m¹A is not readily detected in mRNA

In order to detect m¹A, the modified residue must be reverse transcribed a sufficient number of times during library preparation to generate misincorporations. Since m¹A sites were reported to have on average a 20% stoichiometry³, we set a threshold of 500 unique reads on any given nucleotide to detect m¹A sites. At this stoichiometry, 100 reverse transcription events would encounter m¹A. Of these 100 reverse transcription events, approximately 20% would read through, and most of these would be associated with a misincorporation^10,11. At this read depth, misincorporations should therefore be readily detected in multiple replicates. Thus, to detect m¹A in mRNA, we restricted our search to approximately eight million adenosine residues in the transcriptome that showed a read depth of >500 reads (Supplementary Fig. ^3a).

Analysis of the three billion reads showed 14 high-confidence nucleotide positions across the transcriptome with misincorporations in more than one replicate (see Methods). Of these, 12 occurred at adenosine residues (Supplementary Data ¹ and ²). Most of these modified adenosines were found in mitochondrial tRNAs and occurred at known m¹A positions in mitochondrial tRNAs¹⁵ (Supplementary Data ²). We also detected a modified adenosine in mascRNA (MALAT1-associated small cytoplasmic RNA), a short tRNA-like ncRNA that is derived from endonucleolytic processing of MALAT1 (ref. ¹⁶) (Fig. 1c). Notably, this modified adenosine corresponds to position 58 within the T-loop of tRNAs (Fig. 1d), a conserved m¹A site in tRNAs¹⁷. This m¹A site in MALAT1 may be similarly formed by T-loop-specific m¹A-synthesizing enzymes¹⁷.

Besides these noncoding RNAs, the previously reported⁷ m¹A-containing MT-ND5 mitochondrial mRNA also contained a modified adenosine (Supplementary Data ²). This adenosine exhibited a misincorporation rate of 13.5%. Thus, misincorporation mapping resulted in the same major conclusion as previously reported⁷ that MT-ND5 is the major m1A-modified mRNA in the cell.

We next examined misincorporations at other mitochondrial mRNAs with previously annotated m¹A sites. Safra et al.⁷ and Li et al.⁸ identified 11 and 5 putative mitochondrial m¹A-containing protein-coding genes, respectively. Four mitochondrial mRNAs were common to both studies (MT-ND5, MT-CO1, MT-CO2, and MT-CO3). The misincorporation rates in poly(A) RNA-Seq for these mRNAs were very low (less than 0.7% and 2.1% in the Safra and Li studies^7,8, respectively). However, misincorporations could be detected when m¹A-containing mRNAs were enriched using the m¹A antibody, suggesting that m¹A-containing transcripts are indeed present in cells, but are so rare that they require enrichment to be detected. We therefore leveraged the exceptional mitochondrial read depth in our ultra-deep RNA-seq samples (average ~1 million reads/nucleotide). Here we found that with the exception of MT-ND5 which had a misincorporation rate of 13.5%, the misincorporation rates for all other putative m¹A sites was less than 0.4%, which is close to the background rate (Supplementary Fig. ^3b, Supplementary Data ³). This demonstrates that the baseline stoichiometry of m¹A is very low in all mitochondrial mRNAs except for MT-ND5, and m¹A detection requires a pre-enrichment step due to its exceptionally low stoichiometry.

We also detected three additional sites of modifications in cytosolic mRNAs, with only one site being a modified adenosine (Supplementary Data ²). Therefore, it is unlikely that there are many high-stoichiometry hard-stop nucleotides in mRNA.

Notably, when we examined cytosolic mRNAs reported³ to have the highest stoichiometry of m¹A (i.e. >50%) such as CCDC71, DLST, and STK16, each lacked A→T transitions based on misincorporation mapping (Supplementary Fig. ^3c, Supplementary Data ¹).

m³C is also unlikely to be present at high stoichiometry in mRNA since we could not detect any high-confidence misincorporations at cytidine residues (Supplementary Data ²). The m³C previously detected in mRNA by mass spectrometry¹⁸ may therefore have originated from contaminating tRNA.

Although misincorporation mapping demonstrated that few mRNAs have m¹A or other hard-stop modifications, an important caveat is that the 5′ end of mRNAs cannot be assessed. This is because RNA-seq typically provides less coverage at 5′ ends of RNAs (see Supplementary Fig. ^1a)¹⁹. Overall, the paucity of m¹A sites in mRNA using misincorporation mapping demonstrates that m¹A is not a prevalent high-stoichiometry modification in mRNA.

m¹A-miCLIP detects known m¹A sites at nucleotide resolution

To understand why m¹A antibody-based mapping approaches produce a prominent 5′UTR signal, we developed an approach to detect m¹A at single-nucleotide resolution: m¹A-miCLIP (m¹A-modification individual nucleotide resolution crosslinking and immunoprecipitation) (Fig. 2a). We initially used the AMA-2 m¹A antibody previously used to map m¹A sites in mRNAs^3,4. In m¹A-miCLIP, the m¹A antibody is crosslinked to sheared RNA (Fig. 2a). UV crosslinking with stringent washing reduces nonspecific RNA binding and increases peak resolution in mapping studies²⁰. RNA fragments crosslinked to the antibody are then purified and cloned as a cDNA library. Terminations introduced during reverse transcription are then analyzed to localize precise sites where the m¹A antibody binds throughout the transcriptome. Prior to performing transcriptome-wide m¹A mapping, we confirmed the AMA-2 m¹A antibody binds m¹A, and not other nucleotides (Supplementary Fig. ^4a).

Open in a separate window

Fig. 2

Mapping m¹A in the transcriptome using m¹A-miCLIP. a m¹A-miCLIP workflow. Poly(A) RNA (black lines; black circles, RNA caps) was fragmented, incubated with m¹A antibody, and crosslinked to the antibody with UV. RNA-antibody complexes were immunoprecipitated, ligated to a 3′ linker (blue line), and purified. RNA was released from crosslinked antibody using Proteinase K, generating RNA fragments with peptide adducts. After reverse transcription, first-strand cDNA (gray lines) was circularized to preserve m¹A-induced read terminations (magenta dot) at the cDNA 3′ end. b m¹A-miCLIP detects the known m¹A site in the 28S rRNA. m¹A-miCLIP unique reads (light blue) were aligned to rRNA showing high enrichment at the m¹A site (red circle) at position 1322 of the 28S rRNA. The reads terminated predominantly at the +1 position relative to m¹A (inset, dark blue). c Relative distribution of m¹A-miCLIP clusters. The metagene distribution of m¹A-miCLIP clusters that contained 20 or more stacked reads (n=3877) was normalized to RNA-seq coverage and plotted. Clusters were predominantly enriched in the 5′UTR and next to the transcription-start site (start, start codon; stop, stop codon). The annotation of transcription-start sites is not exact, which results in slight coverage upstream of the transcription-start site. d m¹A-miCLIP coverage is enriched at and near transcription-start sites. m¹A-miCLIP unique reads were mapped relative to annotated transcription-start sites using Deeptools (see Methods). e While RNA-seq reads terminate predominantly at position 0 (i.e. due to the end of the transcript), m¹A-miCLIP reads terminate predominantly at positions 0 and +1, and to a lesser degree, at downstream nucleotides, suggesting that the m¹A antibody binds at or near the transcription-start site and blocks reverse transcription. Error bars represent SEM. Source data are provided as a Source Data file. f m¹A-miCLIP clusters found in 5′UTRs showed significant enrichment of the pyrimidine-rich Initiator sequence motif that facilitates transcription initiation of mRNAs at adenosine (E=3.3×10⁻⁵). g m¹A-miCLIP clusters were overlapped with a collective set of transcription-start sites (TSS) that included RefSeq annotations and m⁶A_m mRNA extended caps⁴⁴. While 33.2% of these collective transcription-start sites were adenosines, the frequency of adenosine transcription-start sites overlapping m¹A-miCLIP clusters was significantly higher (85.2%, n=1,026 TSSs, p<2.2×10^-16, Fisher’s exact test)

m¹A-miCLIP differs from earlier methods^3,4 by preserving the cDNA 3′ ends. Any m¹A antibody-binding site at the transcription-start site would produce a peak that is displaced in the 3′ direction (see Supplementary Fig. ^1a). This can make a m¹A antibody-binding site at the transcription-start site appear to be located at the start codon. To avoid this problem, we chose to generate our libraries in a way that preserves the cDNA ends. Therefore, m¹A-miCLIP reveals exact m¹A antibody-binding sites within transcripts.

We performed m¹A-miCLIP with the AMA-2 antibody using poly(A) RNA from HEK293T cells and examined termination signatures at known m¹A sites in rRNA and tRNA, which typically co-purify to some extent with poly(A) RNA^21–23. The majority of reads truncated at the +1 position relative to the m¹A at position 1322 of the 28S rRNA (Fig. 2b, Supplementary Fig. ^4b) as well as known m¹A sites in tRNA (Supplementary Fig. ^4c). As expected, some read-through was also observed, reflecting the low read-through rate of SuperScript III when it encounters m¹A. These data demonstrate specific detection of m¹A by m¹A-miCLIP using the AMA-2 m¹A antibody. Notably, m¹A-miCLIP showed markedly improved peak resolution compared to the initial peak-based m¹A mapping studies^3,4 (Supplementary Fig. ^4b).

We next asked if m¹A detection in m¹A-miCLIP is adversely affected by m¹A conversion to m⁶A via the Dimroth reaction. m¹A-miCLIP does not use high temperatures and basic pH¹³ which are needed for this conversion (see Supplementary Fig. ^2d). No m⁶A-miCLIP reads²⁴ could be detected at the 28S rRNA m¹A site (Supplementary Fig. ^4d), demonstrating that m¹A does not appreciably converts to m⁶A during the m¹A-miCLIP protocol. Taken together, these data demonstrate that m¹A-miCLIP maps m¹A with high specificity and resolution.

The AMA-2 m¹A antibody binds near the first mRNA nucleotide

We next used m¹A-miCLIP to map m¹A in mRNA (Supplementary Fig. ^5a). To do this, we aligned m¹A-miCLIP unique reads to the genome, generated m¹A-miCLIP clusters (see Methods, Supplementary Data ⁴), and analyzed their distribution. A metagene analysis of m¹A-miCLIP clusters obtained from HEK293T cells showed a marked enrichment in the 5′UTR (Fig. 2c). More precisely, the clusters were located at mRNA transcription-start sites (Fig. 2d). A similar enrichment was seen using mouse mRNA (Supplementary Data ⁵, Supplementary Fig. ^5b).

Crosslinking of the m¹A antibody is expected to cause reverse transcription terminations within several nucleotides of the site of the antibody-RNA adduct²⁰. As expected, we found that in miCLIP, terminations were enriched not only at the transcription-start site, but also prominently at the +1 position relative to the transcription-start site (Fig. 2e), with additional terminations sometimes seen between position +2 and +3 (Fig. 2e). Thus, the AMA-2 m¹A antibody binds at or near mRNA transcription-start nucleotides.

We considered the possibility that terminations near the transcription-start nucleotide could simply reflect general behavior of the reverse transcriptase as it approaches the mRNA 5′ end. To test this, we examined the input RNA fragments in the RNA-seq dataset prepared using the same library cloning strategy as m¹A-miCLIP. In general, RNA-seq reads terminated almost exclusively at the transcription-start nucleotide (Fig. 2e, Supplementary Fig. ^5c). Therefore, read terminations seen in m¹A-miCLIP near the transcription-start nucleotide are likely induced by selective binding and crosslinking of the AMA-2 m¹A antibody rather than an artifact of reverse transcription near mRNA 5′ ends.

m¹A-miCLIP motif analysis revealed a consensus sequence in the upstream genomic region that was pyrimidine-rich (Fig. 2f) and highly similar to Initiator, a transcription-initiating sequence which produces transcripts that initiate with adenosine^25,26. Indeed, 85% of transcripts containing an m¹A-miCLIP cluster at their transcription-start nucleotide initiated with adenosine (Fig. 2g, Supplementary Data ⁶). Thus, we reasoned that adenosine at the transcription-start nucleotide was important for binding of the AMA-2 antibody at or near the transcription-start site.

m¹A and m¹A_m are not detected in extended cap structures

Because the AMA-2 m¹A antibody binds at transcription-start sites, it has been proposed that mRNAs contain a novel mRNA cap structure comprising m⁷G followed by N¹-methylated adenine at the transcription-start nucleotide⁸. The methylated nucleotide would be m¹A or N¹,2′-O-dimethyladenosine (m¹A_m) since the first encoded nucleotide of mRNAs is typically subjected to 2′-O-methylation^27,28. To biochemically validate this, we used mass spectrometry to detect m¹A or m¹A_m in “cap dinucleotides,” i.e., m⁷G-ppp-m¹A_m. We treated cellular RNA with P1 nuclease, which digests internal nucleotides to mononucleotides, but leaves the cap dinucleotide intact (see Methods).

We readily detected diverse cap dinucleotides using high-resolution liquid chromatography and mass spectrometry using positive ion mode detection. We developed a multiple reaction monitoring protocol based on the fragment ion transitions from distinct dinucleotide precursor species (see Methods). To confirm that the N1-methylated adenosine in a cap dinucleotide can be detected, we used synthetic RNA standards (see Methods). Using these standards, we readily detected m⁷G-ppp-m¹A as well as other cap dinucleotides, such as m⁷G-ppp-A_m, and m⁷G-ppp-m⁶A_m (Fig. 3a).

Open in a separate window

Fig. 3

The AMA-2 m¹A antibody does not enrich for a novel m¹A-containing cap structure. a Detection of synthetic 5′ cap dinucleotides using mass spectrometry. Dynamic multiple reaction monitoring (dMRM) chromatograms were recorded in positive ion monitoring. Injected was a mixture of oligonucleotides containing m⁷G-ppp-m⁶A_m, m⁷G-ppp-m¹A, m⁷G-ppp-m⁶A, or m⁷G-ppp-A_m cap structures. Two MRM transitions were recorded for each standard (indicated in key in figure). Note the permanent positive charge on the adenine base of m⁷G-ppp-m¹A resulted in an increased retention time on the ANP resin as expected. Thus, the m⁷G-ppp-m¹A_m dinucleotide would be expected to elute at a retention time later than m⁷G-ppp-m⁶A due to the permanently positive charged adenine, and earlier than m⁷G-ppp-m¹A due to the additional methyl group on the ribose. b Detection of 5′ cap dinucleotides from cellular mRNA using mass spectrometry. Dynamic multiple reaction monitoring (dMRM) chromatograms were recorded in positive ion monitoring. Injected were mRNA samples from HEK293T cells. Two MRM transitions were recorded for each cap structure of interest (indicated in key in figure). We searched for peaks containing MRM transitions and retention times reflecting m⁷G-ppp-m¹A or m⁷G-ppp-m¹A_m (see note in a), but no such signal was detected. c 2D-TLC analysis of capped oligonucleotide standards containing various modified adenines as initiating nucleotides. The migration properties of these various initiating nucleotides are indicated (here, the initiating nucleotides include m¹A, m¹A_m, A_m, m⁶A, and m⁶A_m). Spots reflecting A, C, G, and U are also marked. Unmarked spots most likely reflect RNA dinucleotides that are not completely digested during RNA preparation⁴⁵. d 2D-TLC analysis of mRNA extended caps in antibody-unbound (left panel) or m¹A antibody-bound poly(A) RNA. Known identities of detected spots are marked. Expected positions of potential m¹A or m¹A_m-initiating nucleotides are marked (blue dashed circles) based on standard migrations in a

We next examined endogenous cap dinucleotides prepared by digesting HEK293T poly(A) RNA. We readily detected m⁷G-ppp-m⁶A_m (m/z=815.1) and also m⁷G-ppp-A_m (m/z=801.1), though to a lower degree. m⁷G-ppp-C_m (m/z=777.1), m⁷G-ppp-G_m (m/z=817.1), and m⁷G-ppp-U_m (m/z=778.1) were also detected (Fig. 3b). The identity of each species was confirmed by detection of fragment masses corresponding to 7-methylguanine (m/z=166.1) and the base comprising the first nucleotide (m/z=m⁶A_m 150.1, A_m 136.1, C_m 112.1, G_m 152.1, U_m 112.1) within the extended cap.

Next, we asked whether either m⁷G-ppp-m¹A or m⁷G-ppp-m¹A_m is present in mRNA. Their masses (m/z=801.1 and 815.1, respectively) are identical to cap dinucleotides containing A_m or m⁶A_m. Moreover, the mass of the fragment produced by the N¹-methylated adenine base (m/z=150.1) would be identical to that produced by the m⁶A_m cap. However, m⁷G-ppp-m¹A and m⁷G-ppp-m¹A_m exhibit very distinct retention times from the A_m and m⁶A_m cap dinucleotides based on our synthetic RNA standards (see Fig. 3a). This allows us to differentiate N¹-methyl- and N⁶-methyl-containing adenine. Nevertheless, no N¹-methylated adenine-containing cap dinucleotide was detected in mRNA (Fig. 3b). Thus, m⁷G-ppp-A_m or m⁷G-ppp-m⁶A_m cap structures were readily detected, while m⁷G-ppp-m¹A_(m) was undetectable.

Since mass spectrometry analysis could not validate m¹A at the transcription-start nucleotide, we wanted to directly and sensitively determine the transcription-start nucleotide that is enriched by the m¹A antibody. We used two-dimensional thin layer chromatography (2D-TLC), which can identify and quantify the first encoded nucleotide²⁹. In this approach, mRNA is decapped, and the exposed 5′ RNA end is radiolabeled, permitting sensitive detection of the first transcribed nucleotide. The radiolabeled nucleotide species are then resolved using 2D-TLC²⁹. Based in the mobility of each species, the transcription-start nucleotide can be determined.

We analyzed the transcription-start nucleotide of poly(A) RNA and poly(A) RNA enriched with the m¹A antibody by 2D-TLC. We also used synthetic RNA containing m⁷G-ppp-m¹A and m⁷G-ppp-m¹A_m extended cap structures as standards (see Methods). We first optimized the solvent conditions so that m¹A and m¹A_m migrated to distinct positions by 2D-TLC (Fig. 3c). In poly(A) RNA, no m¹A or m¹A_m was detected at transcription-start nucleotides (Fig. 3d). Since these might be rare nucleotides at transcription-start nucleotides, we enriched for transcription start m¹A or m¹A_m-containing mRNAs by immunoprecipitating poly(A) mRNA with the AMA-2 m¹A antibody before 2D-TLC. Here, we again did not see m¹A or m¹A_m as the transcription-start nucleotide (Fig. 3d).

Taken together, the mass spectrometry and the TLC data suggest that m¹A and m¹A_m are not readily detectable at the transcription-start nucleotide, and m⁷G-ppp-m¹A_m does not constitute a novel and prevalent mRNA cap structure as proposed⁸.

The AMA-2 m¹A antibody recognizes m⁷G-ppp-A cap structures

At this juncture, we had contradictory results: m¹A-miCLIP suggested that m¹A is at the transcription-start nucleotide, but we did not observe m¹A or m¹A_m at this site by either mass spectrometry or TLC. Therefore, we wondered if the AMA-2 antibody binds the transcription-start region in an m¹A-independent manner. When we originally characterized the specificity of the antibody, we performed classic competition studies using nucleosides or nucleotides. However, based on the binding properties of the antibody revealed by mapping studies, we considered the possibility that the AMA-2 m¹A antibody could recognize an epitope comprising the mRNA extended cap.

To test this, we used a dot blot assay to measure binding of the AMA-2 m¹A antibody to an m¹A-containing oligonucleotide in the presence of various competitors. We considered performing the dot blot assay using m⁷G-ppp-A immobilized on the membrane. However, this approach could be misleading since we do not know if m⁷G-ppp-A interacts with the membrane in a way that would prevent antibody binding. We therefore used the classic competition approach. In this approach, a m¹A-containing RNA is immobilized on the membrane and different competitors are added in solution. Competitors that bind the m¹A antibody will prevent the antibody from binding to m¹A on the membrane.

As expected, competition with m¹A inhibited antibody binding, while related nucleotides, including adenosine, m⁶A, ethenoadenosine, and N¹-substituted nucleotides, like N^1,6-dimethyladenosine (m^1,6A), did not compete with binding (Fig. 4a). Surprisingly, a commercially available cap analog, m⁷G-ppp-A, was a relatively effective competitor, with an IC₅₀ of 480nM compared to 100nM for m¹A (Fig. 4b). m⁷G-ppp-G showed weaker inhibition (IC₅₀ ~4 μM) (Fig. 4c). The higher binding to the m⁷G-ppp-A cap analog compared to the m⁷G-ppp-G cap analog may explain the preferential binding of the AMA-2 m¹A antibody to mRNAs that initiate with adenosine.

Open in a separate window

Fig. 4

The AMA-2 m¹A antibody binds to mRNA caps resulting in 5′UTR read enrichment. a Shown is a dot blot analysis of the AMA-2 m¹A antibody’s specificity. An m¹A-containing oligonucleotide was spotted on a nylon membrane in triplicate and the antibody was incubated with the membrane along with m¹A-like nucleosides. Only m¹A prevents antibody binding to the spotted oligonucleotide. m^1,6A N1,6-methyladenosine, εA ethenoadenosine, m¹I 1-methylinosine. b Top, images of dot blot competition assays¹ where binding of the AMA-2 m¹A antibody to an m¹A-containing oligonucleotide was competed with m¹A (m¹ATP), m⁷G-ppp-A, or A (ATP) in triplicate. Bottom, quantification of dot blot competition experiments. m⁷G-ppp-A binds (IC₅₀<1μM) to the antibody, although the binding is ~10-fold weaker than m¹A (IC₅₀=~100nM). Source data are provided as a Source Data file. c The AMA-2 m¹A antibody recognizes the m⁷G-ppp portion of the cap structure. The membrane was probed with the antibody along with m⁷G-ppp or m⁷G-ppp-G. Both m⁷G-ppp and m⁷G-ppp-G competed with the m¹A oligonucleotide for binding to the antibody (IC₅₀=~4μM for m⁷G-ppp-G, 10μM for m⁷G-ppp). Source data are provided as a Source Data file. d The Abcam m¹A antibody (ab208196) does not bind the mRNA cap. Dot blots were performed with either of the two m¹A antibodies (AMA-2 or Ab208196) in the presence of A, m¹A, or m⁷G-ppp-A (A and m¹A were either tri- or mono-phosphorylated). While both antibodies bind m¹A, only the AMA-2 antibody binds m⁷G-ppp-A. e The m¹A antibodies from Abcam (ab208196) and MBL (AMA-2) both detect established m¹A sites. m¹A-miCLIP unique reads from two independent antibodies (in replicate) were aligned to the genome, and peaks at known m¹A sites (red circle) in MT-ND5 and the mitochondrial 16S rRNA (MT-RNR2) exhibited a characteristic truncation at the m¹A site. f The m¹A antibody Ab208196 lacks the prominent 5′UTR metagene peak seen with the AMA-2 antibody. As in e, unique m¹A-miCLIP reads were aligned to the genome and the relative location of the reads in each transcript is shown with a metagene. Each trace is an independent biological replicate. The 5′UTR peak is essentially absent when the m¹A-specific Ab208196 antibody was used compared to the cap-cross-reacting AMA-2 antibody

Notably, the antibody also showed binding to m⁷G-ppp, but not m⁷G or ATP (Fig. 4c, Supplementary Fig. ^4a), suggesting that the antibody’s binding specificity includes recognition of features all along the entire m⁷G-ppp-A extended cap structure.

Since the AMA-2 m¹A antibody binds the cap structure in an m¹A-independent manner, the m¹A peaks seen in the 5′UTR likely reflect binding to the mRNA cap. This would explain why our m¹A-miCLIP shows read enrichment at the transcription-start nucleotide, as has also been seen using other m¹A mapping approaches⁸.

To test this hypothesis further, we used a second commercially available m¹A antibody from Abcam (catalog number ab208196). Unlike the AMA-2 m¹A monoclonal antibody available from MBL, the Abcam antibody did not bind the m⁷G-ppp-A cap analog (Fig. 4d), indicating that it does not exhibit cross-reactivity with the mRNA cap.

We first confirmed by m¹A-miCLIP using HEK293T poly(A) RNA that both ab208196 and AMA-2 antibodies can detect authentic m¹A sites. In each case, we observed a robust peak at the m¹A sites in MT-ND5 and MT-RNR2, the mitochondrially encoded 16S RNA (Fig. 4e), confirming the ability of both antibody to detect validated m¹A sites in mRNA.

We next asked if m¹A-miCLIP performed using the m¹A-specific Ab208196 would produce the same transcriptome-wide 5′UTR enrichment of m¹A-containing fragments observed with the AMA-2 m¹A antibody^3,4. As expected, the metagene of the miCLIP fragments using the AMA-2 antibody showed a prominent 5′UTR enrichment (Fig. 4f). However, a metagene analysis of all the immunoprecipitated reads using ab208196 lacked the 5′UTR enrichment (Fig. 4f). Together, these data demonstrate that only the AMA-2 antibody, which cross-reacts with the mRNA cap, results in a 5′UTR enrichment in read coverage. Overall, these data demonstrate that binding to the 5′UTR regions is not linked to the presence of m¹A at these sites, but rather attributable to cross-reactivity.

Antibodies

The AMA-2 m¹A antibody is a mouse monoclonal antibody (MBL catalog number D345-3). This antibody is documented to react with both m¹A within RNA and the N¹-methylated adenine base, as documented in MBL’s product specifications for AMA-2 (catalog no. D345-3) and prior studies^3,4. The specificity was validated previously^3,4 and in the current study. The Abcam m¹A antibody (ab208196) is a rabbit monoclonal generated against m¹A. Its specificity for m¹A was validated in this study.

Alignment of reads for misincorporation mapping

Raw reads from the ultra-deep RNA-seq dataset used for this study¹² were downloaded from GEO (accession code: GSE33029). This RNA-seq dataset was prepared using standard reverse transcription with SuperScript III, an enzyme expected to produce misincorporations at m¹A positions¹¹. Variants identified in the genomic DNA corresponding to this dataset were acquired from http://snyderome.stanford.edu. Coordinates of other SNPs that may be present in the DNA sequence were downloaded from the SNP database dbSNP (February 2017 build; https://www.ncbi.nlm.nih.gov/projects/SNP/). Read alignment of forward and reverse read mates was performed using STAR (version 2.5.3a) and the hg19 genome build. Alignment incorporated removal of PCR duplicates, and clipping of 10 bases on either end of each read, since the ends of Illumina reads are prone to sequencing error³⁵. Only reads that mapped to a single location in the genome were used for downstream analysis. A maximum of one mismatch per read was permitted for alignment.

Misincorporation mapping

To identify misincorporations, aligned reads were analyzed using Rsamtools Pileup (version 1.27.16). This program was used to determine the frequency of each of the four nucleotides present in mapped reads at every genomic position with read coverage. We limited our analysis to nucleotide positions with a minimum combined read depth of 500 unique reads across the 20 biological replicates to maximize sensitivity of detecting modified nucleotides. To prevent calling genomic variants and SNPs as modification-induced misincorporations, we did not analyze nucleotide positions containing variants discovered in the genomic DNA corresponding to the RNA-seq dataset, or SNPs annotated in dbSNP. Importantly, our analysis could only be performed on transcripts longer than the library insert size of ~250 bases¹². For this reason, analysis of cytosolic tRNAs, which are ~75 nt-long RNAs that contain known conserved m¹A residues, could not be performed. However, short RNAs generated from polycistronic transcripts, like mitochondrial tRNAs³⁶, were represented in the analyzed library. To identify sites of modification throughout the transcriptome, we initially filtered for all nucleotide positions that were covered by at least 500 mapped reads and contained a 1% misincorporation rate, and that were present in at least half of biological replicates (Supplementary Data ¹). To further obtain a high-confidence list of modification positions, we required that within the misincorporation profile at each initially identified position, a minimum of 5% of misincorporations were heterogeneous (i.e. transitions of the reference nucleotide to all three possible alternative nucleotides) in order to minimize detection of adenosine-to-inosine editing, and heterozygous alleles not reported as variants or SNPs. We chose this filter because hard-stop nucleotides have been shown to cause heterogeneous misincorporations, even when one type of misincorporation is predominant¹¹. This resulted in a high-confidence list of sites that were detected at known and novel modification positions (Supplementary Data ²).

m¹A-miCLIP

m¹A-miCLIP was performed as previously described^20,37, briefly described below along with any modifications: C57BL/6 mice (8 weeks) were sacrificed by CO₂ inhalation and cervical dislocation as approved by the Weill Cornell Medicine Institutional Animal Care and Use Committee (IACUC). Total RNA from HEK293T cells (n=2 biological replicates) or whole mouse brain (n=6 biological replicates) was extracted using TRIzol (ThermoFisher) and treated with RNase-free DNase I (Promega). Poly(A) RNA was isolated using one round of selection with oligo(dT)₂₅ magnetic beads (New England Biolabs). This resulted in approximately 10μg of poly(A) RNA for each replicate used in this study. Poly(A) RNA was subjected to fragmentation using RNA Fragmentation Reagents (ThermoFisher) for exactly 12min at 75°C. This fragmentation protocol is identical to the one in m¹A-seq, and has been reported not to facilitate substantial m¹A to m⁶A rearrangement.³ Fragmented RNA was then incubated with 10–15μg of m¹A antibody per replicate and the antibody-RNA complexes were processed for crosslinking, immunoprecipitation, RNA 3′ linker ligation, purification, and reverse transcription^20,37. Following reverse transcription of purified peptide-RNA complexes, first-strand cDNA was circularized using CircLigase II ssDNA Ligase (EpiBio) to preserve the 3′ end of the cDNA, and thus, sites of m¹A-induced terminations of reverse transcription. To generate priming sites for library amplification, the cDNA was cut in the middle of the cDNA primer sequence using a single-stranded DNA oligo complementary to this sequence and FastDigest BamHI (ThermoFisher)^20,37. This generated priming sites for the Illumina P5 and P3 primers on either side of the first-strand cDNA, eliminating the need for a second-strand synthesis step. For library amplification, Accuprime Supermix I (ThermoFisher) and Illumina P5 and P3 primers were used (see Supplementary Data ⁷). Amplified libraries were purified using AMPure XP magnetic beads (Beckman Coulter). Libraries were subjected to next-generation sequencing at the Epigenomics Core of Weill Cornell Medicine. Libraries were sequenced on an Illumina HiSeq 2500 and MiSeq instrument in single-end mode to generate 50-base reads.

RNA-seq

HEK293T cell total RNA was extracted with TRIzol (ThermoFisher), treated with RNase-free DNase I (Promega), and poly(A) RNA was isolated using one round of selection with oligo(dT)₂₅ magnetic beads (New England Biolabs). RNA was then subjected to fragmentation using RNA Fragmentation Reagents (ThermoFisher) for exactly 12min at 75°C. Fragmented RNA was then subjected to RNA 3′ linker ligation using T4 RNA Ligase I (New England Biolabs) and reverse transcription using a primer complementary to the linker sequence and SuperScript III (ThermoFisher) (see Supplementary Data ⁷). First-strand cDNA was gel-purified using denaturing PAGE, and then circularized using Circligase II ssDNA Ligase (EpiBio). Circularized cDNA was then cut and amplified exactly as described above for m¹A-miCLIP. Resulting RNA-seq libraries were subjected to next-generation sequencing at the Epigenomics Core of Weill Cornell Medicine. Libraries were sequenced on an Illumina HiSeq 2500 instrument in single-end mode to generate 50-base reads.

Read processing and alignment

After sequencing, reads from m¹A-miCLIP or RNA-seq libraries were trimmed of the 3′ linker sequence and barcoded reverse transcription primer sequences using Flexbar (version 2.5) (see Supplementary Data ⁷). To demultiplex reads belonging to individual biological replicates, the pyBarcodeFilter.py script of the pyCRAC suite (version 1.2.2) was used. The random portion of the reverse transcription barcode was then moved into the sequence header using a custom awk script (available upon request). PCR duplicates were collapsed using pyFastqDuplicateRemover.py of the pyCRAC suite. Finally, reads were aligned to hg19 for HEK293T cells or mm10 for mouse brain using Bowtie (version 1.1.2).

Generation of m¹A-miCLIP clusters

m¹A-miCLIP clusters of unique reads were generated using the CIMS software package for analysis of HITS-CLIP data^38,39. To generate clusters and determine the cluster score (maximum of stacked reads), the tag2profile.pl, tag2cluster.pl and extractPeak.pl scripts of the CIMS software package were used. A custom awk script was then used to filter for clusters of a minimum score (at least 20 stacked reads; script is available upon request).

Motif analyses

To search for a possible common sequence motif present in our HEK293T cell m¹A-miCLIP dataset, we focused on potential motifs present in m¹A-miCLIP clusters in the 5′UTR, the region of predominant m¹A-miCLIP cluster enrichment. The genomic sequences of these clusters were retrieved using bedtools and subjected to motif discovery using the MEME suite (version 4.11.4).

Metagene distribution analyses

To analyze the metagene distribution of m¹A-miCLIP clusters on mRNAs, MetaPlotR was used⁴⁰, with in-house modifications. The density of m¹A-miCLIP coverage was normalized to that of RNA-seq coverage to reveal any enrichments using a custom R script (available upon request). For the HEK293T cell metagene, the in-house HEK293T cell RNA-seq dataset described above was used for normalization. For the mouse brain metagene, a published whole-brain RNA-seq dataset was used⁴¹ (accession code: GSE52564). To plot the coverage of transcription-start sites by m¹A-miCLIP at higher resolution, the plotProfile tool of the Deeptools suite was used.

Examination of antibody crosslink sites around the transcription-start site

For analysis of antibody crosslinks at the transcription-start sites of mRNAs, we analyzed terminations of reverse transcription (i.e. 5′ ends of reads) around these sites. To do so, the number of terminations was measured around RefSeq-annotated transcription-start sites that had coverage in both m¹A-miCLIP and RNA-seq. Terminations were counted at positions ranging from the transcription-start site to position +4 relative to the transcription-start site. Then, transcription-start sites were filtered for those that contained a minimum coverage of five unique reads at the transcription-start site position in both m¹A-miCLIP and RNA-seq. This filtered set of transcription-start sites was then used to compare the distributions of read terminations in m¹A-miCLIP and RNA-seq. We focused on terminations rather than misincorporations in m¹A-miCLIP is because the misincorporation profile of m¹A is sequence dependent, with both upstream and downstream nucleotides contributing to misincorporation variability¹¹. Thus, we used the presence of terminations as a signature of antibody crosslinking events in our dataset. Additionally, while rare types of reverse transcriptases that read through m¹A have been described⁴², standard reverse transcriptases, like the SuperScript III used in m¹A-miCLIP, produce frequent terminations at m¹A residues^11,43.

Measurement of transcription-start sites enriched by m¹A-miCLIP

To determine the types of transcription-start sites overlapping m¹A-miCLIP clusters, we used a collection of transcription-start sites that included RefSeq transcription-start sites as well as recently-mapped transcription-start site regions containing the m⁶A_m mRNA extended cap²⁴. The frequencies of all transcription-start site types or those overlapping m¹A-miCLIP clusters were thus determined using this collective set.

Synthetic oligonucleotides used in this study

For biochemical analysis of various modifications present within the extended caps of mRNAs, synthetic oligonucleotides were generated as standards for mass spectrometry and/or thin layer chromatography (see below; see Supplementary Data ⁷). Oligonucleotides containing m⁷G-ppp-A_m, m⁷G-ppp-m⁶A, or m⁷G-ppp-m⁶A_m were synthesized chemically⁴⁴. Oligonucleotides containing m⁷G-ppp-m¹A or m⁷G-ppp-m¹A_m were synthesized enzymatically using an oligonucleotide initiating with ppp-m¹A (Trilink). This oligonucleotide was capped using ScriptCap Cap 1 Capping System (CellScript) to generate the m⁷G cap and, in the case of m⁷G-ppp-m¹A_m, 2′-O-methylation of m¹A.

Liquid chromatography and mass spectrometry (LC-MS)

Poly(A) RNA was prepared for mass spectrometry as follows. Total RNA from HEK293T cells was treated with TURBO DNase (ThermoFisher) according to the manufacturer’s instructions, followed by two rounds of poly(A) selection using oligo(dT) magnetic beads (NEB). Small RNAs shorter than 200 nt were then removed from the poly(A) RNA using the RNeasy kit (Qiagen). This size selection was performed to prevent detection of extended cap structures that are known to be present in certain small RNAs, like small nuclear RNAs (snRNAs). Approximately 5μg of DNase-treated, poly(A)-selected, and size-selected RNA was thus generated for each sample for mass spectrometry analysis. To release extended cap structures from the nucleotides comprising the internal portion of the RNA, RNA was digested with 2–4 units of Nuclease P1 (Sigma Aldrich) in a final buffer concentration of 30mM sodium acetate (pH 5.5) for 3h at 37°C. Following digestion, the nuclease was removed from the samples using molecular weight cutoff centrifugal filters (VWR). The digested and purified RNA was finally dried using an Eppendorf Vacufuge and reconstituted with 70% acetonitrile (LC-MS grade; Sigma Aldrich) to a final concentration of 0.5µg/µl. Two microliters of the resulting solution were subjected to MS analysis.

Samples were injected into an LC-MS/MS system comprising an Agilent 1260 HPLC and an Agilent 6460 triple quadrupole mass spectrometer equipped with a JetStream electrospray ionization source. Positive ion monitoring and multiple reaction monitoring was used for detection of extended caps. The caps were resolved on an aqueous normal phase column (ANP, Cogent Diamond Hydride, 4µm particle size, 150mm×2.1mm; Microsolv). To achieve chromatographic separation of the cap structures from mononucleotides, the following gradient was used. The aqueous mobile phase (Buffer A) was 50% isopropanol with 0.025% acetic acid, and the organic mobile phase (Buffer B) was 90% acetonitrile containing 5mM ammonium acetate. EDTA was added to the mobile phase in a final concentration of 6µM. The final gradient applied was 0–1.0min 99% B, 1.0–7.0min to 80% B, 7.0–18.0min to 50% B, 18.0–19.0min to 0% B, and 19.1–29.0min 99% B. The flow rate was 0.4mL/min during data acquisition and 0.6mL/min during column re-equilibration. Data were saved in centroid mode using MassHunter workstation acquisition software (Agilent). Data files were processed with MassHunter Qualitative Analysis Software (Agilent).

Exact operating source parameters for the LC-MS analysis are available upon request.

Biochemical examination of modifications at transcription-start sites

To identify the initiating nucleotide structure in mRNAs bound by the m¹A antibody, we utilized 2D-TLC analysis of mRNA extended caps²⁴. For analysis of cellular RNA, we used HEK293T cell poly(A) RNA or poly(A) RNA enriched using the m¹A antibody^3,4. Oligonucleotide standards, input (antibody-unbound) poly(A) RNA, and m¹A antibody-enriched poly(A) RNA were then subjected to 2D-TLC²⁴. To enhance resolution of m¹A and m¹A_m from other nucleotide species, the first dimension of 2D-TLC was resolved using 66% isobutyric acid and 1% NH₄OH in water, and the second dimension was resolved using 60% ammonium sulfate in 100mM sodium phosphate buffer of pH 6.8 (w/v) with a final concentration of 2% n-propanol. Both dimensions were resolved overnight.

Synthesis of N^1,6-methyladenosine

To synthesize N^1,6-methyladenosine, N⁶-methyladenosine (Selleckchem) was dissolved in dry DMF and followed with addition of iodomethane (Acros Organics; 10:1 molar ratio iodomethane:N⁶-methyladenosine). The mixture was stirred overnight at room temperature. The product was purified by flash chromatography on silica gel (EMD), eluting with methanol and dichloromethane (1:10 to 1:5; ACS or HPLC grade solvents). This resulted in a product yield of 46.3% N^1,6-methyladenosine. Product identity was confirmed by nuclear magnetic resonance (NMR) and high-resolution mass spectrometry (HR-MS).

NMR spectra were recorded using a 500-MHz Bruker DMX-500 instrument at room temperature, and chemical shifts were referenced to the residual solvent peak. Shifts were as follows: ¹H NMR (500MHz, DMSO-d₆) δ 8.11 (s, 1H), 8.04 (s, 1H), 5.76 (d, J=5.8Hz, 1H), 5.15 (s, 1H), 4.43 (t, J=5.4Hz, 1H), 4.15 (m, 2H), 3.92 (d, J=3.7Hz, 1H), 3.63 (dd, J=12.0, 3.9Hz, 1H), 3.56 (m, 2H), 3.50 (s, 3H), 3.45 (d, J=6.5Hz, 1H), 1.23 (s, 3H).

HR-MS data were recorded with Waters LCT-Premier XE at room temperature. For a predicted mass for N^1,6-methyladenosine, or C₁₂H₁₈N₅O₄⁺, of 296.1353, the mass found was 296.1361.

Analysis of Li et al. m¹A sites

The Li et al.⁸ m¹A sites for nuclear-encoded genes was obtained from supplementary Table ² of the published manuscript. We annotated each of the 474 transcriptomic sites with their corresponding genomic coordinates and nucleotide sequences using an annotation file generated from Refseq with MetaPlotR. With a custom R script, we then filtered likely erroneous sites as specified in the Results section. Briefly, sites corresponding to gene IDs missing in Refseq, or that mapped to non-adenosine nucleotides, or with duplicate genomic coordinates were all removed.

Characterization of antibody affinity for various substrates

The specificity or affinity of the m¹A antibody for various nucleosides, nucleotides, or cap structures was determined as follows. To determine the specificity of the m¹A antibody for various nucleosides, two approaches were performed. For testing the specificity of the antibody for various nucleosides in the context of m¹A-miCLIP, the antibody was crosslinked to total cellular RNA in the presence of various competitor nucleotides. Antibody binding, crosslinking, and detection of crosslinked antibody-RNA complexes was performed exactly as in miCLIP, except with the inclusion of the competitor nucleotide during the antibody-binding reaction.

For testing the specificity of the antibody for modified adenines, especially those resembling N¹-methylated adenine, a dot blot assay was performed wherein the competing molecule is added during antibody binding¹. Competition assays are usually used to measure binding, rather than spotting the nucleotides to the membrane, since the manner of interaction of each nucleotide to the membrane is not known and can affect antibody binding. For measuring the affinity, the IC₅₀ of the various nucleotides and cap dinucleotides was measured. In these experiments, a series of the dot blot assays were performed, where serial dilutions of each competitor molecule (ranging from 10µM to 1nM) were used in parallel during antibody binding reactions. The dot blots were performed as follows: 250ng of m¹A-containing synthetic oligonucleotide (see Supplementary Data ⁷) were spotted in triplicate on a BrightStar membrane (ThermoFisher), allowed to briefly air-dry, and auto-crosslinked twice in a Stratalinker 2400 (Stratagene). Each membrane was rinsed briefly in PBST, then blocked for 1h at room temperature in 5% milk in PBST. Each membrane was then placed into a pouch containing a 1:1000 dilution of the m¹A antibody in 0.5% milk in PBST, and an appropriate concentration of competitor molecule. The antibody binding proceeded for 2h at room temperature. Then, each membrane was washed three times in PBST (5min per wash), and then incubated in a dilution of 1:2500 of secondary antibody (anti-mouse, GE # NA931; anti-rabbit, GE NA934) in 0.5% milk in PBST for 1h at room temperature. Finally, the membrane was washed three times in PBST (5min per wash), and developed using ECL Prime (GE). Membranes corresponding to a dilution series of a specific competitor molecule were imaged together using a ChemiDoc Imager (Bio-Rad).

We thank S. Schwartz for helpful discussions, and members of the Jaffrey laboratory for helpful comments and suggestions. We thank J. Mauer and S. Zaccara for experimental contributions that were not included in this manuscript. This work was supported by NIH grants R01DA037755 (to S.R.J.), T32 HD060600 and UL1 TR000457 (to A.V.G.), T32 CA062948, KL2-TR-002385, NRSA 1F32GM120987 (to A.O.O.-G.), U01 HL121828 and P01 HD067244 (to M.S. and S.S.G.), and by a Postdoctoral Enrichment Program Award from the Burroughs Wellcome Fund (to A.O.O.-G.).