Reading between the bases: technological advances in RNA modification detection

Introduction

Chemical modifications of RNA constitute a dynamic regulatory layer of post-transcriptional regulation of the transcriptome, analogous to the epigenome, that fine-tunes RNA function, processing, stability, translation, across all subtypes of cellular coding and non-coding RNA, with profound implications for both physiological and pathological cellular homeostasis (Roundtree et al., 2017; Delaunay et al., 2024). Specific enzymes known as epitranscriptomic writers such as METTL3 (m⁶A), NSUN2 (m⁵C), PUS enzymes (Ψ), TRMT61 (m¹A), METTL1 (m7G), FBL (2’-O-methylation) and ADAR deaminases (A-to-I editing) install these marks post transcriptionally on RNA, often in cooperation with non catalytic subunits required for efficient target identification and enhancing enzymatic activity. In turn, RNA modifications can directly affect RNA function by shaping RNA primary and secondary structures or exert their effect indirectly through reader proteins that dictate RNA fate. The effects of all these post-transcriptional modifications on cellular RNAs establish a complex epitranscriptomic code required for cellular homeostasis and whose dysregulation correlates strongly with diseases including cancer, neurodegeneration, and developmental disorders (Barbieri and Kouzarides, 2020; Delaunay et al., 2024).

The majority of RNA modifications were discovered more than 5 decades ago through low-throughput methods such as thin-layer chromatography and mass spectrometry (Davis and Allen, 1957; Kowalak et al., 1993). These techniques, while precise, offered only limited insights on the functions of modifications, proving inadequate for comprehensive transcriptome-wide analysis (Yoluç et al., 2021).

Several decades later, the advent of next-generation sequencing (NGS) allowed transcriptome-scale epitranscriptomic mapping through complementary strategies. Firstly, thanks to the availability of modification-specific antibodies, antibody-based immunoprecipitation (MeRIP-seq) was developed to specifically enrich modified RNAs followed by next generation sequencing. This approach was used to generate the first transcriptome-wide mapping of m6A (Dominissini et al. 2012). In parallel, techniques coupling chemical derivatization approaches (bisulfite RNA-seq, Pseudo-seq), and retrotranscription signature analysis were developed to identify specific modifications such as m1A, pseudouridine and m5C (Dominissini et al., 2012; Meyer et al., 2012; Squires et al., 2012; Harcourt et al., 2013; Carlile et al., 2014). Together, these methods generated foundational epitranscriptomic maps revealing characteristic modification patterns within specific RNA subtypes and context-dependent dynamics. However, these advances were constrained by fundamental limitations such as antibody specificity, indirect readout biases, reverse transcriptase artifacts, and resolution constraints, which collectively increased false discovery rates and compromised accurate stoichiometric quantification, thereby limiting the depth of biological inference ( Schwartz and Motorin, 2017; Grozhik and Jaffrey, 2018 ). To partially overcome these problems, specific controls, such as enzyme depletion and the use of orthogonal approaches allowed the development of high confidence epitranscriptomic datasets. Subsequently the development of hybrid approaches such as miCLIP allowed to increase specificity and resolutions. Third-generation nanopore sequencing from Oxford Nanopore Technologies (ONT) has since disrupted this paradigm by enabling direct sequencing of native RNA without reverse transcription or amplification (Garalde et al., 2018; Workman et al., 2019; Liu H. et al., 2019). By exploiting modification-induced perturbations in ionic current signals, decoded through machine learning frameworks such as xPore, EpiNano, and CHEUI, nanopore sequencing promises highly quantitative detection of RNA modification on single-molecule at high resolution with inherent multimodality (Pratanwanich et al., 2021; Liu X. M. et al., 2019; Acera Mateos et al., 2024). Nanopore direct RNA sequencing basecaller now includes the identification of 5 modifications, namely m6A, inosine, m5C, pseudouridine and 2′O-ribose-methylation. Nevertheless, the limited signal strength of modification-induced signals and the substantial computational demands of current analytical pipelines indicate that nanopore sequencing, when used in isolation, represents an incremental rather than transformative advance, whose full potential is best realized through integration with orthogonal chemical and genetic validation strategies (Meyer, 2019; Jain et al. 2022). This mini-review moves beyond modification-centric narratives to adopt a technique-focused framework, critically evaluating the core strengths, practical limitations, and applications of current approaches across m⁶A, m⁵C, Ψ, m¹A, m7G, 2’-O-Me and A-to-I editing, while emphasizing how methodological integration can offset individual limitations (Yoluç et al., 2021). Through a rigorous dissection of key trade-offs, we argue that hybrid approaches represent a methodological necessity rather than a convenience, as exemplified by nanopore–MS workflows that markedly reduce false discovery rates while enabling resolution of co-occurrence patterns indicative of combinatorial epitranscriptomic codes (Shaw et al., 2024). Despite persistent challenges in measuring low-stoichiometry events, detecting rare modifications, and establishing causal relationships, this analysis outlines feasible paths toward comprehensive epitranscriptomic atlases capable of decoding RNA’s chemical grammar, thereby laying essential groundwork for therapeutic exploitation.

Immunoprecipitation-based sequencing

Specific RNA immunoprecipitation followed by NGS RNA sequencing emerged as a leading approach in high-throughput epitranscriptomics, leveraging the increased availability of modification-specific polyclonal and monoclonal antibodies (Figure 1A). These tools were used to selectively enrich RNA fragments harboring specific modifications prior to next-generation short-read sequencing. This methodological advance transformed the field from hypothesis-driven biochemistry to transcriptome-scale discovery (Dominissini et al., 2012; Meyer et al., 2012). By fragmenting RNA into 100–200 nucleotide segments, IP-seq produces broad coverage maps that capture modification landscapes ( Grozhik and Jaffrey, 2018; Helm and Motorin, 2017 ).

For N⁶-methyladenosine (m⁶A), MeRIP-seq and m⁶A-seq established foundational maps documenting 12,000–20,000 peaks predominantly near stop codons and 3′UTRs, implicating the mark in both RNA stability and translation regulation via YTH readers (Dominissini et al., 2012; Meyer et al., 2012; Zaccara et al., 2019). However, this design inherently limits resolution to ∼100 nucleotides, obscuring precise site assignments. To overcome this limitation near nucleotide resolution was achieved through the development of techniques, such as m⁶A-CLIP and miCLIP, incorporating UV crosslinking to capture antibody–RNA adducts as truncation or retrotranscription mutation signatures (Ke et al., 2015; Linder et al., 2015).

Similarly, 5-methylcytosine was mapped on the transcriptome using an m5C-RIP-seq approach, revealing widespread but unevenly distributed candidate sites, often occurring in cytosine clusters that likely reflect conversion artifacts rather than true methylation events. Spike-in and immunoprecipitation controls demonstrated that bona fide m5C sites show stronger enrichment compared to unmethylated controls, supporting a more limited and context-dependent m5C landscape and implicating NSUN2 in mRNA methylation (Huang et al., 2019). A pseudouridine-specific antibody was used to map this modification on the transcriptome with limited success, as antibody-based pseudo-RIP suffers from high background, limited reproducibility, and poor single-nucleotide resolution and stoichiometric quantification, particularly at densely modified sites. Successively, hybrid approaches such as photo-crosslinking-assisted Ψ sequencing (PA-Ψ-seq) improved specificity by combining antibody recognition with UV crosslinking, enabling transcriptome-wide mapping of Ψ in both cellular and viral RNAs and linking specific sites to PUS enzymes (Martinez Campos et al., 2021). Integration with eCLIP-seq datasets further refined these maps by correlating pseudouridine sites with RNA-binding protein interactions, enhancing functional interpretation (Martinez et al., 2022). N¹-methyladenosine (m¹A) RIP initially reported widespread levels of this modification on cellular mRNA, specifically within the 5’UTR region of coding mRNAs (Dominissini et al., 2016). These findings were later challenged using structure-disrupting controls (Grozhik et al., 2019) and the presence of m1A on cellular RNA remains controversial. m7G-MeRIP-Seq (Zhang et al., 2019; Pandolfini et al., 2019) was used to identify this modification on tRNA, mRNA and miRNA showing high specificity but limited resolution (∼100 bp). Successively m7G-miCLIP-Seq (Malbec et al., 2019) improved resolution (∼30 bp) by combining antibody enrichment with UV crosslinking, with potential for base-resolution when coupled with motif analysis (Huang et al., 2026).

Despite these advances, IP-seq faces notable limitations: poor reproducibility arising from antibody batch variability, sequence biases, and low resolution. Even CLIP variants, despite increased specificity and resolution, still suffer from UV-induced artifacts, and an inability to accurately quantify low-stoichiometry sites or capture dynamic changes in specific contexts such as cancer(Helm and Motorin, 2017; Wiener and Schwartz, 2021). For these reasons, for many types of modifications RIP-sequencing approaches alone can’t generate high confidence and quantitative transcriptomic mapping of RNA modification. Specific controls such as genetic depletion of RNA writers allows for significant improvement on specificity and sensitivity. Additionally, competition Immunoprecipitation, such as the m6Aseq2 allows for a more quantitative analysis of modifications (Dierks et al., 2021). Moreover, early m¹A datasets were later shown to overestimate the prevalence of this modification, underscoring significant specificity limitations and the need for rigorous validation (Grozhik et al., 2019). This dependence on conventional methods risks hindering the advancement of epitranscriptomics, emphasizing the promise of antibody-free approaches.

Chemical and enzymatic probing

Chemical probing and RT signature methods are complementary antibody-free strategies that can potentially achieve base-resolution precision beyond IP-seq and overcome many shortcomings of antibody-based techniques (Figure 1B). Chemical probing techniques take advantage of modification-specific chemical reactivity to generate secondary marks ultimately detectable through retrotranscription termination, cleavage events, or mutagenesis. These techniques enable high resolution, stoichiometric quantification of epitranscriptomic marks but their reliance on harsh chemistries, context-dependent enzymes and RNA secondary structures can introduce degradation artifacts and false-positive results (Kong et al., 2023).

For m⁶A, MAZTER-seq exploits MazF endonuclease’s inability to cleave methylated ACA motifs, generating fragmentation patterns that precisely map ∼7,000 sites across species with high motif specificity (Garcia-Campos et al., 2019). DART-seq fuses the m⁶A-binding YTH reader domain to ADAR deaminase, enzymatically marking sites via A-to-I edits (20–30% efficiency, sequenced as A-to-G), bypassing antibodies while requiring computational bias correction (Meyer, 2019). For m⁵C, minimal direct polymerase perturbation makes direct RT signatures unreliable, necessitating bisulfite pre-treatment as a key methodological requirement (Fleming et al., 2023). RNA m⁵C sites are protected from bisulfite-mediated C-to-U conversion. Bisulfite RNA sequencing identifies ∼1,000 high-confidence m⁵C mRNA sites via ultrafast protocols that minimize RNA hydrolysis to below 50%, enabling reliable library construction while preserving modification fidelity (Dai et al., 2024), yet structured regions hinder complete reactivity (Johnson et al., 2022). Ψ detection techniques exploit the reactivity of N-cyclohexyl-N′-(2-morpholinoethyl)carbodiimide metho-p-toluenesulfonate (CMCT) toward the N3 position of pseudouridine, forming RT-blocking adducts that are mapped as RT termination points or mutation signatures by Pseudo-seq, identifying ∼200–500 dynamic sites within eukaryotic mRNAs (Carlile et al., 2014; Schwartz et al., 2014). CeU-seq successively combined CMCT treatment with biotin-streptavidin precipitation to improve the sensitivity of Pseudo-seq up to 10-fold (Li et al., 2015). HydraPsiSeq employs a quantitative approach based on specific protection of Ψ residues from hydrazine-induced cleavage at uridine residues: since pseudouridines resist hydrazine-mediated ring opening, their positions appear as negative cleavage signals relative to flanking unmodified uridines. Normalization to in vitro-transcribed unmodified controls enables absolute stoichiometric quantification of Ψ levels at each site (Marchand et al., 2020). Overall, these methods are chemically precise but prone to RT artifacts due to non-Ψ structures that inflate low-occupancy modification calls. A method for quantitative and base-resolution sequencing of pseudouridine termed 2-bromoacrylamide-assisted cyclization sequencing (BACS) was subsequently developed, allowing the accurate quantification of Ψ stoichiometry, precise identification of Ψ positions, and robust detection of densely modified Ψ sites (Xu et al., 2024). Ψ-seq variants capture CMC-induced truncations at ∼90% of known sites (with rRNA depletion), but non-Ψ structural stops necessitate computational filtering (Schwartz et al., 2014; Lovejoy et al., 2014).

Borohydride reduction sequencing (BoRed-seq) has been used to identify m7G sites and several different RNA subtypes. This technique relies on the sensitivity of charged m7G residues to borohydrate reduction at low pH. In turn, this causes the opening of the imidazole ring that can be further exploited for the generation of abasic sites or specific enrichment. BoRed-Seq enabled detection of internal m7G within miRNAs combining borohydride reduction with ARP (aldehyde reactive probe) capture and, together with RNA immunoprecipitation, revealed m7G modifications within a subset of miRNAs implicated in the inhibition of cell migration (Pandolfini et al., 2019). TRAC-seq enables unbiased, single-nucleotide mapping of tRNA m7G: size-selected RNAs (<200 nt) are treated with the demethylase AlkB to remove major tRNA modifications that would otherwise block RT, followed by sodium borohydride (NaBH₄) reduction of m7G to generate dihydro-m7G abasic sites, which are then cleaved by aniline; the resulting cleavage positions are analyzed as cleavage scores across all tRNA positions by high-throughput sequencing (Lin et al., 2019). Bo-seq extended this borohydride reduction strategy to globally map internal m7G at base resolution throughout the transcriptome, identifying m7G in tRNAs and a subset of mRNAs (D’Ambrosi et al., 2024). m7G-MaP-seq detects m7G sites via borohydride-induced abasic site mutations during reverse transcription (Enroth et al., 2019).

For 2′-O-methylation (Nm), Nm-seq exploits the differential sensitivity of methylated versus unmethylated ribose to periodate oxidation, enabling base-resolution mapping and revealing thousands of Nm sites in human mRNA, although the chemical inertness of Nm continues to constrain detection sensitivity (Dai et al., 2017). RiboMethSeq is a complementary approach based on the differential protection of 2′-O-methylated phosphodiester bonds from alkaline hydrolysis: the presence of a 2′-O-methyl group markedly reduces the nucleophilicity of the adjacent ribose, generating a characteristic depletion of RNA fragment 5′-ends at methylated positions that is scored by high-throughput sequencing of randomly fragmented RNA (Krogh et al., 2017). RiboMethSeq has been widely applied to map snoRNA-guided 2′-O-methylation patterns in rRNA and tRNA with single-nucleotide resolution and quantitative stoichiometric output.

The positive charge of m¹A can directly interfere with retrotranscription, triggering RT misincorporations that can be quantified through m¹A-MAP ratios, reliably defining tRNA T-loops despite secondary structure confounds (Tserovski et al., 2016). m¹A produces striking RT signatures (>50% arrests, A-to-G/T misreads), enabling m¹A-seq to identify ∼200 cytoplasmic motifs, later refined to ∼20–50 bona fide sites via structure-disrupting controls, though overlapping charge and folding effects complicate interpretation (Hauenschild et al., 2015; Grozhik et al., 2019). To overcome these limitations, reduction-based m¹A-seq (red-m¹A-seq) leverages NaBH₄ reduction to improve mutation and readthrough rates, while alkaline-catalyzed Dimroth rearrangement converts m¹A to m⁶A as an internal control, enabling higher sensitivity detection of both known and novel m¹A sites (Pajdzik et al., 2024). Despite this, the technique could not reliably be used to identify m¹A on cellular mRNAs, and the presence of m¹A on this RNA subtype remains controversial. A-to-I editing produces A-to-G transitions confirmed by comparison to genomic DNA to exclude polymorphisms, though its dependence on RNA duplex structures limits the range of detectable sites (Bahn et al., 2012; Ramaswami et al., 2013).

Despite these advances, RNA lability, degradation artifacts, and false-positive rates of ∼20–30% from structural artifacts limit standalone chemical probing and RT signatures. Writer-knockout experiments provide the most direct orthogonal control: genetic depletion of an epitranscriptomic writer eliminates its cognate marks, allowing confident discrimination between genuine signals and false positives (Kong et al., 2023; Zhang et al., 2022).

Mass spectrometry approaches

Liquid chromatography-tandem mass spectrometry (LC-MS/MS) provides epitranscriptomics’ gold-standard chemical specificity, hydrolyzing RNA to ribonucleosides separated by HILIC/reverse-phase chromatography, ionized via ESI, and quantified via MRM transitions (e.g., m⁶A: m/z 282→150) achieving femtomolar sensitivity and <5% variance (Jora et al., 2019). Similarly other modifications can be quantified on bulk RNA through MS2 approaches (Wein et al., 2020). This unmatched quantitative accuracy positions LC-MS/MS as the gold standard for absolute stoichiometric quantification of RNA modifications, a critical benchmark that sequencing-based approaches cannot yet replicate (Helm and Motorin, 2017; Yoluç et al., 2021). While nanopore and NGS-based methods yield relative enrichment scores or site-level modification probabilities, LC-MS/MS directly measures molar ratios of modified to unmodified nucleosides, enabling rigorous cross-sample and cross-condition comparisons. This capacity is particularly valuable for calibrating the output of sequencing-based callers and for detecting stoichiometric shifts in response to genetic or pharmacological perturbations of epitranscriptomic writers. This limitation prevents detection of context-dependent RNA modification dynamics, which vary across transcripts, cell states, and environmental conditions. Indeed, studies have shown that identical modifications (e.g., m⁶A) can exert distinct regulatory effects depending on context, underscoring the need for transcript-resolved approaches to capture functional epitranscriptomic regulation (Pandolfini et al., 2019). Beyond global quantification, LC-MS/MS has been applied in hybrid workflows to resolve context-dependent modification stoichiometry: for example, Pandolfini, Barbieri et al. employed refined mass spectrometry methods to map m7G to a single guanosine within the let-7e-5p miRNA, demonstrating that METTL1-dependent m7G methylation is sequence-context-dependent and influences miRNA secondary structure and processing in a position-specific manner (Pandolfini et al., 2019). Such hybrid LC-MS/MS–sequencing workflows powerfully anchor chemical identity and stoichiometry to positional information derived from sequencing, compensating for the bulk-averaging limitation of MS alone (Shaw et al., 2024; Helm and Motorin, 2017).

Despite its quantitative precision, critics highlight its positional blindness, high sample demands, and hydrolysis artifacts that undermine low-input or dynamic applications, persistent limitations even in hybrid workflows where methodological discord can amplify uncertainty (Zhang et al., 2022). These constraints affirm its role as a confirmatory anchor rather than a discovery driver, urging standardized multi-modal benchmarks to mature epitranscriptomic rigor.

Nanopore direct RNA sequencing

Oxford Nanopore long read sequencing represents a new approach to nucleic acid sequencing and the first commercially available technology to allow direct sequencing of native DNA and RNA molecules. This technology represents a step forward to dissect both genomic and transcriptomic features of eukaryotic cells at single molecule level. On top of this, Nanopore direct RNA sequencing (dRNA-seq) represents epitranscriptomics’ boldest recent advance. Initial attempts of direct RNA sequencing reported low basecalling efficiency when compared to direct DNA sequencing due to RNA secondary structures and post-transcriptional chemical modifications. More recently these perturbations of direct RNA sequencing were converted from a bug of RNA basecaller to a feature to specifically identify a number of different RNA modifications. These advancements were possible through the development of machine learning classifiers (xPore, EpiNano, nanoDoc, CHEUI) capable of calling RNA modifications at single-molecule stoichiometry and multi-mark co-occurrence, preserving native context without the need of reverse transcription or other chemical manipulation of RNA (Workman et al., 2019; Wang et al., 2021) (Figure 1C). These computational tools were generated through machine learning approaches where the system was firstly trained to discriminate modified vs. unmodified synthetic RNAs. Next, the basecallers were trained on cellular RNAs derived from wild-type versus knockout cells. These models have been shown the capability to assign modification probabilities at single base resolution from raw direct RNA Nanopore sequencing signals (Leger et al., 2021; Pauli et al., 2025); however, despite achieving recall rates as high as ∼0.92 and strong stoichiometry correlations, per-site false discovery rates remain substantially elevated, reaching ∼40% for m6A and ∼95% for pseudouridine in Dorado, and up to ∼80% for m6A in m6Anet, driven primarily by the low transcriptome-wide prevalence of these modifications rather than sequencing error alone (Zou et al., 2025). A key challenge has been the simultaneous identification of multiple modification types from a single sequencing run. TandemMod addressed this by implementing a transfer learning framework that leverages high-quality m⁶A training data to bootstrap classifiers for m⁵C, m7G, Ψ, and inosine with minimal additional synthetic data (Wu et al., 2024). The m6Aiso method further pushed single-molecule resolution by coupling endogenous DART-seq labelling, in which APOBEC1-YTH fusions introduce C-to-U mutations proximal to m⁶A sites, with nanopore DRS reads, enabling isoform-specific m⁶A quantification and revealing that the same genomic m⁶A site can carry distinct modification stoichiometries on different mRNA isoforms (Guo et al., 2025). Signal alignment accuracy, a limiting factor for all current-based modification callers, was substantially improved by Uncalled4, which provides fast and exact mapping of raw nanopore signals to reference sequences and reduces false-positive calls arising from signal drift (Kovaka et al., 2025). At the multi-mark classification level, ModiDeC demonstrated that m⁶A, inosine, Ψ, 2′-O-methylguanosine, and m¹A can be distinguished simultaneously on a single DRS run using a deep learning classifier trained on synthetic splinted-ligation RNA constructs validated in human cell lines and blood samples (Alagna et al., 2025). Most recently, DirectRM integrated site-level detection with co-occurrence analysis, enabling transcriptome-wide mapping of crosstalk between multiple modification types from a single native RNA dataset and bypassing the IVT-derived artefacts that distort model training when fully substituted synthetic constructs are used (Zhang et al., 2025).

Each modification presents its own detection landscape, with strengths and limitations that reflect the chemical diversity of the epitranscriptome. m⁶A most reliably disrupts translocation, with xPore achieving ∼90% precision/recall via METTL3-knockout benchmarking to estimate 30–70% occupancies; however, chemistry-dependent signal drift across flowcells limits reproducibility across runs and platforms (Pratanwanich et al., 2021; Beslic et al., 2024). m⁵C poses a different challenge, as its ionic signal broadly overlaps with unmodified cytosine noise; CHEUI’s joint modeling partially overcomes this by detecting ∼1,000 co-modified transcripts, yet isolated sites continue to evade confident calls (Acera Mateos et al., 2024). Pseudouridine (ψ) detection via basecalling discrepancies and raw ionic current signals suffers from strong sequence context dependency, such that inferred modification occupancies lack quantitative accuracy and site-level confidence remains variable (Makhamreh et al., 2024). m¹A introduces a further complication, as its positively charged perturbations spectrally mimic m⁶A, necessitating anomaly detection approaches such as IL-AD for de novo identification, a strategy that is innovative but remains error-prone in practice (Wang et al., 2024). Finally, inosine, generated by ADAR-mediated A-to-I editing, has traditionally been inferred from A-to-G mismatches in Illumina RNA-seq data, an approach that is unreliable in short exons, poorly annotated genomes, or mutation-rich disease contexts; nanopore direct RNA sequencing offers a promising alternative, with convolutional neural networks applied to distinguish inosine from adenosine and guanosine and to estimate editing levels at individual sites with improved accuracy (Nguyen et al., 2022).

Using nanopore direct RNA sequencing coupled with machine learning, NanoNm was developed as a predictive model trained on rRNA datasets containing Nm modifications. Application of NanoNm revealed an enrichment of Nm sites on cancer-associated mRNAs in prostate cancer cells compared to healthy controls (Li et al., 2024).

While nanopore direct RNA sequencing enables antibody-free, single-molecule profiling and detection of co-occurring modifications, it is limited by subtle, context-dependent signal shifts and reliance on computational models that affect sensitivity and reproducibility. The primary limitation in many applications is the elevated error rate of nanopore sequencing, which ranges from 6 to 15% for the R9.4 pore—substantially higher than the 0.1–1% typical of Illumina short-read platforms (Wang et al., 2021). Additionally, modifications such as m7G remain difficult to resolve due to weak signal signatures, underscoring the need for orthogonal validation to ensure robust epitranscriptomic mapping. Hybrid approaches such as the combination of chemical treatments discussed earlier and nanopore direct RNA sequencing might be useful to identify low stoichiometry modifications that cannot be directly identified by nanopore basecallers.

Computational approaches

Together with sequencing technologies, the identification of RNA modifications on the transcriptome required the development of key computational tools both sensitive and specific. The interpretation of transcriptome-wide modification data is inseparable from the computational frameworks designed to extract signals from these experiments, and the two domains have co-evolved in sophistication. Based on existing epitranscriptomic datasets, and to overcome experimental limitations, RNA modification predictors were developed, capable of identifying potential sites of modification.

Early sequence-based tools demonstrated that m⁶A sites could be predicted with reasonable accuracy from RNA sequence context alone: SRAMP, one of the first such predictors, used random-forest classifiers trained on DRACH-motif features to prioritise candidate sites independently of experimental enrichment (Zhou et al., 2016). Subsequent deep learning architectures substantially improved both precision and isoform-level resolution, as illustrated by deepSRAMP, which combined Transformer and recurrent network modules to predict m⁶A at individual transcript isoforms across mammalian cell types (Fan et al., 2024), and by geography-aware models that encode the positional context of each adenosine within the full transcript architecture to achieve higher accuracy on low-stoichiometry sites (Huang et al., 2022). Similar machine learning strategies have been extended to modifications beyond m⁶A, exploiting the distinctive ionic current signatures recorded by nanopore direct RNA sequencing. Penguin, for instance, demonstrated that an ensemble of SVM, Random Forest, and neural network classifiers trained on raw ONT signal features and basecalled k-mer representations can identify Ψ sites with over 93% accuracy, substantially outperforming contemporary sequence-only Ψ predictors and establishing nanopore signal-level feature extraction as a viable route for modification-specific site detection (Hassan et al., 2022).

Epitranscriptomic repositories

The accumulation of transcriptome-wide modification data across methods, species, and disease contexts has necessitated the development of specialised public databases capable of standardising, annotating, and cross-referencing these diverse datasets. DirectRMDB was the first repository dedicated to RNA modification sites derived exclusively from nanopore direct RNA sequencing, cataloguing 16 modification types across 904,712 sites in 25 species from 39 independent studies and offering isoform-aware views of modification profiles that short-read databases cannot provide (Zhang et al., 2023). For the m⁶A modification specifically, m6A-Atlas v2.0 consolidated 797,091 high-confidence base-resolution m⁶A sites across multiple species, integrating single-cell profiles and providing writer-eraser-reader target association tools that facilitate direct functional hypothesis generation (Liang et al., 2024). Expanding beyond modification site catalogues, RMBase v3.0 now integrates over 1,880 high-throughput datasets to annotate 73 distinct modification types across 62 species, with modules linking modification sites to RNA-binding protein interactomes, miRNA target overlaps, GWAS-derived disease variants, and histone modification co-occurrence patterns (Xuan et al., 2024). Complementing site-level resources, RMDisease v2.0 specifically addresses the relationship between genetic variation and epitranscriptomic regulation, cataloguing single-nucleotide variants that fall within or proximal to experimentally mapped modification sites and linking these to GWAS-associated traits and disease phenotypes (Song et al., 2023a). For tissue-resolved analysis, m6A-TSHub provides context-specific m⁶A profiles derived from 23 human tissue types, together with a curated collection of m⁶A-affecting single-nucleotide mutations, a resource that enables the study of tissue-specific epitranscriptomic regulation and its perturbation in cancer (Song et al., 2023b). Together these databases constitute a layered infrastructure for the field: from raw modification site catalogues and stoichiometric measurements, through writer-eraser-reader target networks, to disease-variant associations and tissue-resolved landscapes, they collectively transform the fragmented output of individual high-throughput experiments into an integrated, queryable knowledge base that is essential for translating descriptive modification maps into mechanistic and clinical insights.

Discussion

This technique-centered review highlighted the current epitranscriptomic methods to detect the most important post transcriptional RNA modification, namely, m⁶A, m⁵C, Ψ, m¹A, m7G, Nm, and A-to-I editing. Overall each of the approaches described demonstrated to be extremely valuable to identify the distribution of specific modifications within the transcriptome. It is thanks to these techniques that the epitranscriptomic field progressed immensely in the last two decades. Despite this, each technique presents specific limitations that can be resolved only through integration of orthogonal approaches to identify high-confidence modifications transcriptome-wide. A detailed summary of the techniques available for detecting each specific RNA modification is provided in Supplementary Table S1.

IP-seq rapidly maps hotspots (e.g., m⁶A near stop codons) but suffers from coarse resolution (∼100 nt) and limited reproducibility (<50%) due to antibody biases and cross-reactivity (McIntyre et al., 2020; Moon et al., 2025). Chemical and enzymatic probing achieve base-resolution precision (e.g., RBS-seq multiplexing) but require substantial RNA input and may induce degradation. RT-based signatures offer simpler, cost-effective readouts but carry false-positive rates of 20–30%. Nanopore sequencing preserves native RNA context and detects co-occurring marks (e.g., CHEUI m⁶A/m⁵C) but remains limited by model-dependent errors and 5–20% FDR (Acera-Mateos et al., 2024). Finally, RNA mass spectrometry, despite extremely sensitive detection of RNA modifications has limited use in the characterization of epitranscriptomic marks within their RNA sequence context (Figure 1D).

Genetic perturbations, encompassing knockouts, knockdowns, and overexpression systems, represent the most direct framework for causal inference in epitranscriptomics, enabling comparative profiling between perturbed and wild-type states to discriminate bona fide modification sites from technical artifacts while elucidating functional roles (Dominissini et al., 2012; Heissenberger et al., 2019). CRISPR-CAS9 represent an extremely valuable and flexible technology to generate such models while adoption of high-fidelity Cas9 variants (e.g., HiFi Cas9), combined with clonal validation strategies, minimizes off-target effects (Vakulskas et al., 2018; Ortega-Batista et al., 2025).

Functional dissection of specific epitranscriptomic regulators underscores the power of genetic depletion strategies to clean background noise and validate modification-specific signals. METTL3/14 knockout ablates ∼90% of MeRIP-seq peaks, establishing a high-confidence m⁶A baseline and revealing stabilization of oncogenic transcripts (Dominissini et al., 2012). Similarly, NSUN2 and DNMT2 ablation cleanly removes the majority of m⁵C marks across tRNA and mRNA, isolating modification-dependent effects on transcript stability and stress response (Zhang et al., 2012; Tuorto et al., 2015). PUS1, PUS7, and TRUB1 knockouts provide an unambiguous pseudouridine-depleted background to uncover dynamically regulated ψ sites linked to translational control (Schwartz et al., 2014), while TRMT6/61 A/B depletion clarifies m¹A-dependent gene silencing through tRF-3 pathways (Su et al., 2022). Double knockout of ADAR1/2 results in near-complete loss of A-to-I editing, establishing a definitive unedited baseline for inosine detection (Anantharaman et al., 2017). Collectively, these genetic models remain indispensable reference frameworks for benchmarking computational modification callers and distinguishing true epitranscriptomic signals from technical noise.

A second strategy to overcome limitations of epitranscriptomic techniques is their orthogonal integration. Two independent approaches such as antibody-based immunoprecipitation and direct RNA sequencing could be integrated to identify bona fide modifications from background noise. For example, m7G BoRed-seq and m7G RIP were combined to identify modified miRNA in lung cancer cells (Pandolfini et al., 2019); similarly, nanopore direct RNA sequencing-based m6A calls in MOLM13 leukemia cells were intersected with miCLIP-seq profiles from the same cell line, with the overlapping sites representing high-confidence bona fide m6A modifications supported by two independent detection principles (Kim et al., 2025), illustrating how the convergence of signal-based and antibody-based approaches provides a robust filter against platform-specific false discoveries.

Alternatively, two approaches could be combined to increase sensitivity, resolution and specificity of epitranscriptomic techniques. MiCLIP represents a clear example of this hybrid approach, while anti m6A antibody is used to recognise modified bases, UV cross linking was used to generate specific RNA-antibody adducts that can be detected through RT termination and mutagenesis.

Hybrid workflows effectively bridge these gaps. Nanopore sequencing nominates candidate sites, which can then be validated via bisulfite sequencing or LC-MS/MS to confirm chemical identity and stoichiometry ( Burrows and Fleming, 2023 ). Integrative pipelines combine IP peaks, RT signatures, and MS data to resolve co-modifications (Shaw et al., 2024), while candidate sites identified through mapping approaches can guide targeted CRISPR validation ( Liu H. et al., 2019 ).

Key challenges include detecting rare or low-stoichiometry modifications, resolving context-dependent artifacts, achieving reproducible quantification, and linking modifications to functional outcomes. Hybrid approaches integrating orthogonal methods, nanopore sequencing, chemical probing, IP-based mapping, RT signatures, and genetic perturbations, emerge as essential for robust epitranscriptomic mapping. Moving forward, standardized benchmarks, rigorous validation, and computational innovations will be critical to translate descriptive maps into mechanistic insights and therapeutic opportunities.

Emerging trends and future perspectives

Three emerging frontiers are poised to reshape the field beyond its current boundaries. First, single-cell epitranscriptomics is transitioning from proof-of-concept to routine application: adaptations of MeRIP-seq to ultra-low-input formats and integration of nanopore direct RNA sequencing with single-cell barcoding strategies are beginning to reveal cell-type-specific modification landscapes impossible to identify through bulk approaches (Feng et al., 2024; Tegowski et al., 2022). These advances will be indispensable for dissecting the epitranscriptomic heterogeneity underlining cancer cell plasticity, stem cell fate decisions, and therapy resistance. Second, most of the current methods require RNA extraction, which collapses the dynamic modification states that may exist in specific subcellular compartments, polysomal fractions, or stress granules (Kong et al., 2023). The development of cell-permeable chemical probes, building on the logic of bisulfite conversion or CMCT labeling but compatible with living cells, represents a priority goal for the next generation of epitranscriptomic tools. Third, while spatial transcriptomics revolutionized our understanding of cancer and its microenvironment, epistranscriptomic detection techniques are still not capable of spatially resolving modifications with the exception of A to I conversion (Lee et al., 2022). Theoretically, one possible approach would be a spatial DART-seq approach for m6A detection where RNA are edited by the YTHDF1-APOBEC1 fusion in situ, followed by spacial transcriptomic analysis (Feng and Koziol, 2026).

The therapeutic exploitation of the epitranscriptome is moving rapidly from observation to intervention and precise and stoichiometrically accurate detection methods are fundamental for the development of epitranscriptomic therapeutic agents Pharmacological inhibitors targeting METTL3, FTO, and ALKBH5 are advancing through pre-clinical pipelines, with several showing efficacy in AML, glioblastoma, and solid tumor models (Yankova et al., 2021; Su et al., 2020). Beyond small-molecule strategies, epitranscriptomic editing using CRISPR-Cas13-writer or -eraser fusions offers programmable, site-specific modulation of individual modification sites, opening avenues for causal mechanistic dissection and, potentially, therapeutic reprogramming of aberrant epitranscriptomic circuits in disease (Wilson et al., 2020). Together, these directions underscore that detection methodology and biological application are co-evolving: advances in one domain directly enable progress in the other.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. The authors declare that portions of the original manuscript draft were produced with the assistance of a generative AI language model (ChatGPT, OpenAI). All AI-generated content was critically reviewed, fact-checked, and substantially revised by the authors, who take full responsibility for the accuracy of the final text.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frnar.2026.1872456/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Mattia Furlan, Institute for Cancer Research and Treatment (IRCC), Italy

Reviewed by: Kunqi Chen, Fujian Medical University, China

Chun Tang, Peking University, China