Improved detection of RNA changes using direct nanopore sequencing

The central dogma of molecular biology outlines the directional flow of genetic information from DNA to RNA to protein.^[34^] Beyond transcriptional and translational control, post‐transcriptional RNA modifications have emerged as pivotal layers of gene regulation. RNA modifications are chemical alterations to RNA nucleotides that do not change the underlying sequence, which are collectively referred to as the "epitranscriptome". These modifications play critical roles in regulating RNA structure and metabolism, including mRNA stability, translation, splicing, and export.^[35^] To date, over 170 distinct RNA modifications have been identified across various types of RNA species, including mRNAs and non‐coding RNAs (rRNAs, tRNAs, snRNAs, and lncRNAs).^[3^]

Among these modifications, m⁶A, formed by methylation at the nitrogen‐6 position of adenosine, is the most abundant internal modification in eukaryotic mRNAs. It plays critical roles in regulating mRNA stability, splicing and export, thereby influencing key cellular processes.^[6^] It is deposited by a methyltransferase complex comprising METTL3, METTL14, and WTAP, and can be removed by demethylases such as FTO and ALKBH5.^[36, 37, 38^] m⁵C, originally known as a DNA modification, also occurs at low levels in eukaryotic mRNAs. It refers to the methylation of the fifth carbon of cytosine in RNA nucleotides. Although m⁵C represents a small fraction of total cytidines in mRNAs, it plays important regulatory roles, particularly in mRNA stabilization and export.^[39^] ac⁴C is a conserved RNA modification catalyzed by NAT10.^[40^] It was initially characterized in tRNAs and rRNAs but has more recently been identified in human mRNAs. In coding sequences, ac⁴C enhances translation efficiency by stabilizing codon‐anticodon interactions, whereas its presence in 5′ UTRs may inhibit translation initiation.^[41, 42^] Nm is a widespread RNA modification present at both cap‐adjacent and internal positions of mRNAs, where it influences mRNA stability, translation, and splicing.^[43, 44^] Internal Nm can inhibit translation elongation and modulate splicing, whereas cap‐adjacent Nm (Cap‐Nm) serves as a self‐recognition marker to evade innate immune sensing.^[45, 46^]

m⁷G is a critical RNA modification located at the 5′ cap of mRNAs and internal sites of tRNAs, rRNAs, and some miRNAs, playing essential roles in mRNA stability, translation initiation, and nuclear export. Its deposition is mediated by writer complexes such as METTL1/WDR4 in tRNAs and RNMT/RAM in mRNA caps, while QKI proteins have been identified as m⁷G readers.^[47, 48^] Dysregulation of m⁷G methyltransferases leads to aberrant m⁷G methylation and contributes to cancer initiation and progression by affecting gene expression and oncogenic pathways.^[47, 49^] m¹A is a conserved RNA modification at the N¹ position of adenosine that occurs in tRNAs, mRNAs, rRNAs, and lncRNAs. It influences RNA structure and translation.^[50^] The aberrant activity of its dedicated writers, erasers, and readers is increasingly linked to oncogenic proliferation, metabolic reprogramming, and drug resistance, positioning the m¹A machinery as a promising source of both biomarkers and therapeutic targets.^[51, 52^]

Inosine, generated via A‐to‐I RNA editing by ADAR enzymes, introduces transcriptomic diversity by modulating splicing, mRNA stability, and protein function.^[53^] Inosine alters splicing by modifying splice sites and regulates mRNA stability through interactions with miRNAs and RNA‐binding proteins. These regulatory mechanisms play essential roles in brain development, immune regulation, and disease pathogenesis.^[53, 54, 55^] Ψ is catalyzed by multiple pseudouridine synthases and is widely distributed in mRNAs. It influences translation read‐through, alternative splicing, and transcript stability in a context‐dependent manner.^[56^]

Like mammals, m⁶A is the most abundant internal modification in plant mRNAs. It is deposited by a conserved writer complex containing MTA, MTB, VIR, FIP37, and HAKAI, and can be dynamically removed by ALKBH10B and ALKBH9B.^[57^] m⁶A regulates various aspects of mRNA metabolism, including stability, alternative polyadenylation, secondary structure, and translation efficiency, and plays essential roles in biological processes such as embryogenesis, meristem maintenance, floral transition, and responses to environmental stress.^[57^] Ψ is widely distributed across different plant RNAs. It is catalyzed by pseudouridine synthases, including RluA, RsuA, TruA, TruB, PUS10, and TruD.^[58^] Ψ enhances RNA secondary structure and modulates RNA‐protein interactions, thereby impacting ribosome biogenesis, splicing fidelity, and plant stress responses in Arabidopsis thaliana.^[59^] m⁵C is present in diverse RNA species, including mRNAs, tRNAs, rRNAs, and non‐coding RNAs in Arabidopsis thaliana and Oryza sativa, and is catalyzed by conserved methyltransferases such as TRM4B, NSUN2, NSUN5, and TRDMT1.^[60, 61^] m⁵C enhances transcript stability to promote translation efficiency, and facilitates long‐distance RNA transport, particularly for mobile mRNAs critical for inter‐organ communication.^[61, 62^] Dynamic m¹A RNA methylation is widespread in mRNAs, lncRNAs, and circular RNAs (circRNAs), and acts as a key epitranscriptomic regulator of fruit ripening in tomato.^[63, 64^] ac⁴C has recently emerged as a conserved epitranscriptomic marker in angiosperms. It is enriched in coding sequences, particularly near translation start sites, suggesting potential roles in RNA stability, splicing, and translation.^[65^] In plants, three homologs of the human NAT10 acetyltransferase are responsible for catalyzing the N⁴‐acetylation of cytidine (ACYR1/ACYR2 in Arabidopsis^[66^] and OsNAT10 in rice^[67^]). Loss‐of‐function mutants in these genes dramatically reduce ac⁴C levels, and lead to developmental defects, highlighting the essential roles.^[67, 68^] Functionally, ac⁴C promotes the translation of specific transcripts in Arabidopsis thaliana and Oryza sativa, such as members of the light‐harvesting complex (LHC) gene family (LHCB2, LHCB3, LHCB4), and the OsLIR1, a key regulator of photosynthetic electron transfer.^[67, 68^] In response to pathogen infection, OsNAT10 acetylates OsAOC mRNA, which encodes a core enzyme in jasmonic acid biosynthesis, thereby boosting its translation and rapidly activating jasmonate‐mediated immune responses.^[69^] Nm and m⁷G have also been detected in Arabidopsis thaliana rRNAs, but their functions remain unclear.^[70, 71^]

Over the past decade, multiple methodologies have been developed to detect RNA modifications across the transcriptome, each with distinct principles and limitations. Current sequencing strategies for RNA modification detection are commonly classified into four major categories.^[72^] Antibody‐based methods, which are widely adopted for profiling marks such as m⁶A, m⁵C, m⁷G, and ac⁴C, depend on modification‐specific immunoprecipitation.^[40, 48, 62, 73, 74^] However, these approaches typically require large amounts of RNA input and often suffer from limited specificity, batch variability, and lack stoichiometric resolution.^[22^] Reverse transcription signature‐dependent techniques detect RNA modifications through their impact on cDNA synthesis, including the induction of mismatches, premature terminations, or insertions and deletions at modification sites.^[75, 76^] Although relatively straightforward and cost‐effective, these methods are susceptible to false positives due to sequence polymorphisms, sequencing artifacts, and library preparation biases.^[19^] Enzyme‐dependent methods utilize the substrate recognition or cleavage specificity of certain demethylases, endonucleases, or exonucleases to distinguish modified from unmodified nucleotides. These methods are constrained by enzymatic efficiency and context‐dependent sequence preferences. Finally, chemical‐assisted approaches combine selective chemical labeling or derivatization of modified bases with next‐generation sequencing (NGS). These strategies have recently enabled base‐resolution identification and quantification of marks such as m⁶A, Ψ, m⁵C, and ac⁴C.^[77^] Nonetheless, the effectiveness of these methods hinges on reaction conditions that must preserve RNA integrity, which remains a technical challenge.

Nanopore sequencing is a real‐time, single‐molecule sequencing platform that determines nucleic acid sequences by measuring changes in ionic current as individual RNA or DNA molecules translocate through a nanopore embedded in an electrically resistant membrane^[78^] (Figure2). An applied voltage generates an electrochemical gradient that drives nucleic acids through a biological pore,^[79^] while a processive motor protein unwinds and ratchets the strand through the pore at a controlled speed. As the molecule moves forward through the pore, specific combinations of nucleotides (typically 5‐mers) occupy the pore's constriction site and induce characteristic disruptions in the ionic current.^[80^] These current signals are recorded with high temporal resolution and decoded into nucleotide sequences using traditional statistical methods (such as hidden Markov models) or machine learning (ML) methods (such as deep neural networks).^[80, 81^]

In DRS, native RNA molecules are sequenced directly without reverse transcription or amplification, thereby preserving endogenous modifications and avoiding cDNA synthesis biases. Each read retains both sequence and modification‐sensitive signal features, making it possible to analyze transcriptomes while simultaneously capturing information about RNA modifications.^[78, 80^] This fundamental principle forms the basis for the method's growing utility in epitranscriptomics, enabling a more direct and comprehensive investigation of the transcriptome landscape than conventional NGS approaches.

To systematically review recent advances, we first discuss computational algorithms for detecting m⁶A modifications from DRS data (Table1). The core challenge tackled by this diverse and rapidly evolving toolkit is deciphering the subtle signal perturbations caused by m⁶A. Consequently, these methods employ a wide spectrum of computational strategies ranging from statistical tests to sophisticated deep learning architectures, leveraging different signal features to achieve objectives such as site detection or stoichiometry quantification. The following sections provide a detailed exposition of the methodologies, underlying principles, and key innovations that define these significant computational frameworks (Figure3).

While m⁶A has received the most extensive attention in nanopore‐based epitranscriptomic research, a growing body of work has begun to address the detection of other RNA modifications using DRS technology (Figure 3). These include Ψ, Nm, m¹A, 5moU, and adenosine‐to‐inosine (A‐to‐I) editing, each with distinct chemical properties and biological functions. Detecting such modifications poses unique challenges due to their subtle effects on ionic current signals and limited training data. Recent advances in computational methods, particularly those employing supervised and unsupervised ML, have enabled the development of models capable of profiling these modifications with high sensitivity as well as resolution, and in many cases, without the need for chemical derivatization, antibody enrichment, or genetic knockouts. This section highlights representative tools and strategies designed to expand the modification landscape accessible through DRS, underscoring their potential to uncover novel regulatory layers in RNA biology (Table2).

ModQuant is a supervised ML framework designed for accurate and quantitative profiling of Ψ modifications in mRNA using DRS data.^[103^] Trained on synthetic RNA standards containing site‐specific Ψ modifications, ModQuant leverages a comprehensive set of 45 signal‐ and base‐calling‐derived features extracted from ±2 nucleotides surrounding each candidate site. The model demonstrates high predictive accuracy across a wide range of sequence contexts and was robustly validated in seven human cell lines. By coupling synthetic controls with interpretable feature engineering, ModQuant establishes a generalizable and quantitative pipeline for Ψ detection, offering a scalable strategy for epitranscriptomic analyses of native transcriptomes.

NanoPsu is an ML framework for the transcriptome‐wide detection of Ψ modifications from DRS data.^[104^] It employs extremely randomized trees (ExtraTrees, EXT) classifier trained on native ribosomal RNA reads, labeled using Illumina bisulfite sequencing, to identify Ψ‐modified sites. Unlike signal‐based methods, NanoPsu relies solely on 12 engineered features derived from base‐calling error profiles, achieving high predictive performance (AUC ∼0.94). Notably, it enables single‐read stoichiometry estimation and inter‐site linkage analysis, offering insights into modification co‐occurrence patterns along individual transcripts. By eliminating the need for synthetic training controls and direct current signal modeling, NanoPsu provides a quantitative, molecule‐resolved approach to Ψ profiling and represents a key innovation in DRS‐based epitranscriptomic analysis.

Penguin is an ML‐based tool developed for the prediction of Ψ sites from ONT DRS data in human cell lines.^[105^] The framework integrates features derived from both raw nanopore signal data and base‐called k‐mers, enabling accurate Ψ site identification at single‐nucleotide resolution. Penguin achieved high predictive performance (>92% accuracy) across both random data splits and independent validation sets from the HEK293 and HeLa transcriptomes. By offering a comprehensive and scalable pipeline tailored to long‐read RNA data, Penguin addresses the limitations of earlier approaches confined to short‐read sequencing or non‐human models, thereby advancing Ψ detection in human transcriptomic studies.

IndoC is an informatics method for detecting Ψ modifications from nanopore DRS data and is designed to operate without the need for genetic knockouts or synthetic controls.^[106^] The approach performs internal comparisons of trace‐derived values, ionic current intensities, and their statistical distributions between candidate sites and identical 5‐mer sequence contexts within the same dataset. These context‐aware features are integrated into a support vector machine (SVM) classifier to distinguish Ψ‐modified sites robustly from high‐mismatch or noisy positions. By eliminating external controls, IndoC offers a practical, scalable, and species‐agnostic strategy for Ψ mapping, facilitating broad application across diverse biological conditions and sample types.

Mod‐p ID is an innovative computational framework for transcriptome‐wide detection of Ψ modifications using DRS data, eliminating the need for chemical derivatization or cDNA conversion.^[107^] Its core innovation lies in quantifying the Ψ occupancy across cell types by directly analyzing native RNA ionic current signatures. Mod‐p ID compares U‐to‐C base‐calling error rates between native and in vitro transcribed RNA using Fisher's exact test, supplemented by motif analysis to increase site‐level confidence. This approach enables robust, comparative Ψ profiling of native transcriptomes, offering a streamlined and scalable strategy for high‐resolution epitranscriptomic analysis.

NanoPsiPy is a computational tool for the identification and quantification of Ψ modifications from nanopore DRS data, based on statistical analysis of U‐to‐C base‐calling error rates.^[108^] NanoPsiPy enables direct, stoichiometric mapping of Ψ sites in native human transcriptomes without the need for additional treatments. Beyond site‐level detection, this tool uncovers functional interplay between Ψ and other RNA modifications, providing new insights into epitranscriptomic cross‐talk. This streamlined, non‐invasive approach offers a scalable framework for transcriptome‐wide Ψ profiling, with potential applications in elucidating RNA modification dynamics in health and disease.

m¹A‐prediction is the first dedicated computational workflow for the quantitative profiling of m¹A modifications using DRS data.^[109^] The method extracts electrical signal features such as the mean current, standard deviation, and event duration, and applies machine learning models, including XGBoost, trained on in vitro transcribed RNA datasets to predict m¹A sites at single‐molecule resolution. Validation via binomial testing and downstream functional analyses in HEK293 cells revealed putative roles for m¹A in RNA splicing and neurodegenerative pathways. While the tool represents a significant advance in DRS‐based m¹A detection, its reliance on synthetic training data may constrain generalizability across diverse biological contexts.

Dinopore is a DL‐based framework for transcriptome‐wide identification of A‐to‐I RNA editing sites from DRS data.^[110^] Unlike traditional alignment‐dependent methods, which are often confounded by SNPs and mapping artifacts, Dinopore directly models raw ionic current signals, base‐calling errors, and quality metrics using a multi‐dimensional CNN. This architecture enables accurate, single‐nucleotide resolution detection of inosine without requiring matched DNA sequencing or prior genome annotations. When trained on datasets from wild‐type and ADAR1‐knockout human stem cells and validated across multiple species, Dinopore achieves high predictive performance (AUC up to 0.98) and robust cross‐species generalizability. This approach represents a major methodological innovation, offering a robust, annotation‐free solution for RNA editing analysis and broadening the functional utility of nanopore DRS in epitranscriptomics.

Nm‐Nano is an ML‐based framework for the transcriptome‐wide detection of Nm sites from nanopore DRS data at single‐molecule resolution.^[111^] The method integrates XGBoost and RF classifiers trained on signal‐level features and k‐mer embeddings, achieving high predictive accuracy (AUC up to 0.99) in human cell lines. By leveraging long‐read sequencing and advanced ML techniques, Nm‐Nano enables direct, high‐resolution mapping of Nm modifications, highlighting the potential of DRS‐based approaches for exploring Nm dynamics across diverse biological contexts.

NanoML‐5moU is an ML‐based framework for read‐level detection of 5moU modifications from nanopore DRS data.^[112^] By extracting signal‐derived features including mean, median, standard deviation, and dwell time from k‐mers centered on U or 5moU residues, the method achieves single‐read resolution. Leveraging classifiers such as XGBoost, NanoML‐5moU attains high predictive accuracy (AUC up to 0.9567) without requiring matched control samples. This represents a methodological advancement in DRS‐based RNA modification detection, enabling de novo identification of 5moU and providing a valuable tool for quality control in synthetic mRNA therapeutics.

These emerging tools leverage a range of machine learning strategies trained on synthetic standards, native reads, or paired control datasets to enable RNA modification profiling at both single‐nucleotide and single‐molecule resolution. Supervised learning frameworks such as ModQuant, NanoPsu, and Penguin have demonstrated high accuracy for Ψ detection by utilizing signal‐derived or base‐calling error features. In parallel, label‐free approaches including IndoC, Mod‐p ID, and NanoPsiPy exploit intrinsic current distortions and base‐calling discrepancies to identify Ψ sites without the need for knockout models or synthetic controls. Expanding beyond Ψ, Chen et al. developed a workflow for m¹A detection using XGBoost applied to signal‐based features, while Dinopore employs a deep convolutional neural network to detect A‐to‐I editing directly from raw ionic signals, circumventing genome alignment. Other specialized frameworks, such as Nm‐Nano and NanoML‐5moU, target Nm and 5moU, respectively, using ensemble learning models (e.g., XGBoost, RF) to achieve stoichiometric, high‐resolution detection. Collectively, these advances mark the maturation of nanopore DRS‐based epitranscriptomic profiling, offering scalable, quantitative, and modification‐specific alternatives to conventional chemical‐ or antibody‐based assays.

As our understanding of the epitranscriptome deepens, it is increasingly evident that RNA modifications do not act in isolation but often interact in complex, combinatorial patterns to regulate transcript fate and function. This emerging concept of modification cross‐talk, where one modification influences the presence, function, or interpretation of another, has motivated the development of computational frameworks capable of simultaneously detecting multiple RNA modifications based on nanopore DRS data (Figure 3).

Unlike traditional single‐modification tools, these integrative approaches leverage the rich, multidimensional signal space of DRS and apply multi‐label learning or statistical modeling to resolve overlapping or co‐occurring modification signatures (Table3). By capturing co‐modification pattern interactions, these frameworks enable a more holistic view of transcript regulation and open new avenues for studying the layered architecture of the epitranscriptome. This section highlights recent efforts toward joint detection strategies and their potential to uncover mechanistic insights into RNA modification networks in health and disease.

ONT has recently expanded official support for RNA modification detection within its basecalling and analysis ecosystem. The Dorado basecaller includes pretrained models for the detection of multiple RNA modifications, including m⁶A, Ψ, m⁵C, inosine, and Nm. These pretrained models are compatible with the updated RNA004 chemistry and the rna004_130bps_sup@v5.2.0 model, enabling direct, per‐read identification of modifications from native ionic current signals.^[113^] To streamline analysis, ONT also provides a modkit, a toolkit that parses modification calls embedded in the basecalled output.^[114^] In parallel, ONT developed Remora, a lightweight inference engine designed to decouple modification calling from canonical basecalling. Remora operates as a secondary pass following initial basecalling, offering targeted detection of m⁶A, Ψ, and inosine with minimal computational overhead. This modular design allows for independent development and deployment of modification models, increasing flexibility and facilitating updates without retraining the basecaller.^[115^] Together, Dorado and Remora represent an integrated framework for high‐throughput, native RNA modification profiling using nanopore direct RNA sequencing.

TandemMod is a transferable DL framework designed for the simultaneous identification of multiple RNA modification types from single‐sample nanopore DRS data at single‐nucleotide resolution.^[116^] Leveraging in vitro epitranscriptome datasets with diverse sequence contexts and high‐confidence labels, the authors trained models to detect m⁶A, m⁵C, and m¹A modifications, and applied transfer learning to extend detection capabilities to hm⁵C, m⁷G, Ψ, and inosine. TandemMod demonstrated robust accuracy across synthetic, human, and plant datasets, underscoring its broad applicability and generalizability. Validation through complementary molecular assays, including MeRIP‐seq, miCLIP, and m6ACE‐seq, confirmed strong concordance between predicted and experimentally supported modification sites. This integrative approach exemplifies the potential of deep learning frameworks to comprehensively profile diverse epitranscriptomic marks from complex biological samples.

xPore is a statistical framework for detecting differential RNA modifications from nanopore DRS data without requiring an unmodified control.^[81^] By modeling per‐site ionic current distributions as two‐component Gaussian mixtures, xPore estimates modification rates using variational Bayesian inference, enabling quantitative and site‐specific detection. The method was validated using METTL3 knockout and knockdown in HEK293T cells, with orthogonal support from multiple experimental platforms, including m6ACE‐seq, miCLIP, MAZTER‐seq, and RNA mixing assays. Compared with both supervised (e.g., EpiNano) and unsupervised (e.g., Tombo, Nanocompore) approaches, xPore demonstrated superior precision, recall, and computational efficiency. Its applicability has been further demonstrated across diverse biological contexts, including human cancer cell lines and clinical multiple myeloma samples, highlighting its robustness and utility for modification profiling from limited or heterogeneous RNA inputs.

nanoDoc is a DL framework designed to detect a wide spectrum of RNA modifications from DRS data using raw ionic current signals.^[117^] Employing a CNN in combination with deep one‐class (DOC) classification, the nanoDoc is trained exclusively on unmodified synthetic sequences and identifies modifications by quantifying deviations from the learned baseline signal. Without requiring labeled modified data or prior knowledge of modification types, the method achieves high performance (AUC = 0.96) across 23 different RNA modifications. Its reference‐free, unsupervised nature allows broad applicability, and the framework has been successfully applied to rRNA, mRNA, and viral transcriptomes, positioning nanoDoc as a versatile tool for de novo epitranscriptomic profiling.

Nanocompore is a model‐free comparative framework for detecting differential RNA modifications from DRS data.^[22^] It identifies modification‐associated signal changes by statistically comparing the ionic current intensity and dwell time distributions between two experimental conditions. When validated on synthetic RNAs bearing seven distinct modifications and on biological samples, including yeast and METTL3‐knockdown human cells, Nanocompore has demonstrated robust performance across diverse contexts. Orthogonal validation using miCLIP and meRIP‐qPCR further supports its reliability. By avoiding assumptions about specific modification types or sequence motifs, Nanocompore provides a flexible and broadly applicable tool for differential RNA modification analysis.

nanoRMS is a single‐molecule analysis framework designed to quantify RNA modification stoichiometry from DRS data.^[118^] It integrates multiple nanopore‐derived signal features, including ionic current intensity, dwell time, and base probability to enable both supervised and unsupervised detection of modifications such as Ψ and Nm. The method was validated using yeast snoRNA knockout strains and benchmarked against nanoCMC‐seq, demonstrating its capacity to detect and quantify non‐m⁶A RNA modifications with high resolution and flexibility. By providing modification stoichiometry at the per‐read level, nanoRMS offers a valuable tool for quantitative epitranscriptomic profiling beyond canonical marks.

Sequoia is an interactive visual analytics platform designed to explore and distinguish RNA modification signals in DRS data.^[119^] By integrating dynamic time warping (DTW) with t‐distributed stochastic neighbor embedding (t‐SNE), Sequoia enables unsupervised clustering of raw ionic current signals, facilitating the identification of RNA modifications such as m⁶A and m⁵C in HeLa cells. The platform incorporates a multi‐view visualization framework, including boxplots, signal traces, and t‐SNE plots with adjustable parameters (e.g., DTW penalty), allowing users to iteratively refine clustering and detect qualitative signal differences between modified and unmodified reads. Sequoia represents a novel computational approach that complements predictive models by offering intuitive, hypothesis‐generating insights into RNA modification‐associated signal features, and may serve as a foundational tool for the future development of automated and interpretable epitranscriptomic analysis methods.

JACUSA2 is a versatile, platform‐agnostic computational framework for detecting diverse RNA modifications by analyzing base substitution, insertion, deletion, and reverse transcription arrest events across sequencing data.^[120^] Compared with both the Illumina and Oxford Nanopore platforms, JACUSA2 enables the detection of modifications such as m⁶A and Ψ through statistical modeling and unsupervised learning, without relying on raw signal‐level data. A notable innovation is its application of non‐negative matrix factorization (NMF) to deconvolute complex modification‐associated patterns from alignment‐derived features, facilitating robust modification calling across experimental conditions. By abstracting modification signals into a platform‐independent representation, JACUSA2 provides a flexible and scalable solution for comparative epitranscriptomic analyses across datasets and sequencing technologies.

NanoSPA is a DL‐based framework designed for the simultaneous detection of m⁶A and Ψ from DRS data.^[121^] The method employs motif‐specific feedforward neural networks trained on orthogonally validated m6A‐SAC‐seq datasets, integrating features from the raw ionic current, dwell time, and base probability (trace) to capture modification‐specific signatures. Through extensive validation across multiple perturbation experiments and external benchmark datasets, NanoSPA demonstrates high accuracy in the joint profiling of two chemically distinct modifications in human transcriptomes. By enabling concurrent analysis of m⁶A and Ψ from a single dataset, NanoSPA provides an important tool for investigating modification co‐occurrence, interplay, and potential cross‐talk in the epitranscriptome.

ELIGOS is a computational framework for detecting RNA modifications from nanopore direct RNA sequencing by leveraging systematic base‐calling error profiles between native RNA reads and control sequences such as in vitro transcribed or cDNA references.^[122^] Rather than relying on ML or pre‐labeled training data, ELIGOS employs statistical modeling of error signatures, termed reference‐based error models (rBEMs) to infer the presence of modified nucleotides based on recurring signal disruptions. This approach enables the detection of multiple RNA modifications, including m⁶A, m¹A, 5moU, and Ψ at single‐base resolution across synthetic and endogenous RNA samples. While the method excels in its label‐free and reference‐agnostic design, it lacks the ability to resolve specific modification types and performs suboptimally for certain marks such as m⁵C, highlighting trade‐offs between generalizability and biochemical specificity.

The NMP classifier represents a novel enzymatic‐nanopore strategy for RNA modification analysis that departs fundamentally from DRS paradigms.^[123^] In this method, RNA molecules are enzymatically hydrolyzed into individual nucleoside monophosphates (NMPs), which are then analyzed using a PBA‐modified MspA nanopore. Each NMP generates distinct current blockade and noise profiles, enabling accurate classification of a broad spectrum of RNA modifications at the mononucleotide level. Unlike DRS‐based tools that infer modifications within intact transcripts, the NMP classifier offers direct biochemical identification of RNA modifications, albeit without positional or sequence context. This approach underscores a powerful orthogonal strategy for comprehensive epitranscriptomic profiling, which is particularly suited for validating modification identities and expanding the catalog of detectable RNA marks.

CHEUI is a DL‐based framework that enables simultaneous, transcriptome‐wide detection of m⁶A and m⁵C modifications from DRS data at single‐molecule resolution, without the need for control samples.^[124^] The method leverages a two‐stage CNN architecture to first classify individual reads as modified or unmodified and then aggregate predictions to infer site‐level stoichiometry. By directly modeling raw signal features, CHEUI overcomes the limitations of motif dependence and modification specificity that constrain many earlier tools. Importantly, CHEUI allows joint profiling of multiple epitranscriptomic marks within the same RNA molecule, revealing co‐occurrence and potential cross‐talk between m⁶A and m⁵C in human transcripts. This ability to detect combinatorial modification patterns represents a significant advance toward integrative, context‐aware epitranscriptomic analysis.

Diensthuber et al. introduced two supervised DL‐based base‐calling models for nanopore DRS data, which were designed to enhance RNA modification detection through improved base‐calling accuracy.^[125^] The first model utilized a flip‐flop neural network architecture trained on unmodified in vitro transcribed RNA, while the second employed a high‐accuracy model based on connectionist temporal classification with conditional random fields (CTC‐CRF). These models were shown to amplify base‐calling error signatures at known modification sites, including m⁶A, m⁵C, hm⁵C, ac⁴C, Ψ, m¹Ψ, and m⁵U, thereby increasing the sensitivity of downstream detection tools such as ELIGOS2. This work underscores the critical role of the basecaller architecture in shaping the interpretability of modification signals, and highlights a new direction for epitranscriptomic method development: optimizing base‐calling itself as a modifiability‐aware analytical step.

NanoMUD is a DL framework that enables simultaneous detection of Ψ and m¹Ψ from DRS data at single‐molecule resolution.^[126^] The method employs bidirectional long short‐term memory (BiLSTM) neural networks trained on synthetic, motif‐specific RNA sequences, capturing nuanced signal features to differentiate between two chemically similar modifications. NanoMUD integrates a regression‐based stoichiometry estimation module, allowing quantitative profiling of modification levels across transcripts. In terms of exceptional accuracy (AUC up to 0.998), NanoMUD represents the first approach capable of distinguishing Ψ from m¹Ψ directly from raw current signals, providing a powerful tool for studying both native RNA biology and synthetic RNA therapeutics, where m1Ψ is commonly used to increase mRNA stability and translation.

iForest+ is a novel ML‐based method designed for the accurate detection of C‐to‐U RNA editing events from DRS data.^[127^] Leveraging the Isolation Forest (IF) algorithm, iForest+ models genuine editing sites as statistical anomalies within basecalling‐derived features such as base quality, mismatches, and indels. This approach distinguishes true editing events from systematic nanopore sequencing errors without relying on raw current signals or matched knockout controls. When validated across synthetic constructs, mouse macrophage cells, and engineered human HEK293T lines, iForest+ demonstrates high precision, efficiency, and generalizability, offering a fast and lightweight tool for robust RNA editing profiling in diverse biological and biomedical contexts.

IL‐AD is an innovative computational tool developed for detecting RNA modifications such as m⁶A, m¹A, and m⁵C from DRS data at single‐nucleotide and single‐molecule resolutions.^[128^] It uniquely combines incremental learning to refine base‐calling accuracy on modification‐rich reads with anomaly detection to identify signal deviations associated with RNA modifications. This approach allows for sequence‐context‐free and cross‐species detection, enhancing both the accuracy and generalizability of RNA modification profiling and marking a significant advancement in epitranscriptomic research.

DRAP3R is not a novel computational model but rather an integrative framework that combines existing nanopore tools to detect m⁶A and Ψ modifications using base‐called probabilities.^[129^] Its key innovation lies in its experimental design, particularly the use of a custom poly(U)‐targeting adapter that selectively sequences Pol III transcribed RNAs. This approach facilitates the study of co‐ and post‐transcriptional RNA modifications in noncoding transcripts that have been largely overlooked, expanding the scope of epitranscriptomic analysis beyond conventional mRNA targets.

Uncalled4 is an advanced signal‐level alignment toolkit designed to improve RNA modification detection from DRS data.^[130^] It introduces a basecaller‐guided dynamic time warping (bcDTW) algorithm alongside a compact BAM‐based data structure to efficiently align and encode per‐position ionic current features such as the mean, standard deviation, and dwell time. Uncalled4 enhances m⁶A detection sensitivity when integrated with tools like m6Anet, and supports m⁵C detection through a KS‐statistics‐based approach. Compared with the latest Oxford Nanopore Technologies chemistries, Uncalled4 outperforms existing methods, identifying ≈26% more m⁶A sites than Nanopolish does in human cell line datasets, thus representing a significant step forward in high‐resolution epitranscriptomic profiling.

modCnet is a DL framework for the simultaneous detection of ac⁴C and m⁵C RNA modifications from DRS data.^[131^] It combines a CNN, bidirectional long short‐term memory networks (biLSTMs), and attention layers to analyze both raw ionic current signals and base‐level statistical features. This integrated approach enables accurate, single‐molecule resolution profiling of cytidine modifications, facilitating insights into their distribution and co‐occurrence within transcriptomes.

Recent advances in nanopore‐based technologies have enabled the simultaneous detection of multiple RNA modifications, overcoming key limitations of traditional enrichment‐based or single‐modification assays. These approaches leverage diverse strategies from deep learning models such as TandemMod, nanoDoc, CHEUI, modCnet, and NanoMUD, to statistical frameworks like xPore, Nanocompore, and ELIGOS, as well as basecaller‐level innovations exemplified by Uncalled4, Diensthuber et al., and m6Abasecaller to jointly profile RNA modifications including m⁶A, m⁵C, Ψ, m¹A, hm⁵C, and ac⁴C at single‐base or single‐molecule resolution. While some tools (e.g., TandemMod, CHEUI) utilize supervised learning trained on synthetic or control datasets, others (e.g., nanoDoc, Nanocompore) adopt unsupervised or comparative approaches that detect modifications without requiring labeled data. Methods such as iForest+ and IL‐AD extend this landscape by leveraging error profiles or signal deviations in native reads to enable modification‐aware detection even in the absence of control samples. Importantly, these frameworks provide critical insights into the co‐occurrence and cross‐talk between different RNA modifications within individual RNA molecules, allowing researchers to dissect complex epitranscriptomic regulation and interplay at unprecedented resolution. Collectively, these advances not only facilitate accurate stoichiometry estimation and co‐modification analysis but also exhibit robust cross‐sample and cross‐species generalizability, significantly advancing the resolution and scope of epitranscriptomic profiling with nanopore direct RNA sequencing data.