Iterative Polishing: How Many Rounds of Racon and Medaka Are Enough for a High-Quality Genome?

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

Key Takeaways

• For most bacterial and fungal genomes assembled from Nanopore long-read data with adequate coverage (≥30x), one round of Racon followed by one round of Medaka is generally sufficient to reach the practical accuracy ceiling, yielding negligible improvements with additional rounds while consuming substantial computational resources.
• Hybrid polishing approaches incorporating Illumina short reads, typically involving one round of Racon followed by one or two rounds of a short-read polisher like Homopolish, often achieve comparable or superior accuracy and are recommended when high-fidelity reference genomes are required.
• The effectiveness of polishing is critically dependent on read coverage; a minimum of 30x Nanopore coverage is recommended for reliable error correction, with diminishing returns beyond approximately 100x as the polishing tool's error model becomes the limiting factor.
• Tool selection is paramount, with Medaka being highly accurate when its trained base-calling model matches the sequencing data, while Racon offers speed and versatility without model dependency, and Homopolish excels at correcting homopolymer errors common in Nanopore reads.
• Polishing cannot rectify structural errors (misjoins, collapsed repeats) present in the initial assembly; therefore, assessing draft assembly quality using metrics like BUSCO completeness and contiguity (N50) before polishing is essential, as assembly strategy dictates the ultimate achievable quality.
• The "Stop-When-Measured" framework, where polishing ceases when measured improvements in QV (e.g., <1 point) and BUSCO completeness (e.g., <0.5 percentage points) between consecutive rounds fall below predefined thresholds, prevents overpolishing and the introduction of new errors.

Genome assembly polishing is the process of correcting base-level errors in a draft assembly using additional sequence data. For researchers using long-read sequencing platforms such as Oxford Nanopore Technologies, the question of how many polishing rounds are necessary is a practical one that directly affects both final genome quality and computational expenditure. For most bacterial and fungal genomes assembled from Nanopore reads, one round of Racon followed by one round of Medaka is sufficient to reach the practical accuracy ceiling, with additional rounds yielding negligible improvements while consuming substantial compute time. For hybrid approaches that incorporate Illumina short reads, one round of Racon followed by one or two rounds of a short-read polisher such as Homopolish typically achieves comparable or superior accuracy. The evidence base for these recommendations comes from comparative evaluations of polishing tools, published genome assembly projects, and the operational documentation of bioinformatics platforms that support reproducible assembly workflows.

The scope of this guidance covers microbial genomes, including bacteria and fungi, assembled from Nanopore long-read data with moderate to high coverage. The principles extend to larger eukaryotic genomes, but the computational cost considerations scale accordingly. Readers working with plant or animal genomes should expect longer runtimes and may need to adjust their polishing strategies based on available high-performance computing resources. The recommendations here are grounded in published comparative studies and real-world assembly projects instead of theoretical expectations, and they emphasize measurable quality outcomes over procedural habit.

The Polishing Problem in Long-Read Assembly

Long-read sequencing platforms produce reads with higher error rates than short-read platforms. Nanopore sequencing, in particular, generates reads with error rates that vary by base-calling model, pore version, and library preparation method. These errors propagate into the initial assembly, producing a draft genome that may contain insertion, deletion, and substitution errors relative to the true biological sequence. Polishing is the computational correction of these errors using either the original reads or additional sequencing data from a more accurate platform.

The need for polishing is not uniform across all assembly tools. Some assemblers, such as Flye and NECAT, perform internal error correction during the assembly process and produce drafts that are already relatively accurate. Others, such as Canu, may produce drafts with higher error rates that require more aggressive polishing. The choice of assembler therefore influences how many polishing rounds are needed to reach a given quality threshold.

The practical consequence of insufficient polishing is a genome that contains systematic errors in protein-coding regions, leading to incorrect gene predictions, truncated open reading frames, and misleading functional annotations. The practical consequence of excessive polishing is wasted computational time and, in some cases, the introduction of new errors where the polishing tool misinterprets genuine biological variation as sequencing error. Understanding where the balance lies requires examining the empirical evidence from comparative studies and published assembly projects.

What the Comparative Evidence Shows

A comparative evaluation of Nanopore polishing tools for microbial genome assembly provides the most direct evidence on this question. The study, published in Scientific Reports in 2021, assessed eight state-of-the-art polishing tools and combinations using BUSCO completeness scores and Prokka gene prediction as quality metrics. The evaluation used cost-effective Nanopore long-read systems, including Flongle, which produce lower coverage than standard flow cells and therefore represent a challenging polishing scenario.

The study found that Homopolish, PEPPER, and Medaka performed better than other individual tools in terms of assembly quality. In combination polishing, a second round of Homopolish and the PEPPER followed by Medaka combination also showed better results than other approaches. However, the authors noted that individual tools and combinations have specific limitations on usage and results, and that depending on the target organism and the purpose of the downstream research, there remain difficulties in perfectly replacing hybrid polishing carried out by the addition of short reads. The study concluded that through continuous improvement of protein pores, related base-calling algorithms, and polishing tools based on improved error models, a high-quality microbial genome can be achieved using only Nanopore reads without the production of additional short-read data.

The key finding for the practical question of polishing rounds is that the quality gains from additional polishing rounds diminish rapidly. The first round of polishing produces the largest improvement in accuracy, with subsequent rounds producing progressively smaller gains. For most microbial genomes, the difference between one round and two rounds of Medaka is small, and the difference between two and three rounds is typically negligible. The study also demonstrated that the choice of polishing tool matters more than the number of rounds, with a well-chosen single tool outperforming multiple rounds of a poorly chosen tool.

The study's conclusion that high-quality microbial genomes can be achieved using only Nanopore reads without additional short-read data is significant for laboratories that lack access to Illumina sequencing. The authors attributed this to continuous improvement in protein pores, base-calling algorithms, and polishing tools based on improved error models. For researchers who do have access to short-read data, hybrid polishing remains a viable option, but it is not always necessary for achieving acceptable genome quality.

At a Glance: Polishing Round Recommendations

The following table summarizes recommended polishing strategies based on assembly tool, read type, and quality goals. These recommendations apply to bacterial and fungal genomes with coverage of at least 30x from Nanopore reads.

The quality values in this table are indicative based on published assembly projects. The bighead catfish genome assembly, which used high-fidelity long-read sequencing from Pacific Biosciences and Oxford Nanopore Technologies, scaffolded with high-throughput chromosome conformation capture data, and polished with Illumina short-read sequencing, achieved a quality value of 50 and 95.5% BUSCO completeness. This example demonstrates the quality ceiling that hybrid polishing can achieve for a complex eukaryotic genome. For microbial genomes, similar or better quality values are achievable with Nanopore-only polishing when coverage is adequate.

Core Principles of Polishing Strategy

Polishing Does Not Replace Assembly Quality

The quality of the initial assembly sets the ceiling for what polishing can achieve. Polishing corrects base-level errors, but it cannot fix structural errors such as misjoins, misassemblies, or collapsed repeats. If the assembler produces a draft with structural problems, no amount of polishing will produce a high-quality genome. The choice of assembler and the quality of the input reads are therefore more important than the polishing strategy.

The hybrid assembly of the basidiomycete fungus Candolleomyces candolleanus strain CMU-8613 illustrates this point. The assembled genome size ranged from 46.8 Mb using NECAT plus Racon to 59.3 Mb using Canu plus a different genome assembly, depending on the assembly and polishing strategy. This variation of more than 12 Mb in a genome that should be approximately 50 Mb indicates that the assembly strategy, not the polishing strategy, was the primary determinant of genome size and content. The study also found substantial variation in gene content and distribution within the same genome depending on the tools used, with CAZyme-encoding genes varying from 494 in the Flye assembly to 453 in the NECAT assembly. These observations demonstrate that the assembly and annotation pipeline, including the polishing strategy, affects base-level accuracy and the biological conclusions drawn from the genome.

Read Coverage Determines Polishing Effectiveness

Polishing tools work by aligning reads back to the assembly and using the read information to correct errors. The accuracy of this correction depends on the depth of coverage at each position. Low coverage regions may not have enough reads to support confident error correction, and polishing tools may either leave errors uncorrected or introduce new errors by misinterpreting read variation.

For Nanopore-only polishing, coverage of at least 30x is generally recommended for reliable error correction. Higher coverage, in the range of 50-100x, provides more confidence in the polishing calls and may reduce the number of rounds needed. However, coverage beyond approximately 100x provides diminishing returns because the error model of the polishing tool becomes the limiting factor instead of the amount of data.

The comparative evaluation of polishing tools used Flongle data, which produces lower coverage than standard flow cells. The finding that Homopolish, PEPPER, and Medaka performed well even under these conditions suggests that these tools are robust to moderate coverage. However, the authors also noted that individual tools have specific limitations, and the choice of tool should be informed by the expected coverage and error profile of the data.

Error Models and Tool Selection

Each polishing tool uses a different error model to distinguish sequencing errors from genuine biological variation. Racon uses a partial order alignment approach that does not require a trained model, making it fast and applicable to any read type. Medaka uses a neural network trained on specific base-calling models and pore versions, which makes it more accurate when the training data matches the sequencing conditions but less reliable when they do not. Homopolish uses a different approach that is particularly effective at correcting homopolymer errors, which are common in Nanopore data.

The practical implication is that the choice of polishing tool should be informed by the sequencing platform, base-calling model, and expected error profile. Medaka is most effective when the base-calling model used for polishing matches the model used for the original base-calling. If the base-calling model has changed between the initial assembly and the polishing step, the error model may be mismatched, and Racon or Homopolish may be more reliable.

The comparative study found that the second round of Homopolish and the PEPPER followed by Medaka combination showed better results than other combinations. This suggests that combining tools with different error models can be more effective than multiple rounds of the same tool. The specific combination that works best depends on the data, and researchers should be prepared to test a small number of combinations instead of assuming that one approach will work universally.

Practical Workflow for Polishing Decisions

Step 1: Assess the Initial Assembly Quality

Before deciding on a polishing strategy, assess the quality of the initial assembly. Run BUSCO to measure completeness against a lineage-specific set of single-copy orthologs. Run QUAST or a similar tool to measure contiguity statistics such as N50 and the number of contigs. If the assembly has low completeness or poor contiguity, address these issues before polishing instead of expecting polishing to fix them.

For bacterial genomes, a complete assembly should consist of a small number of contigs, ideally one per replicon. For fungal genomes, the number of contigs will be higher, but the assembly should be largely complete with no missing core genes. The Candolleomyces candolleanus study demonstrated that assembly quality varies substantially depending on the tools used, so the initial assessment should inform whether the assembly is worth polishing or whether reassembly with different parameters is warranted.

Step 2: Select the Polishing Tools Based on Data Availability

If Illumina short reads are available, plan for a hybrid polishing approach. One round of Racon using the Nanopore reads followed by one round of a short-read polisher such as Homopolish or Pilon is typically sufficient. The bighead catfish genome project used Illumina short-read polishing after assembly with high-fidelity long reads, achieving a QV of 50, which demonstrates the effectiveness of this approach for complex genomes.

If only Nanopore reads are available, plan for one round of Racon followed by one round of Medaka. If the base-calling model used for the reads is known and matches the Medaka training data, Medaka should perform well. If the base-calling model is uncertain or unusual, consider using Homopolish instead of or in addition to Medaka.

Step 3: Run the First Polishing Round and Measure the Effect

Run the first polishing round and measure the quality improvement. Compare BUSCO completeness, QV, and the number of predicted genes before and after polishing. The first round should produce a substantial improvement in all metrics. If the improvement is small, the initial assembly may already be highly accurate, or the polishing tool may not be well matched to the data.

The Mesorhizobium sp. strain ORM16 genome project, which involved complete genome characterization of a rhizobial microsymbiont from root nodules of Ononis repens, demonstrates the importance of accurate assembly for downstream biological interpretation. The complete genomic characterization enhances understanding of plant-microbe interactions in Mediterranean forest environments and contributes to molecular taxonomy. For such applications, the polishing strategy must produce a genome that supports confident gene prediction and comparative analysis.

Step 4: Decide Whether Additional Rounds Are Warranted

After the first round of polishing, assess whether additional rounds are likely to produce meaningful improvement. If the QV is above 40 and BUSCO completeness is above 95%, additional polishing rounds are unlikely to produce substantial gains. If the QV is below 40 or BUSCO completeness is below 90%, additional polishing may be warranted, but the cause of the low quality should be investigated first.

The diminishing returns from additional polishing rounds are well documented. The comparative evaluation found that the second round of Homopolish showed better results than other combinations, but the improvement over the first round was modest. Additional rounds beyond the second typically produce negligible improvements while consuming significant computational time.

Step 5: Document the Polishing Strategy for Reproducibility

Record the exact polishing commands, tool versions, and parameters used. This documentation is essential for reproducibility and for interpreting the quality metrics of the final assembly. The nf-core documentation emphasizes the importance of reproducible workflow standards, and the Galaxy Training Network provides accessible training on assembly and polishing workflows that include documentation practices.

The StrainCascade workflow, described as an automated modular pipeline for high-throughput long-read bacterial genome reconstruction, integrates genome assembly, accurate annotation, and comprehensive functional profiling into a single reproducible framework. The emphasis on deterministic execution strategies and systematic resolution of strain-level variability highlights the importance of documenting the polishing steps and the entire assembly and analysis pipeline.

Options and Tradeoffs in Polishing Tools

Racon: Fast and Versatile

Racon is a polishing tool that uses partial order alignment to correct errors in a draft assembly. It does not require a trained error model, which makes it applicable to any read type and any sequencing platform. Racon is computationally efficient and can polish a bacterial genome in minutes on a standard workstation.

The primary tradeoff is that Racon is less accurate than tools that use trained error models, particularly for homopolymer regions and other error-prone sequence contexts. Racon is best used as a first polishing round to correct the most obvious errors before applying a more accurate tool such as Medaka.

Medaka: Accurate but Model-Dependent

Medaka uses a neural network trained on specific base-calling models and pore versions to correct errors in a draft assembly. It is more accurate than Racon when the training data matches the sequencing conditions, but it may perform poorly when the base-calling model is mismatched.

The primary tradeoff is that Medaka requires knowledge of the base-calling model used for the reads, and the tool must be updated when new base-calling models are released. For researchers who use the latest base-calling models, Medaka is an excellent choice. For researchers who use older or unusual base-calling models, Medaka may not be the best option.

Homopolish: Specialized for Homopolymer Errors

Homopolish is a polishing tool that is particularly effective at correcting homopolymer errors, which are common in Nanopore data. The comparative evaluation found that Homopolish performed better than other individual tools and that a second round of Homopolish showed better results than other combinations.

The primary tradeoff is that Homopolish may not be as effective at correcting other types of errors, such as substitutions or short indels. Homopolish is best used in combination with other tools, either as a second polishing round after Racon or Medaka or as a final correction step.

PEPPER: Deep Learning Based

PEPPER is a deep learning based polishing tool that uses a different architecture than Medaka. The comparative evaluation found that PEPPER performed well as an individual tool and that the PEPPER followed by Medaka combination showed better results than other combinations.

The primary tradeoff is that PEPPER may require more computational resources than other tools, particularly for larger genomes. PEPPER is a good choice for researchers who have access to GPU resources and who want to use a deep learning approach for polishing.

Hybrid Polishing with Short Reads

Hybrid polishing uses short reads, typically from Illumina sequencing, to correct errors in a long-read assembly. This approach is highly accurate because short reads have much lower error rates than long reads, and the combination of long-read contiguity with short-read accuracy produces high-quality genomes.

The primary tradeoff is the additional cost and time required to generate short-read data. For laboratories that already have Illumina sequencing capacity, hybrid polishing is an attractive option. For laboratories that do not, Nanopore-only polishing can achieve acceptable quality for many applications, as demonstrated by the comparative evaluation.

Observations and Measurements for Polishing Decisions

Quality Metrics to Track

The most useful quality metrics for polishing decisions are BUSCO completeness, QV, and the number of predicted genes. BUSCO completeness measures the fraction of single-copy orthologs from a lineage-specific database that are present in the assembly. QV measures the base-level accuracy of the assembly, with higher values indicating fewer errors. The number of predicted genes provides a functional measure of assembly quality, as assemblies with errors in coding regions will produce truncated or missing gene predictions.

The Candolleomyces candolleanus study used BUSCO and gene prediction to evaluate assembly quality, finding substantial variation in gene content depending on the assembly and annotation tools used. The study identified 15-25 secondary metabolite gene clusters depending on the genome assembly and the tools used for BGC prediction, and CAZyme-encoding genes varied from 494 in the Flye assembly to 453 in the NECAT assembly. These observations demonstrate that polishing decisions affect base-level accuracy and the biological conclusions drawn from the genome.

Recording Polishing Parameters

For each polishing round, record the tool name, version, command, and parameters. Also record the input assembly version and the output assembly version. This documentation allows the polishing process to be reproduced exactly and provides context for interpreting the quality metrics of the final assembly.

The nf-core documentation provides guidance on reproducible workflow standards, and the Galaxy Training Network offers training on assembly and polishing workflows that include documentation practices. The Carpentries lessons provide foundational training on shell, Git, and programming that supports reproducible bioinformatics workflows.

Comparing Polishing Outcomes

When comparing different polishing strategies, use the same quality metrics and the same evaluation tools. Run BUSCO with the same lineage database and the same parameters for all assemblies being compared. Run gene prediction with the same tool and the same parameters. This consistency ensures that differences in quality metrics reflect differences in assembly quality instead of differences in evaluation methodology.

The comparative evaluation of polishing tools used BUSCO and Prokka gene prediction as quality metrics, providing a consistent framework for comparing eight different polishing tools and combinations. The study found that individual tools and combinations have specific limitations on usage and results, and that the best choice depends on the target organism and the purpose of the downstream research.

Common Failure Patterns in Polishing

Overpolishing and Error Introduction

Polishing tools can introduce new errors when they misinterpret genuine biological variation as sequencing error. This is particularly problematic in regions of the genome that contain repetitive sequences, structural variants, or other features that differ from the reference used for training the error model. Overpolishing is more likely when multiple rounds of polishing are applied, as each round provides an opportunity for the tool to introduce new errors.

The practical consequence of overpolishing is an assembly that has high QV according to the polishing tool's own error model but contains systematic errors in biologically important regions. This can lead to incorrect gene predictions and misleading functional annotations. To avoid overpolishing, stop polishing when the quality metrics stop improving substantially between rounds.

Mismatched Error Models

Medaka and other model-based polishing tools are trained on specific base-calling models and pore versions. If the reads were base-called with a different model than the one used for polishing, the error model may be mismatched, and the polishing tool may perform poorly. This is a common problem when researchers use the latest base-calling models but have not updated their polishing tools to match.

The practical consequence of a mismatched error model is that polishing may not correct errors effectively, or may introduce new errors. To avoid this problem, check that the polishing tool version and model match the base-calling model used for the reads. If the base-calling model is unknown, use a tool that does not require a trained error model, such as Racon or Homopolish.

Insufficient Coverage for Polishing

Polishing tools require sufficient read coverage at each position to make confident error corrections. If coverage is too low, the tool may leave errors uncorrected or make incorrect corrections based on limited information. This is particularly problematic for genomes with extreme base composition, such as high GC or high AT content, which may have uneven coverage.

The practical consequence of insufficient coverage is an assembly that retains errors in low coverage regions. To avoid this problem, ensure that the sequencing depth is adequate for the genome size and complexity. For microbial genomes, coverage of at least 30x is generally recommended, with higher coverage providing more confidence in the polishing calls.

Ignoring Assembly Quality Before Polishing

Polishing cannot fix structural errors in the assembly. If the initial assembly has misjoins, collapsed repeats, or other structural problems, polishing will not correct them. Researchers who skip the initial quality assessment and proceed directly to polishing may waste computational time on an assembly that requires reassembly instead of polishing.

The practical consequence of ignoring assembly quality is a polished assembly that still has structural errors, leading to incorrect conclusions about genome structure and content. To avoid this problem, assess the initial assembly quality before polishing and address structural issues before proceeding.

Limitations of Polishing Approaches

Nanopore-Only Polishing Cannot Match Hybrid Polishing for All Applications

The comparative evaluation found that high-quality microbial genomes can be achieved using only Nanopore reads without additional short-read data. However, the authors also noted that there remain some difficulties in perfectly replacing hybrid polishing for certain applications. For applications that require the highest possible accuracy, such as clinical diagnostics or reference genome generation, hybrid polishing with short reads remains the gold standard.

The bighead catfish genome project used Illumina short-read polishing after assembly with high-fidelity long reads, achieving a QV of 50. This level of accuracy is difficult to achieve with Nanopore-only polishing, particularly for complex eukaryotic genomes. Researchers should consider the accuracy requirements of their application when deciding whether to invest in short-read data for hybrid polishing.

Polishing Tools Have Specific Limitations

Each polishing tool has specific limitations on usage and results. Racon is fast but less accurate than model-based tools. Medaka is accurate but requires a matching base-calling model. Homopolish is effective for homopolymer errors but may not correct other error types as well. PEPPER requires more computational resources than other tools.

The practical implication is that no single polishing tool is universally best. Researchers should test a small number of tools and combinations on their data and select the approach that produces the best quality metrics for their specific genome and sequencing conditions.

Quality Metrics Have Interpretation Limits

BUSCO completeness and QV are useful quality metrics, but they have limitations. BUSCO completeness measures the presence of single-copy orthologs, but it does not measure the accuracy of the sequence within those orthologs. QV measures base-level accuracy, but it does not measure structural accuracy or the correctness of gene predictions.

The Candolleomyces candolleanus study found substantial variation in gene content and distribution within the same genome depending on the assembly and annotation tools used. This variation demonstrates that quality metrics based on gene prediction can be sensitive to the choice of tools, and that the biological interpretation of the genome depends on the entire assembly and annotation pipeline, beyond the polishing strategy.

Safety and Regulatory Context for Polishing Decisions

Data Management and Reproducibility

Genome assembly and polishing generate large amounts of data, including raw reads, intermediate assemblies, and final assemblies. Proper data management is essential for reproducibility and for compliance with data sharing requirements from funders and journals. The NCBI provides data resources for sequence data, and the EMBL-EBI Training offers training on data-resource management and practical analysis education.

The nf-core documentation emphasizes the importance of reproducible workflow standards, and the Galaxy Training Network provides accessible workflow training and analysis tutorials. The Carpentries lessons provide foundational training on computing, data, shell, Git, and programming that supports reproducible bioinformatics workflows.

Professional Escalation Criteria

Researchers should escalate to a more experienced bioinformatician or a core facility when they encounter persistent quality problems that cannot be resolved through standard polishing approaches. Specific escalation criteria include:

• BUSCO completeness below 90% after polishing, indicating possible assembly or polishing problems
• QV below 30 after polishing, indicating substantial residual errors
• Large discrepancies between the expected and observed genome size, indicating possible assembly artifacts
• Inconsistent results between different polishing tools, indicating possible error model mismatches
• Poor gene prediction results despite acceptable BUSCO and QV metrics, indicating possible systematic errors in coding regions

The StrainCascade workflow provides an example of a fully automated modular pipeline that integrates genome assembly, accurate annotation, and comprehensive functional profiling into a single reproducible framework. For researchers who need high-throughput processing of multiple genomes, such automated workflows can reduce the risk of inconsistent polishing decisions across samples.

A Practical Decision Framework for Polishing Rounds

The question of how many polishing rounds to run is often answered by habit or by copying parameters from published pipelines. A more defensible approach is to treat the number of rounds as a variable that you adjust based on measured quality metrics at each step. This section presents a decision framework that uses observable assembly statistics to determine when additional polishing is justified and when it is simply consuming compute time without improving the final product.

The Stop-When-Measured Framework

The core principle of this framework is that you stop polishing when the measured improvement between consecutive rounds falls below a predefined threshold. This approach replaces guesswork with a reproducible rule that can be applied consistently across different genomes and sequencing conditions.

Define your stopping threshold before you begin polishing. A practical threshold for microbial genomes is a QV improvement of less than 1 point between rounds, combined with a BUSCO completeness improvement of less than 0.5 percentage points. When both metrics fall below these thresholds, additional rounds of the same polishing tool are unlikely to produce meaningful gains.

The comparative evaluation of Nanopore polishing tools supports this approach. The study found that the first round of polishing produces the largest improvement in assembly quality, with subsequent rounds producing progressively smaller gains. The second round of Homopolish showed better results than other combinations, but the improvement over the first round was modest. Beyond the second round, the quality gains were negligible for most tools.

Implementing the Framework in Practice

To apply this framework, you need a consistent measurement protocol. Run BUSCO with the same lineage database and the same parameters after every polishing round. Run QV estimation with the same tool and settings. Record these values in a simple table alongside the polishing command, tool version, and wall-clock time for each round.

The following steps outline the decision process:

1. Run one round of Racon on the raw assembly. Measure BUSCO completeness and QV.
2. Run one round of Medaka on the Racon-polished assembly. Measure BUSCO completeness and QV.
3. Compare the metrics between step 1 and step 2. If the QV improvement is greater than 1 point or the BUSCO improvement is greater than 0.5 percentage points, proceed to step 4. Otherwise, stop and use the step 2 assembly as the final polished genome.
4. Run a second round of Medaka. Measure BUSCO completeness and QV.
5. Compare the metrics between step 2 and step 4. If the improvement still exceeds the thresholds, you may consider a third round, but be aware that the gains will likely be small. If the improvement is below the thresholds, stop and use the step 2 assembly.

This framework is deliberately conservative. It errs on the side of stopping early instead of overpolishing, because the risk of introducing new errors increases with each additional round. The comparative evaluation noted that individual tools and combinations have specific limitations on usage and results, and that overpolishing can misinterpret genuine biological variation as sequencing error.

A Record System for Polishing Decisions

A simple spreadsheet or text file is sufficient to track polishing decisions across multiple genomes. The following fields capture the essential information for each polishing round:

• Sample identifier
• Assembly tool and version
• Assembly parameters
• Polishing tool and version
• Polishing command
• Base-calling model used for the reads
• Input assembly file name and version
• Output assembly file name and version
• BUSCO completeness percentage and lineage database used
• QV score
• Wall-clock time for the polishing round
• Date and operator name

This record system serves two purposes. First, it provides the data needed to apply the stop-when-measured framework consistently. Second, it creates a reproducible audit trail that supports the documentation requirements emphasized by the nf-core documentation and the Galaxy Training Network. The Carpentries lessons provide foundational training on shell and Git that supports maintaining such records in a version-controlled manner.

Troubleshooting When Quality Does Not Improve

When the first polishing round produces little or no improvement, the problem is usually not the number of rounds. The following diagnostic sequence addresses the most common causes:

First, verify that the polishing tool version matches the base-calling model used for the reads. Medaka is trained on specific base-calling models and pore versions. If the reads were base-called with a newer model than the Medaka version supports, the error model will be mismatched and polishing will be ineffective. Check the Medaka documentation for the supported models and compare against your base-calling settings.

Second, examine the read coverage in regions where errors persist. Low coverage regions may not have enough reads to support confident error correction. The comparative evaluation used Flongle data, which produces lower coverage than standard flow cells, and found that Homopolish, PEPPER, and Medaka performed well even under these conditions. However, the authors also noted that individual tools have specific limitations, and coverage below 30x may require a different approach.

Third, assess whether the initial assembly has structural errors that polishing cannot fix. Polishing corrects base-level errors but cannot repair misjoins, collapsed repeats, or other structural problems. If the assembly has poor contiguity or missing BUSCO genes before polishing, additional polishing rounds will not address these issues. The Candolleomyces candolleanus study demonstrated that assembly quality varies substantially depending on the tools used, with genome size ranging from 46.8 Mb to 59.3 Mb depending on the assembly and polishing strategy. This variation indicates that the assembly step, not the polishing step, was the primary determinant of genome content.

Comparing Polishing Strategies on a Test Genome

For laboratories that process many genomes, a more systematic approach is to benchmark polishing strategies on a single representative genome before applying the chosen strategy to the full batch. Select a genome with typical GC content, repeat structure, and coverage for your study system. Run the following comparisons:

• One round of Racon only
• One round of Racon followed by one round of Medaka
• One round of Racon followed by two rounds of Medaka
• One round of Racon followed by one round of Homopolish
• One round of Racon followed by one round of PEPPER and one round of Medaka

Measure BUSCO completeness, QV, and the number of predicted genes for each strategy. Also record the wall-clock time for each approach. The comparative evaluation found that Homopolish, PEPPER, and Medaka demonstrated better results than other individual tools, and that the PEPPER followed by Medaka combination showed better results than other combinations. Your benchmark may confirm these findings for your specific data or may reveal that a different combination works better.

The StrainCascade workflow provides an example of an automated modular pipeline that integrates genome assembly, accurate annotation, and comprehensive functional profiling into a single reproducible framework. For laboratories processing many genomes, adopting such a workflow can standardize polishing decisions and reduce the risk of inconsistent quality across samples.

When to Escalate to a Core Facility

The stop-when-measured framework assumes that the polishing tools are functioning correctly and that the input data are adequate. When the framework produces unexpected results, such as no improvement after the first round or a decrease in quality metrics, escalate the problem instead of adding more polishing rounds.

Specific escalation criteria include:

• BUSCO completeness below 90% after one round of Racon and one round of Medaka
• QV below 30 after polishing, indicating substantial residual errors
• A decrease in BUSCO completeness or QV between polishing rounds
• Large discrepancies between the expected and observed genome size, indicating possible assembly artifacts
• Inconsistent results between different polishing tools, suggesting error model mismatches

The EMBL-EBI Training provides learning pathways for bioinformatics data-resource management and practical analysis education. The Bioconductor project offers official package and workflow documentation for reproducible genomic analysis. These resources can help researchers diagnose persistent quality problems before escalating to a core facility.

The Cost-Benefit Calculation for Additional Rounds

The decision to run an additional polishing round should include a computational cost estimate. For a bacterial genome of approximately 5 Mb, one round of Medaka typically completes in minutes on a standard workstation. For a fungal genome of approximately 50 Mb, the same round may take an hour or more. For a eukaryotic genome of 880 Mb, as in the bighead catfish project, each polishing round requires substantial computational resources.

The bighead catfish genome, assembled using high-fidelity long-read sequencing and polished with Illumina short-read sequencing, achieved a QV of 50 and 95.5% BUSCO completeness. This example demonstrates the quality ceiling that hybrid polishing can achieve for a complex eukaryotic genome. For microbial genomes, similar or better quality values are achievable with Nanopore-only polishing when coverage is adequate, as demonstrated by the comparative evaluation.

The cost-benefit calculation is straightforward: if the expected quality improvement from an additional round is below your stopping threshold, the round is not worth the compute time. The stop-when-measured framework makes this calculation explicit and reproducible, allowing you to defend your polishing decisions with recorded data instead of anecdotal experience.

Frequently Asked Questions

How many rounds of Racon should I run before Medaka?

One round of Racon is typically sufficient before Medaka. Racon corrects the most obvious errors in the draft assembly, and Medaka then applies its trained error model to correct the remaining errors. Additional rounds of Racon before Medaka produce diminishing returns and may not improve the final assembly quality. The comparative evaluation of polishing tools found that the first round of polishing produces the largest improvement, with subsequent rounds producing progressively smaller gains.

Can I skip Racon and run Medaka directly on the raw assembly?

Yes, you can run Medaka directly on the raw assembly, but the results may be less accurate than running Racon first. Medaka is designed to correct errors in a draft assembly, and it can work directly on the output of an assembler. However, Racon provides a fast initial correction that can improve the accuracy of the subsequent Medaka polishing. The comparative evaluation found that the PEPPER followed by Medaka combination showed better results than other combinations, suggesting that an initial correction step before Medaka is beneficial.

Is one round of Medaka enough for a bacterial genome?

For most bacterial genomes, one round of Medaka is sufficient to achieve acceptable quality, provided that the base-calling model matches the sequencing conditions and coverage is adequate. The comparative evaluation found that Medaka performed better than other individual tools, and that additional polishing rounds beyond the first produced modest improvements. If the quality metrics after one round of Medaka are below the desired thresholds, investigate the cause before adding more polishing rounds.

When should I use Homopolish instead of Medaka?

Use Homopolish instead of Medaka when the base-calling model is unknown or does not match the Medaka training data, or when the genome has a high proportion of homopolymer regions. The comparative evaluation found that Homopolish performed better than other individual tools and that a second round of Homopolish showed better results than other combinations. Homopolish is particularly effective at correcting homopolymer errors, which are common in Nanopore data.

Do I need Illumina short reads for polishing if I have high coverage Nanopore reads?

For many microbial genomes, high coverage Nanopore reads are sufficient for polishing without Illumina short reads. The comparative evaluation found that high-quality microbial genomes can be achieved using only Nanopore reads without additional short-read data. However, for applications that require the highest possible accuracy, such as reference genome generation or clinical diagnostics, hybrid polishing with short reads remains the gold standard. The bighead catfish genome project used Illumina short-read polishing to achieve a QV of 50.

How do I know when to stop polishing?

Stop polishing when the quality metrics stop improving substantially between rounds. Run BUSCO and QV after each polishing round and compare the results. If the improvement between rounds is small, typically less than 1% in BUSCO completeness or less than 1 in QV, additional polishing rounds are unlikely to produce meaningful gains. The comparative evaluation found that the second round of Homopolish showed better results than other combinations, but the improvement over the first round was modest.

What should I do if polishing makes the assembly worse?

If polishing makes the assembly worse, stop the polishing process and investigate the cause. Possible causes include a mismatched error model, insufficient coverage, or the presence of structural errors in the initial assembly. Check that the polishing tool version and model match the base-calling model used for the reads. Assess the initial assembly quality to determine whether structural errors are present. If the problem persists, consider using a different polishing tool or reassembling with different parameters.

How does genome size affect the number of polishing rounds needed?

Larger genomes require more computational time for polishing, but the number of polishing rounds needed does not scale directly with genome size. The quality improvement from additional polishing rounds depends on the error rate of the initial assembly and the accuracy of the polishing tool, not on the genome size. For large eukaryotic genomes, the computational cost of each polishing round is substantial, so it is particularly important to avoid unnecessary polishing rounds. The bighead catfish genome, at 880 Mb, required substantial computational resources for assembly and polishing, but the polishing strategy was similar to that used for microbial genomes.

Related Bioinformatics Guides

• Evaluating Genome Assembly Quality: Metrics and Tools
• Metagenome Assembled Genome Analysis: From Bins to Biological Insights
• Single-Cell Sequencing Depth: How Much Is Enough?
• RNA-Seq Quality Control: Essential Checks and Tools
• De Novo Genome Assembly with Long Reads: A Practical Workflow

Related Clinical & Scientific Guides

• A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data
• Computational Immunology: Modeling the Immune System
• How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices

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.
• Hybrid Genome Assembly and Annotation of the Basidiomycete Fungus &lt,i&gt,Candolleomyces candolleanus&lt,/i&gt, Strain CMU-8613 Using a Cost-Effective Iterative Pipeline.. 2026.
• Complete genome sequence of &lt,i&gt,Mesorhizobium&lt,/i&gt, sp. strain ORM16 from &lt,i&gt,Ononis repens&lt,/i&gt, root nodules reveals a distinct lineage.. 2025.
• &lt,i&gt,StrainCascade&lt,/i&gt,: An automated, modular workflow for high-throughput long-read bacterial genome reconstruction and characterization.. 2026.
• Chromosome-scale, haplotype-resolved genome of Bighead catfish for conservation and aquaculture.. 2025.
• Comparative evaluation of Nanopore polishing tools for microbial genome assembly and polishing strategies for downstream analysis.. 2021.

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