There are more than 170 RNA modifications forming the epitranscriptome.¹ N⁶-methyladenosine (m⁶A) is the most frequent internal modification of messenger RNA (mRNA) within the consensus DRACH motif (D = A/G/U, R = A/G, H = U/A/C) GGACU that is enriched near stop codons and in 3’ untranslated regions.^2–4 m⁶A affects most aspects of RNA regulation, i.e., alternative polyadenylation⁵, splicing⁶, nuclear export⁷, stability⁸, and translation initiation.^3,9 The complex of methyltransferase-like protein 3 (METTL3) and METTL14 (abbreviated as METTL3-14 in the following) is the main m⁶A-RNA methyltransferase (MTase).¹⁰

The METTL3-14 heterodimer is involved in a wide variety of diseases including several types of blood and solid tumors,¹¹ type 2 diabetes¹² and viral infections.¹³ METTL3-14 expression and thus m⁶A levels have been demonstrated to be dysregulated in diverse human malignancies and carcinogenesis.¹⁴ Dysregulated m⁶A deposition is directly involved in the development of acute myeloid leukaemia (AML).^15,16 It has been demonstrated that inhibition of the METTL3 catalytic function is sufficient to induce apoptosis and differentiation in AML cells and in a mouse model of the disease.^17,18 While it has been well established that METTL3-14 dysregulation is related to cancer development, the role of METTL3-14 varies in different cancer types, i.e., it can act as oncogene or tumour suppressor.¹⁹ Despite the growing knowledge on the diverse pathways, the mechanism of how m⁶A regulates gene expression remains poorly understood. Little is known about the recognition of specific RNA transcripts, the binding and release of the adenosine substrate, and the methyl transfer mechanism catalysed by METTL3. Furthermore, inhibiting the MTase function of METTL3-14 is a promising therapeutic strategy for several diseases.^18,20 Hence, understanding the mechanism of this complex would be helpful to develop new therapies.

The METTL3-14 complex is stable and catalytically competent^21–23, though its enzymatic activity is enhanced by a macromolecular assembly of ancillary proteins called MACOM (METTL3-14 associated complex).^24,25 METTL3-14 is the catalytic complex that transfers the methyl group from S-adenosylmethionine (SAM) to the substrate adenosine (Figure 1).^21,26 METTL3 comprises a low-complexity region at the N-terminus, a zinc finger responsible for substrate binding, and the catalytic MTase domain at the C-terminus (Figure 1A).²⁷ METTL14 also has an MTase domain, however, with a degenerate active site of hitherto unknown function, and so-called RGG repeats at its C-terminus essential for RNA binding.²⁶ The METTL3 MTase domain has the catalytically active SAM binding site and adopts a Rossmann fold that is characteristic of Class I SAM-dependent MTases (Figure 1B). METTL14 plays a structural role for complex stabilization and RNA-binding. It forms a positively charged groove at the interface with METTL3 which is predicted to be the RNA binding site.

Thus far, and to the best of our knowledge, no computational study of the conformational landscape or catalytic mechanism of eukaryotic RNA MTases has been reported. Molecular dynamics (MD) studies have mainly focused on protein and bacterial MTases.^28,29,30,31 In the latter, dynamic cross correlation analysis, a technique usually applied for the study of allosteric processes,³² was employed to compare the flexibility of the protein NirE, an MTase of a heme precursor, in the apo and holo states. The study showed that MTase conformational changes can influence the orientation of the substrate. In the former, Chen et al. explored the conformational landscape of SETD8, a histone MTase.³³ With the use of covalent binders and native ligands, several crystal structures in diverse conformations were obtained. Further structural diversity was achieved through structural chimerization between the C’ domain and the rest of the protein from different structures by massively parallel simulations.³⁴ The study showed how slow conformational motions and conformational states of the MTase are relevant to catalysis.

Among RNA MTases, research has focused on viral enzymes. One example is the MD simulation analysis of the binding mechanism of SAH and m⁷GTP to the Zika virus NS5 protein. ³⁵ This enabled a detailed analysis of their interaction and understanding of the effects of MTase inhibitors as antiviral drugs. Another study focused on the mechanisms of the SARS-CoV-2 MTase nsp16 and its heterodimeric partner nsp10 which acts as a stimulator of SAM binding.³⁶ Three systems were simulated: nsp16/nsp10/SAM, nsp16/SAM, and apo nsp10. The contribution to the binding energy of mutated residues in the binding site was calculated by an implicit model of electrostatic solvation and a surface area term for the nonpolar solvation.³⁷ The study provides a comprehensive understanding of the dynamic, thermodynamic, and allosteric processes of MTase complex formation and function.

Only a few structures of RNA-bound MTases are currently available due to the intrinsic instability of RNA and RNA-enzyme complexes and resulting difficulties in obtaining their structures. Of the known m⁶A RNA MTases, only METTL16, a SAM homeostasis factor, has been crystallized in complex with substrate RNA.³⁸ The lack of more such structures results in poorly understood RNA recognition and methyl transfer mechanisms. In contrast, structures are available for many MTases in complex with either the cosubstrate SAM or the coproduct S-adenosylhomocysteine (SAH), allowing for a well-understood cofactor binding within this protein family. Cofactor binding guided the initial design of bisubstrate analogues (BAs) as chemical tools to study the catalytic mechanisms of m⁶A-MTases (Figure 2).³⁹

BAs aim to mimic the transition state in which both the substrate nucleoside and the cosubstrate SAM are bound in the catalytic pocket of the enzyme while the methyl group is transferred from SAM to the adenosine N⁶-atom of the substrate RNA during catalysis (Figure 2A).^39–43 They consist of a SAM analogue (5’-N-SAM) covalently linked to the N⁶-position of an adenosine of a ribonucleotide(-like) fragment (Figure 2B).^40,44 The only structural information on these molecules is their binding mode in the bacterial m⁶A RNA MTase RlmJ.^39,42 There, in the mononucleoside containing compounds BA2 and BA4, the substrate adenosine is positioned in the presumed substrate binding pocket of RlmJ, and for the cofactor moiety, the methionine chain is bound like SAM. However, the structural studies resulted in a (co)substrate conformation that is not always biologically relevant, as the adenosine of the SAM analogue was rotated out of the canonical SAM binding pocket caused by π-stacking with the substrate adenine ring.³⁹ Therefore, the binding modes revealed therein are diverse and not always suitable for mechanistic studies. In a subsequent study using a dinucleotide containing BA (GA* in Figure 2B), the SAM moiety had the correct orientation.⁴² Furthermore, the N⁶-atom of adenosine was positioned, through the alkyl chain of the linker, at 3 Å away from the carbon corresponding to the Cε-atom of the methionine moiety in the cosubstrate SAM. Such a distance, in two non-linked moieties, would allow for an SN2 methyl transfer from SAM to the N⁶-atom of adenosine.⁴⁵ We therefore hypothesized that targeting human RNA MTases, in particular METTL3, with BAs might lead to structures that provide a suitable basis for understanding their MTase mechanism.

In the present study, we use a multidisciplinary approach to study the substrate binding, m⁶A methylation reaction, and release of m⁶A product in the human RNA MTase METTL3-14. To shed light on its catalytic mechanism we combine crystal structures of METTL3-14-bisubstrate analogue complexes with in vitro experiments, multiscale atomistic simulations, namely classical molecular dynamics (MD) and quantum mechanics/molecular mechanics (QM/MM) free energy calculations. Crystal structures show the binding mode of the substrate adenosine in two different conformations, representing an encounter complex of RNA binding and the transition state of catalysis, respectively. These structures are validated through mutational analysis. Classical MD simulations are used to investigate the binding of the substrates SAM and adenosine and dissociation of the products SAH and m⁶A. QM/MM free energy calculations reveal the details of the methylation reaction. Taken together, we elucidate the reaction catalysed by METTL3-14 at atomic level of detail. This knowledge will help in the further investigation of other MTases and the optimization of chemical probes that target their function.

We evaluated a series of bisubstrate analogues (BAs) as catalytic inhibitors and to investigate the structural similarity between their binding mode in METTL3-14 and the putative RNA substrate and SAM cosubstrate during methyl transfer (see Figure 2B). First, we measured the inhibitory activity of the BAs on METTL3-14 using an enzymatic assay that quantifies the N⁶-adenosine methyl transfer (Table 1).⁴⁶ The low micromolar inhibitory activity of the most potent BAs motivated us to conduct a structural investigation on their binding mode to METTL3-14. We conducted crystallization trials and obtained crystal structures of four of the BAs (BA1, BA2, BA4 and BA6) by soaking them into METTL3-14 crystals (see Table 1 and Supplementary Table S1).

All the crystallized BAs bind in the METTL3 active site (Figure 3A). BA1 and BA6, that are missing the methionine part of the SAM analogue or have a polar urea group in the linker, respectively, are the least potent of the crystallized compounds (see Table 1). Furthermore, in their crystal structures with METTL3, they reveal divergent interactions of their SAM-like moiety compared to SAM (Supplementary Figure 1). For BA2 and BA4, however, the interaction of their SAM-like moiety is the same as for SAM (Figure 3B-F). Both ligands bind with their SAM moiety in the active site lined with METTL3 hydrophobic residues and form polar contacts to conserved residues, namely D395, R536, H538, and N539 via the methionine part and D377, I378, F534, N549, and Q550 via the adenosine moiety of 5’-N-SAM. These structures therefore validate their further analysis for the interpretation of the catalytic mechanism of METTL3.

In the METTL3-BA2 and -BA4 structures, the SAM moiety is superimposable with the METTL3-bound conformation of SAM, and adenosine is bound in the presumed catalytic site (Figure 4A). Adenosine, based on BA2 and BA4, is involved in an intricate network of interactions with side chains of METTL3 (Figure 4B).

The classical m⁶A MTase catalytic ³⁹⁵DPPW³⁹⁸ motif (D/N-PP-Y/F/W) of METTL3 in its flexible active site loop 1 (ASL1, METTL3 residues 395-410) participates to the binding of adenosine. D395 forms a salt bridge to SAM and its mutation to alanine was previously shown to abolish METTL3 activity and SAM binding, confirming its involvement in cosubstrate binding.^21–23 The METTL3-BA2 structure reveals that there can also be a hydrogen bond formed between D395 and the N⁶-atom of adenosine in the case of the non-alkylated RNA substrate (Figure 4B). P396 in the ³⁹⁵DPPW³⁹⁸ motif, through its carbonyl group, forms a hydrogen bond to the N⁶-atom of adenosine. In addition, P397 from the ³⁹⁵DPPW³⁹⁸ motif makes hydrophobic contacts with SAM. The BA2 structure further shows that adenosine is stabilized by additional hydrogen bonds from its N⁷- and O2’-atoms to the METTL3 side chains of E481 and K513, respectively. While W398 from the ³⁹⁵DPPW³⁹⁸ motif is not directly involved in adenosine binding, its backbone carbonyl group forms a water-bridged hydrogen bond with the Y406 sidechain in the BA4 structure. The latter is involved in hydrophobic interactions with adenosine in the BA4 structure whereas it remains flexible in the BA2-bound METTL3 structure (Figure 4B).

Taken together, adenosine swaps conformation from solvent exposed in the METTL3-BA4 structure to buried in the METTL3-BA2 structure where it forms hydrogen bonds with residues in the active site loop 2 (ASL2, METTL3 residues 507-515). Each moiety of the BAs is involved in hydrophobic interactions with METTL3 residues, and both adenines (of SAM and adenosine) form hydrogen bonds to conserved residues in METTL3. Alanine mutation of D395, Y406, E481, and K513, which are involved in adenosine binding as seen in the BA2 and BA4 structures, almost completely abolishes the METTL3 catalytic activity (Figure 4C). Importantly, these residues are highly conserved in METTL3 (Supplementary Figure S2). The loss of activity of the Y406, E481, and K513 mutants originates mainly from abolished binding capability to adenosine and not SAM, because they can still bind SAH, as seen from a thermal shift assay (Figure 4D). Upon binding SAH, the thermal shift for the Y406, E481, and K513 mutants is similar as for the WT which suggests that their side chains are not involved in binding SAM/SAH but rather the RNA substrate. In contrast, mutation of D395 impairs SAH binding, which is consistent with the involvement of the D395 side chain in SAM binding.²¹

Because of the different conformations of METTL3 and adenosine in the BA2 and BA4 structures, we went on to characterize the enzyme dynamics in the presence of adenosine. We carried out multiple MD simulations of the METTL3-14 heterodimer in complex with (co)substrates and (co)products. Apo trajectories were generated in a previous study.⁴⁸ The MD simulations were started from the crystal structures of the complex with bisubstrate analogues BA2 and BA4 described above. Based on the position of each BA, the (co)substrates SAM and adenosine monophosphate (AMP) or the (co)products SAH and m⁶AMP were positioned in the protein, aligning them to their respective moiety of the BA. The interaction of each ligand to the protein, and of intramolecular polar contacts, was monitored throughout the trajectories. The dissociation time for each of the substrates and products was analysed and modelled by fitting a single exponential to the fraction of bound ligands (Table 2 and Supplementary Figure S3). The SAM and SAH cofactors remained bound in all but one of the sampled trajectories and thus fitting was not possible.

Simulations that were started with AMP or m⁶AMP in the conformation of BA4 showed immediate dissociation of these ligands (see Table 2 and Supplementary Figures S4 and S5). This could mean that this conformation represents a short-lived intermediate during substrate binding or product release. Possibly, it represents an encounter complex between the METTL3 -SAM holo complex and RNA substrate. This could be promoted through electrostatic steering of the negatively charged RNA backbone by the positively charged sulfonium ion of SAM.

The simulations started from the BA2 conformation show that the AMP and m⁶AMP ligands dissociated in several of the analysed MD simulation runs (Supplementary Figures S6 and S7). Binding of AMP is more stable, while the methylated product dissociates more quickly (see Table 2). This difference originates, at least in part, from the long-range monopole-monopole electrostatic interaction between the positively charged SAM and the negatively charged AMP. After the methyl transfer reaction, SAH is neutral which does not favour interaction with m⁶AMP resulting in its quicker dissociation. One caveat is that we simulate only a mononucleotide, resulting in much quicker dissociation of both the substrate AMP and product m⁶AMP than would be expected for the longer, canonical RNA substrate/product. The long DRACH motif-containing mRNA can interact with the binding groove at the METTL3-14 interface, resulting in a more stable complex. Nevertheless, our model is useful as it emphasizes the differences in binding affinity due to the different electrostatic interactions between the SAM-AMP and SAH-m⁶AMP pairs. Besides the structural stability of the BA2-like pose of AMP, the MD simulations also suggest that AMP can transiently populate a binding mode similar to BA4 (Supplementary Figure S8). This further supports that BA4 could represent an intermediate binding conformation of adenosine.

The crystal structure of the complex with BA2 reveals a string of ionic interactions which involves charged side chains of METTL3 and the amino and carboxyl groups of the SAM analogue. The string consists of seven charged groups (four of which are positive): R536⁺ – SAM COO^- – SAM NH₃⁺ – D395^- – K513⁺ – E481^- – K459⁺ (see Figure 4A,B). We decided to monitor these salt bridges and the following monopole-dipole interactions in the MD simulations: D395^- – AMP-N⁶, K513⁺ – AMP-N⁷, and E481^- – AMP-2’OH (see Supplementary Figures S4 to S7). In the BA2 conformation, the polar interactions between AMP and the charged sidechains D395 and K513 are stronger than the interaction between E481 and the 2’ hydroxyl group of the ribose (see Supplementary Figure S6, black traces). For SAM, the interaction between its positively charged amine and D395 is more stable than the one between its negatively charged carboxy with R536 which can adopt different conformations (see Supplementary Figure S6, blue traces). The intramolecular salt bridges are also stable throughout the sampling (see Supplementary Figure S6, grey traces). The salt bridge D395^- – K513⁺ seems to be the most stable, while E481^- – K459⁺ seems to change frequently between direct and water-separated contacts. In METTL3 apo trajectories, the D395^- – K513⁺ salt bridge is not present in the initial structure but is formed during the course of the simulation (Supplementary Figure S9). The bond between E481 and K513 is stable but separated by water when SAM and AMP are bound. This contrasts with the METTL3 apo simulations, where this bond forms transiently. The simulations with the products reveal similar interactions except for the faster dissociation of m⁶AMP (see Table 2 and Supplementary Figure S7). Another difference is the weaker interaction between SAH and D395 which fluctuates more than with SAM. The intramolecular interactions are also observed for the simulations initiated from the BA4 conformation of adenosine (see Supplementary Figures S4 and S5). These MD results validate the mechanistic interpretation of the BA-bound METTL3-14 crystal structures, and give a dynamic view of the behaviour of the complex before (substrate bound) and after (product bound) the methyl transfer reaction.

In structures of METTL3-14 in the apo state or bound to SAH or SAM, METTL3 residue Y406 is found in different conformations.^21–23 One study suggested that Y406 makes a hydrogen bond with S511 in ASL2 and thereby caps the SAH product.²³ However, another study suggested that Y406 might be important for the interaction with nucleotide bases.²² The different conformations of Y406 seen in our crystal structures with BA2 and BA4 support the latter and suggest an involvement of the Y406 sidechain in RNA nucleotide binding, probably as a first step of RNA recognition (see Figure 4A). Indeed, mutation of Y406 to alanine (in this and a previous study) or a cysteine (in a previous study) abolishes MTase activity (see Figure 4C).^22,23

The flexibility of Y406 in the MD simulations was analysed by monitoring the distance between its hydroxyl oxygen atom and the backbone carbonyl oxygen of METTL3 residue W398. There is a water-bridged polar interaction between these two oxygen atoms in the BA4 crystal structure (see Figure 3E). In the MD simulations, Y406 transitions multiple times (on the µs time scale) from orientations far away from the W398 carbonyl oxygen (distance of ∼15 Å) to shorter distances of ∼6 Å (Figure 5A-C). This corresponds to a transition from the extended ASL1, with Y406 pointing outside the pocket, to a conformation compatible with the water-bridged hydrogen bond observed in the crystal structure with BA4. Figure 5D shows a superposition of frames of a BA2 substrate MD trajectory illustrating this behaviour. We observe that the adenine ring system of AMP can be involved in a π-π interaction with Y406 for several ns before exiting the pocket, and is then captured again later by it, but is not brought into the pocket (Supplementary Movie 1). The full binding mechanism thus probably requires the rest of the substrate RNA, though the role of Y406 emerges already from the present simulations with the mononucleotide. The different conformations of Y406 reflect different steps in the binding and catalysis reaction. In the METTL3 SAM-bound state, the Y406 side chain is flexible. To bind RNA effectively, Y406 needs to stabilize through the water-bridged polar interaction with the backbone of W398. The substrate adenosine can then bind through selection of this conformation. Additionally, Y406 plays a role in positioning the adenosine molecule at the catalytic site.

Given the high concentration of SAM in the cell (60 to 160 μM in the rat liver), the cosubstrate SAM is expected to bind before the RNA substrate.⁴⁹ SAM binding results in large conformational changes of several side chains in the METTL3 active site (Figure 6). In the apo state, the K513 side chain points away from the putative RNA binding site and is involved in intramolecular polar contacts with the side chains of Y518, E532, and, via a water molecule, D395, that probably help to stabilize the apo protein (Figure 6A). SAM binding disrupts the water-mediated hydrogen bond between K513 and D395. This leads to a conformational change in the K513 side chain which then points in the direction of the RNA binding site where it can form a direct salt bridge with D395. In that conformation, the K513 side chain can also readily form a hydrogen bond to the adenosine-N⁷ of an RNA substrate as seen in the BA2 structure (Figure 6B). Furthermore, the side chain of H512, which is part of the SAM binding pocket environment, also undergoes a conformational change upon SAM binding (see Figure 6A). In the apo state, the H512 sidechain points inwards and thus blocks the adenosine binding site. However, once SAM is bound, the H512 side chain is attracted to form a π-π interaction with H538 whose side chain rotates from the apo state to interact with SAM in the holo state. This conformational change of the H512 side chain makes space for adenosine to bind as seen in the BA2 structure (see Figure 6B). MD simulations show that the side chains of both H512 and K513 are flexible in the apo state, but undergo stabilization upon SAM and adenosine binding (Supplementary Figures S6, S9, and S10). Together, the conformational switches of the H512 and K513 side chains upon SAM binding can be seen as priming METTL3 for adenosine binding.

We compared the structure of METTL3-BA2 with the structures of RNA MTases METTL4 and METTL16 bound to their substrates (Figure 7). When the SAM moiety of BA2 is superimposed with SAM bound to METTL4 and METTL16, the adenosine moiety of BA2 is situated in a very similar position as the substrates of the other MTases (Figure 7A). In METTL4 and METTL16, the N⁶-atom of their substrate adenosine is positioned at a distance of 2.8 and 2.1 Å from the methyl group of SAM, respectively (Figure 7B,C). The angle formed between SAM-5’S – Cε – adenosine-N⁶ in the substrate-cosubstrate pairs of METTL4 and METTL16 is 162.5° and 175.4°, respectively. In BA2, the adenosine-N⁶ is situated, through the alkyl chain of the linker, at 1.9 Å away from the carbon corresponding to the Cε-atom of the methionine moiety in the SAM cofactor (Figure 7D). The 1.9 Å distance in the crystal structure with BA2 is similar to the corresponding distance in METTL4 and METTL16, which, in two non-linked moieties, would allow for an SN2 transfer of the methyl group of a METTL3-bound SAM to the N⁶-atom of an adenosine substrate. Strikingly, the angle formed by the SAM-5’S – Cε – adenosine-N⁶ in METTL3 is 160.9° which is similar to the other MTases, and very close to the optimal 180° for an SN2 reaction in which the adenosine-N⁶ attacks the SAM-Cε and SAH becomes the leaving group. Hence, the adenosine of BA2 is in a suitable orientation for methyl transfer and BA2 represents a transition state mimic for METTL3. This is useful atomistic information for setting up QM/MM free energy calculations to study the catalytic reaction (see below).

We carried out hybrid quantum mechanical/molecular mechanical (QM/MM^50–57) free energy simulations to establish the catalytic mechanism of RNA methylation by the METTL3 -14 complex.^23,58,59 The crystal structure with the bisubstrate analogue BA2 (see Figure 3C) was used as the starting point for the QM/MM simulations (Figure 8).

In the simplest mechanism (Figure 8A), the methyl cation in the SAM cofactor is transferred directly to the N⁶ position of the adenosine substrate, prior to the deprotonation of N⁶H₂, which has a very high pK_a of ∼17.⁶⁰ Indeed, DFTB3/MM free energy simulations show that this mechanism is energetically favourable (by about ∼4 kcal/mol) with a barrier of 15-16 kcal/mol (Figure 8B). The turnover as measured by an enzymatic assay is 0.2-0.6 min^-1 at ambient temperature which implies a barrier of ∼20 kcal/mol.⁶¹ Hence, the methyl transfer is not the rate-limiting step. Taken together, the QM/MM and MD simulations suggest that the dissociation of the coproduct SAH and product RNA is likely the rate-limiting step. Compared to the model reaction in solution computed using a continuum solvation model (Table 3), the reaction in the enzyme is substantially more exoergic, suggesting that the enzyme environment stabilizes the product of the methyl transfer reaction. Inspection of the active site structure based on DFTB3/MM simulations suggests that such stabilization primarily comes from the hydrogen-bonding interactions between the adenosine N⁶ group and nearby polar groups, in particular the side chain of D395 and the backbone carbonyl of P396 (Figure 8C).

This catalytic mechanism is similar to N⁶-adenine DNA methyl transferase M·TaqI, in which the adenosine N⁶ group is hydrogen-bonded to an Asn side chain and backbone carbonyl of a Pro in the protein.⁶² In the absence of any catalytic base, it was proposed that the methyl transfer occurs first, leading to a m⁶NH₂⁺ group well stabilized by hydrogen bonding interactions with the Asn side chain and Pro backbone carbonyl. The mechanism was supported by QM/MM free energy simulations with a barrier height of ∼20 kcal/mol.⁶³ On the other hand, the same QM/MM study suggested that when the active site Asn was replaced by an Asp, a mechanism in which deprotonation of the adenine N⁶ group by the Asp preceded the methyl transfer from SAM also had a comparable free energy barrier. Due to the involvement of the proton transfer, the corresponding transition state exhibited rather different charge distributions from that in the WT M·TaqI. DNA N⁶-methyltransferases in the αβ groups feature an Asp in the active site, while those in the γ group have an Asn at the equivalent position.⁶⁴ Hence, the QM/MM computational results suggest that transition states with distinct charge distributions are involved in different groups of enzymes (i.e., αβ vs. γ), giving rise to the opportunity of designing transition state analogues as inhibitors unique to specific classes of methyl transferases.⁶³

This raises the question whether methyl transfer in METTL3-14 may also occur following deprotonation of the adenosine N⁶H₂. The structural features of the METTL3-14 active site do not support this mechanism. The N⁶ position has a very high pK_a of ∼17, and thus its deprotonation requires a particularly strong base, which is absent in the active site of METTL3 -14. For example, while there is a carboxylate nearby (D395), it forms a salt-bridge with K513, and therefore is expected to feature a too low pK_a value to deprotonate the adenosine N⁶H₂. These considerations are congruent with the observation that the computed DFTB3/MM free energy profile without N⁶ deprotonation is possible within the reported experimental kinetics.^61,65,66

The difference in the methylation energetics of adenosine in different protonation states shown in Table 3 is consistent with the pK_a difference of adenosine before and after methylation. Thus, the large difference suggests that N⁶ becomes much more acidic following methylation, which is consistent with literature estimates of the pK_a values of N⁶-protonated adenosine derivatives in the range of -3 to -10.⁶⁷ Close inspection of the active site structure in the product state of DFTB3/MM simulations reveals that the proton release may occur readily through water wires that connect the N⁶ position to the protein-solvent interface (Figure 9). Therefore, favourable salt-bridges in the active site (e.g., D395-K513) do not have to break to allow the proton release following methylation of the substrate.

Finally, we note that the recently solved crystal structures of METTL4, which belongs to a subclade of MT-A70 family members of MTases, showed an active site very similar to that of METTL3.⁶⁸ A nearby ion-pair (D233-K364) is engaged in a hydrogen-bonding network involving both the substrate and SAM. Therefore, we expect that the catalytic mechanism discussed here applies also to METTL4 and potentially other MT-A70 family members of MTases.

The complementarity of the methodologies and the congruence of the experimental data and simulation results allow us to construct a model of the METTL3 catalytic reaction (Figure 10). This model shows that in the apo state (Figure 10, State 1), the sidechain of K513 is involved in intramolecular interactions that stabilize the protein. SAM binding displaces the K513 sidechain and brings it in the right conformation where it can form a hydrogen bond to the N7 of the adenosine substrate (Figure 10, State 2). Space for adenosine to bind is further conditioned by the conformational change of the H512 side chain which blocks the catalytic site in the apo state, but is drawn to interact with H538 in the SAM-bound state. The proper recognition of adenosine is conditioned by the interaction with the aromatic side chain of Y406 to which the adenosine substrate can bind and stabilize the ASL1 loop through interaction in the BA4 conformation (Figure 10, State 3). The Y406 side chain acts as gatekeeper and swaps out to allow the rotation of adenosine into the catalytic site where it is stabilized through hydrogen bonds to E481 and K513 in the BA2 conformation (Figure 10, State 4). The bond between K513 and adenosine is especially stable in the MD simulations started from the binding mode of BA2. Hydrogen bonds with the D395 sidechain and P396 backbone enhance the nucleophilicity of the adenosine-N⁶ and trigger the SN2 reaction with the electrophilic methyl group (Figure 10, State 5). The deprotonation of m⁶NH₂⁺ following the methyl transfer may proceed along multiple water-mediated pathways that lead to the protein-solvent-interface (Figure 10, State 6). The m⁶A product can then flip and slip back into the BA4 conformation (Figure 10, State 7) before the flexibility of the ASL1 loop then facilitates its release (Figure 10, State 8).

The synthesis of the bisubstrate analogues was as previously described: BA1/2/3/4/6, ref. 40; Compound 12, ref. 41; GA*, ref. 42.

For determining the half maximal inhibitory concentration (IC50) with the full-length complex and for crystallization studies with the truncated complex METTL3^MTD:METTL14^MTD containing just the methyltransferase domains (MTD) of METTL3 (residues 354–580) METTL14 (residues 107–395), the recombinant complex constructs were expressed using the baculovirus/Sf9 insect cell expression system and purified as described previously.²³

For mutational analysis, alanine mutants were generated in the baculovirus vector pFastBacDual-StrepII-GFP-TEV-METTL3-His-TEV-METTL14 using the QuikChange site-directed mutagenesis protocol and confirmed by sequencing. Recombinant baculoviruses were generated using the Bac-to-Bac system. For protein expression, suspension cultures of Sf9 cells in Sf-90 II SFM medium (Thermo Fisher) were infected at a density of 2×10⁶ ml⁻¹. Cells were harvested 72 hours post infection, resuspended in Buffer A (50 mM Tris-HCl pH 8.0, 500 mM NaCl) supplemented with Protease Inhibitor Cocktail (Roche Diagnostics GmbH, Germany), phenylmethylsulfonyl fluoride (PMSF), Salt Active Nuclease (Merck), and lysed by sonication. The protein complex was purified by Ni-affinity chromatography on a 5 mL HisTrap HP column (Cytiva) equilibrated and washed with Buffer A and eluted with 250 mM imidazole. Proteins were further purified by Strep-tag purification using a 5 mL StrepTrap XT column (Cytiva) equilibrated and washed with buffer A and eluted with 50 mM biotin. The affinity tags were removed by digestion with TEV protease overnight at 4°C, followed by further purification by size exclusion chromatography using a Superdex 200 Increase 10/300 GL column (Cytiva) in 20 mM Tris-Cl, pH 8.0, and 200 mM KCl. The proteins were concentrated, flash-frozen in liquid nitrogen, and stored at -80°C until further use.

The SAH (S-adenosyl-l-homocysteine)-bound holo protein crystals of METTL3^MTD:METTL14^MTD were obtained as previously described.²³ The bisubstrate analogues (BAs) were dissolved in DMSO at concentrations of 100 mM. Complex structures were solved by soaking BAs into holo protein crystals and replacing the bound SAH in the METTL3 catalytic pocket. First, 1 μL of the BA dissolved in DMSO was left overnight to evaporate the solvent at room temperature. The next day, 1 μL of mother liquor containing 30% PEG-3350 and 200 mM Mg-acetate was added on top of the dried compound stamp. One holo crystal was then transferred into the mother liquor over the target compound stamp. After 16 h of incubation at 22°C, the crystals were harvested and flash-frozen in liquid nitrogen.

Diffraction data were collected at the PXIII beamline at the Swiss Light Source (SLS) of the Paul Scherrer Institute (PSI, Villigen, Switzerland). Data were processed using XDS.⁶⁹ The crystal structures were solved by molecular replacement by employing the 5L6D structure as the search model in the Phaser program (Phenix package).⁷⁰ Crystallographic models were constructed through iterative cycles of manual model building with COOT and refinement with Phenix.refine.^71–74

The inhibitory potencies of the bisubstrate analogues (BAs) for METTL3 were quantified by a homogeneous time resolved fluorescence (HTRF)-based enzyme assay as previously described.⁴⁶ Briefly, the level of m⁶A in an RNA substrate after the reaction catalysed by METTL3-14 was quantified by measuring specific binding to the m⁶A reader domain of YTHDC1 (residues 345-509) by HTRF. BAs that inhibit METTL3 decrease the m⁶A level and thus reduce the HTRF signal. Dose response curves of titrations with the BAs were plotted in OriginLab 2018 and fitted with nonlinear regression “log(inhibitor) vs. normalized response with variable slope” from which IC50 values were determined. Each BA was measured in duplicates.

For the mutational analysis, the HTRF assay was used with some modifications. In the reaction step, METTL3-14 (wildtype or mutant) (40 nM final concentration) methylates the 5′-biotinylated ssRNA (5′-AAGAACCGGACUAAGCU-3′ (Microsynth)) (200 nM final concentration). The co-substrate SAM (Cisbio, 62SAHZLD) (450 nM final concentration) was added as the last component and thus initiated the methylation reaction. The final reaction volume was 15 μL in 20 mM Tris-HCl, pH 7.5, 0.01% (w/v) bovine serum albumin (BSA). The reaction was let to incubate for 1 hour at room temperature (RT) and then stopped by addition of 5 μL detection buffer (50 mM HEPES, pH 7.5, 150 mM NaCl, 200 mM KF, 0.05% (w/v) BSA, 25 nM GST-tagged m⁶A reader YTHDC1(345-509), 3 nM XL665-conjugated streptavidin (Cisbio, 610SAXLB), 1x anti-GST Eu³⁺-labelled antibody (from 400x stock (Cisbio, 61GSTKLB))). Capture of the m⁶A-modified RNA by the m⁶A reader and the biotinylated RNA by Streptavidin was allowed to proceed for 3 hours at RT and in the dark before the HTRF signal was measured using a Tecan Spark plate reader (Tecan). The plate reader recorded with a delay of 100 μs the emission at 620 and 665 nm after the excitation of the HTRF donor with UV light at 320 nm. The emission signal was read over an integration time of 400 μs. The ΔF (((ratio_sample - ratio_background) / ratio_background) * 100) used as background the reaction without SAM as an internal control.

Experiments were conducted as previously described with some modifications.¹⁷ Briefly, METTL3-14 (wildtype or mutants) at a final concentration of 0.5 μM was mixed with SAH at a final concentrations of 500 μM in a final volume of 20 μL in a buffer consisting of 20 mM Tris-Cl, pH 8.0, and 200 mM KCl. DMSO concentration was kept at 1 % (v/v). SYPRO Orange was added at a final dilution of 1 : 1000 (v/v) as a fluorescence probe (ex/em 465/590 nm). Differential scanning fluorimetry was performed on a LightCycler® 480 Instrument II (Roche Diagnostics, Indianapolis, IN). The temperature was raised in steps of 3.6°C per minute from 20°C to 85°C and fluorescence readings were taken at each interval. The T_m values were determined as the transition midpoints of the individual samples. The ΔT_m values were calculated as the difference between the transition midpoints of the individual samples and the reference wells containing the protein and DMSO only from the same plate. Samples were measured in triplicates.

The crystal structures of the METTL3-14 heterodimer in complex with the BA2 and BA4 ligands were used as starting conformation for simulating the heterodimer with its substrates SAM and AMP and products SAH and m⁶AMP. The addition of the phosphate to adenosine and m⁶-adenosine aims to mimic one element of the substrate RNA chain. The ligands were aligned to their respective moiety of the BA, keeping the original coordinates for present atoms, and reconstructing the missing parts. The missing segments of the METTL3-14 crystal structures were reconstructed using the SWISSMODEL web server with the structures as templates. The simulated construct spans residues L354 to L580 of METTL3 and S104 to L289 of METTL14. All MD simulations were performed with GROMACS 2021.5 using the CHARMM36m July 2021 force field. The models were solvated in a 9 nm box and equilibrated with Na⁺ and Cl^- ions to a concentration of 150 mM. Energy minimization was applied and a canonical equilibration under all-atom positional restraints was performed for 5 ns to reach 300 K. A further canonical equilibration was performed for 10 ns with the Cα atoms of the proteins, the adenine moiety of SAM/SAH and of AMP/m⁶AMP under positional restraints. These partial restraints were set to allow a relaxation of the side chains around the ligands. Sixteen independent runs of canonical MD simulations were then started for each of the systems, sampling 500 ns per run.

The existence and behaviour of salt bridges observed in the crystals were monitored throughout the runs. Furthermore, the stability of the binding was calculated by defining a binding pocket for the BA2 and the BA4 conformations. The distance between residues was calculated and contacts were defined as a distance smaller than 5 Å. For each of the simulation groups, the binding pocket was determined as the residues with a mean contact presence higher than the 90 percentiles of mean contacts. A dissociation event was defined as the mean of the ligand to pocket distances surpassing a threshold of 10 Å. The dissociation rates were predicted using exponential fitting, modelled as a single exponential. A single exponential with a multiplicative factor was used to check the quality of the fitting, with a preexponential factor close to 1 indicative of a good fit. No fitting was done for SAM/SAH as they remained bound in almost every single trajectory.

To understand the intrinsic energetics of the methyl transfer reaction, we conduct QM calculations of an infinitely separated model substrate (adenosine) and a truncated model for the cofactor SAM in which the adenosine and amino moieties are replaced by ethyl groups. The N⁶ position of the model substrate is taken to be either protonated (-NH₂) or deprotonated. Calculations are performed in the gas phase using two different density functional theory (DFT) methods (B3LYP with the D3 dispersion correction^75–78 and ωB97XD⁷⁹) and two different basis sets (aug-cc-pVDZ and aug-cc-pVTZ).^80–82 Calculations are also conducted at the DFTB3/3OB level for calibration as the same QM method is used in subsequent QM/MM free energy simulations. To probe the effect of solvation on the methyl transfer energetics, single point calculations are carried out with the conductor-like polarizable continuum model (CPCM^83,84) using the gas-phase optimized structures at both B3LYP-D3 and ωB97XD levels. DFT calculations using B3LYP-D3 and ωB97XD are conducted using the Gaussian16 software⁸⁵, and DFTB3⁸⁶ calculations are carried out using the CHARMM program.⁸⁷

We employ QM/MM free energy simulations to probe the mechanism of adenosine N⁶-methylation catalysed by the METTL3-14 complex. As illustrated in Figure 3C, the crystal structure (at 2.3 Å resolution) of the complex with a transition state analogue (the bisubstrate analogue BA2) suggests a rather straightforward mechanism through which the methyl group is transferred from SAM to the adenosine N⁶ position. The deprotonation of N⁶ by a nearby base in principle may occur either before or after the methyl transfer, but QM/MM calculations strongly suggest that deprotonation occurs after the methyl transfer.

In the QM/MM simulations, the bisubstrate analogue is first converted to a SAM non-covalently bonded to the adenosine: the nitrogen N_BS atom in the crystal structure is replaced by a sulphur atom and the extra carbon atom is deleted. The O5*-PA bond is cut off and a hydrogen atom is patched to the location. The QM region includes the cofactor SAM, the model substrate adenosine (Ade), the carbonyl group of A394, D395, P396, and P397 without the backbone carbonyl group, and the side chain of K513. Link atoms are added between C and C_α of A394 and P397, and between C_β and C_α of K513 to saturate the valence of the QM boundary atoms using the divided frontier charge (DIV) scheme.⁸⁸ The QM atoms are treated with the third-order density functional tight binding (DFTB3) method with the 3OB parameter set^89,90; benchmark calculations using model compounds (see Table 3) indicate that the DFTB3/3OB method describes the energetics of the methyl transfer reaction rather well in comparison to DFT calculations with a large basis set. The MM region is described with the CHARMM36 force field for proteins.⁹¹

In the generalized solvent boundary potential (GSBP^92,93) framework, the inner region contains atoms within a 27 Å-radius sphere centred at the N⁶ in Ade. Newtonian equations of motion are solved for atoms within 25 Å. Protein atoms in the buffer region (25-27 Å) are harmonically restrained with force constants determined from the crystallographic B factors and Langevin equations of motion are solved with a bath temperature of 300 K.⁹⁴ The remaining portion of the system in the outer region is frozen. All water molecules are subject to a weak geometrical (GEO) type of restraining potential to keep them inside the inner sphere.⁸⁷ Weak GEO restraints are added on adenosine to make sure it is well-bounded during the simulations. Electrostatic interactions among inner region atoms are treated with extended electrostatics and a group-based cut-off scheme.⁹⁵ The static field due to the outer region atoms is evaluated with the linearized Poisson-Boltzmann (PB) equations using a focusing scheme, which employs a coarse grid of 1.2 Å and a fine grid of 0.4 Å.⁹⁶ The reaction field matrix is evaluated using spherical harmonics up to the 20th order. In the PB calculations, dielectric constants of the protein and water are set to 1 and 80, respectively, and the salt concentration is set to zero.

To probe the energetics of the methyl transfer reaction catalysed by the METTL3-14 complex, well-tempered multiple-walker metadynamics simulations^97,98 are carried out using the PLUMED-CHARMM interface.^99,100 The antisymmetric stretch that describes the methyl transfer process between the N⁶ in adenosine and the SAM sulphur-methyl group is chosen as the collective variable (CV), i.e., ξ = r(CE−SD) − r(CE−N6). The corresponding C-S distance and C-N distance are also monitored but no bias potential is added. SHAKE is applied to all bonds involving hydrogen and used to avoid undesired proton transfer reactions.¹⁰¹

The first two metadynamic runs are not well-tempered for the efficiency of sampling. In the subsequent well-tempered runs, the bias factor is set to be 35. A new Gaussian biasing potential is added every 0.2 ps with an initial height of 0.3 kJ/mol and a width of 0.05. Twenty-four walkers with different initial velocities are used per simulation in parallel while sharing hill history among all walkers every 1 ps. Each walker is run for 250 ps for a total of 6 ns of sampling, and convergence is evaluated by comparing the PMF as a function of the number of Gaussians added.