# Polishing Long-Read Assemblies: How to Use Racon and Medaka to Fix Residual Errors Long-read sequencing platforms produce reads with higher error rates than short-read platforms, and those errors carry into the draft assembly. Polishing is the procedure used to correct errors in a draft assembly by aligning reads back to the contigs and generating a consensus sequence. Racon and Medaka are two widely used long-read polishing tools, and each has distinct strengths and limitations. This article explains how to use both tools, when to choose one over the other, how to combine them with short-read polishers, and how to evaluate whether polishing actually improved your assembly. ## Understanding Residual Errors in Long-Read Assemblies Oxford Nanopore Technologies and PacBio sequencing platforms generate long reads that enable assembly of complete bacterial genomes and large eukaryotic chromosomes. However, the raw reads contain a higher error rate than short-read platforms. A 2024 benchmarking study in BMC Genomics reported that Oxford Nanopore sequencing can reconstruct complete bacterial genomes with 99.95% accuracy, but even small levels of error can obscure phylogenetic relationships between closely related isolates [8]. For applications such as foodborne illness outbreak tracking, that level of accuracy is insufficient. The errors in a draft assembly are not distributed uniformly across the genome. A 2022 study in Briefings in Bioinformatics noted that existing polishing methods treat all regions of the assembly equally, while there are fundamental differences between the error distributions of different regions [7]. Some regions, such as homopolymers and repetitive sequences, are more prone to errors than others. Understanding where errors occur helps you decide which polishing strategy to apply. Residual errors in long-read assemblies are predominantly insertions and deletions, often called indels. The BlockPolish study found that in whole-genome assemblies of NA12878 assembled by Wtdbg2 and Flye using Nanopore data, errors were predominantly indels [7]. These indels are particularly problematic in homopolymer regions, where the sequencing platform struggles to determine the exact number of repeated bases. ## The Polishing Workflow at a Glance Polishing is an iterative process. You align reads to the draft assembly, use the alignment information to identify discrepancies, and generate a corrected consensus sequence. The corrected assembly becomes the input for the next round of polishing. | Polishing Step | Tool | Input Reads | Primary Use | Key Consideration | | --- | --- | --- | --- | --- | | First round long-read polishing | Racon | Raw long reads | Quick correction of draft assembly | Fast but less accurate than Medaka | | Second round long-read polishing | Medaka | Raw long reads | Higher accuracy correction | More accurate but slower than Racon | | Short-read polishing | NextPolish, Pilon, Polypolish, POLCA | Short reads | Final error correction | Needed for near-perfect accuracy | | Complex region polishing | BlockPolish | Raw long reads | Regions with uneven error distribution | Uses deep learning for complex blocks | The order of polishing tools matters. The 2024 benchmarking study found that using less accurate tools after more accurate ones introduced errors [8]. A typical workflow starts with Racon for rapid initial correction, followed by Medaka for higher accuracy, and then a short-read polisher for final correction. ## Racon: Rapid Iterative Polishing Racon is a long-read polishing tool that works by aligning reads to the draft assembly and generating a consensus sequence. It is designed to be fast and memory-efficient, making it suitable for polishing large genomes. ### How Racon Works Racon takes three inputs: the draft assembly in FASTA format, the raw reads in FASTA or FASTQ format, and an overlap or mapping file in PAF format. The mapping file is typically generated by a read aligner such as Minimap2. Racon then performs a partial order alignment of the reads against the assembly and generates a consensus sequence for each contig. The tool is iterative. You can run Racon multiple times, using the output of one round as the input for the next. Each round of polishing can correct additional errors, though the improvement diminishes with each iteration. ### When to Use Racon Racon is best used as a first-pass polisher. It is fast and can quickly reduce the error rate of a draft assembly. The 2021 evaluation of Mycobacterium tuberculosis genome assembly approaches found that for long-read assemblies, Canu followed by Racon and Medaka polishing was the most accurate approach [9]. This suggests that Racon works well as an initial correction step before a more accurate polisher is applied. Racon is also useful when you need to polish many assemblies quickly. The 2024 benchmarking study tested 132 combinations of assembly and polishing tools and found that Racon was less accurate than Medaka but still improved accuracy compared to no polishing [8]. ### Limitations of Racon Racon has several limitations. It is less accurate than Medaka, particularly in homopolymer regions. The 2024 benchmarking study found that Medaka was a more accurate and efficient long-read polisher than Racon [8]. Racon also does not use information about the quality of the reads, which can limit its ability to distinguish true sequence from sequencing errors. ## Medaka: Neural Network Based Polishing Medaka is a long-read polishing tool developed by Oxford Nanopore Technologies. It uses a neural network to predict the consensus sequence from the alignment of reads to the draft assembly. This approach allows Medaka to learn the error patterns of the sequencing platform and correct them more accurately than alignment-based methods. ### How Medaka Works Medaka takes the draft assembly and the raw reads as input. It aligns the reads to the assembly, then uses a neural network model to predict the most likely consensus sequence at each position. The neural network is trained on data from the specific sequencing platform and chemistry used to generate the reads. Medaka requires a model that matches your sequencing platform and chemistry. Oxford Nanopore Technologies provides models for different flow cells and basecalling versions. Using the wrong model can reduce the accuracy of the polishing. ### When to Use Medaka Medaka is best used after an initial round of Racon polishing. The 2024 benchmarking study found that Medaka was a more accurate and efficient long-read polisher than Racon [8]. The 2020 study of Fusarium oxysporum f. sp. lini found that the assembly performed with Canu and polished with Medaka and POLCA was considered the most full and accurate [11]. Medaka is particularly effective at correcting indels in homopolymer regions. The neural network learns the specific error patterns of the sequencing platform and can predict the correct number of repeated bases more accurately than alignment-based methods. ### Limitations of Medaka Medaka is slower than Racon and requires more computational resources. It also requires a model that matches your sequencing platform and chemistry. If you use a model that does not match your data, the polishing accuracy will be reduced. Medaka is designed for Oxford Nanopore data. While it can be used with PacBio data, the neural network models are trained on Nanopore error patterns and may not perform as well on PacBio data. ## Combining Racon and Medaka in a Polishing Pipeline The most effective polishing pipelines combine multiple tools in a specific order. The 2024 benchmarking study found that among the 5 best performing pipelines, polishing with Medaka followed by NextPolish was the most common combination [8]. The study also found that the order of polishing tools mattered, and using less accurate tools after more accurate ones introduced errors. ### Recommended Pipeline for Nanopore Data For Oxford Nanopore data, a recommended pipeline is: 1. Assemble the genome using your preferred assembler, such as Canu, Flye, or wtdbg2 2. Polish with Racon for 1 to 2 rounds 3. Polish with Medaka for 1 round 4. Polish with a short-read polisher such as NextPolish, Pilon, Polypolish, or POLCA This pipeline was validated in the 2024 benchmarking study, which found that near perfect accuracy was only obtained by pipelines that combined both long- and short-read polishing tools [8]. ### Recommended Pipeline for PacBio Data For PacBio HiFi data, the polishing requirements are different. HiFi reads have lower error rates than Nanopore reads, so less polishing is needed. The 2026 study of Gardenia jasminoides genomes used Medaka for ONT-based polishing and Racon for HiFi-based polishing [10]. The final assemblies resolved all 11 chromosomes at telomere-to-telomere continuity with no gaps. For PacBio HiFi data, a single round of Racon polishing may be sufficient. If you have short-read data available, you can add a short-read polishing step for additional accuracy. ### Hybrid Assembly and Polishing Hybrid assembly uses both long reads and short reads to assemble a genome. The long reads provide continuity, while the short reads provide accuracy. The 2020 study of Fusarium oxysporum f. sp. lini used a combination of Oxford Nanopore long reads and Illumina short reads [11]. The assembly was polished with Medaka and POLCA, and the final assembly had 99.5% completeness according to BUSCO. For hybrid assembly, the polishing pipeline typically involves: 1. Assemble with long reads 2. Polish with long reads using Racon and Medaka 3. Polish with short reads using Pilon, POLCA, or NextPolish The short-read polishing step is critical for correcting errors that long-read polishers cannot fix. ## Short-Read Polishing: The Final Accuracy Step Short-read polishing is often necessary to achieve the accuracy required for downstream analyses. The 2024 benchmarking study found that near perfect accuracy, defined as 99.9999% accuracy or approximately 5 nucleotide errors across a 4.8 Mbp genome, was only obtained by pipelines that combined both long- and short-read polishing tools [8]. ### Short-Read Polishing Tools Several short-read polishing tools are available: - NextPolish: The 2024 benchmarking study found that NextPolish showed the highest accuracy among short-read polishers [8] - Pilon: Performed similarly to NextPolish in the benchmarking study [8] - Polypolish: Performed similarly to NextPolish in the benchmarking study [8] - POLCA: Performed similarly to NextPolish in the benchmarking study [8] The choice of short-read polisher is less critical than the choice of long-read polisher. The 2024 benchmarking study found that all four tools performed similarly [8]. ### When Short-Read Polishing Is Necessary Short-read polishing is necessary when you need near-perfect accuracy. This includes applications such as: - Outbreak investigations and source tracking of foodborne illness - Phylogenetic analysis of closely related isolates - Clinical decision making based on genome sequence The 2021 study of Mycobacterium tuberculosis found that short-read assemblies were more accurate than long-read assemblies, having significantly fewer indels and mismatches per 100 kbp [9]. However, hybrid and long-read assemblies had slightly higher genome fractions. ### Limitations of Short-Read Polishing Short-read polishing has limitations. The 2024 benchmarking study found that indels in homopolymers and repetitive regions, where the short reads could not be uniquely mapped, remained the most challenging errors to correct [8]. Short reads cannot be uniquely mapped to repetitive regions, so errors in those regions persist even after short-read polishing. ## Evaluating Polishing Accuracy After polishing, you need to evaluate whether the polishing improved the assembly. Several metrics can be used to assess assembly accuracy. ### BUSCO Completeness BUSCO (Benchmarking Universal Single-Copy Orthologs) assesses the completeness of an assembly by searching for a set of conserved genes. The 2020 study of Fusarium oxysporum f. sp. lini reported 99.5% completeness according to BUSCO after polishing [11]. BUSCO completeness is a useful metric for assessing whether the assembly contains the expected genes. ### Assembly Statistics Assembly statistics such as N50, total length, and number of contigs provide a basic assessment of assembly quality. The 2020 study of Fusarium oxysporum f. sp. lini reported a total length of 59 Mb and N50 of 3.3 Mb after polishing [11]. These statistics can be compared before and after polishing to assess the impact of polishing. ### Error Rate Estimation The error rate of an assembly can be estimated by aligning reads back to the assembly and counting mismatches and indels. The 2021 study of Mycobacterium tuberculosis reported the number of indels and mismatches per 100 kbp [9]. This metric provides a direct measure of assembly accuracy. ### Mapping to a Reference Genome If a reference genome is available, you can align the polished assembly to the reference and count the number of differences. This provides a direct measure of accuracy. The 2024 benchmarking study used this approach to assess the accuracy of 132 combinations of assembly and polishing tools [8]. ## Practical Implementation Steps Implementing a polishing pipeline requires careful attention to data preparation, tool configuration, and quality assessment. ### Step 1: Prepare Your Data Before polishing, ensure that your reads are in the correct format. Most polishing tools accept FASTA or FASTQ format. If your reads are in BAM format, you may need to convert them. Check the quality of your reads. Low-quality reads can introduce errors during polishing. The 2020 study of Fusarium oxysporum f. sp. lini noted that the quality of DNA is crucial for Nanopore sequencing [11]. They developed a protocol for extraction of pure high-molecular-weight DNA from fungi to improve sequencing quality. ### Step 2: Align Reads to the Assembly Racon requires a mapping file in PAF format. You can generate this using Minimap2. Medaka can generate its own alignments, but you need to specify the correct model for your sequencing platform. For short-read polishing, you need to align the short reads to the assembly using a short-read aligner such as BWA or Bowtie2. ### Step 3: Run Racon Run Racon for 1 to 2 rounds. The first round corrects the most obvious errors. The second round can correct additional errors but provides diminishing returns. Command example for Racon: ``` racon reads.fastq mappings.paf assembly.fasta polished_1.fasta ``` ### Step 4: Run Medaka Run Medaka after Racon. Specify the correct model for your sequencing platform and chemistry. Command example for Medaka: ``` medaka_consensus -i reads.fastq -d polished_1.fasta -o medaka_output -m r941_min_high_g360 ``` ### Step 5: Run Short-Read Polishing If you have short-read data, align the short reads to the Medaka-polished assembly and run a short-read polisher such as NextPolish, Pilon, Polypolish, or POLCA. ### Step 6: Evaluate the Polished Assembly Assess the polished assembly using BUSCO, assembly statistics, and error rate estimation. Compare the polished assembly to the draft assembly to quantify the improvement. ## Records and Measurements Keeping detailed records of your polishing runs is essential for reproducibility and troubleshooting. ### Record the Tool Versions Record the versions of all tools used in the polishing pipeline. Tool versions can affect the results, and knowing the exact versions is necessary for reproducing the analysis. ### Record the Parameters Record all parameters used for each tool, including the model used for Medaka and the number of rounds for Racon. Parameter choices can significantly affect the polishing results. ### Record the Input Data Record the sequencing platform, chemistry, and basecalling version for the reads used in polishing. This information is necessary for selecting the correct Medaka model. ### Record the Quality Metrics Record the assembly statistics and quality metrics before and after each polishing step. This allows you to assess the impact of each step and identify any steps that introduced errors. ## Common Failure Patterns Several common failure patterns can occur during polishing. ### Using the Wrong Medaka Model Using the wrong Medaka model can reduce the accuracy of the polishing. Medaka models are specific to the sequencing platform, flow cell, and basecalling version. Check the Medaka documentation to select the correct model for your data. ### Polishing in the Wrong Order The order of polishing tools matters. The 2024 benchmarking study found that using less accurate tools after more accurate ones introduced errors [8]. Always polish with the less accurate tool first and the more accurate tool second. ### Overpolishing Running too many rounds of polishing can introduce errors. Each round of polishing can introduce new errors, particularly in regions where reads cannot be uniquely mapped. The 2024 benchmarking study found that using less accurate tools after more accurate ones introduced errors [8]. ### Ignoring Short-Read Polishing Long-read polishing alone is not sufficient for near-perfect accuracy. The 2024 benchmarking study found that near perfect accuracy was only obtained by pipelines that combined both long- and short-read polishing tools [8]. If you need high accuracy, include a short-read polishing step. ### Failing to Evaluate the Results Polishing can introduce errors as well as correct them. Always evaluate the polished assembly to ensure that the polishing improved the assembly. Compare the assembly statistics and quality metrics before and after polishing. ## Limitations of Polishing Tools Polishing tools have inherent limitations that cannot be overcome by parameter tuning. ### Homopolymer Regions Homopolymer regions remain challenging for all polishing tools. The 2024 benchmarking study found that indels in homopolymers and repetitive regions, where the short reads could not be uniquely mapped, remained the most challenging errors to correct [8]. No polishing tool can completely resolve errors in these regions. ### Repetitive Regions Repetitive regions are difficult to polish because reads cannot be uniquely mapped to these regions. The 2024 benchmarking study found that errors in repetitive regions persisted even after polishing [8]. If your genome has extensive repetitive regions, you may need to use additional approaches such as optical mapping or Hi-C data. ### Uneven Error Distribution The error distribution across the assembly is not uniform. The BlockPolish study noted that existing methods treat all regions of the assembly equally, while there are fundamental differences between the error distributions of different regions [7]. BlockPolish was developed to address this limitation by dividing contigs into blocks with low complexity and high complexity according to statistics of aligned nucleotide bases [7]. ## Advanced Polishing Approaches Several advanced polishing approaches have been developed to address the limitations of Racon and Medaka. ### BlockPolish BlockPolish is a polishing workflow that divides contigs into blocks with low complexity and high complexity according to statistics of aligned nucleotide bases [7]. Multiple sequence alignment is applied to realign raw reads in complex blocks and optimize the alignment result [7]. Two multitask bidirectional Long short-term memory networks are used to predict the consensus sequences [7]. The BlockPolish study found that in whole-genome assemblies of NA12878 assembled by Wtdbg2 and Flye using Nanopore data, BlockPolish had a higher polishing accuracy than other state-of-the-art tools including Racon, Medaka, and MarginPolish and HELEN [7]. BlockPolish also reduced errors in PacBio assemblies [7]. ### MarginPolish and HELEN MarginPolish and HELEN are another pair of polishing tools that use a graph-based approach and a deep neural network, respectively. The BlockPolish study compared BlockPolish to MarginPolish and HELEN and found that BlockPolish had higher polishing accuracy [7]. ## Professional Escalation Criteria Knowing when to escalate a polishing problem to a more experienced colleague or a specialist is important. ### Persistent Errors in Homopolymer Regions If errors persist in homopolymer regions after multiple rounds of polishing, consider whether the sequencing depth is sufficient. Low sequencing depth can limit the ability of polishing tools to correct errors. You may need to sequence additional data or use a different sequencing platform. ### Unexpected Error Patterns If the error pattern after polishing does not match the expected error pattern for your sequencing platform, there may be a problem with the data or the polishing pipeline. Check the quality of the reads and the parameters used for polishing. ### Large Genome with Complex Repetitive Regions If you are working with a large genome with complex repetitive regions, standard polishing tools may not achieve the desired accuracy. Consider using advanced approaches such as BlockPolish or combining polishing with other data types such as Hi-C or optical mapping. ### Clinical or Regulatory Applications If the assembly will be used for clinical or regulatory applications, the accuracy requirements are higher. The 2021 study of Mycobacterium tuberculosis found that anti-TB drug-resistance prediction, particularly with only long-read sequence data, remains challenging, especially for first-line drugs [9]. If the assembly will be used for clinical decision making, consider using a hybrid assembly approach with both long and short reads. ## Building a Polishing Decision Framework Based on Assembly Context and Downstream Requirements Polishing choices are often made by habit or by copying a pipeline from a tutorial, but the optimal polishing strategy depends on the specific context of your assembly project. The same genome sequenced on the same platform can require different polishing approaches depending on what you plan to do with the finished assembly. A phylogenetic study of outbreak isolates demands near-perfect accuracy, while a structural genomics project may tolerate a higher error rate in exchange for faster turnaround. This section provides a practical decision framework that connects your assembly context and downstream requirements to specific polishing choices, along with a record system for tracking polishing decisions and a troubleshooting method for diagnosing persistent errors. ### Defining Your Accuracy Target Before You Polish The first decision in any polishing project is not which tool to run but what level of accuracy you actually need. The 2024 benchmarking study in BMC Genomics defined near-perfect accuracy as 99.9999% accuracy, which corresponds to approximately 5 nucleotide errors across a 4.8 Mbp bacterial genome, excluding low confidence regions [8]. That standard was necessary for source tracking of foodborne illness outbreaks, where even small errors can obscure phylogenetic relationships between closely related isolates [8]. If your project does not require that level of resolution, you can save substantial compute time and avoid the complexity of multi-tool pipelines. For bacterial isolate comparisons where you need to distinguish strains that differ by only a few single nucleotide polymorphisms, you should target the near-perfect accuracy standard. The 2021 evaluation of Mycobacterium tuberculosis genome assembly approaches found that short-read assemblies were more accurate than long-read assemblies, with significantly fewer indels and mismatches per 100 kbp [9]. However, the same study found that hybrid and long-read assemblies had slightly higher genome fractions [9]. This tradeoff between accuracy and completeness is a central consideration in setting your accuracy target. For eukaryotic genome projects where the goal is gene annotation and structural variation analysis, the accuracy requirements are different. The 2026 study of Gardenia jasminoides genomes used multiple polishing iterations with Medaka for ONT-based polishing and Racon for HiFi-based polishing, then achieved telomere-to-telomere continuity with no gaps [10]. The 2020 study of Fusarium oxysporum f. sp. lini reported 99.5% completeness according to BUSCO after polishing with Medaka and POLCA [11]. These projects demonstrate that for many eukaryotic applications, the combination of long-read polishing with a single short-read polishing step is sufficient. ### Matching Polishing Strategy to Sequencing Platform and Read Type Your sequencing platform determines which polishing tools are appropriate and how many rounds of polishing you need. Oxford Nanopore reads have different error profiles than PacBio HiFi reads, and the polishing tools are designed to address platform-specific error patterns. For Oxford Nanopore data, the standard approach is to start with Racon for rapid initial correction, then apply Medaka for higher accuracy. The 2024 benchmarking study found that Medaka was a more accurate and efficient long-read polisher than Racon [8]. The 2021 Mycobacterium tuberculosis study found that for long-read assemblies, Canu followed by Racon and Medaka polishing was the most accurate approach [9]. This two-step long-read polishing strategy is well validated across multiple studies. For PacBio HiFi data, the polishing requirements are different because HiFi reads have lower error rates than Nanopore reads. The 2026 Gardenia jasminoides study used Racon for HiFi-based polishing [10]. A single round of Racon polishing may be sufficient for HiFi assemblies, though additional rounds can be added if evaluation shows persistent errors. The 2022 BlockPolish study noted that existing polishing methods treat all regions of the assembly equally, while there are fundamental differences between the error distributions of different regions [7]. This observation is important for deciding whether standard polishing tools will be sufficient or whether you need a more sophisticated approach. If your assembly has regions with very high or very low complexity, standard tools may not achieve uniform accuracy across the genome. ### Choosing Between Single-Tool and Multi-Tool Pipelines The decision to use a single polishing tool or a multi-tool pipeline depends on your accuracy target and the complexity of your genome. The 2024 benchmarking study tested 132 combinations of assembly and polishing tools and found that long-read polishing alone improved accuracy, but near-perfect accuracy was only obtained by pipelines that combined both long- and short-read polishing tools [8]. For projects that do not require near-perfect accuracy, a single long-read polishing tool may be sufficient. The 2020 Fusarium oxysporum study found that the assembly performed with Canu and polished with Medaka and POLCA was considered the most full and accurate [11]. This suggests that even for fungal genomes, the combination of one long-read polisher and one short-read polisher is a reliable approach. For projects that require near-perfect accuracy, you should plan for a multi-tool pipeline. The 2024 benchmarking study found that among the 5 best performing pipelines, polishing with Medaka followed by NextPolish was the most common combination [8]. The study also found that the order of polishing tools mattered, and using less accurate tools after more accurate ones introduced errors [8]. This finding has direct implications for pipeline design: always apply the less accurate tool first and the more accurate tool second. ### A Practical Decision Matrix for Polishing Strategy Selection The following decision matrix connects your project context to specific polishing recommendations. This matrix is based on the evidence from the approved sources and provides a starting point for designing your polishing pipeline. | Project Context | Recommended Pipeline | Rationale | Evidence Source | | --- | --- | --- | --- | | Bacterial outbreak investigation, Nanopore data only | Racon then Medaka | Long-read polishing improves accuracy but is insufficient for near-perfect accuracy | [8] | | Bacterial outbreak investigation, Nanopore plus short reads | Racon then Medaka then NextPolish | Near-perfect accuracy requires both long- and short-read polishing | [8] | | Bacterial comparative genomics, Nanopore data | Racon then Medaka | Canu followed by Racon and Medaka was the most accurate approach for LR assemblies | [9] | | Fungal genome assembly, Nanopore plus short reads | Medaka then POLCA | Assembly polished with Medaka and POLCA was most full and accurate | [11] | | Plant genome assembly, ONT plus HiFi | Medaka for ONT, Racon for HiFi | Multiple iterations using Medaka and Racon achieved telomere-to-telomere continuity | [10] | | Complex genome with uneven error distribution | Consider BlockPolish | BlockPolish had higher polishing accuracy than Racon and Medaka in some assemblies | [7] | This matrix is not a substitute for evaluating your own assembly. You should always assess the polished assembly against your accuracy target and adjust the pipeline if the target is not met. ### Building a Polishing Record System for Reproducibility Reproducibility in polishing requires more than saving the commands you ran. You need a systematic record system that captures the context of each polishing decision so that you or a collaborator can reconstruct the pipeline months later. The Carpentries lessons emphasize the importance of reproducible data analysis practices, including version control and documentation [6]. The nf-core documentation similarly stresses the importance of reproducible workflow configuration [5]. A practical polishing record should include the following components for each polishing run: **Assembly metadata.** Record the assembler name and version, the input read set, and the assembly statistics before polishing. This information is necessary for understanding what the polishing tools were working with. **Tool versions and parameters.** Record the exact version of Racon, Medaka, and any short-read polisher you use. Record all parameters, including the Medaka model, the number of Racon rounds, and any quality filtering thresholds. The 2024 benchmarking study demonstrated that tool choice and order significantly affect outcomes, so this information is essential for interpreting your results [8]. **Read data provenance.** Record the sequencing platform, chemistry, flow cell type, and basecalling version for the reads used in polishing. This information is necessary for selecting the correct Medaka model and for troubleshooting if polishing does not achieve the expected accuracy. **Quality metrics at each stage.** Record assembly statistics and quality metrics before polishing, after each polishing step, and after the final polishing round. The 2021 Mycobacterium tuberculosis study reported the number of indels and mismatches per 100 kbp as a measure of assembly accuracy [9]. Tracking these metrics at each stage allows you to identify which polishing step contributed the most improvement and whether any step introduced errors. **Decision rationale.** Record why you chose a particular polishing strategy. This is important for future projects where you may face similar decisions. For example, if you chose to include a short-read polishing step because your accuracy target required it, record that reasoning. A simple spreadsheet or text file for each assembly project is sufficient for this purpose. The goal is to have enough information to reproduce the pipeline and to diagnose problems if the polished assembly does not meet your accuracy target. ### Troubleshooting Persistent Errors After Polishing When polishing does not achieve the expected accuracy, the cause is often identifiable through a systematic troubleshooting process. The following method addresses the most common causes of persistent errors. **Step 1: Verify the error locations.** Map the remaining errors to specific genomic regions. The 2024 benchmarking study found that indels in homopolymers and repetitive regions, where the short reads could not be uniquely mapped, remained the most challenging errors to correct [8]. If the persistent errors are concentrated in these regions, standard polishing tools may not be able to fix them regardless of parameter tuning. **Step 2: Check the Medaka model.** Using the wrong Medaka model can reduce polishing accuracy. Verify that the model matches your sequencing platform, flow cell, and basecalling version. The Medaka documentation provides a list of available models, and selecting the correct one is essential for optimal performance. **Step 3: Assess sequencing depth.** Low sequencing depth can limit the ability of polishing tools to correct errors. The 2020 Fusarium oxysporum study noted that the quality of DNA is crucial for Nanopore sequencing and developed a protocol for extraction of pure high-molecular-weight DNA from fungi [11]. If your sequencing depth is below the recommended threshold for your platform, additional sequencing may be necessary. **Step 4: Evaluate the polishing order.** The 2024 benchmarking study found that using less accurate tools after more accurate ones introduced errors [8]. If you applied Medaka before Racon, or if you applied a short-read polisher before a long-read polisher, the polishing order may have introduced errors. Re-run the pipeline with the correct order. **Step 5: Consider the error distribution.** The BlockPolish study noted that existing methods treat all regions of the assembly equally, while there are fundamental differences between the error distributions of different regions [7]. If your assembly has regions with very different complexity levels, standard polishing tools may not achieve uniform accuracy. BlockPolish was developed to address this limitation by dividing contigs into blocks with low complexity and high complexity according to statistics of aligned nucleotide bases [7]. **Step 6: Determine if the accuracy target is achievable with your data.** The 2021 Mycobacterium tuberculosis study found that anti-TB drug-resistance prediction, particularly with only long-read sequence data, remains challenging, especially for first-line drugs [9]. If your accuracy target requires resolution that is not achievable with your current data, you may need to generate additional data or use a different sequencing platform. ### Common Failure Patterns in Polishing Pipelines Several failure patterns recur across polishing projects. Recognizing these patterns can help you diagnose problems quickly. **Pattern 1: Accuracy improves then degrades with additional rounds.** Running too many rounds of polishing can introduce errors. The 2024 benchmarking study found that using less accurate tools after more accurate ones introduced errors [8]. If you observe accuracy degradation after a certain number of rounds, stop polishing at the point of maximum accuracy. **Pattern 2: Errors persist in specific regions despite overall improvement.** This pattern typically indicates errors in homopolymers or repetitive regions. The 2024 benchmarking study found that these regions remained challenging even after polishing with both long- and short-read tools [8]. If your downstream analysis requires accuracy in these regions, you may need to use targeted approaches or additional data types. **Pattern 3: Polishing improves some metrics but worsens others.** For example, polishing may reduce the number of mismatches but increase the number of indels, or vice versa. The 2021 Mycobacterium tuberculosis study reported both indels and mismatches per 100 kbp as separate metrics [9]. Track both metrics separately to identify this pattern. **Pattern 4: Polishing works for one sample but not another.** This pattern often indicates differences in sequencing depth or read quality between samples. The 2020 Fusarium oxysporum study emphasized that DNA quality is crucial for Nanopore sequencing [11]. If one sample has lower DNA quality, polishing may be less effective for that sample. ### Professional Escalation Criteria for Polishing Problems Knowing when to escalate a polishing problem to a more experienced colleague or a specialist can save substantial time and effort. The following criteria indicate situations where standard troubleshooting is unlikely to resolve the problem. **Escalate when errors persist in homopolymer regions after multiple polishing strategies.** The 2024 benchmarking study found that indels in homopolymers and repetitive regions remained the most challenging errors to correct [8]. If you have tried multiple polishing tools and orders and errors persist in these regions, the problem may require a different sequencing approach or additional data types. **Escalate when the error pattern does not match the expected pattern for your sequencing platform.** If the polished assembly has an unexpected distribution of errors, there may be a problem with the sequencing data or the polishing pipeline. A specialist can help identify whether the issue is in the data generation or the analysis. **Escalate when the assembly will be used for clinical or regulatory applications.** The 2021 Mycobacterium tuberculosis study found that anti-TB drug-resistance prediction, particularly with only long-read sequence data, remains challenging, especially for first-line drugs [9]. If your assembly will inform clinical decisions, the accuracy requirements are higher, and you should seek expert review of your polishing pipeline. **Escalate when you need near-perfect accuracy for outbreak investigations.** The 2024 benchmarking study demonstrated that achieving near-perfect accuracy requires careful pipeline design and validation [8]. If your project requires this level of accuracy and your current pipeline is not achieving it, a specialist can help design a validated pipeline. ### Integrating Polishing Decisions with Reproducible Workflow Practices Polishing is rarely a one-time analysis. You may need to re-polish an assembly when new data become available, when you switch to a different assembler, or when you need to meet a higher accuracy standard. Building your polishing pipeline within a reproducible workflow framework makes these updates easier and reduces the risk of errors. The nf-core documentation provides standards for community pipelines that emphasize reproducibility and configuration management [5]. The Galaxy Training Network offers accessible workflow training that can help you build reproducible polishing pipelines [4]. The EMBL-EBI training resources provide learning pathways for bioinformatics data analysis that include practical exercises in genome assembly and polishing [2]. The Bioconductor project provides official documentation for reproducible genomic-analysis workflows [3]. The Carpentries lessons emphasize foundational computing skills, including shell, Git, and programming, that are essential for building reproducible pipelines [6]. These skills are particularly important for polishing because the pipeline involves multiple tools, multiple parameters, and multiple data files. Version control of your polishing scripts and parameters is essential for reproducing your results. The NCBI provides data resources and search systems that can be used to verify the accuracy of your polished assembly against reference sequences [1]. If a reference genome is available for your organism, aligning the polished assembly to the reference provides a direct measure of accuracy. ### Practical Assessment Steps for Your Polishing Pipeline Before you commit to a final polishing pipeline, run a small-scale assessment to verify that the pipeline achieves your accuracy target. The following steps provide a practical approach to pipeline validation. **Step 1: Select a test sample.** Choose one sample from your project that is representative of your data quality and genome complexity. If your project includes multiple samples with different characteristics, select one sample from each category. **Step 2: Run candidate pipelines.** Run two or three candidate polishing pipelines on the test sample. Include at least one pipeline with long-read polishing only and one pipeline with both long- and short-read polishing. The 2024 benchmarking study found that near-perfect accuracy was only obtained by pipelines that combined both long- and short-read polishing tools [8], so include a combined pipeline if you have short-read data available. **Step 3: Evaluate the results.** Assess each polished assembly using the metrics that are relevant to your accuracy target. Count mismatches and indels per 100 kbp as described in the 2021 Mycobacterium tuberculosis study [9]. If a reference genome is available, align the polished assembly to the reference and count the number of differences. The 2024 benchmarking study used this approach to assess the accuracy of 132 combinations of assembly and polishing tools [8]. **Step 4: Select the best pipeline.** Choose the pipeline that achieves your accuracy target with the least computational cost. If multiple pipelines achieve the target, choose the one with the fewest steps to reduce the risk of introducing errors. **Step 5: Document the pipeline.** Record the pipeline steps, tool versions, parameters, and evaluation results in your polishing record system. This documentation is essential for reproducing the pipeline on other samples and for troubleshooting if problems arise. ### Matching Polishing Effort to Genome Complexity Genome complexity is a major factor in determining how much polishing effort is required. The BlockPolish study noted that existing methods treat all regions of the assembly equally, while there are fundamental differences between the error distributions of different regions [7]. This observation has practical implications for polishing strategy. For simple genomes with low repeat content and few homopolymer regions, a single round of Racon followed by Medaka may be sufficient. The 2021 Mycobacterium tuberculosis study found that Canu followed by Racon and Medaka polishing was the most accurate approach for long-read assemblies [9]. This pipeline is relatively simple and computationally efficient. For complex genomes with high repeat content and many homopolymer regions, you may need additional polishing rounds or more sophisticated tools. The 2024 benchmarking study found that indels in homopolymers and repetitive regions remained the most challenging errors to correct [8]. If your genome has extensive repetitive regions, standard polishing tools may not achieve the accuracy you need, and you may need to consider advanced approaches such as BlockPolish [7]. The 2026 Gardenia jasminoides study demonstrated that multiple polishing iterations can achieve telomere-to-telomete continuity in a plant genome [10]. This study used Medaka for ONT-based polishing and Racon for HiFi-based polishing, with multiple iterations of each [10]. The success of this approach suggests that for complex eukaryotic genomes, multiple polishing rounds are often necessary. ### Balancing Computational Cost and Accuracy Polishing is computationally expensive, and the cost increases with each additional tool and round. The 2024 benchmarking study found that Medaka was a more accurate and efficient long-read polisher than Racon [8]. This finding suggests that for some projects, you may be able to skip Racon and go directly to Medaka, saving computational time without sacrificing accuracy. However, the same study found that the order of polishing tools mattered, and using less accurate tools after more accurate ones introduced errors [8]. This finding suggests that if you use Medaka first, you should not follow it with Racon. The recommended order is Racon first, then Medaka, then a short-read polisher if needed. For projects with limited computational resources, you may need to make tradeoffs between accuracy and cost. The decision matrix provided earlier in this section can help you identify the minimum polishing pipeline that achieves your accuracy target. For projects that do not require near-perfect accuracy, a single round of Medaka may be sufficient, saving the computational cost of Racon and short-read polishing. ### Adapting the Decision Framework to New Evidence The polishing field is evolving rapidly, and new tools and approaches are being developed. The BlockPolish study demonstrated that advanced methods can outperform standard tools in some contexts [7]. As new evidence becomes available, you should update your polishing decision framework to incorporate validated improvements. The approved sources for this article provide a foundation for polishing decisions, but they are not the final word. The EMBL-EBI training resources provide learning pathways for bioinformatics data analysis that can help you stay current with new developments [2]. The Galaxy Training Network offers accessible workflow training that includes updates on new tools and approaches [4]. The nf-core documentation provides standards for community pipelines that incorporate validated tools [5]. When evaluating new polishing tools or approaches, apply the same rigor that the 2024 benchmarking study applied to existing tools [8]. Test the new approach on your own data, compare it to your current pipeline, and evaluate the results using the metrics that are relevant to your accuracy target. Only adopt a new approach if it demonstrably improves accuracy or reduces computational cost without sacrificing accuracy. ## Frequently Asked Questions ### What is the difference between Racon and Medaka? Racon uses partial order alignment to generate a consensus sequence from the alignment of reads to the assembly. Medaka uses a neural network to predict the consensus sequence. The 2024 benchmarking study found that Medaka was a more accurate and efficient long-read polisher than Racon [8]. Racon is faster and requires fewer computational resources, making it useful for initial polishing rounds. ### How many rounds of Racon polishing should I run? Typically 1 to 2 rounds of Racon polishing are sufficient. Each round of polishing corrects additional errors, but the improvement diminishes with each iteration. Running too many rounds can introduce new errors. ### Do I need short-read polishing after Racon and Medaka? If you need near-perfect accuracy, yes. The 2024 benchmarking study found that near perfect accuracy was only obtained by pipelines that combined both long- and short-read polishing tools [8]. Long-read polishing alone improved accuracy but was not sufficient for the accuracy needed for source tracking investigations. ### What Medaka model should I use? The Medaka model must match your sequencing platform, flow cell, and basecalling version. Check the Medaka documentation for the list of available models. Using the wrong model can reduce the accuracy of the polishing. ### Can I use Medaka for PacBio data? Medaka is designed for Oxford Nanopore data. The neural network models are trained on Nanopore error patterns and may not perform as well on PacBio data. For PacBio HiFi data, Racon is a more appropriate choice. The 2026 study of Gardenia jasminoides genomes used Medaka for ONT-based polishing and Racon for HiFi-based polishing [10]. ### How do I evaluate whether polishing improved my assembly? Compare the assembly statistics and quality metrics before and after polishing. Use BUSCO to assess completeness, count mismatches and indels per 100 kbp, and align the polished assembly to a reference genome if one is available. The 2021 study of Mycobacterium tuberculosis reported the number of indels and mismatches per 100 kbp as a measure of assembly accuracy [9]. ### What are the most challenging errors to correct? Indels in homopolymers and repetitive regions are the most challenging errors to correct. The 2024 benchmarking study found that these errors persisted even after polishing with both long- and short-read tools [8]. Short reads cannot be uniquely mapped to repetitive regions, so errors in those regions cannot be corrected by short-read polishing. ### What should I do if errors persist after polishing? If errors persist after polishing, consider whether the sequencing depth is sufficient. Low sequencing depth can limit the ability of polishing tools to correct errors. You may need to sequence additional data or use a different sequencing platform. For complex genomes, consider using advanced polishing approaches such as BlockPolish, which was shown to have higher polishing accuracy than Racon and Medaka in some assemblies [7]. ## Related Bioinformatics Guides - [Long-Read Genome Assembly and Polishing Strategies](/knowledge/bioinformatics/long-read-genome-assembly-and-polishing-strategies) - [Long-Read Metagenome Assembly: Overcoming Challenges with Nanopore and PacBio Data](/knowledge/bioinformatics/long-read-metagenome-assembly-overcoming-challenges-with-nanopore-and-pacbio-data) - [Evaluating Metagenomic Assembly Tools: A Benchmarking Framework for Short-Read and Long-Read Data](/knowledge/bioinformatics/evaluating-metagenomic-assembly-tools-a-benchmarking-framework-for-short-read-and-long-read-data) - [Long-Read Sequencing for De Novo Assembly of Complex Genomes: Case Studies and Best Practices](/knowledge/bioinformatics/long-read-sequencing-for-de-novo-assembly-of-complex-genomes-case-studies-and-best-practices) - [Long-Read Sequencing Cost and Market: What to Expect](/knowledge/bioinformatics/long-read-sequencing-cost-and-market-what-to-expect) ## References and Further Reading - [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information. - [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute. - [Bioconductor](https://bioconductor.org/). Bioconductor Project. - [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project. - [nf-core Documentation](https://nf-co.re/docs). nf-core. - [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries. - [BlockPolish: accurate polishing of long-read assembly via block divide-and-conquer.](https://pubmed.ncbi.nlm.nih.gov/34619757). Briefings in bioinformatics, 2022. - [Benchmarking short and long read polishing tools for nanopore assemblies: achieving near-perfect genomes for outbreak isolates.](https://pubmed.ncbi.nlm.nih.gov/38978005). BMC genomics, 2024. - [Evaluation of whole-genome sequence data analysis approaches for short- and long-read sequencing of Mycobacterium tuberculosis.](https://pubmed.ncbi.nlm.nih.gov/34825880). Microbial genomics, 2021. - [Three Near telomere-to-telomere Genomes for Gardenia jasminoides.](https://pubmed.ncbi.nlm.nih.gov/42026098). Scientific data, 2026. - [High-Quality Genome Assembly of Fusarium oxysporum f. sp. lini.](https://pubmed.ncbi.nlm.nih.gov/33193577). Frontiers in genetics, 2020. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.