Zubair Khalid

Virologist/Molecular Biologist | Veterinarian | Bioinformatician

Conventional & Molecular Virology • Vaccine Development • Computational Biology

Dr. Zubair Khalid is a veterinarian and virologist specializing in conventional and molecular virology, vaccine development, and computational biology. Dedicated to advancing animal health through innovative research and multi-omics approaches.

Dr. Zubair Khalid - Veterinarian, Virologist, and Vaccine Development Researcher specializing in Computational Biology, Multi-omics, Animal Health, and Infectious Disease Research

Category: Guides

Genome Assembly Polishing: Reads, Tools, And Validation

Genome assembly polishing is the process of correcting errors in a draft genome sequence by aligning high quality reads to the assembly and using consensus methods to fix base substitutions, insertions, deletions, and small structural mistakes. This guide is for bioinformaticians, genomics researchers, and graduate students who have produced a draft assembly and need to improve its accuracy before downstream analysis.

Polishing is not optional. Even the best long read assemblers produce errors at rates of 0.1,1% per base, and those errors propagate into gene annotations, variant calls, and comparative genomics Galaxy Training Network. A properly polished assembly can approach Q40 quality (99.99% accuracy), which is essential for confident biological interpretation.

At a Glance

Aspect Key Information
Purpose Correct base level and small indel errors in draft assemblies
Input Draft assembly (FASTA) + aligned reads (BAM/CRAM)
Read types Short reads (Illumina) for high precision, long reads (PacBio HiFi, ONT) for structural correction
Main tools Pilon, Racon, Medaka, NextPolish, Polypolish, FreeBayes + bcftools consensus
Validation metrics BUSCO completeness, read mapping rates, k mer spectrum, QV scores
When to polish After initial assembly, before annotation or variant calling

What Is Genome Assembly Polishing?

Polishing uses high confidence reads to identify and fix errors introduced during assembly. The assembler strings reads into contigs but can miscall bases in homopolymer runs, misassemble repetitive regions, or leave gaps. By realigning raw reads back to the contigs and recalculating the most likely base at each position, polishing reduces error rates dramatically.

The underlying principle is statistical. Each aligned read casts a vote for the true nucleotide. A Bayesian or maximum likelihood model combines the votes, accounting for per base quality scores and known error profiles of the sequencing platform. The result is a consensus sequence that reflects the majority evidence NCBI Bookshelf.

Different sequencing technologies require different polishing strategies. Short reads (Illumina) provide very high per base accuracy but are short. Long reads (PacBio, Oxford Nanopore) span repeats but have higher raw error rates. Many modern workflows combine both: a long read assembly followed by short read polishing.

Key Decision Points

Read Type Selection

If you have only short reads, use tools like Pilon or Polypolish. Pilon requires paired end reads aligned with BWA MEM and works best for single base errors and small indels Bioconductor. Polypolish uses a unique alignment free approach but still depends on short reads.

If you have PacBio HiFi reads (which are themselves highly accurate, >99%), run Racon or Medaka. HiFi reads often need only one polishing round. For Oxford Nanopore, Medaka is preferred because it uses a neural network model trained on ONT error patterns EMBL-EBI Training.

If you have both short and long reads, use a hybrid strategy: first polish with long reads, then with short reads. This catches different error types.

Coverage Depth

Polishing effectiveness plateaus around 50x coverage for short reads and 30x for HiFi. Excess coverage (>100x) can introduce bias from PCR duplicates. Always deduplicate or use mark duplicates before polishing. Low coverage (<15x) may not provide enough support to correct errors reliably, and the polished assembly may remain incorrect.

Tool Selection

  • Pilon (Illumina only): fast, widely validated, handles bacterial and eukaryotic genomes.
  • Racon (any read type): lightweight, good for long reads, requires multiple rounds.
  • Medaka (ONT, PacBio): uses deep learning, outputs consensus with quality scores.
  • NextPolish (short or long): designed for eukaryotic genomes, supports multiple iterations.
  • FreeBayes + bcftools consensus (any): flexible but requires more manual parameter tuning NCBI Sequence Read Archive contains countless examples of these pipelines.

Test two tools on a small region or a single chromosome to see which yields better mapping statistics before polishing the whole assembly.

Practical Workflow

Step 1: Prepare Input Files

Ensure your draft assembly is in FASTA format. Index it with samtools faidx. Align the chosen reads to the assembly using an appropriate aligner. For short reads, use BWA MEM with default parameters. For long reads, use minimap2 with the -x option matching your read type (e.g., map hifi, map ont). Sort and index the resulting BAM file.

Step 2: Run the Polishing Tool

For Pilon (short reads):

java -Xmx16G -jar pilon.jar --genome draft.fa --frags aligned.bam --output polished

For Racon (long reads):

racon reads.fastq aligned.sam draft.fa > polished.fa

Run at least two iterations for Racon, three for ONT data. For Medaka:

medaka consensus aligned.bam polished.fa

Always use the recommended parameters from the tool documentation. Do not change mismatch/gap penalties without testing.

Step 3: Evaluate the First Polish Round

Check the number of corrections made. A good polish will change 0.1,1% of bases depending on the input quality. If corrections exceed 2%, the assembly may be poor quality or the reads have systematic errors. Align the polished assembly against a reference genome (if available) or validate with k mer analysis.

Step 4: Iterate if Necessary

Most bacterial assemblies need one round of short read polish. Eukaryotic genomes may need 2,3 rounds. Use the output of round 1 as input for round 2 with the same reads. Stop when the number of changes between rounds drops below 0.01% of the genome.

Step 5: Final Validation

Run BUSCO to check gene completeness. Compute the consensus QV score using Mercury or similar tools. Remap the original reads to the polished assembly and calculate the mismatch rate PubMed ID 42403179 described validation centered pipelines for bacterial genomes.

Validation and Quality Metrics

A polished assembly is not finished until you have multiple lines of evidence for its accuracy.

Read mapping rate: >99.5% of reads should map properly after polishing. A low rate indicates misassembly or contamination.

Consensus QV (Quality Value): Mercury calculates QV based on k mer counts from short reads. A QV of 40 corresponds to one error per 10,000 bases. QV 50 is one per 100,000. QV above 40 is considered high quality PubMed ID 42391505 used this metric for their butterfly reference genome.

BUSCO completeness: >95% complete single copy BUSCOs for most eukaryotes. Bacterial genomes should exceed 98%. Missing BUSCOs can signal collapsed repeats or misassemblies.

GC content and coverage uniformity: Plot GC content across the genome. Sharp drops or spikes may indicate collapsed repeats or assembly errors. Coverage should be fairly uniform (coefficient of variation < 0.5) after polishing.

Comparison to independent data: If you have optical maps, Hi C, or a different sequencing platform, align and check for structural concordance. Many high quality genome papers, like the southern white rhinoceros assembly PubMed ID 42345780, use multiple orthogonal datasets to confirm polishing success.

Common Mistakes

Polishing only with the same data used for assembly. This is circular and can overcorrect. If you have only one read set, use assembly and polishing as separate steps with different random seeds or subsample.

Running too many polishing rounds. After the third round, improvements are negligible and overpolishing can introduce errors by forcing non consensus variants. Stop when the change rate plateaus.

Ignoring read quality. Low quality reads (Q<20) add noise. Filter reads before alignment. For Illumina, use fastp to trim adapters and low quality tails.

Not removing duplicate reads. PCR duplicates inflate coverage artificially and bias the consensus toward one strand. Mark and remove duplicates with Picard or samtools markdup.

Polishing on unmasked repeats. Repetitive regions often have collapsed reads. Polishing can break these regions or introduce chimeric sequences. Mask known repeats before running the tool, or inspect those regions separately.

Using default parameters for unusual genomes. GC rich genomes (e.g., Streptomyces) require different alignment parameters. Adjust mismatch penalties in BWA MEM for genomes >70% GC.

Forgetting to validate with a second tool. A consensus change detected by Pilon but not by FreeBayes warrants manual inspection. Cross validation catches false positives PubMed ID 42338890 demonstrated this in their MRSA assembly.

Limits of Interpretation

Polishing cannot correct large scale misassemblies. If the assembler placed a contig in the wrong order or orientation, base level polishing will not fix it. You need Hi C or optical map data for scaffolding.

Polishing cannot fill gaps. If the assembly has Ns, polishing will not add sequence. It only modifies existing bases. Closing gaps requires additional sequencing or targeted PCR.

Polishing does not guarantee biological correctness. A consensus may represent the most frequent allele in the sample, but if the sample is heterozygous or polyploid, the polished sequence can be a mosaic. Diploid aware polishing exists (e.g., whatshap) but is not yet standard PubMed ID 42323366 noted this challenge in plant genomes.

Polishing with error prone long reads (R10 ONT) still leaves systematic errors in homopolymers. Use a specialized tool like Homopolish or train a Medaka model specific to your run chemistry.

The polished assembly is a computational product. It should be treated as a hypothesis, not ground truth. Validation with orthogonal data (Sanger sequencing, PCR) for key loci remains essential for high stakes applications like clinical genomics.

Frequently Asked Questions

Q: Do I need to polish if I assembled with PacBio HiFi reads?
A: HiFi reads have >99.9% accuracy, but assemblers still introduce errors at repeat boundaries and homopolymers. A single round of Racon or polishing with the same HiFi reads often improves assembly quality by 0.1,0.5%. It is recommended.

Q: Can I polish with the same reads used for assembly?
A: Yes, but be cautious. Use a different subset of reads or align with different parameters to avoid overfitting. Ideally, polish with a separate read type (e.g., add Illumina reads to a PacBio assembly).

Q: How many polishing rounds are too many?
A: Beyond 3,4 rounds, the change rate typically drops below 0.001% per round. More rounds risk introducing systematic errors from read bias. Stop when improvements plateau.

Q: What is the best tool for a 1 Gb plant genome?
A: NextPolish is designed for large eukaryotic genomes and supports multiprocessor runs. Pilon works but is memory intensive for >500 Mb. Always test on a scaffold first.

References and Further Reading

Related Articles