Step 15: SV Quality Annotation with duphold

What This Does

Adds depth-based quality scores to structural variant VCFs, enabling simple filtering of false positive SVs. duphold annotates each SV call with three scores derived from read-depth evidence around the breakpoints.

Why

Manta (step 4) calls structural variants from paired-end and split-read evidence, but many calls are false positives. duphold adds depth-of-coverage annotations that allow filtering without losing true calls — it removes 63% of false positive deletions while retaining 99% of true ones.

Tool

• duphold (Brent Pedersen)

Docker Image

• DUPHOLD_IMAGE

Pinned in versions.env; Image versions lists the current tag.

Annotations Added

duphold writes these as FORMAT fields (one value per sample), not INFO fields.

Tag	Meaning	Interpretation
DHFC	Fold-change of depth inside the SV vs the rest of the chromosome it is on	General quality indicator
DHBFC	Fold-change of depth inside the SV vs genome bins with similar GC content	> 1.3 for duplications supports a real duplication (depth rises)
DHFFC	Fold-change of depth inside the SV vs its flanking regions	< 0.7 for deletions supports a real deletion (depth drops as expected)

Command

source versions.env   # from the repository root
REF_FASTA=reference/GRCh38_no_alt_analysis_set.fasta   # see 00-reference-setup.md#the-reference-path-on-every-page
docker run --rm \
  --cpus 4 --memory 4g \
  -v ${GENOME_DIR}:/genome \
  "${DUPHOLD_IMAGE}" \
  duphold \
  -v /genome/${SAMPLE}/manta/results/variants/diploidSV.vcf.gz \
  -b /genome/${SAMPLE}/aligned/${SAMPLE}_sorted.bam \
  -f "/genome/${REF_FASTA}" \
  -o /genome/${SAMPLE}/duphold/${SAMPLE}_sv_duphold.vcf

The Depth Filter

The script then writes duphold/${SAMPLE}_sv_filtered.vcf.gz with its index: the same calls without the deletions and duplications the depth does not support. It is the filter of the Nextflow DUPHOLD_FILTER, and step 5 (AnnotSV) annotates this file. The tags are FORMAT fields, so the expression uses FMT/<tag>[0] (the first sample):

# Drop deletions with DHFFC >= 0.7 and duplications with DHBFC <= 1.3;
# other SV types, and records without a value, stay.
docker run --rm -v ${GENOME_DIR}:/genome "${BCFTOOLS_IMAGE}" \
  bcftools view \
  -e '(INFO/SVTYPE="DEL" && FMT/DHFFC[0] >= 0.7) || (INFO/SVTYPE="DUP" && FMT/DHBFC[0] <= 1.3)' \
  -Oz -o /genome/${SAMPLE}/duphold/${SAMPLE}_sv_filtered.vcf.gz \
  /genome/${SAMPLE}/duphold/${SAMPLE}_sv_duphold.vcf
docker run --rm -v ${GENOME_DIR}:/genome "${BCFTOOLS_IMAGE}" \
  bcftools index -t /genome/${SAMPLE}/duphold/${SAMPLE}_sv_filtered.vcf.gz

The script prints how many records it kept. How many SVs the filter removes from a real genome has not been measured here.

Runtime

3 to 4 minutes per genome, peak memory about 1 GB, measured in three 30-40x WGS runs (runtime).

Notes

• Run this AFTER Manta (step 4) and before AnnotSV (step 5), which annotates the filtered file when it is there and Manta's calls otherwise.
• Requires the original BAM and reference FASTA — it re-calculates depth around each SV.
• Outputs: duphold/${SAMPLE}_sv_duphold.vcf (uncompressed), the same VCF with three new FORMAT fields, and duphold/${SAMPLE}_sv_filtered.vcf.gz with its .tbi, after the depth filter above.