rewirebio.iobenchmarks

Use caseDNA and genomesResearch and clinical research

Select a copy-number variant detection and characterisation workflow

Which workflow detects and characterises copy-number changes accurately and helps an analyst assess the evidence?

What the evidence shows

On HG002 deletions, DRAGEN 4.2 with SV-informed CNV calling found 87 to 100% of deletions in every size bin, with precision 0.99 to 1.00; this is the developers' own benchmark.

CNVnator matched it above 10 kb but found only 26% of deletions of 1 to 5 kb and 51% of 5 to 10 kb. On NA12878, LUMPY, DELLY and Manta recovered 88 to 94% of known CNVs, but only 20 to 33% of their calls matched; Manta and CNVnator were used to build this truth set, which can favour them. The other callers recovered 1 to 43%.

Hwang et al. (2026) compared four exome callers with chromosomal microarray CNVs in 39 clinical exomes, scoring deletion and duplication exons together: sensitivity and PPV were 0.758 and 0.025 for ExomeDepth, 0.662 and 0.096 for CNVkit, 0.768 and 0.239 for CoNIFER and 0.170 and 0.041 for cn.MOPS. The truth set keeps only array CNVs with at least ten probes that overlap exome baits, so any other exon call counts as a false positive. After the authors removed regions where each tool called CNVs in more than 10% of the same 180-exome cohort, ExomeDepth reached 0.903 and 0.102 on a smaller reference set of 2,124 exons (Table 4); this filter is the authors' own method.

For characterisation, Zhang et al. (2026) ran five automated ACMG classifiers on 3,031 expert-classified CNVs: five-category accuracy, averaged across classes in a way the paper does not state, ranged from 48.76% (X-CNV) to 69.81% (MarCNV-ISV) (Table 3). With likely and definite classes merged, every tool recognised more pathogenic deletions than pathogenic duplications (AnnotSV 95.16% versus 86.39%) (Table 4). Two of the three authors developed dbCNV and the reference includes their own laboratory's cases, so the dbCNV results are author-reported.

None of the comparisons shown reports duplication detection separately, none covers tumour samples or long reads, and none evaluates a foundation model.

You bring
  • Whole-genome sequencing aligned reads
  • The size range and variant class (deletion, duplication, etc.) of clinical interest
You want
A sourced F-score for one size-stratified deletion class against a truth set; no visualisation-usability or clinical-reporting claim.
Evidence in this release
4 comparisons shown · 3 held for review · 18 tools · 24 configurations
Kept up to date
Evidence last checked 11 October 2026 · re-checked monthly, next due 11 November 2026
Executive Takeaway

For "Select a copy-number variant detection and characterisation workflow", on Known CNVs in NA12878, Recall (NA12878 WGS, 1 bp overlap) (Recall), Lumpy scored 94.08%, reported by A Comparison of Tools for Copy-Number Variation Detection in Germline Whole Exome and Whole Genome Sequencing Data.

Best reported valueLumpyA Comparison of Tools for Copy-Number Variation Detection…
Recall94.08%Known CNVs in NA12878, Recall (NA12878 WGS, 1 bp overlap)
Baseline on this protocolCNVnator43.45%
Ranked tools88 from independent studies

8 tools have a value on this protocol; 8 of them come from studies whose authors did not develop the tool. The use case's baseline, CNVnator, scored 43.45% on the same protocol.

Ranking

Ranked on Known CNVs in NA12878, Recall (NA12878 WGS, 1 bp overlap) (Recall, higher is better). Only results on this protocol are ranked; other tools are listed below the table.

8 tools, from 2 sources. 8 of 8 results (100%) come from an independent study.

Rank▲ModelRecall (higher is better)vs CNVnatorEvidenceCoverageReported byLicense
#194.08%+116.5%Source-checked Independent study1/4A Comparison of Tools for Copy-Number Variation…Not recorded
#290.37%+108.0%Source-checked Independent study1/4A Comparison of Tools for Copy-Number Variation…Not recorded
#387.52%+101.4%Source-checked Independent study1/4A Comparison of Tools for Copy-Number Variation…Not recorded
#4
CNVnator[-] Baseline
43.45%Source-checked Independent study2/4A Comparison of Tools for Copy-Number Variation…Not recorded
#523.22%-46.6%Source-checked Independent study1/4A Comparison of Tools for Copy-Number Variation…Not recorded
#616.96%-61.0%Source-checked Independent study2/4A Comparison of Tools for Copy-Number Variation…Not recorded
#710.69%-75.4%Source-checked Independent study1/4A Comparison of Tools for Copy-Number Variation…Not recorded
#81.16%-97.3%Source-checked Independent study1/4A Comparison of Tools for Copy-Number Variation…Not recorded
10 tools are evaluated elsewhere in this use case but not on this protocol, so not ranked

Baseline

ExomeDepth, CNVnator, AnnotSV and ClassifyCNV. ExomeDepth (exome) and CNVnator (genome) are read-depth callers that fit statistical models of coverage and are measured in most of the linked detection benchmarks, though no benchmark names one standard caller. AnnotSV and ClassifyCNV apply the ACMG CNV scoring rules, the rule-based reference that machine-learning classifiers are compared with in Zhang et al. 2026.

Machine-learning methods are compared with the baseline on the same data in one table: Automated ACMG CNV classification tools against expert labels (Zhang et al. 2026). Baseline rows are marked in each table.

Tools compared

Comparisons

Unified model comparison for this use case. Compare tools against protocol baselines in a single view.

Known CNVs in NA12878

Direct evidenceIndependent study: Reported in a study whose authors did not develop the evaluated tool.

Recall and precision of eight WGS CNV callers on NA12878 against the Haraksingh et al. 2017 gold-standard set (1 bp overlap)

Sorted by Recall (higher is better). The best value in each column is highlighted.

ToolRecall (NA12878 WGS, 1 bp overlap)Precision (NA12878 WGS, 1 bp overlap)
Lumpyconventional
94.08% (best in column)31.00%
DELLYconventional
90.37%19.88%
Mantaconventional
87.52%32.62%
CNVnatorBaselineconventional
43.45%12.76%
GATK_gCNVconventional
23.22%1.86%
cn.MOPSconventional
16.96%12.25%
ControlFREECconventional
10.69%26.68%
CLCconventional
1.16%66.67% (best in column)
Caveats and method
  • Single sample
  • Manta and CNVnator were used to build the truth set
  • 1 bp overlap match
  • Most tool versions unreported

Protocol: NA12878 WGS CNV recall and precision, 1 bp overlap (Gabrielaite et al. Table S2)

Primary benchmark evaluation shown (Known CNVs in NA12878). 3 additional study protocols (Deletions by size, HG002 (GIAB SV v0.6), Exome CNV callers against CMA truth, before and after removing problematic regions (Hwang et al. 2026), Automated ACMG CNV classification tools against expert labels (Zhang et al. 2026)) are cited in individual tool drawer rows and under Question, sources and review.

Held for review

Recorded but not shown as evidence until the reason below is resolved.

Clinical gene panel across 25 cell lines (De La Vega et al. 2025)

Not shown: the article's Results text and its Supplemental Table 3 disagree on whether CNVnator detected 1 to 5 kb events and on which caller had the lowest precision. The conflict is in the HG002 table, not in this panel table, but it is recorded against the whole article, so these precision values are not shown until it is resolved.

Deletions and duplications by size, HG002 (De La Vega et al. 2025)

Not shown: the article's Results text and its Supplemental Table 3 disagree on whether CNVnator detected 1 to 5 kb events and on which caller had the lowest precision. The values here are recorded as printed in the tables, but they are not shown until that conflict is resolved.

Short-read deletion calling by size, HG002 (Nardone et al. 2025)

Not shown: the paper states the read depth and aligner behind Table S1 inconsistently (reads subsampled to 25x and aligned with bwa-mem2 for preliminary evaluations in Methods, the DRAGEN pipeline in Results, 30x reads in Data Availability), and its text and Table S1 disagree for inGAP at 5,000 to 9,999 bp. The table values are recorded as printed, but they are not shown until this is resolved.

Not covered yet

  • Unreported per-bin counts behind the printed F-score.
  • No foundation-model applicability is established for this task.
  • Whole-genome duplication evidence is limited to 10 HG002 duplication truth events in one source (De La Vega et al. 2025); the exome comparison (Hwang et al. 2026) scores duplication and deletion exons together against chromosomal microarray truth.
  • No tumour or somatic CNV result with printed per-caller values was found; the SEQC2 HCC1395 benchmark reports accuracy in figures only.
  • Results are not comparable across sources: each uses its own truth set, matching rule and input data.
  • The De La Vega et al. 2025 and Nardone et al. 2025 mappings are held as drafts because their sources carry evidence concerns (prose and tables disagree; Nardone input conditions are unresolved).
  • The Genome Biology 2026 HG002 T2T-Q100 and CMRG SV benchmark (doi:10.1186/s13059-026-04048-4) is not stored because its supplementary table identity is ambiguous.
  • No machine-learning CNV caller has been compared with ExomeDepth or CNVnator on the same data in the linked evidence.

Contribute evidence or propose a correction

How to cite

Select a copy-number variant detection and characterisation workflow. Rewire biological model benchmark database, release 2026-11-21-c8bea25e5abb. https://benchmarks.rewirebio.io/use-cases/cnv-detection-characterisation/

Release 2026-11-21-c8bea25e5abb. Rankings and values change between releases, so keep the release id in the citation. Cite the original publication as well for any value you reuse: each result links its source.

The compilation and annotations made by Rewire in this release are offered under CC BY 4.0; values and text taken from cited sources stay under the terms of those sources and are not covered by this licence. Data licence

Question, sources and review

Who this is for, setting and exclusions

Inspect the DRAGEN 4.2 CNV/SV benchmarking F-score evidence for 1-5 kb deletions before selecting a workflow; evidence-visualisation usability is not addressed by this bounded intake.

Who this is for
  • Clinical researchers scoping the available diagnostic-genomics evidence
  • Computational researchers comparing exact evaluated configurations
Biological setting
DRAGEN 4.2's CNV/SV benchmarking (HG002, GIAB SV truth set) reports an F-score for 1-5 kb deletion detection (selected Table S4 cells, sheet CNVbenchmarking, H6=.926, L6=.391). The 2026-10-09 pass adds the rest of DRAGEN Table S4 (CNV-only DRAGEN and CNVnator across five deletion bins), Gabrielaite et al. 2021 (eight WGS callers on NA12878), and as drafts De La Vega et al. 2025 (eight WGS caller configurations including the deep-learning caller Cue) and Nardone et al. 2025 (ten short-read callers on HG002 deletions).
Clinical research scope
Clinical applicability is not established beyond this one size-stratified deletion F-score; a conflicting comparator figure (.391 vs 39.20% in source table/prose) is preserved unresolved and not promoted.

Outside this use case

  • Duplications, larger/smaller size strata, and tumour CNV (not ingested in this bounded intake)
  • Evidence-visualisation usability for an analyst (not measured by this F-score endpoint)
Evidence collection plan
Evidence collection planCollecting evidence

Already on record: 4 comparisons shown · 3 held for review · 18 tools, in the evaluated evidence above. The status above describes only the specific comparison in this plan, which remains open; it does not mean no evidence has been collected. The plan defines a comparison to investigate; it does not establish model performance or suitability.

Comparison question

Which workflow detects and characterises copy-number changes accurately and helps an analyst assess the evidence?

Baselines, outcomes and validation requirements

Baselines to include

  • The conventional/author-introduced workflow measured in the linked primary source(s).

Outcomes to measure

  • The declared endpoint in this use case's active mapping(s); see evidence_gaps for what remains open.

Validation requirements

  • Independent held-out population matched to the intended clinical setting.
  • Qualified human scientific review before any clinical-validation claim.

Next collection task

A dedicated rewire-benchmarks protocol/run task covering duplications and tumour CNV, plus an analyst-facing visualisation-usability evaluation.

Sources, review and downloads

Automated source review · 2026-10-07 · Claude Sonnet AMP-integration worker, bounded transcription of Codex-checked primary values; independently reviewed by Codex (workbench/amp-supervision/primary-review.md, integration-review-corrections.md)

Bounded primary-source transcription, independently Codex-checked. No new model execution, independent experimental reproduction, qualified human scientific review or clinical validation.

Release 2026-11-21-c8bea25e5abb · question use-case-cnv-detection-characterisation

Use-case input digest bfbb7ab9303becad1e3cb395568c26f41e72ddb67cd1a89a06d9eaa0f76bd95f

Download questions and review metadata (JSON) (gzip) · Verify release checksums (gzip)