The appeal of nanopore sequencing rests on a simple physical principle. As a native RNA molecule threads through a protein nanopore embedded in an electrically insulated membrane, it disturbs the ionic current flowing through the pore. Because each combination of nucleotides, known as a k-mer, produces a characteristic current signature, the raw squiggle of electrical data can in principle be decoded into sequence. Crucially, because the molecule is read directly rather than converted to complementary DNA and amplified, chemical modifications such as N6-methyladenosine (m6A), 5-methylcytosine (5mC) and pseudouridine leave their own fingerprints in the current trace. This is a capability that short-read platforms like Illumina fundamentally lack, since their reverse transcription and polymerase chain reaction steps erase the very epitranscriptomic marks researchers want to find.

The review traces how the technology matured from the portable MinION released in 2015, through the high-throughput PromethION and the budget-friendly Flongle adapter, to successive chemistry generations from R9.4.1 to the R10 series and the newer RNA004 flow cells. Equally important has been the evolution of the computational pipeline. Base calling, the conversion of raw current into nucleotide letters, has progressed from Hidden Markov Models to LSTM-based recurrent neural networks in ONT’s Guppy, through convolutional architectures in Bonito, to Dorado, which the authors note has been reported to reach around 99.5 percent raw read accuracy. Alignment tools have followed a parallel trajectory, from GraphMap’s gapped seeding strategy through the now-ubiquitous Minimap2, which the study describes as three to four times faster than mainstream short-read mappers at comparable accuracy, to specialists such as Winnowmap2, which achieved false-negative and false-positive rates of just 1.89 percent each on repetitive reference sequences, far outperforming Minimap2’s 39.62 and 5.88 percent respectively in the cited evaluation.

Between base calling and modification detection sits the re-squiggling step, where raw signals are re-aligned to the called sequence so that per-base current features can be extracted. The review dissects Tombo’s five-stage implementation, spanning genome mapping, signal normalization, event detection, sequence-to-signal assignment and the resolution of skipped bases, and notes that most re-squiggling tools share this conceptual skeleton. Quality control software such as NanoPlot, pycoQC, MinIONQC and ToulligQC then guards the integrity of the data before any biological interpretation begins. Together, the authors argue, this pipeline has matured into a sophisticated, community-driven framework that has steadily improved accuracy, scalability and accessibility across a wide range of applications.

The heart of the paper is its survey of modification detection methods, which have evolved through three distinct technological waves. The first wave relied on classical statistics: Nanocompore, xPore and Yanocomp employed Bayesian generative models built on Gaussian mixture models, while DiffErr and DRUMMER used G-tests and ELIGOS applied Fisher’s exact tests to compare error profiles between modified and control samples. The second wave brought machine learning and deep learning: Nanom6A applied XGBoost for single-base m6A detection, MINES used random forests on four specific sequence contexts, EpiNano turned to support vector machines, and m6Anet deployed a neural network for transcriptome-wide m6A quantification. Pseudouridine specialists emerged in parallel, with Penguin combining several machine learning models to reach 93.38 percent accuracy on a HEK293 benchmark, and NanoMUD using bidirectional LSTMs to detect both pseudouridine and its therapeutically important cousin N1-methylpseudouridine.

The third and most recent wave is the most ambitious. Multi-modification detectors such as ModiDeC, which identifies m6A, pseudouridine, Gm, m1A and inosine in a single pass, and RNANO, which leverages attention-enhanced multi-instance learning to detect seven modification types, are pushing toward comprehensive epitranscriptomic profiling. Perhaps most transformative is the emergence of modification-aware base callers, exemplified by m6ABasecaller and ONT’s own Dorado, which now supports modified base calling for m6A and pseudouridine directly. These tools fold modification prediction into the base calling step itself, bypassing computationally expensive re-squiggling and post hoc statistics, and can operate without matched control samples or prior motif knowledge. The catch, the authors caution, is that such models demand training data with precise ground-truth annotations of modification status at individual nucleotides, a resource that remains scarce.

To test whether this proliferation of methods translates into reliable science, the team benchmarked five m6A detection tools, m6Anet, MINES, nanom6A, DRUMMER and DiffErr, against an NGS-derived reference database stratified into four confidence tiers, using the HEK293_WT and HMEC_WT datasets from GEO accession GSE132971. The discrepancies were dramatic. In HEK293_WT, m6Anet predicted 36,688 m6A sites and MINES 35,870, while DRUMMER reported only 36 and DiffErr just 3. Even among the three prolific tools, only 5,088 sites were commonly detected, and a large share of their predictions mapped to the low and very low confidence tiers of the reference: m6Anet, for instance, placed 15.5 percent of its calls in the low group and 22.6 percent in the very low group, with only 1.6 percent reaching high confidence. The pattern repeated in HMEC_WT, where the three leading tools shared just 2,806 sites.

The pseudouridine benchmark was even more sobering. NanoMUD predicted 82,581 pseudouridine sites in HEK293_WT, yet only 57 of them, about 0.1 percent, overlapped with the NGS-based reference, and NanoPsu managed just 31 overlaps. In HMEC_WT, Penguin predicted a staggering 644,248 sites but overlapped the reference at only 74 positions. Cross-tool concordance was similarly dismal, with NanoMUD and NanoPsu sharing only 396 sites in the first dataset and Penguin and NanoPsu agreeing on a single site in the second. The authors conclude that inconsistencies in pseudouridine identification are even more pronounced than for m6A, underscoring an urgent need for improved cross-tool consistency and methodological refinement.

Why does this matter beyond the technical community? RNA modifications are not decorative. m6A, the most prevalent internal modification in eukaryotic messenger RNA, regulates splicing, nuclear export, translation and mRNA stability, and its dysregulation has been implicated in heart failure, cancers and reproductive disorders. Pseudouridine enhances RNA stability and base stacking, and it is the key modification in chemically stabilized mRNA vaccines; NanoMUD has already been applied to quality control in vaccine design contexts. The review also contrasts nanopore approaches with alternatives: liquid chromatography-mass spectrometry offers accurate quantification but no sequence information, antibody-based methods lack single-nucleotide resolution, and newer NGS techniques such as GLORI and eTAM-seq achieve absolute m6A quantification but remain constrained by short reads and indirect chemistry. Nanopore’s direct, long-read, single-molecule approach remains uniquely positioned, but only if the computational interpretation of its signals can be trusted.

The authors are candid about the remaining obstacles. Nanopore error rates still exceed the 0.1 to 1 percent benchmark of Illumina short reads, and because some pipelines infer modifications from base calling errors, this intrinsic noise directly undermines detection. Direct RNA sequencing also demands substantial input, with the RNA004 protocol specifying 300 nanograms of poly(A) RNA or a microgram of total RNA, a barrier for clinical biopsies and single-cell work. Throughput remains modest at one to three gigabases per flow cell, and reliable modification detection requires coverage of at least 30-fold, while deep learning pipelines demand serious GPU infrastructure. Yet the trajectory is encouraging: multiplexed direct RNA protocols, real-time selective sequencing tools like SquiggleNet, signal-level mappers like Sigmap, and the integration of large language models and ensemble learning all point toward a future in which the epitranscriptome can be read as directly, and as reliably, as the genome itself. For now, the study’s message to researchers is clear: choose your tools carefully, validate against orthogonal methods, and treat any single algorithm’s modification calls with healthy skepticism.

Subject of Research: Benchmarking computational tools for detecting RNA modifications using Oxford Nanopore direct RNA sequencing

Article Title: Nanopore direct RNA sequencing for RNA modification analysis: workflow assessment and computational tool benchmarking

Article References: Wu, Z., Li, J., Xia, R., Dai, J., Su, J., Meng, J., & Zhang, Y. (2026). Nanopore direct RNA sequencing for RNA modification analysis: workflow assessment and computational tool benchmarking. Advanced Biotechnology, 4(1), Article 9. https://doi.org/10.1007/s44307-025-00093-5

Image Credits: AI Generated

DOI: 10.1007/s44307-025-00093-5

Keywords: nanopore sequencing, direct RNA sequencing, RNA modifications, m6A, pseudouridine, epitranscriptomics, base calling, machine learning, deep learning, benchmarking, Oxford Nanopore Technologies, computational tools