N⁶-methyladenosine in 5’ UTR does not promote translation initiation

SUMMARY

The most abundant N⁶-methyladenosine modification on mRNAs is installed non-stoichiometrically across transcripts, with 5’ untranslated regions (5’ UTRs) being the least conductive. 5’UTRs are essential for translation initiation, yet the molecular mechanisms orchestrated by m⁶A remain poorly understood. Here, we combined structural, biochemical and single-molecule approaches and show that at the most common position, a single m⁶A does not affect translation yields, the kinetics of translation initiation complex assembly and start site recognition both under permissive growth or following exposure to oxidative stress. Cryo-EM structures of the late preinitiation complex reveal that m⁶A purine ring established stacking interactions with an arginine side chain of the initiation factor eIF2α, however with only a marginal energy contribution, as estimated computationally. These findings provide molecular insights into m⁶A interactions with the initiation complex and suggest that the subtle stabilization is unlikely to affect the translation dynamics under homeostatic conditions or stress.

eTOC blurb

Guca et al. use an integrative approach, including functional, structural, single-molecule and computational approaches, and show that a single m⁶A in the 5’UTR does not affect translation dynamics under homeostatic conditions or stress. They observe interactions between m⁶A with an arginine of eIF2α, however with only a marginal energy contribution.

Graphical Abstract

INTRODUCTION

N⁶-methyladenosine (m⁶A), the most abundant mRNA modification in higher eukaryotes, modulates mRNA stability and metabolism^1,2, and affects diverse physiological and pathological conditions^3,4. m⁶A are installed by a dedicated methyltransferase complex within a conserved DRACH motif, where D = A/G/U; R = A/G; and H = U/A/C⁵. However, only a small subset of all DRACH-confined adenosines (e.g. 0.2–0.6%) in mammalian transcriptomes are methylated with a rather stochastic choice between motifs in a close proximity⁶. Sequencing data across various species have established a conserved asymmetric distribution of m⁶A across transcripts, with a strong positional enrichment towards the end of transcripts and 3’ UTRs^7–10. Two recent studies resolved this enigmatic biased distribution of m⁶A across transcripts and show that m⁶A is introduced in a non-selective manner at eligible consensus DRACH motifs governed by the exclusion from the splice junctions^11,12. The exon junction complexes sterically suppress m⁶A installation within approximately 100 nt of the exon junctions, consequently depleting 5’UTRs and most of the coding sequence of methylation; mammalian mRNAs are composed of short internal exons (90% of all exons are <246 nt)¹¹.

Distinct functional implications have been attributed to m⁶A modification dependent on its position in transcripts. Diverse studies have consistently found that the most pervasively m⁶A modified regions flanking stop codons (i.e. within 3’ ends of the coding sequence and 3’ UTRs) modulate mRNA stability and turnover^13–15. Within the coding sequence, m⁶A modulates tRNA selection and translation-elongation dynamics^16,17 or codon reading accuracy ¹⁸. Compared to other transcript regions, 5’UTRs are the least conductive for methylation with very low modification frequency^9,11,12, yet its functional role in translation initiation remain debatable. An earlier study proposes that m⁶A modifications in the 5’UTRs facilitate translation initiation by direct binding of eukaryotic initiation factor 3 (eIF3) in a manner that is independent of 5’ cap or cap-binding proteins¹⁹. A single m⁶A in the 5’ UTR mediates the assembly of the initiation complex (IC), consisting only of eIF1, eIF1A, eIF2, eIF3 and 40S ribosomal subunit¹⁹. Another study argues that the effect of m⁶A on eIF3 and initiation might be indirect and mediated by a dedicated reader protein YTHDF1 which recognizes 3’UTR m⁶A modifications and through mRNA looping mediates interaction with eIF3¹⁰. Wang and colleagues concluded a similar indirect mechanism with no direct role of m⁶A in 5’UTRs in translation. Synergetic interactions between two readers (YTHDF1 and YTHDF3) recognizing m⁶A in the 3’UTR and the initiating ribosome facilitate translation ²⁰. A recent study shows that an epitranscriptomic m⁶A mark in the 3’UTR and not in the 5’UTR ties translation of insulin on energy homeostasis²¹. Thus, the role of m⁶A in translation initiation and the precise molecular mechanism by which a single m⁶A site in 5’UTR facilitates the IC assembly and translation initiation still remain enigmatic.

Here, we set out to resolve with molecular details the impact of m⁶A in translation initiation dynamics using an integrated approach, combining deep-sequencing analysis in cells and biochemical analysis with structural methods and single-molecule spectroscopy in in vitro reconstituted systems. Our results show that a single m⁶A causes few subtle rearrangements of the late-stage IC, however with no discernible effect on the kinetics of IC assembly and transition into elongation, and on the overall translation efficiency.

RESULTS

m⁶A in the 5’UTR does not enhance translation efficiency in vivo under stress

To determine the m⁶A distribution in 5’UTRs across the transcriptome, we considered three deep sequencing-based approaches: (i) m⁶A-seq ²² and (ii) GLORI-seq ⁹ from HEK293 cells and (iii) eTAM-seq²³ from HeLa cells. GLORI-seq and eTAM-seq use either chemical or enzymatic deamination of unmethylated adenosines, respectively, and thus, dissect m⁶A positions with nucleotide precision. By contrast, the antibody-based m⁶A-seq lacks nucleotide precision, and in the peak calling m⁶A peaks containing at least one DRACH motif were considered as truly positive. A sizeable set of m⁶A sites were common among all three sequencing approaches (Table S1), implying some conservation of the m⁶A modification at gene level. However, much larger number of sites was without a common overlap, thus reflecting the stochastic nature of methylation at neighboring DRACH motifs⁶. We next analyzed the distribution of the m⁶A positions in the different transcript regions, i.e. 5’UTR, coding sequence (CDS) and 3’UTR. For comparison, the detected m⁶A sites for each transcript region were binned into equal number of bins. To compare between the three sequencing approaches, i.e. considering their different resolution, the m⁶A sites were represented as a ratio to the total number of the DRACH motifs found in each region (Figure 1A). Across all transcripts expressed under homeostatic conditions, the 5’UTRs contained the lowest m⁶A amount (Figure 1A), corroborating previous observations^{6,7,9,12,14,22,23}.

Figure 1 – m⁶A is enriched in close proximity of start codons but does not enhance overall translation.

(A) Box plot of m⁶A detected level (ratio of m⁶A over DRACH motifs) in GLORI-seq (HEK293), eTAM (HeLa) and m⁶A-seq (HEK293) from control conditions (black), or hypoxia (gray; GLORI-seq, HeLa), or oxidative stress (gray, m⁶A-seq, HEK293). Statistical significance determined by two-sided Wilcoxon rank-sum test. GLORI-seq: p = 4.2×10⁻³¹ (CDS), p = 1.8×10⁻³ (3’UTR); m⁶A-seq: p = 2.7×10⁻¹³⁸ (5’UTR), p = 1.9×10⁻¹⁴ (CDS) p = 1.3×10⁻¹⁴ (3’UTR).

(B) Superimposition of m⁶A sites in the 5’UTRs in transcripts expressed in permissive growth identified by GLORI-seq (black) and eTAM-seq (red). Zero, first nt of the start codon. m⁶A-containing transcripts were normalized to the total number of detected transcripts.

(C) Constructs for the in vitro translation assays. Dashed, reporter gene; thicker bar, coding sequence.

(D) In vitro translation assay in human lysates. The signal from equal starting amounts of the reporter constructs (C) was normalized to the signal with A(−3)-luc mRNA. Data are means ± s.e.m. (n= 4–7). Two-tailed unpaired t test was used for statistics.

(E) In vitro translation assay of A(−3)-Luc (gray) and m⁶A(−3)-Luc (red) separated by dashed line. The signal from equal starting amounts of the constructs was normalized to that of A(−3)-Luc (left) and m⁷G-A(−3)-Luc (right). Data are means ± s.e.m. (n= 4–7); two-tailed unpaired t test.

See also Figure S1, Table S1, S2 and S3.

The m⁶A levels in 5’ UTR have been shown to be elevated under different types of stress^19,24, thus, we also considered two published data sets from HEK293 cells exposed to oxidative stress (m⁶A-seq²²) and HeLa cells subjected to hypoxia (GLORI-seq ⁹). In hypoxia, the m⁶A levels slightly decreased across all transcript regions (Figure 1A). Following oxidative stress, the m⁶A level also decreased in the 3’UTRs, but increased in 5’UTRs and CDSs (Figure 1A). These results suggest no shared signature between both types of stress, although this comparison might be misleading. The different resolution of the two sequencing methods and the sizeable number of transcripts which underwent degradation following stress (Figure S1A), may distort the comparison under different type of stress.

To assess the effect of m⁶A on gene expression under oxidative stress, we next considered Ribo-seq and RNA-seq from HEK293 cells²² and selected 168 genes that were actively translated under oxidative stress (for details see Method section). We noted that following oxidative stress exposure, the overall ribosome density (RD), also called translation efficiency²⁵, decreased compared to permissive growth (Figure S1B), and the effect was mainly due to a decreased translation (i.e. decreased coverage in the Ribo-seq, Figure S1C), as the mRNA expression remained unchanged (Figure S1D). Notably, translation efficiency decreased equally for both transcript groups with and without m⁶A in the 5’UTRs (Figure S1B,C), implying no clear expression benefit of the m⁶A sites.

m⁶A in the 5’ vicinity of the start codon does not affect translation efficiency

To assess the m⁶A distribution in the 5’UTRs, we considered only GLORI-seq and eTAM-seq^9,23, which, unlike m⁶A antibody-based sequencing approaches, map the m⁶A sites with nucleotide precision. The highest level of the m⁶A sites was at positions −2 and −3 upstream of the start codon (Figure 1B; for details see Method section). We noted that m⁶A at position −3 falls within the Kozak sequence (RNNAUGG, start codon underlined and R=A/G), which is essential for selection of the translation start site^26,27 and in part resembles the DRACH consensus motif. Thus, we next sought to determine whether m⁶A at position −3 would facilitate translation using a well-defined human in vitro translation system. For this, we created four different luciferase reporters containing the 5’UTR and the first few codons of β-globin (Figure 1C). The 5’UTR of β-globin bears a strong Kozak sequence (gacAccAUGG) and at the adenosine at position −3 (underlined), we introduced the N⁶-methyladenosine (m⁶A(−3)) to match the most common m⁶A position (Figure 1B). As expected, cap-dependent translation yielded substantially higher luciferase signal than that of the uncapped variants (Figure 1D). Notably, the m⁶A presence did not affect the luciferase signal of both capped or uncapped mRNA (Figure 1D), implying no clear expression benefit for transcripts with a single m⁶A in the most commonly modified position (−3). This result is in line with an in vivo expression analysis showing that transcripts with m⁶A modification in the 5’UTRs are expressed equally or even slightly worse than transcripts without m⁶A⁹.

We also tested the effect of a single m⁶A on the translation of endogenous transcripts (Figure S1E). We selected three endogenous genes with different strength of the Kozak sequence (gggAcYAUGR, start codon underlined and Y=U/C, Figure S1E)). These transcripts were naturally modified at position −3 and expressed under homeostatic conditions and in stress (Figure S1B,C). We fused their capped 5’UTRs with and without m⁶A at position −3 to a luciferase reporter (Table S2) and compared the translation efficiency in human in vitro translation system. The expression of the GOLIM4-Luc construct with the weakest Kozak sequence was close to the background and did not allow robust measurements. For both TSC22D4-Luc and RP2-Luc, we did not detect any significant difference in the luciferase signal between unmodified and m⁶A modified constructs (Figure S1E), corroborating the observations with β-globin transcript.

The m⁶A modifications in the 5’UTRs have been suggested to facilitate translation initiation in a manner that is independent of 5’ cap¹⁹. An alternative cap-independent initiation could be relevant in stress conditions ²⁴, as activated integrated stress response (ISR) compromises translation of mRNAs with canonical cap-dependent initiation^28,29. Thus, to mimic the stress-triggered ISR activation in cells, we treated the in vitro translation reaction with protein kinase R (PKR)³⁰, which phosphorylates eIF2α (Figure S1F) – a sensor of activated ISR. Compared to the control (i.e. without PKR treatment), in conditions resembling ISR activation, the translation yields decreased equally for both methylated (m⁶A(−3)) and non-methylated (A(−3)) variants (Figure 1E), thus corroborating the observed overall reduction of translation efficiency in vivo following oxidative stress exposure (Figure S1B). Together, our data imply that a single m⁶A in the 5’ UTRs does not affect translation yields and the expression under homeostatic conditions or oxidative stress is comparable.

N⁶-methyladenosine in 5ʹUTR does not impact the dynamics of IC assembly

To investigate in depth the effect of m⁶A in the upstream vicinity of the start codon (i.e. at position −3) on distinct discrete initiation stages, we performed real-time single molecule experiments in a defined human in vitro translation system using the setup described in ³¹. The same 5’UTRs of the A(−3) and m⁶A(−3) reporter constructs (Figure 1C) were extended with 3ʹ-biotinylated oligonucleotide resembling RPL30 coding sequence and tethered to the imaging surface within zero waveguides (ZMWs). Each immobilized mRNA was preincubated with saturating concentrations of eIF4 proteins (4A, 4B, 4E and 4G), ATP and GTP, and then preassembled 43S preinitiation complexes (PIC) with 40S fluorescently labeled with Cy3 on the N terminus of uS19, eIF5B labeled with Cy3.5 on its N terminus and 60S subunit labeled with Cy5 on the C-terminus of uL18 were delivered in real-time (Figure 2A).

Figure 2. Real-time single-molecule analysis reveals m⁶A(−3) does not affect the rate of translation initiation.

(A) Schematic of the smFRET experiment. 43S PIC (including eIF1, eIF1A, eIF3, eIF5 and eIF2–GTP–Met-tRNA_i^Met and labeled 40S–Cy3 (green)), eIF5B-Cy3.5 (orange), and 60S–Cy5 (red) were added to m⁶A(−3), A(−3) or control m⁷G-A(−3) preincubated with saturating concentrations of eIF4A, eIF4B, eIF4G, eIF4E.

(B) Example smFRET traces of sequential association of the 43S PIC (green), eIF5B (orange) and the 60S subunit (red). t₁, 43S PIC association; t₂, eIF5B association; t₃, 60S subunit joining; t₄, eIF5B departure times from the 80S complex. The loss of 60S–Cy5 and 40S-Cy3 fluorescence was due to photobleaching of the dyes.

(C) Cumulative probability distributions of the observed steps of initiation on each mRNA (m⁷G-A(−3) mRNA, n=78; A(−3), n=84; m⁶A(−3), n=91). Data were fit to double exponential functions (dashed lines). Errors represent the 95% confidence interval of the fit. See also Figure S2 and Table S4.

The kinetics of four initiation steps (Figure 2B) were monitored: (i) Cy3-fluorescence appearance indicates 43S PIC loading onto mRNA; (ii) Cy3.5-fluorescence appearance tracks eIF5B association; (iii) Cy3-Cy5 Förster resonance energy transfer (FRET) monitors successful joining of the 60S; and (iv) loss of Cy3.5 fluorescence indicates eIF5B dissociation and successful assembly of the 80S elongation complex. We detected no significant difference in the observed rates of any of these discrete initiation steps between m⁶A(−3) and A(−3) variants (Figure 2C). Moreover, the kinetic constants were comparable to those with a control m⁷G-A(−3) mRNA (Figure 2C). The efficiencies of 40S subunit loading events leading to 80S complex formation were also similar across all mRNAs. We also mimicked the ISR activation and subjected the lysates to PKR treatment, which phosphorylates eIF2α. The kinetic constants and the efficiency of the 80S formation on each mRNA were undistinguishable from each other (Figure S2A) as those in the control (i.e. without PKR treatment; Figure 2).

To simulate the competitive environment in the cell, we also performed the real-time single molecule experiment without preincubation with eIF4 proteins, and added all initiation components simultaneously (Figure S2B,C). Again, we did not detect any differences in the kinetics of 40S loading and 60S joining onto m⁶A(−3) and A(−3) mRNA and the kinetic parameters were comparable to those measured with the control capped m⁷G-A(−3) mRNA (Figure S2D). Together, the real-time single molecule experiments suggest no effect of m⁶A on the kinetics of IC assembly and/or transition into elongation under control conditions or in conditions resembling ISR activation.

N⁶-methyl group does not significantly alter the conformation of the A(−3) purine ring

To probe the effect of m⁶A on IC structure, we next conducted structural investigation of 48S ICs with m⁶A(−3) and A(−3) mRNAs. We applied our previously described approach to stall and synchronize 48S ICs after start-codon recognition (i.e. designated late-stage 48S ICs (LS48S-ICs)) in vitro, in rabbit reticulocyte lysates³². We used A(−3) and m⁶A(−3) synthetic mRNAs, which are a shorter version of the reporters used in the in vitro translation assays (Figure 1C), i.e. covering 10 nts of the 5’UTRs and 28 nts of CDS of β-globin sequence (Table S2). The cryo-EM reconstructions of both A(−3) and m⁶A(−3) LS48S-ICs at resolutions near 3 Å (Figures S3A,B, and S4A–F and Table 1), displayed well-defined electron densities of eIF1A, ternary complex (tRNA_i^Met, eIF2 complex) and ABCE1, which all were bound on the 40S small ribosomal subunit (Figure 3A). Overall the structures of both A(−3) and m⁶A(−3) LS48S-ICs were identical (Figure S4A–F) and the densities for the mRNA nucleotide bases from (+1) to (+3) (i.e. AUG) were consistent with start-codon recognition (Figure S4J–L). We zoomed into the (−3) to (−1) positions observed subtle local differences between the densities surrounding adenosine at position −3 and cytidine at position −2 (Figure 3B, C). The unmodified A(−3) presented less density for the purine ring (Figure 3B). By contrast, the equivalent density of the purine ring in m⁶A(−3) LS48S-IC was slightly better defined (Figure 3C), which may reflect slight differences in the local densities (Figure S4A–F). At lower density threshold, the atomic models of the purine ring of the adenosine in A(−3) and m⁶A(−3) LS48S-ICs were compatible with anti or syn rotamer conformations (Figure 3D). In both anti or syn assignments some groups of atoms remained outside of the density contours (Figure 3D), implying the relatively high rotational flexibility of the A(−3) purine ring with and without the methyl group. At higher density threshold, the anti and syn conformations seem to be more likely for m⁶A(−3) and A(−3), respectively (Figure 3D), however, some atomic groups still fell outside the density contours. Because of the flexibility of the A(−3) purine ring with and without m⁶A, no precise rotamer conformation could be assigned. Furthermore, we could not assign with certainty the precise orientation of the N⁶-methyladenosine (Figure 3C,D). Even at lower threshold no density could be assigned for the N⁶-methyl group of m⁶A(−3), because of the unrestricted rotational freedom of the methyl group.

Table 1.

Cryo-EM data collection, refinement and validation statistics.

Model	A(−3) LS48S-IC	m6A(−3) LS48S-IC
Data collection and EM reconstruction
Microscope	TFS TALOS ARCTICA F200C
Voltage (kV)	200
Camera	Gatan K2 Summit (4k x 4k)
Magnification	36,000
Defocus range (μm)	−0.5 to −2.5
Calibrated pixel size	1.16	1.16
Electron exposure ( e−/Å2)	45
Exposure time (s)	6.5
Number of frames per movie	65
Automation software	SerialEM
Number of micrographs	10,707	11,594
Initial particle number	2,529,429	1,683,066
Final particle number	74,864	103,050
Map sharpening B factor (Å2)	−61.9	−54.9
Map resolution (FSC=0.143)	3.257	3.04
Refinement
Composition
Chains	41	41
Atoms (no H)	86,313	86,354
Residues (amino acids)	5904	5904
Residues (nucleotides)	1825	1827
Water	0	0
Bonds (RMSD)
Ligand (type)	230 (Mg)	227 (Mg)
Length (Å)	0.011	0.008
Angles (Å)	1.474	1.310
MolProbity score	2.76	2.51
Clash score	7.75	6.68
Ramachandran plot (%)
Outliers	0.1%	0.1%
Allowed	11.9%	8.9%
Favored	88.1%	91.0%
Rotamer outliers (%)	9.32%	6.23%
Cβ outliers (%)	0.22%	0.33%
Peptide plane (%)
Cis proline/general	0.9/0.0	0.4/0.0
Twisted proline/general	0.0/0.0	0.0/0.0
C-BLAM outliers (%)	6.57	5.92
ADP (B-factors) (Å2)
Iso/aniso min/max/mean	86253/0	86271/0
Protein	41.3/316.2/95.4	11.7/158.3/58.8
Nucleotide	0.0/391.7/88.2	0.0/237.6/60.2
Ligand	33.0/147.4/61.2	13.9/126.0/34.6
Occupancy (mean)	0.98	0.98
Box
Lengths (Å)	231.0, 237.6, 225.5	231.0, 237.6, 227.7
Angles (°)	90, 90, 90	90, 90, 90
Resolution estimates (Å)
d 99 (full)	3.42 (masked), 3.41 (unmasked)	3.15 (masked), 3.15 (unmasked)
d model	3.3 (masked), 3.3 (unmasked)	3.1 (masked), 3.1 (unmasked)
d FSC model (0/0.143/0.5)	3.02/3.26/3/33 (masked), 3.09/3.27/3.35 (unmasked)	2.70/3.04/3.12 (masked), 2.78/3.06/3.13 (unmasked)
Map min/max/mean	−0.06/0.11/0.00	−0.08/0.16/0.00
Model vs. Data
CC(mask)	0.84	0.83
CC (box)	0.83	0.81
CC (peaks)	0.77	0.78
CC (volume)	0.84	0.82
Mean CC for ligands	0.74	0.72

Figure 3. Structures of LS48S-IC with m⁶A(−3)-mRNA and A(−3) mRNA.

(A) Segmented cryo-EM reconstruction of the m⁶A(−3) mRNA LS48S-IC.

(B, C) Zoom-in into the interactions between the 18S rRNA (G957) and mRNA at position −3, in both LS48S-ICs with A(−3) mRNA (B) and m⁶A(−3) mRNA (C). Density maps were processed with DeepEMhancer at 13 sigma threshold.

(D) Zoom into the density maps surrounding mRNA position (−3) of both complexes with m⁶A (upper row) and A (lower row) at 3.5 and 6 sigma thresholds. The atomic models fitted to anti (left) or syn (right) rotamer conformations.

See also Figure S3 and S4.

To assess the effect of m⁶A on the LS48S-IC assembly under stress, we solved the structure of the m⁶A(−3) LS48S-IC following treatment with PKR³⁰, which phosphorylated eIF2α (eIF2α-P; Figure S1F) and thus, mimicked the stress-triggered ISR activation in cells. Similar number of particles was acquired for m⁶A(−3) LS48S-IC eIF2α-P (Figure S3C), however the local resolution analysis presented clear indications of higher flexibility (Figure S4G–I), probably due the overall lower efficacy of the LS48S-IC formation with phosphorylated eIF2α. The cryo-EM reconstruction of the m⁶A(−3) LS48S-IC eIF2α-P closely resembled that with the unphosphorylated m⁶A(−3) LS48S-IC, at position −3 of the mRNA (Figure S5A) and start codon recognition (Figure S4M). Similarly to m⁶A(−3) LS48S-IC, no density could be assigned for the N⁶-methyl group of the m⁶A(−3) LS48S-IC eIF2α-P. A precise anti or syn rotamer conformation could not be assigned to the m⁶A(−3) purine ring (Figure S5B). Even at higher density threshold, some atomic groups were outside the density contours (Figure S5B), implying its high flexibility. Overall, our structural data suggest that the N⁶-methyl group does not substantially affect the conformation of the m⁶A purine ring.

Parsing the energetic contribution of the stacking interaction with Arg55

The further examination of the atomic models of A(−3) and m⁶A(−3) LS48S-ICs revealed subtle differences in the local conformations of eIF2α and uS7 (Figure 4A–C and S5C,D). In m⁶A(−3) LS48S-IC, we observed a discontinuous local electron density of a segment of loop 3 from domain 1 of eIF2α (i.e. residues 56–58; Figure S5C), suggesting higher flexibility of this loop in presence of m⁶A. The same region presented more solid density in the A(−3) LS48S-IC (Figure S5D). Another noticeable conformational rearrangement was at residue R55 of eIF2α. In the m⁶A(−3) LS48S-IC, the position of the R55 side chain was compatible with establishing stacking interactions with m⁶A (Figure 4D). By contrast, in A(−3) LS48S-IC R55 displayed densities that suggest different interactions, i.e. the distance would be compatible with a hydrogen bond with NH₂ at position 6 of the purine ring of A(−3) and the oxygen atom of the peptide bond of G147 of uS7 (Figure 4E).

Figure 4. Interaction network of eIF2α and structural comparison and relative binding free energies of R-m⁶A and R-A.

(A) Atomic model of LS48S-IC with m⁶A(−3)-mRNA.

(B, C) Molecular contacts surrounding m⁶A(−3) (B) or A(−3) mRNA (C). Magenta arrow points at the structural differences of R55 from eIF2α and m⁶A(−3). Dashed lines and residue numbers in cyan highlight the gaps in the atomic models.

(D, E) Zoom into the interaction mode models of R55 with m⁶A(−3) (D) and A(−3) mRNA (C). (F) Superposition between experimental (exp.) R55-A(−3) and R55-m⁶A(−3) pairs with representative snapshots from the MD simulations (median RMSDs < 1.5 Å among MD conformers). The mesh captures the van der Waals surface.

(G) Relative binding free energies between all amino-acid side chains (except Pro and Gly) and m⁶A versus A in methanol ($\text{ΔΔ}{G}_{m6A-A}^i$). Side chains with a preference for m⁶A (green) and for A (red). See also Figure S5.

We next sought to quantify the energetic contribution of the stacking interaction between R55 of eIF2α and m⁶A(−3). To determine to what extent the adenine methylation reflects the intrinsic interaction preferences of arginine (R) side chain, we analyzed previously reported molecular dynamics (MD) trajectories of binding free energies between the side chain of a free arginine and unmodified A or m⁶A nucleobase^33,34; the trajectories were calculated using umbrella sampling in methanol as a model for the low dielectric environment seen at typical binding interfaces. In the LS48S-ICs atomic models, the relative orientation of R55 with respect to both A(−3) and m⁶A(−3) resembled the behavior of the arginine side chain when interacting with unmodified A or m⁶A nucleobase in isolation (Figure S5E,F). For example, 31% of the simulated conformers with R interacting at the experimentally observed center-of-mass distance with unmodified A and 25% with m⁶A exhibited an all-heavy-atom RMSD<1.5 Å from the respective A(−3) or m⁶A(−3) LS48S-ICs conformations (Figure S5E, F). Representative snapshots with median RMSDs (i.e. RMSD_A-R = 1.19 Å and RMSD_m6A-R = 1.20 Å) matched well with R55-A(−3) or R55-m⁶A(−3) LS48S-ICs (Figure 4F).

We next analyzed the binding free energies between all non-prolyl/non-glycyl amino-acid side chains (i) and m⁶A or A nucleobases ($\text{ΔΔ}{G}_{m6A-A}^i=\text{Δ}{G}_{methanol}^{m6A,i}-\text{Δ}{G}_{methanol}^{A,i}$) (Figure 4G and^33,34). The arginine side chain exhibited the most negative relative binding free energy or the highest discrimination between m⁶A and A ($\text{ΔΔ}{G}_{m6A-A}^{Arg}=-0.8kJ/mol$) (−0.2 kcal/mol; 1 J = 0.239 cal; Figure 4G). Thus, through a favorable stacking interaction between the methyl group of m⁶A and R55, a subtle stabilization of m⁶A(−3) LS48S-IC could be achieved, which, however, is much lower in magnitude than the contribution of a single hydrogen bond (−6 to 30 kJ/mol)^35,36.

In sum, these results suggest that m⁶A triggers some dynamic compensatory rearrangements. Although the newly-established stacking interactions with R55 of eIF2α introduce a subtle thermodynamic stabilization of the m⁶A(−3) LS48S-ICs, this stabilization might be traded off with a loss of interaction with G147 of uS7 and destabilization of loop 3 in eIF2α domain 1, which we detected as absent electron density (Figure 4B, C), thus resulting in overall null net effect.

DISCUSSION

N⁶-methyladenosine modification is asymmetrically distributed across transcripts with several-fold higher frequency towards the end of the coding sequence and 3’ UTR than in the 5’UTRs^6,11–13,37. Despite its low abundance, an earlier study featured a single m⁶A modification in the 5’UTRs as a stand-alone mediator of IC assembly¹⁹. Using diverse experimental techniques, combining structural, biochemical and single-molecule data, we show that the m⁶A at position −3 (i.e. at the most prevalently modified site in 5’UTRs) does not affect the kinetics of IC assembly, start site recognition and translation efficiency in general. The cryo-EM structures reveal stacking interactions of the m⁶A-modified purine ring with the R55 side chain of eIF2α. Among proteinogenic amino acids (except Pro and Gly), arginine side chain exhibits the most favorable interaction with m⁶A nucleobase, with a relative free energy contribution of −0.8 kJ/mol (−0.2 kcal/mol). This subtle m⁶A stabilization of LS48S-ICs was not detected in the smFRET experiments. It is likely that this subtle m⁶A-driven stabilization of LS48S-ICs might be traded off with a loss of interaction with uS7 and the gain of flexibility of loop 3 from domain 1 of eIF2α (i.e. residues 56–59), culminating in no kinetic effect on IC assembly and start site recognition, and the overall translation efficiency under homeostatic conditions and stress. smFRET captures the forward kinetics of IC assembly, whereby the off-rate and LS48S-ICs dissociation remains invisible. Thus, the marginal m⁶A-associated stabilization may confer some benefit and through a decrease of the drop-off rate, would enhance the number of ICs transitioning into elongation. However, this scenario is unlikely as we detected no expression benefit of a single m⁶A under permissive growth conditions and stress.

N⁶-methyladenosine modification exhibits diverse and sometimes opposing effects on mRNA, dependent on whether it occurs in structured mRNA regions or single-stranded stretches. Compared to unmodified adenosine, m⁶A disfavors base pairing with uridines and destabilizes local secondary structures. Depending on the m⁶A localization within the secondary structure, i.e. engaged in base pairs in the middle or at the extremities of secondary structures, m⁶A can destabilize duplexes by 0.5–1.7 kcal/mol ^38,39. By contrast, in single-stranded RNAs methylamino group shows much higher preference for syn rotamer, which is energetically more favored by appr. 1.5 kcal/mol, but interferes with the Watson-Crick interbase hydrogen bonding^40,41. The new stacking interactions between the m⁶A(−3) purine ring and the arginine residue (R55) of eIF2α resemble the regular conformation observed when A is involved in Watson-Crick base pairing with U^39,41. However, this stabilizing thermodynamic contribution of 0.2 kcal/mol is at least an order of magnitude lower than the energetic contribution of the stacking interactions between the m⁷G cap structure and two Trp residues in eIF4E playing a pivotal role in the eIF4E-cap association⁴². Thereby, the main gain in stacking energy is reached by the positive charge of the m⁷G and not by the N⁷-methyl group⁴². The comparatively low stabilization energy of the stacking R55-m⁶A we observed, might be dynamically compensated by the energy penalties by breaking the hydrogen bond with G147 of uS7 (Figure 4D,E), thus resulting in an overall similar flexibility of the purine ring of the unmodified and modified adenosine. It is unlikely that such dynamic and likely transient interactions would significantly stabilize the entire multicomponent LS48S-IC.

Limitations of the study

Our integrative study, combining in-cell deep-sequencing analysis and in vitro structural, biochemical and single-molecule approaches, shows that a single m⁶A modification in the 5’ vicinity of the start codon does not affect the overall translation and the effect is far more subtle than initially proposed^19,24. We cannot exclude a cumulative effect of clustered m⁶A sites. Within the 30nt upstream of the ATG, only 3 transcripts have clustered m⁶A²³, suggesting that if multiple m⁶A sites cumulatively affect initiation, the effect will be restricted to only few transcripts. In regions much further upstream of the start codon, 87 transcripts ²³ display clustered m⁶A sites. Further experiments are needed to determine whether those distant clustered modifications do affect early initiation processes, e.g. scanning. Furthermore, estimates of the relative binding free energies between amino-acid sidechains and A or m⁶A were obtained in methanol as a mimetic of a low-dielectric binding interface, and may be different in other environments. The overall ranking of sidechains, however, likely does not significantly differ from what was reported here (Figure 4G). Prior studies have suggested that m⁶A in the 3’UTRs and not in the 5’UTRs facilitate translation by YTHDF1- and/or YTHDF3-mediated interactions with eIF3^10,20. The role of 3’ m⁶A on translation is a potentially appealing idea given that 3’UTRs are more prevalently modified, as steric occlusion by exon-junction complexes depletes 5’UTRs and CDS of methylation^11,12. Further experiments are likely to reveal the crucial role 3’ m⁶A in translation initiation, while structural studies in cellular environments would reveal the molecular gymnastics of mRNA looping and the cascade of interactions between m⁶A readers, eIF3 and initiating ribosomes.

STAR METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Zoya Ignatova (zoya.ignatova@uni-hamburg.de).

Materials availability

Reagents and materials produced in this study are available from the lead contact Zoya Ignatova (zoya.ignatova@uni-hamburg.de).

Data and code availability

Correspondence and requests for materials should be addressed to Zoya Ignatova (zoya.ignatova@uni-hamburg.de). All sequencing data are available in the Gene Expression Omnibus under accession number GSE201064 (eTAM sequencing²³), and GSE210563 (GLORI-seq⁹), and in the BioSample database of Sequence Read Archive (NCBI) under the accession number SRP121376 (m⁶A-seq²²). The electron densities are deposited in the EMDB under the accession codes: EMD-17330, EMD-17329 and EMD-18510, for the A(−3), m⁶A(−3) and m⁶A(−3) with phosphorylated eIF2α, respectively. The molecular models derived from the maps of the initiation complexes with A(−3) and m⁶A(−3) are available under the PDB codes 8P09 and 8P03, respectively. Accession numbers are listed in the key resources table.

KEY RESOURCE TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
GMP-PNP	Sigma	Cat # G0635
TSY	Pacific Biosciences	100–214-900
Cy3 mono hydrazide	Cytiva	PA13121
Cyanine 3.5 maleimide	AAT Bioquest	Cat # 149
Cy5 mono hydrazide	Cytiva	PA15121
2’-O-methyltransferase	NEB	Cat # M0336
T4 RNA ligase	ThermoFisher Scientific	Cat # EL0021
T4 RNA ligase 2	NEB	Cat # M0242S
Terminator™ 5ʹ-Phosphate-Dependent Exonuclease	Lucigen	Cat # TER51020
RNA 5’ polyphosphatase	Lucigen	Cat # 136120
RNasin® Ribonuclease Inhibitors	Promega	Cat # N251B
DNase I	ThermoFisher Scientific	EN0521
MssI FD	Therno Scientific	Cat # FD1344
Protein kinase R (PKR)	Abcam	Cat # ab32052
Critical commercial assays
T7 megascript kit	Invitrogen	Cat # AM1334
Vaccinia Capping System	NEB	Cat # M2080S
HiScribe RNA synthesis kit	NEB	Cat # E2040S
OneStep Human Coupled IVT Kit	ThermoFisher Scientific	Cat # 88881
Luciferase Assay System	Promega	Cat # E1500
Rabbit reticulocyte lysate	Promega	Cat # L4960
Deposited data
m6A-seq	22	SRP121376
GLORI-seq	9 (Table S1,S5)	GSE210563
eTAM-seq	23	GSE201064
A(−3) mRNA cryo-EM map	This paper	EMDB: EMD-17330
m6A(−3) mRNA cryo-EM map	This paper	EMDB: EMD-17329
m6A(−3) mRNA with P-eIF2α cryo-EM map	This paper	EMDB: EMD-18510
mRNA reconstruction atomic model m6	This paper	PDB: 8P09
m6A(−3) mRNA reconstruction atomic model	This paper	PDB: 8P03
Oligonucleotides
5’ UTR sequences	This paper	Table S1
Primers for qRT-PCR	This paper	Table S1
Recombinant DNA
Plasmid pUC	Addgene	Cat #50005
Plasmid pGL4.51	Promega	Cat # E1320
Plasmid pUC 5’UTR-less	This study	N/A
Plasmid pUC A(−3)-Luc with different 5’UTRs	This study	N/A
Software and algorithms
MATLAB R2021	Mathworks	N/A
RELION	43	https://www2.mrc-lmb.cam.ac.uk/groups/scheres/impact.html
RELION 3.1	56	https://www2.mrc-lmb.cam.ac.uk/groups/scheres/impact.html
MotionCor	44	https://emcore.ucsf.edu/ucsf-motioncor2
Chimera, ISOLDE package	45	https://www.cgl.ucsf.edu/chimera/
DeepEMancer	46	https://tristanic.github.io/isolde/
Bowtie2	47	https://bowtie-bio.sourceforge.net/bowtie2
Cutadapt v1.8.3	48	https://cutadapt.readthedocs.io/en/stable/
STAR	49	https://github.com/alexdobin/STAR
UMItools	50	https://umi-tools.readthedocs.io/en/latest/index.html
SAMtools	51	http://www.htslib.org
HISAT-3N	52	http://daehwankimlab.github.io/hisat2/hisat3n/
HOMER	53	http://homer.ucsd.edu/homer/index.html

• This paper does not report original code.

• Any additional information required to reanalyze the sequencing data is available from Zoya Ignatova (zoya.ignatova@uni-hamburg.de) upon request.

• Any additional information on the structural data is available from Yaser Hashem (yaser.hashem@u-bordeaux.fr) upon request.

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Experimental source material for in vitro work

In vitro translation assays were performed in HeLa cell extracts (One-Step Human Coupled IVT Kit, ThermoFisher Scientific). Cryo-EM samples were prepared from nuclease-treated rabbit reticulocyte lysate (RRL) (Promega L4960). GLORI-seq and m⁶A-seq were performed from HEK293 cells and eTAM-seq from HeLa cells.

METHOD DETAILS

Luciferase reporter constructs

The reporter constructs with unmodified adenine (A(−3)-Luc) contain 51 nt of the 5’ UTR the first 23 nt of either β-globin, TSC22D4, GOLIM4 or RP2 CDS (BC007075) (Table S2), followed by the ORF of firefly luciferase (luc2 derived from pGL4.51 (Promega)) without the first 7 nt to maintain the correct in-frame ORF. To obtain a reporter containing a single m⁶A at position 3 upstream of the start codon (m⁶A(−3)-Luc), we created a construct lacking the 5’UTR (5’UTR-less), comprising the luc2 without the first 7 nt. All reporter constructs were cloned into pUC19 vector under the control of the T7 promoter required for the in vitro transcription. The in vitro transcription was performed with plasmids linearized with MssI FD (ThermoFisher Scientific) using the HiScribe RNA synthesis Kit (NEB) for 2 h at 37°C. Following transcription, the reactions were subjected to DNase I (ThermoFisher Scientific) treatment for 15 min at 37°C to digest the DNA template. The mRNA was purified using Monarch RNA column cleanup kits (NEB).

RNA oligonucleotides representing part of the 5’UTR and CDS of β-globin TSC22D4, GOLIM4 or RP2 CDS (BC007075) (Table S2) with a single m⁶A modification at position −3 were synthesized by Horizon Discovery Biosciences Ltd. The purchased RNA oligonucleotides have a monophosphate group at the 5’ terminus. To obtain m⁶A(−3)-Luc mRNA, the m⁶A oligonucleotide was ligated to the in vitro transcribed 5’UTR-less luc mRNA and an in vitro transcribed fragment with the remaining of the 5’UTR. The in vitro transcription of the 5’UTR-less luc fragment was performed in presence of 8 mM GMP and 2 mM GTP to obtain 5’-monophosphate-modified fragments. The ligation was facilitated by a 80-mer splint DNA oligonucleotide designed to base pair to the complete m6A oligonucleotide and at least 23 nucleotides on each side RNA. The three RNA fragments, 5’UTR transcript, m⁶A(−3) RNA oligonucleotide, 5’UTR-less luc transcript and the splint DNA, were mixed in 2:1:1:2 ratio and incubated with T4 RNA ligase 2 (NEB) for 2h at 37 °C. For β-globin m⁶A(−3)-Luc mRNA, the ligation was done only between a modified m⁶A oligonucleotide containing the complete 5’UTR, and the 5’UTR-less luc fragment using a 58-mer splint DNA, base pairing 29 nt on each side RNA. This RNA oligonucleotide has unmodified 5’ terminus (i.e. 5’OH). Thereafter, the splint DNA was digested with DNase I (ThermoFisher Scientific) for 1h at 37 °C and the ligated RNA product was purified using RNA Clean & Concentrator columns (Zymo Research). The efficiency of the ligation was determined by qRT-PCR using a single reverse primer within the CDS of luc2 and different forward primers (Table S2). The efficiency of the ligation reaction was 15–50% and was determined for every batch. The m⁷G-m⁶A(−3)-Luc and m⁷G-A(−3)-Luc mRNA were created using the Vaccinia Capping System (NEB) for the β-globin constructs and the CleanCap system (NEB) for the endogenous constructs following the manufacturer’s protocol. Uncapped transcripts were digested with RNA 5’ polyphosphatase (Lucigen) for 30 min at 37°C, followed by a treatment with a Terminator^™ 5’-phosphate-dependent exonuclease (Lucigen) for 1h at 30°C. The efficiency of the capping reaction was ~95%.

In vitro translation

In vitro translation assays were performed using HeLa cell extracts (One-Step Human Coupled IVT Kit, ThermoFisher Scientific). In vitro translation reactions were performed in a total volume of 12.5 μL using the protocol suggested by the manufacturer. All reactions were supplemented with 1 mM MgCl₂ and performed with 50 ng of the corresponding RNA template (~7 nM) at 30 °C for 1h. The reaction was terminated by adding 200 μM cycloheximide and immediate cooling on ice. Luciferase expression was quantified using Luciferase Assay System (Promega) in a 96-well plate and the luminescence was recorded with a TECAN plate reader. In each biological replicate, each single measurement was performed in three technical replicates and the averaged luminescence signal was normalized to the signal with an unmodified template (A(−3)-Luc) which was always run simultaneously to each measurement. In some reactions, to test the effect of eIF2α inhibition, human PKR (Abcam) was added to the in vitro translation reaction at a concentration of 65 nM ³⁰.

Single-molecule translation initiation mRNA constructs

DNA template encoding the A(−3) RNA sequence was amplified from a synthetic DNA encoding the human β-globin transcript 1 (NM_000518.5) using forward primer and reverse primers listed in Table S2. The RNA fragment was transcribed in vitro using the T7 Megascript kit (Invitrogen) following the manufacturer’s protocol and purified by phenol-chloroform extraction and ethanol precipitation. The RNA fragment was capped using the Vaccinia Capping System (NEB) and 2ʹ-O-methyltransferase (NEB) following the manufacturer’s one-step protocol and purified again by phenol-chloroform extraction and ethanol precipitation. An RNA oligonucleotide encoding a portion of the S. cerevisiae RPL30 open reading frame with sequence (Table S2) and with 5ʹ-monophosphate and 3ʹ-biotin modifications was purchased from IDT. Capped, uncapped, and m⁶A(−3) RNA fragments (5 μM) were ligated to the RPL30 fragment (10 μM) by incubating with 1 U/μL T4 RNA Ligase (ThermoFisher Scientific) and 20% (w/v) PEG-6000 in the manufacturer’s supplied reaction buffer at 30 °C overnight. Ligation products were treated with Terminator^™ 5ʹ-Phosphate-Dependent Exonuclease (Lucigen) following the manufacturer’s protocol to deplete excess unligated RPL30 oligonucleotide. Ligated mRNAs were purified by phenol-chloroform extraction and ethanol precipitation and stored at −20 °C until use.

Real-time initiation single-molecule assays

All experiments were performed as previously described³¹ using a modified Pacific Biosciences RSII DNA sequencer⁵⁴. Fluorescently labeled ribosomal subunits and initiation factors were purified and prepared as previously described³¹. The reaction buffer for all assays was 20 mM HEPES-KOH, pH 7.3, 70 mM KOAc, 2.5 mM Mg(OAc)₂, 0.25 mM spermidine, 0.2 mg/mL creatine phosphokinase, 1 mM ATP·Mg(OAc)₂, 1 mM GTP·Mg(OAc)₂. Imaging buffer was the reaction buffer supplemented with 62.5 μg/mL casein, 2 mM TSY (Pacific Biosciences), 2 mM protocatechuic acid, and 0.06 units/μL protocatechuate-3,4-dioxygenase. Briefly, 3ʹ-biotinylated mRNA was tethered to a neutravidin-coated zero-mode waveguide (ZMW) chip (Pacific Biosciences). For experiments in which the mRNA was preincubated with eIF4 proteins, the surface was washed with the reaction buffer and incubated with a mixture of 2 μM eIF4A, 440 nM eIF4B, 320 nM eIF4E, and 260 nM eIF4G in imaging buffer for 5–10 min at room temperature. At the start of data acquisition, an equal volume of a mixture containing 10 nM 43S PIC, 80 nM eIF5B-Cy3.5, and 200 nM 60S-Cy5 subunits in imaging buffer was added to the surface. As a result, final concentrations of factors during the movie acquisition were half of the values reported above. For experiments with phosphorylated eIF2α, 65 nM human PKR (Abcam) and 1 mM ATP were added during ternary complex formation. For experiments in which eIF4 proteins were delivered alongside other factors, after mRNA tethering the surface of the ZMW chip was washed with the reaction buffer and incubated with imaging buffer for 5–10 min at room temperature. At the start of data acquisition, an equal volume of a mixture containing 2 μM eIF4A, 440 nM eIF4B, 320 nM eIF4E, 260 nM eIF4G, 10 nM 43S PIC, 80 nM eIF5B-Cy3.5, and 200 nM 60S-Cy5 subunits in imaging buffer was added to the surface. As above, final concentration of all factors are halved from the reported values during data collection. All experiments were performed at 30°C using a 532 nm excitation laser at 0.32 μW/μm², which directly excited Cy3 and Cy3.5 dyes. Cy5 dye was excited via FRET with Cy3. Fluorescence emission was detected at 10 frames per second for 600 s.

Kinetic analysis of the single-molecule traces

Experimental movies that captured fluorescence intensities over time were processed using MATLAB R2021-a as described previously³¹. After filtering ZMWs for desired fluorescence signals the binding events of individual components were assigned manually based on the appearance and disappearance of the respective fluorescence signals. The observed times for events were used to calculate the cumulative probability functions which were fit to single- or double-exponential functions to obtain association and dissociation kinetics. All custom code used to process and analyze single-molecule data are available publicly under an open-source license at https://github.com/puglisilab/Lapointe-2022-Nature.

Sample preparation for Cryo-EM

For cryo-EM, two uncapped oligonucleotides (38-mer) either unmodified (A(−3)) or m⁶A modified (m⁶A(−3)) were synthesized by Dharmacon (Table S2). Both nucleotides resemble 10 nt of the 5’UTR and 28 nt of the CDS of β-globin and are, thus, identical to the constructs used in the in vitro translation assays and single-molecule studies. The LS48S-IC with A(−3) and m⁶A(−3) uncapped mRNA fragments were isolated from nuclease-treated rabbit reticulocyte lysate (RRL) (Promega L4960), according to the protocol described in ³² with some modifications. Briefly, 70 μl of RRL with 20 μl of the amino acid mixture (a final concentration of 0.13 mM) and 80 units of RNasin (Promega) were incubated at 30°C for 5 min, followed by 10 min on ice. Guanylyl imidodiphosphate (GMP-PNP) was then added to a final concentration of 10 mM and incubated for 2 min at 30°C: Thereafter, 11 μM of each previously denatured mRNA fragment (5 min at 65°C) was added to a final reaction volume of 150 μl, incubated for 5 min at 30°C and cooled for 10 min on ice. The reactions in duplicate were loaded on top of 5%–25% linear sucrose gradient (in 25 mM HEPES-KOH pH 7.6, containing 79 mM KOAc, 2 mM Mg(OAc)₂, and 2 mM DTT): The assembled complexes were separated by centrifugation at 37,000 rpm in a SW41Ti rotor for 4.5h at 4°C. LS48S-ICs were collected using BioComp Piston Gradient Fractionator^™ device by monitoring 260 nm absorbance. Fractions containing LS48S-ICs were pooled together, centrifuged at 108,000 rpm (S140AT Sorvall-Hitachi rotor) for 2 h at 4°C and the ribosomal pellet was dissolved in a 10 mM HEPES-KOH pH 7.4, containing 50 mM KOAc, 10 mM NH₄Cl, 5 mM Mg(OAc)₂ and 2 mM DTT. For the 48S IC with phosphorylated eIF2α, the RRL extract was pre-incubated incubated with PKR for 30 min at 30°C prior to adding the mRNA variants.

Grids preparation and data collection parameters

The grids were prepared by applying 4 μL of each complex at ~200 nM to 400 mesh holey carbon Quantifoil 2/2 grids (Quantifoil Micro Tools). The grids were blotted for 1.5 sec at 4°C, 100% humidity, using waiting time 30 s, and blot force 5 (Vitrobot Mark IV). The data acquisition was performed on a Titan Krios S-FEG instrument (FEI) operated at 300 kV acceleration voltage and at a nominal underfocus of Δz = ~ 0.5 to ~ 3.5 μm using the CMOS Summit K2 direct electron detector 4,096 × 4,096 camera and automated data collection with SerialEM at a nominal magnification of 59,000x. The K2 camera was used at super-resolution mode and the output movies were binned twice, resulting in a pixel size of 1.1Å at the specimen level (the calibrated magnification on the 6.35 μm pixel camera is 115,455 x). The camera was set up to collect 20 frames and frames 3 to 20 were aligned. Total collected dose is ~26 e⁻/Ų.

Image processing

MotionCor ⁴⁴ was used for the movie alignment of the movie frames from the LS 48S-ICs with m⁶A(−3) mRNA and the LS 48S-IC with non-methylated A(−3) mRNA and the LS 48S-ICs with m⁶A(−3) mRNA after phosphorylation of eIF2α. CTFFIND4 ⁵⁵ was used for the estimation of the contrast transfer function of an average image of the whole stack. Particles selection and further particle processing was performed in RELION 3.1 ⁵⁶. Approximately 103,000 particles were selected for m⁶A(−3) LS48S-IC, 75,000 particles for A(−3) LS48S-IC, and 78,000 particles for m⁶A(−3) mRNA eIF2α-P. RELION ⁴³ was used for particle sorting through 3D classification. Selected classes were refined using RELION’s 3D autorefine and the final refined classes were then post-processed using the procedure implemented in RELION applied to the final maps for appropriate masking, B-factor sharpening, and resolution validation to avoid over-fitting. The atomic model from ³² was used as a starting model in both LS48S-ICs with m⁶A(−3) mRNA and A(−3) mRNA cryo-EM maps. The m⁶A(−3) with eIF2α-P cryo-EM data processing yielded fewer LS 48S-ICs and thus, required the acquisition and the subsequent merging of three datasets (Figure S3). After particles sorting and 3D reconstruction, the map displayed higher conformational heterogeneity, as suggested by the more variable local resolution levels (Figure S3C), compared to the other two maps, probably because of the lower efficacy of 48S-IC formation due to the eIF2α phosphorylation. In order to avoid any over interpretation of our map, the final refined model of the m⁶A(−3) mRNA was rigid-body fitted into the m⁶A(−3) with eIF2α-P cryo-EM reconstruction to show the similar overall conformation, in particular the m⁶A(−3) base conformation. The mRNA sequences was built manually into the map using Chimera UCSF ⁴⁵ using ISOLDE package ⁴⁶. The final cryo-EM reconstruction of both LS 48S-ICs with m⁶A(−3) mRNA and with A(−3) mRNA were also enhanced using DeepEMancer ⁵⁷ and were used only as display items for Figure 3B,C.

Conformational comparison of simulated and experimental A-Arg and m⁶A-Arg pairs

The trajectories were calculated using umbrella sampling in methanol as a model for the low dielectric environment seen at typical binding interfaces. The conformations of the experimental A-Arg55 and m⁶A-Arg55 pairs in the LS48S-ICs structures were compared with umbrella-sampling MD trajectories of isolated Arg and A^33,34 or m⁶A pairs³⁴ in methanol using all-heavy-atom (i.e. all atoms except H) positional root-mean-square deviation (RMSD) after all-atom superposition. Simulated ensembles were first grouped according to the centre-of-mass distance between Arg and A or Arg and m⁶A groups in windows of the size [d-0.25 Å, d+0.25 Å] with d going from 3.75 to 6.25 in steps of 0.50 Å. The fraction of structures below a given RMSD threshold in a given distance group is reported for A (Figure S5E) and m⁶A (Figure S5F), while representative conformers at the center-of-mass distance window corresponding to the experimentally observed structures are given in Figure 4F. For further details regarding umbrella sampling simulations, please see the original studies ^33,34.

m⁶A-sites identification in deep sequencing data

The following published data sets were used: (i) immunoprecipitation-based m⁶A-seq of HEK293 cells grown under permissive conditions and exposed to moderate oxidative stress (200 μM) for 30 min ²², and (ii) glyoxal and nitrite-mediated deamination of unmethylated adenosine (GLORI-seq) for quantitative m⁶A detection of HEK293 cells grown under permissive conditions and exposed to hypoxia ⁹.

In the m⁶A-seq, the identification of methylation sites was performed individually for each replicate as described ⁵⁸. Briefly, each sequencing set consist of two libraries, i.e. RNAs immunoprecipitated with m⁶A-antibody (IP sample) and total RNA before subjected to immunoprecipitation (input sample). For each transcript in the IP sample, we calculated the peak over median (POM) using a sliding window of 50 nt (with a 25-nt overlapping step), and calculated the ratio between the mean coverage within the sliding window to the median coverage across the entire transcript. Windows with minimal mean coverage of 10 and POM score higher than 3 were retained for the downstream analysis. The POM score calculation was also performed for the input sample. We discarded all shared regions between the IP and input samples. Subsequently, we calculated the peak over input (POI) score, which derives from the ratio of the POM scores of the IP and input samples. The peaks identified individually for each replicate were merged into a single list. Lastly, we identified the DRACH motifs within the reference transcriptome using HOMER algorithm ⁵³. Regions with no DRACH motifs were discarded. Transcripts containing minimum one methylated DRACH motif were considered as methylated. Transcripts with no peaks within a DRACH motif were considered as non-methylated. The methylation levels (Figure 1A) represent the ratios of the methylated DRACH motifs and the total number of DRACH motifs within a given transcript. In the GLORI-seq, we used the already mapped and identified m⁶A sites (listed in the supplementary table 1 and 5 ⁹). We omitted the m⁶A sites in non-annotated regions, and for multiple isoforms of the same transcript we merged them into the longest transcript. In total, 113,744 and 74,067 m⁶A sites were detected within annotated transcripts in the control and hypoxia-exposed cells, respectively. The evolved TadA-assisted N⁶-methyladenosine sequencing (eTAM-seq) data set was downloaded from GEO (GSE201064) and processed as described in ²³. We followed the protocol available in the GitHub repository (https://github.com/shunliubio/eTAM-seq_workflow). Briefly, sequenced reads were depleted from adapters using Cutadapt v1.8.3⁴⁸ and UMI-tools⁵⁰. Reads longer than 39 nt were mapped to the human reference genome (GRCh38.p13) using HISAT-3N⁵². The uniquely mapped reads were deduplicated and kept for further analysis. The replicates 1 and 2 were selected for analysis and to filter true methylation sites we followed the custom scripts used in ²³ and available in the GitHub repository (https://github.com/shunliubio/eTAM-seq_workflow). In total 23,153 m⁶A sites were detected among the annotated transcripts for HeLa mRNA.

We note that while m⁶A level at positions −3 and −2 in the 5’UTRs was the highest in both GLORI-seq and eTAM-seq, the distribution slightly differs between the two sequencing approaches. In GLORI-seq 74% of the detected transcripts were methylated at position −3, and 26% at position −2. By contrast, eTAM-seq detected 62.5% transcripts with methylation at position −2 and 37.5% at position −3. Overall, because of the higher depth of the GLORI-seq, cumulatively for both sequencing methods the m⁶A at position −3 was more abundant.

Identification of actively translated genes under oxidative stress

Ribosome profiling (Ribo-seq) and RNA-seq data sets of HEK293 cells culture at permissive conditions or exposed to moderate oxidative stress (200 μM AS) are published in ²². Sequenced reads were trimmed by fastx-toolkit (0.0.13.2; quality threshold: 20) and adapters removed by cutadapt (1.8.3; minimal overlap: 1 nt). The remaining reads were uniquely mapped to the human reference genome (GRCh38.p13) using STAR mapping algorithm (2.5.4b), allowing one mismatch (--outFilterMismatchNmax 1 --outFilterMultimapNmax 1)⁴⁹. For genes containing more than one isoform, we considered the longest transcript isoform. Mapped reads were normalized as reads per kilobase per million mapped reads (RPKM).

Ribosome-protected fragments (RPF) from two biological Ribo-seq replicates (Pearson correlation coefficient R²= 0.7), were merged into a metagene set. The actively translated genes were identified using the ratio of the RPF in the first 100 nt and total RPF coverage (both as normalized reads per kilobase per million total reads of the library):

$$R_t=\frac{RPFwithinthefirst100nt[RPKM]}{TotalRPFoverfullgene[RPKM]}$$

Oxidative stress alters translation and causes an accumulation of RPFs within first 100 nt²², i.e. 3,505 transcripts presented stalling within the first 100 nt of the CDS with${\text{R}}_{\text{t}}>0.5$, and were considered as non-translated mRNAs. In total, 231 mRNAs had ${\text{R}}_{\text{t}}\le 0.5$ and designated as translated genes. To further filter out false-positives, we compared the coverage of the first and second half of each of those 231 transcripts, reasoning that actively translated genes should have uniform RPF coverage throughout the transcript; thereby, we excluded the first and the last 51 nucleotides to avoid bias due to higher ribosomal occupancy at initiation and termination. In addition, we reasoned that the translated transcripts must be also stably transcribed, thus, we also selected transcripts based on a uniform coverage in the RNA-seq data (merged replicates, Pearson correlation coefficient R²= 0.94). For this, we also compared the coverage in the first and second half of the transcripts and selected those with uniform coverage. We excluded the top 10% and bottom 10%whose mRNA coverage was slightly skewed to the first or the second half and hence we considered them as false positives. These additional filtering procedures allowed determining a set of 168 actively translated transcripts with high precision. Ribosome density (RD) (also known as translation efficiency²⁵) was calculated individually for each of the actively translated transcript, as ratio of RPFs and RNA fragments from the RNA-seq data set:

$$RD=\frac{RPF(RPKM)}{RNA(RPKM)}$$

QUANTIFICATION AND STATISTICAL ANALYSIS

Results were presented as the mean ± s.d. or s.e.m. from the number of replicates indicated in the corresponding figure legends. Details of exact statistical analyses and procedures can be found in the main text, figure legends, and STAR Methods.

Highlights.

• m⁶A in the close proximity of the start codon does not affect translation yields.

• Single m⁶A in the 5’UTR does not alter initiation complex assembly kinetics.

• m⁶A establishes stacking interactions with an arginine side chain of eIF2α.

• The stacking interactions m⁶A-Arg(eIF2α) have marginal energetic contribution.

INCLUSION AND DIVERSITY

We support inclusive, diverse and equitable conduct of research.