The MTase cores of human METTL3 (aa 358 - 580) and METTL14 (aa 116 - 378) form an obligate heterodimer. METTL3 acts as an active SAM-dependent MTase, whereas METTL14, an inactive MTase, stabilizes RNA^14–17. We co-purified the MTase core of METTL3-METTL14 from E. coli. We succeeded in resolving its structure in the presence of N⁶-methyladenosine 5’-monophosphate (m⁶AMP, referred to as m⁶A), a product of methylation reaction, by soaking m⁶A into apo crystals (Fig. 1 a, Extended data Fig. 1a-c). The difference omit map showed clear and unbiased electron density for m⁶A, which was refined well with no discrepancies for the ligand, surrounding regions, or the rest of the protein (Extended data Fig. 1d-g and table 1). METTL3-METTL14-m⁶A model was refined to ~2.5 Å resolution, with excellent stereochemistry and R_free and R_work of ~26.2 and 22.9%, respectively (Extended data Table 1). The final model contains one molecule each of METTL3 (aa 369 – 579), METTL14 (aa 116 – 402), one m⁶A, 90 water, and two ethylene glycol.

The overall fold of METTL3-METTL14 is similar to those reported previously^14–16, except for notable changes in the region around the m⁶A binding pocket. MTases adopt a β-class of MTase fold with a central β-sheet of seven parallel and one antiparallel β-strands flanked by three helices on each side. Three major loops (gate loops 1 and 2 and an interface loop) emanating from the central β-sheet of METTL3 participate in SAM, RNA, and METTL14 binding. While the two gate loops exhibit high flexibility upon SAM or SAH binding and release, the interface loop remains rigid due to extensive protein-protein contacts from METTL14 MTase (Fig. 1a).

While the MTase core of METTL3-METTL14 is not catalytically active form, its structures derived by soaking SAM or SAH into these crystals provided crucial insights into the binding of methyl donor (SAM) and byproduct (SAH) and important conformational changes in the regions of METTL3 (gate loops 1, 2) surrounding these ligands^14–16. The m6A-bound structure, which we report here, is the first nucleotide-bound form of the MTase core of METTL3-14 and provides crucial insights into the mechanism of this critical enzyme that is central to normal homeostasis and diseases. Since a small molecule targeting the catalytic pocket of METTL3 has entered human clinical trials for cancer therapy, our results showing the nucleotide binding in the regions outside of catalytic and SAM-binding pockets will pave the way for the rational design of more specific inhibitors.

Strikingly, m⁶A occupies a cryptic pocket ~16 Å away from the methyl donor SAM pocket with its N⁶-methyl moiety in an energetically favored syn conformation, facing outward (Fig. 1a). Previously, this region was postulated to bind RNA due to its positive charge and polar nature^14–16. m⁶A is stabilized by a vast network of specific interactions, mostly from METTL3 and R298 of METTL14. The purine ring of m⁶A is sandwiched between the side chain of M402 and the backbone atoms of the interface loop residues, R471, T472, G473, and H474. The two arginine residues (R471 of METTL3 and R298 of METTL14) act like a clasp to hold the N⁶-methyl moiety in place through a direct h-bond between R298 and N¹, van der Waals and hydrophobic interactions between N⁶-methyl and its aliphatic portion, and the amino group of the R471 side chain, respectively. The carbonyl oxygen of G473 appears to neutralize the positive charge of the R298 residue. The carbonyl moiety of R471 embraces the N⁶ atom of m⁶A via a direct h-bond, while the opposite side is stabilized by the side chain of H474 via a π-π interaction. Altogether, the arginine clasp, interface loop residues R471-H474 and M402, forms a partial closure around the methylated purine ring of m⁶A. The ribose in the C3’-endo conformation is stacked between the backbone atoms of G473 and H478 and the side chains of I400 and H478. The phosphate group of m⁶A is locked in place by multiple direct h-bonds with its phosphoryl oxygens and side chains of H478, E481, T433, and K459 (water-mediated) – all from METTL3, and another water-mediated interaction with E257 of METTL14. The side chain of H478 holds the m⁶A phosphate on one side and E257 of METTL14 on the other, thus acting as a hinge (Fig 1a, Extended data Fig. 1g).

Strict conservation of the extensive interaction network of m⁶A in human, animal, plant, and yeast suggests that m⁶A sensing by this cryptic pocket is an evolutionarily conserved mechanism (Fig. 1b). Several key residues that partake in m⁶A binding, such as R471 and R298 of the arginine clasp, E481, and H478 that stabilize the N⁶-methyl and phosphate groups are recurrently mutated in endometrioid and adenocarcinoma³² (Fig. 1b). We introduced the R298P mutation in METTL14, a recurrent mutation event in endometrioid carcinoma^32,33, and the R471H, E481A, T433A, K459A, and H478A mutations in METTL3. In addition, we generated two deletion mutants (Δ472-473, Δ472-474) in which three residues of METTL3 (T472, G473, H474) that stack against the purine ring of m⁶A were deleted to shorten the interface loop.

We co-purified the full-length wild-type human METTL3-METTL14 and eight mutant enzymes from insect cells and probed their RNA methylation and binding activities. We used a 30-mer RNA oligo (NEAT2*) consisting of one canonical GGACU motif. Consistently, R298P and R471H mutants significantly reduced (up to 85%), whereas T433A resulted in ~20% loss in methyltransferase activity, agreeing with the reduced m⁶A levels observed in endometrial tumors harboring the R298P mutation³³. The other five mutations in METTL3 (Δ472-473, Δ472-474, K459A, E481A, and H478A) completely abolished the RNA methyltransferase activity of METTL3-METTL14 (Fig. 1c). Thus, the evolutionarily conserved m⁶A binding pocket is essential for efficient conversion of A to m⁶A.

Next, we quantitatively determined the binding affinities of wild-type (WT) and mutant enzymes to the substrate and a product RNA, wherein the target A base within GGACU is replaced by m⁶A. We covalently attached a fluoresceine moiety to the 5’-end of both the substrate NEAT2* (A-RNA) and product (m⁶A-RNA) RNAs and performed fluorescence polarization-based assays. The WT enzyme can still bind the m⁶A-RNA and A-RNA with high affinity (K_d = 9-20 nM) (Fig. 1 d,e), corroborating previous studies attributing a sort of m⁶A-reader function to METTL3 in vivo^23,25,31,34. In contrast, the mutants, including R298P and R471H (both mutated in cancers and belong to the arginine clasp motif that stabilizes the m⁶A) exhibited a loss in binding affinity with varying degrees, with R298P showing weakest binding (Fig. 1e, Extended data Fig. 1h). Thus, inability to sense and distinguish m⁶A properly by the R298P mutation could result in total m6A and promote tumorigenicity and growth of endometrial tumors as observed previously³³. The nanomolar affinity observed in this fluorescence polarization (FP)-based assay for mutant enzymes suggests a significant contribution of accessory motifs such as zinc fingers (ZnFs) of METTL3 and RGG repeats of METTL14 to RNA binding, especially the predicted bulged stem-loop structure of NEAT2* RNA (A-RNA). These motifs are intact in both the wild-type and the mutant enzymes. This could be one of the reasons for not observing the radical change in overall RNA binding as measured by the equilibrium dissociation constant (Kd) for A vs. m6A RNA in an FP-based assay. Moreover, the residue aligning the catalytic and cryptic pockets would interact with the substrate (A) and the product (m6A) nucleotide of the ‘DRACH’ sequence during the pivoting of the base upon conversion to m6A, respectively. Even if the mutations in the cryptic pocket retain overall RNA binding dominated by domains flanking the MTase core (ZnFs and RGG) and the secondary structure of RNA itself (GGACU containing stem-loop RNAs show higher affinity), they could still influence the pivoting of the m6A base and its release after conversion.

We also performed a kinetic analysis by varying the concentration range of RNA substrate from 10 nM to 10 μM in the presence of a saturating concentration of SAM (10 μM) (Fig. 2a). Consistently, these results show that the wild-type enzyme yields the highest methylation activity, whereas the mutant enzymes show reduced methylation. Next, we studied the binding kinetics of the full-length METTL3-METTL14, wild-type (WT), and the two mutant enzymes harboring R298P and R471H mutations in METTL14 and METTL3, respectively. We employed the surface plasmon resonance (SPR) technique, a gold standard for studying binding kinetics that includes ON (K_on) and OFF (K_off) rates of RNA binding. We used our original RNA substrate, NEAT2 (30-mer bulged stem-loop) RNA, and its methylated version (NEAT2-m6A). We also examined a 14-mer linear RNA substrate (r6T) and its methylated form (r6T-m6A) to fully understand the kinetics of enzyme binding to GGACU-containing RNAs with different shapes and sequences. The SPR data of the WT enzyme fit well with a 1:1 binding model. As shown in Fig. 2b and Tables 1-3, the binding affinity and kinetics of METTL3-METTL14 enzymes differ significantly on the two RNA oligos tested. The structured NEAT2 RNA shows a 5-fold tighter affinity than a linear r6T substrate, while their methylated counterparts show reduced binding (1.3-fold for NEAT2-m6A and 2.3-fold for r6T-m6A), but the Kds (dissociation constants) are still in sub and low-micromolar range.

Interestingly, the ON (K_on) and OFF (K_off) rates for a linear substrate RNA differ 2-2.7-fold compared to NEAT2 RNA. A 2.7-fold faster OFF rate on a linear substrate would result in a rapid turnover and higher methylation of the GGACU motif residing in a linear RNA compared to the loop region of a stem-loop RNA. At the same time, the structured elements in RNA can help recruit METTL3-METTL14, as suggested by a 5-fold stronger affinity to NEAT2 RNA. These results suggest the RNA shape surrounding the core ‘GGACU or DRACH’ motif is a crucial determinant of methylation. This observation can explain, in part, why only a fraction of potential DRACH motifs (or perfect GGACU) are methylated in vivo.

The SPR data of the two mutant enzymes (R471H and R298P) revealed a change in Rmax and mode of binding. While the SPR data of the WT enzyme fit well with the 1:1 binding model, the mutant data could only fit well with a two-state reaction model (Table 3). The R298P mutation in METTL14 results in a moderate loss (1.4-1.6-fold) of binding to the stem-loop NEAT2* RNA (Kd=256 vs. 356 nM) and its methylated counterpart (Kd= 337 vs. 538 nM) or the linear r6T RNA (Kd=1360 vs. 2160 nM) but a significant loss (>7-fold) in RNA binding occurred for methylated linear r6T RNA (Kd=3120 vs 22900). On the other hand, the R471H mutant exhibits much slower dissociation after binding, thus hampering the enzyme’s turnover. These data suggest that the two arginines (R471H of METTL3 and R298P of METTL14) in the m6A pocket, which is cryptic in nature, are important for recognizing the methylation status, and the replacement of R298 to proline and R471 to histidine would negatively impact the methylation of the canonical GGACU motif. Of note, striking differences in the mode of binding – two-state binding for R471H/ R298P mutants vs. one-state binding of the WT enzyme – and their dynamics (ON and OFF rates) likely alter the retention time of the enzyme on RNA with varied shapes and sequences or residence time for m6A of these RNAs in the cryptic pocket. Consequently, multi-stage RNA binding and altered dynamics could alter specificity. An independent study by Zhang et al. reported that R298P mutation alters the enzyme’s preference from GGAC to GGAU³⁵. In this context, our structural and binding data provides crucial and timely insights into the existence of a cryptic pocket, the importance of two arginines for methylation of the canonical GGACU motifs, and how mutations at these positions could alter the dynamics of RNA binding of METTL3-METTL14 and ultimately alter enzyme specificity.

To assess the dynamic impact of the mutations in METTL3-METTL14 on m⁶A binding, we conducted supervised molecular dynamics simulations (SuMD) and analyzed the distances between the center of mass (com) of the m⁶A and the product m⁶A binding pocket (defined by residues within 4 Å of m⁶A (see methods) over time for wild-type and each mutant, including T433A, K459A, R471H, Δ472-473, Δ472-474, H478A, E481A, and R298P. The wild-type METTL3-METTL14 complex displayed distances around 1.5 Å for most of the simulation time, highlighting the presence of m6A within the product m6A binding pocket (Extended data Fig. 3a and 4a). Furthermore, this also indicates that the product m⁶A binding pocket in the wild-type is well-structured to accommodate m⁶A securely and thereby maintain a stable binding environment (Extended data Fig. 3b). In the wild-type, interactions between H478, T472, and m⁶A are observed. Some additional interactions are also formed; for example, T433 and E481 form stable interactions with the phosphate group via a water molecule. A salt bridge between K459 and E481 helps to stabilize the water molecule (Extended data Fig. 4a). In H478A, which is no longer able to interact with m⁶A, the phosphate group becomes mobile and is unable to anchor to A478 (Extended data Fig. 4f). In T433A and R298P mutations, the distance between the center of mass of m⁶A and the product m⁶A binding pocket mostly ranged from 2 to 4 Å, suggesting that these mutations, while impactful, do not entirely abolish m⁶A binding (Extended data Fig. 4b, c). Consequently, these mutants could retain m⁶A in the product m⁶A binding pocket for most of the simulation time. It is interesting to note that in the T433A mutation, in spite of the h-bond interaction between the phosphate group and the side chain of T433 is lost, the interaction between H478 and the phosphate group remains intact (Extended data Fig. 4c). Therefore, this T433A mutation does not significantly affect the binding of m⁶A. In the case of R298P (Extended data Fig. 4b), although the arginine clasp was broken, the phosphate group remains stabilized in the same position, making interactions with H478. Moreover, E481 in the R289P mutant can also contribute to stabilizing the phosphate group via a water molecule (Extended data Fig. 4b). However, in the E481A mutant, this water bridge is lost. Similarly, the mutation in K459A results in the loss of a salt bridge with E481, leading to localized destabilization (Extended data Fig. 4d, g). The resulting electrostatic repulsion between the negatively charged side chains of E481 and the phosphate group of m⁶A pushes the m⁶A out of the product m⁶A binding pocket (Extended data Fig. 4g). For the two deletion mutants (Δ472-473, Δ472-474), m⁶A immediately leaves the product m⁶A binding pocket after the removal of the constraint in the equilibration steps, highlighting the essential nature of these residues in maintaining the binding pocket’s structure (Extended data Fig. 4h, i).

The two loops in METTL3 (gate loops 1 and 2) surrounding the methyl donor SAM and acceptor base A pockets show varying degrees of flexibility upon SAM and SAH binding from their original positions in a ligand-free (apo) form^14–16. Thus, we compared the m⁶A structure with three states (SAM, SAH, and apo). These loops also move in opposite directions upon m⁶A binding from their original positions in the SAM-bound METTL3 (Fig. 3a). Gate loop 1 (aa 398-409) moves ~ 5.7Å inward to the direction of m⁶A, whereas the gate loop 2 (aa 506-512) moves ~ 7.8Å outward, with several residues in this region, including H512, that flips ~180º. The invariant T433 and G434 from a small loop between β3 and α2 move ~ 2.1Å with the side chain of T433 rotating ~90º to stabilize the phosphate and ribose of m⁶A (Fig. 3a). This region remains unperturbed in SAH-bound METTL3, suggesting the m⁶A binding to this pocket occurs after hydrolysis of SAM (Fig. 3b). While the gate loop 2 in SAH remains in open confirmation, akin to SAM conformation, the position of gate loop 1 in m⁶A experiences significant repositioning of the M402 side chain (Fig 3c). Although m⁶A-bound METTL3 is most similar to the apo form with the smallest root mean square deviation for superposition of 1539 atoms of METTL3 achieved for apo (1.2), compared to SAH (1.5), and SAM (1.9), we observed notable changes in the m⁶A pocket (Fig. 3c).

The side chain of M402 from gate loop 1 in the m⁶A structure stacks over the purine ring of m⁶A. In the SAM-bound form, this region is moved >5Å away, but in the SAH and apo forms, the M402 side chain will sterically clash with m⁶A ribose (distance between Cε of M402 and C4’ of ribose ~1.2 Å). To avoid this clash, the side chain of M402 in m⁶A-METTL3 rotates > 45º, resulting in a ~3.8Å gain in the distance for the Cε atom compared to its position in the apo structure. As a result of this repositioning, the inter-gate area between interface loop (H474) and gate loop 1 (M402) becomes wider, from 6.8Å in apo to 8.0 Å in the m⁶A structure (Fig. 3d). Thus, gate loop 1 from one side senses the SAM and targets the RNA base at the point of catalysis (³⁹⁵DPPW³⁹⁸ motif). It then swivels after SAM hydrolysis to facilitate the sensing of m⁶A status of the target base at the opposite or exit site.

Another change occurs in how the side chain of invariant R298 (METTL14) orients within the arginine clasp. The R298 side chain rotates ~180º around its C_β, although the guanidino group shifts slightly to form a direct h-bond with N¹ of m⁶A (Fig. 3d, Extended data Fig. 1g). The orientation of the gate loops suggests that m⁶A-METTL3 represents a state of enzyme post-catalysis before release of any product or enzyme reset.

How does m⁶A swivel ~16Å from the point of catalysis to occupy this novel pocket in METTL3? To answer this question regarding the mechanism of base-swiveling, we once again employed supervised molecular dynamics (SuMD) simulations. The simulations successfully captured the transition of the m⁶A from the substrate A pocket to the product m⁶A binding pocket. The phosphate group of m⁶A makes h-bond interactions with H478 (Extended data Fig. 4a). This interaction acts like a hinge and anchors m⁶A, which then allows the nucleotide base in m⁶A to swivel and eventually occupy the product m⁶A binding pocket. Once in the product m⁶A binding pocket, the nucleotide base of m⁶A can form h-bond with R298 (Extended data Fig. 4a). The most stable structure of the m⁶A in the product m⁶A binding pocket displays a root-mean-square deviation (RMSD) of ~ 2 Å with the resolved crystal structure (Extended data Fig. 5). This close structural alignment with the crystallographic data highlights the significance of the observed transition mechanism and corroborates with the H478A mutation where the RNA methyltransferase activity is completely abolished in the METTL3-METTL14 complex. To further validate these interactions, we calculated the interaction energy between m⁶A and the METTL3-METTL14 complex (Extended data Fig. 5-7). The results showed significant binding affinity and stability, supporting the observed h-bond interactions and the overall structural integrity of m⁶A within the m⁶A binding pocket (Extended data Fig. 7).

To verify our findings, we superposed m⁶A-METTL3 over the structure of Arabidopsis METTL4, a member of the subclade of the MTA-70 family that possesses the substrate 2’-O methyladenosine (A_m)³⁶. The central β-sheet and the catalytic motif of the two enzymes (DPPW) overlay very closely. In this model, the acceptor N⁶ atom of Am resides at ~3Å or less from the methylsulfonium group of SAM for S_N2 mechanism of direct methyl transfer. The phosphates of A_m and m⁶A lie in close proximity (~1.2Å). However, their purine and ribose rings are rotated ~120º in opposite directions, suggesting the base (A) pivots after conversion into m⁶A (Fig. 3e). Such a rotation may necessitate the de-stacking of the target base for its presentation to catalytic pocket and or base swiveling.

A water molecule at the putative site of the substrate A base is present in the m⁶A structure to compensate for the loss in binding energy in the emptied site by rotation of m⁶A from this site post-catalysis. This water coordinates with K459, and its mutation to alanine abolishes the methylation activity (Fig. 1c). SAM-dependent DNA methyltransferases, including the ancestral members of MTA-70 family MTases such as EcoP15I, efficiently flip the target adenine base out of the DNA helix into the catalytic pocket³⁷. Although METTL3-METTL14 does not methylate dsDNA and dsRNA^11,12, it can still de-stack the target base from a single-stranded RNA into the catalytic pocket, similar to the m⁶A/m⁶A_m eraser enzyme, FTO^38,39, and the m⁶A reader, YTHDC1⁴⁰. To test this activity, we replaced the target A (6-aminopurine) in a GGACU in a 14-mer ssRNA with 2-aminopurine (2Ap), a fluorescent nucleobase used as a conformational probe due to its high sensitivity to changes in the local environment induced by DNA⁴¹ and RNA MTases⁴². As shown in Fig. 3f, the change in fluorescence intensity (at 371nm) with increasing enzyme concentrations was rapid for WT but not the R298P mutant enzyme, confirming the diminished RNA binding and base de-stacking ability of the mutant enzyme. While the cellular impact of R298P mutation has been studied in the context of cancer³³, our data now uncover a precise mechanistic role of R298 in m⁶A sensing. Thus, an elegant orchestration of loops surrounding the SAM/SAH, substrate A, and product m⁶A binding pockets enables m⁶A base swiveling and sensing (see Movie 1).

To capture an AMP-bound METTL3-14, which could serve as a control structure, we extensively attempted co-crystallization and soaking of AMP into apo crystals but could not observe AMP around the catalytic pocket (DPPW or motif IV) or the m6A binding pocket. We reason the highly mobile nature of the substrate adenine base in the catalytic pocket for this. To gain insights into this phenomenon, we examined the Am binding of the published structure of Arabidopsis METTL4 (PDB: 7CV6)³⁶. The slightly low occupancy coupled with high average B-factors (B = 177 Ų) of Am base in Arabidopsis METTL4 suggest a highly mobile nature of the Am base in the catalytic pocket (Extended data Fig. 2a). Furthermore, Am’s RSCC (real space correlation coefficient) score in Arabidopsis METTL4 is 0.66. A value of RSCC score below 0.8 indicates a modest fit⁴³. In contrast, the m6AMP in our hMETTL3-METTL14 methyltransferase core structure fits nicely into the electron density at full occupancy (1.0) (Extended data Fig. 1 d-f). Consistently, an RSCC score of 0.85 with low average B-factors (B = 70 Ų) confirms a relatively stable mode of m6A binding to hMETTL3-METTL14 (Extended data Table 1 and PDB validation report).

While our work was in preparation, Corbeski et al. reported crystal structures of METTL3-METTL14 MTase core in complex with synthetic bisubstrate analogs (BAs), wherein methyl donor SAM was covalently linked to the substrate adenosine. These analogs may represent a transition state during methyl transfer⁴⁴. This study suggests that the substrate nucleotide-bound structures of METTL3-METTL14 could only be resolved when the N⁶ of adenosine was covalently attached to SAM via a 2-3 carbon linker. Interestingly, despite spatial restriction imposed by covalent linkage, the substrate adenosine moiety still exhibits significant flexibility within and across crystals obtained from soaking two different analogs, e.g., BA2 and BA4 (Extended data Fig. 2b, c). For example, the adenosine moiety of the covalent analog BA2 samples two different orientations; in a ‘buried’ conformation, it occupies the substrate pocket of METTL3 and interacts with E481 and K513, whereas in an ‘alternate conformation,’ it rotates ~90º and exposes to the solvent. In the BA4-METTL3-14 structure, the covalently attached substrate adenosine rotates into solvent-exposed orientation while the cosubstrate SAM remains rigid and occupies the SAM-binding pocket in both structures (Extended data Fig. 2b, c). Importantly, the movement of the adenosine in BA-analogs will be limited due to the covalent linkage to SAM, which naturally has a high affinity. However, the adenosine continuously tries to move in and out of the catalytic pocket. We believe these structures could serve as a control where adenosine is captured in the methyl acceptor state of the N⁶, precisely the way we modeled it by comparing the METTL4-Am structure. Moreover, the gate loop 1 and the interface loop are disordered in METTL3-14 bisubstrate analog structures, most likely due to the lack of stabilizing interactions. In contrast, these regions of METTL3-14 are well resolved in our m6AMP-bound structure due to their stabilizing interaction with m6AMP. Altogether, these observations suggest a highly mobile nature of the substrate adenine (A) base in the catalytic pocket.

The center of mass distance measurements in the simulations revealed that m⁶A consistently maintained the shortest average distance of ~1.5 Å with low variability, indicating a strong and persistent binding interaction. In contrast, other nucleotides exhibited greater and more variable distances, typically exceeding 2 Å and occasionally reaching up to 5 Å, especially AMP, CMP, and GMP, suggesting weaker and less stable interactions. AMP, CMP, and GMP showed the largest RMSD values, > 4 Å, indicating the least stability in their binding positions (Extended data Fig. 3b). This instability can be attributed to the inability of M402 to form hydrophobic interactions with the methyl group of m6A. In AMP, the loss of the methyl group prevents M402 from stabilizing the nucleotide. Although AMP can still form strong interactions with the complex, the RMSD plot indicates significant flexibility within the binding pocket. The results from SuMD are in excellent agreement with our experimental findings, which highlight the specific binding preference of the METTL3-METTL14-core for m⁶A.

The m⁶A-METTL3-METTL14 structure allowed us to gain valuable insights into how m⁶A writer (METTL3-METTL14), eraser (FTO), and reader (YTHDC1) proteins accommodate m⁶A. We examined their m⁶A pocket in detail (Fig. 4a-c). Despite the lack of obvious resemblance at the protein sequence, domain, and structure levels, we observed high similarity in the interaction networks of m⁶A in METTL3-METTL14 to the binding mode of 6mA in FTO (PDB: 5ZMD) and m⁶A in YTHDC1 (PDB: 4R3I) (Fig. 4). Of note, the purine ring of 6mA in FTO stacks between a hydrophobic amino acid, L109 (the equivalent of M402 in m⁶A), from the top and the backbone atoms of V228, S229, and H231 (the equivalent of T472, G473, and H474 in m⁶A) from the bottom. Interestingly, the arginine clasp we found in m⁶A-METTL3-METTL14 is also present in 6mA-FTO. Notably, the side chain of R96 in FTO forms a direct h-bond to N¹ of 6mA, while the guanidino group of its R322 residue forms a van der Waals interaction with N⁶ methyl group, akin to identical interactions by R298 and R471 to stabilize m⁶A in m⁶A-METTL3-METTL14. Stacking interactions that lock the sugar moieties in place also display similarities. For example, the sugar of 6mA in FTO stacks between I85 and H231, whereas the sugar of m⁶A stacks between I400 and H478 of METTL3 (Fig. 4a, b).

We found that m⁶A in METTL3-14 and YTHDC1 (PDB: 4R3I) had many similarities and striking differences, mainly in the orientation of the base (Fig. 4c). The N¹ of m⁶A forms an h-bond with N367, whereas an h-bond with carbonyl of S378 akin to carbonyl of R471 of METTL3 stabilizes the N⁶. Additional hydrophobic interactions from W377 and W428 also support the N⁶ methyl group in YTHDC1. The nature of stacking interactions for the purine ring is also similar, i.e., hydrophobic residues M434, L380, and L439 on one side and backbone atoms of K361, S362, and N363 on the other. However, the orientation of the m⁶A base in YTHDC1 is reversed by 180º compared to 6mA in FTO and m⁶A in METTL3-14. As such, when the direction of sugars and phosphates of modified bases is aligned in three structures (facing downward in Fig. 4a, b, and upper panel of c), the hydrophobic residues (M434/L380/L439) in YTHDC1 stack from the bottom side and the backbone atoms of K361, S362, and N363 stack from the top side, in contrast to the base orientation in FTO and METTL3. A ~180º rotation of YTHDC1 will place the interacting residues in all three proteins in the same plane. However, the orientation of ribose and phosphate of m⁶A in YTHDC1 will also be reversed (facing upward, Fig. 4c lower panel). Thus, a m⁶A reader protein approaches the m⁶A entirely differently than a writer or eraser. This unique geometric difference may allow the reader to avoid clashes with a writer or eraser enzyme acting simultaneously on the same transcript.

There could be two scenarios for stabilizing m6A in the cryptic pocket: a. The pocket may have the capacity to bind to all nucleotides, but m6A can outcompete other nucleotides due to its higher binding affinity to the MTase core. b. The pocket may exclusively interact with m6A while nucleotides flanking A/m6A weakly interact with other RNA binding domains of METTL3-14. Consequently, mutations near the pocket could disrupt the binding to m6A or alter the preference from m6A to other nucleotides. We show that METTL3 possesses features that enable it to act as an atypical m⁶A sensor/reader – a function ideally suited for its emerging non-catalytic functions, including crosstalk with eIF3H to promote mRNA circularization, thereby enhancing RNA translation as observed in lung cancer^23,25 and bone marrow mesenchymal stem cells²¹. Consistently, METTL3 exhibits a strong affinity to a methylated (m⁶A) RNA form, especially those containing secondary structures, a feature that could be important for its non-catalytic roles such as an ‘m⁶A reader’ for an alternative mode of translation initiation during oncogenic translation^23,25 and cellular stress (e.g., heat shock)³⁴.

The gene encoding the MTase domains of human METTL3 (aa 357-580aa) and METTL14 (aa 116-402) were cloned into a pETduet-1 vector and expressed in E. coli NiCo21(DE3) cells. The transformed cells were grown in Terrific Broth medium supplemented with 1 mM ampicillin at 37°C until OD_600nm reached 0.6. Protein expression was then induced by adding 0.4 mM isopropyl β-D-thiogalactopyranoside, and the culture was grown at 18°C for 16 hrs. The cell pellets were harvested by centrifugation at 6000 r.p.m. at 4°C and resuspended in cold lysis buffer containing 25 mM Tris pH 8.0, 0.5 M NaCl, 10% glycerol, 5 mM imidazole, 0.1 mM TCEP, one tablet of protease inhibitor (Roche), lysozyme (0.1 mg/mL), and DNase I (5U/mL) and stirred gently at 4°C until achieving full homogeneity. Resuspended cells were lysed by two passages through a microfluidizer (Analytik, UK) and subjected to centrifugation at 41,000 r.p.m. for 50 min at 4°C. The clarified supernatant was filtered through a 0.22 μm filter and loaded onto a Nuvia IMAC column (Bio-Rad) pre-equilibrated with wash buffer (25 mM Tris pH 8.0, 0.5mM NaCl, 10% glycerol, 5 mM imidazole, and 0.1 mM TCEP). The His-tagged METTL3 was co-eluted with untagged METTL14 by increasing the imidazole concentration. The eluates were dialyzed in a buffer lacking imidazole overnight at 4ºC in the presence of the ULP1 enzyme to remove the His-SUMO tag from METTL3 proteolytically. The dialyzed proteins were then re-loaded onto an IMAC column to remove un-cleaved proteins and the His-SUMO tag. Two successive passages through MonoQ and Hiload Superdex75 columns (Cytiva) further purified the tag-free complex. The fractions of a homogenous peak eluted in 20 mM Tris pH 8.0, 0.2 M NaCl, and 0.1mM TCEP were pooled, concentrated to 15 mg/ml, and either used immediately or flash-frozen in liquid nitrogen and then stored at -80°C.

The full-length human METTL3 and METTL14 (wild-type and mutants) were expressed in insect cells (ExpiSF Expression System, Thermo Fisher) and purified using a protocol published earlier¹². In brief, the METTL3 and METTL14 plasmids were transformed individually into Max Efficiency DH10Bac competent cells (Thermo Fisher) to generate the DNA bacmids. The successful insertion of genes was confirmed by PCR amplification using a pUC/M13 primer (Forward: 5’-CCCAGTCACGTTGTAAAACG -3’, Reverse: 5’ – AGCGGATAACAATTTCACACAGG -3’). The amounts of purified bacmids and ExpiFectamine SF transfection reagent (Thermo Fisher) were optimized as per the manufacturer’s recommendations (Thermo Fisher). The ExpiSf9 insect cells were cultured in ExpiCD medium (Thermo Fisher) at 125 r.p.m. and 27°C in a non-humidified, air-regulated environment. The cells were harvested 72 hrs post-infection by spinning at 300 × g for 5 min. The PBS-washed cells were resuspended in cold lysis buffer containing 0.5% Igepal, two tablets of protease inhibitor (Roche), and DNase I. Cells were lysed by passing through a microfluidizer (Analytick, UK) and clarified by centrifugation at 41,000 r.p.m. for 40 min.

The proteins were purified using a strategy similar to that of the MTase core, except for removing the His-tag from METTL3. This step was achieved by incubating proteins after the affinity column step with TEV protease for 3 hrs at room temperature. A second passage through a nickel IMAC column removed contaminants and any uncleaved fractions. The complex was then successfully purified by successive passages through HiTrap Heparin and Hiload Superdex 200 columns (Cytiva). Eluates from a homogenous peak of a Superdex column run in a buffer of 0.02 M Tris pH 8.0, 0.15 M NaCl, and 5% glycerol were pooled, concentrated to 1-3 mg/ml, and flash-frozen in liquid nitrogen and stored at -80°C. All full-length METTL3-METTL14 mutants (METTL3: T433A, K459A, R471H, Δ472-473; Δ472-474, H478A, E481A; METTL14: R298P) were generated by site-directed mutagenesis and purified by the same method as the wild-type protein.

The crystallization of the human METTL3-METTL14 MTase core (at 10 mg/mL concentration) was carried out by an OryxNano robotic system (Douglas Instruments) using the sitting-drop vapor diffusion method at 20°C. Initial crystals were grown in a solution containing 0.1 M MES pH 6.0, 1.0 M potassium sodium tartrate. After several rounds of optimization by varying pH and salt concentrations, large reproducible crystals were grown in seven days. The N⁶-methyladenosine monophosphate (m⁶AMP; Sigma, M2780) was soaked into native crystals (2.0 mM concentration) for 1 hr at 20°C. A complete diffraction dataset was measured to ~2.5Å at GMCAT 23ID-D beamline at Advanced Photon Source, Chicago, IL. The apo structure of METTL3-METTL14 MTase core (PDB: 5IL0) was used as a search model for molecular replacement in Phenix⁴⁵. The structure was iteratively built and refined using Coot (Version 0.9.8.6)⁴⁶, Phenix (Version 1.15.2-3472) and Buster (Version 2.10.4)⁴⁷, respectively. The ligand geometry restraints were generated by Grade. All structure figures were generated using Pymol (Schrodinger Suite).

5 μM [methyl-³H] SAM (PerkinElmer), 10 μM substrate RNA (NEAT2*: 5’ – GCCUAGUAGCAGAGAGGACUGCUCCUUGGU - 3’), and 2 μM purified WT or mutated METTL3-METTL14 were mixed and incubated at 37°C for 1 hr in a total volume of 5 μL in a reaction buffer (50 mM HEPES pH 7.5, 5 mM NaCl, and 1 mM DTT). The reactions were quenched by blotting 3 μL of each on the Hybond-N+ membrane (Amersham). The methylated substrates were then crosslinked by exposing them to ultraviolet light (254 nm) for 2 min. The membranes were washed three times with 1X PBS, followed by two 95% ethanol washes. Then the membranes were air-dried inside the hood for 15 minutes, and the RNA probe’s count per minute (c.p.m.) on each membrane were measured by a scintillation counter (Beckman LS6500). All results are reported as the means from three independent experiments (n=3) for each group, with one standard deviation (s.d.).

The reactions were carried out in a buffer containing 10 mM HEPES pH 7.5 and 50 mM KCl. The two 30-mer RNA probes (native RNA or A-RNA and its m⁶A-modified version or m⁶A-RNA) were synthesized with a fluoresceine moiety covalently attached to their 5’-end, de-protected, and purified using HPLC (HorizonDiscovery). The sequence of A-RNA was identical except the target A base within the characteristic motif (underlined and bold) in native A-RNA (Fl-NEAT2*: 5’ – [Fl]GCCUAGUAGCAGAGAGGACUGCUCCUUGGU - 3’), was replaced by N⁶-methyladenosine (m⁶A) in the modified RNA (Fl-NEAT2*-m⁶A: 5’ – [Fl]GCCUAGUAGCAGAGAGG[m⁶A]CUGCUCCUUGGU - 3’). A constant 5 nM of RNA probes were incubated with increasing concentrations of the purified WT or mutant METTL3-METTL14 enzymes in a 384-well plate. The fluorescence polarization values (excitation wavelength = 485 nm, emission wavelength = 530 nm) of each reaction were measured by PHERAstar FS (BMG Labtech). The affinity of RNA-protein binding was calculated by a simple one-site specific binding model (Y = Bmax*X/(Kd+X), X = protein concentration, Y = specific binding, Bmax = maximum specific binding, Kd = equilibrium dissociation constant). The results were analyzed and fitted by GraphPad Prism (GraphPad Software, San Diego, CA). Each experiment was repeated three times independently (n=3), and final K_d is reported as the mean of the three replicates with standard deviation (s.d.) for each RNA shown as error bars.

For this experiment, we used a 14-mer single-stranded RNA probe in which the the target adenine base within the m⁶A motif was replaced with 2-aminopurine (2-Ap) (r6T*: 5’ – CUUCGG[2-Ap]CUCUGCU – 3’). In a 384-well plate format, 0.5 μM of RNA probe mixed with increasing concentrations (1 – 5 μM) of full-length human WT or mutant METTL3-METTL14 enzymes in a 20 μL reaction in the buffer containing 50 mM Tris pH 8.0, and 10 mM MgCl₂ and incubated at room temperature. The reaction was excited at 320 nm with a 325 nm cut-off wavelength in a SpectraMax M5 microplate reader (MolecularDevices). The fluorescence emission was measured at 371 nm and 37°C every 5 minutes from 0 - 60 minutes and then every 30 minutes until the end time point (120 minutes). The data were analyzed and fitted by GraphPad Prism (GraphPad Software, San Diego, CA) using the Michaelis-Menten model (Y=Vmax*X/(Km+X), X = protein concentration, Y = enzyme velocity, Vmax = maximum enzyme velocity, Km = Michaelis-Menten constant). All results reported are mean values from three independent experiments with standard deviations (s.d.) shown as error bars.

The surface plasmon resonance experiments were performed using a Biacore 1S+ equipped with a CM5 sensor chip. Recombinant streptavidin (Millipore Sigma, 11721674001) was first immobilized at all six flow cells using amine-coupling chemistry. The surfaces of flow cells were activated for 7 min with a 1:1 mixture of 0.1 M NHS (N-hydroxysuccinimide) and 0.4 M EDC (3-(N, N-dimethylamino) propyl-N-ethylcarbodiimide) at a flow rate of 10 μl/min. The streptavidin at a concentration of 50 μg/ml in 10 mM sodium acetate, pH 5.5, was immobilized at a density of around 6,000 RU on all six flow cells. All surfaces were blocked with a 7-minute injection of 1 M ethanolamine, pH 8.0. The biotinyl-RNA substrates were then diluted to 100nM in running buffer (10mM HEPES, 150mM KCl, 0.05% P20, pH 7.5) and captured in flow cell 3-6 respectively (12s, 5 μl/min) to reach a level around 50RU (see details below). To measure kinetic binding data, the analytes (METTL3-METTL14_WT, METTL3-METTL14_R298P, and METTL3-METTL14_R471H proteins) were diluted in the same running buffer, with five concentrations ranging from 2.4 to 1500 nM. The analytes were then injected over all flow cells at various concentrations in single-cycle kinetics format at a flow rate of 30 μl/min at 25°C. The analyte was allowed to associate and dissociate for 120 seconds for each injection and dissociate for 1200 seconds, respectively. Data were collected at a rate of 10 Hz. The data were fit to a 1:1 binding model or two-state reaction model, as mentioned, using the data analysis option available within Biacore Insight Evaluation Software (Version 5.0.18.22102).