henhaobio
Sign in

GigaScienceBioinformaticsPeer-reviewed journal article

Significantly improving the quality of genome assemblies through curation

Kerstin HoweiD, William ChowiD, Joanna CollinsiD, Sarah E. PelaniD, Damon-Lee B PointoniD, Ying SimsiD, James TorranceiD, Alan TraceyiD, Jonathan M. D. WoodiD

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

Abstract

BackgroundGenome sequence assemblies provide the basis for our understanding of biology. Generating error-free assemblies is therefore the ultimate, but sadly still unachieved goal of a multitude of research projects. Despite the ever-advancing improvements in data generation, assembly algorithms and pipelines, no automated approach has so far reliably generated near error-free genome assemblies for eukaryotes. ResultsWhilst working towards improved data sets and fully automated pipelines, assembly evaluation and curation is actively employed to bridge this shortcoming and significantly reduce the number of assembly errors. In addition to this increase in product value, the insights gained from assembly curation are fed back into the automated assembly strategy and contribute to notable improvements in genome assembly quality. ConclusionsWe describe our tried and tested approach for assembly curation using gEVAL, the genome evaluation browser. We outline the procedures applied to genome curation using gEVAL and also our recommendations for assembly curation in an gEVAL-independent context to facilitate the uptake of genome curation in the wider community.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (13 Aug 2020) · CC BY 4.0

gEVAL-assisted manual curation of 111 VGP and DToL assemblies averaged 221 interventions per Gb, raising mean scaffold N50 by 40% and cutting scaffold counts by 29%.

Automated eukaryotic assembly pipelines still leave false duplications, collapses, misjoins and missed joins that continuity metrics hide and that can be mistaken for biology, while most structural QC tools examine one data type at a time. The work presents gEVAL, a browser showing discordances against many data types at once, the Sanger GRIT curation workflow built around it, and portable curation advice for large-scale genome projects.

Key results

  1. ShownCurators made frequent structural edits across a large, multi-clade set of Vertebrate Genomes Project (VGP) and Darwin Tree of Life (DToL) assemblies. Joins were the most common edit, followed by breaks and removal of false duplications.

    111 assemblies (174 Gb of sequence); on average 221 interventions per Gb, split into 67 breaks, 105 joins and 49 removals of false duplications per Gb. · Impact of assembly curation for high…; Figure legends

  2. ShownOn average, curation improved contiguity without adding new sequence, and most sequence ended up in chromosome-level scaffolds. The effect on N50 depended on how fragmented or over-scaffolded each starting assembly was.

    Mean scaffold N50 rose 40% and scaffold count fell 29%, while total length dropped 2% because no new sequence was generated. N50 rose several hundred-fold in initially fragmented assemblies but was up to halved in over-scaffolded ones. On average 96% of sequence reached chromosome level. · Impact of assembly curation for high…; Figure legends

  3. ShownIn a starfish (Asterias rubens) assembly, Hi-C-guided curation corrected many misjoins and missed joins and placed nearly all sequence into chromosomes.

    75 breaks and 216 joins, plus removal of one erroneously duplicated stretch; 97% of sequence assigned to 22 chromosomes. One scaffold whose order and orientation were ambiguous was submitted as unlocalised for its chromosome. · Figure legends

  4. SuggestedThe Genome Reference Informatics Team (GRIT) pipeline automates data gathering, decontamination and discordance flagging, but leaves edit decisions to experienced curators. The authors report a pace compatible with high-throughput projects.

    Analysis and loading into gEVAL take up to 3 days for a 1 Gb assembly. At around 100 kb resolution, one curator completes 1 Gb in around 3 days. Genome Reference Consortium (GRC) references are curated with no resolution limit. · Assembly curation for high-throughput projects

  5. SuggestedgEVAL's multi-track views can locate an error and trace which pipeline step introduced it. They can also show how a locus improved across successive reference versions.

    In a Taeniopygia guttata (VGP) assembly, a misjoin was flagged by a drop in PacBio coverage, BioNano map discordance and a synteny break, and was traced to the scaff10x step. The zebrafish fbn2b locus was fragmented in Zv9, adjusted in GRCz10 and resolved in GRCz11 using finished clone sequence. · Using gEVAL to assess published assemblies; Figure legends

Methods, data and reproducibility

Assemblies, usually pre-purged, scaffolded and polished, undergo nextflow decontamination, then snakemake/vr-runner pipelines load multi-data-type alignments into gEVAL, an Ensembl-based browser flagging discordances. Curators review these with Hi-C maps where available and edit a version-controlled component path file from which the final assembly is rebuilt.

  • VGP and DToL assemblies from different clades curated by GRIT, built from PacBio CLR, 10X Chromium and Hi-C data, most also with BioNano · multiple species (different clades) · n = 111 assemblies (174 Gb)
  • Starfish assembly from the Sanger 25 genomes for 25 years project, shown in Hi-C maps before and after curation · Asterias rubens (starfish)
  • Bird genome assembly (VGP) used to illustrate a misjoin introduced during scaffolding · Taeniopygia guttata
  • Zebrafish reference assemblies Zv9, GRCz10 and GRCz11, compared at the fbn2b locus · Danio rerio (zebrafish)
  • Publicly available GRC reference assemblies hosted as gEVAL databases with whole-genome alignments to earlier versions
Methods & tools
gEVAL (genome evaluation browser, built on a modified Ensembl framework), nextflow (decontamination pipeline), snakemake, vr-runner, HiGlass, pretext, gap5, Genomicus, purge_dups, BlobToolKit, KAT / KMC / Merqury (k-mer analysis), scaff10x / Access / ASSET / Juicebox (structural QC tools)
Assay conditions & compute
Analysis and loading into gEVAL take up to 3 days for a 1 Gb assembly. Manual curation at around 100 kb resolution takes about 3 days per 1 Gb.
Reproducibility gaps
The authors state that gEVAL and the GRIT pipeline depend on Wellcome Sanger Institute internal infrastructure and cannot be ported. No code repository, software versions or data accessions are given.
Limitations (6)
  • gEVAL and the GRIT curation pipeline depend on Sanger's internal data structures and are not portable. The authors offer general recommendations for curation without gEVAL instead. · stated by the authors
  • Sequence collapses can be resolved only where extra data allow local reassembly; otherwise curation cannot generate new sequence to fix them. · stated by the authors
  • In rare cases, gEVAL and Hi-C maps do not hold enough information to decide on an edit, so curators need further tools such as gap5 or Genomicus. Final decisions still rely on experienced human curators. · stated by the authors
  • To keep throughput up, VGP and DToL curation is generally limited to around 100 kb resolution, unlike GRC reference curation, which has no resolution limit. · stated by the authors
  • Impact is reported only as intervention counts and contiguity metrics. No orthogonal accuracy check (such as held-out maps or k-mer consensus quality) confirms that the curated edits themselves are correct. · noted in this summary
  • Averages pool assemblies with very different starting quality, and N50 changes ranged from several hundred-fold gains to halving. Per-assembly variation is shown only graphically in Figure 3, with no numeric dispersion or per-clade summary in the text. · noted in this summary

Read this if you build or curate eukaryotic reference genomes (for example for VGP, DToL or the Earth Biogenome Project) and want a practical workflow for decontamination, structural error detection and chromosome naming.

Citations
2,123
FWCI
85.20
Field percentile
top 1%

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
THISChow '16Rhie '20Priyam '21Shomorony '16Rhie '21Guan '22
66% semantic matchGenomics
William Chow, Kim Brugger, Mario Caccamo+3 more·bioRxiv·Feb 2, 2016