General Principles and Limitations for Detection of RNA Modifications by Sequencing

Conspectus

Among the many analytical methods applied to RNA modifications, a particularly pronounced surge has occurred in the past decade in the field of modification mapping. The occurrence of modifications such as m⁶A in mRNA, albeit known since the 1980s, became amenable to transcriptome-wide analyses through the advent of next-generation sequencing techniques in a rather sudden manner. The term “mapping” here refers to detection of RNA modifications in a sequence context, which has a dramatic impact on the interpretation of biological functions. As a consequence, an impressive number of mapping techniques were published, most in the perspective of what now has become known as “epitranscriptomics”. While more and more different modifications were reported to occur in mRNA, conflicting reports and controversial results pointed to a number of technical and theoretical problems rooted in analytics, statistics, and reagents. Rather than finding the proverbial needle in a haystack, the tasks were to determine how many needles of what color in what size of a haystack one was looking at.

As the authors of this Account, we think it important to outline the limitations of different mapping methods since many life scientists freshly entering the field confuse the accuracy and precision of modification mapping with that of normal sequencing, which already features numerous caveats by itself. Indeed, we propose here to qualify a specific mapping method by the size of the transcriptome that can be meaningfully analyzed with it.

We here focus on high throughput sequencing by Illumina technology, referred to as RNA-Seq. We noted with interest the development of methods for modification detection by other high throughput sequencing platforms that act directly on RNA, e.g., PacBio SMRT and nanopore sequencing, but those are not considered here.

In contrast to approaches relying on direct RNA sequencing, current Illumina RNA-Seq protocols require prior conversion of RNA into DNA. This conversion relies on reverse transcription (RT) to create cDNA; thereafter, the cDNA undergoes a sequencing-by-synthesis type of analysis. Thus, a particular behavior of RNA modified nucleotides during the RT-step is a prerequisite for their detection (and quantification) by deep sequencing, and RT properties have great influence on the detection efficiency and reliability. Moreover, the RT-step requires annealing of a synthetic primer, a prerequisite with a crucial impact on library preparation. Thus, all RNA-Seq protocols must feature steps for the introduction of primers, primer landing sites, or adapters on both the RNA 3′- and 5′-ends.

Key References

• Hauenschild R.; Tserovski L.; Schmid K.; Thüring K.; Winz M.-L.; Sharma S.; Entian K.-D.; Wacheul L.; Lafontaine D. L. J.; Anderson J.; Alfonzo J.; Hildebrandt A.; Jäschke A.; Motorin Y.; Helm M.. The Reverse Transcription Signature of N-1-Methyladenosine in RNA-Seq Is Sequence Dependent. Nucleic Acids Res. 2015, 43, 9950–9964 .¹ Describes deep sequencing analysis of m¹A RT signatures in rRNA and tRNAs with Random Forest algorithm for classification of hits.

• Werner S.; Schmidt L.; Marchand V.; Kemmer T.; Falschlunger C.; Sednev M. V.; Bec G.; Ennifar E.; Höbartner C.; Micura R.; Motorin Y.; Hildebrandt A.; Helm M.. Machine Learning of Reverse Transcription Signatures of Variegated Polymerases Allows Mapping and Discrimination of Methylated Purines in Limited Transcriptomes. Nucleic Acids Res. 2020, 48, 3734–3746 .² Extended study of RT signatures for m¹A, m⁶⁶A, m¹G, and m²²G in RNA using a set of RT enzymes and Random Forest model.

• Marchand V.; Ayadi L.; Ernst F. G. M.; Hertler J.; Bourguignon-Igel V.; Galvanin A.; Kotter A.; Helm M.; Lafontaine D. L. J.; Motorin Y.. AlkAniline-Seq: Profiling of m⁷G and m³C RNA Modifications at Single Nucleotide Resolution. Angew. Chem., Int. Ed. 201857, 16785–16790.³ Description of highly sensitive protocol for detection of RNA modifications via formation of abasic site and subsequent cleavage.

• Marchand V.; Pichot F.; Neybecker P.; Ayadi L.; Bourguignon-Igel V.; Wacheul L.; Lafontaine D. L. J.; Pinzano A.; Helm M.; Motorin Y.. HydraPsiSeq: A Method for Systematic and Quantitative Mapping of Pseudouridines in RNA. Nucleic Acids Res. 2020, 48, e110 .⁴ Description of the protocol for mapping and quantification of pseudouridines and 5-substituted U residues in RNA.

Introduction

RNA modifications, now known as “epitranscriptomics”,⁵⁻⁸ differ from their parental unmodified nucleotides in their chemical structure and, in consequence, their physicochemical properties and chemical reactivity, including in particular base pairing properties (Figure 1A,B).

Figure 1.

RNA nucleobases and modified nucleotides in RNA. (A) Four canonical RNA purine and pyrimidine nucleobases (A, G, C, U), with atom numbering. Watson–Crick (WC) and Hoogsteen edges as well as donors (red)/acceptors (blue) for H-bond formation are indicated. (B) Some prominent examples of modified RNA nucleobases. Deviations from parental nucleosides are shown in green. Chemical structures are shown for neutral pH (pH 7.0) and protonation status is indicated considering the pK_a values.^9,10 (C) 2′-O-Methylation and (D) m¹A as an example of a modification that affects Watson–Crick pairing.

These features sometimes allow mapping of RNA modifications by RNA-Seq. Altered base pairing, typically linked to the presence of additional chemical groups at the Watson–Crick edge of the base (Figure 1B,D), may be detected by a number of classical methods, by RNA-Seq adaptations, as well as by most other next-generation sequencing (NGS) technologies.¹¹⁻¹⁵ For those methods relying on reverse transcription (RT) prior to sequencing (Figure 2A), modified sites may appear either as strong arrest of primer extension (RT-stop) or, in the majority of cases, as the presence of misincorporations in cDNA (“mutations”) and as base deletions (jumps). These three features in the sequenced cDNA strand together constitute a so-called RT-signature (Figure 2B), which stems from the inability of the enzyme to bind base-pair incoming dNTPs with an RNA modification in the template.

Figure 2.

Synthesis of cDNA by reverse transcriptase (RT) from RNA template. The enzyme requires a hybridized primer and a free 3′-OH to be extended (A). Modified nucleotide in the template (shown as X) can alter the RT-signature, through misincorporations, deletions, or abortive cDNA synthesis (B).

Reverse transcriptases, when moving along an RNA template, employ four different substrates, namely, the four dNTPs, for cDNA synthesis, presumably each with a different affinity and turnover, approximated by different K_m and k_cat values. Based on this alone, one must assume differential catalytic efficiency for each different templated RNA nucleoside. Every single parameter of the in vitro conditions is, in principle, liable to differentially affect incorporation of each different incorporated dNTP. These parameters include, in particular, the individual concentrations of each of the four dNTPs, pH, divalent cations, ionic strength, temperature, and, of course, the type of enzyme and specific activity of each individual enzyme preparation. Biases in efficiency of cDNA synthesis will be further compounded by sequence context, i.e., RNA primary structure, and by stable RNA secondary and tertiary structures that need to be resolved before the RT can process the template. Consequently, RT-stops will occur in an uneven and currently unpredictable distribution, forming a background of events that constitute substantial noise against which modification-dependent RT-stops must be contrasted with statistical significance. Moreover, RNA modifications are installed site-specifically by dedicated modification enzymes, albeit a given site may not be fully modified. For such substoichiometric modification the RT-stop signal is necessarily lower, decreasing the signal-to-noise ratio and thus detection efficiency and confidence.

Before the advent of NGS, the one modification that was easily detectable by normal RT-based sequencing was inosine, a deamination product of adenosine that very faithfully base-pairs with cytidine. This corresponded to a quantitative misincorporation and allowed large scale mapping approaches even with basic sequencing techniques.¹⁶⁻¹⁸ Beyond this, RT-stops caused by different modifications had already been observed and used in an analytical perspective to a limited degree.^19,20 The influence of a selected few in vitro parameters on RT-stops had also already been explored and exploited, for example, for the detection of Nm modifications (Figure 1C) by primer extension at low dNTP concentrations.^21,22 The influence of such parameters on misincorporation in the sequence of the cDNA was, however, only known from DNA polymerases used in PCR.²³ With the onset of the NGS era, altered base pairing properties of RNA modified nucleotides and ensuing misincorporation became more apparent²⁴ and an important source of information to identify positions of RNA in generic RNA-Seq data sets.²⁵⁻²⁷

In more sophisticated approaches, the above-mentioned modulation of RT conditions, e.g., via concentration of metal ions²⁸ and dNTP substrates, were now adapted to NGS, e.g., for the detection of Nm residues.²⁹ This is one example of many where the Watson–Crick edge is unaffected by the modification, causing a weak or invisible RT-signature.^27,30 In the case at hand, it becomes enhanced through modulation of the RT-conditions, namely lowered dNTP conditions which diminish the reverse transcriptase activity to the point where it stalls and produces abortive cDNA. Further sophistication included the combination of stalling and misincorporation as important contributors to the RT-signature,²⁸ which was then shown to vary, e.g., among different enzymes such that the combined RT-signature of two RT-enzymes was more informative than one alone.² This inspired the directed evolution of enzymes whose RT-signature was specifically sensitive to modifications such as m⁶A, Nm, Q and Ψ, which were otherwise considered to be RT-silent.³¹⁻³³ Another sophistication included jumps in the RT-signature, a feature that seems to be fraught with less background.³⁴⁻³⁸ However, it became clear that the number of RNA modifications whose altered base pairing properties allowed comprehensive detection was limited.³⁹

To address this gap, a growing number of methods were and still are being developed that use chemical or enzymatic treatments, which, by selectively acting on modified or unmodified nucleotides enhance the information content of sequencing data. This amounts to a conversion of RT-silent modifications into chemical structures that can now be detected by RT-signature. A pioneering method that was already developed in the 1970s^40,41 uses a carbodiimide, specifically 1-cyclohexyl-(2-morpholinoethyl)carbodiimide metho-p-toluene sulfonate (CMCT), which forms alkaline resistant covalent adducts with pseudouridine. A pseudouridine thus becomes detectable as a CMCT-dependent RT-stop.⁴²⁻⁴⁴ Of note, this method has meanwhile been adapted to NGS-based mapping of pseudouridine by a number of groups.⁴⁵⁻⁴⁷ Among the earliest RNA modifications to become accessible to transcriptome-wide mapping was m⁵C. Bisulfite induced chemical deamination of all unmodified cytosines results in uridines, while methylated m⁵C is resistant to deamination.^48,49

Beyond these two classical examples a large variety of methods using enzymes⁵⁰⁻⁵² and chemicals were developed, some of which derive from very early nucleic acid sequencing.⁵³⁻⁵⁵ While we have reviewed and discussed these methods elsewhere,^7,8,56 we will here focus on methods jointly developed in our laboratories and the conclusions drawn from these developments.

Exploiting the Reverse Transcription Signature of Modifications on the Watson–Crick Edge

At the onset of our work in this field, it was clear that some modifications, in particular those with an altered Watson–Crick edge,²⁰ would block reverse transcription. With the efficiency of that process unknown, we set out to characterize in depth the properties of m¹A in RT reactions. m¹A is a noncanonical nucleoside generated both, by dedicated enzymes⁵⁷ and in alkylation reactions used, e.g., in structural probing.²⁰ Various m¹A-containing RNAs, including tRNAs, rRNAs, and some synthetic RNAs, were subjected to RNA-Seq according to a protocol previously developed with an eye to unbiased amplification, thereby specifically including cDNAs resulting from abortive RT events.^1,58

Importantly m¹A-induced RT-stops turned out to not be quantitative. Rather, a significant fraction of read-through events was consistently observed to contain variegated levels of misincorporation of the three dNTPs noncognate to the unmodified adenosine residue, as well as of dTTP. The latter caused a signal identical to that arising from unmodified RNA, thereby effectively erasing information on the presence of m¹A. Moreover, when compiled from 37 tRNA sequences from yeast, an averaged profile only showed about 50% arrest, suggesting a substantial amount of information to be hidden in what we then termed the RT-signature of m¹A. Data mining of the combination of RT-arrests and misincorporation suggested itself for the detection of thus far unknown or unconfirmed m¹A sites. A follow-up with synthetic m¹A-containing oligos showed the RT-signature to be strongly influenced by the identity of the upstream dNTP. Thus, the residue incorporated just prior to the enzyme happening upon the noncanonical nucleoside in the RNA template influences the efficacy of RT-arrest, misincorporation, or read-through, which is plausible, given its spatial proximity to the catalytic center. Further analysis showed a mild contribution even of the penultimately incorporated dNTP. Hence it was clear that the relative contribution to the RT-signature of m¹A by misincorporation of different dNTPs, by RT-arrest, or by jumps depends on the sequence context. In subsequent work, we characterized the influence of the concentrations of Mg²⁺ and Mn²⁺²⁸ and demonstrated that Mn²⁺ could be employed to enhance the RT-signature of m¹A and therefore facilitate its detection. We furthermore found highly variegated propensities for RT-arrest, mismatch incorporation, and jumps within a panel of 13 reverse transcription enzymes.² By conjecture, we expect that other reaction conditions such as temperature, pH, and salt are very likely to also affect RT-signatures, possibly along with the specific activity of a given enzyme preparation.

In conclusion, RT-signatures are highly variable and correspondingly difficult to employ in the detection and quantification of modifications. However, while the heterogeneous contributions of the above features to the overall RT-signature make it difficult to define rigid thresholds for detection, machine learning turned out to be a viable option for exploiting the information content of RT-signatures for mapping purposes. In the case at hand, we annotated known sites of m¹A in RNA-Seq data as positive learning instances, as well as an equal number of bona fide nonmodified adenosines as negative instances. A fraction of about 80% of the RNA-Seq data was used for training a random forest model, which was then used for predicting occurrence of m¹A in the remaining 20% of the data set. After repetition of the procedure with permutations within the same data set, the obtained performance parameters were averaged to provide a combined assessment.^1,58 Foremost among the various statistical parameters that can be used to characterize the quality of a prediction method, machine learning or other, are false positive and false negative rates, elsewhere discussed in terms of sensitivity and specificity. We will discuss some of these numbers further below.

The principal insights obtained using m¹A as a model modification could be recapitulated for other modifications affecting the Watson–Crick edge of purines. Of interest, two consecutive nucleotides of m⁶⁶A modifications produce a RT-pattern similar to the signature of m¹A,⁵⁸ but those have not been reported outside rRNA. We characterized RT-signatures of m¹G and m²²G² and showed that these could be discriminated based on their signature and that the corresponding signatures could be used for meaningful predictions in a transcriptome of limited size, i.e., restricted to tRNAs and rRNA, or ∼10⁴ nt.

Modulating the RT-Signature by Chemical Treatment

In general, strong signatures are known from modifications that are bulky or have impacts on the Watson–Crick edge of the nucleobase. This can be rationalized with an eye on the catalytic center of a RT enzyme. As part of its catalytic cycle, the enzyme tries to fit incoming dNTPs onto the RNA template, thereby observing Watson–Crick base pairing. Bulky modifications are likely to cause steric clashes in the active site, thereby slowing or entirely preventing catalysis. Similarly, the impossibility of establishing a perfect Watson–Crick pair on a modified nucleobase will lead to rejection rather than the incorporation of a dNTP with higher probability, thereby causing longer pauses and potential abortion at modified sites. Indeed, it seems that under suboptimal reaction conditions, stalling might be more likely, at least for certain modifications. Also, as we have outlined above, even the rather blatant block of the Watson–Crick edge in m¹A causes some incorporation of dTTP into the cDNA, effectively erasing modification information. To circumvent such features of an “incomplete” signature, numerous chemical treatments have been developed to enhance detection. Some date back to the time when so-called “direct” RNA sequencing was developed.^53,55

NO-Seq

Among the earliest was deamination by bisulfite, a reagent that strongly differentiates between canonical cytidine and m⁵C and converts the former to a uridine, which can be read-out in sequencing without additional effort, save some adjustment of the mapping algorithm.⁴⁸ We have developed a similar approach for the detection of m⁶A (Figure 3A,B). The chemical reasoning involves deamination of canonical adenosines to inosines, which behave like a guanosine in reverse transcription. In contrast, m⁶A residues are not converted and instead continue to be read out as adenosines. However, the deamination reaction by nitrous acid (Figure 3B) involved conditions that lead to the concomitant deamination of cytosines to uridines and also of some of the guanosines to xanthosine (Figure 3C). The deaminated RNA template leads to cDNA synthesis and subsequently obtained reads which consist of mostly dA and dC signals, which originate from uridines, inosines, and unreacted guanosines, respectively (Figure 3D). The effect of riboxanthosine on reverse transcription remains ill understood. Our own studies showed that xanthosines in deaminated RNA templates were associated with RT-arrest,⁵⁹ while synthetic xanthosines in oligonucleotides obtained by phosphoramidite chemistry did not.⁶⁰ Comprehensive deamination has several effects on the information contained in the resulting cDNA sequencing. Given that all information on cytidines has been effectively erased at this point and that dT indicates m⁶A sites, the remaining sequence elements are composed of only two nucleotides, namely, dA, incorporated at former C and U sites, and dC, incorporated at inosine, i.e., at former A sites.

Figure 3.

Main features of the NO-Seq protocol for detection of m⁶A. Modified nucleotides (A), reagents used (B), and (C) chemical rationale. m⁶A reacts with the reagent but still produces a T signal after RT. Other deamination products are formed by the reagent. (D) Scheme depicting changes in base composition and sequence upon deamination, RT, and PCR steps. (E) Substitution matrix used by custom alignment. For example, A or G in the sequencing read may correspond to A in the initial RNA. (F) Adenosine deamination is incomplete to warrant RT and PCR amplification. (G) Library preparation protocol consisting of targeted amplification of the region of interest. The signal in NO-Seq is visible as an A residue resistant to deamination.

Among the consequences is a decrease in uniquely mapped reads as the information content is effectively reduced to one-fourth of the original RNA sequence. Mapping algorithms needed to be adapted to account for the lower number of nucleotides that carry information and ensuing ambiguities (Figure 3E). At the same time, complete deamination with nitrous acid reduced the yield of reverse transcription and ultimately of reads in Illumina sequencing to a point where it became impractical for analysis. We therefore resorted to partial deamination (Figure 3F), which significantly improved read numbers and mapping ambiguity, albeit at the expense of sensitivity. Testing the method with a range of synthetic oligonucleotides containing m⁶A in varying stoichiometry at a defined site, it was possible to quantify sites containing down to 30% m⁶A. Overall, the method gives many false positive hits to be used for transcriptome-wide applications and therefore is restricted to analysis of short amplicons (Figure 3G) in the size of 10² nucleotides.⁵⁹ Meanwhile, the introduction of glyoxal as a protection reagent has opened up new perspectives. Glyoxal reversibly masks the exocyclic amine functions in cytidines and guanosines and thus protects them from deamination by nitrous acid. Mild acid hydrolysis after deamination restores the original nucleobases and their information content and thereby has allowed transcriptome-wide mapping applications.⁶¹

The Detection Principle of AlkAniline-Seq Is Independent of RT-Signature

In fundamental contrast to the methods described above, reverse transcription signatures occur only as a byproduct in the Alkaline-Aniline-Sequencing method (AlkAniline-Seq). The central feature of this method is the generation of 5′-phosphates downstream of certain modifications. Because these serve as exclusive entry sites to library preparation, AlkAniline-Seq exhibits significantly lower experimental noise and bias.

The original development of the AlkAniline-Seq protocol was intended for mapping of 7-methylguanosine (m⁷G) residues in RNA (Figure 4A) and was based on a concept that goes back to the characterization of the particular chemical properties of m⁷G several decades ago. Reduction by NaBH₄ under alkaline conditions was demonstrated to promote cleavage of the N-glycosidic bond, resulting in an RNA abasic site.^62,63 Of note, a naked RNA abasic site lacking a nucleobase also shows a distinct RT signature.⁶⁴ The abasic site also features an accessible aldehyde group as part of an equilibrium and can thus undergo condensation with a reactive amine (Figure 4B), typically aniline, to yield a Schiff base. Under alkaline conditions, this in turn leads to cleavage of the RNA chain via β-elimination.^53,55,65 Of note, the RNA fragment 3′-downstream of the elimination site now contains a 5′-phosphate (Figure 4C). Analysis of the cleavage sites by gel electrophoresis or sequencing was used in several perspectives. For example, this reaction was used in early direct (chemical) RNA sequencing protocols, for specific cleavage of RNA at all G residues, after modification of G to m⁷G by chemical methylation with dimethylsulfate (DMS),⁵³ or for DMS-mediated probing of 3D RNA structure.^55,66 Barring the DMS methylation step, the same reaction can also be used for mapping of naturally modified m⁷G residues by RT primer extension.^62,63 Detection of m⁷G residues in RNA can be achieved in approaches using various combinations of the above elements:

• (1)Generation of m⁷G-derived RNA abasic site by NaBH₄, followed by aniline cleavage of the RNA abasic and subsequent detection of RT arrests at the position of cleavage. This approach was used in classical protocols for m⁷G analysis by primer extension.^62,63
• (2)Similar to the previous approach, m⁷G-derived RNA abasic site was generated by treatment with NaBH₄, but the aniline cleavage step was omitted. Instead, RT signatures generated at these positions were analyzed in an RNA-Seq context.^64,67
• (3)Covalent labeling of m⁷G-derived RNA abasic sites with an Aldehyde Reactive Probe (ARP)-like reagent was performed, followed by detection of RT-stops at the bulky ARP structure. In a variation, the ARP contained a biotinylated moiety, enabling prior enrichment of RNA abasic sites. This approach was implemented in high-throughput versions.^68,69
• (4)As mentioned, the 5′-phosphate released specifically upon cleavage of the m⁷G-derived RNA abasic site can be targeted for specific ligation of RNA-Seq adapters. This results in a specific enrichment of modification-derived fragments and is implemented in the AlkAniline-Seq protocol³ as well in TRAC-Seq.⁷⁰

During the development of what eventually became the AlkAniline-Seq protocol (Figure 4D), we compared approaches 1 and 4 described above in an RNA-Seq coupled setting. Indeed, a specific signal corresponding to m⁷G1575 in Saccharomyces cerevisiae 18S rRNA was detected based on RT arrest at the cleavage site according to the “classical” approach 1. However, the high background signal rendered detection of m⁷G residues elsewhere in the transcriptome rather questionable. In contrast, ligation of 5′-adapters to the specifically released 5′-phosphate depressed background and correspondingly increased the signal-to-noise ratio. Omission of NaBH₄, in what were meant to be control reactions, still showed chain scission upon subsequent alkaline RNA fragmentation for library preparation at pH 9.3 and 95 °C.⁷¹ Omission of NaBH₄ avoided side reactions including the reduction of s⁴U,⁷² and further depressed noise. Interestingly the replacement of alkaline hydrolysis by Mg²⁺-driven RNA cleavage at neutral pH produced only very low signals at the m⁷G sites, identifying alkaline conditions as essential.

Figure 4.

Main features of AlkAniline-Seq protocol for detection of m⁷G, m³C, D, and ho⁵C. Modified nucleotides detected (A), reagents used (B), and chemical rationale of the method (C). Decomposition of the RNA abasic site generated 5′-phosphate at the N + 1 nucleotide in the sequence. (D) library preparation protocol. Phosphatase removes all 3′- and 5′-phosphates. (E) Nature of the signal and (F) scoring system for AlkAniline-Seq signals.

Aniline-induced chain scission proceeds via β-elimination, directly forming a 5′-phosphate on the downstream fragment, and indirectly leading to a 3′-phosphate on the upstream fragment. The latter is of little relevance, but the former constitutes a specific molecular entry into library preparation (Figure 4E,F). Therefore, treatment with phosphatase again improved the S/N ratio and in particular the specificity by removing all pre-existing 5′-phosphates resulting from biological RNA processing or degradation events in vivo or in vitro.

Application of this AlkAniline-Seq protocol to bacterial and eukaryotic rRNAs faithfully detected known and anticipated m⁷G sites^3,73,74 with exquisite S/N values. Moreover, inspection of AlkAniline-Seq profiles for E. coli rRNA revealed two additional signals corresponding to modified dihydrouridine (D2449) and 5-hydroxycytosine (ho⁵C2501)⁷⁵ with somewhat lower amplitude. Further signals for D were found in bacterial and eukaryotic tRNAs. A plausible molecular basis for these results is a ring opening of the dihydrouracil base under alkaline conditions to give β-ureidopropionic acid.⁷⁶ The corresponding ring-opened nucleotide derivative is compatible with a β-elimination upon aniline treatment.

Another unanticipated finding consisted of strong signals at all known 3-methylcytosine (m³C) residues in eukaryotic cytoplasmic and mitochondrial tRNAs. While we found m³C to deaminate to m³U under alkaline conditions; m³U sites in rRNA did not show signals. Thus, m³C in RNA must induce chain scission before deamination, apparently activating its downstream phosphate for scission only in the context of the RNA chain.

In summary, AlkAniline-Seq detects at least m⁷G, m³C, D, and ho⁵C (Figure 4A), but our unpublished data indicate further signals presumably arising from noncanonical nucleobases of nonenzymatic origin or from RNA abasic sites. The latter might, in turn, arise from spontaneous loss of RNA bases, including damaged ones. In vivo, RNA damage is thought to mostly result from reactive oxygen species (ROS), either directly leading to oxidized bases or through ROS-mediated lipid oxidation, the latter leading to unspecific alkylation⁷⁷ plausibly including m⁷G and m³C of nonenzymatic origin. Indeed, most of the noise in AlkAniline-Seq originates from guanosines, making 8-oxo-guanosine and its downstream ring-opened oxidation products⁷⁸ prime suspects. The critical dependence of AlkAniline-Seq on unique 5′-phosphates substantially complicates precise and accurate quantification, since calibration curves between AlkAniline-Seq signals and modification content are not linear for all used scoring systems.^73,74,79 Its very low background and exceptional S/N ratio qualify AlkAniline-Seq analyses for full mRNA transcriptomes with a complexity of 10⁶–10⁷ nt.

In the backdrop of large numbers of mapping techniques appearing in the literature, AlkAniline-Seq protocol was successfully validated with known sites in rRNA and tRNA from different species.^73,80,81 However, our analysis did not detect any plausible modification signal for m⁷G, m³C, D, or ho⁵C in any of the studied eukaryotic mRNAs (S. cerevisiae, Arabidopsis thaliana, Mus musculus, Homo sapiens). These observations corroborate the results reported by others on m⁷G in E. coli and S. cerevisiae mRNAs.⁶⁴

HydraPsiSeq

The interest in techniques for mapping and quantification of pseudouridine (ψ) as one of the most widespread and abundant RNA modifications is high. Pseudouridine (Figure 5A) and its derivatives are the only known natural nucleotides with a C–C glycosidic bond. In addition, the isomer of uridine features an additional N–H group, and these features provide particular physicochemical properties and reactivity. The most prominent of several reactions that distinguish between U and ψ is stable adduct formation with CMCT. Resulting methods (cited above) are relatively sensitive but feature a high RT-stop background, which significantly depresses the S/N signal, requiring appropriate spike-in calibration for adequate quantification.⁸² As another differential reactivity, ψ residues are insensitive to hydrazine (Figure 5B,D) even at high concentrations, while unmodified U undergoes a Michael addition at C6, with subsequent destruction of the aromatic pyrimidine base. The resulting damaged site, upon treatment with aniline, leads to chain cleavage via β-elimination (Figure 5B). Consequently, detection of ψ and other hydrazine-insensitive residues can be achieved by measuring noncleavage at presumed uridine sites which are protected by their modification chemistry (Figure 5C). Quantification of such negative signals, if RT-based, is subject to the various sources of noise mentioned above. Despite this, hydrazine hydrolysis was occasionally employed as an auxiliary method for detection of ψ (and also m⁵U) in rRNAs, brome mosaic virus (BMV) viral RNA, and snRNAs.^41,83,84

Figure 5.

Detection and quantification of pseudouridines and m⁵U by HydraPsiSeq. Modified nucleotides detected (A), reagents used (B), and chemical rationale (C). Some modified U nucleotides are resistant to cleavage, while some derivatives of G and C are extensively cleaved (panel A). (D) Decomposition of the RNA abasic site after U cleavage by hydrazine generates 5′-phosphate at the N + 1 nucleotide in the sequence. (E) Library preparation protocol. Removal of only 3′-P residues in RNA is achieved by T4 PNK phosphatase activity. (F) Nature of the signal and (G) scoring system for the HydraPsiSeq signals.

For an adaptation to NGS, we identified two options for the exploitation of random hydrazine cleavage. Option 1, as in the preceding section, leverages RT-stops after primer extension of aniline-cleaved RNA. Option 2 is more specific and conceptually related to AlkAniline-Seq. A graphical overview of HydraPsiSeq^4,85 in Figure 5E shows hydrazine-induced cleavage being followed by extensive T4 polynucleotide kinase (T4 PNK) treatment in order to selectively remove the resulting 3′-phosphates without affecting 5′-phosphate in the process. Consequently, the resulting RNA fragments, which are randomly cleaved at all U residues, exhibit 3′-OH extremities and ligation-competent 5′-phosphate used for adapter ligation and conversion into a sequencing library with simultaneous barcoding. Coverage in the order of 1000 reads per position is required, making the method practical almost exclusively for highly abundant RNAs.

The data analysis pipeline includes adapter trimming, alignment to RNA reference, and tallying the 5′-extremities of the reads for each position in the reference (Figure 5F). In a scoring system named NormUcleavage (Figure 5G), raw data are normalized to the local background of A/C/G cleavage in a rolling 11 nt window. Tallied values are used for constructing uridine profiles and subsequent calculation of a protection score at all U-sites. Because of the latter, data analysis is conceptually similar to the popular RiboMethSeq protocol, which identifies cleavage-protected Nm residues,^86,87 and consequently the same scores are used.

Apart from ψ, 5-substituted uridines, such as m⁵U and some hypermodified uridines, typically found at the wobble position of tRNA, were also found to be resistant to cleavage (Figure 5A). Weak signals at D sites suggest at least partial protection, which might also be related to substoichiometrically modified D sites. Unmodified A, C, and G nucleotides are inert; however, m⁷G and the bacterial wobble modification lysidine (k²C) (Figure 5A) give strong signals.⁸⁵ Interestingly no signals were apparent for m³C, even though this nucleotide was reported to be sensitive to hydrazine,⁸⁸ an early finding that was exploited in so-called “direct” RNA sequencing and structural probing.^53,55 Cleavage conditions substantially differ in hydrazine and salt concentrations between m³C cleavage (10% hydrazine, 3 M NaCl) and HydraPsiSeq (50% hydrazine).

Apart from ψ quantification in noncoding RNA such as rRNA, HydraPsiSeq can be potentially applied for confirmation of candidate ψ residues in coding or other low abundant RNAs;⁸⁹ however, substantial sequencing efforts are required to achieve sufficient coverage. In practice, HydraPsiSeq turned out to be rather sensitive to the quality of chemical reagents, in particular, of hydrazine, requiring some testing to account for batch-to-batch variation. Since HydraPsiSeq is based on detection of “negative” signals, a substantial number of false positive hits may be expected, related to irregularities in RNA cleavage by hydrazine and to ligation biases during the library preparation step. The S/N ratio for HydraPsiSeq is compatible with rRNA/tRNA pool transcriptome size, thus on the order of 10⁴–10⁵ nt.

Error Range and Statistics Should Determine the Application Range

After a given method has run its course in the wet lab (Table 1), data treatment takes over. There is vanishingly little common ground among the methods used for identifying modification sites from Illumina data, to the point that comparative analysis suggests that the influence of data analysis pipelines on the results may dominate over that of the underlying biology.⁹⁰ The developments in RNA modification mapping of the past decade have clearly shown a tendency to overestimate the number of modifications detected,⁹¹ and hence a need to depress the number of false positives, i.e., increase specificity, if necessary at the cost of increasing false negative signals, equivalent to accepting a loss in sensitivity. As the field has seen several open controversies over mapping issues, we suggest that upon publication of a new mapping method, one should state the size of the epitranscriptome (“limited epitranscriptome”) within which the method produces reasonable and verifiable predictions. Thus, rather than stating the number of newly identified modification sites in a full cellular transcriptome, one might state, if the method is rather suited to investigating the ensemble of tRNAs and rRNAs, thus a “limited transcriptome”² of about 10⁴–10⁵ nt in size. From our decade of joint experience, this applies to the majority of methods, whereas only a few are suitable to tackle mRNAs, which invoke an epitranscriptome on the order of 10⁷ nt. Still from our own experience, AlkAniline-Seq is one of very few methods that are up to this task, since this method has a very high S/N ratio and is highly sensitive even for low substoichiometric modification rate. Its particularly high S/N ratio and corresponding sensitivity result, as explained, from its peculiar chemistry that provides privileged entry into sequencing libraries for fragments carrying pertinent information.

Table 1. Comparative Characteristics of the Protocols Discussed in the Review.

Parameters	Protocol
	RT-Seq	NO-Seq	AlkAniline-Seq	HydraPsiSeq
Detected modification	m1A, m1G, m22G, m3U, m3C, m66A, other residues modified at the WC face	m6A	m7G, m3C, D, ho5C/ho5U, RNA abasic sites	Ψ, m5U, 5-modified U*34, k2C (positive cleavage signal), m7G (positive cleavage signal)
Nature of the signal	RT stop + misincorporation/jump signature	Resistance of m6A to deamination compared to unmodified A	Cleavage at the abasic RNA site resulting from decomposition of the modified residue	Resistance of Ψ, m5U, 5-modified U*34 residues to hydrazine cleavage, over cleaved U residues
Detection type	positive	negative	positive	negative
Input RNA required	>100 ng	∼ 500 ng	>50 ng	>150 ng
Molar sensitivity	average	low	very high	average
Quantification	Possible, but complicated	Precise	Possible, but calibration is not linear	Precise
Reagents used	None (different RT enzymes recommended)	NaNO2/H+	Alkaline hydrolysis/Aniline	Hydrazine/Aniline
Library preparation step	Custom protocol	Custom protocol	NEB Small RNA kit	NEB Small RNA kit
Required sequencing depths	Coverage >100 reads/position, better >500	Coverage >50 reads/position	>50 reads/RNA position	>1000 reads/RNA position
Recommended size of ≪ limited transcriptome ≫	10 000 bases rRNA/tRNA size	100 bases Custom amplicons	10 000 000 bases Transcriptome-wide	10 000 bases rRNA/tRNA size
Bioinformatics analysis	straightforward	Complicated (custom alignment algorithm)	straightforward	Intermediate (normalization and score calculations)

Biographies

Yuri Motorin studied chemistry and chemical enzymology at Lomonosov Moscow University, USSR, and obtained his Ph.D. in Biochemistry in 1989, at the A.N. Bakh Institute of Biochemistry in Moscow. After a postdoctoral period in Gif-sur-Yvette, France, he established his group at the University of Lorraine in Nancy. His research aims to develop analytical methodology for RNA modification analysis by combination of specific chemical treatment and deep sequencing.

Mark Helm studied chemistry in Würzburg, Germany, and obtained his Ph.D. in Molecular Biology in Strasbourg, France, in 1999. After postdoctoral fellowships at Caltech, USA, and the Free University of Berlin, Germany, he started his independent research group in 2002 in Heidelberg Germany. Since 2009, he is Professor of Pharmaceutical Medicinal Chemistry in Mainz, Germany, where he is currently Chair of the Research Consortium “RMaP: RNA Modification and Processing”.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.accounts.3c00529.

• Comparative characteristics of the protocols discussed (PDF)

Author Contributions

CRediT: Yuri Motorin conceptualization, writing-original draft; Mark Helm conceptualization, writing-original draft.

The authors declare no competing financial interest.

Special Issue

Published as part of the Accounts of Chemical Research special issue “RNA Modifications”.