# Racon vs. Medaka vs. Pilon: A Comparative Guide to Polishing Tools for Long-Read Assemblies ## Key Takeaways - Medaka leverages neural network error models trained on platform-specific sequencing chemistries (e.g., ONT flow cells and basecalling versions) to achieve high accuracy, outperforming alignment-based methods like Racon, though at a higher computational cost. - Racon employs a fast, platform-agnostic partial order alignment strategy, making it suitable for rapid iterative polishing or as a first-pass correction, but it is generally less accurate than Medaka for long-read polishing. - Pilon is a short-read polishing tool that excels at correcting residual errors, particularly in homopolymers and repetitive regions, that long-read polishers may miss, and is typically applied after long-read polishing for maximum accuracy. - Achieving near-perfect genome accuracy (e.g., <5 errors in a bacterial genome) necessitates a multi-stage polishing pipeline, typically involving a high-accuracy long-read polisher like Medaka followed by a short-read polisher such as Pilon or NextPolish. - Polishing effectiveness is influenced by read coverage depth; for corrected ONT data, coverage of 35× or higher may reduce or eliminate the need for additional polishing, including short-read correction. - Polishing choices significantly impact downstream analyses, affecting genome size estimates, gene content (e.g., secondary metabolite gene clusters, CAZyme-encoding genes), and the resolution of phylogenetic relationships crucial for epidemiological investigations. --- Researchers assembling genomes from long-read sequencing data face a persistent problem: raw assemblies contain base-level errors that obscure biological interpretation. Polishing tools correct these errors, but choosing among Racon, Medaka, and Pilon requires understanding their algorithmic differences, input requirements, and error profiles. This article compares these three tools across practical dimensions, drawing on benchmark studies and assembly projects to help researchers select the appropriate polishing strategy for their specific data. ## The Polishing Problem in Long-Read Assembly Long-read sequencing platforms produce reads that span repetitive regions and structural variants better than short reads, but they introduce base-level errors that short-read platforms do not. Oxford Nanopore Technologies (ONT) sequencing generates reads with error rates that, while improving with newer chemistry, still leave assembled genomes with accuracy below what downstream analyses require. A 2024 benchmarking study of Salmonella enterica serovar Newport isolates noted that ONT assemblies typically achieve 99.95% accuracy, which sounds high but translates to thousands of errors across a bacterial genome. These errors obscure phylogenetic relationships between closely related isolates, complicating outbreak source tracking and epidemiological investigations. Assembly polishing is the computational step that corrects these residual errors. Polishing tools compare the assembled contigs against sequencing reads, identify discrepancies, and produce corrected consensus sequences. The choice of polishing tool affects final genome accuracy, computational cost, and the types of errors that remain uncorrected. Three tools dominate current practice: Racon, Medaka, and Pilon. Each implements a different correction strategy, and each has distinct strengths and limitations that researchers must weigh against their specific assembly data and quality goals. The stakes of this choice extend beyond simple base accuracy. The Candolleomyces candolleanus study demonstrated that assembly and polishing strategies substantially affect genome size estimates, with assembled genome sizes ranging from 46.8 Mb to 59.3 Mb depending on the pipeline. Gene content varied correspondingly, with 15 to 25 secondary metabolite gene clusters detected depending on the assembly and prediction tools. CAZyme-encoding genes varied from 494 in the Flye assembly to 453 in the NECAT assembly. These findings illustrate that polishing choices propagate through the entire analysis pipeline, affecting biological conclusions about gene content, metabolic capacity, and evolutionary relationships. ## Algorithmic Principles of Racon, Medaka, and Pilon ### Racon: Alignment-Based Consensus Correction Racon performs polishing through a straightforward alignment-based approach. It takes an assembly, aligns reads back to the contigs, and constructs a consensus sequence from the aligned reads. The tool uses partial order alignment to build a consensus from the read pileup at each position. This approach is platform-agnostic because it does not rely on platform-specific error models. Racon accepts alignments from multiple read types, including ONT, PacBio, and Illumina short reads, making it a flexible first-pass polisher. The practical consequence of Racon's design is speed. It processes assemblies quickly because partial order alignment is computationally efficient. However, the benchmarking study of Salmonella outbreak isolates found that Racon was less accurate than Medaka as a long-read polisher. The study tested 132 combinations of assembly and polishing tools and reported that Medaka outperformed Racon in both accuracy and efficiency. Racon remains useful for rapid iterative polishing, particularly when researchers need a quick improvement before applying more accurate tools. Racon's alignment-based approach has a specific limitation in regions with dense errors. The AIEdit benchmarking study noted that alignment-based methods achieve high accuracy but incur long run times, while alignment-free k-mer-based tools are scalable but struggle in regions with dense errors. For Racon specifically, the partial order alignment approach can be confounded when reads contain clustered errors that prevent clean alignment to the assembly. ### Medaka: Neural Network Error Model Correction Medaka uses a neural network trained on platform-specific error patterns to correct assemblies. The tool learns the characteristic errors of a particular sequencing chemistry and applies this model to predict the correct base at each position. This approach gives Medaka higher accuracy than alignment-based methods because it accounts for systematic biases in the sequencing data. The tradeoff is computational cost. Medaka requires a trained model matched to the sequencing platform and chemistry used to generate the reads. The neural network inference step is more computationally intensive than partial order alignment, leading to longer runtimes. The AIEdit benchmarking study reported that Medaka took 1.5 or more days to polish a human genome assembly, while achieving a Merquary quality score of 32.7. For bacterial genomes, the runtime is substantially shorter, but the relative computational burden compared to Racon remains. Medaka's accuracy advantage was demonstrated in the Salmonella benchmarking study, where it was identified as the most accurate and efficient long-read polisher among those tested. Pipelines that combined Medaka with short-read polishing tools achieved near-perfect accuracy, defined as 99.9999% accuracy or approximately five nucleotide errors across the 4.8 Mbp genome. This level of accuracy was only reached when long-read polishing with Medaka was followed by short-read polishing with tools such as NextPolish. The platform specificity of Medaka is a critical operational consideration. The tool includes models for various ONT flow cells and basecalling versions, and researchers must select the appropriate model for their data. Using a mismatched model degrades accuracy. The AIEdit benchmarking study noted that machine learning-based polishers often perform well only on specific platforms, which is a limitation of the approach. Researchers using novel or uncommon sequencing configurations may not have an appropriate model available. ### Pilon: Short-Read Polishing with Variant Detection Pilon takes a different approach from Racon and Medaka. It is designed primarily for polishing assemblies with short-read data, using aligned Illumina reads to identify and correct errors. Pilon performs local realignment and variant detection, then applies corrections to the assembly. It also reports features such as local misassemblies and gaps, providing diagnostic information beyond simple base correction. Pilon's strength lies in its ability to correct errors that long-read polishers miss, particularly in homopolymers and repetitive regions. The Salmonella benchmarking study found that Pilon performed similarly to NextPolish, Polypolish, and POLCA among short-read polishers, with NextPolish showing the highest accuracy. The study emphasized that short reads are still needed to correct errors in nanopore assemblies to achieve the accuracy required for source tracking investigations. The practical implication is that Pilon is not a direct alternative to Racon or Medaka. It serves a complementary role in a polishing pipeline, applied after long-read polishing to catch residual errors. Researchers who skip short-read polishing may retain errors that obscure fine-scale genomic differences. Pilon's diagnostic output distinguishes it from the other tools. The tool reports local misassemblies and gaps, providing information that can guide assembly improvement beyond simple base correction. This feature is valuable for researchers who need to understand why certain regions remain problematic after polishing. ## Input Requirements and Data Compatibility ### Read Type Requirements Racon accepts alignments from any read type, including ONT, PacBio, and Illumina reads. This flexibility makes it useful for hybrid assembly projects where multiple data types are available. The tool requires an alignment file in SAM or BAM format, which researchers generate using an aligner such as minimap2. Medaka requires reads from the specific platform and chemistry for which its model was trained. The tool includes models for various ONT flow cells and basecalling versions, and researchers must select the appropriate model for their data. Using a mismatched model degrades accuracy. Medaka also requires an alignment of reads to the assembly, generated with minimap2 or a compatible aligner. Pilon requires aligned short reads, typically Illumina paired-end data. The tool accepts BAM files and performs local realignment internally. Pilon does not accept long reads directly, so researchers using only ONT or PacBio data cannot apply Pilon without also generating short-read data. ### Coverage Considerations Coverage depth affects polishing outcomes for all three tools. The Colletotrichum lini study, which assembled telomere-to-telomere genomes from ONT data, found that at genome coverage of 35× or higher with corrected ONT data, additional polishing did not improve accuracy, even with Illumina data. This finding suggests that high-coverage long-read data may reduce or eliminate the need for short-read polishing. For lower coverage data, polishing becomes more important. The Salmonella benchmarking study used ONT data and found that long-read polishing alone improved accuracy but was insufficient for near-perfect genomes. The study's conclusion that short reads are still needed for source tracking accuracy implies that researchers with lower coverage long-read data should plan for short-read polishing. The relationship between coverage and polishing effectiveness has practical implications for project design. Researchers planning genome assembly projects should estimate their expected coverage and determine whether they need to generate short-read data for polishing. Projects with coverage below 35× should budget for short-read sequencing to achieve the accuracy required for downstream analyses. ### Assembly Quality Input The starting assembly quality influences polishing outcomes. Highly fragmented or misassembled contigs may not polish well because reads cannot align uniquely to repetitive regions. The Candolleomyces candolleanus study demonstrated that assembly and polishing strategies substantially affect genome size estimates, with assembled genome sizes ranging from 46.8 Mb to 59.3 Mb depending on the pipeline. This variation highlights that polishing cannot fully compensate for poor assembly choices. Researchers should assess assembly quality before polishing using metrics such as contiguity (N50), completeness (BUSCO), and read mapping rates. Polishing a poor assembly wastes computational resources and may introduce errors if reads align incorrectly. The Colletotrichum lini study used HERRO to correct ONT reads before assembly, which improved assembly quality. Researchers with high error rates in their raw reads may benefit from read correction before assembly instead of relying solely on post-assembly polishing. This pre-assembly correction step can reduce the polishing burden and improve final accuracy. ## Practical Workflow for Polishing Long-Read Assemblies ### Step 1: Assess Assembly Quality Before Polishing Before selecting a polishing tool, evaluate the raw assembly to establish a baseline. Generate standard metrics including assembly size, contig count, N50, and BUSCO completeness scores. The NCBI provides resources for sequence analysis and quality assessment that researchers can use to understand assembly metrics. Record these baseline values because they allow you to measure the improvement from polishing. Check read mapping rates by aligning the reads back to the assembly. Low mapping rates indicate contamination, misassembly, or parameter issues in the assembler. Address these problems before polishing because polishing tools assume that reads map correctly to the assembly. Document the sequencing chemistry and basecalling version for ONT data, as Medaka models are specific to these parameters. Record the flow cell type, basecalling software, and version for every ONT run. This documentation is essential for selecting the correct Medaka model and for reproducibility. ### Step 2: Select the Polishing Strategy Based on Data Availability Determine which read types are available for polishing. If you have only long reads, choose between Racon and Medaka. Medaka provides higher accuracy but requires a model matched to your sequencing chemistry. Racon is faster and platform-agnostic but less accurate. If you have both long and short reads, plan a two-stage polishing pipeline. The Salmonella benchmarking study found that the order of polishing tools matters, and using less accurate tools after more accurate ones introduced errors. The best-performing pipelines in that study used Medaka followed by NextPolish. For researchers using Pilon as the short-read polisher, apply it after long-read polishing with either Racon or Medaka. Consider the accuracy requirements of your downstream analyses. For applications such as outbreak source tracking or clinical genomics, the Salmonella benchmarking study found that near-perfect accuracy required combined long- and short-read polishing. For less demanding applications, long-read polishing alone may suffice. ### Step 3: Generate Read Alignments All three tools require read alignments to the assembly. Use minimap2 or a comparable aligner to map reads to the assembly. For Racon and Medaka, align the long reads. For Pilon, align the short reads. Ensure that alignment parameters match the read type and expected error rates. The Galaxy Training Network provides accessible tutorials on read alignment and assembly workflows that researchers can follow to ensure correct parameter selection. These tutorials cover the practical steps of generating alignments and running polishing tools within reproducible workflows. The Carpentries lessons provide foundational computing and data skills that support reproducible analysis practices. ### Step 4: Run the Polishing Tool Execute the polishing tool with parameters appropriate for your data. For Racon, specify the assembly, the read alignments, and the reads themselves. For Medaka, select the model matching your sequencing chemistry and provide the assembly and aligned reads. For Pilon, provide the assembly, the short-read BAM file, and specify the genome size to guide variant detection. Record the runtime and resource usage for each polishing step. This information helps you estimate costs for future projects and identify whether a tool is practical for your compute environment. The nf-core documentation provides standards for reproducible workflow configuration that researchers can adapt to their polishing pipelines. ### Step 5: Evaluate Polished Assembly Quality After polishing, regenerate the quality metrics you collected before polishing. Compare assembly size, contiguity, and completeness against the baseline. Use read mapping rates to confirm that reads still map correctly to the polished assembly. For a more detailed accuracy assessment, use k-mer-based quality scores such as Merquary QV. The AIEdit benchmarking study used Merquary QV to compare polishing tools, reporting that AIEdit increased QV from 28.7 to 32.9 on ONT data from the NA24385 human genome, while Medaka achieved QV 32.7. These scores provide a quantitative measure of base-level accuracy that complements assembly statistics. ### Step 6: Iterate if Necessary Polishing can be applied iteratively, particularly with Racon, which is often run multiple times to progressively improve accuracy. However, the Salmonella benchmarking study found that using less accurate tools after more accurate ones introduced errors. If you plan multiple polishing rounds, apply the most accurate tool first and use less accurate tools only for initial passes. The Colletotrichum lini study found that at high coverage with corrected ONT data, additional polishing did not improve accuracy. If your coverage exceeds 35× and you used corrected reads, test whether additional polishing provides any benefit before committing computational resources. ## At a Glance: Tool Comparison | Feature | Racon | Medaka | Pilon | |---------|-------|--------|-------| | Algorithm | Partial order alignment | Neural network error model | Local realignment and variant detection | | Input reads | Any read type (ONT, PacBio, Illumina) | Platform-specific long reads (ONT models) | Short reads (Illumina) | | Accuracy | Moderate | High | High for short-read correction | | Runtime | Fast | Slow | Moderate | | Best use | Rapid first-pass polishing | Primary long-read polishing | Short-read polishing after long-read correction | | Model requirements | None | Must match sequencing chemistry | None | | Output | Corrected assembly | Corrected assembly | Corrected assembly with diagnostic reports | ## Comparative Performance Evidence ### Bacterial Genome Benchmarking The Salmonella enterica serovar Newport benchmarking study provides the most direct comparison of Racon, Medaka, and Pilon available in the approved evidence. The study tested 132 combinations of assembly and polishing tools across 15 highly similar isolates from a 2020 onion outbreak. Several findings from this study directly inform tool selection. First, long-read polishing alone improved accuracy but was insufficient for near-perfect genomes. The study defined near-perfect accuracy as 99.9999% accuracy or approximately five nucleotide errors across the 4.8 Mbp genome, excluding low confidence regions. This level of accuracy was only obtained by pipelines that combined both long- and short-read polishing tools. Second, Medaka was a more accurate and efficient long-read polisher than Racon. This finding supports choosing Medaka over Racon when accuracy is the primary concern and the computational cost is acceptable. Third, among short-read polishers, NextPolish showed the highest accuracy, but Pilon, Polypolish, and POLCA performed similarly. This finding indicates that Pilon remains a viable short-read polishing option, though researchers seeking maximum accuracy might consider NextPolish. Fourth, the order of polishing tools mattered. Using less accurate tools after more accurate ones introduced errors. This finding has direct workflow implications: apply the most accurate tools first and avoid adding lower-accuracy polishing steps afterward. Fifth, indels in homopolymers and repetitive regions remained the most challenging errors to correct, particularly where short reads could not be uniquely mapped. This limitation applies to all polishing tools and reflects fundamental constraints of read mapping in repetitive sequence. ### Fungal Genome Assembly Studies The Candolleomyces candolleanus study used a hybrid assembly and annotation pipeline that combined different assembly and polishing tools. The study found that assembled genome size ranged from 46.8 Mb to 59.3 Mb depending on the assembly and polishing strategy. This substantial variation demonstrates that polishing choices affect also base-level accuracy but also genome size estimates, which has downstream consequences for gene annotation and comparative genomics. The study also found that gene content varied across assemblies, with 15 to 25 secondary metabolite gene clusters detected depending on the assembly and prediction tools. CAZyme-encoding genes varied from 494 in the Flye assembly to 453 in the NECAT assembly. These findings illustrate that polishing and assembly choices propagate through the entire analysis pipeline, affecting biological conclusions. The Colletotrichum lini study compared polishing approaches for telomere-to-telomere assemblies generated from ONT R10.4.1 flow cells. The study found that at genome coverage of 35× or higher with corrected ONT data, additional polishing did not improve accuracy, even with Illumina data. This finding suggests that high-coverage, high-quality long-read data may reduce or eliminate the need for polishing, particularly when reads have been corrected with algorithms such as HERRO. The study also demonstrated that despite significantly different genome coverage with ONT data ranging from 25× to 100×, four assemblies of equal contiguity were obtained, with genome sizes of 53.6 to 54.7 Mb, ten core chromosomes, and two or three accessory chromosomes. This result indicates that coverage variation within this range did not affect the ability to generate complete assemblies when appropriate correction and assembly tools were used. ### Human Genome Polishing The AIEdit benchmarking study provides context for polishing performance on human-scale genomes. The study reported that Medaka achieved a Merquary QV of 32.7 on ONT data from the NA24385 human genome but required 1.5 or more days of runtime. This runtime is substantial and may be prohibitive for researchers without access to high-performance computing. The study also reported that alignment-based methods achieve high accuracy but incur long run times, while alignment-free k-mer-based tools are scalable but struggle in regions with dense errors. Machine learning-based polishers often perform well only on specific platforms and require read-to-assembly alignments. These tradeoffs apply to the tools covered in this article: Racon is alignment-based and fast but less accurate, Medaka is machine learning-based and accurate but slow, and Pilon requires short reads that may not be available. The study introduced AIEdit as an alternative approach that operates alignment-free, using spaced seed matching combined with a neural network trained to detect and correct dense error patterns. On simulated human long-read assemblies with high error rates, AIEdit reduced error rates by 58% compared to ntEdit's 21%, completing in 2.7 hours using 230 GB of memory. This performance was faster than POLCA and Medaka, which required multi-day run times, and used three times less memory than JASPER at 689 GB. On experimental ONT data from the NA24385 human genome, AIEdit increased the Merquary QV from 28.7 to 32.9 in 9.5 hours, achieving comparable accuracy to Medaka at QV 32.7 in a fraction of the time. ## Records and Measurements for Polishing Projects ### Metrics to Record Before Polishing Document the following measurements before starting any polishing project. These records establish the baseline against which you will evaluate polishing effectiveness. Assembly size in base pairs. Record the total length of the assembly, as this can change during polishing if the tool inserts or deletes bases. Contig count and N50. These contiguity metrics describe how fragmented the assembly is. Polishing should not substantially change contiguity, but large changes may indicate problems. BUSCO completeness score. This metric estimates how much of the expected gene content is present in the assembly. Record the complete, fragmented, and missing categories separately. Read mapping rate. Align the reads back to the assembly and record the percentage of reads that map. Low mapping rates indicate problems that polishing cannot fix. Coverage depth. Record the average read coverage for both long and short reads. This measurement helps you interpret polishing results and determine whether additional polishing is likely to help. ### Metrics to Record After Polishing Regenerate all baseline metrics after polishing and record the changes. Pay particular attention to the following. Change in assembly size. Large changes may indicate that the polisher introduced or removed bases incorrectly. Change in BUSCO completeness. Polishing should maintain or improve completeness. Decreases suggest that the polisher introduced errors in coding regions. Change in read mapping rate. The polished assembly should have equal or better read mapping rates than the raw assembly. Merquary QV or equivalent quality score. This k-mer-based metric provides a quantitative measure of base-level accuracy. Record the QV before and after polishing to quantify improvement. Runtime and peak memory usage for each polishing tool. These measurements help you plan future projects and compare tool efficiency. ### Record Keeping for Reproducibility Document the exact versions of all software used, including the assembler, aligner, and polishing tools. Record the parameters used for each step, including alignment settings and polishing tool options. The nf-core documentation provides standards for reproducible workflow configuration that researchers can adapt to their polishing pipelines. Record the sequencing chemistry and basecalling version for ONT data, as Medaka models are specific to these parameters. Using the wrong model degrades accuracy, and documenting the model used allows others to reproduce your results. Store all intermediate files, including raw assemblies, read alignments, and polished assemblies. The EMBL-EBI training resources provide guidance on data management practices that support reproducibility. The Bioconductor project offers additional resources for reproducible genomic-analysis workflows that researchers can incorporate into their polishing pipelines. ## Common Failure Patterns in Polishing ### Failure Pattern 1: Using the Wrong Medaka Model Medaka requires a model matched to the sequencing platform and chemistry. Using a model trained on different flow cells or basecalling versions produces suboptimal corrections. Researchers who do not document their sequencing chemistry may struggle to select the correct model. Prevention: Record the flow cell type, basecalling software, and version for every ONT run. Check the Medaka documentation for the model matching your data. If uncertain, test multiple models on a small region and compare accuracy. ### Failure Pattern 2: Polishing Order Errors The Salmonella benchmarking study found that using less accurate tools after more accurate ones introduced errors. Researchers who apply Racon after Medaka, or who add a low-accuracy polishing step at the end of a pipeline, may degrade their assembly. Prevention: Apply the most accurate polishing tools first. Use Racon only as an initial pass before Medaka, or skip Racon entirely if computational resources allow direct Medaka polishing. Do not add additional polishing steps after the most accurate tool unless you verify that they improve accuracy. ### Failure Pattern 3: Skipping Short-Read Polishing Researchers with only long-read data may skip short-read polishing because they lack Illumina reads. The Salmonella benchmarking study found that long-read polishing alone was insufficient for near-perfect genomes. Errors in homopolymers and repetitive regions remained after long-read polishing. Prevention: If your application requires high accuracy, such as outbreak source tracking or clinical genomics, generate short-read data for polishing. If short-read data are unavailable, acknowledge the accuracy limitations in your reporting and interpret results accordingly. ### Failure Pattern 4: Polishing Low-Quality Assemblies Polishing cannot fix misassemblies or contamination. Researchers who polish poor assemblies waste computational resources and may introduce errors if reads align incorrectly to misassembled regions. Prevention: Assess assembly quality before polishing. Check contiguity, completeness, and read mapping rates. If the assembly has obvious problems, return to the assembly step and adjust parameters or try a different assembler before polishing. ### Failure Pattern 5: Ignoring Coverage Thresholds The Colletotrichum lini study found that at coverage of 35× or higher with corrected ONT data, additional polishing did not improve accuracy. Researchers who polish high-coverage assemblies may waste resources without gaining accuracy. Prevention: Estimate coverage before polishing. If coverage exceeds 35× with corrected ONT data, test whether polishing provides any benefit by comparing QV scores before and after a single polishing round. If no improvement occurs, skip additional polishing. ### Failure Pattern 6: Misinterpreting Polishing Results Researchers may interpret polishing as a guarantee of accuracy. The Salmonella benchmarking study found that even the best pipelines left errors in homopolymers and repetitive regions where short reads could not be uniquely mapped. Polishing reduces errors but does not eliminate them. Prevention: Report the accuracy limitations of your polished assembly. Include QV scores and note the types of errors that remain. For applications requiring high accuracy, validate specific genomic regions of interest with additional sequencing or PCR-based approaches. ## Limitations of Polishing Tools ### Error Correction Ceiling All polishing tools have a ceiling on the accuracy they can achieve. The Salmonella benchmarking study found that near-perfect accuracy required combined long- and short-read polishing, and even then, errors remained in homopolymers and repetitive regions. Researchers should not expect polishing to produce error-free assemblies. The Colletotrichum lini study found that at high coverage with corrected ONT data, additional polishing did not improve accuracy. This finding suggests that polishing has diminishing returns and that researchers should focus on generating high-quality reads instead of relying on polishing to compensate for poor data. ### Platform Specificity Medaka's accuracy depends on using a model matched to the sequencing platform and chemistry. Researchers using novel or uncommon sequencing configurations may not have an appropriate model available. The AIEdit benchmarking study noted that machine learning-based polishers often perform well only on specific platforms, which is a limitation of the approach. Racon and Pilon do not have this platform specificity, but they also do not benefit from platform-specific error models. Racon's alignment-based approach is general but less accurate, while Pilon's short-read approach is limited to data types that may not be available. ### Computational Requirements Medaka's neural network inference requires substantial computational resources. The AIEdit benchmarking study reported that Medaka took 1.5 or more days to polish a human genome assembly. This runtime may be prohibitive for researchers without access to high-performance computing. Racon is computationally efficient but less accurate. The choice between Racon and Medaka involves a tradeoff between speed and accuracy that researchers must make based on their specific needs and resources. ### Repetitive Region Limitations All polishing tools struggle in repetitive regions where reads cannot be uniquely mapped. The Salmonella benchmarking study identified indels in homopolymers and repetitive regions as the most challenging errors to correct. This limitation is fundamental to the read mapping approach that all three tools use. Researchers working with genomes that have large repetitive fractions should expect residual errors in these regions and plan validation strategies accordingly. ### Genome Size Estimation Uncertainty The Candolleomyces candolleanus study demonstrated that assembly and polishing strategies can substantially affect genome size estimates. The study found assembled genome sizes ranging from 46.8 Mb to 59.3 Mb depending on the pipeline used. This variation has direct consequences for downstream analyses that depend on accurate genome size estimates, including gene density calculations and comparative genomics. Researchers should be cautious when comparing genome sizes across studies that used different assembly and polishing pipelines. The variation introduced by pipeline choices can obscure true biological differences between species or strains. ## Safety and Regulatory Context for Polishing in Clinical and Diagnostic Applications ### Accuracy Requirements for Clinical Genomics The AIEdit benchmarking study noted that polishing is critical for variant calling, gene annotation, and clinical genomics applications. In clinical contexts, base-level errors can lead to incorrect variant calls with diagnostic consequences. Researchers working with clinical samples must achieve the highest possible accuracy and document the limitations of their polishing approach. The Salmonella benchmarking study emphasized that even small levels of error can obscure phylogenetic relationships between closely related isolates. For outbreak investigations, this obscurity can affect source tracking and public health responses. Researchers generating genomes for regulatory or public health purposes should use the most accurate polishing pipelines available and report accuracy metrics transparently. ### Veterinary Pathogen Detection The review of nanopore sequencing in veterinary pathogen detection noted that nanopore sequencing enables near-complete genome assembly and identification of plasmid-borne antimicrobial resistance genes. For veterinary diagnostics, polishing accuracy affects the reliability of resistance gene identification and outbreak tracing. The review also noted that routine veterinary deployment faces uncertainty in study design, sample preparation, and interpretation thresholds across diverse hosts and sample matrices. Researchers applying polishing tools in veterinary contexts should validate their pipelines for the specific pathogens and sample types they encounter. The review highlighted that nanopore sequencing offers portability, real-time long-read data generation, and minimal infrastructure requirements, enabling rapid on-site sequencing for veterinary diagnostics and surveillance. These advantages are relevant to polishing decisions because field-based sequencing may have different error profiles than laboratory-based sequencing, affecting the choice of polishing tools and models. ### Reporting Requirements When reporting polished assemblies, document the polishing tools, versions, and parameters used. Report accuracy metrics such as QV scores and note the types of errors that remain. The NCBI provides resources for sequence submission and quality assessment that researchers can use to understand reporting requirements for public databases. For clinical or regulatory applications, consult relevant guidelines and standards before finalizing polished assemblies. The Galaxy Training Network and EMBL-EBI training resources provide guidance on reproducible analysis workflows that support regulatory compliance. ## Professional Escalation Criteria ### When to Seek Additional Expertise Researchers should escalate to bioinformatics specialists or core facility staff when they encounter the following situations. Persistent low read mapping rates after polishing. If reads do not map to the polished assembly, the problem likely lies in the assembly itself, not the polishing step. A specialist can diagnose misassembly or contamination issues. Unexpected changes in assembly size or gene content during polishing. Large changes may indicate that the polisher introduced errors. A specialist can evaluate whether the changes are biologically plausible. Poor accuracy despite following recommended polishing protocols. If QV scores remain low after polishing, the problem may lie in read quality, coverage, or assembly parameters. A specialist can help diagnose the root cause. Need for platform-specific models that are not available. If your sequencing chemistry lacks a Medaka model, a specialist may be able to recommend alternative approaches or help train a custom model. Clinical or regulatory applications requiring validated pipelines. If polished assemblies will be used for diagnostic or regulatory decisions, consult specialists with experience in validated bioinformatics pipelines. ### When to Reconsider the Assembly Strategy Polishing cannot compensate for fundamentally poor assembly data. If polishing does not improve accuracy, consider whether the assembly strategy needs revision. The Candolleomyces candolleanus study demonstrated that assembly and polishing strategies substantially affect genome size estimates and gene content. A different assembler or read correction approach may produce better results than additional polishing. The Colletotrichum lini study used HERRO to correct ONT reads before assembly, which improved assembly quality. Researchers with high error rates in their raw reads may benefit from read correction before assembly instead of relying solely on post-assembly polishing. The study also demonstrated that the choice of assembler affects polishing outcomes. Verkko generated genome assemblies of high completeness but low contiguity, while Hifiasm allowed the generation of telomere-to-telomere assemblies. Researchers should consider whether their assembler choice is compatible with their polishing strategy and accuracy goals. ## Frequently Asked Questions ### What is the main difference between Racon and Medaka? Racon uses partial order alignment to build a consensus from aligned reads, making it fast and platform-agnostic. Medaka uses a neural network trained on platform-specific error patterns, making it more accurate but slower and dependent on having a model matched to your sequencing chemistry. The Salmonella benchmarking study found that Medaka was a more accurate and efficient long-read polisher than Racon. ### Can Pilon be used to polish long-read assemblies without short-read data? No. Pilon requires aligned short reads, typically Illumina paired-end data. It performs local realignment and variant detection using short-read alignments. If you have only long-read data, use Racon or Medaka for polishing. If you need the accuracy that short-read polishing provides, generate short-read data for your sample. ### Which polishing tool should I use for bacterial outbreak investigations? The Salmonella benchmarking study found that near-perfect accuracy for outbreak isolates required combined long- and short-read polishing. The most common combination among the best-performing pipelines was Medaka followed by NextPolish. Pilon performed similarly to other short-read polishers but was slightly less accurate than NextPolish. Use Medaka for long-read polishing followed by a short-read polisher for maximum accuracy. ### How many rounds of polishing should I perform? The optimal number of polishing rounds depends on your data and accuracy goals. Racon is often run multiple times iteratively, but the Salmonella benchmarking study found that using less accurate tools after more accurate ones introduced errors. The Colletotrichum lini study found that at coverage of 35× or higher with corrected ONT data, additional polishing did not improve accuracy. Test whether additional polishing rounds improve your QV score before committing computational resources. ### Does polishing improve assembly contiguity? Polishing corrects base-level errors but does not substantially change assembly contiguity. The Candolleomyces candolleanus study found that assembly and polishing strategies affected genome size estimates, but this variation reflected differences in the assembly pipelines instead of polishing alone. If your assembly has poor contiguity, address the assembly step instead of relying on polishing. ### What accuracy can I expect after polishing? The Salmonella benchmarking study found that combined long- and short-read polishing achieved near-perfect accuracy of 99.9999% or approximately five nucleotide errors across a 4.8 Mbp bacterial genome, excluding low confidence regions. The AIEdit benchmarking study reported that Medaka achieved a Merquary QV of 32.7 on human ONT data. Accuracy depends on read quality, coverage, and the polishing pipeline used. ### Why do errors remain after polishing? Errors remain in homopolymers and repetitive regions where reads cannot be uniquely mapped. The Salmonella benchmarking study identified these regions as the most challenging to correct. Short reads may not bridge repetitive elements, and long reads may have systematic errors in homopolymers. These limitations are fundamental to the read mapping approach used by all polishing tools. ### Should I use Racon before Medaka in my polishing pipeline? The Salmonella benchmarking study found that using less accurate tools after more accurate ones introduced errors. If you plan to use Medaka, applying Racon first may not improve the final result and could introduce errors. Use Racon as a standalone polisher when speed is more important than accuracy, or skip it entirely if you have the computational resources for direct Medaka polishing. ## Related Bioinformatics Guides - [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) - [Genomic Data Analysis Tools: A Comparative Guide for Researchers](/knowledge/bioinformatics/genomic-data-analysis-tools-a-comparative-guide-for-researchers) - [Long-Read Metagenome Assembly: Overcoming Challenges with Nanopore and PacBio Data](/knowledge/bioinformatics/long-read-metagenome-assembly-overcoming-challenges-with-nanopore-and-pacbio-data) - [Long-Read Genome Assembly and Polishing Strategies](/knowledge/bioinformatics/long-read-genome-assembly-and-polishing-strategies) - [How to Choose a Long-Read Sequencing Platform: PacBio vs Oxford Nanopore](/knowledge/bioinformatics/how-to-choose-a-long-read-sequencing-platform-pacbio-vs-oxford-nanopore) ## Related Clinical & Scientific Guides * [A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data](/knowledge/bioinformatics/a-practical-guide-to-detecting-antimicrobial-resistance-genes-in-shotgun-metagenomic-data) * [Computational Immunology: Modeling the Immune System](/knowledge/bioinformatics/computational-immunology-modeling-the-immune-system) * [How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices](/knowledge/bioinformatics/how-to-set-hard-filters-for-germline-variant-calling-a-practical-guide-to-gatk-best-practices) ## 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. - [Hybrid Genome Assembly and Annotation of the Basidiomycete Fungus <,i>,Candolleomyces candolleanus<,/i>, Strain CMU-8613 Using a Cost-Effective Iterative Pipeline.](https://doi.org/10.3390/ijms27010509). 2026. - [Nanopore Data-Driven T2T Genome Assemblies of Colletotrichum lini Strains.](https://doi.org/10.3390/jof10120874). 2024. - [Benchmarking short and long read polishing tools for nanopore assemblies: achieving near-perfect genomes for outbreak isolates.](https://doi.org/10.1186/s12864-024-10582-x). 2024. - [AIEdit: Alignment-free genome assembly polisher trained on spaced seed match patterns.](https://doi.org/10.1371/journal.pcbi.1014245). 2026. - [Nanopore Sequencing in Veterinary Pathogen Detection: A Review of Technologies and Applications.](https://doi.org/10.3390/vetsci13030216). 2026. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.