Skip to content

Step 18: Depth-Based CNV Calling with CNVpytor

What This Does

Detects copy number variants (CNVs) using read-depth analysis — complementary to Manta's paired-end/split-read approach. Especially effective for large CNVs (>1 kb) that Manta may miss.

CNVpytor is the maintained Python successor to CNVnator (same lab — Abyzov — and the same read-depth method), replacing the heavy CERN ROOT dependency with a portable HDF5 .pytor store.

Why

Manta (step 4) detects SVs from discordant read pairs and split reads, which works well for balanced SVs (inversions, translocations) and smaller deletions/duplications. CNVpytor uses read-depth signal only, making it better for: - Large duplications/deletions (>10 kb) - Tandem duplications where breakpoints are ambiguous - Validating Manta/Delly calls with orthogonal evidence

Read depth alone cannot see copy-neutral loss of heterozygosity (LOH): the depth does not change. That needs allele frequencies (CNVpytor's -snp/-baf path), which this pipeline does not run. Step 11 (ROH) covers long homozygous stretches from the VCF.

Tool

  • CNVpytor (Abyzov lab) — Python reimplementation of CNVnator (Abyzov et al., Genome Research 2011)

Docker Image

  • CNVPYTOR_IMAGE

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

Reference resources required. The biocontainer ships without the GC/mask resource files and its built-in -download is broken in 1.3.2. Pinned resource files must be present at ${GENOME_DIR}/reference/cnvpytor/ and are bind-mounted into the container. See 00-reference-setup.md for the one-time download.

Command

source versions.env   # from the repository root
SAMPLE=your_sample
GENOME_DIR=/path/to/your/data
IMG="${CNVPYTOR_IMAGE}"
# The container has no GC/mask data; mount the pinned resource dir onto its data path.
DATA=/usr/local/lib/python3.12/site-packages/cnvpytor/data
MOUNTS="-v ${GENOME_DIR}:/genome -v ${GENOME_DIR}/reference/cnvpytor:${DATA}"
PYTOR=/genome/${SAMPLE}/cnvpytor/${SAMPLE}.pytor
mkdir -p ${GENOME_DIR}/${SAMPLE}/cnvpytor

# 1. Import read depth from the BAM (no reference FASTA needed), canonical chromosomes only
docker run --rm --user root ${MOUNTS} ${IMG} \
  cnvpytor -root ${PYTOR} -rd /genome/${SAMPLE}/aligned/${SAMPLE}_sorted.bam \
  -chrom chr{1..22} chrX chrY -j 4

# 2. Read-depth histogram + automatic GC correction
docker run --rm --user root ${MOUNTS} ${IMG} cnvpytor -root ${PYTOR} -his 1000

# 3. Partition (mean-shift segmentation)
docker run --rm --user root ${MOUNTS} ${IMG} cnvpytor -root ${PYTOR} -partition 1000

# 4. Call CNVs (tab-separated)
docker run --rm --user root ${MOUNTS} ${IMG} cnvpytor -root ${PYTOR} -call 1000 \
  > ${GENOME_DIR}/${SAMPLE}/cnvpytor/${SAMPLE}_cnvs.txt

# 5. Export a VCF (deletions / duplications). `-view` reads commands from stdin, so pass -i.
# The heredoc marker is unquoted so the shell expands ${SAMPLE} in the file name.
docker run --rm --user root -i ${MOUNTS} ${IMG} cnvpytor -root ${PYTOR} -view 1000 <<VIEW
set print_filename /genome/${SAMPLE}/cnvpytor/${SAMPLE}_cnvs.raw.vcf
print calls
VIEW

scripts/18-cnvpytor.sh runs all of the above and then normalizes the VCF (reheaders the full contig set from the reference .fai, sorts, bgzips, and indexes) so it merges cleanly with the other SV callers in step 22.

Bin Size

The 1000 parameter is the bin size in base pairs (must be divisible by 100). Use: - 1000 for 30X WGS (recommended) - 10000/100000 for coarser, faster genome-wide scans - CNVpytor can compute several bin sizes at once: -his 1000 10000 100000

Output

  • ${SAMPLE}.pytor — HDF5 read-depth store (intermediate; a few GB, can be deleted)
  • ${SAMPLE}_cnvs.txt — tab-separated CNV calls, 11 columns:
  • type (deletion / duplication)
  • region (chr:start-end)
  • size (bp)
  • read-depth level (normalized; 1.0 ≈ diploid) 5–8. e-values (t-test / Gaussian tail, full and middle region)
  • q0 (fraction of reads with mapping quality 0)
  • pN (fraction of reference N bases)
  • dG (distance to nearest gap >100 bp)
  • ${SAMPLE}_cnvs.vcf.gz (+ .tbi) — normalized VCF with SVTYPE=DEL/DUP, END, SVLEN, GT, and CN

Filtering

# Keep confident CNVs (e-value in col 5 < 0.05, size in col 3 > 1 kb)
awk -F'\t' '$5 < 0.05 && $3 > 1000' ${SAMPLE}_cnvs.txt
The first call is often a low-level artifact spanning the chromosome-start N gap (pN near 1.0); the pN/q0 columns let you drop such regions.

Runtime

~1-3 hours per 30X WGS genome (CNVpytor is multi-threaded via -j).

Notes

  • Run AFTER alignment (step 2). Independent of Manta/Delly — can run in parallel.
  • CNVpytor and Manta overlap on the majority of true large CNVs; variants called by both have lower false-positive rates than single-caller calls.
  • No CERN ROOT dependency — the .pytor store is standard HDF5, so the container is small and reproducible.
  • Reproducibility: the GC/mask resources are pinned to the CNVpytor v1.3.2 tag and mounted offline; no network access is needed at run time (see 00-reference-setup.md).
  • For maximum sensitivity, intersect CNVpytor calls with Manta and Delly (step 19) for a consensus call set (step 22).
  • CNVpytor can also model B-allele frequency (-snp/-baf) for allele-specific CNV/LOH; this pipeline uses the read-depth path.
  • Read-depth import is restricted to the canonical chromosomes (chr1–chr22, chrX, chrY). The default no-ALT reference still has unplaced and unlocalized scaffolds and chrEBV, and a BAM aligned to a full GRCh38 reference (through REF_FASTA) hundreds of ALT/HLA/decoy contigs. The GC-correction data covers none of them; without this restriction -rd chokes on them and produces no calls.