henhaobio
Sign in

Microbial GenomicsBioinformaticsPeer-reviewed journal article

How low can you go? Short-read polishing of Oxford Nanopore bacterial genome assemblies

George Spyro BourasiD, Louise M. JuddiD, Robert Alan EdwardsiD, Sarah VreugdeiD, Timothy Paul StineariD, Ryan R. WickiD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (10 Mar 2024) · CC BY 4.0

Roughly 25× short-read depth polishes nearly all residual errors in Nanopore-only bacterial assemblies; new '--careful' modes in Pypolca and Polypolish largely avoid introducing low-depth errors.

Q20+ Nanopore-only bacterial assemblies now carry only a few errors, mostly in homopolymers and modified bases, so a short-read polisher's main risk is adding false-positive changes rather than missing true errors. Whether polishing accuracy depends on short-read depth had not been tested for bacteria; this work measures that relationship and adds conservative polishing modes.

Key results

  1. ShownAll polishers except Polypolish-careful raised total errors above the 37-error Nanopore-only baseline at some depth. Most non-conservative tools added very large numbers of errors at low coverage.

    Polypolish-default and Pypolca-careful went above 37 total errors in only 1/500 and 3/500 intervals (peak 42). The other polishers peaked between 1121 errors (Pypolca-default at 3.3×) and 46332 errors (NextPolish at 1.6×). · Results; Figure legends

  2. ShownGenome by genome, Polypolish-careful never made any assembly less accurate at any depth. Pypolca-careful and Polypolish-default did so only occasionally, and only at low-to-moderate depth.

    Pypolca-careful: 45/500 intervals, 51/4500 genomes, never above 16.1×. Polypolish-default: 103/500 intervals, 147/4500 genomes, never above 21.0×. The other tools often worsened at least one genome, even at higher depths (Table 1). · Results; Tables

  3. ShownAbove 25× depth, Pypolca-careful left the fewest residual errors. Its median of 1 remaining error was a hard-to-fix long homopolymer on a Salmonella enterica plasmid. Pilon made assemblies worse on average.

    Pypolca-careful had the lowest mean and median error counts (Table S3). Intervals with every error fixed: Pypolca-default 7, Pypolca-careful 2, NextPolish 1. HyPo never went below 11 remaining errors. Neither Polypolish mode nor NextPolish worsened any assembly above 25×. · Results

  4. ShownMost of the gain from polishing comes below about 25× short-read depth. Non-conservative polishers need much more depth before they stop reducing accuracy.

    For the three best polishers, mean errors in 5× depth bins improved in every bin, with only minor gains past 25× (Table 2). Lowest depth above 2× that gave 37 or fewer errors: 14.9× (Pypolca-default), 17.9× (Pilon), 21.5× (NextPolish), 26.4× (HyPo), 29.1× (FMLRC2). · Results; Discussion; Tables

  5. ShownRunning Pypolca and Polypolish sequentially gave results concordant with single-tool polishing, regardless of order. Hybracter's depth-aware polishing (from v0.7.0) produced error-free assemblies above 25× when the draft had no structural errors.

    Combinations that included Pypolca-careful performed best, and order had no effect (Table S4; Figure S5). Hybracter assemblies were consistently error-free above 25× and sometimes below it (Figure S6). Reference-free ALE scores often did not reflect these gains (Figure S7). · Results; Discussion

Methods, data and reproducibility

Short reads from nine deeply sequenced ATCC isolates were subsampled with seqtk in 0.1× steps from 0.1× to 50× (500 depth intervals per genome). Each subsample was used to polish Trycycler Oxford Nanopore Technologies (ONT)-only assemblies with eight polisher configurations, and remaining errors were counted against curated reference genomes.

  • Nine ATCC bacterial isolates: Salmonella enterica, Vibrio cholerae, V. parahaemolyticus, Listeria ivanovii, L. welshimeri, L. monocytogenes, E. coli, Campylobacter jejuni, C. lari · multiple bacterial species · n = 9 genomes · BioProject PRJNA1042815 (per-sample accessions in Table S1)
  • Illumina NextSeq 2000 150bp paired-end short reads (Illumina DNA Prep, quarter reagents), at least 50× coverage per genome · multiple bacterial species · n = 9 genomes · PRJNA1042815
  • ONT R10.4.1 MinION flow-cell long reads sequenced on a GridION, basecalled with Dorado v0.5.0 (dna_r10.4.1_e8.2_sup@v4.3.0 model) · multiple bacterial species · n = 9 genomes · PRJNA1042815
  • Trycycler v0.5.4 ONT-only assemblies, with errors counted against curated reference genomes · multiple bacterial species · n = 9 assemblies; 37 total errors (0 to 18 per genome)
  • Subsampled short-read sets from 0.1× to 50× estimated genome coverage in 0.1× steps · multiple bacterial species · n = 500 intervals; 4500 genome-interval samples
Methods & tools
Pypolca v0.3.0 (wraps BWA-MEM, Samtools, freebayes), Polypolish v0.6.0, HyPo v1.0.3, FMLRC2 v0.1.8, NextPolish v1.4.1, Pilon v1.24, Trycycler v0.5.4, Dnaapler v0.7.0, fastp v0.23.4, seqtk, Dorado v0.5.0, Hybracter v0.7.0
Software & environment
Pypolca v0.3.0, Polypolish v0.6.0 (Rust), HyPo v1.0.3, FMLRC2 v0.1.8, NextPolish v1.4.1, Pilon v1.24, Trycycler v0.5.4, Dnaapler v0.7.0, fastp v0.23.4, Dorado v0.5.0; Pypolca on Bioconda/PyPI, Polypolish on Bioconda
Limitations (6)
  • The conclusions come from bacterial isolates sequenced with Q20+ Nanopore chemistry. The authors say further studies are needed on how short-read depth affects polishing in extreme-GC genomes, metagenomes and eukaryotic genomes. · stated by the authors
  • Polypolish's '--careful' mode discards reads with multiple alignments. It therefore cannot correct errors inside repeats, and it fixes slightly fewer errors when short-read depth is high. · stated by the authors
  • A high genome-wide depth can hide locally thin coverage, for example from GC extremes, transposase-based library kits or low-abundance metagenomic contigs. In such regions, non-conservative polishers may still introduce errors. · stated by the authors
  • Reference-free ALE scores often did not rise when polishing improved accuracy. This makes ALE of limited use for judging the polishing of near-perfect Nanopore assemblies. · stated by the authors
  • Only 37 baseline errors across nine genomes (0 to 18 per genome) underpin the comparisons of true-error removal. No statistical test is described for the reported variance difference between the Polypolish modes and NextPolish. · noted in this summary
  • The per-genome counts are inconsistent. The Results give 51/4500 for Pypolca-careful and 147/4500 for Polypolish-default, but the Discussion assigns these figures to the two tools the other way round. · noted in this summary

Read this if you polish Nanopore-only bacterial genomes with Illumina reads and need to choose a polisher, a target short-read depth, or a way to avoid polishing-introduced false positives.

Citations
129
FWCI
17.36
Field percentile
top 1%

Abstract

It is now possible to assemble near-perfect bacterial genomes using Oxford Nanopore Technologies (ONT) long reads, but short-read polishing is still required for perfection. However, the effect of short-read depth on polishing performance is not well understood. Here, we introduce Pypolca (with default and careful parameters) and Polypolish v0.6.0 (with a new careful parameter). We then show that: (1) all polishers other than Pypolca-careful, Polypolish-default and Polypolish-careful commonly introduce false-positive errors at low depth; (2) most of the benefit of short-read polishing occurs by 25x depth; (3) Polypolish-careful never introduces false-positive errors at any depth; and (4) Pypolca-careful is the single most effective polisher. Overall, we recommend the following polishing strategies: Polypolish-careful alone when depth is very low (25x). Data SummaryPypolca is open-source and freely available on Bioconda, PyPI, and GitHub (github.com/gbouras13/pypolca). Polypolish is open-source and freely available on Bioconda and GitHub (github.com/rrwick/Polypolish). All code and data required to reproduce analyses and figures are available at github.com/gbouras13/depth_vs_polishing_analysis. All FASTQ sequencing reads are available at BioProject PRJNA1042815. A detailed list of accessions can be found in Table S1.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISWick '22Sereika '21Furqon '26Sanderson '24Sanderson '23Zhang '25
83% semantic matchBioinformatics
Ryan R. Wick, Kathryn E. Holt·PLoS Computational Biology·Jan 24, 2022· 715 citations