Exploring the interaction dynamics of eukaryotic translation initiation factor 2 Open Access

eIF2 is a heterotrimeric GTPase, composed of α, β, and γ subunits (called eIF2S1, S2, and S3, respectively, in online databases). The eIF2γ subunit is the actual GTPase, while eIF2α and β have accessory and regulatory functions. Phosphorylation of eIF2α under a number of stress conditions by any one of the stress kinases GCN2, PKR, PERK, and HRI turns eIF2 into an inhibitor of its GEF eIF2B, which triggers the ISR (Figure 1, reviewed in [1,3,4,13]). Because of its role as an ISR marker, eIF2α is arguably the best known eIF and also the most frequently misnamed – it is often called elF2a, but sometimes even elF2A, which is a completely unrelated protein. eIF2, eIF1, eIF3, eIF5, and the Met-tRNA_i assemble into the multifactor complex (MFC). Within the MFC, all proteins interact with each other, although the strength of individual interactions varies between mammals and S. cerevisiae [14,15]. The interactions of eIF1 with eIF2 and eIF3c have been observed within the PIC [16-20], although only a single helix from eIF3c is observed interacting with eIF1 in PIC structures while at least in S. cerevisiae, eIF3c has multiple binding sites for eIF1 [21]. eIF5-CTD binds to eIF2β, eIF3c, and weakly to eIF1 [14,22-26], as well as to eIF1A [24,25,27]. Unfortunately, eIF5-CTD is not visible in Cryo-EM structures of yeast or mammalian 43S or 48S PICs; thus, none of its interactions could be observed, either. The only case where eIF5-CTD was seen was the structure of the T. cruzi 43S PIC, where it was observed contacting eIF2 [28].

If the 48S PIC encounters an AUG codon in optimal Kozak sequence context, the Met-tRNA_i anticodon base-pairs with the start codon; the 48S PIC converts from an ‘open’, scanning-competent, to a ‘closed’, scanning-incompetent conformation; and translation almost always initiates from that start codon. The probability of initiation at a suboptimal start codon depends on the stringency of start codon selection. If the 48S PIC encounters an AUG in a suboptimal context or a near-cognate start codon (e.g., CUG or GUG) in an optimal context, the 48S PIC may recognize it and adopt a ‘closed’ conformation, but it would be less stable, and the 48S PIC may either initiate at that start codon or revert back to the open state and continue scanning (Figure 1, right). While all PIC components influence the stringency of start codon selection, either directly or indirectly, the central players in this process are eIF1 and eIF5. eIF1 is bound next to the Met-tRNA_i in the ribosomal P-site and is displaced from there when the tRNA base-pairs with the start codon. The N-terminal domain (NTD) of eIF5 then takes the place of eIF1 on the 40S subunit. eIF5 also promotes GTP hydrolysis by eIF2, which commits the 48S PIC to initiate at that start codon (reviewed in [1-4]). Higher eIF1 concentrations increase the stringency of start codon selection, whereas higher eIF5 concentrations lower it. In fact, eIF1 and eIF5 regulate their own and each other’s translation by modulating the stringency of start codon selection: the eIF1 start codon is in a suboptimal context, whereas the 5′-untranslated region (5′-UTR) of eIF5 contains inhibitory upstream open reading frames (uORFs) in a suboptimal context. Thus, higher eIF1 concentration decreases translation of its mRNA while increasing translation of eIF5, whereas higher eIF5 concentration has the opposite effect [29-31]. The regulatory eIF5-mimic protein 5MP competes with eIF5 for binding to eIF2 and the PIC (Figure 1) [32-34]. Mammals have two 5MP isoforms: 5MP1 and 5MP2, previously known as basic leucine zipper – W2 domain 2 (BZW2) and BZW1, respectively; but their NTD was predicted by us [35] and AlphaFold [36,37] to be an MA3 HEAT domain homologous to the second HEAT domain of eIF4G, not a leucine zipper. The C-terminal domains (CTD) of eIF5, eIF2Bε, and the 5MPs are W2 HEAT domains homologous to each other and to the third HEAT domain of eIF4G [38]. By competing with eIF5, the 5MPs increase the stringency of start codon selection, similar to eIF1, and their translation efficiency is auto- and cross-regulated with those of eIF1 and eIF5 [34,39]. While the regulation of eIF1 translation is conserved in S. cerevisiae, this does not appear to be the case for eIF5. Most Ascomycetes (yeast), including S. cerevisiae, do not have a 5MP homolog, whereas Basidiomycetes (mushrooms) do have a 5MP [40,41].

eIF2-GDP and eIF5 bind to each other with high affinity [15,42-44] and eIF2-GDP may be released from the 48S PIC together with its GAP eIF5, at least in S. cerevisiae [43,44]. The GEF eIF2B has to displace eIF5 before it can recycle eIF2-GDP back to eIF2-GTP and eIF2-GTP•Met-tRNA_i TC (Figure 1). Thus, eIF5 also plays the role of a GDP-dissociation inhibitor (GDI), and eIF2B also plays the role of a GDI-dissociating factor (GDF) [43-45]. According to the GDI definition used in ref. [44] that a GDI directly slows down GDP release, only S. cerevisiae eIF5, but not human eIF5 is a GDI [15,44]. However, the main inhibitory effect of eIF5 in both yeast and mammals is mediated by preventing eIF2B binding to eIF2, not further slowing down the already slow spontaneous exchange rate of free eIF2-GDP. According to the original definition that the GDI interferes with the GEF binding [46], both eIF5′s are GDIs. 5MP1 and 5MP2 compete for eIF2 binding not only with eIF5 but also with eIF2B [32,47]. Thus, like eIF5, they can also act as GDIs.

The CTDs of eIF5, eIF2Bε, and the 5MPs bind to the eIF2β N-terminal tail (eIF2β-NTT). eIF2β-NTT has three lysine repeats (K-boxes) important for the interactions with eIF5, eIF2Bε, and the 5MPs [14,23,32,34]. We recently found that eIF2β-NTT contains three tandem binding sites centered on the three K-boxes. eIF5, eIF2Bε, and 5MP1 can each bind to all three K-boxes with similar but not identical affinities. Three proteins can simultaneously bind to the same eIF2β-NTT, but when they do, they reduce each other’s affinity [47]. This led us to propose that the simultaneous binding and mutual destabilization allow eIF5 and eIF2B to speed up each other’s displacement from eIF2 (as in the GDF function of eIF2B) and mediate the regulatory functions of the 5MPs, which compete with eIF5 and eIF2B for eIF2 and modulate their activities [47].

CK2 is important for cell proliferation and differentiation, has hundreds of substrates, and is constitutively active in most cells under most conditions [48-51], but inactive under certain chronic stress conditions [48,52-54]. In mammals, CK2 phosphorylates eIF5 [49], eIF2Bε [55], and the 5MPs [47] within their eIF2β-binding regions, which increases their affinities for eIF2β-NTT in vitro [47,56], stimulates the activities of eIF2B in vitro [55], eIF5 in cell culture [49], and likely also the 5MPs. The observation that CK2 phosphorylation stabilizes multiple, often competing interactions, suggests that it remodels the entire system, making it more efficient, although the effects of phosphorylation have not yet been examined in the context of the PIC. This raises the possibility that under chronic stress, when overall translation is suppressed and the concentrations of ribosomes and translation factors are likely proportionally reduced, the stability and even composition of translation complexes may be significantly modulated. Mammalian mTORC1 and CK2 bind to eIF2β and phosphorylate its NTT [48] on the periphery of the binding sites for eIF5, eIF2Bε, and the 5MPs, allowing NCK1 to bind and recruit PP1, which in turn dephosphorylates eIF2α and blunts the ISR when mTORC1 and/or CK2 are active [48].

When (and how stably) are eIF1 and eIF5 bound to the scanning 48S PIC? eIF1 is required for proper scanning and start codon selection. Therefore, it must be present on the scanning 48S PIC, at least most of the time, on most PICs. eIF5, on the other hand, is not required for scanning but is required at the start codon, where it replaces eIF1 at the P-site and promotes GTP hydrolysis by eIF2, which is required for the 48S PIC to commit to initiating at this start codon. On the other hand, the stringency of start codon selection depends on eIF5 concentration; therefore, the eIF5 occupancy at the start codon depends on free eIF5 concentration [30,57,58]. Therefore, eIF5 must be present on every 48S PIC at the start codon, at least transiently to promote initiation at that codon, but not 100% of the time, because then its occupancy would not have been dependent on free eIF5 concentration. eIF5-NTD is first observed in Cryo-EM structures of PICs once it replaces eIF1 upon start codon selection [4,19,59]. The only PIC structure that contains eIF5-CTD (and the only instance when eIF5 was observed on the PIC prior to start codon selection) is the T. cruzi 43S PIC structure [28]. This is also the only instance when a putative portion of the intrinsically disordered eIF2β-NTT was observed – as density contacting eIF5-CTD. At least at suboptimal start codons, the release of eIF1 is reversible, and it can compete with eIF5-NTD for binding to the 40S. A recent report using SM-FRET provided direct evidence that mammalian eIF1 and eIF5 binding to the 48S PIC is indeed dynamic and that they compete for the binding site on the 40S upon start codon selection [60]. Remarkably, the authors did not observe eIF5 on the 48S PIC until start codon selection and proposed that in mammals, eIF5 is not present on the scanning 48S PIC [60]. While the failure to observe eIF5 in a given Cryo-EM structure may mean that it is tethered to the PIC via disordered segments and remains invisible, failure to detect it on the scanning 48S PIC using SM-FRET or even with direct excitation is not easy to rationalize. This finding is highly unexpected and difficult to reconcile with many previous reports, raising questions we do not seem to have the answer to. The discussion below relies on the assumption that if an interaction exists in vitro, it also occurs in vivo on and/or off the PIC, and it has a function, or it would not have been conserved during evolution. (1) eIF5 is a stable component of the MFC, not only in S. cerevisiae [14], but also in mammals [15]. Of course, not all interactions within the MFC are retained within the PIC. However, if eIF5 is part of the MFC and falls off when the MFC binds to the 40S, why was it part of the MFC in the first place? At present, there are no known functions of eIF5 within the MFC or of the MFC itself off the PIC. (2) What would prevent eIF5 from interacting dynamically with the eIF2β-NTT, which is not visible in PIC structures and likely remains disordered? A possible explanation is that the authors used recombinant bacterially expressed eIF5, which would not be phosphorylated by CK2 and thus have at least an order of magnitude lower affinity for eIF2β [47]. (3) The interactions of eIF5 with eIFs 1, 1A, and 3c are conserved from fungi to mammals [14,15,24,25] and such ‘late arrival’ of mammalian eIF5 is also difficult to reconcile with the model that eIF1A and eIF3c displace eIF2β-NTT from eIF5-CTD on the PIC [24,25] – if eIF5 first binds when the 48S PIC is already at the start codon, does the incoming eIF5 bind to eIF2β first or to eIF1A/3c first? Furthermore, the eIF1A N-terminus may no longer be accessible in the 48S PIC at the start codon [19]; then, when and where does eIF5-CTD bind to eIF1A? It appears that our understanding of the dynamics of eIF5 binding to the PIC is shifting from eIF5 being always present in the PIC, to coming later, after scanning. Or the truth could be somewhere in between. As discussed earlier, the dependence of the stringency of start codon selection on eIF5 concentration indicates that eIF5 is not present on the PIC 100% of the time; therefore, even if it bound initially to the 43S PIC (whether alone or as part of the MFC), this interaction is likely dynamic, involving rounds of binding and release. eIF5 may bind dynamically (on and off) to the scanning 48S PIC by contacting eIF2β-NTT and likely also eIF3c-NTT. Then, upon start codon selection, eIF5-NTD replaces eIF1 from its binding site on the 40S (and becomes visible in Cryo-EM structures, while eIF5-CTD remains dynamic and invisible). However, this may not be the case under certain chronic stress conditions, where CK2 is inactive, eIF5-CTD becomes dephosphorylated and its binding affinity for eIF2β weakens significantly (to a Kd in the μM range [47,56]). Under those conditions, eIF5 may indeed arrive after start codon selection as observed by Grosely and co-authors [60].

When (and how stably) is the eIF2-GTP•Met-tRNA_i TC bound to the scanning 48S PIC? For the purposes of this discussion, we focus on PICs, where eIF2 (and not alternative factors) recruits the Met-tRNA_i. Also, a 48S PIC missing only the Met-tRNA_i would be functionally similar to a 48S PIC missing the entire TC. The general notion is that under normal conditions the TC is always present on the scanning 48S PIC and absent only under special conditions. The TC is not required for scanning, as evidenced by the mechanism of translation re-initiation after an uORF [61]. However, it is required for start codon selection – without the Met-tRNA_i, the 48S PIC cannot discriminate between a start codon and any other codon. Therefore, it must be present on every 48S PIC before it can select a start codon. However, it is possible that, during ISR, when the TC levels are lower, incomplete 43S PICs can be assembled – containing eIF2-GDP, but not Met-tRNA_i, and/or PICs missing the entire TC. This possibility was suggested, for example, by Sokabe and co-authors, who found that the eIF2 interaction with the MFC is independent of the bound nucleotide, while the MFC strongly discriminates against Met-tRNA_i binding to eIF2-GDP [15]. If such a Met-tRNA_i-less 43S PIC can bind to the mRNA and start scanning, it would enable regulatory mechanisms similar to those mediated by translation re-initiation after an uORF but distinct from them and complementary.

When is eIF2-GDP released from the 48S PIC, and when does eIF5B come? Upon start codon selection and GTP hydrolysis, eIF2-GDP is replaced by eIF5B. But is it replaced or displaced? It has long been known that the eIF2-GDP affinity for Met-tRNA_i is not negligible and that eIF2-GDP release may not be instantaneous [62,63]. Furthermore, if there is a long delay between eIF2 release and eIF5B binding, there is a risk of the labile bond between the tRNA and the methionine getting hydrolyzed. Previously, it was reported that eIF2-GDP hangs around on the 48S PIC for a while, some even surviving ultracentrifugation, and that eIF5B promotes its release [63]. Later, we reported that in mammals (but not S. cerevisiae), eIF5 binds directly to eIF5B and proposed that eIF5B is recruited by eIF5 before it is needed (before GTP hydrolysis by eIF2) and displaces eIF2-GDP upon GTP hydrolysis [64]. New late-48S PIC structures show that eIF2-GDP does indeed remain on the 48S PIC and is indeed displaced from the Met-tRNA_i by eIF5B, and even remains associated with the 48S PIC after it has lost contact with the Met-RNAi. The authors did not observe eIF5B on the 48S PIC until after eIF5-NTD replaced eIF1 on the 40S [19]. Also, if as discussed above, eIF5 itself does indeed arrive only after start codon selection, that is the earliest it can bring eIF5B with itself. Thus, like with eIF5, the ‘eIF5B pendulum’ may have swung back and forth, too – from the original idea that eIF5B comes after eIF2-GDP has left, to eIF5B being present already on the scanning 48S PIC, back to maybe coming before GTP hydrolysis by eIF2 but after start codon selection. If eIF5 does bind dynamically to the 48S PIC before start codon selection, the known interaction between eIF5-CTT and eIF5B-CTD (Kd about 500 nM) [64] may not be sufficient to keep eIF5B on the 48S PIC without additional contacts between eIF5B and other PIC components. But since eIF5B is not observed in early-stage PICs and its binding site on the 40S subunit may be occupied by ABCE1 [17], such putative interactions would need to be dynamic as well – perhaps mediated by the mostly disordered N-terminal half of eIF5B? It is intriguing that in the new late-48S PIC structure [19], eIF5B-CTD contacts eIF5-NTD – a novel interaction that has never before been observed.

Another set of outstanding questions relate to interactions observed in solution but not in PIC structures. For example, while it is safe to assume that eIF2β-NTT binds to eIF5-CTD and 5MP-CTD both off and on the PIC (before and/or only after start codon selection), it is not clear when eIF5-CTD binds to eIF1, eIF1A, and eIF3c, when 5MP binds to eIF3c, or whether the timing of these interactions is conserved between mammals and yeast. The interaction between eIF1 and eIF5 is rather weak, mediated by eIF5-CTD, and the respective surface on eIF1 is occupied by other interactions when it is bound to the PIC; therefore, eIF1 and eIF5 probably only contact each other within the MFC off the ribosome. However, the interactions of eIF5-CTD and 5MP-CTD with eIF3c were reported to be important for the stringency of start codon selection and thus must be happening on the 48S PIC (but remain invisible in structures; thus, must be tethered via flexible linkers). Likewise, eIF1A has not been reported to be part of the MFC. Therefore, it would likely be binding eIF5 within the PIC. The main interaction involves the N-terminus of eIF1A; thus, it would be possible at stages, where the N-terminus of eIF1A remains disordered within the PIC (otherwise, eIF5-CTD would have been observed bound to eIF1A in the PIC structure). It is even harder to guess when the eIF5-CTD interaction with the folded domain of eIF1A could occur. Finally, while a single helix from eIF3c is seen in PIC structures contacting eIF1 [16-20], eIF3c appears to have multiple eIF1- and eIF5-binding sites in S. cerevisiae [21]. The location of the eIF1-binding region in mammalian eIF3c is different than in S. cerevisiae, while the sequences appear similar, and this difference could be due to translocation, as opposed to convergent evolution [16,17]. Thus, it remains to be elucidated how many eIF1- or eIF5-binding sites exist in eIF3c from different kingdoms of life and which interactions can occur on the PIC, fully or partially.

In summary, it is now well established that many more interactions within the PIC are dynamic and that the PICs themselves are more heterogeneous than previously thought. In fact, we may be better off assuming that the association of a given eIF with the PIC may be dynamic until proven otherwise. Figure 2 attempts to illustrate most points in translation initiation at which each of the proteins discussed here binds to the PIC or dissociates from it. This emphasizes the need for quantitative approaches to better understand the process of translation initiation. In this context, kinetics experiments measuring the lifetimes of individual complexes are invaluable. Another quantitative approach with great potential is considering the concentrations of ribosomes and eIFs in different cells and under different conditions. Unfortunately, the cellular concentrations of proteins and other molecules are notoriously difficult to measure accurately, and the available numbers are contradictory in proteomics databases and even in studies specifically focused on ribosomes and translation factors, varying by up to an order of magnitude [65-67]. Although some recent studies report more realistic numbers and ratios of translation factors to ribosomes [66,68].

Another factor to consider is the presence of membrane-less protein condensates, which appear to serve a wide range of functions, including sequestering mRNAs and translation apparatus components, degradation, and translation factories, where certain mRNAs and translation complexes are concentrated (reviewed in [69,70]). Protein condensates are often nucleated by intrinsically disordered regions (IDRs) in proteins. A number of eIFs contain extensive IDRs, including eIF3 subunits, eIF4G, and eIF5B, some of which have already been reported to form protein condensates involved in regulation of translation (see e.g., [71,72]). Such condensates pose potential challenges because they make it difficult to estimate the actual local mRNA and protein concentrations at the site, where translation takes place but also open exciting new avenues for research.

Combining PIC structures, ribosome and eIF concentrations, and SM-FRET and other kinetics data will allow a better mechanistic understanding of the process by performing an ‘accounting’ of the translation apparatus – how many molecules of a given component are there, who are they bound to, for how long, what are they doing, and how long does every step of the process take on average?

The author declares no competing interest.

This work was supported by the National Institute of Health [GM134113 to A.M.].

AM: Conceptualization; Writing – Original Draft; Writing – Review & Editing; Funding Acquisition; Supervision.

GAP

GTPase-activating protein

GEF

guanine nucleotide exchange factor

ISR

integrated stress response

MFC

multifactor complex

PIC

pre-initiation complex

TC

ternary complex

eIF2

eukaryotic translation initiation factor 2

uORF

upstream open reading frame