# Detecting 6mA and Other Non-Canonical DNA Modifications with Nanopore Sequencing: A Practical Workflow Nanopore sequencing can detect DNA modifications beyond 5-methylcytosine (5mC), including N6-methyladenine (6mA), by analyzing raw electrical current signals as DNA strands pass through protein pores. This workflow guides researchers studying bacterial or plant epigenetics who need to identify non-canonical modifications using Oxford Nanopore Technologies platforms. The practical outcome is a validated set of 6mA calls with known false discovery rates, supported by appropriate negative controls and cross-validation with orthogonal methods where available. Researchers should expect to invest time in model selection, training data preparation, and validation before trusting modification calls for biological interpretation. ## Scope and Reader Context This workflow addresses researchers who have access to nanopore sequencing data and need to detect 6mA and related modifications in bacterial or plant genomes. The methods described apply to both R9 and R10 pore chemistries, though performance characteristics differ between them. A critical assessment of nanopore sequencing for detecting multiple forms of DNA modifications published in 2024 emphasizes that false positive calls can mislead biological interpretations, particularly for low-abundance modifications such as 6mA in most mammalian cell types. The authors of that assessment advocate prioritizing nanopore sequencing for mapping abundant modifications over rare ones in biomedical applications. For bacterial systems where 6mA is an intrinsic and principal epigenetic marker, the situation differs substantially because 6mA abundance is typically much higher than in mammalian cells. The workflow covers data acquisition, basecalling with modification-aware models, specialized detection tools, model training with limited labeled data, validation strategies, and common pitfalls that lead to false positive calls. Researchers working with plant genomes face additional complexity because multiple modification types coexist, requiring detection methods that can distinguish between them. The practical decisions in this workflow apply to both research contexts, with specific guidance for bacterial samples where 6mA is the dominant modification and plant samples where modification diversity is higher. ## At a Glance | Workflow Component | Primary Tools | Key Consideration | Evidence Context | | --- | --- | --- | --- | | Basecalling with modification detection | Dorado, Megalodon | Dorado delivers consistently strong performance for bacterial 6mA profiling across R9 and R10 data | Comprehensive comparison of eight tools for bacterial 6mA profiling published in Nature Communications (2025) | | Specialized 6mA detection | DeepMod, DeepSignal, Nanopolish | Deep neural network approaches outperform hidden Markov models when training data are incomplete | DeepMod achieves approximately 0.9 average precision for 6mA on Escherichia coli data, DeepSignal outperforms Nanopolish in incomplete data settings | | Visualization and analysis | Methylartist | Supports all detectable methylation types including 5mCpG, 5hmC, and 6mA | Methylartist is a consolidated suite for processing, visualizing, and analyzing nanopore-derived modified base calls | | Validation framework | Custom negative controls, false discovery rate estimation | Rationally designed negative controls capturing general background and confounding modifications are essential | Critical assessment framework highlights false discovery rate with negative controls for rigorous evaluation | | Cross-reference validation | SMRT sequencing, 6mA-IP-seq, DR-6mA-seq | Multi-dimensional evaluation across six bacterial strains reveals performance variation at single-base resolution | Third-generation sequencing tools comparison includes cross-reference with orthogonal methods | ## Understanding Non-Canonical DNA Modifications in Nanopore Data ### The Biological Relevance of 6mA and Other Modifications DNA N6-methyladenine serves as an intrinsic and principal epigenetic marker in prokaryotes, impacting various biological processes including replication, repair, transcription, and host-pathogen interactions. Unlike 5mC which is the dominant modification in mammalian genomes, 6mA is the primary methylation type in many bacterial species. Plant genomes present a more complex picture with multiple modification types present simultaneously, requiring detection methods that can distinguish between them. The detection challenge stems from the physical mechanism of nanopore sequencing. As a DNA strand translocates through a protein pore, the electrical current is modulated by the specific nucleotide sequence and any chemical modifications present. Modified bases produce characteristic current signatures that differ from their unmodified counterparts. These signatures are context-dependent, meaning the same modification can produce different current patterns depending on the surrounding k-mer sequence. ### Why Standard Basecalling May Miss Non-Canonical Modifications Standard basecalling models are trained to identify canonical bases (A, C, G, T) and often include 5mC and 4mC detection as built-in features. Non-canonical modifications such as 6mA require either modification-aware basecalling models or specialized downstream analysis tools. The choice between these approaches depends on your specific research question, the abundance of the modification in your sample, and the availability of training data for your organism. A comprehensive comparison of third-generation sequencing tools for bacterial 6mA profiling evaluated eight tools designed for 6mA identification or de novo methylation detection. This assessment included Nanopore R9 and R10 chemistries, Single-Molecule Real-Time (SMRT) sequencing, and cross-reference with 6mA-IP-seq and DR-6mA-seq. The multi-dimensional evaluation covered motif discovery, site-level accuracy, single-molecule accuracy, and outlier detection across six bacterial strains. While most tools correctly identified motifs, their performance varied at single-base resolution, with SMRT and Dorado consistently delivering strong performance. The study also noted that existing tools cannot accurately detect low-abundance methylation sites. ## Core Principles of Nanopore Modification Detection ### Raw Signal Analysis and K-mer Context Nanopore modification detection relies on analyzing the raw electrical current signal instead of the basecalled sequence alone. Each k-mer context produces a characteristic current distribution, and the presence of a modification shifts this distribution in a detectable way. The relationship between modification status and current signal is learned by machine learning models during training. DeepMod, a bidirectional recurrent neural network with long short-term memory units, was designed to detect DNA modifications using raw electric signals from Oxford Nanopore long-read sequencing data. The developers sequenced a human genome and a Chlamydomonas reinhardtii genome using nanopore sequencing and evaluated DeepMod on three genome types: Escherichia coli, Chlamydomonas reinhardtii, and human. For 5mC detection, DeepMod achieved average precision up to 0.99 for both synthetically introduced and naturally occurring modifications. For 6mA detection, DeepMod achieved approximately 0.9 average precision on Escherichia coli data and demonstrated improved performance compared to existing methods on Chlamydomonas reinhardtii data. ### The Training Data Completeness Problem Nanopore modification callers have typically been trained on data that span all modifications on all possible DNA k-mer backgrounds, representing a complete training dataset. As nanopore technology expands to more epigenetic modifications, such complete training data become infeasible to obtain. This limitation is particularly acute for non-canonical modifications like 6mA where reference datasets with known modification status are limited. Traditional hidden Markov model (HMM) approaches cannot make successful calls for k-mer contexts not seen during training because of their independent emission distributions. Deep neural networks (DNNs), which share parameters across contexts, can generalize better to unseen k-mer contexts. Research published in Open Biology demonstrates that a common DNN approach (DeepSignal) outperforms a common HMM approach (Nanopolish) in the incomplete data setting. The same study proposed a novel hybrid HMM-DNN approach called amortized-HMM that outperforms both pure HMM and DNN approaches on 5mC calling when training data are incomplete. This type of approach is expected to be useful for calling other base modifications such as 5-hydroxymethylcytosine and for simultaneous calling of different modifications. ### False Discovery Rate and Negative Controls The 2024 critical assessment of nanopore sequencing for detecting multiple forms of DNA modifications describes a framework for rigorous evaluation that highlights the use of false discovery rate with rationally designed negative controls. These negative controls should capture both general background and confounding modifications. The assessment found that while nanopore sequencing performs reliably for high-abundance modifications including 5mC at CpG sites in mammalian cells and 5hmC in mammalian brain cells, it makes a significant proportion of false positive detections for low-abundance modifications such as 5mC at CpH sites, 5hmC, and 6mA in most mammalian cell types. For bacterial samples where 6mA is abundant, the false positive problem is less severe but still requires attention. The Nature Communications comparison of eight tools for bacterial 6mA profiling found that while most tools correctly identify motifs, their performance varies at single-base resolution. The study introduced an optimized method for advancing 6mA prediction that substantially improves the detection performance of Dorado. ## Practical Workflow for 6mA Detection ### Step 1: Data Acquisition and Quality Assessment Begin with raw nanopore data in FAST5 or POD5 format. Ensure that your sequencing run has adequate coverage for modification detection. For bacterial genomes, 50x to 100x coverage is typically sufficient for reliable modification calling at abundant sites. For plant genomes with larger size and more complex modification patterns, higher coverage may be necessary. Assess data quality using standard metrics including read length distribution, quality scores, and throughput. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible workflow training and analysis tutorials for nanopore data processing that can help establish quality assessment protocols. The [European Bioinformatics Institute](https://www.ebi.ac.uk/training) offers bioinformatics learning pathways and data-resource training that cover fundamental concepts in sequence analysis. ### Step 2: Basecalling with Modification-Aware Models The choice of basecalling software and model significantly impacts downstream modification detection. Dorado, the official ONT basecaller, includes modification-aware models that can detect 6mA in addition to canonical bases. The Nature Communications comparison found that Dorado consistently delivers strong performance for bacterial 6mA profiling across both R9 and R10 pore chemistries. When selecting a basecalling model, consider the following factors: - Pore chemistry version (R9.4 or R10.4) - Whether you need simplex or duplex basecalling - The specific modifications you need to detect - Whether you plan to use downstream specialized tools that require specific basecall formats For researchers using specialized detection tools like DeepMod or DeepSignal, basecalling with a standard model (without modification detection) may be appropriate, as these tools analyze raw signals directly. For researchers using Dorado's built-in modification detection, the basecalling step produces both sequence and modification calls simultaneously. ### Step 3: Selecting Specialized Detection Tools The choice of detection tool depends on your organism, modification abundance, and available computational resources. The Nature Communications comparison evaluated eight tools for 6mA identification or de novo methylation detection, providing a useful framework for tool selection. For bacterial samples with abundant 6mA, Dorado with its optimized 6mA prediction method offers strong performance. The optimized method introduced in the comparison study substantially improves Dorado's detection performance, making it a recommended first choice for bacterial 6mA profiling. For research contexts where training data are incomplete or where you need to detect multiple modification types simultaneously, deep neural network approaches like DeepSignal or hybrid approaches like amortized-HMM may be more appropriate. These methods share parameters across k-mer contexts and can generalize to unseen contexts better than HMM-based approaches. For visualization and integrated analysis of modified base calls, [Methylartist](https://pubmed.ncbi.nlm.nih.gov/35482479) provides a consolidated suite of tools for processing, visualizing, and analyzing nanopore-derived modified base calls. All detectable methylation types including 5mCpG, 5hmC, and 6mA are supported, enabling integrated study of base pairs when modified naturally or as part of an experimental protocol. Methylartist is implemented in Python and is installable via PyPI and bioconda. ### Step 4: Model Training with Limited Labeled Data When working with organisms that lack established modification reference data, you may need to train or fine-tune detection models. The [Open Biology study](https://pubmed.ncbi.nlm.nih.gov/38862018) on effective training of nanopore callers with limited labeled data provides practical guidance for this scenario. The key insight from this research is that deep neural network approaches outperform hidden Markov model approaches when training data are incomplete. If you have limited labeled data, prioritize DNN-based callers over HMM-based callers. The hybrid HMM-DNN approach (amortized-HMM) offers additional improvements for 5mC calling and is expected to be useful for other modifications. For training data preparation, consider the following strategies: - Use organisms with well-characterized modification profiles as positive controls - Generate negative controls through PCR amplification or enzymatic removal of modifications - Use publicly available datasets from [NCBI](https://www.ncbi.nlm.nih.gov/) for organisms with established modification maps - Consider synthetic DNA with known modification status for controlled training The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide official descriptions of databases, search systems, sequence resources, and analysis services that can help locate appropriate training data. ### Step 5: Validation with Negative Controls and Orthogonal Methods Validation is the most critical step in the 6mA detection workflow. The [2024 critical assessment framework](https://pubmed.ncbi.nlm.nih.gov/39605700) emphasizes the use of false discovery rate with rationally designed negative controls that capture both general background and confounding modifications. Design negative controls that match your experimental samples as closely as possible. For bacterial 6mA detection, appropriate negative controls might include: - A strain with known methylation genes knocked out - PCR-amplified DNA that lacks native modifications - DNA treated with enzymes that remove specific modifications - Synthetic DNA with known modification status Cross-reference your nanopore-based 6mA calls with orthogonal methods where available. The [Nature Communications comparison](https://pubmed.ncbi.nlm.nih.gov/40295502) used 6mA-IP-seq and DR-6mA-seq as cross-reference methods for validating nanopore and SMRT sequencing results. SMRT sequencing provides an independent long-read platform for modification detection and can serve as a valuable validation method. The multi-dimensional evaluation approach described in the Nature Communications study encompasses motif discovery, site-level accuracy, single-molecule accuracy, and outlier detection. Apply these evaluation dimensions to your own data to understand the strengths and limitations of your detection approach. ### Step 6: Interpretation and Reporting When interpreting 6mA detection results, consider the abundance of the modification in your sample. The critical assessment found that nanopore sequencing performs reliably for high-abundance modifications but makes significant proportions of false positive detections for low-abundance modifications. For bacterial samples where 6mA is abundant, confidence in positive calls is higher than for mammalian samples where 6mA is rare. Report your results with appropriate caveats about false discovery rates. Include information about: - The specific tools and models used for detection - Coverage and data quality metrics - Negative control results - False discovery rate estimates - Validation with orthogonal methods where performed The [Bioconductor project](https://bioconductor.org/) provides official package, workflow, installation, and reproducible genomic-analysis documentation that can support reproducible analysis practices. The [nf-core documentation](https://nf-co.re/docs) describes community pipeline standards, usage, configuration, and reproducible workflow context that can help structure your analysis pipeline. ## Options and Tradeoffs in Tool Selection ### Dorado versus Specialized Tools Dorado offers the advantage of integrated basecalling and modification detection in a single step. The Nature Communications comparison found that Dorado consistently delivers strong performance for bacterial 6mA profiling, and the optimized method introduced in that study substantially improves its detection performance. For researchers who want a streamlined workflow with minimal tool integration effort, Dorado is a strong choice. Specialized tools like DeepMod and DeepSignal offer more flexibility for research contexts where standard models may not perform well. [DeepMod](https://pubmed.ncbi.nlm.nih.gov/31164644) achieves approximately 0.9 average precision for 6mA detection on Escherichia coli data and has demonstrated improved performance compared to existing methods on Chlamydomonas reinhardtii data. These tools require raw signal data and may require more computational resources and bioinformatics expertise to implement. ### R9 versus R10 Pore Chemistry The Nature Communications comparison included both Nanopore R9 and R10 chemistries in its evaluation of 6mA detection tools. Performance characteristics differ between these chemistries, and the choice of pore chemistry affects both basecalling accuracy and modification detection sensitivity. R10 chemistry offers improved basecalling accuracy compared to R9, which can improve modification detection by reducing sequence context errors. However, modification detection models may be more mature for R9 data due to longer history of use. Consider your specific research question and the availability of validated models for your pore chemistry when making this choice. ### HMM versus Deep Neural Network Approaches The Open Biology study provides clear guidance on the tradeoff between HMM and DNN approaches. HMM approaches like Nanopolish cannot make successful calls for k-mer contexts not seen during training because of their independent emission distributions. DNN approaches like DeepSignal share parameters across contexts and can generalize better to unseen contexts. For research contexts where complete training data are not available, prioritize DNN approaches. The hybrid HMM-DNN approach (amortized-HMM) offers additional improvements and is expected to be useful for calling modifications where complete training data are not likely to be available. ## Records and Measurements for Quality Control ### Essential Metrics to Track Maintain detailed records of the following metrics throughout your 6mA detection workflow: - Sequencing coverage and read depth at modification sites - Basecalling quality scores and read length distributions - Modification call confidence scores - False discovery rate estimates from negative controls - Motif enrichment statistics for known methylation motifs - Site-level accuracy compared to orthogonal methods ### Coverage Requirements Coverage requirements for 6mA detection depend on the abundance of the modification and the detection tool used. For abundant modifications in bacterial genomes, lower coverage may be sufficient. For low-abundance modifications or for tools with higher false positive rates, higher coverage is necessary to achieve statistical confidence. The critical assessment of nanopore sequencing for detecting multiple forms of DNA modifications emphasizes that false positive calls can mislead biological interpretations. Adequate coverage combined with appropriate negative controls is essential for distinguishing true modification signals from background noise. ### Quality Control Checkpoints Establish quality control checkpoints at each stage of the workflow: 1. After basecalling: verify read quality and alignment statistics 2. After modification calling: check modification frequency distributions 3. After validation: compare positive and negative control results 4. Before biological interpretation: confirm motif enrichment patterns match expected methylation motifs The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible workflow training and analysis tutorials that can help establish reproducible quality control protocols. The [Carpentries lessons](https://carpentries.org/lessons) offer foundational computing, data, shell, Git, and programming training that supports the computational skills needed for rigorous quality control. ## Common Failure Patterns and Troubleshooting ### False Positive Calls in Low-Abundance Modifications The most significant failure pattern in nanopore modification detection is the generation of false positive calls for low-abundance modifications. The [2024 critical assessment](https://pubmed.ncbi.nlm.nih.gov/39605700) found that nanopore sequencing makes a significant proportion of false positive detections for 6mA in most mammalian cell types. This problem is less severe in bacterial samples where 6mA is abundant, but it still requires attention. If you observe unexpected modification calls in samples where the modification should be absent, investigate the following: - Are your negative controls properly designed to capture background signal? - Is your detection tool appropriate for the abundance level of your modification? - Are there confounding modifications that could produce similar signals? ### Poor Performance at Single-Base Resolution The [Nature Communications comparison](https://pubmed.ncbi.nlm.nih.gov/40295502) found that while most tools correctly identify motifs, their performance varies at single-base resolution. If your analysis requires single-base resolution, you may need to use multiple tools and compare results to identify reliable calls. The study also found that existing tools cannot accurately detect low-abundance methylation sites. If your research question involves low-abundance sites, consider whether nanopore sequencing is the appropriate method or whether orthogonal approaches like SMRT sequencing might be more suitable. ### Incomplete Training Data When training data are incomplete, HMM-based approaches fail to make successful calls for k-mer contexts not seen during training. If you observe poor performance in specific sequence contexts, consider whether your training data adequately cover those contexts. The [Open Biology study](https://pubmed.ncbi.nlm.nih.gov/38862018) demonstrates that DNN approaches outperform HMM approaches in the incomplete data setting. If you are using an HMM-based approach and observing context-specific failures, consider switching to a DNN-based approach or the hybrid amortized-HMM approach. ### Motif Detection Success with Site-Level Failure A common pattern in bacterial 6mA detection is correct motif identification with incorrect site-level calls. The Nature Communications comparison found that most tools correctly identify motifs, but their performance varies at single-base resolution. This pattern suggests that the detection tool can recognize the sequence context of methylation but struggles to determine which specific sites within the motif are modified. If you observe this pattern, consider using multiple detection tools and requiring consensus calls across tools for high-confidence site-level predictions. Cross-reference with orthogonal methods like SMRT sequencing can provide additional validation. ## Limitations and Interpretation Boundaries ### Abundance-Dependent Reliability The reliability of nanopore-based modification detection depends heavily on modification abundance. The [2024 critical assessment](https://pubmed.ncbi.nlm.nih.gov/39605700) advocates prioritizing nanopore sequencing for mapping abundant over rare modifications in biomedical applications. For bacterial 6mA where the modification is abundant, nanopore sequencing can provide reliable results. For mammalian 6mA where the modification is rare, nanopore sequencing produces significant false positive detections and should be interpreted with extreme caution. ### Tool-Specific Performance Variation The Nature Communications comparison of eight tools for bacterial 6mA profiling revealed substantial performance variation across tools. While most tools correctly identify motifs, their performance varies at single-base resolution. SMRT and Dorado consistently deliver strong performance, but other tools may be more appropriate for specific research questions or organism types. The optimized method introduced in the Nature Communications study substantially improves the detection performance of Dorado. If you are using Dorado for bacterial 6mA profiling, consider implementing this optimized method to improve detection performance. ### Cross-Species Transferability Detection models trained on one organism may not perform well on other organisms due to differences in k-mer composition, modification patterns, and sequence context. The [DeepMod study](https://pubmed.ncbi.nlm.nih.gov/31164644) evaluated the tool on three genome types (Escherichia coli, Chlamydomonas reinhardtii, and human genomes) and found performance varied across organisms. When applying detection tools to a new organism, validate performance using appropriate controls and consider whether model retraining or fine-tuning is necessary. ## Safety and Regulatory Context ### Data Management and Reproducibility Nanopore modification detection involves large datasets and complex analysis workflows. Implement reproducible analysis practices using version-controlled pipelines and documented parameters. The [nf-core documentation](https://nf-co.re/docs) describes community pipeline standards, usage, configuration, and reproducible workflow context that can help structure your analysis. The [Bioconductor project](https://bioconductor.org/) provides official package, workflow, installation, and reproducible genomic-analysis documentation. ### Ethical Use of Epigenetic Data Epigenetic data can reveal information about organism health, development, and environmental exposures. When working with human samples, ensure appropriate ethical approvals and data protection measures are in place. For bacterial and plant research, consider any relevant biosafety regulations for your organisms and modifications. ### Professional Escalation Criteria Seek expert consultation when: - You observe unexpected modification patterns that could indicate sample contamination or technical artifacts - Your validation results show high false discovery rates that cannot be resolved through standard troubleshooting - You need to make biological interpretations that depend on low-abundance modification calls - You are applying detection tools to a new organism or modification type without established protocols The [European Bioinformatics Institute](https://www.ebi.ac.uk/training) offers bioinformatics learning pathways and data-resource training that can help build the expertise needed for advanced modification detection analysis. ## Decision Framework for Selecting Detection Tools Based on Modification Abundance and Research Goals Selecting the right detection tool for 6mA and other non-canonical modifications requires a structured decision process that accounts for modification abundance, genome complexity, available training data, and the specific biological question being addressed. The 2025 Nature Communications comparison of eight tools for bacterial 6mA profiling provides a useful evidence base for this decision, but the findings must be interpreted within the context of your specific experimental system. This section presents a practical decision framework that researchers can apply before committing computational resources to a particular detection strategy. ### Tier 1 Decision: Modification Abundance Assessment The first decision point concerns the expected abundance of the target modification in your sample. The 2024 critical assessment of nanopore sequencing for detecting multiple forms of DNA modifications found that nanopore sequencing performs reliably for high-abundance modifications including 5mC at CpG sites in mammalian cells and 5hmC in mammalian brain cells, but makes a significant proportion of false positive detections for low-abundance modifications such as 5mC at CpH sites, 5hmC, and 6mA in most mammalian cell types. This abundance-dependent reliability directly informs tool selection. For bacterial samples where 6mA serves as an intrinsic and principal epigenetic marker, abundance is typically high enough that nanopore-based detection can produce reliable results. The Nature Communications comparison evaluated eight tools across six bacterial strains and found that most tools correctly identify motifs, with SMRT and Dorado consistently delivering strong performance. However, the same study noted that existing tools cannot accurately detect low-abundance methylation sites. If your research question involves rare methylation events within a bacterial genome, you should expect reduced accuracy regardless of tool choice. For plant genomes where multiple modification types coexist, abundance assessment becomes more complex because you must consider the relative abundance of each modification type you intend to detect. A modification that is abundant in one sequence context may be rare in another, and the detection tool must handle this variation. The DeepMod study demonstrated that a bidirectional recurrent neural network with long short-term memory units achieved approximately 0.9 average precision for 6mA detection on Escherichia coli data and improved performance compared to existing methods on Chlamydomonas reinhardtii data, suggesting that neural network approaches can handle the complexity of plant genomes where multiple modifications coexist. ### Tier 2 Decision: Training Data Availability The second decision point concerns the availability of labeled training data for your organism and modification type. The Open Biology study on effective training of nanopore callers with limited labeled data provides critical guidance here. Nanopore modification callers have typically been trained on data that span all modifications on all possible DNA k-mer backgrounds, representing a complete training dataset. As nanopore technology expands to more epigenetic modifications, such complete training data become infeasible to obtain. If you have access to complete training data for your organism and modification type, you can use any detection tool with confidence. If your training data are incomplete, the choice of tool becomes more consequential. The Open Biology study demonstrated that a common deep neural network approach (DeepSignal) outperforms a common hidden Markov model approach (Nanopolish) in the incomplete data setting. This is because HMM approaches cannot make successful calls for k-mer contexts not seen during training due to their independent emission distributions, while DNN approaches share parameters across contexts and can generalize better to unseen contexts. The same study proposed a novel hybrid HMM-DNN approach called amortized-HMM that outperforms both pure HMM and DNN approaches on 5mC calling when training data are incomplete. This hybrid approach is expected to be useful for calling other base modifications such as 5-hydroxymethylcytosine and for simultaneous calling of different modifications, settings in which complete training data are not likely to be available. For practical decision making, assess your training data situation honestly. If you are working with a well-characterized bacterial strain with established methylation maps, you may have access to sufficient labeled data for any tool. If you are working with a less characterized organism or a modification type with limited reference data, prioritize DNN-based approaches or the hybrid amortized-HMM approach. ### Tier 3 Decision: Required Resolution and Validation Capacity The third decision point concerns the resolution required for your biological question and your capacity for orthogonal validation. The Nature Communications comparison found that while most tools correctly identify motifs, their performance varies at single-base resolution. If your research question requires single-base resolution, you need tools that perform well at this level. The same study found that SMRT and Dorado consistently deliver strong performance, making them preferred choices when single-base resolution is critical. However, single-base resolution comes with validation requirements. The 2024 critical assessment framework emphasizes the use of false discovery rate with rationally designed negative controls that capture both general background and confounding modifications. If you cannot implement robust negative controls, you should be more conservative in your tool selection and interpretation. Consider the following validation capacity levels: - Level 1 validation: You have access to a strain with known methylation genes knocked out, providing a definitive negative control. This supports confident interpretation of positive calls. - Level 2 validation: You can generate PCR-amplified DNA that lacks native modifications or treat DNA with enzymes that remove specific modifications. This provides useful but less definitive controls. - Level 3 validation: You have access to orthogonal methods such as SMRT sequencing, 6mA-IP-seq, or DR-6mA-seq for cross-reference. The Nature Communications comparison used these methods for cross-reference validation across six bacterial strains. - Level 4 validation: You have no independent validation capacity and must rely on statistical measures and motif enrichment patterns. Match your tool selection to your validation capacity. If you have limited validation capacity, choose tools with demonstrated strong performance such as Dorado with the optimized method introduced in the Nature Communications study. If you have robust validation capacity, you can explore a wider range of tools and use consensus approaches across multiple tools. ### Practical Decision Matrix The following decision matrix summarizes the tool selection process based on the three tiers described above. This matrix is designed for researchers who need to make practical decisions before committing computational resources. | Scenario | Modification Abundance | Training Data | Validation Capacity | Recommended Approach | | --- | --- | --- | --- | --- | | Bacterial strain with abundant 6mA | High | Complete or incomplete | Level 1 or 2 | Dorado with optimized 6mA prediction method | | Bacterial strain with rare methylation sites | Low at specific sites | Complete | Level 3 | Multiple tools with consensus calls, cross-reference with SMRT | | Plant genome with multiple modifications | Variable by context | Incomplete | Level 2 or 3 | DeepSignal or amortized-HMM for generalization | | New organism without established protocols | Unknown | Limited or absent | Level 4 | Start with Dorado, validate with motif enrichment, escalate for expert consultation | | Mammalian sample with rare 6mA | Low | Complete | Level 3 | Consider whether nanopore is appropriate, prioritize abundant modifications | ### Implementation Steps for the Decision Framework Apply the following steps when implementing this decision framework for your research project: 1. Document the expected abundance of your target modification based on published literature for your organism or closely related species. For bacterial systems, consult established methylation maps available through [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) which provide official descriptions of databases, search systems, sequence resources, and analysis services. 2. Inventory your available training data. Determine whether you have access to labeled data spanning all k-mer contexts for your organism and modification type. If not, note which contexts are likely underrepresented. 3. Assess your validation capacity honestly. List the negative controls and orthogonal methods available to you and assign yourself a validation level from 1 to 4. 4. Select your primary detection tool based on the decision matrix above. Document your selection rationale in your laboratory notebook or project documentation. 5. Define success criteria before running your analysis. Specify acceptable false discovery rates based on your negative controls and the abundance of your target modification. 6. Plan for troubleshooting by identifying which failure patterns are most likely given your scenario. The Nature Communications comparison found that most tools correctly identify motifs but vary at single-base resolution, so plan for potential site-level discrepancies. 7. Schedule validation experiments early in your workflow instead of after completing the full analysis. The 2024 critical assessment framework emphasizes that false positive calls can mislead biological interpretations, so validation should be integrated throughout the workflow instead of added at the end. ### Record System for Tool Selection Decisions Maintain a structured record of your tool selection decisions and their outcomes. This record supports reproducibility and provides a basis for troubleshooting when results are unexpected. Include the following elements in your record: - Organism and strain information including genome assembly version - Expected modification types and their approximate abundance - Pore chemistry version (R9 or R10) and basecalling model used - Primary detection tool and version number - Training data source and completeness assessment - Validation level and specific negative controls used - False discovery rate estimates from negative controls - Motif enrichment results compared to expected methylation motifs - Site-level accuracy compared to orthogonal methods where available - Any troubleshooting actions taken and their outcomes The [Bioconductor project](https://bioconductor.org/) provides official package, workflow, installation, and reproducible genomic-analysis documentation that can support structured record keeping for genomic analyses. The [nf-core documentation](https://nf-co.re/docs) describes community pipeline standards, usage, configuration, and reproducible workflow context that can help structure your analysis pipeline and associated records. ### Common Decision Errors and Their Consequences Several common errors in tool selection lead to poor outcomes in 6mA detection projects. Recognizing these patterns helps researchers avoid them. The first error is selecting a tool based on familiarity instead of suitability for the modification abundance in the sample. A tool that performs well for abundant 5mC in mammalian cells may not perform adequately for 6mA in a bacterial strain with different abundance characteristics. The 2024 critical assessment found that nanopore sequencing makes significant false positive detections for low-abundance modifications, so tool selection must account for abundance. The second error is assuming that training data from one organism transfer to another without validation. The DeepMod study evaluated the tool on three genome types and found performance varied across organisms. Detection models trained on one bacterial species may not perform optimally on other species due to differences in k-mer composition and sequence context. The third error is proceeding with a single tool when the research question requires high confidence at single-base resolution. The Nature Communications comparison found that while most tools correctly identify motifs, their performance varies at single-base resolution. For research questions requiring site-level accuracy, consider using multiple tools and requiring consensus calls across tools. The fourth error is neglecting negative controls until after the main analysis is complete. The 2024 critical assessment framework emphasizes the use of false discovery rate with rationally designed negative controls that capture both general background and confounding modifications. Negative controls should be designed and sequenced alongside experimental samples, not added as an afterthought. ### Escalation Criteria for Tool Selection Uncertainty Seek expert consultation when you encounter the following situations during tool selection: - You cannot determine the expected abundance of your target modification from published literature or preliminary data - Your training data are so incomplete that even DNN-based approaches may struggle to generalize - Your validation capacity is limited to Level 4 and your research question requires single-base resolution - You observe conflicting results between candidate tools during preliminary testing - You are applying detection tools to a modification type or organism combination without established protocols The [European Bioinformatics Institute](https://www.ebi.ac.uk/training) offers bioinformatics learning pathways and data-resource training that can help build the expertise needed for advanced modification detection analysis. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible workflow training and analysis tutorials that can help establish reproducible analysis protocols. The [Carpentries lessons](https://carpentries.org/lessons) offer foundational computing, data, shell, Git, and programming training that supports the computational skills needed for rigorous tool evaluation and selection. ### Integrating the Decision Framework with Existing Workflow Steps This decision framework operates at the beginning of the practical workflow described in the main article, specifically between the data acquisition step and the basecalling step. The framework informs which basecalling model and detection tool you should use, which in turn affects downstream analysis steps. For researchers using Dorado, the decision framework helps determine whether to use the standard modification-aware model or the optimized 6mA prediction method introduced in the Nature Communications study. The optimized method substantially improves Dorado detection performance, making it the preferred choice for bacterial 6mA profiling when available. For researchers using specialized tools like DeepMod or DeepSignal, the decision framework helps determine whether these tools are appropriate given your training data situation and validation capacity. DeepMod achieves approximately 0.9 average precision for 6mA detection on Escherichia coli data, but performance varies across organisms. DeepSignal outperforms Nanopolish in incomplete data settings, making it a better choice when training data are limited. For researchers considering the hybrid amortized-HMM approach, the decision framework helps identify situations where this approach is most beneficial. The Open Biology study found that amortized-HMM outperforms both pure HMM and DNN approaches on 5mC calling when training data are incomplete. This approach is expected to be useful for calling other base modifications and for simultaneous calling of different modifications. ### Practical Example of Framework Application Consider a researcher working with a plant pathogenic bacterium where 6mA is known to be abundant based on published methylation maps. The researcher has access to a knockout strain lacking the primary methylation enzyme, providing a Level 1 negative control. Training data for this specific strain are incomplete, but the researcher has access to data from closely related strains. Applying the decision framework, the researcher would classify the scenario as high modification abundance, incomplete training data, and Level 1 validation capacity. The decision matrix recommends Dorado with the optimized 6mA prediction method as the primary approach. The researcher would document this selection, define success criteria including acceptable false discovery rates based on the knockout strain control, and plan validation experiments using the knockout strain alongside experimental samples. If the researcher later observes unexpected modification calls in the knockout strain, the troubleshooting process would investigate whether the detection tool is appropriate for the abundance level, whether confounding modifications produce similar signals, and whether the optimized method is correctly implemented. The Nature Communications comparison provides context for interpreting these results, noting that while most tools correctly identify motifs, performance varies at single-base resolution. ### Limitations of the Decision Framework This decision framework has several limitations that researchers should acknowledge. First, the framework relies on published evidence from specific organisms and tools, and performance may differ in other contexts. The Nature Communications comparison evaluated eight tools across six bacterial strains, but this does not cover all possible organism-tool combinations. Second, the framework assumes that researchers can accurately assess modification abundance before running their analysis. In practice, abundance may vary across strains, growth conditions, and sequence contexts. Preliminary data or published literature may not fully capture this variation. Third, the framework does not account for computational resource constraints that may limit tool choices. Some tools require substantial computational resources and bioinformatics expertise to implement. Researchers should consider their available infrastructure when applying the framework. Fourth, the framework is designed for research contexts where 6mA is the primary modification of interest. For research questions involving multiple modification types simultaneously, additional considerations apply. The Open Biology study notes that simultaneous calling of different modifications is a setting where complete training data are not likely to be available, making DNN and hybrid approaches particularly relevant. ## Frequently Asked Questions ### What is the difference between 5mC and 6mA detection in nanopore sequencing? 5mC and 6mA produce different electrical current signatures in nanopore sequencing because they are chemically distinct modifications attached to different bases. 5mC is a methyl group added to the fifth carbon of cytosine, while 6mA is a methyl group added to the sixth nitrogen of adenine. Detection models are trained separately for each modification type, and some tools can detect multiple modification types simultaneously. The [2024 critical assessment](https://pubmed.ncbi.nlm.nih.gov/39605700) found that nanopore sequencing performs reliably for high-abundance 5mC at CpG sites in mammalian cells but makes significant false positive detections for 6mA in most mammalian cell types where the modification is rare. ### How much sequencing coverage do I need for reliable 6mA detection in bacteria? Coverage requirements depend on the abundance of 6mA in your bacterial strain and the detection tool you use. For abundant 6mA in bacteria, coverage of 50x to 100x is typically sufficient for reliable detection at most sites. The [Nature Communications comparison](https://pubmed.ncbi.nlm.nih.gov/40295502) of eight tools for bacterial 6mA profiling evaluated performance across six bacterial strains and found that most tools correctly identify motifs, though performance varies at single-base resolution. Higher coverage may be needed for low-abundance methylation sites or when using tools with higher false positive rates. ### Can I use the same detection model for different bacterial species? Detection models trained on one bacterial species may not perform optimally on other species due to differences in k-mer composition and sequence context. The [DeepMod study](https://pubmed.ncbi.nlm.nih.gov/31164644) evaluated the tool on Escherichia coli, Chlamydomonas reinhardtii, and human genomes and found performance varied across organisms. When applying a detection tool to a new species, validate performance using appropriate controls and consider whether model retraining or fine-tuning is necessary. ### What negative controls should I use for 6mA detection validation? Appropriate negative controls for 6mA detection include strains with methylation genes knocked out, PCR-amplified DNA that lacks native modifications, DNA treated with enzymes that remove specific modifications, and synthetic DNA with known modification status. The [2024 critical assessment framework](https://pubmed.ncbi.nlm.nih.gov/39605700) emphasizes the use of rationally designed negative controls that capture both general background and confounding modifications. These controls should match your experimental samples as closely as possible to provide meaningful false discovery rate estimates. ### How do I choose between Dorado and specialized tools like DeepMod? Dorado offers integrated basecalling and modification detection in a single step and consistently delivers strong performance for bacterial 6mA profiling according to the [Nature Communications comparison](https://pubmed.ncbi.nlm.nih.gov/40295502). Specialized tools like DeepMod offer more flexibility for research contexts where standard models may not perform well. [DeepMod](https://pubmed.ncbi.nlm.nih.gov/31164644) achieves approximately 0.9 average precision for 6mA detection on Escherichia coli data. Consider your specific research question, computational resources, and bioinformatics expertise when making this choice. ### What should I do if I detect 6mA in a sample where it should be absent? If you detect 6mA in a sample where it should be absent, first verify that your negative controls are properly designed and that your detection tool is appropriate for the abundance level of your modification. The [2024 critical assessment](https://pubmed.ncbi.nlm.nih.gov/39605700) found that nanopore sequencing makes significant false positive detections for low-abundance modifications. Investigate whether confounding modifications could produce similar signals and consider cross-validation with orthogonal methods like SMRT sequencing. ### How does R10 pore chemistry compare to R9 for 6mA detection? The [Nature Communications comparison](https://pubmed.ncbi.nlm.nih.gov/40295502) included both R9 and R10 pore chemistries in its evaluation of 6mA detection tools. R10 chemistry offers improved basecalling accuracy compared to R9, which can improve modification detection by reducing sequence context errors. However, modification detection models may be more mature for R9 data due to longer history of use. Consider your specific research question and the availability of validated models for your pore chemistry when making this choice. ### Can nanopore sequencing detect 6mA in plant genomes? Nanopore sequencing can detect 6mA in plant genomes, but plant genomes present a more complex picture with multiple modification types present simultaneously. The [DeepMod study](https://pubmed.ncbi.nlm.nih.gov/31164644) evaluated the tool on Chlamydomonas reinhardtii and demonstrated improved performance compared to existing methods for 6mA detection. However, the [2024 critical assessment](https://pubmed.ncbi.nlm.nih.gov/39605700) emphasizes that false positive calls can mislead biological interpretations, particularly for low-abundance modifications. Validate your detection approach with appropriate controls and consider cross-reference with orthogonal methods. ## Related Bioinformatics Guides - [Metabolomics Data Analysis in R: A Practical Workflow](/knowledge/bioinformatics/metabolomics-data-analysis-in-r-a-practical-workflow) - [Single-Cell DNA Sequencing: Applications and Workflow Considerations](/knowledge/bioinformatics/single-cell-dna-sequencing-applications-and-workflow-considerations) - [Single-Cell Sequencing Workflow: From Sample Preparation to Data Analysis](/knowledge/bioinformatics/single-cell-sequencing-workflow-from-sample-preparation-to-data-analysis) - [Proteomics Data Analysis in R: A Practical Workflow for Differential Expression and Visualization](/knowledge/bioinformatics/proteomics-data-analysis-in-r-a-practical-workflow-for-differential-expression-and-visualization) - [Spatial Transcriptomics Data Analysis: A Practical Workflow from Raw Data to Biological Insights](/knowledge/bioinformatics/spatial-transcriptomics-data-analysis-a-practical-workflow-from-raw-data-to-biological-insights) ## References and Further Reading - [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information. - [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute. - [Bioconductor](https://bioconductor.org/). Bioconductor Project. - [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project. - [nf-core Documentation](https://nf-co.re/docs). nf-core. - [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries. - [Critical assessment of nanopore sequencing for the detection of multiple forms of DNA modifications.](https://pubmed.ncbi.nlm.nih.gov/39605700). bioRxiv : the preprint server for biology, 2024. - [Comprehensive comparison of the third-generation sequencing tools for bacterial 6mA profiling.](https://pubmed.ncbi.nlm.nih.gov/40295502). Nature communications, 2025. - [Detection of DNA base modifications by deep recurrent neural network on Oxford Nanopore sequencing data.](https://pubmed.ncbi.nlm.nih.gov/31164644). Nature communications, 2019. - [Effective training of nanopore callers for epigenetic marks with limited labelled data.](https://pubmed.ncbi.nlm.nih.gov/38862018). Open biology, 2024. - [Methylartist: tools for visualizing modified bases from nanopore sequence data.](https://pubmed.ncbi.nlm.nih.gov/35482479). Bioinformatics (Oxford, England), 2022. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.