# Assembly Polishing for Complex Genomes: How to Correct Errors in Repeats and Heterozygous Regions

Genome assembly polishing corrects base-level errors in a draft assembly using additional sequencing data. For complex genomes with abundant repetitive elements and heterozygous regions, standard polishing approaches can introduce new errors or fail to resolve existing ones. This article provides a practical workflow for researchers working with such genomes, covering tool selection, parameter choices, quality assessment, and common failure patterns.

## Understanding Why Complex Genomes Require Specialized Polishing

Draft assemblies produced from long-read data alone typically contain residual errors in homopolymer runs, tandem repeats, and regions where sequencing coverage is uneven. Short-read polishing tools such as Pilon use Illumina reads to correct these errors, while long-read polishers like Racon and Medaka use the original or additional long-read data. The choice of polishing strategy becomes critical when the genome contains high repeat content or elevated heterozygosity.

The Qiantang River Megalobrama draft assembly illustrates the challenges of complex genomes. This non-model fish genome was estimated to contain 34.7% repeat content and 1.18-1.20% heterozygosity based on k-mer spectra analysis. The Illumina-only draft assembly was fragmented, spanning 1.072 Gb across 519,765 scaffolds. When researchers applied NextPolish refinement, they preserved draft-level BUSCO completeness at 34.1% while improving consensus and read-backed support. This example demonstrates that polishing can improve base-level accuracy without necessarily resolving larger structural issues in the assembly.

The bighead catfish genome provides another relevant case. This 880 Mb genome was assembled using high-fidelity long-read sequencing from Pacific Biosciences and Oxford Nanopore Technologies, scaffolded with Hi-C data, and polished with Illumina short-read sequencing. The final assembly achieved a quality value of 50, which corresponds to approximately one error per 100,000 bases. This result shows that combining long-read assembly with short-read polishing can produce highly accurate genomes when the underlying data quality is sufficient.

The sponge genome assembly of Halisarca dujardinii demonstrates the additional challenges posed by repeat-rich genomes. This 176 Mb genome was assembled using Oxford Nanopore long reads followed by polishing with Illumina short reads. The analysis revealed that more than half of the genome for Demospongiae is occupied by repeats, most of which are evolutionarily young. In such repeat-rich genomes, the distinction between true sequence variation and assembly error becomes difficult to establish, and aggressive polishing can reduce accuracy instead of improve it.

## Core Principles of Polishing Complex Genomes

### Error Types in Draft Assemblies

Polishing addresses several distinct error classes. Base substitutions occur when the assembler places the wrong nucleotide at a position. Insertions and deletions, often called indels, are particularly common in homopolymer runs and simple sequence repeats. Structural errors involve larger misarrangements that polishing tools generally cannot fix because they operate at the base level instead of the scaffold level.

For repetitive regions, the primary challenge is that reads from different copies of a repeat can be collapsed into a single consensus sequence during assembly. This collapse creates a chimeric sequence that does not accurately represent any single genomic location. Polishing tools may then correct this chimeric sequence based on reads from multiple repeat copies, producing a consensus that matches none of them perfectly.

### Heterozygosity and Polishing Decisions

Heterozygous regions present a different problem. In a diploid genome, two homologous chromosomes may differ at many positions. A haploid assembly represents only one version of each locus. When polishing tools align reads from both haplotypes to this haploid reference, they may introduce errors by incorporating variants from the unrepresented haplotype.

The Megalobrama study addressed this issue by projecting polished assemblies onto chromosome-level references and defining a callable core. The analysis showed that callable fractions were nearly identical between two reference choices at 77.24-77.35%, and callable core sequence reached 724.85-727.35 Mb. This concept of a callable core, where confident variant calling is possible, applies directly to polishing decisions. Regions outside the callable core, such as complex repeats or highly divergent haplotypes, may not benefit from standard polishing approaches.

### The Risk of Over-Polishing

Over-polishing occurs when correction tools introduce errors into regions that were initially correct. This problem is especially acute in repetitive regions where read mapping is ambiguous. A read originating from one repeat copy may align equally well to several copies, and the polishing tool may then apply corrections based on the wrong copy.

The Halisarca dujardinii genome provides a cautionary example. With more than half of the genome consisting of evolutionarily young repeats, the assembly required careful polishing to avoid introducing errors in repeat families. The distinction between true sequence variation and assembly error becomes difficult to establish in such genomes, and aggressive polishing can reduce accuracy instead of improve it.

## At a Glance: Polishing Strategy Decision Table

| Genome Characteristic | Recommended Approach | Primary Risk | Quality Control Focus |
|---|---|---|---|
| Low repeat content, low heterozygosity | Single round of short-read polishing with Pilon | Minimal over-polishing risk | BUSCO completeness, QV score |
| High repeat content, low heterozygosity | Long-read polishing with Racon or Medaka, then limited short-read polishing | Repeat collapse and chimeric correction | Read-backed support in repeat regions, k-mer completeness |
| High heterozygosity, any repeat content | Haplotype-aware polishing or careful parameter tuning | Introducing variants from unrepresented haplotype | Callable fraction, variant concordance between haplotypes |
| Very high repeat content, young repeats | Conservative polishing with high mapping quality thresholds | Over-polishing of young repeat families | Repeat-specific error assessment, comparison to unpolished assembly |

## Practical Polishing Workflow

### Step 1: Assess Assembly Quality Before Polishing

Before selecting polishing tools, evaluate the draft assembly to identify problem regions. Run BUSCO to measure gene-space completeness. Calculate contig N50 and scaffold N50 to understand assembly continuity. Generate k-mer spectra to estimate genome size, heterozygosity, and repeat content. These metrics inform whether the assembly is suitable for polishing and which regions may require special attention.

For the Piptanthus nepalensis genome, researchers reported a chromosome-scale assembly spanning approximately 1.04 Gb with a contig N50 of 39.5 Mb and a scaffold N50 of 111.5 Mb. BUSCO analysis indicated 99.3% completeness for the genome and 99.2% for the predicted protein set. Approximately 99.0% of assembled sequences were anchored onto nine pseudo-chromosomes. This level of completeness suggests that polishing was applied to an already high-quality assembly, and the polishing step served to refine base-level accuracy instead of rescue a fragmented assembly.

The Megalobrama study used 670,751,604 filtered reads totaling 100.14 Gb for their analysis. K-mer spectra estimated a 906.7-909.0 Mb haploid genome with 1.18-1.20% heterozygosity and 34.7% repeat content. These estimates provided the context for understanding why the Illumina-only draft was fragmented and guided the polishing strategy.

### Step 2: Select Polishing Tools Based on Data Availability

Racon operates as a standalone consensus tool that uses alignments from long or short reads to correct assemblies. It is fast and memory-efficient, making it suitable for initial polishing rounds. Medaka uses a neural network model trained on specific sequencing platforms and produces high-quality consensus sequences from Oxford Nanopore data. Pilon uses Illumina short reads to correct base errors, fix local misassemblies, and identify gaps.

The choice between these tools depends on available data. If only long reads are available, Racon or Medaka are appropriate. If short reads are available, Pilon can provide complementary correction. Many workflows apply multiple rounds of polishing, alternating between long-read and short-read tools.

The bighead catfish genome project used high-fidelity long-read sequencing from both Pacific Biosciences and Oxford Nanopore Technologies, followed by Illumina short-read polishing. This multi-platform approach leverages the strengths of each data type. Long reads provide continuity across repeats, while short reads offer high per-base accuracy for correction.

The Halisarca dujardinii genome project used Oxford Nanopore long reads for assembly followed by polishing with Illumina short reads. This combination allowed the researchers to leverage the long-read continuity for repeat resolution while using short-read accuracy for base-level correction.

### Step 3: Configure Mapping Parameters for Complex Regions

The success of polishing depends on read mapping quality. For repetitive regions, use mapping tools that report multiple alignments and apply filtering based on mapping quality. Low-quality mappings in repeat regions should be excluded from polishing to avoid introducing errors from incorrectly placed reads.

For heterozygous regions, consider whether to use a haploid or diploid mode. Some polishing tools can account for heterozygous variants, but this capability varies by tool. When heterozygosity is high, as in the Megalobrama genome at 1.18-1.20%, standard haploid polishing may introduce errors at heterozygous sites.

The Megalobrama study demonstrated the importance of careful parameter configuration. The NextPolish refinement preserved draft-level BUSCO completeness at 34.1% while improving consensus and read-backed support. This outcome required appropriate parameter choices that balanced correction power against the risk of introducing errors.

### Step 4: Apply Polishing Iteratively with Quality Checks

Apply polishing in rounds, checking assembly quality after each round. A typical workflow might involve two to three rounds of Racon with long reads, followed by one round of Medaka, then one round of Pilon with short reads. After each round, run BUSCO and calculate QV scores to determine whether polishing improved or degraded the assembly.

The Megalobrama study demonstrated that NextPolish refinement preserved BUSCO completeness while improving consensus and read-backed support. This outcome shows that polishing can improve base-level accuracy without changing the overall assembly structure. If BUSCO completeness drops after polishing, the polishing parameters may be too aggressive for the genome.

The bighead catfish genome achieved a QV of 50 after polishing with Illumina short-read sequencing. This quality value corresponds to approximately one error per 100,000 bases, which is a common target for high-quality assemblies. The achievement of this QV demonstrates that iterative polishing with appropriate tools can produce highly accurate genomes.

### Step 5: Evaluate Polishing Success in Repeat and Heterozygous Regions

Standard quality metrics may not capture errors in repetitive regions. Use repeat-specific assessments to evaluate polishing success. Compare the assembly before and after polishing to identify regions where the sequence changed. Examine read-backed support in repeat regions to determine whether corrections are supported by multiple independent reads.

For heterozygous regions, assess whether polishing introduced variants that are not supported by the underlying data. The callable core concept from the Megalobrama study provides a framework for this assessment. Regions within the callable core can be confidently evaluated, while regions outside it require additional validation.

The Megalobrama study found that reference choice affected anchored span much more than the correction step. Reference A anchored 919.17-919.24 Mb, whereas reference B anchored 878.98-879.06 Mb. Yet 50-kb concordance remained high at 92.47-96.95% one-to-one, callable fractions were nearly identical at 77.24-77.35%, and callable core sequence reached 724.85-727.35 Mb. Gene-interval partitioning further placed most transferred gene-bearing sequence in the callable core.

## Options and Tradeoffs in Polishing Tools

### Racon for Rapid Iterative Correction

Racon accepts alignments in SAM format and produces a corrected consensus sequence. It is designed for speed and can process large genomes in reasonable time. Racon is often used in the first polishing rounds because it quickly fixes the most obvious errors. However, Racon may not fully resolve errors in complex repeats because it relies on the alignments provided to it, and ambiguous alignments in repeats can lead to incorrect corrections.

### Medaka for Platform-Specific Accuracy

Medaka uses neural network models trained on specific sequencing platforms and basecaller versions. This platform-specific training allows Medaka to correct systematic errors characteristic of particular sequencing technologies. For Oxford Nanopore data, Medaka typically produces higher accuracy than Racon. The tradeoff is that Medaka requires the correct model for the specific data type, and using an incorrect model can reduce accuracy.

### Pilon for Short-Read Correction

Pilon aligns Illumina short reads to the assembly and identifies discrepancies between the reads and the assembly. It can correct base errors, fill gaps, and identify local misassemblies. Pilon is particularly effective for fixing errors in homopolymer runs and other regions where long-read technologies have systematic biases. However, Pilon requires sufficient short-read coverage, and low coverage regions may not be corrected.

### NextPolish for Integrated Correction

NextPolish combines multiple correction strategies in a single workflow. It can use both short and long reads and applies iterative correction rounds. The Megalobrama study used NextPolish and found that it preserved BUSCO completeness while improving consensus quality. This integrated approach can be more convenient than running separate tools, but it may offer less control over individual correction steps.

## Observations and Measurements for Polishing Assessment

### BUSCO Completeness as a Primary Metric

BUSCO assesses assembly completeness by searching for a set of single-copy orthologs expected to be present in the target taxonomic group. The Piptanthus nepalensis genome achieved 99.3% BUSCO completeness, while the bighead catfish genome achieved 95.5%. These values indicate that polishing did not compromise gene-space completeness. When BUSCO completeness decreases after polishing, the polishing process has likely introduced errors in genic regions.

The Megalobrama study reported that NextPolish refinement preserved draft-level BUSCO completeness at 34.1%. This relatively low completeness value reflects the fragmented nature of the Illumina-only draft assembly. The polishing step improved consensus and read-backed support without changing the overall completeness, demonstrating that polishing and assembly improvement are separate processes.

### Quality Value Scores for Base-Level Accuracy

The quality value, or QV, estimates the number of errors per 100,000 bases. A QV of 50 corresponds to one error per 100,000 bases, which is a common target for high-quality assemblies. The bighead catfish genome achieved a QV of 50 after polishing with Illumina short reads. QV scores can be calculated using tools that compare the assembly to available reference data or by using k-mer-based methods that assess assembly consistency with the raw reads.

### Read-Backed Support in Problem Regions

For repetitive and heterozygous regions, assess whether polishing corrections are supported by multiple independent reads. A correction supported by a single read may reflect a sequencing error instead of a true assembly error. Require a minimum read depth and mapping quality for corrections in complex regions.

### K-Mer Completeness and Consistency

K-mer-based methods compare the k-mer spectrum of the assembly to that of the raw reads. A polished assembly should contain most k-mers present in the raw reads, and it should not contain excessive k-mers that are absent from the reads. The Megalobrama study used k-mer spectra to estimate genome characteristics, and similar approaches can evaluate polishing success.

## Records and Documentation for Polishing Workflows

### Documenting Tool Versions and Parameters

Record the exact versions of all polishing tools and the parameters used for each round. Tool behavior can change between versions, and parameter choices significantly affect polishing outcomes. This documentation supports reproducibility and allows other researchers to understand the polishing history of the assembly.

### Tracking Changes Between Polishing Rounds

Compare the assembly before and after each polishing round to identify regions where the sequence changed. This comparison reveals which regions are being corrected and whether corrections are consistent across rounds. Large-scale changes between rounds may indicate that the polishing parameters are too aggressive.

### Maintaining Read Alignment Records

Store the read alignments used for polishing, including mapping quality information. These alignments support downstream analysis and allow researchers to evaluate whether specific corrections are well supported. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide official systems for storing and accessing genomic data, including assemblies and associated sequencing reads.

### Using Reproducible Workflow Standards

Community workflow standards support reproducible polishing pipelines. The [nf-core Documentation](https://nf-co.re/docs) describes community pipeline standards, usage, configuration, and reproducible workflow context. Adopting such standards helps ensure that polishing workflows can be reliably repeated and shared across research groups.

## Common Failure Patterns in Polishing Complex Genomes

### Over-Correction in Repetitive Regions

The most common polishing failure in repeat-rich genomes is over-correction. When reads from multiple repeat copies align to a single assembly region, the polishing tool may apply corrections based on a mixture of reads from different copies. The resulting sequence may not match any actual genomic location. This problem is particularly severe for young repeat families, as seen in the Halisarca dujardinii genome where more than half of the genome consists of evolutionarily young repeats.

### Introduction of Haplotype-Specific Variants

In heterozygous genomes, polishing can introduce variants from the unrepresented haplotype. If the assembly represents one haplotype and polishing tools align reads from both haplotypes, the corrected sequence may contain a mixture of both haplotypes. This problem is difficult to detect because the resulting sequence may have high read support from both haplotypes.

### Failure to Correct Errors in Low-Complexity Regions

Low-complexity regions, such as homopolymer runs and simple sequence repeats, are prone to errors in long-read assemblies. Polishing tools may fail to correct these errors if the reads used for correction have similar systematic biases. For example, Oxford Nanopore reads have known errors in homopolymer regions, and polishing with these reads may not fix homopolymer errors.

### Polishing-Induced Fragmentation

Aggressive polishing can break contigs by introducing errors that disrupt the assembly graph. This problem is more common with tools that make large-scale changes instead of base-level corrections. If contig N50 decreases after polishing, the polishing parameters should be adjusted to be more conservative.

## Limitations of Polishing Approaches

### Polishing Cannot Fix Structural Errors

Polishing operates at the base level and cannot correct large-scale structural errors such as misjoins, inversions, or duplications. If the draft assembly has structural errors, polishing may make the sequence more accurate at the base level while preserving the incorrect structure. Detecting structural errors requires additional approaches such as comparing the assembly to a reference genome or using optical mapping data.

### Short-Read Polishing Requires Sufficient Coverage

Pilon and similar short-read polishing tools require adequate short-read coverage to make confident corrections. Low coverage regions may remain uncorrected, and very low coverage regions may be incorrectly corrected based on insufficient data. The Megalobrama study used 670,751,604 filtered reads totaling 100.14 Gb, which provided substantial coverage for polishing.

### Reference Choice Affects Polishing Evaluation

When evaluating polishing success by comparing to a reference genome, the choice of reference affects the results. The Megalobrama study found that reference choice affected anchored span much more than the correction step. Reference A anchored 919.17-919.24 Mb, whereas reference B anchored 878.98-879.06 Mb. This difference of approximately 40 Mb demonstrates that reference-based evaluation can be misleading if the reference is not closely related to the target genome.

### Polishing Cannot Recover Missing Sequence

If the draft assembly is missing sequence due to assembly gaps or collapsed repeats, polishing cannot recover this missing sequence. Polishing only corrects errors in existing sequence. The Megalobrama draft assembly spanned 1.072 Gb across 519,765 scaffolds, but the estimated haploid genome size was 906.7-909.0 Mb. The assembly contained more sequence than the estimated genome size, suggesting that some sequence was duplicated or misassembled instead of missing.

## Quality Control and Safety Context for Polishing

### Establishing Quality Thresholds Before Polishing

Define quality thresholds before starting the polishing workflow. These thresholds should include minimum BUSCO completeness, maximum acceptable error rate, and minimum read support for corrections. Establishing thresholds in advance prevents the common failure of polishing until the assembly appears good without objective criteria.

### Monitoring Computational Resource Usage

Polishing large genomes requires substantial computational resources. Racon is relatively fast, while Medaka and Pilon can be computationally intensive. Monitor memory usage and runtime to ensure that the polishing workflow completes within available resources. The sponge genome of Halisarca dujardinii at 176 Mb is relatively small, but the bighead catfish genome at 880 Mb and the Piptanthus nepalensis genome at 1.04 Gb require more substantial resources.

### Professional Escalation Criteria

Escalate to more specialized approaches when standard polishing fails to improve assembly quality. Indicators that escalation is needed include persistent BUSCO incompleteness, high error rates in repeat regions, or assembly fragmentation that does not improve with additional polishing rounds. Specialized approaches may include haplotype-aware assembly, trio-binning, or using additional data types such as Hi-C or optical mapping.

The Oreobates chiquitanus study demonstrates the value of comprehensive genomic resources for conservation. Researchers sequenced all available type specimens and provided a reference genome from one of the paratopotypes. This approach required careful attention to assembly quality, including polishing, to ensure that the reference genome accurately represented the species.

## Training and Reproducibility Resources

### Foundational Bioinformatics Training

Researchers new to genome assembly polishing should build foundational skills in bioinformatics. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides learning pathways for bioinformatics, data-resource training, and practical analysis education. These resources help researchers understand the underlying concepts of sequence analysis and quality assessment.

### Reproducible Analysis Environments

The [Bioconductor Project](https://bioconductor.org/) provides official packages, workflows, installation documentation, and reproducible genomic-analysis resources. These tools support the statistical analysis and visualization components of assembly quality assessment.

### Accessible Workflow Training

The [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible workflow training, analysis tutorials, and reproducibility context. These resources are particularly useful for researchers who prefer graphical interfaces over command-line tools.

### Foundational Computing Skills

The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational computing, data, shell, Git, and programming training. These skills are essential for managing the computational aspects of genome assembly polishing workflows.

## A Decision Framework for Polishing Rounds Based on Assembly Error Profiles

Standard polishing workflows often prescribe a fixed number of rounds or a generic tool order without considering the specific error profile of the draft assembly. For complex genomes with high repeat content and heterozygosity, this one-size-fits-all approach frequently leads to over-polishing in repeats or under-correction in unique regions. This section provides a practical decision framework that uses measurable assembly characteristics to determine when to polish, which tools to apply, when to stop, and how to document the process for reproducibility.

### Classifying Assembly Error Profiles Before Polishing

Before selecting any polishing tool, classify the draft assembly into one of four error profiles based on measurable characteristics. This classification determines the polishing strategy and the stopping criteria for each round.

**Profile A: Homopolymer and Indel Dominated**

This profile is typical of Oxford Nanopore assemblies where the primary error types are insertions and deletions in homopolymer runs and simple sequence repeats. The assembly may have high base-level accuracy in unique regions but systematic errors in low-complexity sequence. To identify this profile, align a subset of reads back to the assembly and examine error types at positions with low mapping quality. If indels account for more than 70% of observed discrepancies, the assembly fits this profile.

For this profile, use Racon for the first two rounds to correct indel errors quickly, then apply Medaka with the platform-specific model for the final long-read correction. The Megalobrama study demonstrated that NextPolish refinement improved consensus and read-backed support while preserving BUSCO completeness, showing that indel correction can be achieved without disrupting gene space.

**Profile B: Base Substitution Dominated**

This profile is typical of Pacific Biosciences assemblies where residual errors are primarily single nucleotide substitutions instead of indels. The assembly may have good continuity but base-level errors distributed throughout unique regions. To identify this profile, compare the assembly to available short-read data and calculate the substitution-to-indel ratio. If substitutions account for more than 60% of discrepancies, the assembly fits this profile.

For this profile, short-read polishing with Pilon is the primary correction strategy. The bighead catfish genome achieved a QV of 50 after polishing with Illumina short-read sequencing, demonstrating that substitution correction with short reads can produce high-quality assemblies. Apply Pilon in one or two rounds, checking QV scores between rounds to determine whether additional correction is beneficial.

**Profile C: Repeat-Induced Chimeric Errors**

This profile occurs when reads from multiple repeat copies are collapsed into a single consensus during assembly, creating chimeric sequences. The assembly may have high BUSCO completeness but errors concentrated in repeat families. To identify this profile, examine read-backed support in repeat regions. If reads from different repeat copies align to the same assembly position with high mapping quality but disagree on the consensus sequence, the assembly fits this profile.

For this profile, standard polishing tools may make the problem worse by incorporating variants from multiple repeat copies. The Halisarca dujardinii genome, where more than half of the genome consists of evolutionarily young repeats, exemplifies this challenge. Apply conservative polishing with high mapping quality thresholds and consider masking known repeat families before polishing to prevent incorrect corrections.

**Profile D: Heterozygosity-Induced Haplotype Mixture**

This profile occurs in diploid or polyploid genomes where the assembly represents one haplotype but polishing tools incorporate variants from the unrepresented haplotype. The assembly may have high read support for corrected positions because reads from both haplotypes align, but the resulting sequence is a mixture that does not match either haplotype. To identify this profile, examine heterozygous positions in the raw read data and compare them to the polished assembly. If polished positions show a mixture of alleles from both haplotypes, the assembly fits this profile.

The Megalobrama genome had 1.18-1.20% heterozygosity, which is moderate but sufficient to cause haplotype mixture problems. The study addressed this by defining a callable core where confident variant calling is possible. For this profile, consider haplotype-aware polishing approaches or accept that some heterozygous positions may be incorrectly represented.

### Determining Polishing Stopping Points Using Measurable Criteria

The decision to stop polishing should be based on measurable criteria instead of a fixed number of rounds. Define stopping thresholds before beginning the polishing workflow and evaluate them after each round.

**BUSCO Completeness Stability**

Run BUSCO after each polishing round and compare completeness values. If BUSCO completeness decreases by more than 1% between rounds, stop polishing immediately. The Piptanthus nepalensis genome achieved 99.3% BUSCO completeness, and the bighead catfish genome achieved 95.5%. These values represent the expected range for high-quality assemblies. The Megalobrama study preserved draft-level BUSCO completeness at 34.1% through NextPolish refinement, demonstrating that polishing should not reduce completeness even in fragmented assemblies.

**QV Score Improvement Rate**

Calculate QV scores after each round using k-mer-based methods or comparison to available reference data. Track the improvement rate between rounds. If QV improvement between consecutive rounds is less than 1 point, additional polishing rounds are unlikely to provide meaningful benefit. The bighead catfish genome achieved a QV of 50, which corresponds to one error per 100,000 bases. This value represents a practical ceiling for most polishing approaches.

**Read-Backed Support Consistency**

After each polishing round, examine read-backed support in repeat and heterozygous regions. A correction is considered well supported if at least three independent reads support the corrected base with mapping quality above a defined threshold. If corrections in consecutive rounds are not supported by additional reads, the polishing process has likely reached its limit for the available data.

**K-Mer Spectrum Concordance**

Compare the k-mer spectrum of the polished assembly to that of the raw reads. A well-polished assembly should contain most k-mers present in the raw reads and should not contain excessive k-mers absent from the reads. The Megalobrama study used k-mer spectra to estimate genome characteristics, and similar approaches can evaluate polishing success. If k-mer concordance does not improve between rounds, stop polishing.

### Tool Selection Matrix Based on Error Profile and Data Availability

The following matrix guides tool selection based on the error profile classification and available sequencing data.

| Error Profile | Long Reads Only | Short Reads Only | Both Data Types |
|---|---|---|---|
| A: Indel dominated | Racon two rounds, then Medaka | Pilon with caution, may not correct all indels | Racon two rounds, Medaka, then Pilon one round |
| B: Substitution dominated | Racon one round, limited benefit | Pilon two rounds | Racon one round, then Pilon two rounds |
| C: Repeat-induced chimeric | Conservative Racon with high mapping quality | Pilon with repeat masking | Mask repeats, Racon one round, Pilon one round |
| D: Haplotype mixture | Haplotype-aware tools or accept limitations | Pilon with haplotype filtering | Long-read polishing only, avoid short-read mixing |

For Profile C, repeat masking before polishing is critical. Identify known repeat families using tools such as RepeatModeler or RepeatMasker, mask these regions in the assembly, polish the unmasked regions, then restore the original repeat sequence. This approach prevents polishing tools from incorporating variants from multiple repeat copies.

For Profile D, the safest approach is to use long-read polishing only and avoid short-read polishing that may introduce haplotype mixture. The Megalobrama study demonstrated that reference choice affected anchored span much more than the correction step, suggesting that the correction step itself was not the primary source of variation.

### Implementing a Polishing Decision Log

Maintain a structured decision log for each polishing round. This log supports reproducibility and provides evidence for the final assembly quality. The log should include the following fields for each round.

**Round Identifier and Date**

Record the round number and the date the polishing was performed. This information supports version tracking and helps identify when specific corrections were introduced.

**Tool Version and Parameters**

Record the exact version of the polishing tool and all parameters used. Tool behavior can change between versions, and parameter choices significantly affect polishing outcomes. The [nf-core Documentation](https://nf-co.re/docs) describes community pipeline standards that support reproducible workflow configuration.

**Input Assembly Version**

Record which assembly version was used as input for the polishing round. This information is critical for tracking changes between rounds and for reproducing the polishing history.

**Read Data Used**

Record which read dataset was used for polishing, including the sequencing platform, basecaller version, and read filtering parameters. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide official systems for storing and accessing sequencing reads associated with assemblies.

**Quality Metrics Before and After**

Record BUSCO completeness, QV scores, and k-mer concordance before and after the polishing round. These metrics provide objective evidence of whether the polishing round improved or degraded the assembly.

**Regions Changed**

Record the genomic regions where the sequence changed during the polishing round. This information helps identify whether corrections are concentrated in specific repeat families or heterozygous regions.

**Decision for Next Round**

Record whether to continue polishing, change parameters, or stop. This decision should be based on the measurable criteria described above instead of a fixed number of rounds.

### Troubleshooting Common Polishing Failures Using the Decision Framework

**Failure Pattern: BUSCO Completeness Decreases After Polishing**

If BUSCO completeness decreases after a polishing round, the polishing parameters are likely too aggressive for the genome. Check whether the decrease is concentrated in specific gene families or distributed across the genome. If concentrated, examine whether those genes are located in repeat-rich or heterozygous regions. Reduce the number of polishing rounds, increase mapping quality thresholds, or mask repeat regions before polishing.

**Failure Pattern: QV Score Does Not Improve Between Rounds**

If QV scores plateau between rounds, the polishing process has likely reached the limit of what the available data can correct. Check whether the remaining errors are concentrated in specific regions. If errors are in low-complexity regions, additional polishing with the same data will not help. Consider whether additional sequencing data would address the remaining errors.

**Failure Pattern: Polishing Introduces Errors in Repeat Regions**

If polishing introduces errors in repeat regions, the reads used for correction are likely mapping ambiguously. Check mapping quality in repeat regions and increase the mapping quality threshold for corrections. Consider masking known repeat families before polishing. The Halisarca dujardinii genome, with more than half of the genome consisting of evolutionarily young repeats, required careful polishing to avoid introducing errors in repeat families.

**Failure Pattern: Polishing Creates Haplotype Mixture**

If polishing creates a haplotype mixture in heterozygous regions, the polishing tool is incorporating variants from both haplotypes. Check whether the corrected positions show a mixture of alleles from both haplotypes. Consider using haplotype-aware polishing approaches or accept that some heterozygous positions may be incorrectly represented. The Megalobrama study defined a callable core of 724.85-727.35 Mb where confident variant calling is possible, and regions outside this core may not benefit from standard polishing.

**Failure Pattern: Polishing Does Not Correct Known Errors**

If polishing does not correct known errors, the reads used for correction may have the same systematic biases as the assembly. For example, Oxford Nanopore reads have known errors in homopolymer regions, and polishing with these reads may not fix homopolymer errors. Consider using complementary data types for correction. The bighead catfish genome used both Pacific Biosciences and Oxford Nanopore long reads, which can provide complementary error profiles.

### Integrating the Decision Framework with Reproducible Workflow Standards

The decision framework described in this section should be implemented within a reproducible workflow system. The [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible workflow training and analysis tutorials that support reproducible analysis. The [Bioconductor Project](https://bioconductor.org/) provides official packages and workflows for reproducible genomic analysis. The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational computing skills for managing the computational aspects of polishing workflows.

When implementing the decision framework, record all decisions in the polishing decision log. This log provides the evidence needed to justify the final assembly quality and supports future analysis of the assembly. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides learning pathways for bioinformatics that help researchers understand the underlying concepts of sequence analysis and quality assessment.

### Professional Escalation Criteria for Polishing Workflows

Escalate to more specialized approaches when the decision framework indicates that standard polishing cannot achieve the desired quality. Indicators that escalation is needed include persistent BUSCO incompleteness, high error rates in repeat regions that do not improve with conservative parameters, or assembly fragmentation that does not improve with additional polishing rounds.

Specialized approaches may include haplotype-aware assembly, trio-binning, or using additional data types such as Hi-C or optical mapping. The Oreobates chiquitanus study demonstrated the value of comprehensive genomic resources for conservation, requiring careful attention to assembly quality including polishing. The Piptanthus nepalensis genome achieved high completeness with a contig N50 of 39.5 Mb and a scaffold N50 of 111.5 Mb, demonstrating that high-quality assemblies are achievable with appropriate polishing strategies.

When escalating, document the reasons for escalation in the polishing decision log. This documentation supports transparency about the limitations of standard polishing approaches and provides context for the specialized approaches used.

## Frequently Asked Questions

### What is the difference between polishing and scaffolding?

Polishing corrects base-level errors in an assembly using additional sequencing data. Scaffolding orders and orients contigs into larger structures using information from Hi-C, optical mapping, or other long-range data. Polishing does not change the order or orientation of contigs, while scaffolding does not change the base sequence. Both processes are typically applied to a draft assembly, but they address different types of errors.

### How many rounds of polishing should I apply?

The optimal number of polishing rounds depends on the genome and the tools used. Many workflows apply two to three rounds of Racon, followed by one round of Medaka, then one round of Pilon. After each round, assess assembly quality to determine whether additional polishing is beneficial. If BUSCO completeness decreases or QV scores do not improve, additional polishing rounds are unlikely to help.

### Can I polish with long reads only?

Yes, Racon and Medaka can polish assemblies using long reads alone. This approach is necessary when short-read data are unavailable. However, long-read polishing may not correct all error types, particularly if the long reads have systematic biases. The bighead catfish genome used both Pacific Biosciences and Oxford Nanopore long reads, which can provide complementary error profiles.

### Why does polishing sometimes make my assembly worse?

Polishing can introduce errors when reads are incorrectly mapped, particularly in repetitive regions. Over-polishing occurs when corrections are applied based on insufficient or ambiguous evidence. To avoid this problem, use conservative parameters, require high mapping quality, and check assembly quality after each polishing round.

### How do I know if my assembly has too many repeats for standard polishing?

K-mer analysis can estimate repeat content before polishing. The Megalobrama study estimated 34.7% repeat content using k-mer spectra. Genomes with repeat content above approximately 50%, such as the Halisarca dujardinii genome, require careful polishing strategies. If standard polishing produces inconsistent results across repeat copies, the genome may require specialized approaches.

### What is the callable core and why does it matter for polishing?

The callable core is the portion of the genome where confident variant calling is possible. The Megalobrama study defined a callable core of 724.85-727.35 Mb out of approximately 900 Mb total genome size. Regions outside the callable core, such as complex repeats, cannot be confidently evaluated. Polishing decisions should focus on the callable core, and regions outside it may require specialized validation.

### Should I use a reference genome to evaluate polishing success?

Reference-based evaluation can be useful but has limitations. The Megalobrama study found that reference choice affected anchored span by approximately 40 Mb, demonstrating that reference-based metrics can vary substantially. Use reference-based evaluation cautiously and complement it with reference-free metrics such as BUSCO completeness and k-mer consistency.

### How does heterozygosity affect polishing decisions?

High heterozygosity means that the two haplotypes differ at many positions. A haploid assembly represents only one haplotype, and polishing tools may introduce variants from the unrepresented haplotype. The Megalobrama genome had 1.18-1.20% heterozygosity, which is moderate. For genomes with higher heterozygosity, consider haplotype-aware polishing approaches or accept that some heterozygous positions may be incorrectly represented.

## Related Bioinformatics Guides

- [Metagenomic Assembly and Binning: A Practical Workflow for Recovering Genomes from Complex Microbial Communities](/knowledge/bioinformatics/metagenomic-assembly-and-binning-a-practical-workflow-for-recovering-genomes-from-complex-microb)
- [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)
- [Metagenomics Assembly: Strategies for Reconstructing Microbial Genomes](/knowledge/bioinformatics/metagenomics-assembly-strategies-for-reconstructing-microbial-genomes)
- [Evaluating Genome Assembly Quality: Metrics and Tools](/knowledge/bioinformatics/evaluating-genome-assembly-quality-metrics-and-tools)
- [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)

## 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.
- [Dual-reference projection defines a callable core and reference-sensitive shell in a Qiantang River Megalobrama draft assembly.](https://pubmed.ncbi.nlm.nih.gov/42538492). Functional & integrative genomics, 2026.
- [Genomics as a time capsule: insights from Oreobates chiquitanus type specimens.](https://doi.org/10.1186/s12864-026-12984-5). 2026.
- [Chromosome-Level Genome Assembly and Annotation of Piptanthus nepalensis (Hook.) Sweet.](https://doi.org/10.1038/s41597-026-07134-1). 2026.
- [Chromosome-scale, haplotype-resolved genome of Bighead catfish for conservation and aquaculture.](https://doi.org/10.1038/s41597-025-06325-6). 2025.
- [First draft genome assembly of sponge Halisarca dujardinii reveals key components of basement membrane and broad repertoire of aggregation factors.](https://doi.org/10.1038/s41598-025-28962-y). 2025.

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