# A Benchmark of Modified Base Callers for Oxford Nanopore: Megalodon, Dorado, Guppy, and Remora Compared

Oxford Nanopore sequencing platforms produce raw electrical current signals that require computational interpretation to translate into nucleotide sequences. Modified base calling adds a second layer of inference, attempting to identify chemical modifications such as 5-methylcytosine (5mC), 5-hydroxymethylcytosine (5hmC), and N6-methyladenosine (m6A) directly from the same signal data. This article compares four base calling tools that support modified base detection: Megalodon, Dorado, Guppy, and Remora. The comparison focuses on practical outcomes for researchers who need to select a base caller for their specific experimental goals, balancing accuracy, runtime, and ease of use. The benchmark draws on public datasets and published evaluations to provide systematic guidance instead of anecdotal recommendations.

The reader of this article is assumed to be a biology student, researcher, laboratory professional, or life-science practitioner who has generated or plans to generate Oxford Nanopore data and needs to make an informed choice among modified base callers. The primary decision problem is straightforward: multiple tools exist, each with different strengths, and the choice materially affects downstream analysis quality and computational cost. This article provides a structured comparison to support that decision.

## At a Glance

The table below summarizes the four base callers compared in this article. The information reflects published benchmarks and official documentation available at the time of writing. Specific version numbers and performance metrics change rapidly in this field, so readers should verify current details before committing to a workflow.

| Tool | Primary Use Case | Modified Base Support | Computational Profile | Ease of Use | Best Suited For |
|------|------------------|----------------------|----------------------|-------------|-----------------|
| Guppy | Production base calling with integrated modification detection | 5mC, 5hmC, m6A (model dependent) | Moderate to high GPU usage, CPU mode available | High, command line with clear options | Routine sequencing runs where speed matters and standard models suffice |
| Dorado | Current generation base caller with simplified workflow | 5mC, 5hmC, m6A (model dependent) | High GPU usage, optimized for newer flow cells | High, single command with automatic model selection | Newer R10 flow cell data and users who want a streamlined pipeline |
| Megalodon | Modified base calling with per-read and per-site output | 5mC, 5hmC, m6A (model dependent) | Very high computational demand, CPU and GPU modes | Moderate, requires more configuration | Research projects needing detailed modification probabilities and aggregate reports |
| Remora | Modified base calling module that works with Guppy or Dorado | 5mC, 5hmC, m6A (model dependent) | Moderate, runs as a post-processing step | Moderate, requires separate installation and model training | Custom modification detection and users who need to train models on their own data |

The choice among these tools depends on three factors: the sequencing platform generation (R9 or R10 flow cells), the required output granularity (per-read probabilities versus aggregate methylation calls), and the available computational resources. Dorado represents the current recommended path for most users, while Megalodon remains relevant for specific research questions that require its detailed output format. Guppy continues to appear in published workflows, particularly in older datasets and in tools that were built around its output format. Remora serves as a modular component that can be integrated into custom pipelines.

## Oxford Nanopore Sequencing and Signal Interpretation

Oxford Nanopore sequencing operates by passing a single strand of DNA or RNA through a protein nanopore embedded in a membrane. An electrical potential across the membrane drives the strand through the pore, and the resulting current fluctuations are measured thousands of times per second. The pattern of current changes corresponds to the sequence of nucleotides passing through the pore, but the relationship is not a simple one-to-one mapping. Each k-mer of nucleotides (typically 5 to 6 bases) produces a characteristic current level, and the raw signal must be decoded through computational methods.

The core analytical workflow for Oxford Nanopore data follows a consistent sequence: base calling, alignment, re-squiggling, and quality control. Base calling converts raw electrical signals into nucleotide sequences. Alignment maps those sequences to a reference genome. Re-squiggling aligns the raw signal back to the called sequence to enable detailed analysis. Quality control assesses data quality at each stage. This workflow is described in detail in a recent review of Nanopore direct RNA sequencing for RNA modification analysis, which emphasizes that each stage has dedicated computational tools and that choices made early in the pipeline affect all subsequent analyses.

Modified base detection adds a specialized step to this workflow. When a nucleotide carries a chemical modification, the electrical current profile through the nanopore shifts in a characteristic way. Base callers that support modified base detection use machine learning models trained on known modification patterns to identify these shifts. The accuracy of modification detection depends on the quality of the training data, the model architecture, and the sequencing conditions. The same review notes that benchmark analyses of m6A and pseudouridine detection across publicly available datasets demonstrate substantial variability across different tools, underscoring the inherent difficulties in reliably detecting modifications from Oxford Nanopore signals.

The practical implication for researchers is that modified base calling is not a solved problem. Different tools produce different results on the same data, and the choice of base caller can change the biological conclusions drawn from an experiment. This variability is a central motivation for the benchmark comparison presented in this article.

## Base Calling Fundamentals and Modified Base Detection

Base calling for Oxford Nanopore data has evolved through several generations of software. Early tools used hidden Markov models to decode raw signals. Modern tools use neural networks, specifically recurrent neural networks and transformer architectures, to achieve higher accuracy. The transition from Guppy to Dorado represents this evolution, with Dorado incorporating newer model architectures and optimizations for current flow cell chemistry.

The relationship between base calling and modified base detection is important to understand. Some tools perform both tasks simultaneously, producing a canonical sequence and modification calls in a single pass. Other tools separate the tasks, first producing a canonical sequence and then analyzing the raw signal or the aligned reads for modification evidence. This architectural difference affects both accuracy and computational efficiency.

Guppy was the standard base caller for Oxford Nanopore data for several years. It supports modified base detection through specialized models that can be selected at runtime. Guppy models are available for 5mC detection in CpG contexts, 5hmC detection, and m6A detection in RNA data. The tool operates in both CPU and GPU modes, with GPU acceleration providing substantial speed improvements. Many published workflows and downstream tools were built around Guppy output, and the tool remains relevant for reproducing or extending existing analyses.

Dorado is the current generation base caller from Oxford Nanopore Technologies. It was designed to replace Guppy and offers several advantages: simplified command line interface, automatic model selection based on flow cell and chemistry, and improved accuracy on newer R10 flow cell data. Dorado also supports modified base detection with models for 5mC, 5hmC, and m6A. The tool is actively maintained and receives regular updates. For new projects, Dorado is the recommended starting point.

Megalodon is a base caller that was developed for research applications requiring detailed modification information. It produces per-read modification probabilities and can generate aggregate modification reports at the level of individual sites. Megalodon uses a different architecture than Guppy and Dorado, and it offers more configuration options. The computational cost is substantially higher, making it less suitable for routine production runs. However, for research questions that require per-read modification states or detailed probability distributions, Megalodon provides capabilities that other tools lack.

Remora is a modular tool that performs modified base calling as a post-processing step. It can work with base called sequences from Guppy or Dorado, adding modification information to existing outputs. Remora also supports training custom models on user-provided data, which is valuable for detecting modifications that are not covered by standard models. The modular design allows researchers to integrate Remora into existing pipelines without changing their primary base calling workflow.

The choice among these tools should be guided by the specific requirements of the experiment. A clinical diagnostic workflow that prioritizes speed and standardization might choose Dorado with standard models. A research project investigating a novel modification might choose Remora with custom model training. A study requiring detailed per-read modification states might choose Megalodon. The benchmark data presented in the following sections provides quantitative evidence to support these decisions.

## Benchmark Methodology and Public Datasets

Benchmarking modified base callers requires carefully designed datasets and evaluation criteria. The ideal benchmark uses data from samples with known modification states, allowing direct measurement of accuracy. Publicly available datasets from the National Center for Biotechnology Information (NCBI) provide a foundation for such benchmarks. NCBI maintains sequence databases, search systems, and analysis services that researchers can use to access and process sequencing data for benchmarking purposes.

The benchmark approach used in published evaluations typically involves three components. First, a set of reference samples with known modification states, either from in vitro transcribed RNA with controlled modifications or from cell lines with well-characterized modification profiles. Second, a standardized evaluation pipeline that applies each base caller to the same raw data and measures accuracy against the known truth. Third, a set of performance metrics that capture both accuracy and computational efficiency.

A recent benchmark of sequencing technologies for genetic variant detection provides a useful framework for understanding the strengths and limitations of different platforms. The study compared short-read and long-read sequencing across diverse genomic contexts and sequencing depths. Long-read platforms demonstrated clear advantages in detecting structural variants and resolving small variants in difficult genomic regions. Among long-read pipelines, Oxford Nanopore R10 data with DeepVariant performed particularly well in clinically relevant loci. Coverage analyses indicated that long-read sequencing reached accuracy saturation between 20x and 45x coverage, while short-read sequencing required more than 60x coverage to achieve comparable results.

These findings have direct implications for modified base calling benchmarks. The coverage level affects the statistical power for detecting modifications, and the sequencing platform generation affects the base accuracy that underlies modification detection. A benchmark that uses low coverage data may underestimate the accuracy of all tools, while a benchmark that uses only one platform generation may not generalize to other configurations.

The review of Nanopore direct RNA sequencing for RNA modification analysis provides additional benchmark context. The authors carried out benchmark analyses of m6A and pseudouridine detection across two publicly available datasets. Their results demonstrated substantial variability across different tools, highlighting the difficulty of reliably detecting modifications from Oxford Nanopore signals. This variability is an important consideration for researchers who need to compare results across studies that used different base callers.

For practical benchmarking, researchers should consider the following steps. First, identify publicly available datasets that match the experimental context of interest, using resources such as NCBI to locate suitable data. Second, establish a reference truth for modification states, either through orthogonal methods or through known biological controls. Third, apply each base caller to the same raw data using recommended settings. Fourth, evaluate accuracy using appropriate metrics, including sensitivity, precision, and area under the receiver operating characteristic curve. Fifth, measure runtime and computational resource usage to inform practical workflow decisions.

## Accuracy Comparison Across Base Callers

Accuracy is the primary criterion for selecting a modified base caller. The published literature provides several sources of accuracy data, though direct head-to-head comparisons of all four tools on identical datasets are limited. The available evidence allows for informed conclusions about relative performance.

The Nanopore-m6A-finder study provides accuracy data for m6A detection using Guppy basecalled direct RNA sequencing data. The tool, NP-mFinder, uses an XGBoost model for mRNA exonic regions and a hard-voting ensemble of XGBoost and random forest models for poly(A) regions. After training with direct RNA sequencing data from in vitro transcribed RNA, the tool achieved high performance on held-out test datasets, with area under the curve around 0.90 and accuracy, precision, recall, and F1-score above 0.80. This result demonstrates that Guppy basecalled data can support accurate m6A detection when paired with appropriate downstream analysis tools.

The same study reported that NP-mFinder achieved 93 percent precision in detecting m6A within the "AAAAA" sequence context in the mRNA exonic region of HEK293T direct RNA sequencing data when compared to high-confidence m6A site annotations from GLORI v2.0. This finding is notable because homopolymer regions are challenging for base callers, and the high precision in this context suggests that Guppy basecalled data retains sufficient signal information for accurate modification detection even in difficult sequence contexts.

The SegPore study provides comparative data on raw signal segmentation, which is a critical step for modification detection. The authors developed SegPore, a white-box model that improves raw signal segmentation by leveraging a molecular jiggling translocation hypothesis. They demonstrated that SegPore outperforms state-of-the-art methods such as Nanopolish and Tombo in raw signal segmentation across three large benchmark datasets. The improved segmentation enabled SegPore combined with m6Anet to deliver state-of-the-art performance in site-level m6A identification. This finding underscores the importance of the signal processing steps that precede or accompany base calling.

The review of Nanopore direct RNA sequencing for RNA modification analysis reported benchmark analyses of m6A and pseudouridine detection across two publicly available datasets. The results demonstrated substantial variability across different tools, with no single tool consistently outperforming all others across all metrics and datasets. This variability reflects differences in model architecture, training data, and the specific modifications being detected.

For 5mC detection in DNA, the accuracy comparison is somewhat different. Dorado and Guppy both offer 5mC models, and published benchmarks suggest that Dorado achieves higher accuracy on R10 flow cell data. The improved base accuracy of R10 chemistry contributes to better modification detection, as the underlying sequence context is more reliably determined. Megalodon also supports 5mC detection but requires more computational resources and produces output in a format that requires additional processing.

The practical recommendation based on available evidence is to use Dorado for new projects involving standard modifications (5mC, 5hmC, m6A) on current flow cell chemistry. For projects that must use Guppy for compatibility with existing pipelines or published protocols, the accuracy is sufficient for many applications, particularly when paired with robust downstream analysis tools. Megalodon should be reserved for research questions that require its detailed output format. Remora is the appropriate choice when custom model training is needed.

## Runtime and Computational Resource Requirements

Computational efficiency is a practical consideration that often determines whether a base calling strategy is feasible for a given project. The four tools compared in this article have substantially different computational profiles, and these differences affect both cost and turnaround time.

Guppy offers both CPU and GPU modes. In CPU mode, Guppy can process data on standard laboratory workstations, though throughput is limited. A typical CPU-based Guppy run might process data at a rate that is adequate for small projects but becomes a bottleneck for large-scale sequencing. GPU mode provides substantial acceleration, with throughput improvements of 10 to 50 times depending on the GPU model and the specific Guppy version. The GPU requirement adds hardware cost but is often justified for production sequencing facilities.

Dorado is optimized for GPU acceleration and is designed to take advantage of newer NVIDIA GPUs. The tool achieves higher throughput than Guppy on equivalent hardware, particularly for R10 flow cell data. Dorado also supports CPU mode, but the performance difference between CPU and GPU modes is more pronounced than with Guppy. For laboratories that do not have access to suitable GPUs, Dorado in CPU mode may be impractical for large datasets.

Megalodon has the highest computational demand of the four tools. The detailed per-read modification analysis requires substantial processing time, and the tool is not designed for high-throughput production use. Megalodon can run in both CPU and GPU modes, but even with GPU acceleration, the runtime is substantially longer than Guppy or Dorado. The computational cost is justified when the detailed output is required for the research question.

Remora has moderate computational requirements. As a post-processing step, it adds processing time to an existing base calling workflow but does not require re-processing the raw signal. The additional time depends on the number of reads and the complexity of the model. Remora can run on CPU for small datasets, but GPU acceleration is recommended for larger projects.

The choice of base caller affects also the base calling step but also downstream analysis. Some downstream tools expect specific input formats or specific modification call formats. For example, tools that were developed to work with Guppy output may require format conversion when used with Dorado or Megalodon output. These compatibility considerations can add hidden computational costs to a workflow.

For practical planning, researchers should estimate the computational requirements of their project before selecting a base caller. The number of reads, the read length distribution, the flow cell chemistry, and the available hardware all affect the runtime. A benchmark on a small subset of the data can provide a reliable estimate of the full run time. The Galaxy Training Network provides accessible workflow training and analysis tutorials that can help researchers develop and test their pipelines before committing to full-scale runs.

## Ease of Use and Workflow Integration

The practical usability of a base caller depends on installation complexity, command line interface design, documentation quality, and compatibility with existing workflows. These factors affect the time required to implement a base calling strategy and the likelihood of user errors.

Guppy has been widely used for several years, and extensive documentation and community support are available. The command line interface is straightforward, with options for selecting models, specifying output formats, and controlling computational resources. Guppy integrates with many existing pipelines and downstream tools, and published protocols often include specific Guppy commands. The main usability challenge is selecting the correct model for the flow cell and chemistry, as incorrect model selection can produce poor results.

Dorado was designed with usability improvements over Guppy. The command line interface is simpler, and the tool automatically selects appropriate models based on the flow cell and chemistry information embedded in the sequencing data. This automatic selection reduces the risk of user error. Dorado also provides clearer error messages and more consistent output formats. The documentation is current and includes examples for common use cases. For new users, Dorado is the easiest tool to get started with.

Megalodon has a steeper learning curve than Guppy or Dorado. The tool requires more configuration, and the output formats are more complex. Megalodon documentation is available but assumes familiarity with Oxford Nanopore data analysis concepts. The tool is best suited for researchers who need its specific capabilities and are willing to invest time in learning its interface.

Remora requires separate installation and configuration. The tool is designed to work with Guppy or Dorado output, so users must first set up their primary base calling workflow. Remora also supports custom model training, which adds complexity but provides flexibility. The documentation includes examples for common use cases, and the tool is actively maintained.

Workflow integration is an important consideration for laboratories that use standardized pipelines. The nf-core documentation describes community pipeline standards that emphasize reproducibility and best practices. Pipelines built on these standards often include specific base calling steps, and the choice of base caller may be constrained by the pipeline design. Researchers should check whether their preferred pipeline supports the base caller they want to use.

The Carpentries lessons provide foundational training in computing, data analysis, and programming that can help researchers develop the skills needed to implement and troubleshoot base calling workflows. These lessons cover shell scripting, version control with Git, and programming fundamentals that are directly applicable to bioinformatics pipeline development.

## Practical Implementation Steps

Implementing a modified base calling workflow requires careful planning and systematic execution. The following steps provide a practical framework for selecting and using a base caller for Oxford Nanopore modified base detection.

First, define the experimental requirements. Identify the specific modification or modifications of interest, the expected modification frequency, and the required accuracy. Consider whether per-read modification states are needed or whether aggregate modification levels at specific sites are sufficient. These requirements determine which base caller and which output format are appropriate.

Second, assess the available computational resources. Determine whether a suitable GPU is available and what model it is. Check the available storage space, as base calling produces large intermediate files. Estimate the total computational time required for the dataset and verify that it fits within project timelines.

Third, select the base caller and model. For standard modifications on current flow cell chemistry, Dorado with the appropriate model is the recommended choice. For compatibility with existing pipelines or published protocols, Guppy may be necessary. For custom modification detection, Remora with a trained model is appropriate. For detailed per-read analysis, Megalodon should be considered.

Fourth, test the workflow on a small subset of the data. Run the base caller on a representative sample to verify that the output format is correct, the modification calls are reasonable, and the runtime is as expected. Compare the results with known controls if available. This testing step can identify problems before committing to a full run.

Fifth, execute the full base calling run. Monitor the run for errors and resource usage. Verify that the output files are complete and correctly formatted. Document the base caller version, model, and settings used for the run, as this information is essential for reproducibility.

Sixth, validate the modification calls. Compare the results with orthogonal methods if available, such as antibody-based enrichment followed by short-read sequencing or mass spectrometry. Check that the modification calls are biologically plausible given the sample and experimental context. The review of Nanopore direct RNA sequencing for RNA modification analysis emphasizes that substantial variability across tools underscores the need for validation.

Seventh, document and report the results. Include the base caller version, model, and settings in the methods section of any publication. Report the validation results and any limitations of the modification calls. This documentation supports reproducibility and allows other researchers to interpret the results appropriately.

## Records and Measurements for Quality Control

Maintaining detailed records of base calling runs is essential for quality control and reproducibility. The following measurements should be recorded for each base calling run.

The base caller version and build number should be recorded, as different versions can produce different results. The model name and version should also be recorded, including the specific modification model used. The flow cell type and chemistry version should be noted, as these affect the raw signal characteristics and the appropriate model selection.

The runtime and computational resource usage should be recorded, including the total processing time, the peak memory usage, and the GPU utilization. These measurements help with capacity planning for future runs and can identify performance issues.

The output statistics should be recorded, including the number of reads processed, the number of bases called, the read length distribution, and the base calling quality scores. These statistics provide a baseline for comparing runs and detecting anomalies.

For modified base calling, the modification call statistics should be recorded, including the number of modification calls, the modification frequency at specific sites, and the quality scores associated with the calls. These measurements provide a basis for comparing results across runs and for validating the modification calls.

The National Center for Biotechnology Information provides resources for storing and sharing sequencing data and analysis results. Researchers should consider depositing their raw data and base calling outputs in appropriate repositories to support reproducibility and data sharing.

The EMBL-EBI Training program offers learning pathways for bioinformatics data resources and practical analysis education. These resources can help researchers develop the skills needed to implement robust quality control procedures for their base calling workflows.

## Common Failure Patterns and Troubleshooting

Several failure patterns recur in modified base calling workflows. Recognizing these patterns and understanding their causes can help researchers troubleshoot problems efficiently.

The first common failure pattern is incorrect model selection. Using a model that does not match the flow cell chemistry or the modification of interest produces poor results. Symptoms include low base calling accuracy, unexpected modification frequencies, or errors during the base calling run. The solution is to verify the model selection against the flow cell and chemistry information and to consult the base caller documentation for the correct model.

The second failure pattern is insufficient data quality. Poor quality raw signal data, whether from degraded samples, suboptimal library preparation, or flow cell issues, produces unreliable base calling and modification detection. Symptoms include low read yields, short read lengths, and poor quality scores. The solution is to assess the raw data quality before base calling and to address any sample or sequencing issues.

The third failure pattern is computational resource exhaustion. Base calling can require substantial memory and storage, and runs may fail when resources are insufficient. Symptoms include out-of-memory errors, disk full errors, and crashes during the run. The solution is to estimate resource requirements before the run and to allocate sufficient resources.

The fourth failure pattern is format incompatibility. Downstream tools may expect specific input formats that differ from the base caller output. Symptoms include errors when loading base calling output into downstream tools or unexpected results from downstream analysis. The solution is to verify format compatibility and to use format conversion tools when necessary.

The fifth failure pattern is modification call inconsistency. Different base callers or different runs of the same base caller may produce different modification calls on the same data. This variability is expected given the inherent difficulty of modification detection from Oxford Nanopore signals. The solution is to validate modification calls with orthogonal methods and to report the variability in publications.

The sixth failure pattern is model overfitting or underfitting in custom training. When training custom models with Remora, the training data may not be representative of the experimental data, leading to poor performance. Symptoms include high accuracy on training data but poor accuracy on experimental data. The solution is to use diverse training data and to validate the trained model on held-out data.

## Limitations of Modified Base Calling

Modified base calling from Oxford Nanopore data has inherent limitations that researchers should understand before designing experiments or interpreting results. These limitations affect all four base callers compared in this article.

The first limitation is sequencing error rates. Oxford Nanopore sequencing has higher error rates than short-read sequencing, particularly in homopolymer regions and repetitive sequences. These errors affect the accuracy of modification detection, as the sequence context is not always reliably determined. The review of Nanopore direct RNA sequencing for RNA modification analysis identifies sequencing error rates as a persistent limitation.

The second limitation is the complexity of multi-modification inference. A single position may carry multiple modifications, and the current signal may reflect the combined effect. Disentangling these effects is challenging, and most base callers are trained to detect one modification at a time. The review identifies the complexity of multi-modification inference as a persistent limitation.

The third limitation is the dependence on training data. Base callers are trained on specific modification types in specific sequence contexts. Modifications that are not represented in the training data may not be detected accurately. This limitation is particularly relevant for less common modifications or for modifications in unusual sequence contexts.

The fourth limitation is the variability across tools and datasets. Benchmark analyses demonstrate substantial variability in modification detection across different tools and datasets. This variability means that results from different studies may not be directly comparable, and that modification calls should be interpreted with caution.

The fifth limitation is the computational demand. Modified base calling requires substantial computational resources, particularly for detailed per-read analysis. This demand can be a barrier for laboratories with limited computational infrastructure.

The sixth limitation is the lack of standardized evaluation criteria. Different studies use different metrics and different reference standards, making it difficult to compare results across studies. The development of standardized benchmarks would improve the field.

The SegPore study provides an example of how methodological improvements can address some of these limitations. By improving raw signal segmentation, SegPore reduces structured noise and improves modification detection accuracy. This finding suggests that continued methodological development can mitigate some of the inherent limitations of modified base calling.

## Safety and Regulatory Context

Modified base calling is a computational analysis step and does not involve direct safety hazards. However, the results of modified base calling can have regulatory and clinical implications, particularly when used in diagnostic or precision medicine contexts.

The review of Oxford Nanopore sequencing in pediatric emergency infectious diseases describes the use of the platform for rapid diagnosis and precision medicine. The authors note that the platform enables comprehensive microbial characterization directly at the point of care, including pathogen identification, antimicrobial resistance profiling, and strain typing. They also address current challenges, including bioinformatics complexity and analytical standardization.

For clinical applications, the accuracy and reproducibility of modified base calling are critical. Regulatory frameworks for diagnostic tests require validation of analytical performance, including accuracy, precision, and reproducibility. Researchers and clinicians who use modified base calling in clinical contexts should ensure that their workflows meet applicable regulatory requirements and that the limitations of the technology are clearly understood.

The same review proposes pathways toward integration within a precision infectious disease framework. These pathways include standardization of analytical workflows, validation of diagnostic performance, and integration with clinical decision support systems. Modified base calling is one component of these workflows, and its performance must be validated in the context of the full diagnostic pathway.

For research applications, the regulatory context is less stringent, but ethical considerations still apply. Researchers should ensure that their use of modified base calling is appropriate for their research questions and that the limitations of the technology are accurately represented in publications and presentations.

## Professional Escalation Criteria

Researchers should escalate issues to appropriate professionals when they encounter problems that exceed their expertise or when the results have implications that require specialized interpretation.

If base calling runs consistently fail or produce anomalous results despite troubleshooting, escalate to the base caller support team or to experienced bioinformatics colleagues. Provide detailed records of the run, including the base caller version, model, settings, and error messages.

If modification calls are inconsistent with biological expectations or with orthogonal validation results, escalate to a specialist in the relevant biological modification or to a bioinformatics consultant with expertise in Oxford Nanopore data analysis. The variability across tools and datasets documented in the literature means that unexpected results should be investigated instead of accepted at face value.

If modified base calling results are intended for clinical or diagnostic use, escalate to the appropriate regulatory and clinical professionals. The analytical validation requirements for clinical applications are more stringent than for research applications, and specialized expertise is needed to ensure compliance.

If custom model training with Remora produces poor results, escalate to the Remora development team or to researchers with experience in training modification detection models. The training data requirements and model evaluation procedures are specialized, and expert guidance can help avoid common pitfalls.

If the computational requirements of a base calling workflow exceed available resources, escalate to institutional computing support or to cloud computing providers with experience in bioinformatics workloads. The runtime and resource measurements recorded during the workflow can inform capacity planning and resource allocation decisions.

## Frequently Asked Questions

### What is the difference between Guppy and Dorado for modified base calling?

Guppy is the older base caller that has been widely used for Oxford Nanopore data. It supports modified base detection through specialized models and offers both CPU and GPU modes. Dorado is the current generation base caller that was designed to replace Guppy. Dorado offers a simplified command line interface, automatic model selection based on flow cell and chemistry, and improved accuracy on newer R10 flow cell data. For new projects, Dorado is the recommended choice, while Guppy remains relevant for reproducing or extending existing analyses that used Guppy output.

### When should I use Megalodon instead of Dorado or Guppy?

Megalodon is appropriate when you need detailed per-read modification probabilities or aggregate modification reports at the level of individual sites. The tool produces output formats that support these analyses, but the computational cost is substantially higher than Dorado or Guppy. If your research question requires per-read modification states or detailed probability distributions, Megalodon provides capabilities that other tools lack. For routine production runs where speed matters and standard models suffice, Dorado or Guppy are more practical choices.

### How does Remora differ from the other base callers?

Remora is a modular tool that performs modified base calling as a post-processing step. It works with base called sequences from Guppy or Dorado, adding modification information to existing outputs. Remora also supports training custom models on user-provided data, which is valuable for detecting modifications that are not covered by standard models. The modular design allows researchers to integrate Remora into existing pipelines without changing their primary base calling workflow.

### What accuracy can I expect from modified base calling?

Accuracy depends on the modification type, the sequence context, the flow cell chemistry, and the base caller and model used. Published benchmarks demonstrate substantial variability across tools and datasets. For m6A detection in direct RNA sequencing data, one study reported area under the curve around 0.90 and accuracy, precision, recall, and F1-score above 0.80 for a downstream analysis tool using Guppy basecalled data. The same study reported 93 percent precision in detecting m6A within the "AAAAA" sequence context in a specific cell line. These results are context specific and should not be generalized without validation.

### How much computational resources do I need for modified base calling?

The computational requirements vary substantially across tools. Guppy offers both CPU and GPU modes, with GPU acceleration providing substantial speed improvements. Dorado is optimized for GPU acceleration and achieves higher throughput than Guppy on equivalent hardware. Megalodon has the highest computational demand of the four tools. Remora has moderate requirements as a post-processing step. The number of reads, read length distribution, flow cell chemistry, and available hardware all affect runtime. Testing on a small subset of data can provide a reliable estimate of full run time.

### Can I train custom models for detecting novel modifications?

Yes, Remora supports training custom models on user-provided data. This capability is valuable for detecting modifications that are not covered by standard models. The training process requires carefully prepared training data with known modification states. The quality of the training data directly affects the accuracy of the trained model. Validation on held-out data is essential to assess model performance. The review of Nanopore direct RNA sequencing for RNA modification analysis discusses the range of machine learning and deep learning frameworks used for modification detection.

### How should I validate modified base calling results?

Validation should use orthogonal methods when available, such as antibody-based enrichment followed by short-read sequencing or mass spectrometry. Comparison with known biological controls is also valuable. The substantial variability across tools and datasets documented in the literature means that modification calls should be interpreted with caution. Reporting the base caller version, model, and settings in publications supports reproducibility and allows other researchers to interpret results appropriately.

### What are the main limitations of modified base calling from Oxford Nanopore data?

The main limitations include sequencing error rates, particularly in homopolymer regions and repetitive sequences, the complexity of multi-modification inference, the dependence on training data that may not cover all modifications of interest, variability across tools and datasets, computational demand, and the lack of standardized evaluation criteria. The review of Nanopore direct RNA sequencing for RNA modification analysis identifies these as persistent limitations and proposes future directions aimed at improving accuracy, robustness, and biological interpretability.

## Related Bioinformatics Guides

- [Oxford Nanopore Sequencing: From Sample to Base Calls](/knowledge/bioinformatics/oxford-nanopore-sequencing-from-sample-to-base-calls)
- [How to Choose a Long-Read Sequencing Platform: PacBio vs Oxford Nanopore](/knowledge/bioinformatics/how-to-choose-a-long-read-sequencing-platform-pacbio-vs-oxford-nanopore)
- [Metagenomic Binning Tools Benchmark: How to Evaluate and Choose](/knowledge/bioinformatics/metagenomic-binning-tools-benchmark-how-to-evaluate-and-choose)
- [Single-Cell Sequencing Services: How to Choose a Provider](/knowledge/bioinformatics/single-cell-sequencing-services-how-to-choose-a-provider)
- [RNA-Seq Batch Effect Detection and Correction](/knowledge/bioinformatics/rna-seq-batch-effect-detection-and-correction)

## 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.
- [Nanopore-m(6)A-finder, a novel m(6)A site caller for Nanopore DRS data.](https://pubmed.ncbi.nlm.nih.gov/42095024). Frontiers in genetics, 2026.
- [Nanopore direct RNA sequencing for RNA modification analysis: workflow assessment and computational tool benchmarking.](https://doi.org/10.1007/s44307-025-00093-5). 2026.
- [Benchmarking of sequencing technologies defines optimal strategies for genetic variants detection in a human genome.](https://doi.org/10.1186/s13059-026-04048-4). 2026.
- [Raw signal segmentation for estimating RNA modification from Nanopore direct RNA sequencing data.](https://doi.org/10.7554/elife.104618). 2026.
- [Oxford Nanopore Sequencing in pediatric emergency infectious diseases: from rapid diagnosis to precision medicine.](https://doi.org/10.3389/fcimb.2026.1793808). 2026.

> This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.