henhaobio
Sign in

medRxivGenetic & Genomic MedicinePreprint

SVkhor: a unified framework for structural variant integration across long-read, short-read, and optical genome mapping data

Edris Sharif RahmaniiD, Quentin ThomasiD, Émilie Tisserant, Valentin VautrotiD, Anthony Auclair, Fridolin Hounnondaho, Estella CastilloniD, Laurence Olivier FaivreiD, Christel Thauvin‐RobinetiD, Antonio VitobelloiD, Yannis DuffourdiD

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

Abstract

SummaryMulti-technology human genome structural variant (SV) discovery is challenged by differences in breakpoint resolution, allele representation, SV annotation, and VCF structure across various callers and platforms. Here, we present SVkhor, a software framework designed to merge outputs from multiple callers within each technology and integrate SV callsets across available short-read sequencing, long-read sequencing, and optical genome mapping data. SVkhor addresses these challenges through caller-aware normalization, within-technology merging, and cross-technology integration, producing compact, source-annotated SV catalogs suitable for benchmarking and downstream interpretation. Benchmarking using HG002 and analysis of a clinical trio demonstrate that SVkhor reduces redundant caller-level complexity while preserving technology-specific evidence, enabling the transition from heterogeneous SV callsets to interpretable sample- and family-level SV catalogs. Availability and implementationSVkhor is implemented as a Linux command-line workflow. The source code, documentation, and example workflows are accessible at http://gitlab.gad-bioinfo.org/gad-public/svkhor under the MIT license. ContactAntonio.Vitobello@u-bourgogne.fr or yannis.duffourd@u-bourgogne.fr Supplementary informationSupplementary data are accessible online.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (2 Aug 2026) · CC BY 4.0

SVkhor normalizes, merges and integrates long-read, short-read and optical-mapping structural variant calls; on HG002 it retained 98.6% of naive concatenation's true positives with 58.2% fewer false positives.

Long-read (LR), short-read (SR) and optical genome mapping (OGM) structural variant (SV) callsets are hard to combine because callers and platforms differ in breakpoint resolution, allele representation and VCF conventions. Earlier mergers cluster by position (SURVIVOR, Jasmine) or sequence (SVanalyzer, PanPop), whereas SVkhor adds caller-aware normalization and source-annotated integration across technologies that also covers symbolic records and interval-level OGM evidence.

Key results

  1. ShownMerging all three HG002 technologies kept nearly all the true positives that plain concatenation recovers and cut its false positives by 58.2%. Each merged record notes which technologies support it.

    techmerge produced 110,941 deletion/insertion (DEL/INS) records, 17.2% supported by more than one technology. Truvari vs GIAB HG002 v5.0q (28,123 high-confidence events): 26,151 true positives, recall 0.930, F1 0.499. Compared with bcftools concat of the normalized VCFs, SVkhor kept 98.6% of true positives and had 58.2% fewer false positives. · 2.6 Benchmarking and tool comparison; 3.2 HG002 benchmarking and tool comparison

  2. ShownLong-read-only merging with svmerge had the top recall among the single-technology merging runs. SVkhor beat Jasmine and SVanalyzer on sensitivity. Every tool did worse on short-read-only input.

    LR-only svmerge: 76,650 records (60,320 after benchmark filtering), recall 0.923, F1 0.587. Jasmine and SVanalyzer held their recall in both the LR-only and LR+SR+OGM settings, but SVkhor had higher sensitivity and a better precision-recall balance (Figure 2A). SURVIVOR had fewer false positives and lower sensitivity. PanPop produced no benchmarkable output for the combined input. · 3.2 HG002 benchmarking and tool comparison

  3. ShownAmong tools whose outputs carry informative genotypes, SVkhor had the highest genotype concordance in both the long-read-only and the integrated runs.

    Truvari genotype concordance was 0.776 (LR-only) and 0.777 (LR+SR+OGM). Site-only or missing-genotype outputs from other tools were marked not applicable (Supplementary Figure 2). · 2.6 Benchmarking and tool comparison; 3.2 HG002 benchmarking and tool comparison

  4. ShownSVkhor produces a sorted, indexed, benchmark-ready VCF in three steps. It ran the integrated analysis in under a minute and used less memory than Jasmine, SVanalyzer and PanPop, but it was not the fastest tool by wall-clock time.

    Three steps vs six to eight for the external tools, which needed header repair, multiallelic decomposition, sorting, compression and indexing. Integrated LR+SR+OGM run on up to four threads: 54.4 s and about 2.1 GB peak memory, about 41 times faster than SVanalyzer and 19 times faster than the LR-only PanPop run. SVanalyzer and PanPop used about 1.7- and 10.5-fold more peak memory. · 3.3 Workflow complexity and computational…

  5. SuggestedIn a neurodevelopmental-disorder trio, integrating first per sample and then per family reduced the caller output to one non-redundant catalog. That catalog was then sorted into proband-centred inheritance classes.

    949,302 caller-level records became 607,121 after normalization, 391,557 after cross-technology integration and 317,857 trio-wide events. These fell into 14 classes: 61.2% reference in the proband, 23.6% of unresolved parental origin, 12.9% maternally, paternally or biparentally inherited. Recessive, mosaic and de novo candidates together made up under 1%. · 2.7 Clinical use case; 3.4 Clinical trio SV integration with SVkhor

Methods, data and reproducibility

svnorm harmonizes caller VCFs (type, span, contigs, multiallelic splitting, DEL/INS left-alignment); svmerge graphs compatible same-type records per technology (breakpoint/interval, Edlib sequence, orientation, copy-number checks), one record per component; techmerge matches SR calls to a long-read backbone, then OGM calls to nearby LR or SR events in the same interval, and annotates technology support.

  • Genome in a Bottle (GIAB) HG002 Illumina short-read sequencing, aligned with BWA-MEM to GRCh38 with decoys · human · n = 5.89 billion reads after filtering; 267× depth
  • HG002 Oxford Nanopore long reads (≥300 bp, Ratatosk-corrected; also assembled with Flye) · human · n = 156.5 GB; read N50 48.4 kb; 41.3× depth
  • HG002 Bionano optical genome mapping molecules · human · n = 1.28 million molecules retained, 780,185 aligned; 63.56× effective coverage
  • GIAB HG002 v5.0q GRCh38 DEL/INS benchmark (high-confidence BED regions) · human · n = 28,123 high-confidence events
  • Neurodevelopmental-disorder trio with LR, SR and OGM data from the authors' laboratory · human · n = three family members
Methods & tools
SVkhor, Truvari v5.4.0, Jasmine v1.1.5, SURVIVOR v1.0.7, SVanalyzer v0.36, PanPop, Edlib, Manta v1.6.0, cuteSV2 v2.1.1, Dysgu v1.8.7, SVIM-asm v1.0.3, Bionano Solve v3.8.2.1
Software & environment
Versions are given for most tools, e.g. fastp v1.3.2, minimap2 v2.30, Ratatosk v0.9.0, Flye v2.9.6, samtools v1.22.1, CNVnator v0.4.1, SVision v1.4, LongTR v1.2. Callers ran at default settings with a 50 bp minimum. SVkhor is MIT-licensed.
Assay conditions & compute
Runtime and memory were profiled on up to four threads where supported. The integrated LR+SR+OGM run took 54.4 s with about 2.1 GB peak memory. Hardware is not specified.
Reproducibility gaps
Individual-level trio data are withheld for participant confidentiality and data protection (summary results only). No accession numbers are given for the HG002 datasets. No container or environment file is described, and the profiling hardware is not stated.
Limitations (4)
  • Benchmarking covered only deletions and insertions of at least 50 bp. The HG002 truth set has no comparable reference for the other SV classes SVkhor handles, such as copy number variants and breakend-type events. · stated by the authors
  • Evaluation used one benchmark genome. The authors say it should be extended to independent benchmark samples and to more SV classes. · stated by the authors
  • The integrated LR+SR+OGM set had lower F1 (0.499) than LR-only merging (0.587), with recall rising only from 0.923 to 0.930. Adding SR and OGM therefore costs precision on this benchmark. All long-read data were Oxford Nanopore. · noted in this summary
  • The trio inheritance classes, including the de novo and mosaic candidates, were not checked against a truth set or by an independent assay, and the data are not public. The clinical use case is therefore illustrative only. · noted in this summary

Read this if you need to combine SV calls from several callers across long-read, short-read and Bionano optical mapping data into one per-sample or per-family catalog that records which technologies support each event.

Citations
0

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
THISMugoya '25Sarwal '24Li '25Brown '25Denti '22Chaisson '19
69% semantic matchBioinformatics
Trevor Mugoya, Arun Sethuraman·bioRxiv·Nov 26, 2025