experimentally identified crosslink dn between two lysines i and j, the forward model predicts the presence of the crosslink given the structural coordinates X. Because of the multi-scale representation, residues are mapped into coarse-grained beads with variable radius, which represents the uncertainty of their positions. Furthermore, the rigid-body representation of crystallographic structures does not account for the flexibility of residues placed in loops. Therefore, we introduced the parameters {σk}, one for each crosslinked lysine, that define the unknown radii of the spheres, centered on the beads representing the lysines, within which the position of the lysine Cα is uniformly distributed. Given ri and rj (the coordinates of the Cα of the two lysines) and the maximum length for the crosslinker lXL, including the lysine side chains, we defined the forward model fn as the probability of randomly picking two points and within the spheres centered on ri and rj, with radii σi and σj, such that the distance between them is lower than lXL. The forward model value is one (ie, the identified crosslink is present in the given structure) when the spheres are closer than lXL and the radii are sufficiently small. It has a value of zero (i.e., the identified crosslink is not present in the given structure) if the distance between the spheres exceeds lXL. To reduce the number of parameters in the model, we utilized a single structural uncertainty σ, whose value estimates the average of {σk} for all crosslinked residues. We estimated lXL = 21 Å from potential of mean force calculations of a system comprising of the DSS crosslinker covalently attached to the two lysine sidechains (Yannick Spill and Michael Nilges, personal communication).The likelihood function for an identified crosslink dn is defined by the following data noise model:where ψn ∈[0,0.5] is the uncertainty associated with the crosslink identification. When the value of ψn is 0.5, the uncertainty is maximal and the likelihood does not depend on the forward model fn or on the distance between the lysines. When ψn is 0.0, the uncertainty is minimal, and the likelihood is identical to the forward model, i.e., it reaches a maximum when the distance between the spheres rij is below the crosslink length lXL. The joint likelihood function p(D|M,I) for a data set D = {dn} of NXL of independently observed crosslinks is the product of likelihood functions for each data point. Crosslinks observed multiple times are considered as independent observations, and therefore considered multiple times in the joint likelihood function. Analysis of the intra-40S crosslink distances (Figure 5C) allowed us to group the crosslinks into three classes, including a class with an estimated ∼5% false-positive rate (high confidence crosslinks with Id-Score > 36.0; FDR < 0.05), a class with a ∼10% false-positive rate (medium confidence with 32.0 < Id Score < 36.0; 0.05 < FDR < 0.1) and a class with ∼30% false-positive rate (low confidence with 28.0 < Id Score < 32.0; 0.1 < FDR < 0.3) (Figures 5D, S5 and Table S5). We reduced the number of ψn parameters, by defining uncertainties ψH (358 crosslinks), ψM (263 crosslinks) and ψL (344 crosslinks) for the three classes.The model prior p(M|I) is defined as a product of the individual priors p(X), p(σ), p(ψH), p(ψM), and p(ψL) on the structural coordinates, structural uncertainties, and the three data point uncertainties, respectively. The prior p(X) included the excluded volume and terms to maintain sequence connectivity. The excluded volume restraint was implemented as a pairwise hard-sphere repulsive potential, where the bead radii were determined using the statistical relationship between the volume and the number of residues. The sequence connectivity terms are upper harmonic distance restraints, with a threshold distance set to be four times the sum of the radii of the two connected beads. To facilitate configurational sampling, we applied a weak restraint whose score depended linearly on the distance between cross-linked residues, with a slope of 0.01 Å−1. p(σ), p(ψH), p(ψM), and p(ψL) are uniform distributions, that bound σ to interval [0,100], and the ψ parameters to the interval [0,5].Structural models of the 40S⋅eIF1⋅eIF3 complex were computed by Replica Exchange Gibbs sampling, based on Metropolis Monte Carlo sampling. This sampling was used to generate a collection of configurations for the system as well as values for the uncertainty parameters. The Monte Carlo movements included random translation and rotation of rigid bodies (1 Å and 0.025 radian maximum, respectively), random translation of individual beads in the flexible segments (1 Å maximum), as well as Gaussian perturbation of the uncertainty parameters. The sampling was run on 64 replicas, with temperatures ranging between 1.0 and 2.5. Each replica produced between 10,000 and 30,000 models. 12 independent sampling calculations were run, each one starting with a random initial configuration, for a total of 180,000 models. This set of models was divided into two ensembles of the same size, to assess sampling convergence (Figure S8). Furthermore, 6 additional runs were carried to assess the robustness of the modeling approach by removing approximately 10% and 20% of the crosslinks, respectively.For each ensemble, the solutions (ie, the 500 best scoring models) were grouped using k-means clustering using the root-mean-square deviation (rmsd) of all eIF3 subunits as a structural similarity criterion when the models were superposed on the 40S subunit structure. The rmsd distance matrix identified two dominant clusters of similar structures (Figure S8A), here named cluster 1 and 2, with cluster 1 being more populated and scoring better than cluster 2 (Figure S8). The overall precision (calculated as the average pairwise rmsd) for cluster 1 and 2 was ∼36 and ∼41 Å, respectively, for both ensembles. The superposed structures in the cluster were converted into the probability of any volume element being occupied by a given protein (that is, the ‘localization density’) (Figure 6C). While the eIF3b/eIF3a/eIF3c/eIF3j subunits localize on the 40S surface similarly in the two clusters, the eIF3i subunit displayed two alternate positions (Figure S8B and S8C): one toward the 40S head (cluster 1) and one toward the bottom of the body (cluster 2). Localization precision calculated for each individual domain (Table S6) indicated that the eIF3b β propeller was the domain with the most precisely determined position (∼15 Å), followed by the eIF3a/c PCI and the eIF3b-RRM. eIF3j and the eIF3b helical linker positions were determined less precisely, and eIF3i localized with different precision within the two clusters (∼38 Å in cluster 1 and ∼58 Å in cluster 2). eIF3g-RRM was the least well determined component, with a localization precision of ∼90 Å.

35Proteins were cloned into pET-derived vectors using ligation-independent-cloning and expressed in E. coli BL21-DE3 pRIL cells in 2xYT media overnight at 18°C. Cells were lysed in 50 mM HEPES pH 7.5, 800 mM KCl, 40 mM imidazole, 0.5 mM Tris-(2-Carboxyethyl) phosphine hydrochloride (TCEP) and 10% Glycerol. S. cer. or L. klu. His-MBP-tagged eIF3c and His-tagged eIF3a were independently purified over Ni-affinity columns, then combined and purified over a Heparin column. D. han. His-eIF3b/eIF3i/eIF3g was expressed from a polycistronic construct and purified over Ni-Sepharose and Heparin columns. Bacterial cells expressing S. cer. or L. klu. eIF3i were mixed with cells expressing S. cer. or L. klu His-MBP-tagged eIF3g before lysis and purification over Ni-Sepharose and HiTrap Q. L. klu. His-eIF3b, His-MBP-eIF3j and His-eIF1 were expressed separately and purified over Ni-Sepharose and Heparin columns. Samples for crystallography were assembled and digested with TEV protease overnight to remove all tags before final purification over a Superdex 200 gel-filtration column. The 6-subunit eIF3 core complex was assembled from purified L. klu. eIF3a/eIF3c, eIF3b, eIF3i/eIF3g (or eIF3b/eIF3i/eIF3g in the case of D. han.) and eIF3j before overnight TEV cleavage to remove all tags. After orthogonal purification to remove uncleaved proteins, tags and TEV protease, the sample was applied onto a Superdex 200 column to yield the full eIF3 complex. Se-Met-derivatized proteins were expressed in minimal media supplemented with an amino acid cocktail including Se-Met.

3640S ribosomal subunits were purified from Lachancea kluyveri (DSMZ strain 70517). Cells were grown to an OD of 10 in a 50-l fermenter (INFORS) at 24°C, harvested by centrifugation, washed with 1x PBS, resuspended in lysis buffer and snap frozen in liquid nitrogen. Yeast cells were lysed using a cell cracker at 40 kpsi, the lysate cleared by centrifugation (Sorvall SS-34) and then applied onto a high-density sucrose cushion (60% w/v) in low salt for a 24 hr ultracentrifugation at 50k rpm in a TI-70 rotor. Pelleted ribosomal particles were resuspended in high-salt buffer and the subunits separated on 10%–47% w/v sucrose gradients by ultracentrifugation. 40S bands were carefully removed and pooled, then buffer exchanged into 50 mM HEPES pH 7.5, 10 mM MgCl2, 75 mM KCl and 0.5 mM TCEP.

37Because of the propensity of S. cer. eIF3b to form dimers, we focused on two related budding yeasts for our EM and CX-MS experiments. The full 6-component eIF3 complex, lacking only the disordered C terminus of eIF3a (equivalent to S. cer. residues 870-964) was assembled from individual components. While we were able to assemble the full eIF3 complex from Lachancea kluyveri, we were unable to purify eIF3a and eIF3c from Debaromyces hansenii. Because our polycistronic construct for D. han. eIF3b/eIF3i/eIF3g expressed large quantities of the trimeric complex, and because of the high degree of conservation among these related budding yeasts, we decided to test whether we could obtain a hybrid complex of L. klu. and D. han. to increase the variability of the samples for CX-MS. Hybrid eIF3 complexes behaved similarly to the nonhybrid L. klu. complex and showed the same areas of extra density in negative-stain single particle reconstructions (data not shown). 40S⋅eIF1⋅eIF3 (10 μM) was assembled with a 1.2x molar excess of initiation factors in 25 mM HEPES, 15 mM MgCl2, 125 mM KCl and 2 mM TCEP prior to crosslinking. For each round of experiments, several different salt concentrations (40 mM to 175 mM KCl) as well as sample concentrations (2 μM to 10 μM) were explored to find optimal conditions for stable complex formation prior to crosslinking.

38All proteins were exchanged into crystallization buffer (20 mM HEPES, 75 mM KCl and 0.5 mM TCEP) prior to crystallization. The eIF3a-PCI domain (18 mg/ml) was crystallized by sitting-drop vapor diffusion against a solution of 14%–18% PEG 3350, 400-600 mM KSCN in a broad pH range (6.8-8.8) using bis-tris-propane, HEPES or bicine as buffers. Crystals belonging to space group C2 grew over 2-5 days and were cryo-stabilized with the addition of 25% ethylene glycol. The eIF3a/eIF3c dimer (20 mg/ml) was crystallized by sitting-drop vapor diffusion against a solution of bis-tris-propane pH 6.0-6.3, 9%–11% PEG 6000 and 400-500 mM KSCN. Crystals belonging to space group P3221 appeared overnight and achieved maximum size after 3 days. Crystals were stabilized and dehydrated over a 2 hr period by the addition of 25% ethylene glycol and 12% PEG 20K over 4 equal concentration steps. This dehydration treatment was essential for diffraction beyond 4.5Å and to promote isomorphicity along the crystallographic c-axis between native and Se-Met crystals. The eIF3b β propeller (20 mg/ml) was crystallized in space group P21 after 2-5 days in 10%–12% PEG MME 5k, 100-200 mM KSCN or KH2PO4 and pH 8 – 9.5 using Tris-Acetate or CHES as buffers. Crystals were cryo-stabilized with the addition of 25% Glycerol and flash frozen in liquid nitrogen. The trimeric eIF3b-CTD/eIF3i/eIF3g-NTD complex (15 mg/ml) was crystallized in 5%–7% PEG 20K, 25%–28% PEG MME 550 at pH 6.5 with Imidazole or MES buffers and belonged to space group C2221. The structures of eIF3a, eIF3a/eIF3c and eIF3b were solved by single-wavelength anomalous dispersion techniques using Se-Met substituted proteins. eIF3b-CTD/eIF3i/eIF3g-NTD was solved using the yeast eIF3b/eIF3i structure as a molecular replacement model.

39The eIF3a, eIF3b β propeller and eIF3b-CTD/eIF3i/eIF3g-NTD structures were solved and built using standard methodologies as described in Experimental Procedures. The structure of the eIF3a/eIF3c heterodimer presented more of a challenge due to its lower resolution, prevalence of split spots and unusually high overall thermal factors (Wilson B = 170 Å2; Average refined B factor = 175 Å2). Initial models were built into the 3.7 Å experimental SAD maps obtained from AutoSol, as they were more informative than the phase-extended 3.5 Å maps. After extensive rounds of manual rebuilding in Coot and refinement in Phenix, eventually incorporating the higher-resolution data, a high-confidence model was built, although the presence of split lattices and the high B-factors prevented further improvement of the R-factors.

40H. sapiens eIF3 subunits were threaded onto the model templates (Figure S2) using Phyre2. Models are represented as poly-alanine but labeled according to the predicted human residue number, except for the helical bundle, which is entirely poly-alanine. The model was fitted into the 43S⋅IRES EM map, using automated fitting with manual adjustments in Chimera.

41Throughout the paper, we use S. cer. residue numbering for all data. All crosslink data are therefore mapped onto the S. cer. 40S and eIF3 sequences based on multiple sequence alignments. S. cer. ribosomal proteins are identified by their historical names as well as the newly proposed naming system.

42EIF3 samples from L. klu. or L. klu. and D. han. were assembled from purified, recombinant subunits as described above. 60 μl of purified 40S⋅eIF1⋅eIF3 sample was cross-linked with 1 mM disuccinimidyl suberate d0/d12 (DSS, Creative Molecules Inc.) at room temperature (RT) for 30 min and quenched for 10 min at RT with ammonium bicarbonate (50 mM final concentration). Samples assembled in high salt, where no significant ribosome oligomerization occurred, were processed immediately, while low salt samples, which showed a significant level of ribosome oligomerization, were applied to 10%–47% sucrose gradients to separate monomeric 40S fractions from oligomeric assemblies and unbound eIF3 before further processing. There are strength and weaknesses in each of these approaches, as the extra purification steps lower the final peptide yields, while the nonpurified samples retain a background of eIF3 interlinks from the unbound eIF3 fraction of the sample. To ensure a large and unbiased set of identified crosslinked peptides, we included crosslinks from both data types in the final data set, although we did exclude eIF3⋅eIF3 interlinks from our nonpurified samples, as there is no unbiased way to determine whether the crosslink was generated when eIF3 was bound to 40S or free in solution.

43After 1 hr treatment with 10 μg/ml RNase A (Roche), the samples were reduced with 2.5 mM TCEP in 8 M urea at 37°C for 30 min and alkylated with 5 mM iodoacetamide (Sigma-Aldrich) for 30 min at RT in the dark. Before digestion, the samples were diluted with ammonium bicarbonate to a 1 M final concentration of urea. 2% w/w trypsin (Promega) was added overnight at 37°C and inactivated by acidification with 2% (w/v) trifluoroacetic acid (TFA). Peptides were purified with Sep-Pak C18 MicroSpin columns (Waters, Milford, MA), according to the manufacturer’s protocol, followed by enrichment of crosslinked peptides using size exclusion chromatography. LC-MS/MS analysis was carried out on an Orbitrap Elite mass spectrometer (Thermo Electron, San Jose, CA) and samples were shot in duplicates. Data were searched using xQuest in iontag mode against a database containing the protein sequences of all eIF3 subunits, eIF1 and all 40S ribosomal proteins with a precursor mass tolerance of 10 ppm. For matching of fragment ions tolerances of 0.2 Da for common-ions and 0.3 Da for crosslink ions were used and only peptides with a minimal length of 4 amino acids were considered. False discovery rates (FDR) of crosslinked peptides were assigned using xProphet. Crosslinked peptides were identified with a delta score of < 0.95 and a linear discriminant (ld) score of > 28 as described in the text. Mass spectra of the most abundant peaks were additionally analyzed by visual inspection in order to ensure good matches of ion series for both cross-linked peptides.

44The integrative approach to determine the S. cerevisiae 40S⋅eIF1⋅eIF3 complex architecture proceeds through four stages: (1) gathering of data, (2) representation of subunits and translation of the crosslinking data and the prior knowledge into a Bayesian scoring function, (3) configurational sampling to produce an ensemble of models that minimize the Bayesian scoring function, and (4) analysis of the ensemble (Figure S6). All residue ranges described below refer to the S. cerevisiae sequence.

45We used the crystallographic structures of the 40S ribosome subunit (PDB 3U5F and 3U5G) as well as the eIF3 subunit domains described in the main text (the eIF3a/eIF3c PCI heterodimer, the eIF3b β propeller, and the eIF3b-CTD/eIF3i/eIF3g-NTD complex). We removed chain h (the Stm1 protein) from the 40S structure, and rigid-body adjusted the position of helix ES6A to better fit the 43S EM density. We increased the structural coverage of the complex by adding comparative models for the eIF3b-RRM (residues 37-626, human homolog PDB 2KRB) in complex with a short eIF3j fragment (residues 30-41), eIF3g-RRM (residues 183-282, human homolog, PDB 2CQ0), eIF3j-CTD helical domain (residues 146-210, human homolog, PDB 3BPJ), and eIF1 (human homolog, PDB 2IF1). Comparative models were built with MODELER 9.12. The S. cerevisiae eIF1⋅40S interface was modeled using the Tetrahymena thermophila 40S⋅eIF1 complex structure. Sequence segments without a template structure included eIF3a-CTD (residues 496-964), eIF3c-NTD (residues 1-251), eIF3g linker (residues 97-183), eIF3j disordered regions (41-145) as well as other missing loops or terminal domains (Figure S7). The 965 crosslink observations (Table S4) were identified by mass spectrometry from eight independent experiments. We used the crosslink Id-Score and FDR to define three crosslink confidence classes, allowing us to construct a noise model for the crosslinking data (Figures 5D, S5 and Table S5).

46The domains of the 40S⋅eIF1⋅eIF3 subunits were represented by beads, arranged into either a rigid body or a flexible string, based on the available crystallographic structures and comparative models (Figure S7). To balance the thoroughness of configurational sampling and precision of model representation, we represented the structures in a multi-scale fashion. For the crosslink and excluded volume scoring terms, the crystallographic structures and comparative models were coarse-grained by representing each consecutive segments of 1 and 10 residues by a bead, centered on the center of mass, respectively (Figure S7). Sequence segments missing in the crystal structures were substituted by a single or multiple beads of the corresponding size (Figure S7). For each domain and interface with an atomic model, the beads representing a structured region were kept rigid with respect to each other during configurational sampling (i.e., rigid bodies). The rigid bodies include: the 40S⋅eIF1 complex, the eIF3a/eIF3c PCI heterodimer, eIF3b(655-699)/eIF3i/eIF3g(43-86) complex, eIF3b-RRM/eIF3j(30-41), eIF3g-RRM, eIF3b short helix (631-655), eIF3g(8-38) N-terminal double-stranded β sheet, eIF3g(86-97) short helix, and the C-terminal helices eIF3j(146-167), eIF3j(168-187), and eIF3j(187-210). Segments without a crystallographic structure or comparative model (i.e., segments predicted to be disordered or structured without a known homolog) were represented by a flexible string of large beads corresponding to a maximum of 30 residues each (Figure S7).

47The Bayesian approach estimates the probability of a model, given information available about the system, including both prior knowledge and newly acquired experimental data. The model M ≡ (X, {αi}) includes the structure coordinates X and additional parameters {αi} (below). Using Bayes’ theorem, the posterior probability p(M|D,I), given data D and prior knowledge I, iswhere the likelihood function p(D|M,I) is the probability of observing data D, given I and M; and the prior is the probability of model M, given I. To define the likelihood function, one needs a forward model that predicts the data point (i.e., the presence of a crosslink between two given residues) given the structure, and a noise model that specifies the distribution of the deviation between the observed and predicted data points. The Bayesian scoring function is the negative logarithm of p(D|M,I)p(M|I), which ranks the models identically to the posterior probability. All information gathered in Stage 1 is encoded into the Bayesian scoring function as a likelihood function (crosslinking data), as well as priors (sequence connectivity restraints to enforce proximity between beads representing consecutive sequence, and the excluded volume between all pairs of beads).

48For a given experimentally identified crosslink dn between two lysines i and j, the forward model predicts the presence of the crosslink given the structural coordinates X. Because of the multi-scale representation, residues are mapped into coarse-grained beads with variable radius, which represents the uncertainty of their positions. Furthermore, the rigid-body representation of crystallographic structures does not account for the flexibility of residues placed in loops. Therefore, we introduced the parameters {σk}, one for each crosslinked lysine, that define the unknown radii of the spheres, centered on the beads representing the lysines, within which the position of the lysine Cα is uniformly distributed. Given ri and rj (the coordinates of the Cα of the two lysines) and the maximum length for the crosslinker lXL, including the lysine side chains, we defined the forward model fn as the probability of randomly picking two points and within the spheres centered on ri and rj, with radii σi and σj, such that the distance between them is lower than lXL. The forward model value is one (ie, the identified crosslink is present in the given structure) when the spheres are closer than lXL and the radii are sufficiently small. It has a value of zero (i.e., the identified crosslink is not present in the given structure) if the distance between the spheres exceeds lXL. To reduce the number of parameters in the model, we utilized a single structural uncertainty σ, whose value estimates the average of {σk} for all crosslinked residues. We estimated lXL = 21 Å from potential of mean force calculations of a system comprising of the DSS crosslinker covalently attached to the two lysine sidechains (Yannick Spill and Michael Nilges, personal communication).

49The likelihood function for an identified crosslink dn is defined by the following data noise model:where ψn ∈[0,0.5] is the uncertainty associated with the crosslink identification. When the value of ψn is 0.5, the uncertainty is maximal and the likelihood does not depend on the forward model fn or on the distance between the lysines. When ψn is 0.0, the uncertainty is minimal, and the likelihood is identical to the forward model, i.e., it reaches a maximum when the distance between the spheres rij is below the crosslink length lXL. The joint likelihood function p(D|M,I) for a data set D = {dn} of NXL of independently observed crosslinks is the product of likelihood functions for each data point. Crosslinks observed multiple times are considered as independent observations, and therefore considered multiple times in the joint likelihood function. Analysis of the intra-40S crosslink distances (Figure 5C) allowed us to group the crosslinks into three classes, including a class with an estimated ∼5% false-positive rate (high confidence crosslinks with Id-Score > 36.0; FDR < 0.05), a class with a ∼10% false-positive rate (medium confidence with 32.0 < Id Score < 36.0; 0.05 < FDR < 0.1) and a class with ∼30% false-positive rate (low confidence with 28.0 < Id Score < 32.0; 0.1 < FDR < 0.3) (Figures 5D, S5 and Table S5). We reduced the number of ψn parameters, by defining uncertainties ψH (358 crosslinks), ψM (263 crosslinks) and ψL (344 crosslinks) for the three classes.

50The model prior p(M|I) is defined as a product of the individual priors p(X), p(σ), p(ψH), p(ψM), and p(ψL) on the structural coordinates, structural uncertainties, and the three data point uncertainties, respectively. The prior p(X) included the excluded volume and terms to maintain sequence connectivity. The excluded volume restraint was implemented as a pairwise hard-sphere repulsive potential, where the bead radii were determined using the statistical relationship between the volume and the number of residues. The sequence connectivity terms are upper harmonic distance restraints, with a threshold distance set to be four times the sum of the radii of the two connected beads. To facilitate configurational sampling, we applied a weak restraint whose score depended linearly on the distance between cross-linked residues, with a slope of 0.01 Å−1. p(σ), p(ψH), p(ψM), and p(ψL) are uniform distributions, that bound σ to interval [0,100], and the ψ parameters to the interval [0,5].

51Structural models of the 40S⋅eIF1⋅eIF3 complex were computed by Replica Exchange Gibbs sampling, based on Metropolis Monte Carlo sampling. This sampling was used to generate a collection of configurations for the system as well as values for the uncertainty parameters. The Monte Carlo movements included random translation and rotation of rigid bodies (1 Å and 0.025 radian maximum, respectively), random translation of individual beads in the flexible segments (1 Å maximum), as well as Gaussian perturbation of the uncertainty parameters. The sampling was run on 64 replicas, with temperatures ranging between 1.0 and 2.5. Each replica produced between 10,000 and 30,000 models. 12 independent sampling calculations were run, each one starting with a random initial configuration, for a total of 180,000 models. This set of models was divided into two ensembles of the same size, to assess sampling convergence (Figure S8). Furthermore, 6 additional runs were carried to assess the robustness of the modeling approach by removing approximately 10% and 20% of the crosslinks, respectively.

52For each ensemble, the solutions (ie, the 500 best scoring models) were grouped using k-means clustering using the root-mean-square deviation (rmsd) of all eIF3 subunits as a structural similarity criterion when the models were superposed on the 40S subunit structure. The rmsd distance matrix identified two dominant clusters of similar structures (Figure S8A), here named cluster 1 and 2, with cluster 1 being more populated and scoring better than cluster 2 (Figure S8). The overall precision (calculated as the average pairwise rmsd) for cluster 1 and 2 was ∼36 and ∼41 Å, respectively, for both ensembles. The superposed structures in the cluster were converted into the probability of any volume element being occupied by a given protein (that is, the ‘localization density’) (Figure 6C). While the eIF3b/eIF3a/eIF3c/eIF3j subunits localize on the 40S surface similarly in the two clusters, the eIF3i subunit displayed two alternate positions (Figure S8B and S8C): one toward the 40S head (cluster 1) and one toward the bottom of the body (cluster 2). Localization precision calculated for each individual domain (Table S6) indicated that the eIF3b β propeller was the domain with the most precisely determined position (∼15 Å), followed by the eIF3a/c PCI and the eIF3b-RRM. eIF3j and the eIF3b helical linker positions were determined less precisely, and eIF3i localized with different precision within the two clusters (∼38 Å in cluster 1 and ∼58 Å in cluster 2). eIF3g-RRM was the least well determined component, with a localization precision of ∼90 Å.

53J.P.E., F.S., R.P., and N.B. conceived the study and experimental approach; J.P.E., S.Z., and T.S. cloned, purified, and crystallized eIF3 samples, collected and processed X-ray diffraction data, solved the structures, and built the atomic models; J.P.E., T.S., F.S., and C.H.S.A. prepared samples for EM and CX-MS experiments. C.H.S.A. and D.B. collected EM data and calculated the EM reconstructions; F.S. performed CX-MS experiments; R.P. carried out the integrative modeling with the help of P.C.; J.P.E., F.S., and R.P. analyzed the integrative modeling results, and J.P.E., F.S., R.P., A.S., R.A., and N.B. wrote the paper with input from all authors.