The Expanding Chemical Alphabet of the Human RNome

1. Introduction

The nucleic acid alphabet of life extends far beyond the canonical nucleotides of adenine, cytosine, guanine, and thymine in DNA and uracil in RNA [1,2]. While transcription copies the basic genetic code from DNA, the resulting primary RNA transcripts are typically unstable and structurally immature [3,4]. This necessitates significant post-transcriptional enzymatic modification of RNA, including splicing and A-to-I editing as well as the hundreds of modifications ranging from simple methylations to complex multi-step hypermodifications, which shape RNA conformation and function and enable dynamic and granular control of gene expression [5]. Since the discovery of pseudouridine (Ψ), often referred to as the ‘fifth RNA nucleotide’ [6], in calf liver RNA hydrolysate by Cohn and Volkin in 1951 [7], the world of RNA modifications has expanded exponentially. To date, over 180 modified ribonucleosides have been reported across all kingdoms of life [8], with new modifications being discovered regularly and this number expected to climb continuously [9].

The human epitranscriptome represents the complete landscape of post-transcriptional modifications found on all RNA molecules, such as transfer RNA (tRNA), ribosomal RNA (rRNA), messenger RNA (mRNA), and various non-coding RNA species [10]. These are much more than chemical decorations and govern every stage of the RNA lifecycle, including splicing, nuclear export, stability, translation efficiency, and environmental stress adaptation [8]. Rather than acting as static structural features, these modifications are dynamically orchestrated by a broad network of ‘writers’ that install the modifications, ‘readers’ that recognize and bind to them, and ‘erasers’ that remove them in response to cellular demands [11]. Consequently, this dynamic chemical layer serves as a master regulator of cellular fate, and its dysregulation is now associated with a broad spectrum of human diseases, from neurological disorder to cancer [12,13]. For example, in glioblastoma, overexpression of the epitranscriptome writer METTL1 installs m⁷G46 modifications across a subset of tRNAs including the isodecoder tRNA-Arg-TCT-4-1, which stabilizes and increases its copy numbers, ultimately boosting the translation of cancer-driving mRNAs enriched in corresponding cognate codons [13,14]. Given the critical roles these modifications play in cellular function, human health, and disease, and because the field is still in its relative infancy, further research is essential to fully decipher the spectrum of modifications across various RNA species, elucidate their underlying molecular mechanisms, and facilitate the development of novel diagnostic and therapeutic methods.

Not all the 180 reported modifications, however, rest on equally solid evidence. Beyond the most abundant, well-characterized epitranscriptome marks such as Ψ, m⁶A, and 2′-O-methylation (Nm) in human RNAs [15], recent years have witnessed the discovery of novel RNA modifications of low abundance from highly limited sample materials. Typical examples include 5-taurinomethyluridine (τm⁵U) and 5-taurinomethyl-2-thiouridine (τm⁵s²U) in human mitochondrial tRNAs (mt-tRNAs) [16,17], the identification of which was facilitated by breakthroughs in RNA purification, isolation, and enrichment in tandem with continuous improvement in the sensitivity and specificity of modern analytical instrumentation. As many of these newly reported modifications often occur at ultra-low abundance, lack orthogonal validation, and rely on unique, non-commercially available synthetic standards, the literature remains complicated by conflicting datasets, inadvertent technical artifacts, and false positives generated during sample handling or enzymatic hydrolysis. This growing gap between discovery and validation highlights an urgent need for rigorous validation frameworks and standardized criteria to cross-examine and validate novel epitranscriptomic discoveries.

To address these pressing challenges, this review provides a systematic and critical evaluation of recent additions to the human epitranscriptome, comprehensively cataloging all modified ribonucleosides reported across the entire spectrum of human RNA species. While existing literature typically addresses RNA modifications broadly across all kingdoms of life or focuses exclusively on the human tRNA epitranscriptome, a unified reference documenting all modified ribonucleosides across all human RNA classes has been conspicuously absent. To bridge this gap, we build on the 52 RNA modifications identified in the landmark review by Tsutomu Suzuki and coworkers on tRNA modifications [18], the publication by Taoka et al. on rRNA modifications [19], and the MODOMICS database to update the inventory of the human epitranscriptome [20].

We then expand this baseline through a systematic literature search to capture newly discovered and previously overlooked modifications. More importantly, we establish a standardized set of analytical validation criteria to score the detection confidence of these novel ribonucleosides and discuss their biological significance and associations with diseases. We further compile all human RNA modifications reported as of July 2026 into two ready-to-use tables for the field: one listing their detection confidence tier and RNA species of occurrence, and another detailing their instrumental parameters. Ultimately, we hope this review serves as a definitive resource and a rigorous benchmark for high-fidelity epitranscriptomic research, paving the way for future discoveries and the development of dependable diagnostic and therapeutic innovations.

2. Methods for Mapping RNA Modifications

The accurate mapping of the human epitranscriptome requires the strategic integration of targeted sample preparation and complementary analytical instrumentation (Figure 1). Historically, candidate modifications have emerged either from targeted hypotheses formulated by researchers or as unassigned peaks discovered during unrelated experiments. Regardless of their origin, authenticating a novel RNA modification requires resolving a complex, multi-tiered analytical puzzle. Because no single method can independently validate a new mark, researchers must first isolate or enrich the specific RNA species of interest, combine multiple orthogonal analytical technologies, synthesize chemical standards to cross-validate chemical identities and structures, and deploy sequencing platforms to capture exact transcriptomic maps. Ultimately, proving the presence of a novel mark requires satisfying two criteria: establishing its absolute chemical structure and identifying its precise location a specific transcript.

Liquid chromatography-coupled mass spectrometry (LC-MS) has routinely been used for identifying, structurally characterizing, and quantifying nucleic acid modifications given its ability to chromatographically resolve complex mixtures, uniquely select specific molecules as ions, and break them into predictable fragments, often with exact mass measurements [21]. Reassembly of the resulting data provides rigorous chemical identification and quantification. In epitranscriptome studies, researchers generally hydrolyze enriched RNA samples into individual monoribonucleosides using a combination of endonucleases, phosphodiesterases, and alkaline phosphatases to maximize analytical sensitivity and chromatographic resolution [22], then analyze the resulting hydrolysates by an MS¹ full scan to determine accurate molecular mass. Candidate peaks are then subjected to tandem mass spectrometry (MS² or MSⁿ), in which collision-induced dissociation (CID) or higher-energy collisional dissociation (HCD) generates characteristic product ions for deep structural elucidation [23]. To corroborate these mass-centric findings, researchers can pre-purify target candidate peaks or employ complementary assays, such as ultraviolet spectrophotometry, infrared spectrophotometry, pulse labeling, and nuclear magnetic resonance (NMR) spectroscopy [23,24], to confirm the presence and exact stereochemical positions of specific functional groups within the core structure. Once a structure is proposed, researchers synthesize matching chemical standards to rigorously cross-validate the endogenous candidate mark [24]. High-confidence confirmation demands that the candidate peak in the sample and the synthetic standard exhibit identical chromatographic retention times, high-resolution exact masses, and CID/HCD tandem mass spectra.

After successfully validating the chemical identity of a novel modification, completing the epitranscriptomic puzzle still requires mapping its precise coordinates within the transcriptome. To achieve this, the field has deployed various RNA sequencing (RNA-seq) methodologies, including short-read sequencing (SRS), long-read sequencing (LRS), and LC-MS-based RNA-seq (Figure 1) [4]. It is important to note that none of these methods provide the unequivocal identification or quantification provided by LC-MS ribonucleoside analysis.

Historically anchored in SRS, this methodology is widely regarded as an indirect approach because RNA must first be reverse-transcribed into cDNA before sequencing: it is the cDNA, not the modified ribonucleotide itself, that is actually read [4,25]. What differs between methods is how a modification is rendered visible within that cDNA. Reverse transcriptase (RT) signature-based methods are applied to modifications that intrinsically disrupt RT through misincorporation, stalling, or abortive cDNA synthesis, thereby leaving specific and interpretable signatures at the modified sites in the cDNA [25,26]. These signatures arise because the modified base no longer forms normal Watson-Crick pairs with its canonical partner, or because bulky or highly hydrophobic substituents slow the rate of cDNA synthesis [26]. For RT-silent modifications, chemical- and biological-based methods can instead be used to react selectively with the modified base and introduce a distinctive RT signature or cleave modified and unmodified bases differently to convert chemical identity into a positional cut site [25,26]. In addition, antibody-based methods are used to enrich modification-bound RNA fragments prior to RT and sequencing [27], but this enrichment generally sacrifices single-nucleotide resolution as they typically achieve peak-level resolution on the order of ~100 – 200 nt [26]. However, most SRS methods are optimized to interrogate one specific modification at a time, and each modification requires substantial effort to develop a dedicated antibody, probe, or enzyme. This specialization creates a major bottleneck for using SRS in the discovery of novel RNA modifications.

In contrast to SRS, LRS can sequence native RNA molecules as they are threaded through a protein nanopore, without prior fragmentation, amplification, or chemical labeling [28]. As each short stretch of RNA resides within the pore, it produces a characteristic disruption in the ionic current passing through. RNA modifications perturb this expected current signature and can alter the time each stretch spends in the pore (the dwell time), providing a direct readout of modification presence at single-nucleotide resolution [29]. This signal can be interpreted in three main ways: (1) as “errors” when a canonical basecaller is forced to represent a modified base as one of the four standard nucleotides; (2) as a direct comparison of raw current intensity and dwell-time distributions against an unmodified control sample; or (3) through basecalling models trained on known modifications to call a specific modification [30]. Current LRS methods are therefore restricted to modifications for which suitable training data exist. A novel modification will produce, at best, an unassigned current or dwell-time anomaly relative to a control sample and will still require orthogonal methods to validate its presence and define its chemical identity.

Alternatively, researchers can leverage LC-MS-based RNA-seq methods, which represent a direct, modification-type-independent approach to simultaneously and accurately identify and locate multiple modifications in RNA samples at single-base resolution [31]. Instead of fully digesting the RNA down to monoribonucleosides, this approach utilizes site-specific endoribonucleases, such as RNase T1, to generate oligonucleotide fragments of varying lengths [32]. By coupling data-dependent/data-independent acquisition scans for these oligonucleotides with specialized data-processing software suites, such as Thermo BioPharma Finder and the OpenMS Nucleic Acid Search Engine (NASE) [33], researchers can simultaneously decode the nucleotide sequences and pinpoint the exact transcriptomic coordinate of a modification based on fragment ions. However, LC-MS-based RNA-seq is comparatively weak at distinguishing isobaric modifications and at pinpointing the exact functional group and position responsible for a given mass shift [33]. For example, a +14 Da shift from methylation could arise from a methyl group on the base (m⁶A), on the 2′-O position of the ribose (Am), or even at different positions on the same base (m¹A). Achieving that level of structural resolution still requires cross-validation against monoribonucleoside-level LC–MS and the dedicated analytical chemistry methods.

Taken together, LC-MS ribonucleoside analysis is essential for identifying a modified ribonucleoside and establishing whether its structure actually exists in the human epitranscriptome. Once a modification has been structurally validated, it must then be localized within RNA sequences by SRS, LRS, or LC-MS-based RNA-seq. Because each analytical method and sequencing platform has its own strengths and weaknesses, orthogonal validation across multiple platforms is critical for increasing rigor and confidence. Replication by independent laboratories is also highly desirable to further strengthen detection claims. By integrating ribonucleoside-level structural validation, high-resolution spatial mapping data, and orthogonal validation, we can confidently authenticate a new component of the human epitranscriptome.

3. A Scoring System for Confidence in Identifying RNA Modifications

Many of the modifications described in the literature vary widely in the rigor with which their structures and RNA locations have been established. To assess the confidence for the recently reported additions to the human epitranscriptome, we introduce a standardized analytical validation scoring system that evaluates the integration of monoribonucleoside characterization, synthetic standard cross-validation, and transcriptomic mapping. Rather than treating all literature reports with equal weight, this framework classifies modifications into three progressive confidence tiers (Table 1), based on the depth and stringency of the analytical evidence provided in the original discovery papers.

Authenticated modifications represent the highest confidence tier: both chemical identity and transcriptomic location must be established beyond analytical ambiguity. Identification requires a full precursor and product ion scan, complementary structural methods such as NMR, cross-validation against a synthetic standard, and orthogonal confirmation by an independent laboratory. Transcriptomic mapping requires single-base or fragment-level mapping via LC-MS-based RNA-seq. Provisional modifications meet nearly the same identification bar, but complementary structural methods and independent orthogonal confirmation become optional. The larger distinction from Authenticated status lies in localization, where mapping is confined to bulk RNA fractions rather than a precise sequence coordinate, and no single sequencing platform is required to reach this tier. Putative modifications fall short on both fronts. Identification rests on only a low-resolution precursor ion scan, with the product ion scan, complementary structural methods, synthetic-standard cross-validation, and orthogonal confirmation all optional, and mapping data are absent entirely. This tier is not necessarily a dead end, as a Putative candidate can be promoted to Provisional or Authenticated status with further structural validation and precise mapping, while others may instead be reclassified as technical artifacts or false positives upon closer scrutiny.

4. The Expanding Repertoire of Human RNA Modifications and Their Biological Significance

The purpose of this review is to provide an accurate inventory of the human RNA modifications, building on previous reviews of tRNA modifications [18], analysis of rRNA modifications [19], and the MODOMICS database [20]. As depicted in Figure 2, we expand the repertoire by 24 RNA modifications, assessing the degree to which the structures have been validated. Supplementary Table 1 summarizes these 24 RNA modifications and the other 52 that together comprise the human epitranscriptome.

1,N⁶-Dimethyladenosine (m^1,6A). m^1,6A was first identified in 1991 in the urine of a patient with stage IV lung adenocarcinoma, characterized as its trimethylsilyl derivative by gas chromatography-MS and confirmed against a synthetic standard [34]. This work established its chemical identity but not biological origin, as the authors concluded the study by noting that m^1,6A had not been identified as a component of tRNA, though they speculated that a methyltransferase might install a second methyl group to m¹A or m⁶A [34]. This modification remained unassigned to an RNA species or enzymatic pathway for over thirty years, until 2022, when You et al. unambiguously identified m^1,6A in human tRNAs [35].

You et al. synthesized an m^1,6A standard and cross-validated its structure by high-resolution MS/MS and MS³ fragmentation as well as NMR [35]. Its presence in cellular RNA was confirmed both by spiking the synthetic standard into tRNA hydrolysate and by stable-isotope tracing with deuterated methionine, and m^1,6A was undetectable in the other RNA species tested such as mRNA and 28S/18S/5.8S/5S rRNAs. m^1,6A occurs at frequencies ranging from 0.0049% to 0.0084% in tRNAs from human cell lines, including HEK293T, HeLa, HepG2, K562, Jurkat-T, MCF-7, and HL-7702 [35]. Site-specific mapping was equally rigorous: using a targeted DNA hybridization protection/S1 nuclease digestion strategy on the purified tRNA^Gly(GCC) to isolate the exact oligonucleotide fragment containing the modification, the authors located m^1,6A at position 58 of the tRNA, the same site occupied by the well-characterized m¹A58 mark. While You et al. closed the gap the 1991 urinary finding had left open, no other publication has provided orthogonal evidence to show the presence of m^1,6A in human RNA. On this basis, we assign m^1,6A as a Provisional modification: definitive identification and location but no independent validation.

Mechanistically, the authors incubated purified tRNAs with TRMT6/61A and observed an increased level of m^1,6A, confirming TRMT6/61A is responsible for m^1,6A formation. Because no dedicated methyltransferase for m⁶A58 has been identified, they propose a two-step pathway in which a fraction of m¹A58 first undergoes Dimroth rearrangement to m⁶A58, which TRMT6/61A then methylates at N1 position to form m^1,6A58. The authors also demonstrated that m^1,6A can be reversed by the demethylase ALKBH3. To assess its disease relevance, the authors compared m^1,6A levels in tRNAs from human breast cancer tissues against matched adjacent normal tissues. They found that m^1,6A, along with m¹A and m⁶A, was significantly elevated in cancer tissue and that this increase correlated with reduced mRNA expression of the demethylases ALKBH1 and ALKBH3, whereas mRNA levels of TRMT6 and TRMT61A were only slightly increased. These observations point to the demethylase loss, rather than increased methyltransferase activity, as the more likely driver of elevated m^1,6A, m¹A, and m⁶A levels in cancer tissue. Finally, they noted that m^1,6A may represent either an RNA damage product or a genuine regulatory mark, since its proposed biosynthetic route involves a non-enzymatic Dimroth rearrangement step.

3,2′-O-Dimethylcytidine (m³Cm) and N4,2′-O-dimethylcytidine (m⁴Cm). Cytidine dual methylations occurring at both the nucleobase and the ribose 2′-hydroxyl were already known in two forms before the discovery of m³Cm: m⁴Cm, found at position 1402 of the 16S rRNA of Escherichia coli, where it fine-tunes the local structure of the decoding site [36], and 5,2′-O-dimethylcytidine (m⁵Cm), found in archaeal and mammalian tRNA [37]. However, the third possible combination, m³Cm, had never been reported in any domain of life and m⁴Cm itself had never been identified outside bacteria. Cheng et al. addressed both gaps in a single study by synthesizing an m³Cm standard, characterizing it with high-resolution MS/MS and NMR, and developing an LC-MS/MS method capable of resolving the three structural isomers m³Cm, m⁴Cm, and m⁵Cm, which share identical molecular weights and multiple reaction monitoring (MRM) transitions [38].

The authors detected both m³Cm and m⁴Cm in the total RNA of HeLa cells, confirmed each identity by high-resolution MS/MS, spiking synthetic standards into RNA hydrolysate, and stable-isotope tracing with deuterated methionine. Following experiments showed that m³Cm and m⁴Cm are localized to small RNA (<200 nt) and 18S rRNA in HeLa cells. To rule out contamination from bacterial 16S rRNA as the source of the signal, the authors performed additional confirmation by agarose-gel purification of 18S rRNA to over 99.999% purity. Across four human cell lines (HeLa, HepG2, HL-7702, and MCF-7), m³Cm was quantified at 0.0003–0.0008% in total RNA and 0.002–0.007% in small RNA, while m⁴Cm was quantified at 0.0004–0.001% in total RNA and 0.0009–0.003% in 18S rRNA. In both cases, the mapping stopped at the RNA species level, and for m³Cm, the authors infer it most likely originates from tRNA, since tRNA accounts for roughly 90% of the small RNA pool. As this is the only study reporting their presence in human RNA and the modifications are only localized to a bulk RNA species, both m³Cm and m⁴Cm are Provisional modifications under our scoring framework.

To date, no enzyme has been experimentally linked to the formation or removal of both modifications in human cells. The authors propose METTL6 and FTSJ1 as candidate writers that could act sequentially to generate m³C and then m³Cm, since METTL6 is known to install m³C in serine tRNA and FTSJ1 carries out 2′-O-methylation in the tRNA anticodon region, respectively. No analogous candidate pathway is proposed for m⁴Cm formation. The authors also speculate that both modifications could enhance the structural stability of tRNA and 18S rRNA. Future studies will be necessary to better understand the biological roles and disease associations of both m³Cm and m⁴Cm.

5,2′-O-Dimethylcytidine (m⁵Cm). m⁵Cm was first structurally characterized in 1987 by Edmonds et al., who identified it in the tRNA of three extremely thermophilic archaebacteria, Sulfolobus solfataricus, Thermoproteus neutrophilus, and Pyrodictium occultum [39]. Three decades later, Huber et al. used a commercially obtained m⁵Cm standard and LC-MS to report the first detection of m⁵Cm in total RNA from HEK293T cells, as part of a study primarily focused on a related oxidative derivative, 2′-O-methyl-5-hydroxymethylcytidine (hm⁵Cm); no attempt to detect m⁵Cm in different RNA fractions was made [37]. Using a similar commercial standard and LC-MS, Feng et al. and Cheng et al. later verified that m⁵Cm is present in small RNA (<200 nt) in human cells [38,40].

Feng et al. developed a chemical-labeling method with 2-bromo-1-(3,4-dimethoxyphenyl)-ethanone (BDMOPE) that improved detection sensitivity of m⁵Cm by over 100-fold over unlabeled analysis [40]. They then confirmed the identity of endogenous m⁵Cm using its chemical standard and high-resolution MS/MS, quantified m⁵Cm in total RNA across six human cell lines (HEK293T, HL-7702, HepG2, Huh7.5, HeLa, and MCF-7; 0.0001%–0.0047%), and later localizing it predominantly to small RNA (<200 nt) in HEK293T and HeLa cells (0.0022%–0.044%). Cheng et al., by contrast, analyzed m⁵Cm in unlabeled form and confirmed its presence in both total RNA and small RNA (<200 nt) across four human cell lines (HeLa, HepG2, HL-7702, and MCF-7) by spiking the standard into RNA hydrolysate and by high-resolution MS/MS [38]. In total RNA, m⁵Cm was detected at 0.0001% to 0.0003%, and at 0.0006%–0.002% in small RNA, comparable with the findings of Feng et al.. However, mapping in both papers stopped at the RNA species level. The only site-specific mapping data in the literature comes from Kawarada et al., who used oligonucleotide MS and detected m⁵Cm at position 34 of cytoplasmic tRNA^Leu(CAA) in ALKBH1-KO HEK293T cells but not in wild-type cells [41]. Because this finding reflects a single tRNA species under an artificial KO condition, it cannot be used to confirm the specific source of the m⁵Cm detected in studies by Feng et al. and Cheng et al. On this basis, the characterization of m⁵Cm in human cells meets Provisional standard under our scoring framework.

Mechanistically, Kawarada et al. showed that in ALKBH1-KO cells, NSUN2 methylates C34 to form m⁵C34 in the primary transcript of cytoplasmic tRNA^Leu(CAA) in the nucleus, which is then exported to the cytoplasm and methylated by FTSJ1 to form m⁵Cm34 [41]. Under normal conditions, the same m⁵C34 intermediate is instead oxidized by ALKBH1 before 2′-O-methylation occurs, so the biosynthetic route to m⁵Cm specifically under physiological conditions still requires further validation. The clearest functional signal for m⁵Cm comes from Feng et al.’s comparison of 12 paired human lung carcinoma tissues against matched tumor-adjacent normal tissue, which found that m⁵Cm was significantly elevated in tumor tissue, while its downstream oxidation products hm⁵Cm and 5-formyl-2′-O-methylcytidine (f⁵Cm) were both significantly decreased [40]. The authors linked this observation to decreased expression of ALKBH1 in tumor tissue, proposing that reduced ALKBH1 activity could cause m⁵Cm to accumulate without being converted into hm⁵Cm and then f⁵Cm, as demonstrated in Kawarada et al. that ALKBH1 is responsible for this conversion [41]. This positions m⁵Cm as one node in a multi-step oxidative cascade (m⁵Cm → hm⁵Cm → f⁵Cm) whose balance shifts measurably in cancer tissue, suggesting that altered levels of these three modifications may contribute to lung carcinoma development and could serve as a candidate diagnostic indicator.

5-Carboxylcytidine (ca⁵C). Driven by TET family of enzymes, DNA demethylation occurs through sequential oxidation of m⁵C, generating three intermediates, hm⁵C, f⁵C, and ca⁵C, in DNA [42]. Given the structural similarity between DNA and RNA, researchers speculated that RNA might undergo the same demethylation pathway. This hypothesis gained support when Fu et al. showed that TET proteins also oxidize m⁵C to hm⁵C in RNA from human cells [43], and Zhang et al. later also confirmed the presence of f⁵C in human RNA [44], leaving ca⁵C unexplored. Huang et al. closed this gap by developing a chemical-labeling method using 2-bromo-1-(4-diethylaminophenyl)-ethanone, which improves detection sensitivity for ca⁵C for 313-fold in LC-MS/MS analysis [45].

Using this method, the authors identified endogenous ca⁵C in total RNA from HEK293T cells and mouse liver tissue, confirming its identity by comparing retention times and high-resolution MS/MS fragmentation against its commercial standard [45]. Of note, ca⁵C was detected at substantially lower frequency (~1 per 10⁶ rG) than m⁵C, f⁵C, and hm⁵C (0.1 – 20 per 10⁴ rG) in mouse liver total RNA. By examining different RNA species from mouse liver tissue, the authors found that ca⁵C, similar to f⁵C and hm⁵C, was concentrated almost exclusively in mRNA (~20 per 10⁶ rG), with little to no signal in 28S rRNA, 18S rRNA, or small RNA (<200 nt). The trace level of ca⁵C was also reported by Arguello et al., who were unable to reliably detect ca⁵C in total RNA or small RNA from HEK293T cells by LC-MS/MS. The mapping stopped at RNA species level. On this basis, the discovery of ca⁵C in human RNA is Provisional under our scoring framework.

A plausible biosynthetic route for this endogenous ca⁵C was supported by Basanta-Sanchez et al. [46], who demonstrated that the catalytic domain of TET1 can oxidize f⁵C to ca⁵C in vitro across single-stranded, double-stranded, and hairpin RNA substrates, though conversion yields were substantially lower (<7% in the most permissive RNA context) than those observed during TET-mediated oxidation in DNA. The authors noted that the oxidative activities of other TET enzymes warrant further investigation, given their distinct RNA-binding domains. To date, ca⁵C in RNA has not shown a significant disease association. Although Huang et al. detected ca⁵C in total RNA from 20 paired human colorectal carcinoma tissues and 8 paired human hepatocellular carcinoma (HCC) tissues [45], they found no significant difference in ca⁵C levels between tumor and adjacent normal tissues; the same held true when the comparison was narrowed specifically to mRNA from paired HCC tissues. While the existence of ca⁵C in RNA is now established, further studies are required to define its functional and clinical significance.

5-Formyluridine (f⁵U) and 2′-O-methyl-5-formyluridine (f⁵Um). Following the discovery of DNA formylation, shown by the TET-mediated conversion of m⁵dC to hm⁵dC to f⁵dC [42], and the finding that f5C exists in human mt-tRNA^Met [47], Jiang et al. questioned whether other types of endogenous DNA and RNA formylation exist in humans [48]. In this study, they specifically searched for f⁵C, f⁵U, f⁵Cm, and f⁵Um in RNA. Given the known extremely low abundance of f⁵dC and f⁵C in DNA and RNA [42,49,50], they developed a chemical-labeling strategy pairing Girard’s P reagent (GirP) with in-tube solid-phase microextraction and LC-MS/MS, improving method sensitivity for f⁵C and f⁵U by 350- and 880-fold, respectively, relative to unlabeled analysis [48].

The authors synthesized standards for f⁵U and its GirP-labeled derivative and confirmed their identities by high-resolution MS/MS [48]. Using the developed approach, they detected f⁵U for the first time in total RNA from three cultured human cell lines (HEK293T, MCF-7, and HeLa) and human thyroid tissue, with identity confirmed by matched retention time and high-resolution MS/MS product-ion spectra against the synthesized standard. Detected levels of f⁵U ranged from 0.8 to 1.6 per 10⁶ nucleosides in cultured cells and from 0.4 to 3.9 per 10⁶ nucleosides in thyroid tissues, comparable to the levels detected for f⁵C. f⁵Um, by contrast, was detected without any standards in either its native or GirP-labeled form. The authors monitored its predicted MRM transition (m/z 420.3→274.2) and observed a single chromatographic peak in total RNA hydrolysate from human thyroid tissue, further supported by a high-resolution MS/MS spectrum consistent with the predicted fragment masses. Considering all the data presented and the fact that this is the only study reporting their occurrence in human RNA, we assign f⁵U as a Provisional modification and f⁵Um as a Putative modification.

The authors did not investigate the endogenous biosynthetic pathway of f⁵U or f⁵Um, and to our knowledge no subsequent study has identified the enzyme responsible for either modification. The clearest functional signal comes from the comparison of paired human thyroid carcinoma and tumor-adjacent normal tissue: f⁵U and f⁵C in RNA was significantly elevated in tumor tissue [48], while f⁵Um was not included in this quantitative comparison. The authors interpreted this increase in RNA formylation as potentially contributing to tumor formation and development, though the underlying mechanism was not established. Independent replication, a synthetic f⁵Um standard, and RNA-species- or site-resolved mapping for both marks remain outstanding needs before their biological significance can be considered established.

2′-Thiouridine (s²U). 2-Thiolation at the wobble position (position 34) of the anticodon loop is among the most extensively characterized modifications in the human tRNA epitranscriptome, occurring in tRNA^Gln, tRNA^Glu, and tRNA^Lys [51,52]^. At this position, s²U is rarely found in its standalone form and is almost universally further decorated at C5 to generate hypermodified 5-substituted-2-thiouridine derivatives (xm⁵s²U) [53], such as 5-methoxycarbonylmethyl-2-thiouridine (mcm⁵s²U) in cytoplasmic tRNAs and τm⁵s²U in mt-tRNAs [16,17,54]. Whether standalone s²U persists as a steady-state modification in human tRNAs has remained a subject of debate, and two independent studies have approached this question using different analytical approaches, providing preliminary evidence for its existence.

Dai et al. developed a chemical derivatization strategy using 2-bromo-1-(3,4-dimethoxyphenyl)-ethanone (BDMOPE) coupled with LC-MS/MS, achieving a 27-fold improvement in s²U detection sensitivity over underivatized analysis [55]. By comparing chromatographic retention time and high-resolution MS/MS fragmentation against a synthesized BDMOPE-s²U standard, they detected s²U in total RNA and small RNA from five human cell lines (HEK293T, HeLa, HepG2, HL-7702, and MCF-7) at levels ranging from 0.0057 to 0.0125% (s²U/U) in total RNA and 0.0380 to 0.0860% (s²U/U) in small RNA. BDMOPE-s²U was not detectable in isolated 18S rRNA, 28S rRNA, or mRNA fractions, consistent with s²U and its hypermodified derivatives being confined to tRNAs. The authors did not proceed to map the transcriptomic coordinates of the detected s²U signal. Using a different strategy, Suzuki et al. isolated mt-tRNA^Gln from human placenta, digested it with RNase T1, and analyzed the resulting oligonucleotide fragments by high-resolution MS/MS. At the wobble position, they resolved a mixture of modification states: 67% τm⁵s²U, 12% τm⁵U, 11% s²U, 1.3% cmnm⁵s²U, and 8% unmodified U [17]. This quantitative stoichiometric breakdown demonstrates that standalone s²U exists as a genuine, steady-state species at the wobble position of human mt-tRNA^Gln. On this basis, we assign standalone s²U as a Provisional modification, as nucleoside-level detection and site-specific oligonucleotide mapping do not support each other.

The biosynthetic pathways leading to mcm⁵s²U and τm⁵s²U are enzymatically independent: cytoplasmic 2-thiolation is carried out by the CTU1–CTU2 complex, which draws sulfur through an upstream relay involving NFS1, TUM1, UBA4, and the sulfur carrier protein URM1 [56], while mitochondrial 2-thiolation is catalyzed by TRMU and NFS1 [57]. Notably, neither pathway has been demonstrated to generate standalone s²U as an endpoint, and s²U is therefore best understood as a biosynthetic intermediate en route to the fully hypermodified xm⁵s²U species. No disease association has yet been attributed specifically to standalone s²U as distinct from the broader xm⁵s²U hypermodification landscape. Nevertheless, given that CTU1, CTU2, and TRMU are individually linked to cancer, neurological disorders, and mitochondrial disease, respectively [58], it remains possible that aberrant accumulation of unprocessed s²U intermediates contributes to these phenotypes.

5-Hydroxyuridine (ho⁵U). ho⁵U is a naturally occurring modified ribonucleoside enzymatically installed at the wobble position of bacterial tRNAs [59]. Depending on the type of bacteria, ho⁵U is further modified to cmo⁵U (Gram-negative bacteria) or mo⁵U (Gram-positive bacteria) to expand the decoding capacity of the modified tRNA to read codons from fourfold degenerate codon boxes [60]. In eukaryotes, by contrast, ho⁵U has historically been characterized as an oxidation product alongside 8-hydroxyguanosine and 8-hydroxyadenosine in yeast RNA, rather than as a naturally occurring modification [61]. More recently, in view of reports of two other hydroxylated ribonucleosides, N⁶-hydroxymethyladenosine (hm⁶A) and hm⁵C [43,62], both of which act as intermediates in RNA demethylation pathways, Dai et al. explored whether a parallel uridine hydroxylation modification, ho⁵U, might exist in human RNA, using chemical derivatization with BDMOPE coupled with LC-MS/MS [55]. This approach achieved a 32-fold improvement in ho⁵U detection sensitivity over the underivatized approach.

Using an authentic BDMOPE-ho⁵U standard as reference, Dai et al. detected a chromatographic peak with matching retention time and high-resolution MS/MS product-ion spectra in total RNA and small RNA isolated from five human cell lines (HEK293T, HeLa, HepG2, HL-7702, MCF-7), confirming the presence of ho⁵U in RNA of unstressed human cells. ho⁵U was quantified at 0.0011 – 0.0017% (ho⁵U/U) in total RNA and 0.0052 – 0.0062% (ho⁵U/U) in small RNA across the five cell lines, but was absent in 18S rRNA, 28S rRNA, and mRNA. Furthermore, Fleming et al. also detected ho⁵U in total RNA from both untreated and hydrogen peroxide-treated HEK293T cell using HPLC coupled with ultraviolet (UV) and electrochemical detection, which enables sensitive detection of ho⁵U based on UV absorbance and redox properties [63]. However, no orthogonal analysis, such as high-resolution MS/MS, was performed to verify the identities of the observed chromatographic peaks. Also, neither study performed additional experiments to define the transcriptomic coordinate of the detected ho⁵U. We therefore assign ho⁵U as a Provisional modification.

To date, no mammalian writer enzyme for ho⁵U has been identified, and no disease-association study specific to endogenous ho⁵U in human RNA has been reported.

N⁶-Formyladenosine (f⁶A) and N⁶-hydroxymethyladenosine (hm⁶A).N⁶-Methyladenosine (m⁶A) is a well-characterized and abundant internal modification on human mRNA, constituting 1–2% of all adenosines in mRNA and installed by the METTL3/METTL14 methyltransferase complex [62,64]. Studies have shown that m⁶A plays broad regulatory roles in RNA metabolism, including pre-mRNA processing, nuclear export, stability, translation efficiency, and non-canonical translation initiation [65]. Jia et al. had previously discovered that the human fat-mass-and-obesity-associated protein (FTO) possesses efficient oxidative demethylation activity toward this mark in human cells [66], providing the first example of reversible methylation in RNA and establishing m⁶A as a dynamic epitranscriptomic mark. Motivated by the fact that FTO is a Fe(II)- and α-ketoglutarate–dependent dioxygenase with versatile catalytic activity [67], Fu et al. sought to further dissect the mechanism of FTO-mediated m6A removal and discovered that FTO oxidizes m⁶A to hm⁶A and then to f⁶A in a stepwise manner [62].

Using high-resolution MALDI-TOF MS to monitor FTO-mediated oxidation of a synthetic m⁶A-RNA oligonucleotide, Fu et al. observed the expected demethylation product, unmodified A, along with two unexpected species: an m⁶A−2 Da species and an m⁶A+14 Da species [62]. They proposed that the m⁶A−2 Da species corresponded to N⁶-methyleneadenosine, formed by dehydration of hm6A and serving as indirect evidence for hm6A itself, while the m⁶A+14 Da species corresponded to f⁶A, arising from a second oxidation of hm⁶A, analogous to TET-mediated oxidation of m⁵C to hm⁵C to f⁵C during DNA demethylation. To confirm the identity of both species, the authors synthesized hm⁶A and f⁶A standards, characterizing hm⁶A by high-resolution MS/MS and f⁶A by both high-resolution MS/MS and NMR. On the ribonucleoside level, both FTO reaction products co-eluted with their respective synthetic standards and produced identical MS/MS fragmentation patterns, providing rigorous chemical confirmation. The authors then used an RNase T1/nuclease P1 digestion protocol at neutral pH to detect both hm⁶A and f⁶A in poly(A)-selected mRNA from cultured human HeLa cells and mouse liver tissue. Peak identity was confirmed by matched retention time and MRM transition (m/z 298.1→136.0 for hm⁶A; m/z 296.1→164.1 for f⁶A) against the synthetic standards. Both hm⁶A and f⁶A were estimated at a minimum of 0.5–1% of total m⁶A, since both may decompose to unmodified A during RNA isolation. No further mapping on poly(A)-selected RNA was performed. On this basis, hm⁶A and f⁶A are Provisional modifications.

Both modifications proved stable under simulated physiological conditions, with half-lives of 186 min (hm⁶A) and 188 min (f⁶A), comparable to the average half-life of mammalian mRNA. Mechanistically, both arise as sequential products in the two-step FTO-mediated oxidative demethylation of m⁶A, with oxidation of hm⁶A to f⁶A proceeding at a relatively slower rate than the oxidation of m⁶A to hm⁶A. The authors also pointed out that because the m⁶A writers (METTL3/METTL14 complex) and erasers (FTO and ALKBH5) partially colocalize in nuclear speckles, demethylated RNA could in principle be remethylated in a directly reversible manner. The transient hydroxymethyl and formyl intermediates at the N6 position of adenosine, however, may allow oxidized RNA to diffuse or be exported from nuclear speckles before remethylation can occur. The authors further showed that the m⁶A reader protein YTHDF2 does not bind either hm⁶A- or f⁶A-RNA, suggesting that they could serve as distinct marks recruiting different sets of RNA-binding proteins and regulating downstream RNA-related pathways. Current knowledge mostly links disease to m⁶A itself and its regulators in cancer, neurodegeneration, inflammatory disease, and metabolic disorders [65,68,69,70], but has not yet identified hm⁶A or f⁶A themselves as separate, functionally distinct modifications linked to human disease.

5-Methyl-4-pyrimidinone (h²U), its derivatives (xm⁵h²U), 5-carboxymethyluridine (cm⁵U), and 5-carboxymethyl-2-thiouridine (cm⁵s²U). As discussed earlier, s²U and its derivatives xm⁵s²U are wobble modifications found on both human cytoplasmic and mitochondrial tRNA^Glu, tRNA^Gln, and tRNA^Lys, where they are essential for efficient tRNA aminoacylation, accurate and efficient decoding of A/G-ending codons, and rapid translocation on the ribosome [16,17,53,54]. However, the thiocarbonyl group of the s²U moiety is chemically unstable and has been shown to be susceptible to oxidative desulfuration under in vitro oxidative stress conditions, generating the 4-pyrimidinone nucleoside (h²U) and unmodified U as end products [71]. While earlier studies remained largely confined to in vitro conditions, whether cellular oxidative stress could drive natural desulfuration in vivo went untested until two recent, independent studies demonstrated the presence of xm⁵h²U in RNA from human cells.

Sierant et al. applied LC-TOF-MS/MS with synthetic standards, including s²U, h²U, mcm⁵s²U, mcm⁵h²U, mcm⁵U, cm⁵s²U, cm⁵h²U, and cm⁵U, to test for the presence of xm⁵s²U and their desulfuration products in small RNA isolated from untreated human cells, including HEK293, HeLa, A375, MCF-7, MOLT-4, K562, U-87 MG, and A431 [72]. Their analysis revealed that mcm⁵s²U and cm⁵s²U were the most abundant of all the targets, at 29–53% and 33–57% of the modified nucleoside respectively, followed by mcm⁵h²U (2–11%), mcm⁵U (37% in HeLa and 4–8% in other cell lines), and cm⁵U (4–8% and detectable only in four cell lines); cm⁵h²U was detectable (3–5%) only when cells were treated with oxidants. While LC and MS parameters were provided in this paper, chromatograms and MS/MS spectra comparing sample analyses with synthetic standards were not shown. The presence of cm⁵U in human cells has also been independently reported by two studies. Fu et al. detected cm⁵U in small RNA from HEK293T cells using LC-MS/MS and confirmed the peak identity using high-resolution QTOF MS [73], while Qian et al. detected cm⁵U in tRNA from HCT116 cells by matching LC-MS/MS retention times to its synthetic standard [74].

More recently, Mo et al. prepared authentic xm⁵h²U by in vitro oxidation of mouse tRNA and pig mt-tRNA^Lys, then co-injected nucleosides from untreated and oxidized tRNA for high-resolution MS/MS analysis [75]. Their analysis detected peaks at m/z 301 and 366, i.e., 32 Da less than mcm⁵s²U and τm⁵s²U, respectively, in both untreated and oxidized tRNA, with a marked increase in signal in oxidized tRNA. CID of the two peaks generated a BH₂⁺ ion with loss of the unmodified ribose, indicating that the 32 Da loss originates from the base and confirming the presence of mcm⁵h²U and τm⁵h²U in mammalian tissue. Importantly, a spike-in experiment using E. coli tRNA as a tracer demonstrated that artificial formation of h²U derivatives during tRNA extraction and purification was negligible. The authors further mapped mcm⁵h²U, τm⁵h²U, and h²U to position 34 of individual mouse tRNA species at single-nucleoside resolution using RNase T1 oligonucleotide MS and fragment ion analysis. Quantitatively, mcm⁵h²U was present at 9.9% and 7.2% in mouse cytoplasmic tRNA^Lys and tRNA^Gln, respectively, and τm⁵h²U was detected at 2.3% and 2.9% in mouse mitochondrial tRNA^Lys and tRNA^Glu, respectively; h²U itself was detected at 0.4% and 1.3% in the same two mitochondrial tRNAs. In human cells, mcm⁵h²U was detected at ~15% of tRNA^Lys in HeLa cells, based on peak intensities of the RNase T1-digested fragments containing the anticodon, but at much lower levels in MOLM-13, RAW 264.7, MKN45, A549, T24, MCF-7, HuH-7, and HepG2. On this basis, we assign mcm⁵h²U as an authenticated modification, and cm⁵U and cm⁵s²U as Provisional modifications. τm5h2U and h2U are assigned as Putative modifications as they lack direct evidence in human RNA. cm⁵h²U is also assigned as a Putative modification as it was detected only under oxidative stress conditions.

Mechanistically, cm⁵U at wobble U34 is formed by the Elongator complex (ELP1-6) and serves as the first intermediate in the biosynthesis of mcm⁵U and mcm⁵s²U, whereas cm⁵s²U arises when this cm⁵-modified base is further subjected to 2-thiolation, probably by CTU1-CTU2 complex and associated factors, and it functions as an intermediate on the pathway toward mature xm⁵s²U [58,76]. To date, no human disease has been attributed specifically to cm⁵U or cm⁵s²U; instead, disease associations involve defects in the broader Elongator/ALKBH8- and CTU1–CTU2-dependent wobble-U modification pathways and their fully matured products such as mcm⁵s²U. For h²U derivatives, Mo et al. further demonstrated that mcm⁵h²U at the wobble position reduces aminoacylation efficiency for tRNA^Lys, tRNA^Glu, and tRNA^Gln, and impairs recognition of AAA/AAG codons at the ribosomal A-site, therefore negatively affecting codon-specific translation efficiency [75]. Whether xm⁵h²U accumulation is sufficient to trigger downstream consequences such as ribosome stalling or stress signaling cascades under oxidative stress conditions remains to be tested. Together, these observations support a model in which s²U-to-h²U conversion at the wobble position functions as an oxidative stress sensor and dynamically regulates codon recognition and protein synthesis during oxidative stress.

N⁴-Methylcytidine (m⁴C).N⁴-Cytosine DNA methylation is a widespread epigenetic modification in prokaryotes and serves important roles in bacterial evolution and gene regulation [77,78,79]. m⁴C has been found in the small subunit rRNAs of bacteria and in mitochondria of eukaryotic cells [79]. In E. coli, the SAM-dependent methyltransferase RsmH installs m⁴C at C1402 in the decoding center of 16S rRNA [36]. Although m⁴C had long been regarded as a bacteria-specific mark, an equivalent modification was reported in hamster mitochondrial rRNA in 1978 by Dubin et al. [80], but with the responsible enzyme unidentified. Given this conservation, and the presence of a human protein, METTL15, with clear sequence and structural homology to RsmH, Van Haute et al. set out to test whether METTL15 is responsible for introducing m⁴C at C839 in human 12S mt-rRNA and to characterize its functional role [81].

The authors generated METTL15 knockout (KO) HAP1 cells and quantified m⁴C in total RNA and mitochondria-enriched RNA from wild-type (WT) and KO cells by LC-MS/MS [81]. Using an LC method that resolves the positional isomers m³C, m⁴C, and m⁵C, they detected m⁴C readily in WT cells but not in METTL15 KO cells, while total RNA showed no m4C signal in either genotype. This indicates that mitochondrial RNA is the major cellular source of m⁴C and that METTL15 is required for its formation. Peak identity was confirmed by matched retention time against a synthetic m4C standard, and with no high-resolution MS/MS spectra were provided. The authors further used targeted RNA bisulfite sequencing, a method that is widely used to study m⁵C but also works for m⁴C, to confirm a modified base at C839 in 12S mt-rRNA from WT HAP1 and HeLa cells, consistent with the presence of m⁴C at this position. On this basis, we assign m⁴C as a Provisional modification.

By analogy to the N⁴-methylation of C1402 by RsmH in E. coli, the authors proposed that m⁴C839 also sits near the mRNA P-site codon and may help fine-tune P-site geometry, thereby improving decoding fidelity, particularly correct recognition of the initiation codon, though this remains to be tested [81]. They further showed that loss of functional METTL15 impairs mitoribosome biogenesis, mitochondrial translation, and mitochondrial oxidative phosphorylation performance. Because the loss of METTL15 also disrupts assembly of the small subunit of the mitochondrial ribosome, and given its position in the ribosome, the authors speculate that m⁴C at C839 may act together with the neighboring m⁵C841 to stabilize 12S rRNA folding, thereby facilitating the late stages of mitoribosome assembly.

N^2,7-Dimethylguanosine (m^2,7G) and N²,N²-7-trimethylguanosine (m^2,2,7G). Human mRNAs are capped co-transcriptionally at their 5′ end with a 7-methylguanosine (m⁷G) linked to the first transcribed nucleotide via a reverse 5′–5′ triphosphate bridge [82], as soon as the nascent 5′ end emerges from RNA polymerase II (RNAP II) [83]. This m⁷G cap protects transcripts from 5′-exoribonuclease-mediated degradation, promotes proper splicing and polyadenylation, and signals for nuclear export. It is also the key determinant of eIF4E-dependent translation initiation, and thereby strongly influences mRNA stability and protein production [84]. Besides protein-coding mRNAs, a subset of RNAP II transcripts, including spliceosomal small nuclear RNAs (snRNAs) and small nucleolar RNAs (snoRNAs), undergo similar 5′ capping but are further hypermethylated to m^2,7G cap and m^2,2,7G cap [84].

Abdelhamid et al. combined orthogonal detection strategies on snRNAs and snoRNAs isolated from THP-1 cells [84]. They first released 5′ cap structures by treating size-fractionated RNA (<50 nt, 50–100 nt, and 100–200 nt) with tobacco acid pyrophosphatase, radiolabeled the products, and resolved them by two-dimensional thin-layer chromatography (TLC) against radiolabeled m^2,7Gp and m^2,2,7Gp standards. Spots consistent with both m^2,7Gp and m^2,2,7Gp were observed across all three length fractions, with m^2,2,7Gp emerging as the most intense signal. To corroborate the TLC assignments, the authors digested fractionated RNA with RNase T2 and nuclease P1 to release intact 5′–5′ cap dinucleotides and analyzed the products by high-resolution MS. With mass error ≤ 3 ppm between observed masses and calculated masses, they confirmed the m^2,2,7G cap structure in all three length fractions, while the m^2,7G cap structure was found primarily in the fraction longer than 100 nt.

However, it should be noted that the 5′–5′ triphosphate bridge linking these caps may not be efficiently hydrolyzed by the endonucleases and alkaline phosphatases conventionally used for ribonucleoside analysis, raising some uncertainty as to whether m^2,7G and m^2,2,7G could be released as free ribonucleosides and detected during ribonucleoside analysis. Two earlier papers, by Ro-Choi et al. and Shibata et al. in 1975 [85,86], isolated mouse U2 snRNA and digested it into a 5′-oligonucleotide (m^2,2,7GpppAmUmC) using pancreatic RNase followed by snake venom phosphodiesterase and alkaline phosphatase, and analyzed by LC-UV. They detected a peak matching the retention time of an m^2,2,7G nucleoside standard, however, this remains the only report of m^2,2,7G released in free ribonucleoside, and no m^2,7G ribonucleoside has been reported in the literature. It is possible that the enzyme preparations used in 1970s were co-purified with contaminants capable of cleaving the 5′–5′ triphosphate bridge, allowing the alkaline phosphatase to further dephosphorylate m^2,2,7Gp to its ribonucleoside, similar to the well-documented contamination of commercial alkaline phosphatases with adenosine deaminase, which causes artifactual deamination of adenosine to inosine in nucleoside analyses [22]. Further experiments will be needed to confirm the presence of m^2,7G and m^2,2,7G in RNA hydrolysates. As m^2,2,7G cap has been reported in multiple instances, we assign m^2,2,7G cap as an Authenticated modification, while m^2,7G cap as a Provisional modification.

Benarroch et al. showed that the catalytic domain of human TGS1 is capable of two separable transmethylation steps: conversion of m⁷G to m^2,7G, and conversion of m^2,7G to m^2,2,7G [87]. Using purified recombinant enzyme and a panel of synthetic cap-like nucleotide and dinucleotide substrates, they found that TGS1 requires prior guanine-N⁷ alkylation for both methylation steps and was completely unreactive toward unmethylated G or N²-monomethylated nucleotides. Functionally, m^2,2,7G is well established as playing roles in splicing regulation and long RNA editing, and the cap methylation status influences RNA turnover, nuclear export, and translational competence [11]. By contrast, no biological function has been assigned specifically to m^2,7G-capped transcripts.

5-Methyl-2-thiouridine (m⁵s²U) and 5-carbamoylmethyl-2′-O-methyluridine (ncm⁵Um). Both m⁵s²U and ncm⁵Um have been mentioned in the literature as human RNA modifications, but the supporting evidence for each is weak. The claim that m⁵s²U occurs in mammalian tRNA rests on a single study by Kimura-Harada et al. (1971), who reported detecting m⁵s²U in rat liver tRNA^Glu and tRNA^Lys using TLC with UV detection, identified by comparing its migration profile to a synthetic standard [88]. MODOMICS lists m⁵s²U as a naturally occurring modification in eukaryotic tRNA, installed by TrmU (Homo sapiens) [20]. However, we were unable to find any additional, independent evidence to corroborate its presence in human RNA. Similarly, ncm⁵Um is listed as a human cytoplasmic tRNA modification in a review by Chujo et al., formed by 2′-O-methylation of ncm⁵U at position 34 by FTSJ1-WDR6 [89]. Notably, we could not locate any primary reference either within this review or in the literature that demonstrates the presence of ncm⁵Um at the anticodon wobble position of human tRNA. Taken together, both modifications appear in secondary sources without verified primary sources establishing their occurrence in human RNA. Further investigation is needed to clarify whether these two modifications genuinely exist in human RNA.

Wybutosine (yW) and derivatives. yW is one of the most extensively characterized hypermodified nucleosides in eukaryotic tRNA, sitting at position 37 in tRNA^Phe [90]. Its bulky, rigid tricyclic structure stabilizes the codon-anticodon interaction and helps maintain the reading frame during translation, preventing frameshifting on Phe codons [91]. While yW itself was first structurally defined in yeast, several publications over the years have suggested that yW is also present in human tRNA, likely by analogy to the yeast yW biosynthetic pathway, as direct evidence confirming its presence in human cells has not been reported [58,92]. Work from Noma et al. clarified this ambiguity by showing that human TYW5 recognizes the N4-methyl group on yW-72 and hydroxylates the β-carbon of the propyl side chain to form OHyW-72, which TYW4 then further methylates and methoxycarbonylates OHyW-72 to form OHyW [18,93]. Because hydroxylation occurs before methylation and methoxycarbonylation of the side chain, yW is unlikely to form as a discrete intermediate along the normal human pathway. Unmodified yW was only detected in TYW5 knockout cells in this study, where TYW4 acts directly on yW-72 in the absence of hydroxylation on the propyl side chain, producing yW rather than OHyW. Moreover, the side-chain hydroxyl group of OHyW was also reported to undergo further peroxidation by unidentified enzymes to form peroxywybutosine (o2yW) in human tRNA [18,94]. Besides OHyW-72, MODOMICS has listed another intermediate OHyWy in the biosynthesis of OHyW, i.e., OHyW-72 (named as OHyWx in MODOMICS) → OHyWy → OHyW [20], however, we are not able to locate the primary reference for this intermediate. On this basis, we assign yW and OHyWy as Putative modifications.

5. Overview of the Current Landscape of Human Transcriptome (as of July 2026)

Besides the 24 modifications critically evaluated here in their original publications, a substantial number of RNA modifications summarized in earlier review articles and modification databases (e.g., MODOMICS) have not been subjected to this level of systematic scrutiny. To extend our assessment to this broader landscape, we applied the same scoring framework to these RNA modifications and listed their assigned confidence tiers in Table 2. Taken together, 76 total modifications represent the current state human epitranscriptome as of July 2026.

Concluding Remarks and Perspectives

Advances in RNA enrichment, chemical derivatization, and MS sensitivity have enabled the detection of modifications present at ≤ 1 per 10⁶ nt, while also lowering the threshold for reporting modifications whose chemical identity, biological origin, or reproducibility remain uncertain. The analytical validation framework introduced in this review is intended to provide the field with a practical guide for assessing how confidently we should assume that a reported modification exists in humans, and to clarify the evidence that should be presented when researchers submit a manuscript describing a novel modification. As a resource for the field, we have critically evaluated 24 modified ribonucleosides that expand the landscape of the human epitranscriptome to 76 modifications (Supplementary Table 1).

Several patterns emerged from applying the validation framework (Table 1) across the modifications we evaluated. First, most of the modifications discussed lack independent replication by at least one other research group, with the entirety of the evidence for their existence in human RNA traces back to a single publication. This does not necessarily imply a reliability concern, but it does reflect the inherent difficulty of detecting, characterizing, and mapping modifications present at ultra-trace abundance or in unstable or intermediate forms in cells. As a field, there is an urgent need to establish a public repository of synthetic standards and raw MS spectra for novel modifications, which would substantially accelerate the pace at which Provisional and Putative modifications can be orthogonally validated.

Second, most of the modifications discussed are mapped to bulk RNA fractions, with their transcriptomic coordinates at single-base or fragment resolution yet to be precisely determined. Chemical identity validation and site-specific mapping inherently require orthogonal approaches, instruments, and platforms. While it is understandable that not every laboratory is equipped with a full suite of instruments and methodologies to generate the full body of evidence needed to authenticate a novel modification, we encourage collaboration among research groups to strengthen the integration of ribonucleoside-level LC-MS analysis with LC-MS-based RNA-seq or LRS platforms within the same study. Through such collaborative endeavors, we expect the long tail of Provisional and Putative modifications to be reclassified into the Authenticated tier, with others identified as technical artifacts.

Finally, there is an evident gap between what is technically detectable and what is biologically present, as not every new addition to the epitranscriptome has clear genetic validation. Most of the modifications discussed here were confirmed by matching retention time, exact mass, and fragmentation pattern of a candidate peak to its synthetic standard, but comparatively few were also shown to disappear upon knockdown or knockout of a candidate writer enzyme. Without this genetic layer of evidence, a chemically well-defined peak in RNA cannot be fully distinguished from one arising through a non-enzymatic process, a co-purifying contaminant, or non-human origin RNA, most notably from microbial sources such as mycoplasma, a common and frequently overlooked contaminant of cultured human cell lines. We therefore encourage researchers to routinely inspect and report contamination controls alongside genetic perturbation of the candidate writer or eraser enzyme, as complementary lines of evidence to the identification and mapping of a modification.

Taken together, we hope the validation framework, the tiered list, and the priorities outlined above provide the field with both a practical checklist for reporting new modifications and a shared benchmark for revisiting existing ones, so that the human epitranscriptome can be built upon with confidence. Once these epitranscriptomic marks are validated, studies on their biological functions and disease associations will bloom, thus accelerating research and development toward their clinical and therapeutic utilities.