henhaobio
Sign in

Genome ResearchBioinformaticsPeer-reviewed journal article

Canu: scalable and accurate long-read assembly via adaptive k-mer weighting and repeat separation

Sergey KoreniD, Brian P. WalenziD, Konstantin BerliniD, Jason Rafe MilleriD, Nicholas H. Bergman, Adam M. PhillippyiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Long-read single-molecule sequencing has revolutionized de novo genome assembly and enabled the automated reconstruction of reference-quality genomes. However, given the relatively high error rates of such technologies, efficient and accurate assembly of large repeats and closely related haplotypes remains challenging. We address these issues with Canu, a successor of Celera Assembler that is specifically designed for noisy single-molecule sequences. Canu introduces support for nanopore sequencing, halves depth-of-coverage requirements, and improves assembly continuity while simultaneously reducing runtime by an order of magnitude on large genomes versus Celera Assembler 8.2. These advances result from new overlapping and assembly algorithms, including an adaptive overlapping strategy based on tf-idf weighted MinHash and a sparse assembly graph construction that avoids collapsing diverged repeats and haplotypes. We demonstrate that Canu can reliably assemble complete microbial genomes and near-complete eukaryotic chromosomes using either PacBio or Oxford Nanopore technologies, and achieves a contig NG50 of greater than 21 Mbp on both human and Drosophila melanogaster PacBio datasets. For assembly structures that cannot be linearly represented, Canu provides graph-based assembly outputs in graphical fragment assembly (GFA) format for analysis or integration with complementary phasing and scaffolding techniques. The combination of such highly resolved assembly graphs with long-range scaffolding information promises the complete and automated assembly of complex genomes.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (4 Jan 2017) · CC BY 4.0

Canu's tf-idf weighted MinHash overlapping and error-filtered sparse overlap graph yield complete Nanopore microbial genomes and contig NG50 above 21 Mbp for PacBio human and Drosophila.

Single-molecule reads are long but noisy (10 to 20% raw error): repeats inflate candidate overlaps, fixed repetitive k-mer cutoffs used by Celera Assembler, Falcon and Miniasm miss some true overlaps, and noise blurs near-identical repeat copies and alleles. Canu replaces the unsupported Celera Assembler with one PacBio and Nanopore pipeline adding probabilistic k-mer down-weighting, data-estimated overlap error cutoffs and Graphical Fragment Assembly (GFA) output.

Key results

  1. ShownOn PacBio eukaryotes Canu was usually the most contiguous assembler and the fastest route to a polished assembly; for human it cut CPU time well below Celera Assembler 8.2 and Falcon, avoided Miniasm's memory failure and raised contig NG50, including on diploid HX1.

    Most contiguous on three of four PacBio eukaryotes; a single Quiver round sufficed vs four for Miniasm. Human: ~20,000 CPU hours vs ~60,000 (Falcon) and >250,000 (Celera Assembler v8.2); Miniasm exceeded 1 TB memory vs <36 GB for Canu. CHM1 P5-C3 NG50 4 Mbp (Celera) to over 7 Mbp; P6-C4 over 20 Mbp; HX1 9.00 Mbp vs 7.61 Mbp (published Falcon). · Architecture; PacBio sequence assembly

  2. ShownFrom Nanopore reads alone Canu equaled or beat every tested assembler in continuity on seven 2D runs and produced far more accurate drafts than Miniasm; with ten correction rounds it also assembled template-only 1D E. coli reads of 70% raw accuracy.

    Drafts: Canu typically >99% identity vs <90% for Minimap/Miniasm; Nanopolish reached up to 99.85% (Miniasm polishing 750% slower). After Pilon, Canu and Falcon, like hybrid SPAdes, routinely exceeded 99.99%; Miniasm stayed below 99.9%. E. coli 1D: ten contigs, NG50 619 kbp, 89% of reference at 85.52% identity vs under 10% for Miniasm; Pilon gave 99.72% identity. · Nanopore sequence assembly; Nanopore 1D sequence assembly

  3. ShownIn a PacBio coverage titration on Arabidopsis thaliana Ler-0, Canu at 20× already out-assembled matched-coverage hybrid SPAdes in continuity and kept improving with depth, whereas hybrid continuity saturated early.

    Subsamples of 10× to 150× PacBio P5-C3: at 20×, Canu was more contiguous than hybrid SPAdes using 20× PacBio plus 100× Illumina and comparable to the best hybrid run with 150× PacBio; SPAdes plateaued after 30× while Canu peaked near 50× (prior advice for hierarchical assembly: at least 50×). Short-read polishing is still advised for low-coverage consensus. · Low-coverage hierarchical assembly; Figure legends

  4. ShownOn Nanopore Bacillus anthracis reads, whose pXO1 plasmid sits at about 6-fold the chromosome's depth, tf-idf weighted MinHash sketches recovered most plasmid overlaps at high precision with no parameter tuning, unlike fixed k-mer count filters.

    Plasmid overlap sensitivity 89% with positive predictive value (PPV) 99.5% and no added runtime or memory, vs 26% for MHAP with a fixed repeat cutoff and 17% (Minimap) or 60% (DALIGNER) at defaults; raised thresholds gave Minimap 94% at lower PPV and DALIGNER 76% with 1.6-fold memory and 2-fold runtime costs. Coverage 487× plasmid vs 76× chromosome. · Adaptive MinHash k-mer weighting

  5. SuggestedBogart's data-driven overlap error cutoff (six median absolute deviations above the median best-overlap error) let Canu's greedy best overlap graph separate less-diverged repeat copies than Falcon or Miniasm in a simulation, without manual tuning.

    Simulated two-copy repeats (0% to 15% divergence, no spanning reads): Canu resolved copies from 3% divergence, Falcon only at 5%, Miniasm only at 13%. Cutoffs are typically below 2% versus the 5% default of the earlier best overlap graph (BOG); on D. melanogaster PacBio, median error 0.25% (MAD 0.15%) yielded a 1.6% cutoff. · Best overlap graph; Graph Construction; Figure legends

Methods, data and reproducibility

Canu chains read correction, overlap-based trimming and assembly: the MinHash Alignment Process (MHAP) finds correction-stage overlaps via tf-idf weighted sketches, and Bogart builds a greedy best overlap graph with a data-estimated error cutoff, splits unsupported repeat joins and outputs GFA; Falcon, Miniasm and hybrid SPAdes served as baselines.

  • Bacillus anthracis Sterne 34F2 Oxford Nanopore MinION reads, used for the pXO1 plasmid overlap test and Nanopore assembly benchmarks · Bacillus anthracis · n = pXO1 at 487× vs chromosome at 76× coverage · BioProject PRJNA357857
  • Yersinia pestis 195/P sequencing data, a repetitive genome in the Nanopore-only and hybrid assembly comparisons · Yersinia pestis · BioProject PRJNA357858
  • Seven recent 2D Oxford Nanopore runs (Table 2), incl. public lab.loman.net releases, with Illumina for Pilon and hybrid SPAdes; E. coli MAP006-1 1D template reads also assembled · microbial, incl. Escherichia coli · n = seven 2D runs
  • Arabidopsis thaliana Ler-0 PacBio P5-C3 reads randomly subsampled for a coverage titration, plus Illumina reads for Illumina-only and hybrid SPAdes · Arabidopsis thaliana · n = 10× to 150× PacBio subsamples; 100× Illumina
  • PacBio bacterial and eukaryotic datasets from external releases such as PacBio DevNet, incl. Caenorhabditis elegans and Drosophila melanogaster · multiple species (incl. C. elegans, D. melanogaster) · n = four eukaryotic genomes compared in Table 1
  • Human PacBio data: effectively haploid CHM1 (P5-C3 and P6-C4 chemistries) and the diploid Chinese genome HX1 at about twice the CHM1 P5-C3 coverage · human
  • Saccharomyces cerevisiae R6 and R7 MinION reads, assembled with iterative correction and Illumina-polished with Pilon · Saccharomyces cerevisiae · n = 20× coverage of 2D reads; average identity 70%
  • Simulated dataset with two repeat copies at known divergence (0% to 15%) and no spanning reads, used to test repeat separation
Methods & tools
Canu v1.3 (MHAP overlapper, Bogart graph module, GFA output), Celera Assembler 8.2, Falcon v0.4.1, Minimap / Miniasm (commits of 2016-03-16), SPAdes v3.7.1 (Illumina-only and hybrid), DALIGNER, Quiver, Nanopolish, Pilon, Racon, MUMmer alignment with GAGE accuracy metrics, Bandage
Software & environment
Canu v1.3 (GPLv2); baselines Falcon v0.4.1, Minimap/Miniasm (2016-03-16 commits), SPAdes v3.7.1; supports SGE, SLURM, LSF and PBS/Torque schedulers or cloud via StarCluster; MHAP defaults idfmax = 3, a = 0.9, 1,500 k-mers per bottom sketch.
Assay conditions & compute
About 20,000 CPU hours per human genome (vs ~60,000 for Falcon, >250,000 for Celera Assembler v8.2); best overlap graph fits in 64 GB for a mammalian genome and Canu used <36 GB on human, where Miniasm exceeded 1 TB.
Reproducibility gaps
Only the B. anthracis and Y. pestis data were deposited at NCBI (PRJNA357857, PRJNA357858); all other datasets come from external sources itemized in Supplementary Note 2, and run settings and per-assembly tables sit in supplementary material.
Limitations (6)
  • Human haplotypes are collapsed because heterozygosity is low, so producing a haplotype-resolved human assembly (as for HX1) would need dedicated phasing. · stated by the authors
  • Alleles diverged beyond the post-correction overlap error rate (typically 1.5% for recent PacBio data) are split into separate contigs, inflating assembly size above the haploid genome, and such bubbles are not yet annotated or phased. · stated by the authors
  • Nanopore-only consensus tops out at 99.9% (QV30) with Nanopolish owing to systematic errors, so finished-quality (>99.99%) Nanopore assemblies needed Illumina polishing with Pilon; low-coverage single-molecule consensus is also limited and short-read polishing is advised. · stated by the authors
  • Repeat arrays too large to span, such as the histone gene cluster (over 500 kbp) on D. melanogaster 2L, stay unresolved, and near-identical repeats inside the error tolerance may still create spurious edges that Bogart must detect and split. · stated by the authors
  • Raw 1D Nanopore reads near 70% identity needed ten iterative correction rounds and gave reduced consensus quality, reaching only 98% identity for E. coli after Nanopolish. · stated by the authors
  • Key comparisons rest on narrow tests: tf-idf sensitivity on one plasmid-bearing B. anthracis dataset, repeat separation on one simulated two-copy repeat, and the HX1 contrast on a previously published Falcon assembly rather than a matched re-run. · noted in this summary

Read this if you assemble PacBio or Oxford Nanopore genomes and want to know how tf-idf weighted MinHash overlapping and error-adaptive best overlap graphs affect runtime, coverage needs and repeat or haplotype separation.

Citations
7,329

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
THISGonzález-Ga '23Nurk '20Nie '24Dierckxsens '25Hu '23Guiglielmon '21
72% semantic matchBioinformatics
Laura Natalia González-García, David Guevara-Barrientos, Daniela Lozano‐Arce+13 more·Life Science Alliance·Feb 22, 2023· 12 citations