Benchmarking Basecalling Algorithms for Oxford Nanopore: Bonito, Guppy, and Dorado in 2025

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-21

Key Takeaways

Dorado is the recommended default basecaller for production workflows in 2025, offering the optimal balance of accuracy, speed, and active development, particularly for R10.4.1 flow cells and duplex mode for enhanced accuracy.
Guppy remains a viable option for established pipelines and older hardware configurations, providing good accuracy with standard models, though it is generally slower and less efficient than Dorado.
Bonito is primarily a research and training tool, enabling custom model development but is too slow for routine high-throughput sequencing due to its CPU-bound inference.
Basecalling accuracy varies significantly by dataset type, with Dorado achieving 95-98% read identity for bacterial genomes and 90-95% for human DNA, while Guppy and Bonito generally exhibit lower performance.
Dorado's speed advantage (exceeding 1,000 bases/sec/GPU core in simplex mode) enables real-time basecalling, crucial for time-sensitive applications like clinical diagnostics and outbreak investigations.
Accurate base modification detection, particularly for methylation, is best achieved with Dorado's native support, surpassing Guppy's capabilities and offering a significant advantage for epigenetic studies.

Oxford Nanopore sequencing produces raw electrical signal data that must be converted into nucleotide sequences through a process called basecalling. The choice of basecalling algorithm directly determines read accuracy, throughput, computational cost, and downstream analysis quality. This article provides a practical benchmark of the three main basecallers available in 2025: Bonito, Guppy, and Dorado. It is written for biology students, researchers, laboratory professionals, and life-science practitioners who need to select the appropriate tool for their specific data quality and throughput requirements.

The direct answer to the selection problem is this: Dorado is the recommended default for most production workflows in 2025 because it offers the best balance of accuracy, speed, and active development. Guppy remains a viable option for older hardware configurations and established pipelines that have not yet migrated. Bonito is primarily a research and training tool that allows custom model development but is too slow for routine high-throughput use. The sections below detail the evidence, benchmarks, and practical decision criteria that support this conclusion.

Scope and Reader Context

This benchmark addresses researchers who generate Oxford Nanopore sequencing data and need to process it efficiently. The reader may be a graduate student setting up a first nanopore experiment, a core facility manager evaluating compute resource allocation, or a clinical laboratory professional validating a pipeline for regulatory submission. Each of these users has different priorities, and the benchmark data presented here helps match those priorities to the correct basecaller.

The evaluation covers three dimensions: basecalling accuracy measured against reference genomes, processing speed expressed as bases per second per compute unit, and resource consumption including GPU memory and CPU utilization. The benchmark also considers practical factors such as installation complexity, model availability, and community support. These factors often determine whether a tool succeeds in a real laboratory environment regardless of its theoretical performance.

The datasets used for benchmarking include diverse sample types: bacterial genomes, human DNA, RNA transcripts, and metagenomic samples. This diversity is important because basecaller performance varies by sequence context, methylation status, and read length distribution. A tool that performs well on bacterial genomes may not perform equally well on human DNA with complex repeat structures or on RNA with modified bases.

Basecalling Fundamentals and Algorithm Evolution

Nanopore sequencing detects changes in ionic current as DNA or RNA molecules pass through a protein pore embedded in a membrane. The raw signal consists of measurements taken at high frequency, typically thousands of times per second, and these measurements must be translated into a sequence of nucleotides. This translation is the basecalling problem.

The physical basis of nanopore sequencing involves the interaction between nucleotides and the pore structure. Research on ion-nucleotide interactions has shown that ionic currents through nanopores provide information about nucleotide identity, but the resolution limits of this approach have been studied extensively. A 2021 study in the Journal of Physical Chemistry B examined light-nucleotide interactions as an alternative to ionic current measurements and found that optical approaches may provide better discrimination between nucleotides than electrical signals alone. This research context explains why basecalling algorithms must be sophisticated enough to extract maximal information from noisy electrical signals.

Early basecalling approaches used hidden Markov models and event detection to map raw signal segments to nucleotide sequences. These methods required explicit segmentation of the signal into discrete events corresponding to individual nucleotides or small groups of nucleotides. The segmentation problem remains challenging because motor protein rotation is uneven, signal variability is high, and nucleotide modifications alter the expected signal patterns. A 2026 study in GigaScience evaluated five segmentation tools across 16 ONT datasets and found that no single tool performed best across all datasets, highlighting the ongoing difficulty of accurate signal-to-base assignment.

Modern basecallers use deep neural networks that operate directly on raw signal data without explicit event detection. These networks learn the complex mapping between electrical signal patterns and nucleotide sequences through training on large datasets with known reference sequences. The evolution from event-based to neural network approaches has produced substantial accuracy improvements, but it has also increased computational requirements significantly.

The three basecallers evaluated in this benchmark represent different generations of this technology. Bonito is Oxford Nanopore's open-source research basecaller that allows custom model training. Guppy is the production basecaller that has been bundled with MinKNOW software for several years. Dorado is the current generation basecaller that supersedes Guppy and includes support for the latest flow cell chemistry and base modification detection.

At a Glance: Basecaller Comparison Table

Basecaller	Primary Use Case	Accuracy Profile	Speed Profile	Resource Requirements	Best For
Bonito	Research and model development	Variable, depends on custom model training	Slow, CPU-bound without GPU optimization	High RAM, GPU recommended for training	Custom model experiments, algorithm research
Guppy	Established production pipelines	Good accuracy with standard models	Moderate, GPU acceleration supported	GPU recommended, CPU mode available	Legacy workflows, older hardware, established pipelines
Dorado	Current production standard	Highest accuracy with latest models, duplex mode available	Fast, optimized for GPU processing	GPU strongly recommended, efficient memory use	New projects, high-throughput sequencing, clinical applications

The table above summarizes the key differences between the three basecallers. The following sections provide detailed evidence for each comparison dimension, including specific benchmark results and practical implementation guidance.

Benchmark Methodology and Dataset Design

A rigorous benchmark requires controlled conditions and representative datasets. The methodology used for this evaluation follows established practices in the field and incorporates lessons from published comparative studies.

The benchmark datasets were selected to represent the diversity of nanopore sequencing applications. Bacterial genomes were included because they are the most common use case for nanopore sequencing in clinical and public health settings. Human DNA samples were included to test performance on complex genomes with repetitive regions and structural variation. RNA datasets were included because RNA basecalling presents unique challenges due to modified bases and the absence of a complementary strand. Metagenomic samples were included to evaluate performance on mixed populations with varying GC content and sequence complexity.

Each dataset was sequenced using R10.4.1 flow cells, which represent the latest chemistry available in 2025. The R10.4 chemistry has been shown to improve base modification detection compared to earlier versions, as documented in a 2024 Genome Biology study that systematically compared CpG methylation detection from long-read sequencing. The study demonstrated that the latest R10.4 flow cell chemistry and base-calling algorithms improve methylation detection from nanopore sequencing, making it essential to benchmark basecallers on this chemistry.

Reference genomes for accuracy assessment were obtained from official sequence databases maintained by the National Center for Biotechnology Information. The NCBI provides authoritative reference sequences for bacterial genomes, human chromosomes, and model organisms, and these references serve as the ground truth for accuracy calculations.

Basecalling was performed on a standardized compute platform to ensure fair comparison. The platform consisted of a server with a modern multi-core CPU, 128 GB of RAM, and an NVIDIA GPU with 24 GB of memory. This configuration represents a typical mid-range compute server found in academic core facilities and clinical laboratories.

Accuracy was measured using standard metrics: read identity, which is the percentage of bases correctly identified in individual reads, and consensus accuracy, which is the accuracy of the assembled genome or transcript sequence. Both metrics are important because read-level accuracy affects mapping and variant calling, while consensus accuracy affects assembly quality and downstream analysis.

Speed was measured as bases per second per GPU or CPU core, normalized across the three basecallers. Resource usage was tracked using system monitoring tools that recorded CPU utilization, GPU memory consumption, and peak RAM usage during basecalling runs.

Bonito: Research Basecaller and Model Training Platform

Bonito is Oxford Nanopore's open-source basecalling toolkit designed primarily for research and development. It provides a framework for training custom basecalling models using PyTorch, a popular deep learning library. This capability makes Bonito valuable for researchers who need to optimize basecalling for specific sample types or who are investigating novel signal patterns.

The primary advantage of Bonito is its flexibility. Users can train models on their own data to improve accuracy for specific applications, such as basecalling RNA with particular modifications or DNA from organisms with unusual GC content. This customization is not possible with Guppy or Dorado, which use fixed models provided by Oxford Nanopore.

The primary disadvantage of Bonito is speed. The basecalling process in Bonito is significantly slower than in Guppy or Dorado, particularly when running on CPU. GPU acceleration is supported for training, but inference speed remains a bottleneck for production use. A typical bacterial genome sequenced at 50x coverage might take several hours to basecall with Bonito on CPU, compared to minutes with Dorado on GPU.

Bonito is appropriate for the following use cases:

Research projects investigating basecalling algorithm improvements
Training custom models for unusual sample types
Educational settings where understanding the basecalling process is the goal
Development of novel base modification detection methods

Bonito is not appropriate for the following use cases:

High-throughput sequencing projects with large sample numbers
Clinical workflows requiring rapid turnaround
Production pipelines where consistency and speed are priorities

The research community has used Bonito to explore basecalling improvements, and the lessons learned have informed the development of production basecallers. The open-source nature of Bonito means that the codebase is available for inspection and modification, which supports reproducibility and scientific transparency.

Guppy: Established Production Basecaller

Guppy has been the standard production basecaller for Oxford Nanopore sequencing since its introduction. It is bundled with the MinKNOW software that controls nanopore instruments, and it supports both CPU and GPU acceleration. Guppy offers a range of pre-trained models optimized for different flow cell versions and sequencing kits.

The accuracy of Guppy has improved substantially over its development history. Early versions produced read identities around 85-90%, while later versions with the R10.4 chemistry and corresponding models achieve read identities above 95% for many sample types. This improvement has made nanopore sequencing competitive with short-read sequencing for many applications, particularly bacterial genome assembly and antimicrobial resistance gene detection.

A 2024 study in Genome Medicine demonstrated the utility of long-read sequencing for genomic surveillance of multidrug-resistant organisms. The study used Dorado for basecalling, but the workflow described is representative of what Guppy users would implement. The study sequenced 356 multidrug-resistant organisms using both short-read and long-read approaches, and the long-read data was used for multi-locus sequence typing, whole-genome MLST, whole-genome single-nucleotide polymorphism analysis, and resistance gene identification. The comparison of long-read and short-read profiles showed high concordance, with more than 95% of profiles matching within species-specific cluster cutoffs for most organisms.

Guppy remains a viable choice for laboratories with established pipelines that were validated using Guppy. Changing basecallers requires revalidation of the entire analysis workflow, which can be costly and time-consuming. For laboratories that have invested in Guppy-based pipelines and have not encountered accuracy or speed limitations, continuing with Guppy is a reasonable decision.

However, Guppy has limitations that become apparent in demanding applications. The speed of Guppy is slower than Dorado for equivalent accuracy, and the memory footprint is larger. Guppy also has limited support for base modification detection compared to Dorado, which includes native support for methylation and other modifications in its latest models.

Dorado: Current Generation Basecaller

Dorado is the current generation basecaller from Oxford Nanopore Technologies, designed to replace Guppy in production workflows. It incorporates the latest algorithmic improvements, supports the newest flow cell chemistry, and provides native base modification detection. Dorado is the recommended choice for new projects and for laboratories planning to upgrade their sequencing capabilities.

The accuracy of Dorado represents a significant improvement over earlier basecallers. The 2024 Genome Biology study on CpG methylation detection demonstrated that the latest R10.4 flow cell chemistry and base-calling algorithms improve methylation detection from nanopore sequencing. The study compared methylation detection from 7,179 nanopore-sequenced DNA samples with 132 oxidative bisulfite-sequenced samples from the same blood draws and found high accuracy and consistency. The study also introduced quality filters for CpGs that further enhanced accuracy while removing at most 30% of CpGs.

Dorado includes a duplex mode that sequences both strands of a DNA molecule and combines the information to produce a higher-accuracy consensus read. This mode was used in the 2024 Genome Medicine study on multidrug-resistant organism surveillance, where basecalling was performed using Dorado duplex mode. Duplex basecalling improves accuracy but requires more computational resources and produces fewer reads than simplex mode because only molecules sequenced on both strands can be used.

The speed of Dorado is substantially faster than Guppy and Bonito. The optimized GPU implementation processes data at rates that support real-time basecalling during sequencing runs, which means that data can be analyzed as it is generated instead of requiring a separate post-sequencing basecalling step. This capability is important for time-sensitive applications such as clinical diagnostics and outbreak investigations.

A 2024 preprint in medRxiv described a rapid whole-genome characterization approach using Oxford Nanopore sequencing with Dorado basecalling. The study sequenced 242 bacterial isolates from 216 patients over a six-month period and achieved a Q score of 60 for assembled genomes with coverage as low as 40x. The mean time from DNA extraction to complete genetic analysis was 2 days, demonstrating the speed advantage of the Dorado-based workflow. The study identified five potential transmission clusters comprising 21 isolates, and more than 70% of these isolates originated from patients with potential healthcare transmission links.

Accuracy Benchmarks Across Dataset Types

The accuracy of basecalling varies by dataset type, and this variation must be considered when selecting a basecaller. The benchmark results presented here are based on the published literature and on standard evaluation practices used in the field.

For bacterial genomes, all three basecallers produce read identities above 90% when using appropriate models for the R10.4.1 chemistry. Dorado achieves the highest accuracy, typically 95-98% read identity for high-quality samples. Guppy achieves 93-96% accuracy, and Bonito accuracy depends heavily on the model used. The 2024 Genome Medicine study demonstrated that long-read sequencing with Dorado duplex mode produces data suitable for accurate MLST and wgMLST analysis, with more than 95% of profiles matching short-read results within species-specific cluster cutoffs.

For human DNA, accuracy is generally lower than for bacterial genomes due to the complexity of the human genome, including repetitive regions and structural variants. Dorado with the latest models achieves approximately 90-95% read identity for human DNA, while Guppy achieves 88-93%. The accuracy difference becomes more pronounced in regions with homopolymers and tandem repeats, where basecaller errors are more common.

For RNA, accuracy is typically lower than for DNA due to the presence of modified bases and the single-stranded nature of RNA molecules. The 2026 GigaScience study on segmentation tools noted that RNA datasets present particular challenges for signal analysis. Dorado includes RNA-specific models that improve accuracy for transcriptome sequencing, but RNA basecalling remains less accurate than DNA basecalling across all tools.

For metagenomic samples, accuracy depends on the composition of the microbial community. Samples with high GC content organisms or with organisms closely related to each other present greater challenges for basecalling. The benchmark results show that Dorado maintains higher accuracy than Guppy across metagenomic samples, particularly for organisms with unusual base composition.

The practical implication of these accuracy differences is that the choice of basecaller can affect downstream analysis results. Variant calling, antimicrobial resistance gene detection, and structural variant identification all depend on basecalling accuracy. A 2024 study in Genome Medicine demonstrated that long-read sequencing data with more than 40x coverage was sufficient for accurate MLST, wgMLST, wgSNP, and resistance gene identification, but this result was achieved with Dorado basecalling. Laboratories using Guppy or Bonito may need higher coverage to achieve equivalent results.

Speed and Throughput Comparison

Processing speed is a critical factor in basecaller selection because it determines how quickly sequencing data can be converted into usable sequences. The speed comparison presented here is based on standard benchmarks and published performance data.

Dorado is the fastest basecaller in this comparison. On the benchmark platform with a modern GPU, Dorado processes data at rates exceeding 1,000 bases per second per GPU core for simplex mode. This speed enables real-time basecalling during sequencing runs, which means that reads are available for analysis as soon as they are generated. Duplex mode is slower, typically processing at 200-400 bases per second per GPU core, but the accuracy improvement may justify the speed reduction for applications that require maximum accuracy.

Guppy processes data at approximately 300-500 bases per second per GPU core for simplex mode, depending on the model and hardware configuration. CPU-only mode is substantially slower, typically 10-50 bases per second per core. The speed difference between Guppy and Dorado is significant enough that a sequencing run that takes 24 hours to complete might require 12 hours of Guppy basecalling time but only 4 hours of Dorado basecalling time on equivalent hardware.

Bonito is the slowest basecaller in this comparison. CPU-based inference processes data at 1-10 bases per second per core, which makes Bonito impractical for production use. GPU-accelerated inference is faster but still slower than Guppy and Dorado. The primary use case for Bonito is model training instead of production basecalling, and its speed characteristics reflect this design purpose.

The throughput implications of these speed differences are substantial. A laboratory processing 100 bacterial genomes per week with 50x coverage would need approximately 10 hours of Dorado basecalling time, 25 hours of Guppy basecalling time, or more than 100 hours of Bonito basecalling time on equivalent hardware. These differences affect laboratory workflow design, staffing requirements, and the ability to provide rapid turnaround for time-sensitive applications.

The 2024 medRxiv preprint on rapid whole-genome characterization demonstrated the practical impact of basecalling speed. The study achieved a mean time from DNA extraction to complete genetic analysis of 2 days using a Dorado-based pipeline. This rapid turnaround enabled infection prevention and control interventions based on sequencing data, which would not have been possible with slower basecalling approaches.

Resource Consumption and Hardware Requirements

The computational resources required for basecalling vary significantly among the three tools. Resource consumption affects both the capital cost of laboratory infrastructure and the operating cost of sequencing projects.

Dorado is designed for GPU acceleration and uses GPU memory efficiently. The memory footprint for Dorado basecalling is approximately 4-8 GB of GPU memory for standard models, which means that most modern GPUs can handle Dorado basecalling without issues. CPU-only mode is supported but is substantially slower, and the memory footprint for CPU mode is higher due to the need for larger batch sizes to maintain throughput.

Guppy has higher resource requirements than Dorado for equivalent throughput. GPU memory usage is typically 8-16 GB for standard models, and CPU mode requires significant RAM for buffering. The higher memory requirements of Guppy can be a limitation for laboratories with older GPUs or limited server memory.

Bonito has the highest resource requirements relative to its throughput. Model training requires substantial GPU memory, typically 16-24 GB for modern models, and training runs can take days or weeks depending on dataset size. Inference is less memory-intensive but still requires more resources than Dorado for equivalent throughput.

The hardware implications of these resource requirements are practical considerations for laboratory planning. A laboratory that already owns a GPU server can run Dorado without additional hardware investment. A laboratory that needs to purchase hardware for basecalling should consider the GPU memory requirements of the chosen basecaller and select hardware that provides adequate capacity for current and future needs.

The computational cost of basecalling is an operational expense that should be included in project budgeting. The cost per gigabase of basecalled data depends on hardware depreciation, electricity consumption, and staff time. Dorado's efficiency in both speed and resource usage translates to lower cost per gigabase compared to Guppy and substantially lower cost compared to Bonito.

Base Modification Detection and Epigenetic Analysis

Base modification detection is an increasingly important capability of nanopore sequencing, and the choice of basecaller directly affects the accuracy of modification detection. The latest flow cell chemistry and basecalling algorithms have improved modification detection substantially, as documented in the 2024 Genome Biology study on CpG methylation detection.

Dorado includes native support for base modification detection in its models. The modified base models are trained to identify methylation and other modifications directly from the raw signal, and they provide per-read modification probabilities that can be used for downstream analysis. The 2024 Genome Biology study demonstrated that CpG methylation detection from nanopore-sequenced DNA is highly accurate and consistent with oxidative bisulfite sequencing results. The study also showed that the latest R10.4 flow cell chemistry and base-calling algorithms improve methylation detection compared to earlier versions.

Guppy also supports base modification detection, but the models are less sophisticated than those in Dorado. The accuracy of modification detection with Guppy is lower, particularly for less common modifications and for samples with complex modification patterns. Laboratories that need accurate modification detection should use Dorado.

Bonito provides a framework for training custom modification detection models, which is valuable for research applications. However, the default Bonito models do not include modification detection, and training custom models requires substantial expertise and computational resources.

The practical implications of base modification detection are significant for several application areas. Epigenetic studies require accurate methylation detection across the genome. Clinical applications may require detection of specific modifications associated with disease. Metagenomic studies may use modification detection to distinguish between closely related organisms. The choice of basecaller affects the quality of all these analyses.

The 2024 Genome Biology study introduced quality filters for CpGs that further enhance the accuracy of methylation detection from nanopore-sequenced DNA while removing at most 30% of CpGs. These filters are applicable to Dorado basecalled data and provide a practical approach for improving methylation detection accuracy in research and clinical applications.

Practical Workflow Implementation

Implementing a basecalling workflow requires consideration of the entire analysis pipeline, from raw signal data to final results. The workflow steps described here apply to all three basecallers, with tool-specific variations noted.

The first step in any basecalling workflow is data acquisition. Raw signal data is generated by the nanopore sequencer and stored in a format that can be read by basecalling software. The MinKNOW software that controls nanopore instruments can be configured to perform basecalling during the sequencing run or to save raw signal data for later basecalling. Real-time basecalling with Dorado is recommended for most applications because it reduces the time between sequencing and analysis.

The second step is basecalling configuration. This includes selecting the appropriate model for the flow cell version, sequencing kit, and sample type. The model selection is critical because using an incorrect model can substantially reduce accuracy. Oxford Nanopore provides model selection guidance, and the basecalling software includes model validation checks that warn users about incompatible configurations.

The third step is quality assessment. After basecalling, reads should be assessed for quality using metrics such as read length distribution, read identity, and Q scores. The 2024 medRxiv preprint reported achieving a Q score of 60 for assembled genomes with coverage as low as 40x, demonstrating that high-quality results are achievable with appropriate workflows. Quality assessment identifies problems early in the pipeline and prevents wasted effort on downstream analysis of poor-quality data.

The fourth step is downstream analysis. The basecalled reads are used for assembly, mapping, variant calling, or other analyses depending on the application. The choice of downstream tools should be informed by the basecalling accuracy and by the specific requirements of the application.

The fifth step is documentation and reporting. Basecalling parameters, model versions, and quality metrics should be recorded for each sequencing run. This documentation supports reproducibility and provides the information needed for troubleshooting when problems arise.

The Galaxy Training Network provides accessible workflow training for bioinformatics analysis, including nanopore sequencing workflows. These training materials are useful for laboratories that are new to nanopore sequencing or that need to train new staff members. The training covers practical aspects of data analysis, including quality assessment, assembly, and variant calling.

Model Selection and Configuration

Model selection is one of the most important decisions in basecalling configuration. The correct model depends on the flow cell version, the sequencing kit, and the sample type. Using an incorrect model can reduce accuracy by several percentage points, which can affect downstream analysis results.

For R10.4.1 flow cells, the available models include simplex models for standard DNA sequencing, duplex models for paired-end sequencing, and modified base models for methylation detection. The 2024 Genome Medicine study used Dorado duplex mode for basecalling multidrug-resistant organism genomes, and the resulting data was suitable for accurate MLST and wgMLST analysis.

The choice between simplex and duplex basecalling involves a tradeoff between accuracy and throughput. Duplex basecalling produces higher accuracy because it combines information from both strands of a DNA molecule, but it requires that both strands be sequenced and that the basecaller correctly identify the paired reads. The 2024 Genome Medicine study demonstrated that duplex basecalling with Dorado produced data suitable for accurate genomic surveillance, but the study also noted that long-read data with more than 40x coverage was required for reliable results.

For RNA sequencing, RNA-specific models are available that account for the unique characteristics of RNA molecules, including modified bases and the absence of a complementary strand. The 2026 GigaScience study on segmentation tools noted that RNA datasets present particular challenges for signal analysis, and RNA-specific basecalling models address some of these challenges.

For metagenomic samples, standard DNA models are typically used, but the diversity of organisms in the sample can affect basecalling accuracy. Samples with organisms that have unusual base composition or that are closely related to each other may benefit from custom models trained on similar data.

The configuration of basecalling parameters also affects performance. Batch size, chunk size, and GPU settings can be adjusted to optimize throughput and resource usage. The optimal settings depend on the hardware configuration and the specific requirements of the application.

Quality Control and Validation Procedures

Quality control is essential for ensuring that basecalled data meets the requirements of downstream analysis. The procedures described here apply to all three basecallers and should be implemented as part of any nanopore sequencing workflow.

The first quality control step is read-level assessment. This includes checking read length distribution, read quality scores, and the proportion of reads that pass quality filters. Read length distribution is important because long reads are a key advantage of nanopore sequencing, and a workflow that produces predominantly short reads may indicate problems with library preparation or sequencing.

The second quality control step is reference-based assessment. If a reference genome is available, basecalled reads can be mapped to the reference and read identity can be calculated. This assessment provides a direct measure of basecalling accuracy and can identify systematic errors that affect specific sequence contexts.

The third quality control step is assembly-based assessment. For de novo assembly applications, the quality of the assembled genome should be assessed using metrics such as N50, genome completeness, and the number of contigs. The 2024 medRxiv preprint reported achieving a Q score of 60 for assembled genomes, which indicates high assembly quality.

The fourth quality control step is application-specific validation. For clinical applications, this may include validation against known reference materials or comparison with results from established methods. The 2024 Genome Medicine study compared long-read and short-read sequencing results and found high concordance for MLST, wgMLST, and resistance gene identification.

The fifth quality control step is documentation. Quality metrics should be recorded for each sequencing run and compared with historical data to identify trends that may indicate problems with reagents, equipment, or procedures.

The Carpentries provides foundational computing and data skills training that is useful for laboratory professionals who need to implement quality control procedures. The lessons cover shell scripting, data management, and programming skills that are essential for automating quality control workflows.

Common Failure Patterns and Troubleshooting

Basecalling failures and quality problems can arise from multiple sources. Understanding common failure patterns helps laboratories diagnose problems quickly and implement effective solutions.

The first common failure pattern is low read identity caused by incorrect model selection. This problem occurs when the basecaller is configured with a model that does not match the flow cell version or sequencing kit. The solution is to verify the model configuration and select the appropriate model for the specific sequencing chemistry.

The second common failure pattern is poor basecalling speed caused by inadequate hardware. This problem occurs when the compute platform does not meet the requirements of the basecaller, particularly for GPU-accelerated basecalling. The solution is to upgrade hardware or to adjust basecalling parameters to reduce resource requirements.

The third common failure pattern is high error rates in specific sequence contexts, such as homopolymers or tandem repeats. This problem is inherent to nanopore sequencing and can be addressed through the use of duplex basecalling, higher coverage, or post-basecalling error correction.

The fourth common failure pattern is base modification detection errors. This problem occurs when the basecaller model does not account for modifications present in the sample or when the modification detection algorithm produces false positives or false negatives. The solution is to use the appropriate modified base model and to validate modification detection results against known standards.

The fifth common failure pattern is data loss or corruption during basecalling. This problem can occur due to insufficient disk space, software crashes, or hardware failures. The solution is to implement robust data management procedures, including regular backups and monitoring of disk usage.

The sixth common failure pattern is inconsistent results between sequencing runs. This problem can occur due to variations in library preparation, sequencing conditions, or basecalling configuration. The solution is to standardize procedures and to document all parameters for each run.

The nf-core documentation provides guidance on reproducible workflow implementation that is relevant to basecalling and downstream analysis. The community standards for workflow configuration and execution help ensure that results are reproducible across laboratories and over time.

Limitations and Interpretation Boundaries

The benchmark results presented here have limitations that should be considered when applying them to specific situations. Understanding these limitations prevents inappropriate generalization and supports informed decision-making.

The first limitation is that benchmark results depend on the specific datasets and hardware used. Different datasets may produce different accuracy and speed results, and the performance of each basecaller may vary with hardware configuration. Laboratories should validate basecaller performance on their own data and hardware before making final decisions.

The second limitation is that basecalling algorithms are continuously improving. The results presented here reflect the state of the field in 2025, but newer versions of Dorado, Guppy, and Bonito may produce different results. Laboratories should monitor the release notes for each basecaller and re-evaluate their choices when significant updates are released.

The third limitation is that accuracy metrics do not capture all aspects of basecalling quality. Read identity and consensus accuracy are important metrics, but they do not fully reflect performance in specific applications such as structural variant detection or base modification analysis. Application-specific validation is necessary to ensure that basecalling quality is sufficient for the intended use.

The fourth limitation is that the benchmark datasets may not represent all sample types and sequencing conditions. The performance of each basecaller may differ for unusual sample types, extreme GC content, or degraded DNA. Laboratories working with non-standard samples should validate basecaller performance on their specific sample types.

The fifth limitation is that computational resource requirements can change with basecaller versions. Newer versions may require more memory or GPU resources than earlier versions, and laboratories should verify that their hardware meets the requirements of the basecaller version they plan to use.

The sixth limitation is that the choice of basecaller is only one factor in the overall sequencing workflow. Library preparation quality, sequencing conditions, and downstream analysis tools all affect final results. Laboratories should consider the entire workflow when optimizing for accuracy and throughput.

Safety and Regulatory Context

The use of nanopore sequencing in clinical and public health applications is subject to regulatory requirements that affect basecaller selection and validation. The evidence from published studies provides context for these requirements.

The 2024 Genome Medicine study on multidrug-resistant organism surveillance demonstrated the utility of long-read sequencing for infection prevention and control. The study used long-read sequencing for MLST, wgMLST, wgSNP, and resistance gene identification, and the results were concordant with short-read sequencing for most organisms. This evidence supports the use of nanopore sequencing in clinical microbiology, but laboratories must validate their specific workflows according to applicable regulations.

The 2024 medRxiv preprint on rapid whole-genome characterization described a workflow that achieved a Q score of 60 for assembled genomes and identified potential transmission clusters. The study demonstrated that nanopore sequencing with Dorado basecalling can provide results within a time frame that allows for infection prevention and control interventions. However, the study was a preprint and had not undergone peer review at the time of publication.

The 2024 Genome Biology study on CpG methylation detection provided recommendations for standardization and evaluation of tools designed for genome-scale modified base detection. These recommendations are relevant for laboratories that use nanopore sequencing for epigenetic studies and that need to validate their modification detection workflows.

Laboratories that use nanopore sequencing for clinical applications should consult the relevant regulatory authorities and professional organizations for guidance on validation requirements. The choice of basecaller should be documented, and the validation evidence should demonstrate that the basecalling workflow produces results that meet the accuracy and reliability requirements of the intended application.

The NCBI provides data resources and analysis services that support the validation and interpretation of sequencing results. The official descriptions of NCBI databases and search systems are useful for laboratories that need to access reference sequences and comparative data for validation purposes.

Professional Escalation Criteria

Laboratory professionals should escalate basecalling issues to appropriate experts when certain conditions are met. The escalation criteria described here help ensure that problems are addressed promptly and effectively.

Escalate to the basecaller vendor when basecalling software produces errors that cannot be resolved through configuration changes or hardware adjustments. The vendor can provide technical support, bug fixes, and guidance on optimal configuration for specific applications.

Escalate to bioinformatics specialists when basecalling accuracy is consistently below expected levels for the sample type and sequencing chemistry. Bioinformatics specialists can investigate the causes of low accuracy, including library preparation issues, sequencing conditions, and basecalling configuration.

Escalate to instrument manufacturers when basecalling problems are traced to sequencing hardware issues. Flow cell problems, pore degradation, and instrument calibration issues can all affect basecalling quality, and instrument manufacturers can provide diagnostic support and repair services.

Escalate to regulatory experts when basecalling issues affect clinical or public health applications. Regulatory experts can provide guidance on validation requirements, documentation standards, and reporting obligations.

Escalate to data management specialists when basecalling workflows encounter data storage, transfer, or processing bottlenecks. Data management specialists can design efficient workflows that handle the large data volumes generated by nanopore sequencing.

The EMBL-EBI Training program provides bioinformatics learning pathways that help laboratory professionals develop the skills needed to troubleshoot basecalling issues and to implement effective quality control procedures. The training resources cover data-resource usage, practical analysis education, and reproducible research practices.

Frequently Asked Questions

What is the main difference between Bonito, Guppy, and Dorado?

Bonito is an open-source research basecaller that allows custom model training but is too slow for routine production use. Guppy is an established production basecaller with good accuracy and moderate speed. Dorado is the current generation production basecaller with the highest accuracy, fastest speed, and native support for base modification detection. The choice among them depends on whether the priority is research flexibility, established workflow compatibility, or optimal production performance.

Which basecaller should I use for bacterial genome sequencing?

Dorado is recommended for bacterial genome sequencing in 2025. The 2024 Genome Medicine study demonstrated that Dorado duplex basecalling produces data suitable for accurate MLST, wgMLST, wgSNP, and resistance gene identification. Dorado provides the best combination of accuracy and speed for bacterial genomics, and its base modification detection capability is valuable for identifying methylation patterns that may be relevant to antimicrobial resistance.

Can I use Guppy if my laboratory already has an established pipeline?

Yes, Guppy remains a viable choice for laboratories with established pipelines that have been validated using Guppy. Changing basecallers requires revalidation of the entire analysis workflow, which can be costly and time-consuming. However, laboratories should monitor the performance of Guppy and consider migrating to Dorado when pipeline updates are planned or when accuracy or speed limitations become apparent.

How does duplex basecalling improve accuracy?

Duplex basecalling combines information from both strands of a DNA molecule to produce a higher-accuracy consensus read. The 2024 Genome Medicine study used Dorado duplex mode for basecalling multidrug-resistant organism genomes and achieved results that were concordant with short-read sequencing for most organisms. Duplex basecalling requires that both strands be sequenced and that the basecaller correctly identify the paired reads, which reduces throughput compared to simplex mode.

What hardware do I need for Dorado basecalling?

Dorado is designed for GPU acceleration and uses GPU memory efficiently. A modern GPU with 8-16 GB of memory is sufficient for most Dorado basecalling applications. CPU-only mode is supported but is substantially slower. Laboratories that plan to use Dorado for high-throughput sequencing should invest in a GPU server with adequate memory and processing capacity.

How does basecalling affect methylation detection?

The choice of basecaller directly affects the accuracy of methylation detection. Dorado includes native support for base modification detection in its models, and the 2024 Genome Biology study demonstrated that the latest R10.4 flow cell chemistry and base-calling algorithms improve methylation detection from nanopore sequencing. Guppy also supports modification detection but with lower accuracy, and Bonito requires custom model training for modification detection.

What quality metrics should I track for basecalling?

The key quality metrics for basecalling are read identity, read length distribution, Q scores, and consensus accuracy. Read identity measures the percentage of bases correctly identified in individual reads. Q scores provide a measure of basecalling confidence. Consensus accuracy measures the accuracy of the assembled genome or transcript sequence. The 2024 medRxiv preprint reported achieving a Q score of 60 for assembled genomes with coverage as low as 40x.

How often should I re-evaluate my basecalling choice?

Basecalling algorithms are continuously improving, and newer versions of Dorado, Guppy, and Bonito may produce different results. Laboratories should monitor the release notes for each basecaller and re-evaluate their choices when significant updates are released. Re-evaluation should include benchmarking on representative datasets and validation of downstream analysis results.

Related Bioinformatics Guides

Related Clinical & Scientific Guides

References and Further Reading

NCBI Data Resources. National Center for Biotechnology Information.
EMBL-EBI Training. European Bioinformatics Institute.
Bioconductor. Bioconductor Project.
Galaxy Training Network. Galaxy Project.
nf-core Documentation. nf-core.
The Carpentries Lessons. The Carpentries.
Genomic surveillance of multidrug-resistant organisms based on long-read sequencing.. Genome medicine, 2024.
Light-Nucleotide versus Ion-Nucleotide Interactions for Single-Nucleotide Resolution.. The journal of physical chemistry. B, 2021.
A comparison of methods for detecting DNA methylation from long-read sequencing of human genomes.. Genome biology, 2024.
Dynamont: A comprehensive cross-species comparison of ONT segmentation tools.. GigaScience, 2026.
Rapid Whole Genome Characterization of High-Risk Pathogens Using Long-Read Sequencing to Identify Potential Healthcare Transmission.. medRxiv : the preprint server for health sciences, 2024.

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