henhaobio
Sign in

CellGenomicsPeer-reviewed journal article

A complete diploid human genome benchmark for personalized genomics

Nancy Fisher HanseniD, Nathan J. DwarshuisiD, Hyun Joo JiiD, Arang RhieiD, Hailey LoucksiD, Glennis A. LogsdoniD, Mitchell Robert VollgeriD, Jessica M. StoreriD, Juhyun KimiD, Eleni Adam, Nicolas AltemoseiD, Dmitry AntipoviD, Mobin AsriiD, Sofia N. BarreiraiD, Stephanie C. BohaczukiD, Andrey V. BzikadzeiD, Sara A. CariosciaiD, Andrew W. CarrolliD, Kuan-Hao ChaoiD, Yanan ChuiD, Arun DasiD, Peter EbertiD, Adam C. EnglishiD, Mark FlehartyiD, Laura E. Fleming, , Adam M. PhillippyiD

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

Abstract

Human genome resequencing typically involves mapping reads to a reference genome to call variants; however, this approach suffers from both technical and reference biases, leaving many duplicated and structurally polymorphic regions of the genome unmapped. Consequently, existing variant benchmarks, generated by the same methods, fail to assess these complex regions. To address this limitation, we present a telomere-to-telomere genome benchmark that achieves near-perfect accuracy (i.e. no detectable errors) across 99.4% of the complete, diploid HG002 genome. This benchmark adds 701.4 Mb of autosomal sequence and both sex chromosomes (216.8 Mb), totaling 15.3% of the genome that was absent from prior benchmarks. We also provide a diploid annotation of genes, transposable elements, segmental duplications, and satellite repeats, including 39,144 protein-coding genes across both haplotypes. To facilitate application of the benchmark, we developed tools for measuring the accuracy of sequencing reads, phased variant call sets, and genome assemblies against a diploid reference. Genome-wide analyses show that state-of-the-art de novo assembly methods resolve 2-7% more sequence and outperform variant calling accuracy by an order of magnitude, yielding just one error per 100 kb across 99.9% of the benchmark regions. Adoption of genome-based benchmarking is expected to accelerate the development of cost-effective methods for complete genome sequencing, expanding the reach of genomic medicine to the entire genome and enabling a new era of personalized genomics.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (21 Sept 2025) · CC0 1.0

T2T-HG002v1.1, a complete diploid HG002 assembly lacking detectable errors over 99.4% of its length, enables genome-based benchmarking of assemblies, sequencing reads, and phased variant calls.

Mapping-based variant benchmarks such as GIAB v4.2.1 suffer reference bias and exclude the sex chromosomes plus many duplicated and structurally polymorphic regions, so methods go untested exactly where pangenomes should help most. Treating the sample's complete diploid sequence as truth brings 701.4 Mb of autosomal sequence, plus 216.8 Mb of sex chromosomes, into benchmarkable territory and ends the circular dependency on incomplete references.

Key results

  1. ShownT2T-HG002v1.1 resolves all 46 chromosomes telomere-to-telomere (rDNA array interiors unfinished) with 99.35% of diploid length free of detectable errors, and covers far more of the genome than GIAB variant benchmarks; residual disagreements with GIAB v4.2.1 were mostly GIAB errors.

    3,172 low-confidence regions total 38.8 Mb (0.65%) of 6.0 Gb, 86.3% rDNA; Merqury QV rose from Q63.1 (v0.7) to Q68.9 (v1.1); the 216.8 Mb sex chromosomes and 701.4 Mb (11.7%) of autosomes lacking from GIAB v4.2.1 are now covered; excluding rDNA, 99.9% vs 85.2% high-confidence coverage; shared-region discrepancies fell from 5,972 to 219. · A complete, nearly perfect assembly of the HG002…; Introduction; Discussion; Abstract

  2. ShownDe novo assembly now outperforms reference-based variant calling by about an order of magnitude in genome-wide consensus accuracy, and covers more of HG002 than calling against even the complete T2T-CHM13 reference.

    HiFi DeepVariant+HapCut2 variant constructed genomes (min GQ 10–40) covered 93.8–92.9% (GRCh38) and 97.8–95.7% (CHM13) of HG002 with QV 37–40 and 35–39; CHM13 raised substitutions (72.3 vs 41.1 /Mb at GQ ≥ 40); the ONT LC24 assembly reached QV 51 at 99.94% coverage. · Benchmarking variant call sets; Abstract

  3. ShownGenome Quality Checker (GQC) tracks five years of HG002 assemblies against the truth, showing steady gains in continuity, base accuracy and phasing, and that Merqury k-mer QV overrates PacBio-only assemblies while underrating ONT-including ones.

    Across the five assemblies spanning 2020 to 2024, substitution rates fell by more than two orders of magnitude and indels about tenfold; Merqury minus GQC QV was +7.9 (Ash1, Hifiasm) and +11.7 (HPRC) but -0.4 (Verkko) and -1.4 (LC24); homopolymers over 20 bp stayed error-prone. · Benchmarking genome assemblies; Figure legends

  4. ShownOver half the errors that genome benchmarking finds in a prior HPRC diploid assembly fall outside what GIAB variant benchmarking can assess; within confident regions, single repeat errors often surface as multiple variant false positives or negatives.

    Of 66,544 genome-benchmark errors in the Jarvis et al./HPRCv1 assembly, 50,843 lifted to GRCh38 and 30,505 fell in small-variant confident regions; there, variant FP+FN rates tracked genome error rates across repeat strata but were higher, except similar for SNVs in segmental duplications. · Relationship between genome and variant…; Figure legends

  5. ShownA diploid annotation exposes haplotype differences in gene content: broken protein-coding genes concentrate in segmental duplications, and copy-number variable genes such as GSTM1, GSTT1, CFHR1/CFHR3 and DUSP22 sit on only one haplotype.

    59,380 MAT and 57,728 PAT genes, 39,144 protein-coding across both; 129 MAT and 121 PAT broken protein-coding genes, 28% (69 / 250) overlapping segmental duplications vs 8% (2996 / 36,957) of unbroken genes; 14 MAT-only and 12 PAT-only autosomal genes; only 39 genes absent from both haplotypes vs CHM13. · Personalized genome annotation; Abstract

Methods, data and reproducibility

Verkko assembly of roughly 170× HiFi and 209× ONT ultra-long reads (parental Illumina for phasing, Strand-seq/Hi-C for rDNA scaffolding), then three rounds of curated polishing; the GQC tool aligns assemblies, reads, or variant-constructed genomes to the diploid truth, assigning haplotype blocks via a two-state HMM over haplotype-specific k-mers.

  • T2T-HG002v1.1 diploid assembly of HG002, son in an Ashkenazi trio (Coriell LCL GM24385, NIST RM 8391) · human · n = 46 chromosomes; 6.0 Gb diploid · BioProject PRJNA730823; GCA_018852605.3 (paternal), GCA_018852615.3 (maternal)
  • HG002 PacBio HiFi and ONT ultra-long reads (HPRC, GIAB) used to build the assembly; inventory in Table S1 · human · n = roughly 170× HiFi; 209× ONT
  • Parental HG003/HG004 Illumina reads (phasing, switch-error k-mers) and HG002 Strand-seq and Hi-C reads (rDNA scaffolding, inversion check) · human
  • HG002 read sets benchmarked with GQC: ONT Q28 Ultra-Long V14, PacBio Revio HiFi ICSv13, Element Aviti UltraQ 2×150, Illumina NovaSeq PCR-Free 2×151 · human · n = four datasets
  • Earlier HG002 assemblies evaluated: Ash1v2, Hifiasm, HPRCv1, Verkko, ONT LC24, plus Jarvis et al. · human · n = five assemblies plus Jarvis et al.
  • GIAB HG002 v4.2.1 small variant benchmark and draft GRCh38 benchmark derived from T2T-HG002 (defrabb V0.019) · human
  • GRCh38 and T2T-CHM13v2.0 references (annotation liftover, variant constructed genomes), with MANE v1.4 transcripts · human
  • Matched HG002 long-read functional tracks in the browser hub: HiFi/ONT 5mC, Fiber-seq, Iso-Seq Kinnex, MAS-Seq · human
Methods & tools
Verkko (diploid T2T assembly), Ribotin (rDNA consensus), Merqury (k-mer QV and phasing), DeepVariant, DeepTrio, DeepPolisher, HapCut2 (phasing), Liftoff, LiftOn, miniprot (annotation), Genome Quality Checker (GQC), BCFtools (variant constructed genomes), Sniffles, NucFreq, T2T-Polish (validation), Strand-seq and Hi-C, Fluorescence in situ hybridization (rDNA), UCSC Genome Browser hub
Software & environment
Stated versions: T2T-CHM13v2.0, MANE v1.4, GIAB v4.2.1, PacBio HiFi ICSv13 with DeepConsensus v1.2; versions of most other tools and any containers are not given in the main text.
Reproducibility gaps
No new sequencing data; inputs are catalogued in Table S1 and linked from the project GitHub. 3,172 low-confidence regions (mostly rDNA, nine gaps) are excluded; further reanalysis details only from the lead contact on request.
Limitations (6)
  • rDNA arrays remain unresolved: nine gaps spanning 31.1 Mb plus flanking rDNA are flagged low-confidence and excluded, and rRNA genes are omitted from annotation; the authors expect better sequencing is needed given rDNA size and possible mitotic recombination. · stated by the authors
  • Most data came from cultured LCL cells, so somatic differences from the NIST reference DNA batch are expected (85 known mosaic SNVs at 5–30% allele fraction), and V(D)J recombination forced exclusion of immunoglobulin loci from haplotype-switch analysis. · stated by the authors
  • k-mer QV estimates miss loss of heterozygosity, long-homopolymer, and repeat expansion/collapse errors and give false positives when k-mer databases are incomplete; homopolymers over 20 bp and short-read-unmappable repeats remain lower-confidence. · stated by the authors
  • Genome-benchmark error rates use different denominators from precision/recall and are not yet standardized; phase switches and haplotype collapses were easier to interpret in variant-benchmark output, and the clinical meaning of genome metrics is unresolved. · stated by the authors
  • The assembly-versus-variant-calling comparison rests on a single pipeline (HiFi DeepVariant with HapCut2 phasing) for one sample; other callers and multi-technology or pangenome-based calling were not tested, so the order-of-magnitude gap may not generalize. · noted in this summary
  • The truth set derives from one Ashkenazi male and was assembled from HiFi and ONT data overlapping inputs of some evaluated assemblies; residual shared errors could favor similar pipelines, and performance in other ancestries' polymorphic regions is untested. · noted in this summary

Read this if you develop or evaluate sequencing, assembly, phasing, variant-calling, or pangenome inference methods and need a whole-genome diploid truth set for HG002, including segmental duplications, satellites, and sex chromosomes.

Citations
42
FWCI
75.93
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
THISOlson '26Huang '26Jarvis '22Zook '19Sarashetti '25Rautiainen '23
69% semantic matchGenomics
N. D. Olson, N. Dwarshuis, N. F. Hansen+42 more·bioRxiv·Oct 1, 2026