Long-read sequencing technologies and bioinformatics: a new perspective for decoding DNA methylation modifications Authors: Hanjing Hou, Yuanfeng Zhang, Yu Ma, Yanxi Han, Jinming Li, Rui Zhang Published: 2026 ABSTRACT AbstractBackgroundDNA methylation is a crucial epigenetic regulatory mechanism in eukaryotes, and its dysregulation is closely associated with numerous diseases. Recent advances in long-read sequencing (LRS) have transformed the ability to comprehensively characterize the methylation modifications in DNA, including the technically difficult genomic regions. However, the translation of this technological potential into reliable biological insights relies heavily on accurate computational analysis of raw signal data. Currently, a growing number of bioinformatic tools have emerged, demonstrating superior performance in LRS-based methylation detection.Main bodyOur review first briefly describes the detection principles and technological improvement of LRS, and then provides a comprehensive overview of the existing LRS-based methylation detection tools and related benchmarking studies. We further explore the applications in biomedical research, the current challenges, and the future perspective of LRS-based methylation detection.ConclusionsBy highlighting the research progress and key issues in the field, this review aims to provide researchers with an essential framework to advance the further development and application of LRS-based methylation detection.Keywords: Long-read sequencing, Epigenetics, DNA methylation Introduction DNA methylation is a key epigenetic modification in eukaryotes, including the most common 5-methylcytosine (5mC), as well as other less common types such as 5-hydroxymethylcytosine (5hmC), N6-methyladenine (6 mA), and N4-methylcytosine (4mC) [1]. These methylation modifications can affect gene regulation without changing the nucleotide sequence, thereby profoundly influencing multiple biological processes, including embryonic development, genomic imprinting, and X chromosome inactivation [2, 3]. And their aberrations are associated with the occurrence and progression of various diseases, such as cancer, neurological disorders, and autoimmune conditions [4–7]. Therefore, accurate detection of DNA methylation is crucial for revealing the fundamental principles of life processes and the molecular mechanisms of disease. While bisulfite sequencing (BS-seq) has long been regarded as the gold standard for DNA methylation detection, it still has significant limitations. The time-consuming and rigorous chemical conversion process leads to DNA damage, reduced sequence complexity, and amplification bias. Furthermore, the method depends on next-generation sequencing (NGS), which generates short reads (only a few hundred base pairs in length), making it difficult to resolve methylation states in genomic repetitive regions (Table 1) [8, 9]. Table 1Comparison of DNA methylation detection approaches using different sequencing technologiesLong-read sequencingShort-read sequencingSequencing technologyPacBio sequencingONT sequencingIllumina sequencing(e.g., BS-seq, EM-seq, TAPS)Data type and read lengthCLR: typically 5-60 kb, up to 200 kbHiFi read: 1-25 kbStandard long read:10-100 kbUltra-long read: typically greater than 100 kb, up to 4 MbShort read:<300 bpRead accuracyHiFi read: >99.9%Single read accuracy: >99%> 99.9%Methylation detection principleDirect detection based on polymerase kinetics(no base conversion)Direct detection based on electrical signals(no base conversion)Indirect detection based on base conversionAdvantages of methylation detectionAbsence of DNA damage introduced by base conversion and following amplification bias from PCR;Simultaneous detection of sequence and multiple DNA modifications;Effective detection in repetitive genomic regions;Reliable methylation detection with low systematic bias.Absence of DNA damage introduced by base conversion and following amplification bias from PCR;Simultaneous detection of sequence and multiple DNA modifications;Effective detection in repetitive genomic regions;Real-time methylation detection during sequencing.High throughput;Low cost;Mature workflow.Main limitations of methylation detectionHigh cost;Shorter read length.Frequent updates of hardware/software.DNA damage;Amplification bias;Reduced sequence complexity;Cannot detect multiple modifications;Cannot detect methylation modifications in repetitive regions of the genome.Abbreviations: BS-seq: bisulfite sequencing; CLR, continuous long read; EM-seq: enzymatic methyl-seq; HiFi, high-fidelity; TAPS: TET-assisted pyridine borane sequencing Comparison of DNA methylation detection approaches using different sequencing technologies Illumina sequencing (e.g., BS-seq, EM-seq, TAPS) CLR: typically 5-60 kb, up to 200 kb HiFi read: 1-25 kb Standard long read:10-100 kb Ultra-long read: typically greater than 100 kb, up to 4 Mb Direct detection based on polymerase kinetics (no base conversion) Direct detection based on electrical signals (no base conversion) Absence of DNA damage introduced by base conversion and following amplification bias from PCR; Simultaneous detection of sequence and multiple DNA modifications; Effective detection in repetitive genomic regions; Reliable methylation detection with low systematic bias. Absence of DNA damage introduced by base conversion and following amplification bias from PCR; Simultaneous detection of sequence and multiple DNA modifications; Effective detection in repetitive genomic regions; Real-time methylation detection during sequencing. High throughput; Low cost; Mature workflow. High cost; Shorter read length. DNA damage; Amplification bias; Reduced sequence complexity; Cannot detect multiple modifications; Cannot detect methylation modifications in repetitive regions of the genome. Abbreviations: BS-seq: bisulfite sequencing; CLR, continuous long read; EM-seq: enzymatic methyl-seq; HiFi, high-fidelity; TAPS: TET-assisted pyridine borane sequencing The emergence of LRS technologies has provided a powerful solution for fundamentally overcoming these challenges. By analyzing enzyme kinetics in Pacific Biosciences (PacBio) sequencing or current signals in Oxford Nanopore Technologies (ONT) sequencing, both nucleotide sequences and methylation information can be obtained simultaneously from native DNA molecules. Since no destructive chemical conversion and PCR amplification are required, DNA damage and systematic biases introduced by these steps are effectively avoided. Moreover, the long reads generated by LRS can reach kilobase (kb) or even megabase (Mb) lengths, thoroughly resolving the methylation landscape in previously inaccessible repetitive regions, and revealing the critical roles of methylation patterns in these regions for maintaining genomic stability and contributing to disease onset (e.g., cancer, neurological disorders) [10–12]. Notably, these advantages have enabled a range of applications that were previously difficult to achieve. For example, methylation phasing enabled by LRS technologies allows the direct, efficient, and accurate identification of parent-of-origin methylation errors in imprinting disorders [13, 14]. Additionally, studies have demonstrated that simultaneous detection of methylation patterns and genetic variants using ONT sequencing enables rapid molecular subtyping of complex brain tumors within hours [15]. These comprehensive and high-resolution analytical capabilities are propelling LRS-based methylation detection to the forefront of biomedical research. However, translating these revolutionary technological potentials into reliable biological discoveries relies on bioinformatics tools capable of accurately interpreting complex raw signals. With the rapid technological evolution of LRS and the continuous expansion of its application scenarios, a series of methylation detection tools, ranging from statistical models to deep learning (DL) models, have been developed [9, 16]. Therefore, a comprehensive review of this rapid development has become crucial. To address this pressing need, with the basics of the detection principles of PacBio and ONT sequencing, we summarize the existing LRS-based methylation detection tools and their characteristics, as well as related benchmarking studies. Furthermore, we explored the potential value of LRS-based methylation detection in biomedical research, outlined the major challenges, and the future perspectives. In summary, this review aims to provide a comprehensive overview of LRS-based methylation detection, driving the optimization of its technologies and algorithms as well as its broader application in biomedical research. Detection principles and technological advances of LRS The core technological advancements of PacBio and ONT sequencing have substantially enhanced their performance, thereby enabling more efficient and accurate detection of DNA methylation. In this section, we describe their detection principles and core technological advancements. PacBio sequencing In 2011, PacBio introduced the PacBio RS, the first commercially available sequencer to enable LRS. The unique circular molecular templates used in PacBio sequencing are called SMRTbells, generated by attaching hairpin adapters to both ends of the target DNA fragment. After binding to the polymerase and primers, these templates are loaded onto the SMRT cell. The surface of the SMRT cell contains millions of tiny wells with nanometer diameters, known as zero-mode waveguides (ZMWs), and each well can accommodate an independent sequencing reaction [17, 18]. During sequencing, the DNA polymerase located at the bottom of the ZMWs catalyzes the incorporation of fluorescently labeled dNTPs, and sequence information is obtained through fluorescence signals. Notably, due to the nanoscale structure of ZMWs, only the fluorescent groups carried by the dNTPs that are bound to the template in the extremely small area at the bottom of the pores can be activated and detected, thereby significantly reducing the interference from background fluorescence [19, 20]. Additionally, since the various modifications affect polymerase kinetics differently, DNA methylation modifications are identified by analyzing unique kinetic signatures, including interpulse duration (IPD) and pulse width (PW) (Fig. 1A) [21, 22]. IPD refers to the time interval between two consecutive nucleotide incorporation events, and PW refers to the duration of the fluorescence signal generated by a single nucleotide incorporation event. Overview of LRS-based methylation detection: principles, major challenges, and proposed solutions. (A) PacBio sequencing detects DNA methylation by analyzing the characteristic changes in polymerase kinetic parameters: IPD and PW. (B) ONT sequencing identifies DNA methylation by detecting the characteristic current signal changes as DNA passes through a nanopore. (C) Overview of major challenges in LRS-based methylation detection (e.g., high sequencing cost, high input requirement, limited modification detection capability, computational burden, lack of standardization) and corresponding solutions. (A: adenine; C: cytosine; cfDNA: circulating cell-free DNA; dNTPs: deoxynucleotide triphosphates; EM-seq: enzymatic methylation sequencing; FFPE: formalin-fixed and paraffin-embedded; G: guanine; IPD: interpulse period; LRS: long-read sequencing; PCR: polymerase chain reaction; PW: pulse width; QC: quality control; SOPs: standard operating procedures; T: thymine; TAPS: ten-eleven translocation (TET)-assisted pyridine borane sequencing; ZMW: zero-mode waveguides; 5hmC: 5-hydroxymethylcytosine; 5mC: 5-methylcytosine). Created in BioRender. Zhang, R. (2026) https://BioRender.com/bzxqlwq Continuous advancement in PacBio sequencing has significantly enhanced its capability for DNA methylation detection. Continuous long read (CLR), as the main data type in the early stage of PacBio sequencing, has a relatively long read length, but its accuracy rate is relatively low, only 85–92% [23]. In 2019, PacBio introduced the circular consensus sequencing (CCS) technology, also known as high-fidelity (HiFi) sequencing technology. It performs multiple sequencing on the same DNA template to generate multiple subreads, which are then integrated through algorithms to produce a HiFi read with long read lengths (1–25 kb) and exceptional accuracy (≥ 99.9%) [24–26]. These highly accurate reads provide a solid foundation for the precise identification of methylation states. For microorganisms, PacBio has successfully mapped over 2,000 bacterial methylomes to date [27]. And HiFi sequencing currently supports the detection of 6 mA, 4mC, and 5mC in microorganisms. In contrast, for humans and other eukaryotes, PacBio 5-base HiFi sequencing currently supports 5mC detection in the CpG context [28]. However, it is worth noting that HiFiMeth, based on HiFi sequencing, was the first to achieve the detection of 5mC in all sequence contexts (CpG, CHG, and CHH). However, current validation mainly focuses on plant species, and its applicability and detection accuracy in human and other mammalian genomes still require further independent research to verify. It has not yet become the mainstream choice for eukaryotic methylation analysis [29]. ONT sequencing In 2014, ONT launched its first nanopore sequencer, MinION, which pioneered portable real-time sequencing [30]. The core principle of ONT sequencing is a biosensing system based on the protein nanopore. By detecting the characteristic current signal of DNA as it passes through the nanopore, sequence and methylation modification information can be resolved simultaneously (Fig. 1B) [31–33]. The continuous improvements in both accuracy and read length of ONT sequencing have collectively enabled high-quality DNA methylation detection. Early ONT sequencing had relatively low average accuracy, with most reads having an accuracy between 87% and 98% [23]. ONT has continuously improved its flow cells to improve the sequencing accuracy [34–36]. In 2022, the R10.4.1 flow cell was released, compatible with Kit 14 chemistry. Using the latest product portfolio, it achieves > 99% single-read accuracy (Q20 chemistry), outperforming the R9.4.1 [37, 38]. Improved accuracy directly enhances the reliability of identifying methylation modifications from raw electrical signals. Meanwhile, ONT sequencing retains its inherent advantage of ultra-long read lengths, theoretically constrained only by the length of the DNA molecule itself. To date, the longest read length reported has exceeded 4 Mb [39]. Notably, the ultra-long read feature of ONT sequencing helps to construct more complete genome assemblies, which allows for better probing of the epigenetic profile of the human genome. In 2022, based on T2T-CHM13 [40], the first complete reference sequence of the human genome released by the Telomere-to-Telomere (T2T) Consortium, Gershman et al. [11] revealed the epigenetic patterns in previously unresolved challenging regions of the human genome utilizing ONT sequencing, providing the most complete human methylome ever produced. Advances in LRS-based methylation detection tools LRS technologies have provided important technical support for epigenetic research with their unique advantages. However, decoding the raw, complex electrical signals or enzyme kinetics into reliable methylation information is the critical step in realizing its scientific value. This has driven the rapid development of bioinformatics tools. In this section, we review the existing PacBio-based methylation detection tools and ONT-based methylation detection tools, respectively. PacBio-based methylation detection tools At present, various PacBio-based methylation detection tools have been developed (Fig. 2; Table 2). The development of these tools is closely related to the iterative evolution of PacBio sequencing technology, spanning from the early CLR era to the current HiFi era. Evolution of PacBio-based methylation detection tools. The gray section illustrates the evolution of PacBio sequencing from the early CLR era to the HiFi era. During the CLR era, the sequencing accuracy was relatively low (85–92%), while HiFi sequencing achieved high accuracy (≥ 99.9%) and supported five-base sequencing (A, C, G, T, and 5mC), providing a technical foundation for more reliable DNA methylation detection. The following section summarizes the PacBio-based methylation detection tools and their performance evolution. The blue section summarizes the early PacBio CLR-based methylation detection tools, including CRF model, seqPatch, SMALR, qDNAMod, and AgIn, which were mainly used for 6 mA and 4mC detection, while 5mC detection remained challenging. These tools have not been maintained or updated for over 8 years, are not suitable for HiFi reads, and are no longer recommended for current analyses. The green section summarizes the HiFi-based methylation detection tools in recent years, including HK model, Pb-CpG-tools, Primrose, Jasmine, Fibertools, Ccsmeth, and HiFiMeth. These tools support accurate 5mCpG detection and are currently the mainstream tools for PacBio methylation analysis. Among them, Jasmine (replacing Primrose), Pb-CpG-tools, and Fibertools are currently official tools provided by PacBio. HiFiMeth pioneered 5mC detection in multiple sequence contexts (CpG, CHG, CHH), although current validations are mainly focused on plant species. (A: adenine; C: cytosine; CCS: circular consensus sequencing; CLR: continuous long read; CRF: conditional random field; FPR: False positive rate; G: guanine; HiFi: high-fidelity; T: thymine; 4mC: N4-methylcytosine; 5hmC: 5-hydroxymethylcytosine; 5mC: 5-methylcytosine; 5mCpG: 5-methylcytosine in CpG; 6 mA: N6-methyladenine; 8-oxoG: 8-oxoguanine). Created in BioRender. Zhang, R. (2026) https://BioRender.com/7cxkb7b Table 2Current PacBio-based methylation detection toolsToolMethodMethylation typeDatasetReported performanceReferenceCRF modelCRF model5mC, 6 mA, 8-oxoGM.Sau3AI plasmid (5mC);Synthetic template (8-oxoG); M.EcoKdam plasmid (6 mA).Supervised: AUROC = 0.93 (5mC, at the site level);Unsupervised: only 8-oxoG detectable, 5mC undetectable (at the site level);6 mA not evaluated.[141]seqPatchEmpirical Bayesian hierarchical model6 mA, 4mCE. coli K-12 (6 mA);Plasmid (4mC).With control: Sensitivity = 95% (6 mA, at the site level);Without control: 6 mA detectable, 4mC undetectable.[42]qDNAmodEmpirical Bayesian hierarchical model5mC, 6 mA, 4mCSimulated data (6 mA);S. pneumoniae ST556 (6 mA)Only validated on 6 mA.Significantly better than conventional statistical methods at low modification ratios, with comparable performance at medium to high ratios.[43]SMALRSingle-molecule models (SMSN and SMP)6 mA, 4mCE. coli C227 (6 mA);C. jejuni (4mC)Sensitivity = 0.985, specificity = 0.995 (6 mA, at the read level);4mC detectable (no numerical data).[44]FibertoolsSemi-supervised CNNs6 mAK562 Fiber-seqPrecision > 90%, recall > 90% (at the read level).[142]AgInLinear-time dynamic programming algorithm and linear discriminant analysis5mCpGHuman (2 individuals)The correlation with Illumina BeadChip was 0.816, 75.4% concordance within 0.25.[143]HK modelCNN5mCpGHuman genomic DNA (WGA vs. M.SssI control);11 human tissuesSensitivity = 0.90, specificity = 0.94 (at the read level);The correlation with BS-seq was 0.99 (P < 0.0001).[47]JasmineCNN5mCpG, 5hmC(CpG), 6 mAHG002 WGS (5mCpG, 5hmC at CpG);Human WGS(6 mA)FPR = 11% (5mCpG), 0.9% (5hmC at CpG), 0.6%(6 mA) (at the read level).[48]Pb-CpG-toolsCNN5mCpGN/AThe accuracy of the 5mCpG prediction was significantly improved at the genome-wide site level, but the exact value is unknown.[50]ccsmethA DL model combining bidirectional GRU and Bahdanau attention network.5mCpGNA12898 (PCR vs. M.SssI control);HG002Accuracy = 0.9, AUROC = 0.97 (at the read level);The correlation with BS-seq was > 0.9 (only 10×reads).[16]HiFiMethDL5mCTen plant speciesThe correlation with BS-seq was 0.900-0.973 (CHG) and 0.755-0.800 (CHH) in plant species.[29]Abbreviations: AUROC: area under the receiver operating characteristic curve; BS-seq: bisulfite sequencing; C. jejuni: Campylobacter jejuni; CNN: convolutional neural network; CpG: cytosine-phosphate-guanine; CRF: conditional random field; DL: deep learning; E. coli: Escherichia coli; FPR: False positive rate; GRU: gated recurrent unit; HK: holistic kinetic; M.EcoKdam: Methyltransferase EcoK Dam; M.Sau3AI: Methyltransferase Sau3AI; M.SssI: M.SssI methyltransferase; N/A: not available; PCR: polymerase chain reaction; SMP: single molecule, pooled; SMSN: single molecule, single-nucleotide; S. pneumoniae: Streptococcus pneumoniae; WGA: whole genome amplification; 4mC: N4-methylcytosine; 5hmC: 5-hydroxymethylcytosine; 5mC: 5-methylcytosine; 5mCpG: 5-methylcytosine in CpG; 6 mA: N6-methyladenine; 8-oxoG: 8-oxoguanine Current PacBio-based methylation detection tools M.Sau3AI plasmid (5mC); Synthetic template (8-oxoG); M.EcoKdam plasmid (6 mA). Supervised: AUROC = 0.93 (5mC, at the site level); Unsupervised: only 8-oxoG detectable, 5mC undetectable (at the site level); 6 mA not evaluated. E. coli K-12 (6 mA); Plasmid (4mC). With control: Sensitivity = 95% (6 mA, at the site level); Without control: 6 mA detectable, 4mC undetectable. Simulated data (6 mA); S. pneumoniae ST556 (6 mA) E. coli C227 (6 mA); C. jejuni (4mC) Sensitivity = 0.985, specificity = 0.995 (6 mA, at the read level); 4mC detectable (no numerical data). Human genomic DNA (WGA vs. M.SssI control); 11 human tissues Sensitivity = 0.90, specificity = 0.94 (at the read level); The correlation with BS-seq was 0.99 (P < 0.0001). HG002 WGS (5mCpG, 5hmC at CpG); Human WGS(6 mA) NA12898 (PCR vs. M.SssI control); HG002 Accuracy = 0.9, AUROC = 0.97 (at the read level); The correlation with BS-seq was > 0.9 (only 10×reads). Abbreviations: AUROC: area under the receiver operating characteristic curve; BS-seq: bisulfite sequencing; C. jejuni: Campylobacter jejuni; CNN: convolutional neural network; CpG: cytosine-phosphate-guanine; CRF: conditional random field; DL: deep learning; E. coli: Escherichia coli; FPR: False positive rate; GRU: gated recurrent unit; HK: holistic kinetic; M.EcoKdam: Methyltransferase EcoK Dam; M.Sau3AI: Methyltransferase Sau3AI; M.SssI: M.SssI methyltransferase; N/A: not available; PCR: polymerase chain reaction; SMP: single molecule, pooled; SMSN: single molecule, single-nucleotide; S. pneumoniae: Streptococcus pneumoniae; WGA: whole genome amplification; 4mC: N4-methylcytosine; 5hmC: 5-hydroxymethylcytosine; 5mC: 5-methylcytosine; 5mCpG: 5-methylcytosine in CpG; 6 mA: N6-methyladenine; 8-oxoG: 8-oxoguanine Back in 2010, the research team from PacBio first demonstrated that the differences in kinetic signatures (e.g., IPD, PW) between modified and unmodified bases can be used to detect the presence of DNA methylation modifications, such as 6 mA, 4mC, and 5mC [22]. However, it is noteworthy that the early PacBio CLR sequencing mode had significantly different requirements for sequencing coverage requirements for detecting different DNA methylation types. For example, 4mC and 6 mA can be accurately detected at low coverage (25×), whereas 5mC tends to require very high coverage reads (up to 250×) to ensure its detection accuracy due to its very low signal-to-noise ratio, which makes the detection costly [41]. Therefore, the early research on PacBio CLR-based methylation detection tools mainly focused on 6mC and 4mC, which have distinct kinetic signatures [42–44]. The HiFi sequencing technology has reduced the coverage required for reliable 5mC detection from 250× (CLR) to 10–20× [16, 45, 46]. In recent years, several teams have developed DL models for highly accurate 5mC detection using HiFi sequencing. In 2021, Tse et al. [47] achieved the first detection of eukaryotic genome-wide 5-methylcytosine in CpG (5mCpG) modifications by PacBio sequencing. They developed a convolutional neural network (CNN)-based holistic kinetic (HK) model that integrates the kinetic signals of DNA polymerase as well as sequence context to detect 5mCpG. On a genome-wide scale, the model achieved 90% sensitivity and 94% specificity for 5mCpG detection at single-base resolution. It is also noteworthy that the HK model can accurately detect 5mCpG, achieving an AUROC of 0.93 with a sequencing depth of only 10×. Similarly, Primrose, developed by PacBio, also used CNN to improve the accuracy of methylation detection, but has since been replaced by Jasmine [48, 49]. Notably, Jasmine is the current mainstream methylation detection solution of PacBio, and it has been integrated into the HiFi sequencing basecalling process (that is, converting the raw signals generated by the sequencer into base sequence information). This eliminates the need for repeated basecalling or extraction of intermediate signals from subreads files when performing DNA methylation analysis, effectively reducing redundant computation and storage space. Additionally, PacBio provides the downstream analysis toolkit pb-CpG-tools, which integrates these HiFi reads annotated with methylation information to calculate the methylation frequencies at specific sites. In addition, it also supports haplotype resolution analysis [50]. In 2023, Ni et al. [16] constructed a DL model, ccsmeth, using bidirectional gated recurrent unit (GRU) and attention neural networks to better extract signal features from HiFi reads. The model takes into account the symmetric properties of 5mCpG methylation sites at the single-molecule level and the aggregation properties of methylation sites on the genome at the genomic site level, which enhances the accuracy of 5mCpG detection. The validation results in different datasets showed that at the read level, the detection performance of ccsmeth was superior to existing CNN-based models. At the genome-wide site level, the ccsmeth model assays showed a higher correlation with BS-seq, especially at low coverage, where a correlation of > 0.9 was achieved using only 10×reads. Recently, HiFiMeth has expanded the application of PacBio HiFi sequencing in DNA methylation detection. Unlike the aforementioned models that mainly focus on CpG sites, HiFiMeth enables the detection of 5mC in CpG, CHG, and CHH sequence contexts. This model has been validated in multiple plant species, and its correlations with BS-seq were 0.900–0.973 (CHG sequence context) and 0.755–0.800 (CHH sequence context) [29]. Because validation has mainly focused on plant species, it has not yet become the mainstream choice for eukaryotic DNA methylation analysis. It should be noted that the PacBio CLR-based methylation detection tools have not been maintained or updated for at least 8 years and are no longer applicable to current mainstream HiFi sequencing. In contrast, most DL models developed for HiFi sequencing have shown clear advantages and have become the preferred approach for current 5mCpG detection. However, their performance in detecting non-CpG methylation remains unclear. HiFiMeth was the first to extend detection to non-CpG sites, but its reliability in mammalian genomes has yet to be fully validated. ONT-based methylation detection tools Published ONT-based methylation detection tools can be broadly classified into three categories according to their detection strategies (Fig. 3; Table 3): (1) tools based on basecalling, which directly detect methylation signals during the basecalling process; (2) tools based on statistical tests, which detect methylation by comparing the sequencing signals of natural samples with those of unmodified control samples after basecalling and alignment, without prior knowledge of methylation modification (de novo detection); and (3) tools based on pre-trained models, which detect methylation using pre-trained models after basecalling and alignment. Each of these three categories has distinct strengths and limitations, which are discussed in detail below. Evolution of ONT-based methylation detection tools. The blue section represents tools based on basecalling. These tools can not only support canonical basecalling but also support methylation calling. Guppy was widely used in the early stage, but has been replaced by Dorado, which is currently the official basecaller provided by ONT. The green section represents tools based on statistical tests. These tools do not require training data, but their accuracy is lower than that of other types of tools. Moreover, these tools have not been maintained or updated for over 3 years and are not recommended for current analysis. The red and yellow sections represent tools based on pre-trained models. The red section represents tools based on ML models (HMM). These tools rely on manual feature extraction, have strong interpretability, but have bottlenecks in handling large-scale data and limited generalization ability for complex datasets. Nanopolish and signalAlign have not been maintained for over 3 years, only f5C is still actively updated and supports the R10.4.1 flow cell. The yellow section represents tools based on DL models. These tools enable automated feature extraction, scale effectively to large datasets, and achieve better accuracy than statistical tests and traditional ML models. They are currently the most widely supported among all tool categories. The current best-performing tools are DeepMod2 and Rockfish (F1 > 0.97), compatible with the R10.4.1 flow cell. Additionally, it is worth noting that DeepSignal-plant enables 5mC detection in multiple sequence contexts (CpG, CHG, CHH) in plants and is also compatible with the R10.4.1 flow cell. (AUROC: area under the receiver operating characteristic curve; DL: deep learning; ML: machine learning; ONT: Oxford Nanopore Technologies; 4mC: N4-methylcytosine; 5hmC: 5-hydroxymethylcytosine; 5mC: 5-methylcytosine; 5mCpG: 5-methylcytosine in CpG; 6 mA: N6-methyladenine). Created in BioRender. Zhang, R. (2026) https://BioRender.com/jt7pbf3 Table 3Current ONT-based methylation detection toolsToolStrategyAlgorithmMethylation typeFlow cells compatibilityDatasetReported performanceReferenceGuppyBasecallingDNN5mC, 5hmC, 6 mAR7.3, R9, R9.4,R9.4.1, R9.5, R10,R10.3, R10.4, R10.4.1N/AN/A[54]BonitoDNN5mC, 5hmC, 6 mAR10.4.1N/AN/A[59]DoradoDNN5mC, 5hmC, 6 mA, 4mCR9.4.1, R10.4.1N/AAccuracy = 0.994 (5mC), 0.987(5hmC), 0.997(6 mA), 0.976(4mC)(at the read level).[144]Nanorow(later integrated into TomBo)Statistical testMann-Whitney U-testde novoR9, R9.4Synthetic (WGA E. coli DNA with 7 methylases)AUROC = 0.62–0.97 (at the site level).[63]NanoModKolmogorov-Smirnov testde novoR7.3, R9E. coli with M.SssIPrecision > 0.90 (at the site level).[65]NanopolishPre-trained modelHMM5mCpGR7.3, R9NA12878 (PCR vs. M.SssI control)Accuracy = 0.87 (at the read level).[67]signalAlignA variable-order HMM with an HDP (HMM-HDP)5mC, 5hmC, 6 mAR7.3, R9pUC19 plasmid (5mC, 6 mA)Accuracy = 0.86 (6 mA) and 0.96 (5mC) (at the site level).[68]f5CNanopolish-based optimization model5mCpGR9.4, R10.4, R10.4.1NA12878f5c is 3-5x faster than Nanopolish and shows a 0.99999 correlation with comparable accuracy.[69]mCallerNeural network5mCpG, 6 mAR9, R9.4, R9.5E. coli MG1655 K12 (6 mA);E. coli with M.SssI (5mCpG).Accuracy = 0.813 (6 mA), 0.822 (5mCpG) (at the read level).[77]DeepSignalCNN and BRNN5mC, 6 mAR9, R9.4, R9.4.1, R10.4.1NA12878 with M.SssI (5mC);pUC19 plasmid (6 mA).Accuracy = 0.947, AUROC = 0.988 (5mC, at the read level);Accuracy = 0.999 (6 mA, at the site level).[70]DeepModA BRNN with LSTM5mC, 6 mAR9, R9.4, R9.4.1NA12878 (5mC);E. coli (6 mA).Precision = 0.99 (5mC) and ~ 0.90 (6 mA) (at the site level).[71]MegalodonRNN5mC, 6 mAR9.4.1, R10.3N/AN/A[72]DeepSignal-plantA BRNN with LSTM5mCR9.4.1, R10.4.1A. thaliana (5mC)The correlation with BS-seq was 0.94 (CpG), 0.91 (CHG), and 0.78 (CHH).[75]METEOREA consensus approach that uses RF and multiple linear regression (REG)5mCpG, 6 mAR9.4, R9.4.1NA12878 (5mCpG)The correlation with BS-seq was 0.9262 with an RMSE of 0.1607 (5mCpG).[66]DeepMPCNN5mCpG, 6 mAR7.3, R9, R9.4NA12878 (5mC);pUC19 plasmid (6 mA).F1 score = 0.932 (5mC), 0.874(6 mA) (at the read level).[78]methBERTmethBERT5mC, 6 mAR9E. coli & H. sapiens (5mC);E. coli (6 mA).AUROC = 0.82–0.97 (5mC), 0.68–0.93 (6 mA) (at the read level).[73]DeepMod2BiLSTM and Transformer5mCpGR9.4.1, R10.4.1HG002, HG003, HG004 (R10.4.1)F1 scores = 0.975–0.982 (at the read level);The correlation with BS-seq was 0.95–0.97.[56]RockfishTransformer5mCpGR9.4.1, R10.4.1NA12878 (R10.4.1)F1 scores = 0.974 (at the read level);The correlation with BS-seq was 0.909.[9]DeepBAMBiLSTM5mCpGR10.4HG001, HG002AUROC = 0.957–0.98, F1 scores ~ 0.90–0.97 (at the read level).[145]NanoConA deep hybrid network combining transformer and Bi-GRU with a contrastive learning strategy.5mCR9.4A. thaliana, O. sativa, NA12878AUROC = 0.942–0.993, F1 scores = 0.656–0.925 (at the read level).[74]amortized-HMMA novel hybrid HMM–DNN approach5mCpGR9.4NA12878Significantly outperforms DeepSignal and Nanopolish when the training data is incomplete, with higher generalization ability.[81]Abbreviations: AUROC: area under the receiver operating characteristic curve; A. thaliana: Arabidopsis thaliana; BERT: Bidirectional Encoder Representations from Transformers; Bi-GRU: bidirectional gated recurrent unit; BiLSTM: bidirectional long short-term memory; BRNN: bidirectional recurrent neural network; BS-seq: bisulfite sequencing; CNN: convolutional neural network; DNN: deep neural network; E. coli: Escherichia coli; HDP: hierarchical Dirichlet process; HMM: hidden Markov model; LSTM: long short-term memory; M.SssI: M.SssI methyltransferase; N/A: not available; O. sativa: Oryza sativa; PCR: polymerase chain reaction; RNN: recurrent neural network; WGA: whole genome amplification; 5mC: 5-methylcytosine; 5mCpG: 5-methylcytosine in CpG; 5hmC: 5-hydroxymethylcytosine; 6 mA: N6-methyladenine Current ONT-based methylation detection tools R7.3, R9, R9.4, R9.4.1, R9.5, R10, R10.3, R10.4, R10.4.1 Accuracy = 0.994 (5mC), 0.987(5hmC), 0.997(6 mA), 0.976(4mC) (at the read level). Nanorow (later integrated into TomBo) E. coli MG1655 K12 (6 mA); E. coli with M.SssI (5mCpG). NA12878 with M.SssI (5mC); pUC19 plasmid (6 mA). Accuracy = 0.947, AUROC = 0.988 (5mC, at the read level); Accuracy = 0.999 (6 mA, at the site level). NA12878 (5mC); E. coli (6 mA). NA12878 (5mC); pUC19 plasmid (6 mA). E. coli & H. sapiens (5mC); E. coli (6 mA). F1 scores = 0.975–0.982 (at the read level); The correlation with BS-seq was 0.95–0.97. F1 scores = 0.974 (at the read level); The correlation with BS-seq was 0.909. Abbreviations: AUROC: area under the receiver operating characteristic curve; A. thaliana: Arabidopsis thaliana; BERT: Bidirectional Encoder Representations from Transformers; Bi-GRU: bidirectional gated recurrent unit; BiLSTM: bidirectional long short-term memory; BRNN: bidirectional recurrent neural network; BS-seq: bisulfite sequencing; CNN: convolutional neural network; DNN: deep neural network; E. coli: Escherichia coli; HDP: hierarchical Dirichlet process; HMM: hidden Markov model; LSTM: long short-term memory; M.SssI: M.SssI methyltransferase; N/A: not available; O. sativa: Oryza sativa; PCR: polymerase chain reaction; RNN: recurrent neural network; WGA: whole genome amplification; 5mC: 5-methylcytosine; 5mCpG: 5-methylcytosine in CpG; 5hmC: 5-hydroxymethylcytosine; 6 mA: N6-methyladenine Tools based on basecalling In ONT sequencing, earlier basecallers performed basecalling from pre-segmented current signals (i.e., events) by hidden Markov model (HMM) or recurrent neural network (RNN) [51, 52]. It was not until 2018 that Chiron, the first open-source tool that incorporates multiple DL networks, appeared, allowing sequence translation directly from raw current data without relying on current signal division information [53]. In 2019, specific basecallers for different kinds of data began to appear. Currently, some basecallers developed by ONT not only support canonical basecalling but also support methylation calling. Among them, Guppy was widely used in earlier ONT workflows [54]. However, it has been superseded by Dorado, ONT’s current official basecaller. Unlike Guppy, which relies on the central processing units (CPUs) for methylation calling, Dorado uses graphics processing units (GPUs) to perform this process, requiring almost no additional running time [2, 55]. In addition, Dorado can perform methylation calling directly from unmapped reads, without the need for reference genome or aligned reads [55, 56]. The methylation information is recorded in the BAM files in the form of MM/ML tags with respect to the read coordinates. Users can choose to conduct the subsequent alignment by themselves or specify the reference genome in the Dorado command line, allowing Dorado to call minimap2 for alignment. Notably, Dorado has been integrated into the MinKNOW software of ONT sequencing equipment, enabling live basecalling, and can also be used as an independent tool [57]. Meanwhile, the research infrastructure used for training the basecaller model has also been updated. Taiyaki was a research software developed by ONT for training the basecaller models. This software was officially archived and is no longer supported in 2022 [58]. Currently, Bonito has replaced Taiyaki as the open-source research basecaller of ONT. It supports the training and evaluation of basecalling models, providing a flexible and efficient research platform for the development of algorithms [59]. As with these basecaller updates, the field remains vibrant and rapidly evolving. It should be noted that Guppy and Taiyaki are no longer supported. Currently, Dorado is the official basecaller provided by ONT, while Bonito serves as an open-source research platform for basecalling model training, which is not recommended for routine data analysis. Although these basecallers with methylation calling capability show great promise by reducing the need for downstream complex computational analyses, there are still several limitations at present. First, due to limited diversity in the available training datasets, existing basecallers are mainly focused on specific species (e.g., human or bacterial). Secondly, it is difficult to balance the speed and accuracy of the basecaller. For example, while Bonito offers over 1% higher basecalling accuracy than Guppy, its slower processing speed substantially limits its practical utility [60]. Third, such tools are highly dependent on computing resources. However, it is worth noting that ONT’s sequencers, such as the PromethION 24/48 and PromethION 2 Integrated, are equipped with GPUs, enabling basecalling and methylation calling on the device, thereby reducing the dependence on external computing resources to some extent [61, 62]. Tools based on statistical tests Tools based on statistical tests do not require prior knowledge, but they require two sets of sequencing data: one from the sample to be tested (which may contain modifications), and the other from the matched non-modified control sample. By conducting statistical tests on the differences in electrical signals at the same genomic site of the two groups of samples, the methylation status of the target site can be inferred. As early as 2017, Stoiber et al. [63] proposed nanorow, the first open-source software that can identify 5mC, 6 mA, and 4mC from ONT sequencing data. It employs the Mann-Whitney U-test to determine whether a site is modified or not. In the synthetic sample dataset, nanoraw was able to identify three methylation modifications, 4mC, 5mC, and 6 mA, in seven sequence contexts, with an area under the receiver operating characteristic curve (AUROC) of 0.62–0.97. Notably, nanoraw was subsequently integrated into a more comprehensive and powerful toolset called Tombo [64]. Building on the sample comparison strategy inherited from nanoraw, Tombo introduced two additional strategies: alternative model and de novo, enabling detection in more complex scenarios such as the lack of control samples. In 2019, Liu et al. [65] proposed a new computational tool named NanoMod, which employed the Kolmogorov-Smirnov test to identify base modifications. On the E. coli methylation dataset, NanoMod significantly outperformed nanoraw in detecting 5mC. It should be noted that the study only used nanoraw as a comparison object and lacked a direct performance comparison with Tombo. Later benchmarking studies reported that Tombo achieved a per-read AUROC of 0.90 and a BS-seq correlation of 0.78 on the NA12878 dataset, both lower than those of tools based on pre-trained models, including Nanopolish, DeepSignal, and Megalodon [66]. Overall, while the tools based on statistical tests mentioned above can detect all detectable methylation modifications present in a sample, they cannot further identify the specific types of these modifications. Moreover, these tools tend to be less accurate than those based on basecalling and pre-trained models, and having not been maintained or updated for 3 to 8 years, they are no longer recommended for current analysis. Tools based on pre-trained models Tools based on pre-trained models typically focus on specific methylation contexts (e.g., CpG sites). These tools are trained on labeled datasets to learn the relationship between electrical signal patterns and known methylation states, and subsequently apply the trained models to predict methylation in new sequencing data. Depending on the algorithms used, such tools can be further classified into those based on conventional machine learning (ML) models, e.g., HMM, and those based on DL models. Tools based on conventional machine learning models In the early stages, such tools typically employed traditional ML models, such as HMM, to identify DNA modification sites. The advantage of this approach lies in the simplicity of the algorithm and the low computational resource requirements, making it suitable for handling small-scale datasets. Back in 2017, a tool based on the HMM model, Nanopolish, was proposed to directly detect DNA 5mC by analyzing current events recorded by ONT sequencers [67]. In the homo sapiens ONT sequencing dataset, Nanopolish achieved 87% accuracy for methylation detection at individual CpG sites. In the same year, Rand et al. [68] proposed signalAlign, combining the HMM and the hierarchical Dirichlet process (HDP), which allows distinguishing three cytosine variants (C, 5mC, 5hmC) and two adenine variants (A, 6 mA) from the current signals of ONT sequencing. Subsequently, in 2020, Gamaarachchi et al. [69] developed an optimization tool f5C based on nanopolish. This tool dramatically reduced the runtime and significantly improved the efficiency of methylation detection by improving the signal comparison algorithm and using GPU acceleration. At the same time, f5C also maintains a methylation detection accuracy comparable to that of nanopolish, with a high correlation of 0.99999. However, notably, these conventional ML models typically rely on manual selection and extraction of features, which limits their ability to generalize to complex sequencing datasets containing diverse modification types. In addition, the performance of conventional ML models is limited as the size of the dataset increases. More importantly, Nanopolish and signalAlign have not been updated for over 3 years and are incompatible with current ONT chemistries, only f5C is still actively maintained and adapted to the latest R10.4.1 flow cell. Tools based on deep learning models Tools based on DL models use deep neural networks to automatically extract features from sequencing signals and predict DNA modification sites. Not only do they better adapt to large-scale data, but they also demonstrate higher accuracy and robustness in detecting complex modification types. They are currently the mainstream tools for LRS-based methylation detection. Back in 2019, two DL tools, DeepMod and DeepSignal, both using a bidirectional recurrent neural network (BRNN) architecture, were proposed for detecting 5mC and 6mA [70, 71]. In the same year, in addition to the tools proposed by these third-party research teams, ONT similarly proposed Megalodon, an RNN-based methylation detection tool [72]. However, it is noteworthy that the computational sequential nature of the sequence models (e.g., RNN, BRNN) used by the above tools makes them difficult to optimize in parallel, leading to slower model inference. Facing this situation, subsequently, in 2021, Zhang et al. [73] proposed an alternative fast detection model to the existing BRNN, methBERT, a non-recursive neural network based on Bidirectional Encoder Representations from Transformers (BERT). It captures the contextual information in the sequence through a self-attentive mechanism and can process all the positional information in the sequence simultaneously. On several benchmark datasets, BERT performs comparably to or outperforms BRNN, with inference speeds 1.49 to 7.17 times faster than BRNN, making it more efficient in practice. Subsequently, some detection tools (e.g., DeepMod2 [56], Rockfish [9], NanoCon [74]) have emerged that also employ the Transformer architecture based on the self-attention mechanism, showing great potential in accurately detecting DNA methylation. In addition, although many ONT-based methylation detection tools have been developed, the difference among flow cell versions is only covered by two of them-DeepMod [56, 71] and DeepSignal [70, 75, 76]. In contrast, mCaller [77], methBERT [73], and DeepMP [78] have not been maintained or updated for 3–4 years. However, it is noteworthy that while the latest version of DeepSignal adapts to the R10.4.1 flow cell, it has not yet been officially released. Until 2024, DeepMod2 became the first open-source tool adapted to the R10.4.1 flow cell for predicting the methylation status of CpG sites [56]. Similar to methBERT, DeepMod2 is a DL framework that contains a bidirectional long short-term memory (BiLSTM) model and a Transformer model. Compared to DeepMod, which contains a three-layer BiLSTM architecture, DeepMod2 is better able to capture long-range dependencies. The results of the study show that in the human dataset, DeepMod2 has an F1 score of approximately 95% at the per-read level and 99% at the per-site level. Its performance is comparable to that of the current state-of-the-art tool Dorado from ONT. In the same year, Stanojević et al. [9] developed a DL tool, Rockfish, based on the Transformer architecture, for the detection of 5mC. It can also be used to analyze R10.4.1 flow cell datasets. The study showed that Rockfish performed well on the human dataset (R10.4.1), with an accuracy of 97.25% and an F1 score of 97.39% at the read level. In addition, notably, the adaptable architectural design of DeepMod2 and Rockfish allows them to be applied to signal data from different new versions of flow cells in the future. However, although existing DL-based tools have realized significant advantages in DNA modification detection of ONT sequencing data, there are still some limitations. First, in practical DNA methylation site detection tasks, there is usually a significant imbalance in the distribution of methylated and unmethylated sites, and their ratio varies depending on the species and genomic contexts. For example, in plants, unmethylated sites are usually far more numerous than methylated sites, while in human CpG site detection, the methylation levels differ significantly across different genomic regions, resulting in a similar imbalance in the distribution of methylated and unmethylated sites [75, 79]. This imbalance reduces the model’s ability to identify minority classes of sites. Therefore, existing detection tools often perform poorly when processing real-world data [80]. To address this problem, Yin et al. [74] proposed NanoCon, a deep hybrid network based on contrast learning for 5mC detection. It was shown that the feature representations between methylated and unmethylated sites were enhanced by introducing a contrast learning strategy, which improved the detection accuracy and robustness of the model under unbalanced data. Second, these tools perform well when the training data is complete, but their performance degrades significantly when the training data is incomplete (i.e., some k-mer context is not present in the training data). As ONT sequencing continues to expand to detect more kinds of methylation modifications, obtaining complete training data sets will become increasingly difficult. Therefore, to effectively detect DNA methylation modifications with limited labeled training data, Yao et al. [81] proposed a new hybrid model, amortized-HMM, combining HMM and deep neural network (DNN). The ability to predict unseen k-mers is improved by using DNN to predict missing emission distribution parameters in HMM. The results show that it significantly outperforms DeepSignal and the conventional HMM method (Nanopolish) when the training data is incomplete (e.g., k-mer completeness is less than 30%), especially when dealing with unseen k-mer. In conclusion, this hybrid model has stronger generalization ability and robustness under limited marker data, providing a more practical solution for epigenetic modification detection. Advances in systematic benchmarking Systematic benchmarking is crucial for evaluating the reliability and accuracy of various LRS-based methylation detection tools, thereby guiding the selection of optimal analytical protocols. Existing benchmarking studies have provided valuable references in this regard, but due to the differences in experimental designs and evaluation indicators, the reported performance results are often difficult to directly compare. Because early PacBio sequencing suffered from low signal-to-noise in methylation detection, benchmarking studies of LRS-based methylation detection tools have focused on ONT sequencing, but the datasets used for these studies were limited in both scale and diversity [13, 66]. There is still a lack of systematic evaluation and published guidelines for LRS-based methylation detection tools on a human epigenome-wide scale. To fill this gap, in 2021, Liu et al. [80] systematically evaluated the performance of seven mainstream CpG methylation detection tools (Nanopolish, Megalodon, DeepSignal, Guppy, Tombo, DeepMod, METEORE) across different genomic contexts at per-read and per-site levels for the first time at the human epigenome-wide scale using ONT sequencing datasets from four human samples (NA19240, NA12878, APL, K562). This study reveals significant regional dependence in tool performance, with detection capabilities declining across all tools in regions of inconsistent non-haplotype CpGs, intergenic regions, low CG density areas, and repetitive sequences. And METEORE (a consensus algorithm that integrates the prediction results of multiple tools) also performs lower than Megalodon, Nanopolish, DeepSignal, and Guppy on the human genome, which is different from the results evaluated by Yuen et al. [66] based on small training and validation datasets (containing only 100 CpG sites in E. coli and less than 2,000 CpG sites in humans). In addition, this study quantifies computational resource consumption, providing an important reference for large-scale data analysis: Guppy and Nanopolish are used for methylation detection when computational resources are limited, while Megalodon is preferred in high-performance computing environments. Since 2022, benchmark studies incorporating the PacBio platform have begun to emerge to compare the performance of different sequencing technologies in CpG methylation detection [21, 82]. However, these cross-platform studies have a significant limitation in that the PacBio platform only incorporates Primrose and fails to cover other mainstream PacBio-based methylation detection tools. This restricts a comprehensive understanding of the optimal detection tools for the PacBio platform. Notably, the methylation detection results produced by Primrose exhibit a systematic bias, tending to avoid extreme methylation states (i.e., 0% or 100%). This bias can lead to substantial underestimation of fully methylated or completely unmethylated sites within key genomic regions, such as imprinting control regions and promoters, thereby misleading downstream biological conclusions [21]. In conclusion, systematic benchmarking provides an important basis for the selection and optimization of LRS-based methylation detection tools. However, there is currently a lack of universally applicable best practice recommendations. The number of relevant benchmark studies is limited, and they have several significant shortcomings. First, there is a lack of comprehensive evaluation for newly developed platforms and tools. Existing benchmark studies primarily focus on older tools that are no longer maintained or updated and are no longer recommended for current analytical work. Second, their reliance on standardized cell lines makes it difficult to capture the heterogeneity observed in real clinical samples. More importantly, the widespread use of BS-seq as the gold standard may introduce systematic bias. In the future, to establish reliable best practice recommendations, benchmark studies must incorporate more current mainstream tools. It is also urgently necessary to establish diverse standard datasets and a unified evaluation framework to improve the comprehensiveness and reliability of system benchmarking, thereby promoting the development of this field [83]. Applications of LRS-based methylation detection in biomedical research In the preceding section, we reviewed the evolutionary trajectory of the core algorithms underlying LRS-based methylation detection tools. The steady advancement of these tools provides a reliable foundation for downstream biological discoveries. Looking back at this field, it is clear that many pioneering studies in this field were driven by robust and widely adopted methylation detection tools available at the time (such as the core suite of ONT platforms, Guppy, and Dorado). In the following section, we highlight representative case studies that illustrate how LRS-based methylation detection is reshaping our understanding of human diseases by elucidating pathogenic mechanisms and facilitating precision diagnostics and therapeutics. Multiple studies have demonstrated that ONT-based methylation detection has played a pivotal role in the research and diagnosis of neurodevelopmental disorders associated with repetitive amplification. By using basecallers such as Guppy or Dorado to directly identify methylation signals during sequencing, researchers can simultaneously resolve repetitive sequences and methylation modification states in a single sequencing run. This technology not only enables precise identification of pathogenic repeat expansions but also uncovers allele-specific methylation and methylation mosaicism, thereby extending our understanding of disease mechanisms from genetic variation to epigenetic regulation [6, 84–86]. These findings provide an essential molecular basis for explaining phenomena such as incomplete penetrance, clinical heterogeneity, and progressive disease trajectories, marking a shift in the field from simple genetic diagnosis to comprehensive assessment of gene functional states. For imprinting disorders, LRS-based methylation detection is capable of detecting methylation in key imprinted regions and determining parental methylation phasing without requiring parental samples. This effectively overcomes the inherent limitations of traditional methods, which often involve cumbersome workflows and yield incomplete or ambiguous information [14, 87]. In addition, DNA methylation detection also holds significant value in tumor diagnosis. Abnormalities in the genomic methylation pattern are one of the important molecular characteristics of tumors, and these changes can be used for tumor tissue tracing, molecular classification, and prognosis assessment [88, 89]. At present, the tumor molecular typing technology based on methylation profiles has demonstrated clinical application value in brain tumors, hepatobiliary-pancreatic tumors, and hematological malignancies [90–93]. Among these, brain tumors exhibit high heterogeneity. The treatment strategies for different subtypes (especially the extent of surgical resection) vary significantly. Therefore, precise molecular classification directly determines clinical decisions. The World Health Organization (WHO) classification guidelines for central nervous system tumors have established DNA methylation profiling analysis as an important basis for the identification of brain tumor molecular subtypes [94]. However, the current diagnostic methods have obvious limitations. Preoperative imaging and intraoperative frozen section histology are unable to distinguish the molecular subtypes. Although the widely used Illumina EPIC array (a high-throughput DNA methylation detection chip) can complete the classification, it has a complex procedure and takes several days, which cannot meet the requirements for real-time decision-making during surgery [95, 96]. Therefore, a rapid and accurate in-surgery DNA methylation detection technology holds significant clinical importance. In response to this demand, in recent years, studies have attempted to apply the ONT-based methylation detection technology to the intraoperative classification of brain tumors. For instance, the ONT methylation typing process established by Emiliani et al. [97] and Deacon et al. [98] can complete the classification of brain tumors within 2 h. They respectively verified it in 15 intraoperative samples and 50 prospective cohorts, and the accuracy rate was over 90% in both cases. Vermeulen et al. [96] reduced the detection turnaround time to less than 90 min and achieved an accuracy rate of 90% in 50 retrospective samples. Brändl et al. [99] further compressed the process to 60 min, achieving an accuracy rate of 94.5% in 75 clinical samples. Notably, Patel et al. [100] were able to complete tumor typing based on methylation and obtain DNA copy number information within 30 min during the operation through their self-developed adaptive sampling ONT sequencing method, Rapid-CNS2, with an accuracy rate of 94.6% in 301 clinical samples. Overall, the existing studies have all shown that the ONT-based methylation detection technology can significantly shorten the detection time while maintaining a high accuracy rate, thus having the potential to be applied in real-time to achieve rapid and accurate tumor classification during surgery. Beyond ONT sequencing, PacBio sequencing offers unique advantages in single-molecule epigenomic tracing due to its high accuracy. In maternal plasma analysis, researchers employed the HK model to detect methylation states on long cfDNA molecules, thereby constructing single-molecule methylation haplotypes. This approach successfully distinguished long cfDNA molecules as either fetal or maternal in origin [101]. This approach directly addresses the challenge of determining maternal haplotype inheritance in non-invasive prenatal testing. Looking ahead, as LRS-based methylation detection tools continue to be refined, we can expect to achieve more precise molecular analysis in increasingly complex biological scenarios, thereby unraveling more intricate biological questions. Challenges of LRS-based methylation detection In this review, we describe LRS-based methylation detection tools. Emerging tools enhance the accuracy and efficiency of methylation detection, laying the foundation for more refined downstream analysis. Furthermore, although LRS-based methylation detection has demonstrated revolutionary potential in biomedical research, it still faces numerous challenges. Only by confronting these challenges can we clarify the direction for future development and drive its broader application (Fig. 1C). First, the high cost makes many scientific institutes and hospitals flinch. The initial investment for PacBio’s equipment is relatively high. The current mainstream sequencers cost several hundred thousand dollars or more. For example, PacBio Revio typically costs about $779,000 [102]. In contrast, ONT offers more flexible options in terms of equipment investment. The price range of its sequencers is quite wide, ranging from the portable sequencer MinION Mk1D (approximately $3150) to the PromethION P48 (approximately $310000), which to some extent lowers the initial entry barrier [20, 103]. However, a lower equipment threshold does not equate to an advantage in terms of operating costs. According to the public service pricing of the Cancer Research Technology Program (CRTP), when normalized to the data amount required for a single human whole genome methylation assay (~ 100 Gb of raw data), the actual running costs (including library preparation and sequencing) for a single sample on the three platforms vary significantly [104]. The PacBio Revio costs approximately $2,300, the ONT costs about $950, and the Illumina NovaSeq X Plus costs around $840. The main reason for this disparity lies in the different single-run throughput of each platform. Revio’s single-run typically only meets the demand of a single sample, while ONT’s single-run can support approximately 3 sample pooled runs, and NovaSeq X Plus’s single-run can support dozens of sample pooled runs, significantly reducing the cost per single sample. Thus, the single-sample run costs of PacBio Revio and ONT are higher than those of Illumina NovaSeq X. Combined with the relatively limited base quality of LRS, which requires deeper sequencing coverage to enhance confidence. These multiple factors jointly contribute to the fact that the cost of LRS-based methylation detection is higher than that of BS-seq. In recent years, researchers have attempted to reduce error rates in raw signals and decrease the need for repeated sequencing by adopting new biosensors and improving the control of the speed of DNA translocation, including molecular ping-pong methodology and other technical advances [105–107]. Second, the strict requirement of large quantities of high-quality, complete DNA material further raises the barrier of methylation detection using LRS. The standard PacBio HiFi sequencing workflow typically requires starting DNA > 3 µg per 1 Gb genome and recommends that most DNA fragments be > 20–30 kb to achieve an insertion size distribution of 10–25 kb after shearing [108]. The standard procedure for the ONT Ligation Sequencing Kit recommends starting with 1 µg of DNA [109, 110]. In contrast, BS-seq has relatively lower requirements for the amount of starting DNA. Taking the widely used Zymo EZ DNA Methylation-Gold Kit in the literature as an example, its manual suggests inputting 200–500 ng of DNA to achieve a more stable bisulfite conversion effect [111, 112]. Some clinical samples, such as cfDNA and formalin-fixed and paraffin-embedded (FFPE), which are of low quantity or degraded, usually have insufficient DNA content and integrity to meet the requirements of the standard LRS library construction process [113, 114]. To address this issue, PacBio and ONT have introduced low-input or ultra-low-input library preparation protocols in recent years. Additionally, there have been approaches combining enzymatic methylation sequencing (EM-seq) and TET-assisted pyridine borane sequencing (TAPS) with LRS technologies (i.e., long-read EM-seq, long-read TAPS). EM-seq utilizes tet methylcytosine dioxygenase 2 (TET2)-mediated oxidation combined with apolipoprotein B mRNA editing enzyme catalytic subunit 3A (APOBEC3A)-mediated deamination to convert unmethylated cytosine (C) to uracil (U) (read as T in sequencing), while methylated C remains as C, thus enabling DNA methylation detection [115]. TAPS utilizes ten-eleven translocation (TET) oxidation and pyridineborane reduction to convert methylated C to dihydrouracil (DHU) (read as T in sequencing), while unmethylated C remains unchanged, thus enabling DNA methylation detection [116]. By combining with LRS, it is possible to preserve the adequate length of DNA fragments and enable the PCR amplification safely, reducing the amount of DNA or cell input [117–120]. In addition, a previous study suggested that applying the bead beating approach in library preparation could reduce the input needed by 93.75% in an ideal situation [121]. Third, unbalanced detection capability across DNA methylation types. Although LRS theoretically supports multiple methylation modifications, the development and validation of the tool have been highly focused on 5mCpG. 5mCpG is currently the most thoroughly studied and clinically validated epigenetic marker, the detection kits based on CpG methylation in the promoter regions of SEPT9 and NDRG4/BMP3 genes have been approved by the Food and Drug Administration (FDA) for colorectal cancer screening [122, 123]. Although 5mCpG remains the main target for current clinical translation, the independent diagnostic value of modifications such as 5hmC and non-CpG methylation is increasingly drawing attention, but the development of related detection tools is lagging. 5hmC is an important intermediate product of active demethylation. In recent years, it has shown certain independent diagnostic potential in cancer liquid biopsy. By differentiating and detecting 5mC and 5hmC in cfDNA, the AUROC for detecting stage I colorectal cancer can be improved from 0.66 (combined detection of 5mC and 5hmC) to 0.95 [124]. Its independent diagnostic value has also been verified in multiple other cancer types, such as lung cancer and liver cancer [125]. However, currently only a few tools (such as Dorado) support 5hmC detection, and their accuracy and generalization performance in different samples still require systematic benchmark evaluation. The detection gap for non-CpG methylation is also significant. Non-CpG methylation accounts for over 50% of the total methylated cytosines in mature neurons. It constitutes an independent neural regulatory pathway that is independent of CpG methylation and is closely related to neurodevelopmental disorders such as Rett syndrome [126, 127]. However, the existing non-CpG methylation detection tools are mainly developed for plants, and their validation is still lacking in the mammalian system. Overall, the current LRS-based methylation detection mainly focuses on 5mCpG, which basically meets the existing application requirements. However, the detection capabilities for 5hmC and non-CpG methylation modifications are still significantly lagging. While 5mCpG detection remains important, future tool development should also pay more attention to enhancing detection capabilities for other modifications with clear biological and clinical significance. Although current 5mCpG detection tools are widely used, most of them are based on DL models, which are inclined to overfitting and were built based on limited overlapped datasets, the extrapolation of them is questionable [66, 80]. For this reason, it’s essential to establish a reliable and unbiased ground truth dataset for conducting systematic benchmarking to evaluate the performance of LRS-based methylation detection. It should cover multiple tissues or commonly used cell lines from species such as human and mouse, and get validated through cross-comparison with multiple base modification detection technologies, including NGS, LRS, quantitative polymerase chain reaction (qPCR), pyrosequencing, and liquid chromatography-mass spectrometry (LC-MS). Fourth, the LRS data analysis is computing- and resources-intensive. Even with GPU acceleration, the implementation remains expensive and lengthy, which urges the need for algorithm innovation [69]. In addition, data storage formats can also be optimized. ONT has replaced the FAST5 format with POD5, which offers faster read/write speeds, lower computational overhead, and lower disk usage than FAST5 [128, 129]. However, the original signal data volume of ONT remains extremely large. A single PromethION flow cell operating at full capacity can generate approximately 2 TB of POD5 raw data, and even a small 10 Gbases run requires about 70 GB of disk usage, posing significant challenges to the data management and analysis infrastructure [130, 131]. Compared with POD5, the new raw data formats SLOW5/BLOW5 achieve comparable disk usage while further increasing analysis speed by approximately 4.7-fold and reducing memory consumption by approximately 12.5-fold (from 374 GB to 30 GB), thereby significantly alleviating the computing resource pressure without increasing the storage burden [132]. Fifth, the lack of standardization and harmonization. While the wet-bench and dry-bench experiments of LRS-based methylation detection are growing rapidly, there is no study establishing the objective criteria or quality control systems and addressing the core issues, such as the quantitative accuracy of single-base methylation detection, which results in the dilemma of choosing and applying. Specifically, in wet-bench experiments, variations in DNA extraction, library preparation, and sequencing parameter settings across different laboratories limit the comparability and reproducibility of results. In dry-bench experiments, the algorithms and tools used for basecalling, quality control, sequence alignment, methylation detection, and other processes are still undergoing continuous advancement and have yet to establish a consensus standard [133]. Therefore, it is important to establish standard operating procedures (SOPs) and quality control systems that cover the entire process. Looking ahead, the realization of the potential of LRS technologies in epigenetics requires several breakthroughs through interdisciplinary collaboration. The sequencing cost and amount of DNA material should be addressed first. For example, offering a flexible combination of interested targets [134]. Additionally, future research should focus on developing methylation detection algorithms that enhance efficiency while ensuring accuracy. For the development of wet-bench and dry-bench solutions, unbiased ground truth datasets and comprehensive evaluation systems should be established as soon as possible [83]. Future perspectives The development of LRS-based methylation detection is moving towards breaking through the limitations of traditional technologies. Currently, the main progress is taking place in three key areas. New technologies that simplify the process and enhance accessibility, reference genomes that are increasingly complete and representative, and multi-omics strategies that add functional association. These developments collectively determine how LRS-based methylation detection transforms from a powerful detection tool into a technology with wide applicability. First, multi-faceted coordination efforts are facilitating the accessibility of LRS-based methylation detection. The throughput and accuracy of the two major LRS platforms have significantly improved compared to earlier versions, providing a solid performance foundation for the large-scale application of methylation detection. In addition, simplifying the experimental process and lowering the application threshold have also become important development directions. For example, the ONT adaptive sampling technology eliminates the cumbersome steps of experimentally capturing or amplifying the target region. By simply pre-defining the target sequence, the system can immediately complete the screening during the hardware operation through signal processing and algorithms, without any additional costs or time consumption [56, 135, 136]. In the field of methylation detection, the clinical application value of this technology has been proven. It can complete the methylation typing of brain tumors within 30 min during the operation, with an accuracy rate of 94.6% [100]. Therefore, the evaluation of LRS-based methylation detection no longer solely focuses on technical performance, but also needs to take into account factors such as accessibility and cost, to promote the wider application of long-read methylation detection. Second, the evolution of reference genome frameworks is expanding the boundaries of methylome analysis. The current LRS-based methylation analysis process still uses a single linear reference genome (such as GRCh38, T2T-CHM13) as the framework for read alignment. GRCh38 contains numerous sequence gaps, making it difficult to accurately capture the methylation signals in the highly repetitive regions. The T2T-CHM13 released in 2022 filled these gaps. It added approximately 225 million bp (Mbp) of sequence compared to GRCh38, increasing the number of detectable CpG sites by approximately 10% (~ 3.2 million) in LRS and enabling methylation profiling of previously unresolved regions [11, 40]. However, T2T-CHM13 is essentially still a single linear reference, which is unable to cover the entire genetic diversity of the human genome. Especially in regions with structural variations, some methylation signals may fail to be accurately interpreted due to the absence of the corresponding allelic sequences, thereby resulting in systematic omissions of epigenetic information. The comparative analysis based on the same short-read sequencing data indicates that, compared with T2T-CHM13, each sample using the genome graph consisting of 94 haplotypes can further resolve approximately 3.9% more CpG sites [137]. In addition, an independent WGBS study also demonstrated that when the same methylation reads were aligned to GRCh38 and a genome graph containing 44 diploid genomes, the latter could additionally resolve approximately 2.2–2.6 million previously inaccessible CpG sites per sample [138]. Currently, there is a lack of long-read methylation data analysis tools based on the genome graph. However, given the inherent advantages of LRS in spanning structural variation regions, the future integration of the genome graph into LRS-based methylation analysis holds promise for substantially enhancing the completeness of methylome characterization. Third, LRS facilitates the simultaneous acquisition and functional association of methylation with multi-omics information. Traditional research usually relies on different experimental platforms to obtain methylome, genome, and transcriptome information separately, and then integrates these data to infer their functional associations. However, this strategy lacks corresponding relationships at the same molecular level and is prone to being affected by batch effects. Recently, the Fiber-seq method developed based on the PacBio HiFi platform has the ability to simultaneously obtain genome, CpG methylome, chromatin epigenome, and transcriptome information at the haplotype level during sequencing on a single Revio chip [139, 140]. Its true advantage lies not only in the simultaneous acquisition of multi-omics data, but also in the ability to directly establish functional connections between genetic variations, epigenetic regulation, and transcriptional output on the same DNA molecule. This enables the development of a new research framework for understanding the pathogenic mechanisms of Mendelian diseases and complex disorders. However, this method still requires multiple independent tools to handle different data types separately. The integration and optimization of the analysis process will be an important consideration for improving efficiency and repeatability. In conclusion, these advancements collectively point to a single direction. LRS-based methylation detection will no longer merely be a technical branch of epigenetic research, instead, it will become a routine and scalable core tool in clinical diagnosis and basic research. Concluding remarks LRS technologies have great potential in epigenetic research. The question has shifted from “Are LRS technologies useful?” to “When and how is it useful?” Biological objectives, sample quality, timing, and cost considerations must all be weighed together. It is critical to address the challenges faced by LRS-based methylation detection to achieve early clinical translation and make it a powerful tool for disease diagnosis and enable individualized treatment. In summary, it is foreseeable that with technological innovation and the advancement of methylation detection algorithms, exciting discoveries will undoubtedly emerge in the future. Acknowledgements The authors gratefully acknowledge the financial support from Beijing Natural Science Foundation (L254025), the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0533100/2024ZD0533102), and the National Key R&D Program of China (2023YFC3402503). Illustrations in this review were created by BioRender. Abbreviations Apolipoprotein B mRNA editing enzyme catalytic subunit 3 A Allele-specific methylation Area under the receiver operating characteristic curve Bidirectional Encoder Representations from Transformers Bidirectional long short-term memory Bidirectional recurrent neural network Bisulfite sequencing Cytosine Cyclic coherence sequencing Cell-free DNA Convolutional neural network Cytosine-phosphate-guanine Central processing units Cancer Research Technology Program Dihydrouracil Deep learning Differential methylation region Deep neural network Enzymatic methylation sequencing Food and Drug Administration Graphics processing units Gated recurrent unit High-fidelity Hierarchical Dirichlet process Holistic kinetic Hidden Markov model Interpulse duration Kilobase Liquid chromatography-mass spectrometry Long-read sequencing Megabase Machine learning Next-generation sequencing Oxford Nanopore Technologies Pacific Biosciences Pulse width Quantitative polymerase chain reaction Recurrent neural network Ten-eleven translocation Tet methylcytosine dioxygenase 2 TET-assisted pyridine borane sequencing Uracil World Health Organization Whole genome bisulfite sequencing Zero-mode waveguides N4-methylcytosine 5-methylcytosine 5-hydroxymethylcytosine 5-methylcytosine in CpG N6-methyladenine Author contributions R. Z. and J. M. L. conceived the study and revised the final manuscript. H. J. H. and Y. F. Z. collected relevant references, drafted the manuscript. Y. M. and Y. X. H. offered crucial content revision and language polishing. Funding This work was supported by Beijing Natural Science Foundation (L254025), the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0533100/2024ZD0533102), and the National Key R&D Program of China (2023YFC3402503). Data availability Not applicable. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare that they have no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Hanjing Hou and Yuanfeng Zhang contributed equally to this work. Contributor Information Jinming Li, Email: jmli@nccl.org.cn. Rui Zhang, Email: ruizhang@nccl.org.cn. References Associated Data Data Availability Statement Not applicable.