Skip to content

Structural Variant Triage

Parse, annotate, score, and classify structural variants (DEL, DUP, INV, INS, BND, CNV) using the ClinGen 2020 technical standards for CNV interpretation.

Quick start

vartriage sv --sv-vcf sv_calls.vcf.gz --output sv_report.json \
  --gene-annotation gencode.v46.gtf \
  --dosage-sensitivity clingen_dosage.tsv \
  --gnomad-sv gnomad_sv.bed \
  --pathogenic-regions clingen_pathogenic_regions.bed \
  --benign-regions clingen_benign_regions.bed

Pipeline stages

SVParser → SVAnnotator → SVScorer → SVClassifier → Report

  1. SVParser - Streams SV records from VCF, extracting SVTYPE, END/SVLEN, confidence intervals, BND notation. Supports Manta, DELLY, GATK-SV, GRIDSS, and LUMPY INFO field conventions.

  2. SVAnnotator - Determines gene overlap (whole-gene vs partial), attaches ClinGen dosage sensitivity (HI/TS scores), and matches against gnomAD-SV population frequencies via reciprocal overlap.

  3. SVScorer - Computes a composite pathogenicity score (0-1) from four weighted components:

  4. Gene impact severity (35%)
  5. Dosage sensitivity (30%)
  6. Population frequency rarity (20%)

  7. SV size (15%)

  8. SVClassifier - Applies ClinGen 2020 evidence framework (Sections 1-4) to produce a 5-tier classification: Pathogenic, Likely Pathogenic, VUS, Likely Benign, Benign.

CLI reference

Standalone SV triage

vartriage sv [OPTIONS]
Flag Default Description
--sv-vcf required Input VCF with SV calls
--output required Output report path
--output-format json json or csv
--gene-annotation None GTF/GFF for gene overlap
--dosage-sensitivity None ClinGen dosage TSV
--gnomad-sv None gnomAD-SV BED (chrom, start, end, type, af)
--pathogenic-regions None Known pathogenic CNV regions BED
--benign-regions None Known benign CNV regions BED
--min-sv-size 50 Minimum SV size (bp)
--max-sv-size 0 Maximum SV size (0 = no limit)
--sv-types all Comma-separated: DEL,DUP,INV,INS,BND,CNV
--max-af 0.01 Maximum population frequency
--min-quality 20.0 Minimum QUAL threshold
--reciprocal-overlap 0.5 Minimum reciprocal overlap for frequency match
--whole-gene-threshold 0.8 Fraction to classify as whole-gene event
--include-benign false Include Benign/Likely_Benign in output

Combined SNV + SV analysis

Add --sv-vcf to the main vartriage command to run both pipelines:

vartriage --vcf snv.vcf.gz --sv-vcf sv_calls.vcf.gz \
  --output report.json --gene-annotation gencode.v46.gtf \
  --gnomad gnomad.tsv

This runs the standard SNV pipeline and the SV triage pipeline, producing two output files: report.json (SNVs) and report_sv.json (SVs).

Reference data

ClinGen dosage sensitivity

Tab-separated file with columns: gene_symbol, hi_score, ts_score

Score scale (ClinGen curation levels):

  • 0 = No evidence for dosage sensitivity
  • 1 = Little evidence
  • 2 = Emerging evidence
  • 3 = Sufficient evidence (gene is established HI or TS)

A bundled subset ships with the package at vartriage/data/clingen_dosage.tsv. To download the full dataset:

python scripts/download_clingen_dosage.py --output vartriage/data/clingen_dosage.tsv

gnomAD-SV frequency database

Tab-separated BED format: chrom, start, end, sv_type, allele_frequency

Available as a bundle:

vartriage bundle download --bundle gnomad-sv

Pathogenic/benign region BED files

Four-column BED: chrom, start, end, syndrome_name

Bundled files include curated ClinGen regions for common microdeletion/microduplication syndromes (22q11.2, Williams-Beuren, Prader-Willi/Angelman, etc.).

Classification framework

The classifier implements the ACMG/ClinGen Technical Standards (Riggs et al. 2020):

Section 1 - Genomic content assessment

  • 1A: Contains protein-coding genes
  • 1B: Contains established HI/TS gene

Section 2 - Overlap with established regions

  • 2A: Complete overlap with pathogenic region
  • 2B-2C: Partial overlap
  • 2D-2F: Benign region overlap

Section 3 - Gene-level evaluation (losses)

  • 3A: HI gene fully contained
  • 3B: HI gene partially deleted
  • 3C: Breakpoint within gene

Section 4 - Duplication-specific

  • 4F: TS gene fully contained
  • 4G: Gene disrupted by dup breakpoint
  • 4H: Intragenic dup without disruption

Evidence points accumulate and map to classification. The thresholds are calibrated to the additive strength of the evidence lines the classifier accumulates, where a strong line contributes about 0.45, a moderate line about 0.25, and a supporting line about 0.10:

  • = 0.99: Pathogenic

  • 0.90 - 0.98: Likely Pathogenic
  • -0.89 to 0.89: VUS
  • -0.98 to -0.90: Likely Benign
  • <= -0.99: Benign

The benign side is symmetric with the pathogenic side, so every tier is reachable from the evidence the pipeline accumulates. A gnomAD-SV population frequency at or above 5 percent is a strong benign line on its own: it reaches Likely Benign without other evidence, and reaches Benign when combined with a benign-region overlap. An SV with no gnomAD-SV frequency match scores neutral for the frequency component, and a gene overlap with no ClinGen dosage record scores neutral for the dosage component; absence of a lookup result is treated as absence of evidence, not as evidence in either direction.

Python API

from pathlib import Path
from vartriage.structural import SVTriagePipeline, SVTriageConfig

config = SVTriageConfig(
    vcf_path=Path("sv_calls.vcf.gz"),
    output_path=Path("sv_report.json"),
    gene_annotation_path=Path("gencode.v46.gtf"),
    dosage_sensitivity_path=Path("clingen_dosage.tsv"),
    gnomad_sv_path=Path("gnomad_sv.bed"),
    pathogenic_regions_path=Path("clingen_pathogenic_regions.bed"),
    benign_regions_path=Path("clingen_benign_regions.bed"),
)

pipeline = SVTriagePipeline(config)
output = pipeline.run()

Using individual components

from vartriage.structural import SVParser, SVAnnotator, SVScorer, SVClassifier

# Parse
with SVParser(Path("sv_calls.vcf.gz"), min_size=50) as parser:
    for sv in parser:
        print(sv.chrom, sv.start, sv.end, sv.sv_type.value)

# Annotate
annotator = SVAnnotator(
    gene_annotation_path=Path("gencode.v46.gtf"),
    dosage_sensitivity_path=Path("clingen_dosage.tsv"),
)
annotated_stream = annotator.annotate(iter(parser))

# Score
scorer = SVScorer(max_allele_frequency=0.01)
scored_stream = scorer.score(annotated_stream)

# Classify
classifier = SVClassifier(
    pathogenic_regions=[("chr22", 18916842, 21465659)],
)
for classified in classifier.classify(scored_stream):
    print(classified.classification.value, classified.syndrome_name)

Output format

JSON

{
  "pipeline": "structural_variant_triage",
  "version": "0.13.0",
  "total_variants": 3,
  "variants": [
    {
      "chrom": "chr22",
      "start": 18916842,
      "end": 21465659,
      "sv_type": "DEL",
      "length": 2548818,
      "consequence": "Whole_Gene_Deletion",
      "classification": "Pathogenic",
      "pathogenicity_score": 0.92,
      "evidence_score": 1.35,
      "evidence_categories": ["1A", "1B", "2A", "3A"],
      "genes_affected": 5,
      "hi_genes_affected": 1,
      "syndrome_name": "22q11.2 deletion syndrome (DiGeorge)",
      "gene_details": [...],
      "scoring": {"dosage_score": 1.0, "size_score": 0.9, "frequency_score": 1.0}
    }
  ]
}

CSV

Flat format with columns: chrom, start, end, sv_type, length, consequence, classification, pathogenicity_score, evidence_score, genes_affected, hi_genes_affected, population_frequency, evidence_categories, gene_symbols.

Supported SV callers

The parser handles caller-specific INFO field conventions:

Caller SVTYPE END field Length field Copy number Mate ID
Manta SVTYPE END SVLEN, INSLEN - MATEID
DELLY SVTYPE END SVLEN - MATEID
GATK-SV SVTYPE END SVLEN CN, CNVAL -
GRIDSS SVTYPE END, END2 SVLEN - PARID
LUMPY SVTYPE END SVLEN - MATEID

All standard VCF 4.3 SV fields are supported. The parser auto-detects the available fields and resolves coordinates accordingly.