m⁶A level and isoform characterization sequencing (m⁶A-LAIC-seq) reveals the census and complexity of the m⁶A epitranscriptome

Abstract

N⁶-Methyladenosine (m⁶A) is a widespread, reversible chemical modification of RNA molecules, implicated in many aspects of RNA metabolism. Little quantitative information exists as to either how many transcript copies of particular genes are m⁶A modified (‘m⁶A levels’) or the relationship of m⁶A modification(s) to alternative RNA isoforms. To deconvolute the m⁶A epitranscriptome, we developed m⁶A-level and isoform-characterization sequencing (m⁶A-LAIC-seq). We found that cells exhibit a broad range of nonstoichiometric m⁶A levels with cell-type specificity. At the level of isoform characterization, we discovered widespread differences in the use of tandem alternative polyadenylation (APA) sites by methylated and nonmethylated transcript isoforms of individual genes. Strikingly, there is a strong bias for methylated transcripts to be coupled with proximal APA sites, resulting in shortened 3′ untranslated regions, while nonmethylated transcript isoforms tend to use distal APA sites. m⁶A-LAIC-seq yields a new perspective on transcriptome complexity and links APA usage to m⁶A modifications.

Introduction

m⁶A is the most abundant known internal chemical modification of mRNAs and long noncoding RNAs (lncRNAs)^{1, 2, 3, 4, 5, 6}. m⁶A has been implicated in all forms of RNA metabolism^{2, 3, 4, 7, 8, 9, 10, 11}. The ‘writing’ of m⁶A RNA modification is accomplished via the m⁶A methyltransferases METTL3 and METTL14 (refs. 7,12, 13,14,15). Furthermore, m⁶A modification is reversible or ‘erased’ by the FTO and ALKBH5 m⁶A demethylases and ‘read’ by the m⁶A-binding family of YTH-domain RNA-binding proteins (RBPs), which promote degradation and translational regulation of m⁶A-modified transcripts^{8, 9, 10, 11, 16, 17}. Recent findings that METTL3 is required for efficient embryonic-stem-cell differentiation highlight the functional importance of m⁶A modification^{18, 19}.

Although transcriptome-wide m⁶A location analysis by m⁶A-seq (or MeRIP-seq) has led to significant insights, methods to quantify the ratio of methylated to nonmethylated transcripts (defined here as m⁶A level) on a transcriptome-wide scale have been lacking^{3, 15, 20, 21}. The proportion of transcripts methylated per gene has been measured directly for only a handful of genes, and in those instances m⁶A levels were found to be nonstoichiometric (i.e., fewer than 100% of transcript copies were methylated)^{22, 23}. Genome-wide quantitative measurements of m⁶A levels are critical to understanding the proportion of m⁶A-modified versus unmodified transcripts for each gene to assess the magnitude of potential m⁶A regulatory impact. Here, we report a new method, m⁶A-LAIC-seq, which quantitatively deconvolutes methylated versus nonmethylated transcripts. Unlike m⁶A-seq, our protocol does not fragment the RNA before anti-m⁶A RNA immunoprecipitation (RIP); instead relying on sequencing intact full-length transcripts in both m⁶A-positive and m⁶A-negative fractions post-RIP. While allowing quantification of m⁶A levels, the isolation and sequencing of intact full-length RNAs also allowed us to examine differential isoform usage in methylated and nonmethylated transcripts of each gene.

Results

In developing m⁶A-LAIC-seq (Fig. 1a), we conducted initial experiments to establish the saturation curve, sensitivity, and specificity for anti-m⁶A RIP of full-length transcripts. The anti-m⁶A antibody became saturated at 500 ng of RNA per 1 μg of antibody (500 ng μg⁻¹) under our anti-m⁶A RIP conditions, which used in vitro–synthesized GAPDH transcripts containing 100% m⁶A-modified adenosines (Fig. 1b). We started m⁶A-LAIC-seq with 2 μg of 2× polyadenylated (poly(A)⁺) RNA, using an anti-m⁶A antibody at 500 ng μg⁻¹. We employed a vast excess of antibody, as a second round of anti-m⁶A RIP on the m⁶A-negative fraction recovered no further RNA (Supplementary Fig. 1a–c). We also found that anti-m⁶A RIP efficiency was independent of the number of modifications per transcript (Supplementary Fig. 1d,e). To examine the sensitivity of the anti-m⁶A antibody in vitro, we performed anti-m⁶A RIP experiments using in vitro–synthesized GAPDH with, on average, two random m⁶A modifications per transcript as well as unmodified GAPDH transcripts, both of which were radiolabeled with CTP32. We then mixed the unmodified and m⁶A-modified GAPDH transcripts at varying percentages. We detected as few as one m⁶A-modified transcript mixed with 1,000 unmodified transcripts, which suggested a very high sensitivity of the anti-m⁶A RIP (Fig. 1c).

Figure 1.

Next, we added various spike-in controls to our starting input poly(A)⁺ RNA to evaluate the specificity, sensitivity, and quantitation of the actual m⁶A-LAIC-seq protocol (Fig. 1a). As intended, our protocol sequenced intact full-length RNAs with no local enrichment of m⁶A sites (Fig. 1d). Furthermore, no reads were detected of the three nonmethylated RNA transcript spike ins in the eluate (m⁶A-positive fraction), but they were readily detected in the supernatant (m⁶A-negative fraction), confirming the high specificity. We quantified m⁶A levels per gene by the ratio of RNA abundances, (eluate)/(eluate + supernatant). RNA abundances were represented by the RNA-seq fragment counts normalized across input, eluate, and supernatant using the 96 synthetic ERCC control RNAs equally spiked in just before library construction (see Online Methods). To assess m⁶A-LAIC-seq’s ability to quantify m⁶A levels or stoichiometry, we mixed a series of nonmammalian m⁶A-modified RNAs generated in vitro, each with one or two modifications, with an unmodified counterpart at ratios ranging from 0–80%. Quantitative agreement between the known stoichiometry and its m⁶A-LAIC-seq quantification was excellent (Pearson r = 0.995; P = 3.7 × 10⁻⁵; Fig. 1e).

m⁶A-LAIC-seq reveals the transcriptome-wide patterns of m⁶A levels in H1-ESC

We recently characterized the m⁶A-seq profile of the hESC line H1-ESC¹⁸. Here we applied m⁶A-LAIC-seq to further characterize its m⁶A epitranscriptome. On a transcriptome-wide scale, we observed a strong concordance of m⁶A levels in two biological replicates (r = 0.98, P < 2.2 × 10⁻¹⁶; Fig. 2a), with m⁶A levels following an almost bimodal distribution whereby most genes exhibited less than 50% methylation levels (Fig. 2b). The analysis of the relationship between the number of m⁶A peaks and m⁶A levels revealed that the more peaks present, the higher the m⁶A levels (Supplementary Fig. 2a). Since m⁶A RIP efficiency is independent of the number of m⁶A sites and peaks (see above paragraph; Supplementary Fig. 1d,e), the increase in m⁶A levels with number of peaks is consistent with at least some m⁶A modifications occurring at different sites on separate transcripts, as opposed to all occurring in cis on the same transcripts. Notably, m⁶A-seq-derived peak intensities correlated only modestly with m⁶A levels (r = 0.6) (Supplementary Fig. 2b,c). Importantly, m⁶A levels as measured by qRT-PCR correlated strongly with m⁶A-LAIC-seq data (r = 0.96; P < 2.2 × 10⁻¹⁶, Supplementary Fig. 2d), crossvalidating the m⁶A-LAIC-seq approach. Figure 2c and Supplementary Figure 2e show examples of m⁶A-LAIC-seq profiles of genes with various m⁶A levels. Our results offer a quantitative road map as to which genes are most (SOX2 and LINC-ROR) or least (POU5F1 and TCF3) likely to be influenced by m⁶A-dependent regulatory networks (Supplementary Table 1).

Figure 2.

m⁶A-level-linked biological processes and protein domains

To understand potential links between biological pathways and m⁶A levels, we analyzed Gene Ontology (GO) terms. The 1,000 most highly methylated genes based on m⁶A levels were enriched in terms such as regulation of transcription (false discovery rate (FDR) = 2.9 × 10⁻²⁸) and regulation of RNA metabolic process (FDR = 7.5 × 10⁻²⁴) (Fig. 2d). This was consistent with the observation that high m⁶A levels were enriched in transcription factor (TF) domains, the strongest being TFs containing the repressive KRAB domain (FDR = 8.3 × 10⁻³⁵; n = 92) and the ZNF-C2H2-type domain (FDR = 4.0 × 10⁻³⁶; n = 143). Conversely, the 1,000 genes with the lowest methylation levels were enriched in processes such as translational elongation (FDR = 5.7 × 10⁻⁴⁶) and generation of precursors of metabolites and energy (FDR = 4.2 × 10⁻¹²). These results provided evidence that transcriptional regulators, especially transcriptional repressors, are likely the most susceptible to m⁶A regulation.

m⁶A levels are inversely correlated with steady-state RNA levels and half life

As m⁶A is implicated in promoting RNA degradation^{8, 18, 19}, we asked how m⁶A levels were related to RNA steady-state expression levels and half-life measurements^{8, 18, 19}. We found an inverse correlation between m⁶A levels and steady-state expression levels (r = −0.39; P < 2.2 × 10⁻¹⁶; Supplementary Fig. 2f). An inverse correlation was also identified between m⁶A levels in human embryonic stem cells (hESCs) and half-life measurements in induced pluripotent stem cells (iPSCs) (r = −0.14; P < 2.2 × 10⁻¹⁶; Supplementary Fig. 2g)²⁴. Furthermore, long intergenic noncoding RNAs (lincRNAs) had significantly higher m⁶A levels than mRNAs or pseudogenes (P = 3.0 × 10⁻¹⁶; two-sided Wilcoxon test; Supplementary Fig. 2h and Supplementary Table 2).

m⁶A-LAIC-seq reveals the extent of differential m⁶A levels between cell types

To examine variation of m⁶A levels among cell types, we performed m⁶A-LAIC-seq on the B-cell lymphoblastoid cell line GM12878, which also exhibited highly correlated m⁶A levels in replicates (Supplementary Fig. 2i). GM12878 cells also followed a non-Gaussian bimodal m⁶A-level distribution, but exhibited higher methylation levels than H1-ESC (P < 2.2 × 10⁻¹⁶; two-sided Wilcoxon test; Fig. 3a and Supplementary Fig. 2j). We found that 1,509 genes exhibited a reproducible variation in m⁶A levels of at least ±10% between cell types, but almost none showed variation larger than ±20%. Furthermore, such genes were found to largely exhibit a unidirectional increase in m⁶A levels in GM12878 as compared to H1-ESC (ratio, 1,465:44) (Fig. 3b). qRT-PCR analyses of genes with differential m⁶A levels between cell types showed highly concordant results compared to m⁶A-LAIC-seq (Fig. 3c). These results imply that cell-type-specific m⁶A levels of individual genes exhibit limited bidirectional variability, at least in the two cell types under the examined conditions. However, m⁶A levels appear to follow more global shifts, perhaps representing different m⁶A methyltransferase, demethylase, and/or reader expression and activity among conditions and cell types. Indeed, we found significantly higher mRNA levels of the YTH-domain-containing RBPs in H1-ESC (Fig. 3d), which would be expected to lead to higher rates of m⁶A-targeted RNA degradation and thus to a global reduction of H1-ESC m⁶A levels, compared to GM12878 as observed⁸.

Figure 3.

N⁶,2′-O-Dimethyladenosine is a low-abundance modification

In transcripts starting with an adenosine, methylation of the 2′-O and the N-6 positions can lead to a modification structurally related to m⁶A, N⁶,2′-O-dimethyladenosine (m⁶Am), that appears to be recognized by anti-m⁶A antibodies. m⁶Am is thought to occur exclusively on first nucleotides of mRNAs, whereas m⁶A has not been found at the first nucleotide position^{25, 26, 27, 28}. To evaluate the impact of m⁶Am on the quantification of m⁶A level using m⁶A-LAIC-seq, we analyzed mRNAs that were unlikely to carry m⁶Am given that they lacked m⁶A-seq peak signals within the first 200 nt of any annotated isoforms. We found very similar m⁶A-level distributions between all mRNAs and the list of mRNAs lacking m⁶Am in H1-ESC, suggesting that the presence of m⁶Am was unlikely to cause significant distortions in m⁶A-level quantification (Supplementary Fig. 3a). However, to our knowledge, there is no report of absolute quantitative measurement of m⁶Am level in RNA. Thus, we synthesized m⁶Am standards and used liquid chromatography–mass spectrometry spectrometry (LC-MS) (Supplementary Fig. 3b) to quantify m⁶Am levels in poly(A)⁺ RNA²⁹. We found that H1-ESC poly(A)⁺ RNAs contained ~3 m⁶Am nucleotides per 10⁵ nucleotides compared to ~100 m⁶A nucleotides per 10⁵ nucleotides, revealing 33 times more m⁶A than m⁶Am (Supplementary Fig. 3c). GM12878 also exhibited low levels of m⁶Am compared to m⁶A, suggesting that in both cell types m⁶Am is unlikely to skew m⁶A-LAIC-seq level results significantly (Supplementary Fig. 3c). Furthermore, multiple studies showed that fewer than 20% of methylated genes harbor m⁶A peaks in their 5′ UTRs^{18, 25}. Even with a liberal estimate that all of these m⁶A peaks are m⁶Am, the vast majority (>80%) of genes do not have any potential m⁶Am signal, which further suggests that the impact of m⁶Am on m⁶A-LAIC-seq is limited.

Internal splicing differs in m⁶A and non-m⁶A RNA fractions

The m⁶A methyltransferase METTL3 has been implicated in splicing of internal exons, although the true extent of involvement is unclear^{19, 21, 30, 31, 32}. Unlike traditional m⁶A-seq, m⁶A-LAIC-seq can directly compare alternatively spliced isoforms between methylated and nonmethylated RNA. By analyzing differential splicing events between these fractions using our rMATS software^{33, 34}, at our current sequencing depth we detected only a limited number of differential splicing events in internal exons, including alternative cassette exons (n = 49 in H1-ESC; n = 77 in GM12878), and retained introns (n = 21 in H1-ESC; n = 78 in GM12878). There was a significant trend for alternative cassette exons included in the m⁶A-positive fraction of H1-ESC to harbor m⁶A sites as detected by m⁶A-seq (P = 2.7 × 10⁻¹⁷; Fisher’s exact test; Supplementary Fig. 4a,b and Supplementary Table 3). We independently validated an example of alternative-cassette-exon inclusion in the m⁶A-positive fraction by RT-PCR (Supplementary Fig. 4c). We concluded that internal isoform-specific m⁶A targeting may allow for differential regulation of individual isoforms by m⁶A modification.

m⁶A modification is associated with proximal APA usage

A prevalent form of mRNA isoform regulation outside of internal splicing is 3′-UTR length variation via APA. Visual inspection of m⁶A-LAIC-seq data tracks revealed numerous examples of genes with significantly differential 3′-UTR signals in m⁶A-positive versus m⁶A-negative fractions. A systematic evaluation of APA based on m⁶A-LAIC-seq (see Online Methods) revealed many genes (H1-ESC, n= 2,512; and GM12878, n = 2,260) in which methylated transcripts significantly used the proximal rather than the distal APA site and thus harbored short 3′ UTRs, compared to their nonmethylated transcripts (FDR < 0.01) (Fig. 4a, Supplementary Fig. 5a, and Supplementary Table 4). By contrast, we found a much smaller number of genes (H1-ESC, n = 410; and GM12878, n = 413) in which the reverse was true (Fig. 4a, Supplementary Fig. 5a, and Supplementary Table 4). Furthermore, the m⁶A-seq peak signals surrounding proximal poly(A) sites suggest that for genes with preferential m⁶A modification of the shorter 3′-UTR isoform, the m⁶A sites are located upstream of the proximal poly(A)⁺ sites and are thus shared by both the shorter and longer 3′-UTR isoforms (Fig. 4b). Figure 4c and Supplementary Figure 5b show examples of m⁶A-LAIC-seq results for genes with or without differential APA usage. We confirmed our findings in these examples, calculating the change in the percentage of distal polyadenylation (dPAS) usage index (DPDUI) by qRT-PCR (Fig. 4c and Supplementary Fig. 5b)³⁵. Of note, genes in which the proximal APA usage in the m⁶A-positive fraction is greater than that in the m⁶A-negative fraction, designated P/D, m⁶A⁺ > m⁶A⁻, also exhibited higher m⁶A levels than genes without differential usage (P = 1.8 × 10⁻¹⁵², two-sided Wilcoxon test; Supplementary Fig. 5c,d). A recent study reported that genes with longer last exons have a higher density of m⁶A peaks³⁶. Consistent with this observation, we found that m⁶A levels of genes are positively correlated with maximum 3′-UTR lengths as measured in input steady-state RNA-seq data (Supplementary Fig. 5e). Detailed analyses showed that genes with higher proximal APA usage in the m⁶A-positive fraction (P/D; m⁶A⁺ > m⁶A⁻) had significantly longer 3′ UTRs (Supplementary Fig. 5f) and lower ratios of proximal versus distal signals (Supplementary Fig. 5g,h) in the input RNA-seq data as compared to genes with higher distal APA usage in the m⁶A-positive fraction (P/D; m⁶A⁺ < m⁶A⁻) or no change in APA usage between fractions (P/D; m⁶A⁺ = m⁶A⁻). Given that m⁶A modification is strongly coupled with RNA degradation^{3, 15, 20, 21}, our data suggest one possible model in which preferential m⁶A methylation and degradation of shorter 3′-UTR isoforms using proximal poly(A) sites simultaneously result in their increase in the m⁶A-positive fraction and decrease in the total steady-state population of RNA.

Figure 4.

One possible interpretation of our findings is that m⁶A and/or METTL3 may directly dictate APA choice. To test this possibility, we analyzed 3′-UTR read distribution of published RNA-seq or 3p-seq data in wild-type versus knockout or knockdown of m⁶A writers (METTL3, METTL14, and WTAP) from multiple human and mouse cell lines^{14, 15, 18, 19}. We observed no or limited numbers of significant APA changes in all data sets, with a trend for increased signals of proximal poly(A) sites upon depletion of m⁶A writers (Supplementary Table 5). Thus, compared to thousands of differential APA usage events in methylated versus nonmethylated transcripts of the same genes, the direct effect of m⁶A on APA was considerably smaller, suggesting that it is unlikely to explain the widespread association between m⁶A and proximal APA usage revealed by m⁶A-LAIC-seq.

N⁶,2-O-Dimethyladenosine (m⁶Am) does not correlate with APA usage

To remove the potential confounding effects of m⁶Am, we repeated the APA analyses of m⁶A-LAIC-seq data in genes without any peak within the first 200 nt of transcripts as detected by m⁶A-seq and obtained similar results (Supplementary Fig. 6a). We also examined APA usage upon stratification of genes based on whether their transcription start sites (TSSs) are adenosine in H1-ESC and GM12878. The TSSs were annotated based on Ensembl annotation or cap analysis gene expression (CAGE) data. We found similarly strong association of m⁶A with proximal APA usage in genes with or without A at TSSs based on TSS annotation (Supplementary Fig. 6b). These results suggest that the association between m⁶A and proximal APA usage revealed by m⁶A-LAIC-seq is not dictated by the potential presence of m⁶Am at the starting nucleotide position.

APA of m⁶A versus non-m⁶A fractions affects regulatory elements

The 3′ UTRs of transcripts serve as post-transcriptional nexuses of regulation by miRNAs as well as RBPs³⁷. To assess the consequence of differential 3′-UTR usage patterns between modified and unmodified m⁶A isoforms, we analyzed potential miRNA–mRNA and RBP–mRNA interactions. To do so, we analyzed predicted miRNA target sites as well as RBP-binding motifs in the extended segment of 3′ UTR between the proximal and distal poly(A)⁺ sites in genes with differential APA usage, while using genes with m⁶A modifications and tandem APA sites but no difference in APA usage between methylated and nonmethylated fractions as controls. Genes with higher proximal APA usage in the m⁶A-positive fraction as compared to the m⁶A-negative fraction (P/D; m⁶A⁺ > m⁶A⁻) show a higher degree of transcript regulation in the extended segment by miRNAs and RBPs, as compared to genes with higher distal APA usage in the m⁶A-positive fraction (P/D; m⁶A⁺ < m⁶A⁻), or no change in APA usage between fractions (P/D; m⁶A⁺ = m⁶A⁻) (Fig. 5a,b). By contrast, genes with higher distal APA usage in the m⁶A-positive fraction (P/D; m⁶A⁺ < m⁶A⁻) had a higher density of m⁶A RRACU motif in the extended region (Fig. 5c).

Figure 5.

Discussion

We envision that m⁶A-LAIC-seq complements standard m⁶A-seq identification of methylation sites and adds to our armamentarium for the study of m⁶A biology. As demonstrated in this work, the combined use of these two technologies provides new insights into the dynamic range and isoform complexity of the m⁶A epitranscriptome. Factors that determine m⁶A levels remain to be elucidated, but they may be related to a number of interacting and stochastic processes ranging from the concentration of m⁶A readers, erasers, and writers to the in vivo folding and accessibility of RNA. Future application of m⁶A-LAIC-seq to a larger number of tissue types and cellular states will help further elucidate tissue or cell-state specificity of m⁶A levels and identify the regulatory logic controlling m⁶A levels. Theoretically, RNA secondary structures might affect m⁶A-LAIC-seq results if the structures in full-length RNAs shield the m⁶A sites from being recognized by the antibody. However, this scenario is unlikely, given that previous studies have shown that m⁶A modifications occur in unstructured regions^{31, 32}.

A unique application of m⁶A-LAIC-seq is to deconvolute and contrast the isoform complexity of m⁶A-methylated versus nonmethylated RNA fractions in a given cell type. We found that thousands of genes with m⁶A-modified transcripts exhibited a tendency to utilize proximal APA sites, while for hundreds of genes the converse was true. These data suggest that many genes generate at least two distinct pools of transcripts, encoding the same protein in the case of mRNAs, with differing metabolic fates based on m⁶A coupled with differential binding of miRNAs and RBPs in their UTR regions. Indeed, our findings are consistent with a potential model that the m⁶A fraction of genes that often harbor shorter 3′ UTRs turns over more rapidly, leading to the observation in the steady-state RNA population that such genes have longer 3′ UTRs. Collectively, these observations may reflect the need to fine-tune the amount of RNA or protein under different circumstances using distinct post-transcriptional mechanisms³⁸. For genes with tandem APA sites in the 3′ UTR, the shorter transcript isoforms have less regulatory sites at the 3′ UTR than the longer isoforms. However, they preferentially acquire m⁶A modifications and potentially turn on m⁶A-mediated regulatory pathways. Furthermore, our results suggest that rather than m⁶A METTL3 dictating APA choice, perhaps the opposite is true. APA factors may help determine m⁶A site distribution along transcripts, consistent with the predominance of m⁶A sites in 3′ UTR or last exon, along with data linking APA factors to internal exon splicing³⁹. Further studies are needed to investigate the role of the APA machinery on the installment of m⁶A along transcripts⁴⁰.

m⁶A-LAIC-seq provides a new vantage point from which to dissect and classify the human transcriptome. For example, using m⁶A-LAIC-seq as a road map, it should now be possible to study the effects of differentially depleting m⁶A-modified versus m⁶A-unmodified transcripts for the same gene by designing UTR-segment-specific shRNA or siRNAs. Furthermore, strategies employing RNA-targeting Cas9 systems will allow the rationale design of experiments based on LAIC-seq maps to track and manipulate methylated versus nonmethylated subfractions of a transcript based on their differential UTRs⁴¹. In addition, our work sets the foundation for subphenotyping methylated versus nonmethylated transcripts with respect to parameters such as their three-dimensional structures, half lives, and the binding specificity of RBPs on genome-wide scales.

In summary, m⁶A-LAIC-seq has revealed a new design principle embedded in the human transcriptome based on m⁶A modifications and 3′-UTR length combinatorics. Given the increasing recognition of other RNA modifications, such as 5mC, m1A, and pseudouridylation (ψψ)⁴², our results may point to an even greater heterogeneity and isoform complexity of the transcriptome as part of an RNA epitranscriptome code.

Methods

Antibodies

Anti-m⁶A antibody: the anti-m⁶A antibody was obtained from Synaptic Systems (cat# 202 003) m⁶A rabbit polyclonal, affinity purified. Normal rabbit serum R9133-5ML, Sigma (cat# R9133). Donkey secondary anti-rabbit antibody was obtained from Amersham (cat# NA934) (antibody dilution factor, 500 ng μg⁻¹).

Cell line

The GM12878 cell line was obtained by the Coriell Institute for Medical Research (cat# GM12878). The species of origin was confirmed by nucleoside phosphorylase, glucose-6-phosphate dehydrogenase, and lactate dehydrogenase. The cell lines were genotyped based on microsatellite polymorphism assay used by Coriell Institute. The GM12878 cell line tested negative for mycoplasma contamination using the polymerase chain reaction (PCR) method developed by Applied Biosystems and the MycoSEQ Mycoplasma Detection System (4460625, Life Technologies).

The H1 embryonic stem cell line was obtained from WiCell (name: WA1, alias: H1). The cell tested negative for mycoplasma infection based on PCR method.

Synthesis of m⁶Am

The synthesis of N⁶, 2′-O-dimethyladenosine was analogous to a published procedure for the synthesis of 2′-O-methyladenosine, except for the use of N⁶-methyladenosine (Berry & Associates) as starting material²⁹. Purification was performed on an Agilent 1200 series HPLC equipped with a diode array detector (DAD), fraction collector, and a Synergy Fusion RP column (4 μm particle size, 80 Å pore size, 250 mm length, and 4.6 mm inner diameter) from Phenomenex at 35 °C. The solvents consisted of 5 mM ammonium acetate buffer adjusted to pH 5.5 with acetic acid (solvent A) and acetonitrile (solvent B). The elution at a flow rate of 1 mL/min started with 90% solvent A followed by a linear gradient to 21% solvent B at 15 min. Solvent B was increased further to 50% over 5 min, followed by a decrease to 10% over 3 min. The desired fractions were dried under vacuum.

Analysis of m⁶A and m⁶Am by liquid chromatography-coupled mass spectrometric analysis (LC-MS)

Prior to mass spectrometric analysis, all RNA samples were hydrolyzed enzymatically to ribonucleosides as described previously⁴⁴. Briefly, the digestion was carried out in 0.1 mM Tris buffer (pH 8), 5 mM MgCl2, 0.0375 U μL⁻¹ benzonase, 0.17 U μL⁻¹ alkaline phosphatase, 1 U mL⁻¹ phosphodiesterase, 3 mM desferroxamine (antioxidant), 0.3 mM butylated hydroxytoluene (antioxidant), 0.05 μM [¹⁵N5]-2′-deoxyadenosine (internal standard), and 5 μM 2′-deoxyinosine (internal standard) at 37 °C for 2 h. Enzymes were then removed using a YM-10 centrifugal spin column (Millipore).

Quantitative analyses of m⁶A and m⁶Am were achieved using an Agilent 1200 HPLC coupled to an Agilent 6430 triple quadrupole mass spectrometer in positive-ion mode using dynamic multiple reaction monitoring (MRM). The ribonucleosides in the hydrolyzed RNA samples were resolved on a Phenomenex C18 HPLC column (1.7 μm particle size, 100 Å pore size, 2.1 × 150 mm; 25 °C) at 330 μL/min using a solvent system consisting of 10 mM ammonium acetate in H2O (A) and acetonitrile (B). The elution profile was 0% B for 3 min, 0–7% B over 20 min, then to 7–40% B over 4 min, followed by a column washing at 80% B and column equilibration. The operating parameters for the mass spectrometer were as follows: gas temperature 350 °C; gas flow 10 l/min; nebulizer 50 psi and capillary voltage 3500 V, fragmentor voltage 100 V, and collision energy 15 V. The quantification of a ribonucleoside can be achieved using mass-to-charge ratio (m/z) of the parent ribonucleoside ion and m/z of the deglycosylated product ion. The nucleosides were quantified based on the transition of the parent ribonucleoside to the deglycosylated base ion: m/z, 282.1/150.1 for m⁶A and m/z, 296.1/150.1 for m⁶Am. Absolute quantities of each m⁶A and m⁶Am were determined respectively using an external calibration curve prepared with synthetic standards for each ribonucleoside.

m⁶A-level dot blots

Amersham Hybond-XL (cat# RPN303s) membrane was rehydrated in H2O for 3 min. The membrane was then ‘sandwiched’ in Bio-Dot Microfiltration Apparatus (Bio-Rad, cat# 170-6545). Each well was then filled with H2O and flushed by gentle suction vacuum until it appeared dry. 5 μl of H2O alone was then applied to the membrane in each well followed by addition of indicated amount of RNA and this was allowed to bind to the membrane by gravity. The apparatus was disassembled and the membrane was crosslinked in a UV STRATALINKER 1800 using the automatic function, and then the membrane was placed back into the apparatus. The membrane was then blocked 10 min using sterile RNase DNase free TBST + 5% milk. The m⁶A primary antibody (Anti-m⁶A, Synaptic Systems, cat# 202 003) was then added at a concentration of 1:500 at room temperature for 1 h in TBST + 5% milk. The membrane was then washed four times in PBST. The membrane was then incubated with the secondary anti-rabbit antibody (1:5,000 dilution) for 30 min in TBST + 5% milk. The membrane was washed 4 times for 5 min in TBST and exposed on an autoradiographic film using Pierce ECL Western Blotting Substrate.

cDNA synthesis for qRT-PCR analysis

First, a mix was made of 100 ng of RNA in 5 μl volume, 2 μl of random hexamers (Roche), 1 μl of dNTP Mix (10 mM each) and 5 μl of ultrapure H2O was first generated, heated at 65 °C for 5 min and immediately put on ice. 4 μl of 5× first-strand buffer was added along with 1 μl of 0.1 M DTT, 1 μl RNase inhibitor and 1 μl of Superscript III reverse transcriptase (Invitrogen). The 20 μl reverse transcription reaction was then incubated 5 min at room temperature, then 60 min at 50 °C then 15 min at 70 °C. The freshly synthesized cDNA was treated with 1 μl of RNase H at 37 °C for 20 min.

SYBR Green qRT-PCR

To calculate m⁶A levels by qPCR for Figures 2d and 3c and Supplementary Figures 2d and 3e, SYBR Green quantitative real-time PCR assays were performed utilizing a spike-in synthetic control called Drosophila E (see below) for normalization. Each PCR reaction was done in a 10 μl volume made of 5 μl of master mix (SYBR GreenER qPCR SuperMix for iCycler, Invitrogen), 2.5 μl of primer mix at 1.2 μM (each) and 2.5 μl of cDNA template at 20 ng μl⁻¹. The PCR was carried out using a standard protocol with melting curve. The amount of target was calculated using the formula: amount of target = 2^−ΔΔCT (ref. 45).

Calculation of ΔPDUI

All ΔPDUI calculations were based on the method described in Masamha et al.³⁵, adopted to compare UTR length in positive and negative fractions of the m⁶A IP of m⁶A-LAIC-seq. The following was calculated using primer sets designed in the common part of the isoforms of a transcript (called ‘common primer set’) as well as in the distal part of long isoform of the same transcripts (called ‘distal primer set’): ΔCT (common and distal) was normalized to an artificial spike-in transcript DrosoE (described in spike-ins controls for m⁶A-LAIC-seq). ΔΔCT = ΔCT distal − ΔCT common. ΔΔΔCT = ΔΔCT negative fraction m⁶A IP − ΔΔCT positive-fraction m⁶A IP. The increase or decrease in ΔΔΔCT was given by ±2^Δ−ΔCT (ref. 45).

RNA extraction, DNase I treatment, and poly(A)⁺ selection

Total RNA was isolated from cells according to manufacturer’s instructions using TRIzol LS reagent (Ambion). Total RNA was treated using DNase I (Promega) for 20 min at 37 °C. The treated RNA was then acid-phenol-chloroform extracted and chloroform extracted. The RNA was precipitated using 300 mM final concentration of NaCl2 spiked with 1 μl of 50 mg/ml of Ultra Pure Glycogen (Promega) and 2.5 volume of 100% ethanol at −20 °C either for 2 h or overnight. The precipitated RNA was then centrifuged using a refrigerated tabletop at maximum speed (>13,000g) at 4 °C for 20 min. The precipitated RNA was then washed with 70 °C ethanol and centrifuged at maximum speed for an additional 10 min. The final pellet was then resuspended in ultrapure H2O. Poly(A)⁺ RNA selection was performed twice using Dynabeads mRNA Purification Kit (Invitrogen cat. # 610.06) according to the manufacturer’s protocol. The second poly(A)⁺ RNA selection was performed using the eluate of the first poly(A)⁺ RNA selection as starting material according to the manufacture’s instruction. The obtained poly(A)⁺ RNA was evaluated both by NanoDrop and Bioanalyzer. For all RNA samples, the concentration, purity, and integrity of the RNA were verified using a NanoDrop and Bioanalyzer.

H1-ESC cell and GM12878 cell culture

H1 (WA01) cells were cultured in feeder-free condition using mTESR1 media (Stem Cell Technologies cat. # 05850) on 6-well plates coated with Matrigel (BD Biosciences, cat. # 354603), as described⁴⁶. GM12878 cells were cultured on complete RPMI media (Invitrogen) supplemented with 15% FBS and 10% Pen–Strep antibiotic.

m⁶A level and isoform-characterization sequencing (m⁶A-LAIC-seq)

Of note for each biological replicate for m⁶A-LAIC-seq, we started with 150 μg of total RNA yielding approximately 3.8 μg of double poly(A)-selected RNA which was resuspended in a final volume of 50 μl using UltraPure H2O (Life Technologies cat# 10977-015). The 2× poly(A)⁺ RNA was then heated at 65 °C for 5 min and immediately put on ice. 50 μl of m⁶A–DynaBeads (the m⁶A antibody, Synaptic Systems cat# 202 003, was coupled to Dynabeads using the Life Technologies coupling kit cat# 14311D) were equilibrated by washing twice for 5 min in 500 μl of m⁶A–Binding Buffer (50 mM Tris-HCl, 150 mM NaCl2, 1% NP-40, 0.05% EDTA). The RNA and the spike-ins controls were then added to the equilibrated m⁶A–DynaBeads. The RNA was allowed to bind to the m⁶A–Dynabeads (in 500 μl volume of m⁶A–Dynabeads and m⁶A–Binding Buffer at room temperature while rotating (tail over head) at 7 rotations per min for 1 h). The tubes containing the samples were placed on a magnet, allowing the beads complexes to cluster for 1 min or until the solution become clear. The liquid phase was carefully collected and placed on ice as this 500 μl fraction represented the ‘supernatant’ of the m⁶A IP. Following the collection of the supernatant fraction, series of washes were performed using various buffers (see as follows). For all wash steps, with the exception of the elution step, the beads were washed for 3 min then placed on a magnet and the wash buffers were discarded. Following the supernatant collection, Wash step 1: the remaining fractions bound to the beads were washed twice in 500 μl of m⁶A–Binding Buffer (50 mM Tris-HCl, 150 mM NaCl2, 1% NP-40, 0.05% EDTA). Wash step 2: the RNA–beads complexes were washed once in 500 μl of low-salt buffer (0.25× SSPE, 0.001 M EDTA, 0.05% Tween-20, 37.5 mM NaCl). Wash step 3: the RNA–beads complexes were washed once in 500 μl of high-salt buffer (0.25× SSPE, 0.001 M EDTA, 0.05% Tween-20, 137.5 mM NaCl). Wash step 4: the RNA–beads complexes were washed twice in 500 μl of TET (T.E. + 0.05% Tween-20). Elution step: The m⁶A RNA was eluted from the beads by repeating the following four times: 125 μl of elution buffer (0.02 M DTT, 0.150 M NaCl, pH 7.5 0.05 M Tris-HCl, 0.001 M EDTA, 0.10% SDS) was added to the beads and incubated at 42 °C for 5 min. At the end of the 5 min the beads were gently placed on the magnet. The liquid phase was collected and transferred to a fresh tube as this will represent the eluate fraction containing the m⁶A ‘enriched RNA’. An additional 125 μl of elution buffer was then added to the beads and the process was repeated. The liquid phase obtained at each step was added to the ‘fresh tube’ containing the 125 μl of eluate from the previous step so the total final eluate volume was 500 μl.

All RNA fractions were extracted as follow. 500 μl of acid phenol-chloroform (acid-phenol:chloroform), pH 4.5 (with IAA, 125:24:1; Ambion) was added to the 500 μl sample. The sample was centrifuged at 4 °C at 10,000g for 7.5 min. The upper phase was carefully collected, ensuring that it did not touch the interphase, and it was transferred to a clean 1.5 ml tube. 500 ml of chloroform was added to the fresh tube, mixed by gentle manual shaking, and centrifuged at 4 °C at 10,000g for 7.5 min. The upper phase was transferred to a fresh 1.5 ml tube and NaCl2 ethanol precipitated overnight at −20 °C in the presence of 1 μl of (20 mg/ml) Ultra Pure Glycogen. The following day the sample was centrifuged at 4 °C for 20 min at 16,000g. The pellet was then washed in 70% ethanol and centrifuged an additional 10 min at 4 °C at 16,000g. The pellet was then left to dry at room temperature for 10 min before it was resuspended in the desired volume of Ultra-Pure H20 (Invitrogen cat# 10977-015). Note that only one round of anti-m⁶A RNA immunoprecipitation for m⁶A-LAIC-seq was performed given our data that under the conditions we used there was essentially 100% efficiency in pulling down m⁶A-positive transcripts (Fig. 1 and Supplementary Fig. 1).

Spike-in controls for m⁶A-LAIC-seq

For each sample after double-poly(A)-RNA selection, but before anti-m⁶A RIP, in vitro–transcribed transcripts with and without m⁶A modification were mixed into the 2× poly(A)⁺ RNA as spike-in controls at the indicated percentage of m⁶A-modified to m⁶A-unmodified transcript, the indicated number of copies, and the indicated percentage of m⁶A in the modified transcript. For all samples these spike ins included: 80% mCHERRY (10⁸ copies, 0.2% m⁶A), 60% XEF (10⁸ copies, 0.2% m⁶A), 20% eGFP (10⁸ copies, 0.2% m⁶A), luciferase unmodified (10⁸ copies, 0% m⁶A), and Kanamycin unmodified (10⁸ copies, 0% m⁶A). An additional unmodified transcript originating from Drosophila called transcript DrosC (EST:RH10804; 10⁸ copies, 0% m⁶A) was also added before anti-m⁶A RIP. Post-m⁶A RIP 5 μl of 1:100 dilution of the universal ERCC spike-in control A (Invitrogen) was added to each fraction (input fraction, m⁶A⁻ fraction and m⁶A⁺ fraction) as well as transcript originating from Drosophila called transcript-unmodified DrosE (EST:IP03042; 10⁸ copies, 0% m⁶A).

Library construction for RNA-seq

100 ng of RNA was used for library construction for m⁶A-LAIC-seq input, supernatant (m⁶A-negative fractions), and eluate (m⁶A-positive fractions) using the TrueSeq Stranded mRNA Sample Preparation Guide, entering the protocol by adding the fragment, prime, finish mix, skipping the elution step and proceeding immediately to the synthesis of the first strand cDNA. From that point on, the exact steps of the Illumina TruSeq Stranded mRNA Sample Preparation Guide were followed to the end.

m⁶A-antibody-binding saturation and sensitivity titration

We generated an m⁶A antibody titration curve to identify the point of saturation of the anti-m⁶A antibody in the context of performing m⁶A RIPs (Fig. 1b). To do so, we used an in vitro–generated transcript from a plasmid containing cDNA of GAPDH transcript (1,008 bps; 233 ATPs). The plasmid was first linearized by restriction digest using SalI just downstream of the GAPDH cDNA cloning site. The linearized plasmid was gel purified and in vitro T7-mediated transcription was performed using the Ambion MEGAscript Kit (AM1334) as described in the user manual using 100% N⁶-methyladenosine-5′-triphosphate (cat# N1013) nucleotide during synthesis adding α-32P-CTP as tracer (PerkinElmer cat# BLU008H250UC). The anti-m⁶A RIP was performed as described in the m⁶A-seq section, using anti-m⁶A antibody (Synaptic system # 202 003) or normal rabbit serum on respectively 10 μg, 2 μg, and 400 ng of in vitro–transcribed transcript (antibody dilution factor, 500 ng μg⁻¹). m⁶A dot blots were performed on the m⁶A-positive (eluate) and m⁶A-negative (supernatant) fractions resulting from the RIP as shown in Figure 1b. To test for sensitivity of anti-m⁶A RIP in vitro on full-length transcripts, we synthesized α-32P-CTP GAPDH transcripts with on average 2 m6A modifications per transcript, as well as unmodified α-32P-CTP GAPDH transcripts. Subsequently, various ratios (shown as percentages) of unmodified versus m⁶A-modified GAPDH transcripts were mixed (e.g., 0.1% indicates 1,000 unmodified to 1 modified transcripts). The various percentage ratios of mixed transcripts were then subjected to anti-m⁶A RNA immunoprecipitation (m⁶A-RIP) using anti-m⁶A antibody or normal rabbit serum (NRS-RIP). The indicated fractions were subjected to gel electrophoresis and radiographically visualized (Fig. 1c).

RNA sequencing

All libraries were prepared as previously described for RNA-seq. Each individual library fragment size was verified on Agilent Bioanalyzer 2100 with high-sensitivity chip. Final quantification was done by qPCR on PerkinElmer 2500Fast with Kapa library quantification kit (# KK4824). Libraries were pooled at equimolar concentrations according to the manufacturer guidelines (TruSeq Stranded mRNA Sample Preparation Guide, September, 2012). After clustering on Illumina cBot, samples were run on Illumina HiSeq 2000.

Anti-m⁶A RNA immunoprecipitation is efficient and independent of the number of m⁶A modifications

Total RNA was first isolated using TRIzol reagent. Following RNA isolation the total RNA was further treated using RQ1 RNase free DNase (PROMEGA # M610A). The RNA was then poly(A)-selected twice using Ambion Dynabeads mRNA Purification Kit. The quality and the integrity of the RNA was checked by BioAnalyzer. The 2× poly(A)⁺ RNA was split into 5 equal quantities. Each tube was spiked with in vitro–generated GFP transcripts containing one of five variant percentiles of incorporated m⁶A nucleotides (100% m⁶A, 20% m⁶A, 2% m⁶A, 0.2% m⁶A, or 0% m⁶A). Note that every percentage variant of m⁶A spike in underwent one round of m⁶A RIP and eluate was obtained to ensure the purity of each transcript after in vitro synthesis before being used as a spike in. Each of the five poly(A)⁺ RNA pools with unique spike in were subject to m⁶A RIP and fractions were subjected to a qRT-PCR for GFP to calculate the percentage of recovery of GFP post-m⁶A-RIP. To check for overall efficiency of anti-m⁶A RIP a similar experiment was performed with total RNA and a single spike in of 2% m⁶A GFP and subjected to anti-m⁶A RIP followed by qRT-PCR of GFP and anti-m⁶A dot blot of input, m⁶A positive (eluate), and m⁶A negative (supernatant). The anti-m⁶A RIP was performed a second time on the supernatant from the first round of anti-m⁶A RIP and each subsequent fraction was also subject to qRT-PCR and m⁶A dot blotting.

Read-mapping and gene-expression analyses

We mapped strand-specific paired-end RNA-seq reads first to Ensembl genes (release 74) and then to hg19 human genome incorporated with all spike-in sequences using tophat (v 1.4.1)⁴⁷. Only the properly paired and uniquely mapped reads were used in future analyses. FPKMs of input fractions were calculated using Cufflinks (v 2.0.2)⁴⁸. Differentially expressed genes (DEGs) between H1-ESC input and GM12878 input were determined by Cuffdiff (v 2.0.2)⁴⁸.

Read coverage distribution comparison

To compare the read coverage distributions around m⁶A peaks during the three fractions of m⁶A-LAIC-seq and m⁶A-seq in H1-ESC, for each peak, we calculated the RNA-seq read coverage of the upstream and downstream 1,500 bp regions on the longest isoform of the Ensembl gene. The peaks with upstream or downstream transcript regions less than 1,500 bp were removed from the analyses. For each 3,000 bp region, the read coverage of each site was normalized by the average of the whole region. To plot Figure 1d, the normalized read coverage of all the 3,000 bp regions around m⁶A peaks were averaged at each site.

m⁶A-level calculation

The ratio of RNA abundance (copies) between eluate and supernatant was calculated first in order to calculate m⁶A level for each gene or spike-in RNA before m⁶A-RIP. Since equal amount of ERCC RNA Spike-In Mix were added to eluate and supernatant fractions right after m⁶A RIP, the abundance of each gene or spike-in RNA in eluate or supernatant was represented by the RNA-seq fragment (read pair) counts normalized by the fragment counts of the ERCC RNAs. So, for each gene or spike-in RNA before m⁶A RIP, the ratio of RNA abundance between eluate and supernatant could be calculated using the ratio of its RNA-seq fragment counts and the ratio of fragment counts of ERCC RNAs. Because there were 92 RNAs with a variety of concentrations in ERCC mixture, the ratio of ERCC RNA fragment counts between eluate and supernatant in each replicate could be accurately calculated using these RNAs together. As shown in Supplementary Figure 7, we used the log2-transformed counts of ERCC RNAs to fit a linear regression model (eluate counts as a function of supernatant counts) with coefficient of 1, and the log2 ratio between ERCC eluate counts and supernatant counts was indicated by the residual of the regression formula. To avoid using unreliable counts, only the ERCC RNAs with at least 100 counts were used to fit the regression model.

The m⁶A levels of a specific Ensembl-annotated gene or spike-in RNA before m⁶A-RIP were calculated using eluate counts (E), supernatant counts (S), and the residual of ERCC regression (R) according to the following formula:

$${\mathrm{m}}^6\mathrm{A}\text{level}=\frac{E}{E+S\times 2^R}$$

m⁶A-level analyses

For all the analyses of m⁶A levels, we required eluate counts ≥50 or supernatant counts ≥50 in both replicates to obtain reliable m⁶A levels. To compare the m⁶A levels with m⁶A peaks, we used our previously published¹⁸ m⁶A peaks of H1 ESC cell line. The peak intensities of genes were represented by the maximum peak intensities of the peaks in the genes, and we also required the input RPKMs of the peaks ≥5 in both replicates to obtain reliable peak intensities. Gene ontology (GO) analyses for top 1,000 highly methylated genes and bottom 1,000 weakly methylated genes were conducted using DAVID⁴⁹; all genes with the above required read counts for level calculation were used as background. Published half-life measurements of iPSC²⁴ were used to compare with m⁶A levels, and the genes with average FPKMs of the two H1 replicates <5 were removed from the analyses.

Comparison of m⁶A levels between cell lines

For all the comparisons of m⁶A levels between H1-ESC and GM12878 cell lines, we required eluate counts ≥50 or supernatant counts ≥50 in both replicates of both H1-ESC and GM12878 cell lines to obtain reliable m⁶A levels. Differentially methylated genes were determined as the genes with m⁶A-level differences between cell types of > 10% or 20%.

Splicing analyses

We used rMATS (v 3.0.8)³⁴ to determine the differential alternative splicing events between different libraries. Four types of alternative splicing events (skipped exon (SE), alternative 5′ splice site (A5SS), alternative 3′ splice site (A3SS), and retained intron (RI)) based on annotated Ensembl genes were tested for inclusion-level difference ≥5%, and the events with FDR < 0.05 were determined as differential alternative splicing events.

Differential usage of alternative poly(A) sites

Poly(A) sites from the APADB database⁵⁰ were identified within the 3′-UTR regions of Ensembl protein-coding genes. For genes with multiple poly(A) sites, any pairs of poly(A) sites are defined as the end of the 3′ termini of common and extended 3′ UTR (the different region between longer 3′ UTR and shorter 3′ UTR) segments. If the lengths of the 3′-UTR segments are both greater than 100 nt (to rule out the regions determined by two close poly(A) sites), they are considered a valid pair of tandem UTRs. For each pair of tandem UTRs, the number of RNA-seq reads mapped within the common and extended 3′ UTRs were calculated for each replicate of supernatant and eluate, respectively. The P:D (proximal:distal) ratio plus 1 (P:D + 1) was calculated as the ratio of region-length normalized read counts between common and extended 3′ UTR. We then used the statistic model of rMATS³⁴ to test the differences of the read counts ratio of common and extended 3′ UTRs between supernatant and eluate; the 3′-UTR pairs with read depths (total reads in common and extended 3′ UTR) <5 in any replicate of supernatant or eluate were removed from the calculation. The 3′-UTR pairs with FDR < 0.01 were then identified as differentially used alternative poly(A) sites between supernatant and eluate; we then took the most significant 3′-UTR pair of each gene for the downstream analyses. Based on the above analyses, we classified three categories of genes for further analyses: group 1, eluate has higher proximal poly(A) site usage; group 2, supernatant has higher proximal poly(A) site usage; group 3, eluate and supernatant have no significant (FDR > 0.5) difference of APA usage.

We then studied the m⁶A peak coverage around proximal poly(A) sites for genes in group 1 and group 2, respectively, in H1-ESC. For the transcript with the extended 3′ UTR of each gene, the peak coverage of upstream and downstream 750 bp flanking the last nucleotide of proximal poly(A) site was calculated. The transcripts with either flanking regions <750 bp were removed from the analyses. To study the relationship between m⁶A level and 3′-UTR length, we used the longest 3′ UTR of all isoforms for each gene.

m⁶Am analyses

To create a list of protein-coding genes that were unlikely to harbor m⁶Am, we removed the genes with m⁶A peaks within the first 200 nt of at least one Ensembl-annotated transcript isoform in H1-ESC and GM12878, respectively. The m⁶A peaks of GM12878 were obtained from previous publication⁵¹. To determine whether the genes had A at TSSs, we used the Ensembl-annotated TSSs of the longest protein-coding isoforms or TSSs annotated by ENCODE CAGE data of poly(A)⁺ mRNAs of H1-ESC and GM12878. The processed data summarizing CAGE tags were downloaded from http://ccg.vital-it.ch/mga/hg19/encode/GSE34448/GSE34448.html. For the CAGE data, we only used the TSSs with CAGE tags ≥10 in both replicates and the TSSs must have been located in the 5′ UTR of first exon of the longest protein-coding isoforms. For each gene, the TSS with maximum average CAGE tags was used in the analyses.

miRNA and RBP analyses

We compare the densities of miRNA targets and RBP binding sites in the extended 3′ UTRs in the 3 categories of genes in terms of poly(A) usage as described above. miRNA targets predicted by TargetScanHuman 5.1 were downloaded from the UCSC genome browser⁵². RBP binding sites were predicted through matching RNAcompete-derived motifs of 110 RNA-binding proteins and other well-known RNA-binding motifs^{53, 54, 55}. Only the genes with extended 3′-UTR lengths between 500 nt and 2,000 nt were used in RNA-binding-protein analyses.

Data visualization

UCSC custom data tracks were generated using the uniquely mapped reads. For Figure 2c,e and Supplementary Figure 2e, to make the read coverage of supernatant and eluate reflect the real ratio of their RNA abundance, we used the raw reads coverage of eluate, but we multiplied the supernatant read coverage by 2^R (as described above, R was the residue of ERCC regression formula).