henhaobio
Sign in

Genome biologyBioinformaticsPeer-reviewed journal article

Merqury: reference-free quality, completeness, and phasing assessment for genome assemblies

Arang RhieiD, Brian P. WalenziD, Sergey KoreniD, Adam M. PhillippyiD

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

Abstract

Recent long-read assemblies often exceed the quality and completeness of available reference genomes, making validation challenging. Here we present Merqury, a novel tool for reference-free assembly evaluation based on efficient k-mer set operations. By comparing k-mers in a de novo assembly to those found in unassembled high-accuracy reads, Merqury estimates base-level accuracy and completeness. For trios, Merqury can also evaluate haplotype-specific accuracy, completeness, phase block continuity, and switch errors. Multiple visualizations, such as k-mer spectrum plots, can be generated for evaluation. We demonstrate on both human and plant genomes that Merqury is a fast and robust method for assembly validation. Availability of data and materialProject name: Merqury Project home page: https://github.com/marbl/merqury, https://github.com/marbl/meryl Archived version: https://github.com/marbl/merqury/releases/tag/v1.0 Operating system(s): Platform independent Programming language: C++, Java, Perl Other requirements: gcc 4.8 or higher, java 1.6 or higher License: Public domain (see https://github.com/marbl/merqury/blob/master/README.license) Any restrictions to use by non-academics: No restrictions applied

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (31 Mar 2020) · CC0 1.0

Merqury estimates assembly consensus quality, k-mer completeness, and trio-based phasing from read k-mers alone, avoiding reference alignment; validated on Arabidopsis F1 hybrid and human NA12878 assemblies.

De novo diploid assemblies have no ground truth to check against: short-read mapping misses repetitive or low-accuracy regions, BUSCO covers only conserved single-copy genes, and phasing metrics depend on incomplete reference variant sets. Building on KAT's k-mer spectra ideas, Merqury derives quality, completeness and phase-block metrics without any reference, so it also applies to species that lack curated genomes.

Key results

  1. ShownOn the human NA12878 trio-binned assembly, the reference-free k-mer QV is lower than the mapping-based QV. Read mapping drops regions with very low or very high coverage, which are likely enriched for errors. The k-mer QV is also far cheaper to compute.

    NA12878 TrioCanu: Merqury Q31 per haplotype vs 34.1 from read mapping plus variant calling (2.1 million error bases), which excluded 212.2 Mbp at <3x or >600x coverage; 14.9 CPU h vs 338.3 CPU h (308 h of it for mapping). · Benchmarking on a human genome (NA12878); QV estimates

  2. ShownInherited parental hap-mers give haplotype recall and precision for NA12878 without a reference. Recall is close to the SNP-based sensitivity, and precision is much higher than the SNP-based estimate, which the authors attribute to consensus errors probably being called as false heterozygous SNPs.

    From 18.4 and 19.9 million inherited paternal and maternal hap-mers: recall 90.7% (maternal) and 91.0% (paternal), precision 99.1% and 98.8%; a platinum-callset comparison (3.4 million heterozygous SNPs vs hg38) gave 93.5% sensitivity but only 86.1% precision. · Benchmarking on a human genome (NA12878)

  3. ShownOn an A. thaliana F1 hybrid, copy-number (spectra-cn) and assembly (spectra-asm) spectra reveal partial haplotype collapse, false duplications and missing heterozygous k-mers in an unpolished Canu assembly. They also show an imbalanced FALCON-Unzip primary that absorbs sequence from both haplotypes.

    A. thaliana F1 reads at 45x: Canu k-mer completeness 95.7% vs 98.3% and 97.8% for the other two diploid assemblies; 42.7 Mbp of homozygous sequence left un-collapsed and 10.8 million artificially duplicated k-mers; QV 29.3 for Canu vs 35.7 (TrioCanu) and 36 (FALCON-Unzip) combined. · Copy number spectrum; Assembly spectrum

  4. ShownPhase blocks computed from hap-mers directly on each assembly separate the trio-binned (TrioCanu), pseudo-haplotype (FALCON-Unzip) and mixed-haplotype (Canu) A. thaliana assemblies. A stricter short-range switch allowance gives shorter but more reliable blocks.

    At the default of 100 switches per 20 kbp: TrioCanu phase-block NG50 3.6 and 5.5 Mbp (0.3% switch rate), FALCON-Unzip 3.1 and 2.5 Mbp (0.3%), Canu 100 kbp. Tightening to 10 per 20 kbp on Canu cut NG50 to 33 kbp and the per-block switch error from 3.4% to 0.47%. · Evaluating phasing completeness with hap-mers

  5. ShownExact k-mer counting is the main resource cost when evaluating a human-scale genome. Subsampling with Mash Screen is a lighter alternative that gives QV only, and it slightly underestimates QV because it ignores copy number.

    On NA12878, exact counting peaked at 21 GB memory and 25.5 GB disk (9.1 h counting, 4.7 h merging, 1.1 h statistics). Mash Screen QVs (sketch size 1000000, k=21) were slightly lower per haplotype and lower still for the combined assembly. · Benchmarking on a human genome (NA12878); Mash-based QV

Methods, data and reproducibility

Meryl counts canonical k-mers (k=18 for Arabidopsis, k=21 for human) in high-accuracy Illumina reads and in the assemblies. Set operations derive reliable k-mers and inherited parental hap-mers, from which Merqury computes QV (via Mash-style k-mer containment), completeness, copy-number spectra, hap-mer blob plots, and phase blocks with switch error.

  • Illumina whole-genome reads of an A. thaliana F1 hybrid of Col-0 and Cvi-0, used as the k-mer read set (k=18) · Arabidopsis thaliana · n = 45x coverage
  • A. thaliana F1 assemblies: TrioCanu (Koren et al.), FALCON-Unzip (Chin et al.) and an unpolished Canu 1.9 assembly, all built from the same PacBio reads · Arabidopsis thaliana · n = three assemblies
  • Parental Col-0 and Cvi-0 genome assemblies (Chin et al.) used to derive parental-specific k-mers · Arabidopsis thaliana
  • TrioCanu haplotype-resolved maternal and paternal assemblies of NA12878 (Koren et al.) · human
  • Illumina whole-genome sequencing of NA12878 (Illumina Platinum Genomes) · human · PRJEB3381
  • Parental Illumina WGS of NA12891 and NA12892 from the 1000 Genomes high-coverage data, combined with Platinum Genomes reads · human · PRJEB3381
  • Illumina platinum heterozygous SNP callset for NA12878 against hg38, used as the mapping-based comparator for haplotype completeness and precision · human · n = 3.4 million heterozygous SNPs
Methods & tools
Merqury, Meryl v1.0, Mash Screen, KAT (k-mer spectra concept), BUSCO v3, TrioCanu, FALCON-Unzip, Canu 1.9, BWA, FreeBayes v1.3.1, GenomeScope, IGV
Software & environment
Meryl v1.0; Canu 1.9; BUSCO v3 (embryophyta_odb9, mammalia_odb9); FreeBayes v1.3.1; other requirements are gcc 4.8 or higher and java 1.6 or higher; written in C++, Java and Perl.
Assay conditions & compute
Intel Xeon Gold 6140 CPU @ 2.30GHz node with up to 24 CPUs (NIH HPC). NA12878 QVs took 14.9 CPU h, with peaks of 21 GB memory and 25.5 GB disk during k-mer counting.
Reproducibility gaps
Several command lines are announced but not shown in the provided text. No accession is printed for the A. thaliana F1 PacBio reads (cited only to Chin et al.), and no source is given for the F1 Illumina reads.
Limitations (6)
  • The k-mer QV depends on the read set. Missing true k-mers from low or biased coverage push it down, and abundant false k-mers push it up. The survival model also treats overlapping k-mers and clustered errors as independent, so the choice of k shifts the estimates. · stated by the authors
  • Inherited hap-mers only approximate the child's true haplotype markers: de novo variants and heterozygous variants inherited differently from each parent distort them. They are also scarce in low-heterozygosity genomes such as human (0.12% vs 0.99% for the A. thaliana F1). · stated by the authors
  • A clean spectrum is necessary but not sufficient: some structural mis-assemblies leave the copy-number spectra unchanged. In repetitive genomes, recurring errors in high-copy repeats can also pass the reliable-k-mer cutoff. · stated by the authors
  • The low-memory Mash Screen route ignores copy number. It therefore underestimates QV, especially for combined haplotypes, and cannot localize base errors or support most other Merqury analyses. · stated by the authors
  • Validation covers only one A. thaliana F1 and one human trio, and no assembly has a known error content. The claim that k-mer QV is more accurate than mapping-based QV therefore rests on argument rather than on a ground-truth comparison. · noted in this summary
  • Phase-block NG50 and switch rates depend strongly on the user-set short-range switch allowance (Canu NG50 fell from 100 to 33 kbp). Comparing results across studies would require matched thresholds. · noted in this summary

Read this if you assemble diploid or trio-sequenced genomes and need reference-free QV, k-mer completeness, or phase-block and switch-error metrics, or if you want Meryl's k-mer set operations.

Citations
3,648
FWCI
62.99
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
THISMapleson '16Formenti '22O’Donnell '26He '20Chow '16Priyam '21
66% semantic matchBioinformatics
Daniel Mapleson, Gonzalo Garcia Accinelli, George Kettleborough+2 more·Bioinformatics·Oct 17, 2016· 541 citations