# Detecting 5-Methylcytosine in Nanopore Sequencing: A Step-by-Step Guide to Methylation Calling with Megalodon and Dorado Nanopore sequencing detects 5-methylcytosine (5mC) directly from native DNA by measuring ionic current shifts as the molecule passes through a protein pore. This protocol guides researchers who need to call 5mC modifications from raw nanopore signal data using Megalodon or Dorado, covering model selection, command-line execution, and output interpretation. The workflow assumes you have basecalled FAST5 or POD5 files from an Oxford Nanopore Technologies (ONT) sequencer and need to produce per-read or per-site methylation calls for downstream analysis. ## Understanding the Biological and Technical Basis of 5mC Detection 5-methylcytosine is an epigenetic modification where a methyl group attaches to the fifth carbon of cytosine. In nanopore sequencing, this modification alters the electrical current signature as the modified base passes through the pore, creating a detectable signal difference from unmodified cytosine. The key advantage of nanopore-based methylation detection is that it works on native DNA without bisulfite conversion or PCR amplification, preserving the original modification state. The detection principle relies on comparing observed current signals against expected signals from a reference model. Basecallers trained on known methylation states can identify 5mC by recognizing the characteristic current perturbations. This approach differs fundamentally from short-read bisulfite sequencing, which converts unmethylated cytosines to uracil and infers methylation from sequence changes. Nanopore methods measure the modification directly at single-molecule resolution. Recent comparisons between ONT and Pacific Biosciences (PacBio) platforms have shown that methylated motif detection from ONT R10.4.1 data can achieve results comparable to PacBio when using appropriate analysis tools. A 2024 study on bacterial genome reconstruction demonstrated that ONT data, particularly with advanced tools like Nanomotif, can reliably detect methylation patterns across the genome. This validates the biological relevance of nanopore methylation calling for research applications. The practical implication for your workflow is that the quality of methylation calls depends on three factors: the sequencing chemistry and flow cell version, the basecaller model you select, and the downstream analysis tool you use for motif detection or differential methylation analysis. Each of these components requires deliberate choices based on your experimental goals. ## Prerequisites and System Requirements Before starting methylation calling, verify that your computing environment meets the requirements for both Megalodon and Dorado. These tools process raw signal data and require substantial computational resources, particularly GPU acceleration for reasonable processing times. ### Hardware Considerations GPU acceleration is strongly recommended for both Megalodon and Dorado. A CUDA-compatible NVIDIA GPU with at least 8 GB of VRAM handles typical flow cell outputs within practical timeframes. CPU-only processing is possible but significantly slower, often by an order of magnitude or more. For a typical PromethION flow cell producing hundreds of gigabytes of raw data, CPU-only processing may take days to weeks, while GPU processing completes in hours. RAM requirements depend on the size of your dataset and the reference genome. A minimum of 32 GB RAM is advisable, with 64 GB or more recommended for human genome-scale projects. Storage requirements are substantial because raw signal files, basecalled reads, and methylation output files all consume significant disk space. Plan for at least three times the raw data size to accommodate intermediate files. ### Software Installation Megalodon is available through the ONT software repository and can be installed via conda or pip. The tool requires Python 3.6 or later and depends on ont-fast5-api, ont-pyguppy-client-lite, and other ONT libraries. Dorado is distributed as a standalone binary and is the current recommended tool from ONT, with regular updates that improve accuracy and add new features. For reproducible analysis environments, consider using containerized installations. The [nf-core documentation](https://nf-co.re/docs) describes community standards for workflow configuration and reproducibility that apply to bioinformatics tool installations. Similarly, [Bioconductor](https://bioconductor.org/) provides package management and reproducible analysis frameworks that can complement your methylation calling pipeline. ### Data Format Requirements Megalodon requires raw signal data in FAST5 format, either multi-read or single-read files. If your sequencer produced POD5 format, you must convert to FAST5 using the pod5-file-format tools before running Megalodon. Dorado accepts both POD5 and FAST5 inputs, making it more flexible for current sequencing platforms. Basecalling can be performed within Megalodon or Dorado, or you can use pre-basecalled data. For methylation calling, the tools need access to the raw signal, so you cannot use basecalled FASTQ files alone. The raw signal files must be organized in a directory structure that the tools can traverse, typically with one directory per flow cell or experiment. ## Selecting the Appropriate Basecalling and Methylation Model Model selection is the most consequential decision in your methylation calling workflow. The model determines how the tool interprets electrical current signals and assigns methylation probabilities. Using an incorrect model for your sequencing chemistry will produce unreliable calls. ### Model Types and Compatibility ONT provides several model categories for methylation detection. The 5mC models are trained to detect 5-methylcytosine in CpG context, which is the primary methylation context in mammalian genomes. Some models also detect 5mC in all contexts, which is relevant for plant genomes and certain microbial systems where non-CpG methylation occurs. The model must match your flow cell version and chemistry. R9.4.1 flow cells use different models than R10.4.1 flow cells because the pore chemistry produces different current signatures. Using an R9.4.1 model with R10.4.1 data will produce incorrect methylation calls, even if basecalling appears to work correctly. Dorado provides a model download command that fetches the appropriate models from ONT servers. The model naming convention includes the flow cell type, chemistry version, and modification type. For example, a model name might indicate R10.4.1 chemistry with 5mC detection in CpG context. Megalodon similarly requires you to specify the model, either by downloading a pre-trained model or using the remora model format. ### Model Selection Decision Framework Consider your experimental context when choosing a model. For human or mammalian DNA samples, select a model trained for CpG context 5mC detection. For plant samples or microbial communities where non-CpG methylation is biologically relevant, select a model that detects 5mC in all sequence contexts. The sequencing platform also matters. A 2024 study comparing ONT and PacBio for bacterial genome reconstruction found that ONT R10.4.1 data with appropriate analysis tools produced methylation results comparable to PacBio. This suggests that current ONT models are suitable for bacterial methylation studies, but you should verify model compatibility with your specific flow cell version. For RNA modifications, note that 5mC detection models are designed for DNA. RNA methylation detection requires different models and analysis approaches. A 2025 study on direct RNA sequencing described tools like SingleMod for detecting N6-methyladenosine (m6A) on individual RNA molecules, but these are separate from the DNA 5mC workflow described here. ### Testing Model Performance Before processing a full dataset, run a small test on a subset of reads to verify that the model produces sensible results. Check that the methylation frequency at known methylated sites matches expected values. For human DNA, imprinted regions or known differentially methylated regions serve as positive controls. For bacterial samples, known methylated motifs from the literature or from PacBio data can validate your calls. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible tutorials on quality assessment and workflow validation that apply to this testing step. Similarly, [The Carpentries](https://carpentries.org/lessons) offers foundational computing lessons that help with scripting the validation steps. ## Running Megalodon for Methylation Calling Megalodon is the established tool for methylation calling from nanopore raw signal data. While ONT has transitioned to Dorado as the primary tool, Megalodon remains in use for existing pipelines and offers features that some researchers prefer. ### Basic Megalodon Command Structure The core Megalodon command specifies the input data directory, the output directory, the reference genome, and the model. A minimal command looks like this: ``` megalodon /path/to/fast5_files \ --output-directory /path/to/output \ --reference /path/to/reference.fasta \ --mod-motif m 5mC CG 0 \ --outputs basecalls mods mod_mappings \ --guppy-config dna_r9.4.1_450bps_hac.cfg \ --remora-model /path/to/remora_model \ --device cuda:0 ``` The `--mod-motif` parameter specifies the modification type, the sequence motif, and the position within the motif. For CpG methylation, the motif is CG and the modified base is at position 0. This parameter tells Megalodon where to look for modified bases in the sequence context. The `--outputs` parameter controls which output files Megalodon generates. The `mods` output produces per-read modification probabilities, while `mod_mappings` produces modification calls mapped to the reference genome. The `basecalls` output provides the sequence reads. ### Megalodon Output Files Megalodon produces several output files in the specified output directory. The per-read modification probabilities are stored in a text file with one line per read, containing the read ID, the position, the modification probability, and the canonical base. The mapped modification calls are stored in BAM format with MM and ML tags that encode the modification status. The BAM file uses the SAMtags specification for modified bases. The MM tag lists the modification type and positions, while the ML tag provides the probability scores. These tags are compatible with downstream analysis tools that understand the modified base specification. For per-site methylation analysis, you will typically aggregate the per-read probabilities across all reads covering each genomic position. This produces a methylation frequency estimate for each CpG site. Tools like ModKit or custom Python scripts can perform this aggregation. ### Megalodon Limitations and Considerations Megalodon is no longer actively developed by ONT, and support for newer flow cells and chemistry versions may be limited. The tool requires FAST5 input, so POD5 data must be converted first. Processing speed is slower than Dorado, particularly for large datasets. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide background on sequence analysis concepts that help interpret Megalodon outputs correctly. Understanding the difference between per-read modification calls and aggregated methylation frequencies is essential for proper analysis. ## Running Dorado for Methylation Calling Dorado is the current recommended tool from ONT for basecalling and methylation detection. It offers improved speed, support for POD5 format, and regular updates that incorporate the latest model improvements. ### Dorado Installation and Model Download Dorado is distributed as a standalone binary for Linux, macOS, and Windows. After downloading and extracting the binary, you must download the appropriate models. The `dorado download` command fetches models from ONT servers: ``` dorado download --model dna_r10.4.1_e8.2_400bps_hac@v5.0.0 ``` This command downloads the specified model to the local models directory. The model name encodes the flow cell version, chemistry, and basecalling mode. The `hac` designation indicates high-accuracy mode, while `fast` and `sup` indicate faster but less accurate and super-accurate modes respectively. ### Dorado Basecalling with Methylation Detection Dorado performs methylation detection during basecalling. The command structure is: ``` dorado basecaller \ dna_r10.4.1_e8.2_400bps_hac@v5.0.0 \ /path/to/pod5_or_fast5_files \ --modified-bases 5mC \ --reference /path/to/reference.fasta \ --output-dir /path/to/output ``` The `--modified-bases 5mC` parameter enables 5mC detection. Dorado outputs BAM files with MM and ML tags containing the methylation calls. The `--reference` parameter enables alignment during basecalling, which produces mapped reads directly. For simplex basecalling, Dorado processes each read independently. For duplex basecalling, which uses information from both strands of the DNA molecule, Dorado requires additional steps. Duplex basecalling improves accuracy but requires that both strands of the same molecule are sequenced, which is not always the case. ### Dorado Output Interpretation The BAM output from Dorado contains the same MM and ML tags as Megalodon. The MM tag uses the SAMtags specification to encode modification positions. For 5mC in CpG context, the tag appears as `MM:Z:C+m?` where the question mark indicates the position of the modified base relative to the motif. The ML tag provides the probability of modification at each position listed in the MM tag. These probabilities range from 0 to 255, where 255 corresponds to a probability of 1.0. Converting these values to methylation frequencies requires dividing by 255. For downstream analysis, you can use the [Bioconductor](https://bioconductor.org/) packages that support modified base BAM files. These packages provide functions for reading MM and ML tags, aggregating per-read calls, and performing differential methylation analysis. ## Comparing Megalodon and Dorado for Your Workflow The choice between Megalodon and Dorado depends on your specific requirements, including data format, flow cell version, and pipeline compatibility. ### Performance Comparison Dorado offers substantially faster processing than Megalodon, particularly with GPU acceleration. For a typical flow cell producing 50 to 100 GB of raw data, Dorado completes basecalling and methylation detection in hours, while Megalodon may take a day or more. This speed difference matters for projects with multiple samples or time-sensitive analyses. Dorado also supports POD5 format natively, eliminating the conversion step required for Megalodon. As ONT transitions to POD5 as the standard format, this compatibility becomes increasingly important. ### Accuracy Considerations Both tools use similar underlying models for methylation detection, so accuracy is comparable when using equivalent model versions. However, Dorado receives regular model updates that incorporate improvements from ONT's ongoing research. For the latest flow cells like R10.4.1, Dorado models are more current than Megalodon models. A 2024 study on target adaptive sampling long-read sequencing demonstrated that ONT workflows can identify genetic variants with accuracy comparable to short-read platforms while also detecting epigenetic alterations. This study used current ONT tools and validated the biological relevance of nanopore methylation calls in a clinical research context. ### Pipeline Integration Consider how the methylation calling tool integrates with your existing analysis pipeline. If you have scripts and workflows built around Megalodon output formats, migrating to Dorado requires updating those scripts. Conversely, if you are starting a new project, adopting Dorado from the beginning avoids future migration work. The [nf-core documentation](https://nf-co.re/docs) describes community standards for pipeline development that help with tool integration and reproducibility. Following these standards ensures that your methylation calling workflow remains maintainable and shareable. ## Quality Control and Validation of Methylation Calls Quality control is essential for reliable methylation analysis. Poor quality calls can arise from sequencing errors, model mismatches, or insufficient coverage. Implementing systematic quality checks prevents downstream analysis errors. ### Coverage and Depth Requirements Methylation frequency estimates require sufficient read coverage at each CpG site. For reliable frequency estimates, a minimum of 10 reads per site is commonly used, though deeper coverage improves confidence. Low coverage sites should be flagged or excluded from downstream analysis. The relationship between coverage and confidence depends on the methylation level. Sites with intermediate methylation levels around 50 percent require more coverage to distinguish true intermediate methylation from sampling noise. Sites with very high or very low methylation levels can be called with fewer reads. ### Control Samples and Known Methylation Standards Include control samples with known methylation status in your experiment. Fully methylated and fully unmethylated controls help calibrate your methylation calls and identify systematic biases. Commercially available methylated and unmethylated DNA standards provide reference materials for this purpose. For human samples, imprinted genes with known parent-of-origin methylation patterns serve as internal controls. For example, the SNRPN gene shows allele-specific methylation that can validate your calls. A 2024 study identified allele-specific MLH1 promoter hypermethylation in a Lynch syndrome patient using long-read sequencing, demonstrating the biological relevance of allele-specific methylation detection. ### Metrics for Assessing Call Quality Several metrics help assess the quality of your methylation calls. The fraction of reads with modification calls at known methylated sites indicates sensitivity. The fraction of reads with modification calls at known unmethylated sites indicates specificity. The balance between sensitivity and specificity depends on the probability threshold you apply. The per-read modification probability distribution provides another quality metric. High-confidence calls cluster near 0 or 1, while low-confidence calls distribute across intermediate values. Examining this distribution helps identify systematic issues with the model or data quality. The [Galaxy Training Network](https://training.galaxyproject.org/) provides tutorials on quality assessment that apply to methylation data. These resources help you implement systematic quality checks in your workflow. ## Aggregating Per-Read Calls to Methylation Frequencies The raw output from Megalodon and Dorado provides per-read modification calls. Most biological analyses require site-level methylation frequencies, which aggregate these per-read calls across all reads covering each genomic position. ### Computing Methylation Frequencies For each CpG site in the reference genome, count the number of reads with a modification call and the total number of reads covering that site. The methylation frequency is the ratio of modified reads to total reads. This calculation requires parsing the MM and ML tags from the BAM file. Several tools perform this aggregation. ModKit from ONT provides a `pileup` command that computes per-site methylation frequencies from modified base BAM files. Custom Python scripts using the pysam library can also perform this calculation, giving you control over filtering and output formats. ### Handling Multi-Mapping Reads and Repetitive Regions Reads that map to multiple genomic locations create ambiguity in methylation frequency calculations. For unique regions, all reads can be assigned confidently. For repetitive regions, multi-mapping reads should be handled carefully, either by assigning them randomly, weighting them, or excluding them from analysis. The choice of multi-mapping read handling depends on your analysis goals. For genome-wide methylation profiles, excluding multi-mapping reads reduces noise but loses information from repetitive regions. For targeted analysis of specific loci, you may want to include multi-mapping reads with appropriate weighting. ### Strand-Specific Methylation Analysis Methylation is often strand-specific, particularly at imprinted loci and in certain disease contexts. The MM and ML tags in BAM files retain strand information through the read alignment. For strand-specific analysis, separate methylation frequencies for forward and reverse strand reads. Allele-specific methylation analysis requires distinguishing reads from different haplotypes. This is possible when heterozygous variants distinguish the parental chromosomes. A 2024 study demonstrated allele-specific MLH1 promoter hypermethylation detection using long-read sequencing, showing that this analysis is feasible with appropriate bioinformatics approaches. ## Common Failure Patterns and Troubleshooting Methylation calling workflows can fail in predictable ways. Recognizing these failure patterns helps you diagnose problems quickly and implement corrective actions. ### Model Mismatch Errors The most common failure is using a model that does not match the sequencing chemistry. Symptoms include low modification call rates, unusual methylation frequency distributions, or complete absence of modification calls. Verify that the model name matches your flow cell version and chemistry. If you are unsure which flow cell version was used, check the sequencing run metadata or the FAST5 file headers. The flow cell version is recorded in the run information and can be extracted using the ont-fast5-api tools. ### Insufficient Coverage at Target Sites Low coverage at specific genomic regions produces unreliable methylation frequency estimates. This often occurs in GC-rich regions or regions with extreme base composition that sequence poorly. Solutions include increasing sequencing depth, using adaptive sampling to enrich target regions, or accepting lower confidence at these sites. A 2024 study on target adaptive sampling long-read sequencing demonstrated that this approach can enrich specific gene regions for intensive analysis. The study applied this method to hereditary cancer patient genomes and identified complex structural variations and epigenetic alterations in target genes. ### Basecalling Errors Affecting Methylation Calls Basecalling errors can create false methylation calls when the sequence context is misidentified. If the basecaller misreads a cytosine as another base, the methylation calling model may not recognize the CpG context. This is more common in homopolymer regions and other challenging sequence contexts. High-accuracy basecalling modes reduce this problem but require more computational resources. The tradeoff between speed and accuracy should be balanced based on your analysis requirements. ### Batch Effects Across Flow Cells or Runs When processing multiple samples across different flow cells or runs, batch effects can introduce systematic differences in methylation calls. These effects arise from variations in sequencing conditions, reagent lots, or instrument performance. Including control samples in each batch helps identify and correct for batch effects. The [Bioconductor](https://bioconductor.org/) project provides statistical methods for batch effect correction that apply to methylation data. These methods help ensure that biological differences are not confounded with technical variation. ## Limitations and Interpretation Boundaries Understanding the limitations of nanopore methylation calling prevents overinterpretation of results and guides appropriate experimental design. ### Detection Context and Specificity The standard 5mC models detect methylation in CpG context for mammalian genomes. Non-CpG methylation, which is biologically relevant in plants, embryonic stem cells, and certain disease states, requires different models or analysis approaches. Verify that your model detects the methylation context relevant to your biological question. The models also detect 5mC specifically, not other cytosine modifications like 5-hydroxymethylcytosine (5hmC). If 5hmC is biologically relevant to your system, you need specialized models or additional experimental approaches to distinguish these modifications. ### Quantitative Accuracy Nanopore methylation calls provide relative methylation levels, not absolute quantitative measurements. The relationship between current signal perturbation and methylation fraction is approximately linear but may have systematic biases. Calibration with known methylation standards helps correct these biases. The accuracy of methylation frequency estimates depends on coverage depth and the true methylation level. Sites with very low or very high methylation levels are called more accurately than sites with intermediate levels. Confidence intervals should be reported for methylation frequency estimates, particularly at intermediate levels. ### Single-Molecule Resolution A key advantage of nanopore sequencing is single-molecule resolution. Each read provides methylation information for an individual DNA molecule, revealing heterogeneity that is masked in bulk measurements. A 2025 study on direct RNA sequencing demonstrated single-molecule detection of RNA modifications, highlighting the power of this approach for understanding modification heterogeneity. However, single-molecule resolution also means that individual reads have higher error rates than aggregated measurements. Per-read methylation calls should be interpreted with appropriate caution, and biological conclusions should be based on aggregated data across multiple reads. ### Reference Genome Dependence Methylation calling requires a reference genome for read mapping and context identification. The quality of the reference genome affects the accuracy of methylation calls. For non-model organisms, a high-quality reference may not be available, limiting the accuracy of methylation analysis. A 2025 metagenomics study on the MAST-3 stramenopile Incisomonas marina demonstrated that complete genome assembly from long-read data enables downstream analysis including methylation detection. This highlights the importance of genome assembly quality for epigenetic analysis. ## Professional Escalation Criteria Certain situations warrant escalation to specialized support or additional expertise. Recognizing these situations prevents wasted effort and ensures appropriate resolution. ### When to Consult ONT Support Contact ONT technical support when you encounter persistent errors in basecalling or methylation detection that you cannot resolve through troubleshooting. This includes model download failures, unexpected tool crashes, or systematic errors across multiple runs. ONT support can also provide guidance on model selection for unusual sample types or experimental designs. Their expertise with the underlying pore chemistry and signal processing helps resolve issues that are not apparent from documentation. ### When to Seek Bioinformatics Consultation Consult a bioinformatics specialist when you need to develop custom analysis approaches beyond standard methylation frequency calculations. This includes allele-specific methylation analysis, integration with other omics data, or development of novel statistical methods. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide pathways for developing bioinformatics skills, but complex analyses may require specialized expertise. Recognizing the limits of your own expertise and seeking appropriate consultation improves the quality of your analysis. ### When to Validate with Orthogonal Methods Validate nanopore methylation calls with orthogonal methods when your results have critical implications or when you observe unexpected patterns. Bisulfite sequencing, methylation-specific PCR, or PacBio methylation detection can confirm nanopore results. A 2024 study comparing ONT and PacBio for methylated motif detection found comparable results between platforms, suggesting that orthogonal validation may not be necessary for routine applications. However, for clinical or high-stakes research applications, independent validation provides additional confidence. ## Records and Documentation for Reproducibility Maintaining detailed records of your methylation calling workflow ensures reproducibility and facilitates troubleshooting. Document all parameters, model versions, and software versions used in each analysis. ### Required Documentation Elements Record the software versions for Megalodon or Dorado, the model names and versions, the reference genome version, and all command-line parameters. Document the sequencing chemistry, flow cell version, and basecalling mode. Record the date of analysis and the computing environment. The [nf-core documentation](https://nf-co.re/docs) emphasizes the importance of version tracking and parameter documentation for reproducible workflows. Following these practices ensures that your analysis can be reproduced by collaborators or reviewers. ### Data Storage and Archiving Store raw signal data, basecalled reads, and methylation calls in organized directory structures with clear naming conventions. Archive intermediate files that may be needed for re-analysis. Document file formats and compression methods used. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide repositories for depositing sequencing data and analysis results. Depositing your data ensures long-term preservation and enables data sharing with the research community. ## At a Glance | Decision Point | Megalodon | Dorado | Recommendation | |----------------|-----------|--------|----------------| | Input format | FAST5 only | POD5 and FAST5 | Use Dorado for POD5 data, Megalodon for existing FAST5 pipelines | | Model updates | Limited, legacy models | Regular updates for current chemistry | Use Dorado for R10.4.1 or newer flow cells | | Processing speed | Slower, CPU or GPU | Faster, GPU optimized | Use Dorado for large datasets or time-sensitive projects | | Output format | Text and BAM with MM/ML tags | BAM with MM/ML tags | Both compatible with downstream analysis tools | | Active development | Discontinued | Active | Adopt Dorado for new projects | ## Practical Implementation Steps Implementing a methylation calling workflow requires systematic planning and execution. Follow these steps to establish a reliable pipeline. ### Step 1: Verify Data Compatibility Check that your raw signal data is in a format compatible with your chosen tool. For Megalodon, convert POD5 to FAST5 if needed. For Dorado, verify that the POD5 or FAST5 files are accessible and complete. Check the flow cell version and chemistry from the run metadata. ### Step 2: Download and Test Models Download the appropriate model for your flow cell and chemistry. Run a test on a small subset of reads to verify that the model produces sensible basecalls and methylation calls. Check that known methylated sites show modification calls and known unmethylated sites do not. ### Step 3: Configure the Analysis Environment Set up the computing environment with sufficient GPU, RAM, and storage resources. Install the chosen tool and verify that it runs correctly. Configure the output directory structure and naming conventions. ### Step 4: Run Methylation Calling Execute the methylation calling command with appropriate parameters. Monitor the progress and check for errors. For large datasets, consider running in batches to manage resource usage and facilitate error recovery. ### Step 5: Validate Output Quality Check the output files for completeness and quality. Verify that the BAM files contain MM and ML tags. Examine the distribution of modification probabilities and coverage across the genome. Compare methylation frequencies at known control sites with expected values. ### Step 6: Aggregate and Analyze Aggregate per-read calls to methylation frequencies using appropriate tools. Perform quality filtering based on coverage and probability thresholds. Conduct downstream analysis appropriate to your biological question. ### Step 7: Document and Archive Document all parameters, versions, and decisions in a laboratory notebook or electronic record. Archive raw data, intermediate files, and final results. Deposit data in appropriate repositories for long-term preservation. ## Common Failure Patterns and Corrective Actions | Failure Pattern | Symptoms | Corrective Action | |-----------------|----------|-------------------| | Model mismatch | No modification calls or unusual frequency distributions | Verify model matches flow cell and chemistry | | Low coverage at target sites | Unreliable frequency estimates at specific loci | Increase sequencing depth or use adaptive sampling | | Basecalling errors | False methylation calls in homopolymer regions | Use higher accuracy basecalling mode | | Batch effects | Systematic differences across flow cells | Include controls in each batch and apply correction methods | | Software version issues | Unexpected errors or output format changes | Document versions and test updates before full analysis | ## Building a Methylation Calling Decision Framework for Mixed Datasets and Multi-Project Laboratories When your laboratory processes samples across multiple flow cell versions, chemistry releases, or organism types, the single most common source of analysis failure is not tool malfunction but inconsistent decision-making about model selection and data handling. A structured decision framework prevents these errors by forcing explicit documentation of experimental context before any command is executed. ### Establishing a Sample-to-Model Matching Protocol Before running Megalodon or Dorado, classify each sequencing run according to three fixed parameters: flow cell version, chemistry version, and expected methylation context. These parameters determine the universe of valid models for your data. Record them in a standardized run manifest that accompanies the raw data through the entire analysis pipeline. The flow cell version is the first filter. R9.4.1 and R10.4.1 flow cells produce fundamentally different current signatures, and models trained on one cannot interpret the other. The chemistry version, such as V14 for current R10.4.1 runs, further refines model compatibility. The methylation context requirement depends on your organism: mammalian samples typically need CpG-context 5mC detection, while plant or microbial samples may require all-context models. A 2024 study comparing ONT and PacBio for bacterial genome reconstruction demonstrated that ONT R10.4.1 data with appropriate analysis tools achieved methylation motif detection comparable to PacBio. This result depended on using models matched to the specific flow cell and chemistry, reinforcing that model selection is not a minor parameter but the primary determinant of call reliability. ### Creating a Run Manifest Template Implement a simple tabular manifest for each sequencing run. Include columns for run identifier, flow cell version, chemistry version, basecalling mode, model name and version, organism, expected methylation context, and the analyst responsible. This manifest serves three purposes: it forces documentation at the time of analysis, it enables rapid troubleshooting when results look unusual, and it provides the audit trail needed for reproducible research. Store the manifest alongside the raw data in a plain text or CSV format. The [nf-core documentation](https://nf-co.re/docs) emphasizes version tracking and parameter documentation as core reproducibility standards, and a run manifest operationalizes these principles at the laboratory level. When you revisit a dataset months later, the manifest eliminates guesswork about which model was used and why. ### Implementing a Tiered Validation System instead of applying a single validation approach to all samples, implement a tiered system based on the stakes of the analysis. Tier 1 applies to exploratory or screening experiments where approximate methylation patterns suffice. Tier 2 applies to confirmatory experiments where quantitative accuracy matters. Tier 3 applies to clinical or high-impact research where results may influence patient care or major scientific conclusions. For Tier 1 validation, check that known methylated control sites show modification calls and known unmethylated sites do not. For Tier 2 validation, add quantitative checks: compare methylation frequencies at control sites against expected values and verify that the distribution of per-read modification probabilities clusters near 0 and 1. For Tier 3 validation, include orthogonal confirmation with bisulfite sequencing or methylation-specific PCR, and consider consulting a bioinformatics specialist before finalizing conclusions. A 2024 study on target adaptive sampling long-read sequencing demonstrated that ONT workflows can identify allele-specific MLH1 promoter hypermethylation in a Lynch syndrome patient. This finding carried clinical implications and would warrant Tier 3 validation in a diagnostic context. The tiered system ensures that validation effort scales with the consequences of error. ### Recording Model Performance Metrics Across Runs Maintain a cumulative record of model performance metrics across all runs processed in your laboratory. Track the fraction of reads with modification calls, the distribution of modification probabilities, and the methylation frequency at standard control sites. This historical record serves two functions: it establishes baseline expectations for normal performance, and it enables early detection of drift or degradation in model performance. When a new model version is released, process a standard control dataset before applying it to experimental samples. Compare the new model's performance against the historical baseline. If the new model produces substantially different methylation frequencies at control sites, investigate whether the difference reflects improved accuracy or a systematic shift that requires recalibration. The [Galaxy Training Network](https://training.galaxyproject.org/) provides tutorials on quality assessment workflows that can be adapted for tracking these metrics systematically. Similarly, [Bioconductor](https://bioconductor.org/) packages offer statistical frameworks for monitoring batch effects and performance drift over time. ### Handling Mixed-Format Data Archives Laboratories often accumulate data in both FAST5 and POD5 formats as they transition between sequencing platforms. Establish a clear policy for format handling before analysis begins. If you use Megalodon, convert all POD5 files to FAST5 using the pod5-file-format tools and document the conversion parameters. If you use Dorado, you can process both formats natively, but verify that the tool version supports the specific POD5 version produced by your sequencer. For multi-project laboratories, consider maintaining separate analysis directories for each project with their own model configurations and manifests. This prevents cross-project contamination of parameters and simplifies troubleshooting when issues arise. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide guidance on organizing bioinformatics projects and maintaining clear documentation practices. ### Establishing Escalation Triggers for Unusual Results Define specific conditions that trigger escalation to specialized support or orthogonal validation. These triggers include: methylation frequencies at control sites deviating by more than 20 percent from expected values, complete absence of modification calls across an entire run, modification probability distributions that are uniformly intermediate instead of bimodal, and unexpected methylation patterns at loci with well-established epigenetic status. When any trigger condition is met, stop the analysis pipeline and document the observation. Check the run manifest for model mismatches or documentation errors. If the manifest is correct, escalate to ONT support for persistent tool errors or to a bioinformatics specialist for unexpected biological patterns. A 2025 study on direct RNA sequencing demonstrated that single-molecule modification detection can reveal heterogeneity invisible to bulk methods, so some unexpected patterns may reflect genuine biology instead of technical error. The escalation framework helps distinguish these cases systematically. ### Integrating the Framework with Existing Pipelines The decision framework described here complements instead of replaces existing analysis pipelines. Integrate the run manifest as a required input to your analysis scripts, so that no methylation calling job executes without documented model selection parameters. Add the tiered validation checks as post-processing steps that generate quality reports for each sample. The [nf-core documentation](https://nf-co.re/docs) describes community standards for pipeline development that support this integration approach. By embedding the decision framework into your pipeline infrastructure, you ensure that documentation and validation occur automatically instead of relying on individual analyst diligence. ### Common Framework Implementation Failures The most frequent implementation failure is treating the run manifest as optional paperwork instead of an integral analysis component. Laboratories that skip manifest documentation inevitably encounter model mismatches when revisiting older datasets. A second common failure is applying a single validation tier to all samples, either over-validating exploratory work or under-validating high-stakes analyses. A third failure is neglecting to update the historical performance record when new models are adopted, losing the baseline needed for drift detection. Address these failures by assigning manifest completion as a prerequisite for analysis initiation, by explicitly classifying each project's validation tier at the experimental design stage, and by scheduling quarterly reviews of cumulative performance metrics. The [Carpentries lessons](https://carpentries.org/lessons) provide foundational training on data management practices that support consistent implementation of these documentation habits. ## Frequently Asked Questions ### What is the difference between Megalodon and Dorado for methylation calling? Megalodon is the older ONT tool that performs basecalling and methylation detection from FAST5 raw signal data. Dorado is the current recommended tool that supports both POD5 and FAST5 formats, offers faster processing with GPU acceleration, and receives regular model updates for current sequencing chemistry. For new projects, Dorado is the preferred choice, while Megalodon remains useful for existing pipelines that require its specific output formats. ### How do I choose the correct methylation model for my data? The model must match your flow cell version and chemistry. Check the run metadata to identify the flow cell type, then select a model with the corresponding designation in its name. For R10.4.1 flow cells, use models labeled for that chemistry. For R9.4.1 flow cells, use the appropriate legacy models. The model also specifies the methylation context, such as 5mC in CpG context for mammalian DNA. ### What coverage depth is needed for reliable methylation frequency estimates? A minimum of 10 reads per CpG site is commonly used for reliable frequency estimates, though deeper coverage improves confidence. Sites with intermediate methylation levels require more coverage to distinguish true intermediate methylation from sampling noise. For genome-wide analysis, aim for at least 20 to 30 reads per site when possible, and flag or exclude sites with lower coverage. ### Can I detect 5mC in non-CpG contexts with nanopore sequencing? Standard 5mC models detect methylation in CpG context for mammalian genomes. Non-CpG methylation requires models trained for all-context detection, which are available for certain applications. Verify that your chosen model detects the methylation context relevant to your biological question. Plant genomes and certain microbial systems have biologically significant non-CpG methylation that requires appropriate model selection. ### How do I convert per-read methylation calls to site-level frequencies? For each CpG site, count the number of reads with a modification call and divide by the total number of reads covering that site. This calculation requires parsing the MM and ML tags from the BAM file. Tools like ModKit provide a pileup command for this purpose, or you can write custom scripts using the pysam library. ### What quality controls should I include in my methylation analysis? Include control samples with known methylation status, such as fully methylated and fully unmethylated DNA standards. For human samples, imprinted genes with known allele-specific methylation serve as internal controls. Check the distribution of modification probabilities and coverage across the genome. Compare methylation frequencies at known control sites with expected values. ### How does nanopore methylation calling compare to bisulfite sequencing? Nanopore methylation calling works on native DNA without bisulfite conversion, preserving the original modification state and avoiding the DNA damage and bias introduced by bisulfite treatment. Nanopore provides single-molecule resolution and long reads that span repetitive regions. However, bisulfite sequencing is more established and may provide more quantitative accuracy for certain applications. The choice depends on your specific experimental requirements. ### What should I do if my methylation calls do not match expected patterns? First, verify that you used the correct model for your flow cell and chemistry. Check the coverage at the sites of interest and ensure sufficient depth for reliable calls. Validate your results with orthogonal methods such as bisulfite sequencing or methylation-specific PCR. If the discrepancy persists, consult ONT support or a bioinformatics specialist for further investigation. ## Related Bioinformatics Guides - [Genomic Data Analysis Tools: A Comparative Guide for Researchers](/knowledge/bioinformatics/genomic-data-analysis-tools-a-comparative-guide-for-researchers) - [Oxford Nanopore Sequencing: From Sample to Base Calls](/knowledge/bioinformatics/oxford-nanopore-sequencing-from-sample-to-base-calls) - [Genomic Data Processing: From Raw Sequencing to Analysis-Ready Files](/knowledge/bioinformatics/genomic-data-processing-from-raw-sequencing-to-analysis-ready-files) - [RNA Sequencing Data Analysis: From Raw Reads to Differential Expression](/knowledge/bioinformatics/rna-sequencing-data-analysis-from-raw-reads-to-differential-expression) - [Single-Cell Sequencing Analysis Pipeline: From Raw Data to Biological Insights](/knowledge/bioinformatics/single-cell-sequencing-analysis-pipeline-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. - [Matching excellence: Oxford Nanopore Technologies' rise to parity with Pacific Biosciences in genome reconstruction of non-model bacterium with high G+C content.](https://doi.org/10.1099/mgen.0.001316). 2024. - [Assessing the efficacy of target adaptive sampling long-read sequencing through hereditary cancer patient genomes.](https://doi.org/10.1038/s41525-024-00394-z). 2024. - [Single-molecule direct RNA sequencing reveals the shaping of epitranscriptome across multiple species.](https://doi.org/10.1038/s41467-025-60447-4). 2025. - [Metagenomics of the MAST-3 stramenopile, <,i>,Incisomonas<,/i>,, and its associated microbiome reveals unexpected metabolic attributes and extensive nutrient dependencies.](https://doi.org/10.1099/mgen.0.001510). 2025. - [Nanopore Methylation Calling from Limited Training Data](https://www.semanticscholar.org/paper/6ee4926786ee20aecca0dd5b2920d1dc93553abe). 2022. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.