Heterogeneity of chemical modifications on RNA

Abstract

The chemical modifications of RNAs broadly impact almost all cellular events and influence various diseases. The rapid advance of sequencing and other technologies opened the door to global methods for profiling all RNA modifications, namely the “epitranscriptome.” The mapping of epitranscriptomes in different cells and tissues unveiled that RNA modifications exhibit extensive heterogeneity, in type, amount, and in location. In this mini review, we first introduce the current understanding of modifications on major types of RNAs and the methods that enabled their discovery. We next discuss the tissue and cell heterogeneity of RNA modifications and briefly address the limitations of current technologies. With much still remaining unknown, the development of the epitranscriptomic field lies in the further developments of novel technologies.

Introduction

The study of RNA modification began in 1957, when pseudouridine (Ψ) was first identified from the hydrolysate of bulk yeast RNA (Cohn and Volkin 1951). In the following decades, more and more modified RNAs have been identified, along with the discovery of RNA modifying enzymes (Grembecka et al. 2012; Zuber et al. 2011). For example, the first Ψ synthase TruA was found to modify uridine into Ψ on tRNA (Ciampi et al. 1977; Kammen et al. 1988). Increasing evidence points to a new concept that co-transcriptional and post-transcriptional modification by enzymes occurs on RNA from all kingdoms of life. The term “epitranscriptome” soon emerged (Meyer et al. 2012), denoting the collective biochemical modifications on RNA within a cell. Novel enrichment strategies coupled with advanced sequencing technologies have revolutionized the scope of understanding of epitranscriptomics, leading to the transcriptome-wide mapping of a handful of RNA modifications in different organisms. The drastic burst of systematic investigation of RNA modifications points out that the RNA modifications directly impact RNA biology in nearly every aspect, such as transportation, localization, stability, and translation regulation (Roundtree et al. 2017). Accumulating evidence suggests that RNA modifications precisely regulate RNA functions to heavily influence cell fate; thus, the type and location of many RNA modifications are heterogeneous among different tissues and cells, changing dynamically in response to stresses and disease conditions (Gu et al. 2014). Thus, RNA modifications play fundamental roles in a broad spectrum of cellular events and influence various disease conditions, including cancer, obesity, and neurological disorders. In this review, we provide an overview of RNA modifications and their current mapping methods before highlighting recent growing interests in exploring heterogeneity of RNA modifications.

RNA modifications landscape

The early studies on RNA modifications focused on highly abundant types of modification, such as Ψ and N6-methyladenosine (m6A). Driven by the technological advances, less abundant and even rare modifications emerged. To date, more than 170 types of RNA modifications have been recorded on RNAs of various organisms (Boccaletto, et al. 2017). While majority of the RNA modifications were found on tRNA (Pan 2018) and rRNA (Starr et al. 1964), many modifications were also found on mRNAs (Harcourt et al. 2017) as well as other small non-coding RNAs (Li et al. 2021). In this section, we will discuss the modifications found on mRNA, rRNA, and tRNA (Fig. 1), as their heterogeneity profile is the most well studied.

Fig. 1.

Representative RNA modifications classified by their reference nucleotide. The occurrence of the modifications on the three major types of RNAs was indicated in the bracket: t for tRNA, r for rRNA, and m for mRNA. The data were extracted from the MODOMICS database (Boccaletto, et al. 2021)

mRNA

N6-methyladenosine (m6A), 5-methylcytidine (m5C), and ribose 2′OMe modifications (Nm) are the first modifications discovered on mRNA in the 1970s (Dubin and Taylor 1975; Schibler and Perry 1977). Other modifications found on mRNA include N1-methyladenosine (m1A), N6,2′-O-dimethyladenosine (m6Am), and Ψ (Dominissini et al. 2016; Eyler et al. 2019; Lim et al. 2018). Among these modifications, m6A is the most prevalent and the well-studied internal modification on mRNA. Various sequencing methods have been developed to precisely locate m6A on mRNAs at single-nucleotide resolution. These transcriptome-wide m6A analysis revealed that one-third of mammalian mRNAs contain 3 to 5 m6A modifications, with most m6As located in conserved motifs that are frequently identified around the stop codon of mRNAs (Jiang et al. 2021). Besides the mapping of m6A itself, a list of proteins that metabolize m6A have been discovered: “Writer” proteins are enzymes that modify adenine into m6A, including METTL3/14/16 and RBM15/15B (Jiang et al. 2021). “Eraser” proteins are enzymes that remove the modification, such as FTO and ALKBH5 (Jia et al. 2011; Zheng et al. 2013). “Reader” proteins recognize and bind specifically to m6A and depending on the identity of the reader, the modified mRNA is directed towards different downstream pathways. For instance, YTHDF1/2/3 directs modified mRNA towards degradation by the CCR4-NOT complex; YTHDC1/2 recruits other factors involved in remodeling the epigenome proximal to nascent m6A-modified RNA; IGF2BP1/2/3 acts to stabilize m6A-modified mRNA (Zaccara and Jaffrey 2020; Xu et al. 2014; Wojtas et al. 2017; Bailey, et al. 2017; Huang et al. 2018).

rRNA

The rRNAs in the cells of all kingdoms of life are extensively modified. The main modifications found on rRNA can be categorized into three types. Ribose 2′-OMe methylation and pseudouridylation counts for the majority of rRNA modifications, with base modifications being the third type of modification (Sharma and Lafontaine 2015). Besides sequencing, single particle cryo-electron microscopy (cryo-EM) can also directly visualize the modifications on rRNA, as rRNA folds into stable 3D structures in ribosomes (Natchiar et al. 2017). Both sequencing data and the cryo-EM data agree that the majority of modifications of rRNA are found on the 18S and 28S rRNAs. With a few modifications found on the 5.8S rRNA, no modifications were seen on the 5S rRNA. Interestingly, some modifications predicted by sequencing data, such as Gm75 (ribose 2′-OMe) on 5.8S rRNA, are absent from the cryo-EM result, while some modification detected by cryo-EM are not found from sequencing, including m5C4447 and Gm1883 (ribose 2′-OMe) on 28S rRNA. RNA sequencing analyzes bulk rRNA samples from large numbers of cells and the cryo-EM data is obtained from visualizing limited numbers of rRNAs. The disagreement on many modifications between sequencing and cryo-EM data is a strong evidence that rRNA modifications could be ribosome-specific (Natchiar et al. 2017). Several studies on differently specialized ribosomes and cancer-related ribosomes also show the differential modifications among ribosomes, suggesting the ribosome heterogeneity can be a signature of rRNA modifications (Sloan et al. 2017; Slavov et al. 2015; Preiss 2016; Marcel et al. 2015; Vanden Broeck and Klinge 2023). With more than 130 rRNA modifications found on rRNAs, their molecular mechanisms and exact impact to rRNA are largely unknown. To date, increasing evidence suggests that the extensive modifications strongly affect rRNA structure and stability. As the major role of rRNA is to ensure proper binding and alignment of ribosome to mRNA and tRNA, the modifications are thought to modulate the catalytic activity of ribosome, contributing to the regulation of protein synthesis. Some rRNA modifications are found to be associated with the dysregulated protein synthesis in pathological conditions, including inherited human disorders and cancer (Janin et al. 2020; Meyer et al. 2011).

tRNA

To date, 16 types of modifications were found on tRNA from different organisms (Suzuki 2021), including m1A, N1-methylguanosine (m1G), N2,2-dimethylguanosine (m2,2G), N3-methylcytidine (m3C), m5C, and Ψ. With only 73–93 nucleotides in length, tRNAs share a highly conserved characteristic three-leafed clover tertiary structure, with one loop structure hosting the trinucleotide anticodon. Many modifications are found on the anticodon loop of tRNA, affecting the codon recognition and the reading frame maintenance. The remaining part of the tRNA is also heavily modified, which affects the shape, the stability, localization, and quality check of tRNA (Lyons et al. 2018). Recent studies found that the tRNA modifications change dynamically under stress or in response to cellular conditions. For instance, hydrogen peroxide can alter the level of 2′OMe-C (Cm), m5C, and m1G (Chan, et al. 2010; Duechler et al. 2016). Such dynamic alteration of tRNA modification is essential for optimal translation of stress response-associated mRNAs.

RNA modification profiling

The initial studies on RNA modification heavily rely on chemical approaches to isolate and identify the modified nucleotides. Ψ was first isolated by paper chromatography (Cohn 1960; Davis and Allen 1957) and its structure was determined using NMR and periodate oxidation experiments (Davis 1995). The early research identified the structure of dozens of modified nucleotides but are unable to reveal the locations of the modifications on RNAs. As we learnt more about RNA modifications, it became clear that their distributions provide an additional layer of the complexity of the epitranscriptome. With the rapid advance of high throughput analysis methods, we are now able to map many types of RNA modifications on a transcriptome-wide scale, which enable the discovery of modification heterogeneity. In this section, we will discuss the major techniques for transcriptome-wide RNA modification profiling (Table 1).

Table 1.

Second-generation sequencing based transcriptome-wide RNA modification profiling methods

Short nomenclature	Full name	Transcriptome-wide detection method
I	Inosine	RNA-Seq (A-to-G conversion) ice-Seq
m6A	N6-methyladenosine	MeRIP/m6A-seq miCLIP SCARLET m6A-LAIC-Seq m6ACE-Seq m6A-CLIP-seq m6A-REF-Seq DART-Seq m6A-SAC-seq GLORI-Seq eTAM-Seq
m1A	1-methyladenosine	RIP-Seq
m6Am	N6,2′-O-dimethyladenosine	miCLIP m6ACE-Seq
Ψ	Pseudouridine	Pseudo-Seq
m5C	5-methylcytidine	RNA-Seq (C-to-U conversion) m5C-RIP AZA-IP
hm5C	5-hydroxymethylcytosine	hMeRIP-Seq
m3C	3-methylcytidine	AlkAniline-Seq
m7G	7-methylguanosine	AlkAniline-Seq
2′-OMe	2′-O-methylation	Nm-REP-seq

Mass spectrometry- and cryo-electron microscopy-based methods

Liquid chromatography coupled with tandem mass spectrometry (LC–MS/MS) directly reports RNA modification based on their physical properties, without the need for conversion of RNA to cDNA as in NGS (Chan, et al. 2010). Around a decade ago, LC–MS/MS-based approaches emerged to profile specific modifications. tRNAs are treated with base-specific ribonucleases to produce small oligonucleotides. LC–MS/MS analyzes the molecular weight of these oligonucleotides and their ionized fragments to identify the type and location of RNA modification. Even the mass-silent Ψ can be detected by detecting its unique and hydrolysis-resistant C–C glycosidic bond (Addepalli and Limbach 2016). Recently, multi-modification profiling of tRNA using LC–MS/MS is made possible through the use of site-specific RNase H digestion (Yan et al. 2021). This approach is unbiased to tRNAs with different base composition, significantly improving the information content of collision-induced dissociation (CID) spectra. This approach is universally applicable to any tRNA species and can simultaneously detect five species of modifications.

As mentioned above in the rRNA section, cryo-electron microscopy (cryo-EM) has been extensively used for visualization of rRNA modifications, as rRNA folds into stable structure inside the ribosome (Zhang, et al. 2019; Saletore et al. 2012). Single particle cryo-EM with focused refinement techniques can push the resolution of 60S ribosomal subunit to 2.9 Å, with local resolution at some regions reaching 2.5 Å. Such high resolution presents unprecedented detail in rRNA structure to visualize more than 130 rRNA modifications on the human ribosome. The advantage of cryo-EM lies in that it can directly resolve the 3D environment of the modification sites, presenting a huge advantage for studying the impact of modification to rRNA structure and for discovering specific ligands. The limitation of cryo-EM lies in that not all modifications can be resolved. For example, Natchiar and co-workers found that due to the isomeric nature of Ψ, its presence can only be confirmed in the structure when the N1 position forms a hydrogen bond (Natchiar et al. 2017), meaning that the information of the Ψ with free N1 position cannot be captured by cryo-EM.

Next-generation sequencing-based methods

Next-generation sequencing (NGS) or second-generation sequencing technologies have evolved rapidly over the past decade. It is a powerful tool for high-throughput and massively parallel sequencing (Saletore et al. 2012). As RNA molecules must be converted into cDNA for NGS, the key of RNA modification detection lies in how to preserve the modification information during library preparation. Based on the strategy for library preparation, there are mainly three categories of NGS-based RNA modification sequencing.

The first category uses modification-specific antibodies to immunoprecipitate fragmented RNA oligos (RIP-seq). Sequencing of the enriched RNA fragments followed by alignment of the fragments can report the location of the modifications. The first generation of RIP-seq simply relies on immunoprecipitation, such as m6A-seq and MeRIP-seq that were developed in 2012 (Meyer et al. 2012; Dominissini et al. 2012). The resolution of these methods is 100–200 nt, which was the range of the fragment size. Strategies involving photo-crosslinking-assisted library preparation protocols such as miCLIP and m6ACE-Seq have further refined the resolution to as high as single-base resolution (Linder et al. 2015; Ke et al. 2015; Koh et al. 2019). With the further development of modification-specific antibodies, a handful of modified nucleotides can now be detected at high resolution using RIP-seq.

The second category uses chemical compounds that selectively react with specific ribonucleotides, converting it into other bases. The most widely used method uses bisulfite treatment to convert a cytosine to a uridine, leaving methylated cytosine unchanged (Schaefer et al. 2008). Comparison of bisulfite-treated library with untreated library reveals the methylation site of cytosine. Other methods in this category include GLORI-Seq, Pseudo-Seq, and AlkAniline-Seq (Carlile et al. 2015; Marchand et al. 2018; Liu et al. 2023). Incomplete conversion during chemical treatment and treatment resistance due to the presence of other modifications are the two main factors that limit the accuracy of chemical-based sequencing. Nonetheless, they are widely used as they are efficient and cost-effective.

The third category uses enzyme treatment to selectively convert specific ribonucleotides into other bases for detection. These methods include Nm-REP-seq for 2′-O-methylation detection (Dimitrova et al. 2019) and Mazter-Seq, m6A-REF-Seq, DART-Seq, and eTAM-Seq for m6A detection (Garcia-Campos et al. 2019; Meyer 2019; Hu et al. 2022; Xiao et al. 2023). Nm-REP-seq uses RNA exoribonuclease and periodate oxidation reactivity to remove 2′-O-methylation sites. The comparison of enzyme treated and untreated library produces single-base resolution of 2′-O-methylation at the 3′-end of small non-coding RNAs, such as snoRNAs, snRNAs, tRNAs, piRNAs, and miRNAs. Hybrid protocol that combines enzyme treatment with RIP-seq has been developed to allow single-base resolution detection of modification, such as Aza-IP for the detection of m5C (Khoddami and Cairns 2013).

Nanopore-based methods

Next-generation sequencing methods can only indirectly report RNA modifications as they are blind to nucleotide modifications. Although they provide valuable information on epitranscriptome, their applications are limited by the availability of modification-specific antibody, chemical or enzyme reactions. Oxford Nanopore direct RNA sequencing, which is a form of third-generation sequencing, can directly report RNA modification in native RNA sequences. As RNA passes through the pore, distinct features of the disruptions in the electrical signal reveal the identity of each nucleotide. To date, several Nanopore-based methods have been developed to accurately report m6A. Some of the methods, such as EpiNano, MINES, and nanom6A (Lorenz et al. 2019), use training data to identify m6A using the increased “errors” and decreased qualities caused by m6A in the current intensity. Another strategy, including xPore (Pratanwanich et al. 2021) and m6Anet (Hendra et al. 2022), uses neural network to compare signals from m6A-containing and m6A-free samples to determine the location of m6A. Similar strategies have been applied to sequence other modifications using Nanopore, including m7G and Ψ (Smith et al. 2019). Since Nanopore-based methods can directly capture RNA modification information in RNA molecules with high accuracy and high throughput, they are promising tools for deepening the understanding of epitranscriptome. However, Nanopore-based methods suffer from several drawbacks. First, the amount of input RNA required is still at the scale of hundreds of nanograms of mRNA so clinical samples might not have enough material for direct RNA sequencing. Second, direct RNA sequencing still suffers from a high base-calling error rate compared to second-generation sequencing. Fortunately, improved chemistry and consumables in upcoming versions of the library preparation/sequencing kits promise higher sequencing accuracy and reduced input requirements. Together with community-driven efforts to improve base/modification-calling software, direct RNA sequencing holds great potential to become the de facto method for profiling RNA modifications.

Layers of RNA modification heterogeneity

Heterogeneity of writer and host RNA species

Heterogeneity of RNA modifications can be defined at many levels. The most obvious one is the different types of RNA modifications, with the current count at over 170 different RNA modifications (Boccaletto, et al. 2021). While this review focuses on the m6A RNA methylation, even this single RNA modification alone exemplifies 2 additional levels of heterogeneity: the writer that methylates A to give m6A and the RNA type that m6A is hosted on. Here, a specific m6A writer only generates m6A on specific target RNAs: Mettl3 and its co-factors methylate mRNA mostly in the 3′UTR; Mettl16 methylates Mat2a 3′UTR and U6 snRNA; Mettl5 methylates 18S rRNA; Zcchc4 methylates 28S rRNA (Meyer et al. 2012; Dominissini et al. 2012; Tran et al. 2019; Ma et al. 2019; Pinto et al. 2020). Another RNA methylation, m6Am exemplifies that RNA modifications can also be categorized not just by their writers but also by the erasers that regulate them: Mettl4 generates m6Am internally within U2 snRNA, while Pcif1 generates m6Am on both mRNAs and snRNA 5′ termini (Koh et al. 2019; Chen et al. 2020; Goh et al. 2020; Akichika, et al. 2019; Sun et al. 2019; Sendinc et al. 2019; Boulias et al. 2019). The latter group of m6Am can be further divided by the fact that the FTO eraser can demethylate snRNA 5′m6Am but not demethylate mRNA 5′m6Am (Koh et al. 2019; Mauer et al. 2019). In summary, RNA modification type, host RNA, writer, and eraser dependencies are just some of the RNA modification heterogeneity levels that have been explored so far.

Tissue heterogeneity

Early on in the field of epitranscriptomics, limitations in methylome sequencing technology limited further exploration of RNA modification heterogeneity to only the tissue level. For most of the initial years where mRNA methylation was being profiled, only methylomes of a small number of mammalian cell lines and mouse tissues were generated (Meyer et al. 2012; Dominissini et al. 2012). One early effort in cataloging a comprehensive tissue-level RNA methylome landscape profiled RNA m6A and m6Am landscape across various human and mouse tissues (Fig. 2) (Liu et al. 2020). Within each species, the greatest distinction in tissue-specific methylome patterns was between brain and non-brain tissues. This observation was also echoed by another recent study, albeit for fetal tissues (Xiao, et al. 2019). The observed methylome heterogeneity could be slightly explained by positive correlation of m6A levels with m6A writers’ expression and negative correlation with m6A eraser expression across the assayed tissues. Likewise, Liu et al. observed positive correlation between m6Am levels and m6Am writer expression but not negative correlation with the purported mRNA m6Am eraser, FTO. This is likely due to the previous observation that though FTO has clear in vitro m6Am demethylation activity, it likely does not demethylate mRNA m6Am in vivo (Koh et al. 2019; Mauer et al. 2019). Overall, Liu et al. were able to uncover many transcripts that exhibit tissue-specific m6A patterns, demonstrating that RNA methylation can be used to distinguish different cell subtypes and thus defines a new layer of tissue heterogeneity. Recent studies have reported that m6A patterns are largely determined by the METTL3 writer complex only methylating DRACH sites not occluded by the exon-junction complex (EJC) (Yang et al. 2022; He et al. 2023; Uzonyi et al. 2023). This suggests that tissue-specific alternative splicing pattern differences is one underlying factor that contributes to tissue-specific m6A methylomes.

Fig. 2.

Cross-tissue m6A and m6Am methylomes highlight tissue heterogeneity of m6A (Adapted from Liu et al. 2020; Rightslink order:5483470919073)

Since the previous efforts in mapping RNA methylomes in different tissues were performed using low-resolution RNA methylome mapping techniques, our group has made use of the single-base-resolution m6A sequencing technology we developed, xPore, to see if a higher methylome profiling precision can reveal further insights regarding heterogeneity of RNA modifications (Pratanwanich et al. 2021). Focusing on various human cell lines, we found that m6A levels were generally stable across different cell lines. However, the single-base-resolution of xPore afforded the opportunity to compare m6A differences at specific m6A motifs (Fig. 3). This revealed that specifically at “GGACU” motifs, K562 lymphoblast cells and HCT116 colon cancer cells exhibited higher methylation, while HEPG2 liver cancer cells and MCF7 breast cancer cells exhibited lower methylation than HEK293T embryonic kidney cell lines (Fig. 3). This difference was specific to “GGACU” and not to other common m6A motifs such as “AGACU” or “GGACA,” suggesting that motif-specific heterogeneity underlies cell identity. This highlights the importance of requiring sequencing precision in future endeavors for mapping additional levels of RNA modification-based heterogeneities.

Fig. 3.

High-resolution m6A sequencing reveals cell-specific motif hypermethylation (Adapted from Pratanwanich et al. 2021 Nature Biotechnology (Pratanwanich et al. 2021)

Single-cell heterogeneity

Unlike efforts to map RNA modification heterogeneity across bulk tissues or bulk cell lines, efforts to map RNA modification heterogeneity at the single-cell level are less frequent. The only effort so far was by Tegowski et al. who adapted their previously developed DART-seq for single-cell m6A sequencing (Tegowski et al. 2022). DART-seq exogenously expressed m6A-binding YTH domain fused to an APOBEC1 cytidine deaminase to direct C-to-U editing near m6A sites. Sequencing and mapping these edited RNA sites allows for transcriptome-wide mapping of m6A sites at high resolution. Tegowski et al. combined DART-seq with droplet-based single-cell (scRNA) sequencing to develop single-cell-DART-seq (scDART-seq) for single-cell m6A sequencing (Fig. 4). Using scDART-seq, Tegowski et al. found a discrepancy between single-cell-level and bulk population-level m6A prevalence, which they attributed to the high m6A heterogeneity amongst the single cells. Interestingly, m6A sites determined to have low stoichiometry were found to be abundant in a subset of single cells within the population. Furthermore, they found that cells were best clustered based on m6A patterns found on mRNAs encoding proteins involved in RNA processing and RNA binding. Together, these findings underscore the utility of single-cell m6A sequencing over bulk m6A sequencing and its potential for identifying unique cell subtypes.

Fig. 4.

scDART-seq for detecting m6A in single cells reveals single-cell heterogeneity of m6A (Adapted from Tegowski et al. 2022 Molecular Cell (Tegowski et al. 2022); Rightslink order:5483470360033).

Future direction

Moving forward, we envision interest in the field will focus on exploring the single-cell heterogeneity aspect of RNA modifications, especially given that this aspect also encompasses the tissue heterogeneity aspect. A major hurdle against this exploration is the limitation of current single-cell RNA modification sequencing methods. Despite its utility, the aforementioned scDART-seq possesses some restrictions that limit its adoption as a general single-cell m6A sequencing technique. This includes the requirement to express APOBEC1-YTH in cells, preventing its use in clinical samples. Furthermore, without having a APOBEC1-mutantYTH expression control (which edits RNA in a non-specific manner) as the authors did for bulk DART-seq (Meyer 2019), it is hard to rule out that any identified m6A is not a false-positive. These problems will hopefully be overcome with further optimization.

On the other hand, third-generation sequencing in the form of Nanopore direct RNA sequencing provides an alternative strategy to profile m6A at the single-cell resolution. We had already demonstrated xPore’s ability to map differential m6A at single-base resolution (Pratanwanich et al. 2021). By combining xPore with droplet-based methods for isolating single cells as well as a method to barcode and distinguish the RNA from each single cell, we envision that direct RNA sequencing can be adopted to develop a new single-cell m6A sequencing method that can be universally applied to both experimental cell culture systems and clinical samples in a highly scalable manner. To conclude, the field of heterogeneity in RNA modification, especially of single-cell heterogeneity, remains a vast area to be explored and further developments in epitranscriptome sequencing technologies will be the key to this exploration.

Author contribution

All authors contributed to the writing of the manuscript. All authors read and approved the final manuscript.

Funding

We thank the General Research Fund from Research Grants Council (RGC) of Hong Kong (16101720) and the National Natural Science Foundation, Research Fund for International Excellent Scientists (32250610210) for funding.

Declarations

Ethics approval

There is no involvement of human or animal subjects.

Consent to participate

There is no involvement of human subjects.

Consent for publication

There is no involvement of human subjects.

Competing interests

W.S.S.G. owns shares in Oxford Nanopore Technologies. The remaining author hereby declares no conflict of interest relating to this manuscript.