pbsv

PacBio structural variant (SV) calling and analysis tools

---

pbsv is a suite of tools to call and analyze structural variants in diploid genomes from PacBio single molecule real-time sequencing (SMRT) reads. The tools power the Structural Variant Calling analysis workflow in PacBio's SMRT Link GUI.

pbsv calls insertions, deletions, inversions, duplications, and translocations. Both single-sample calling and joint (multi-sample) calling are provided. pbsv is most effective for:

• insertions 20 bp to 10 kb
• deletions 20 bp to 100 kb
• inversions 200 bp to 10 kb
• duplications 20 bp to 10 kb
• translocations between different chromosomes or further than 100kb apart on a single chromosome

Availability

Latest version can be installed via bioconda package pbsv.

Please refer to our official pbbioconda page for information on Installation, Support, License, Copyright, and Disclaimer.

Latest Version

Version 2.9.0: Full changelog here

Workflow

The general pbsv workflow is:

1. Align PacBio reads to a reference genome, per movie. (.subreads.bam/.ccs.fastq to .bam)
2. Discover signatures of structural variation. (.bam to .svsig.gz)
3. Call structural variants and assign genotypes, all samples. (.svsig.gz to .vcf)

1. Align PacBio reads to a reference genome

The recommended aligner is pbmm2 that can be installed via conda install pbmm2. For each movie (.subreads.bam or .ccs.fq) align records to a reference genome (ref.fa).

Subreads BAM input:

pbmm2 align ref.fa movie1.subreads.bam ref.movie1.bam --sort --median-filter --sample sample1

CCS BAM input:

pbmm2 align ref.fa movie1.ccs.bam ref.movie1.bam --sort --preset CCS --sample sample1

CCS FASTQ input:

pbmm2 align ref.fa movie1.Q20.fastq ref.movie1.bam --sort --preset CCS --sample sample1 --rg '@RG\tID:movie1'

The sample name, stored in the SM tag of the read groups, associates aligned reads with a particular sample. It is required for downstream joint calling.

2. Discover signatures of structural variation

For each aligned BAM or set of aligned BAMs, identify signatures of structural variation. This reduces all aligned reads to those that are relevant to calling structural variants. The signatures are stored in a .svsig.gz file.

pbsv discover ref.movie1.bam ref.sample1.svsig.gz
pbsv discover ref.movie2.bam ref.sample2.svsig.gz

# optionally index svsig.gz to allow random access via `pbsv call -r`
tabix -c '#' -s 3 -b 4 -e 4 ref.sample1.svsig.gz
tabix -c '#' -s 3 -b 4 -e 4 ref.sample2.svsig.gz

It is highly recommended to provide one tandem repeat annotation .bed file of your reference to pbsv discover via --tandem-repeats. This increases sensitivity and recall. Feel free to use the following for human SV calling: GRCh38 or hs37d5/hg19.

Sample names are transferred from the RG headers to the .svsig.gz file.

3. Call structural variants and assign genotypes

Call structural variants from structural variant signatures, jointly for all samples of interest. One or more .svsig.gz files are accepted, including multiple .svsig.gz for a single sample and/or svsig.gz for multiple samples. If the input is CCS reads, please add --ccs to the following call:

pbsv call ref.fa ref.sample1.svsig.gz ref.sample2.svsig.gz ref.var.vcf

Variant calls for all samples are output in a single .vcf file.

Parallel processing per chromosome

For large genomes with high sequencing coverage, it is recommended to process chromosomes separately. After aligning each movie:

Generate separate .svsig.gz files per chromosome

for i in $(samtools view -H hg38.movie1.bam | grep '^@SQ' | cut -f2 | cut -d':' -f2); do
    pbsv discover --region $i hg38.movie1.bam hg38.sample1.$i.svsig.gz
done

Call SVs

# -j is number of threads
pbsv call -j 8 hg38.fa hg38.sample1.*.svsig.gz hg38.sample1.vcf

Algorithm Overview and Advanced Parameters

Deletions

Cluster options used during pbsv call:

SV Signature Cluster Options:
  --cluster-max-length-perc-diff   Do not cluster signatures with difference in length > P%. [25]
  --cluster-max-ref-pos-diff       Do not cluster signatures > N bp apart in reference. [200]

Number of flanks used for consensus generation:

Consensus Options:
  -x,--max-consensus-coverage      Limit to N reads for variant consensus. [20]

Split deletions: Deletions that are not fully aligned using the D cigar are recovered up to a size of 100kb. Deletions greater than 100kb are currently called as translocations.

Insertions

Insertion calling workflow is identical to the above described deletion workflow, except for one additional criteria, the inserted sequence similarity check during clustering:

SV Signature Cluster Options:
  --cluster-min-basepair-perc-id   Do not cluster signatures with basepair identity < P%. [10]

Split insertions: Insertions can be recovered from split split reads, if the reference overlap of those split reads is less than -k,--max-skip-split [100] and if the query distance is larger than 500 bp.

The explicit upper limit on the insertion size can be adjusted in pbsv call, but be aware that predicting larger insertions will consume more memory!

  --max-ins-length   Ignore insertions with length > N bp. ["15K"]

Inversions

An inversion signature is detected if a single read is split into three alignments with different orientations / strands, either +-+ or -+-. Gaps in the reference, as depicted as A, are not constrained. The maximum permitted reference overlap B, between consecutive alignments, is configured in pbsv discover:

  -k,--max-skip-split   Ignore alignment pairs separated by > N bp of a read or reference. ["100"]

Clustering is performed on the inverted segment and uses the same criteria as deletion clustering.

The VCF call marks the most likely position and size of the inverted segment, as shown in this IGV screenshot:

Translocations

Translocations are identified using breakends of individual split reads with a query skip of less than -k,--max-skip-split [100]. The minimum reads that support a BND (total over all samples), can be defined with --call-min-bnd-reads-all-samples [2]. All four breakend combinations are supported:

Duplications

From split reads

Duplications can be identified from the following split-read signatures:

whereas a duplication has to include on fully spanning read to be flagged as PASS; otherwise, it is filtered with NotFullySpanned.

The maximum size can be configures in pbsv call:

  --max-dup-length   Ignore duplications with length > N bp. ["1M"]

From insertion

In addition, each insertion is tested against its neighboring reference regions and labeled as a duplication if it matches the reference with 80%. Caution: Activating duplication calling has a negative impact when comparing to GIAB, as GIAB labels everything as insertion or deletion.

Calling and Genotyping

An variant is output if it passes all of the following criteria:

• supported by at least -A,--call-min-reads-all-samples [3] reads total across samples,
• supported by at least -O,--call-min-reads-one-samples [3] in a sample,
• supported by at least -S,--call-min-read-per-one-sample [1] percent of reads in a sample,
• supported by at least -B,--call-min-bnd-reads-all-samples [2] reads total across samples for BNDs,
• supported by at least -P,--call-min-reads-per-strand-all-samples [20] reads per strand total across samples,
• assigned a non-reference genotype in at least one sample; a sample is assigned a non-reference genotype for a variant if at least --gt-min-reads [1] reads support the variant.

For CCS input, using the --ccs mode in pbsv call, thresholds are relaxed to -A 3 -B 2 -O 3 -S 0 -P 10.

Filtering

The VCF filter column is

1. PASS
2. NearReferenceGap: variant is near (< --filter-near-reference-gap [1K]) from a gap (run of >= 50 Ns in the reference assembly)
3. Decoy: variant involves a decoy sequence, where the chromosome name contains decoy, hs37d5, or hs38d1
4. NearContigEnd: variant is near (< --filter-near-contig-end [1K]) from a contig end
5. InsufficientStrandEvidence: variant is not supported by at least (--call-min-reads-per-strand-all-samples [1]) reads in forward and reverse orientation
6. NotFullySpanned: duplication variant does not have any fully spanning reads

Performance benchmarks

Using the publicly available HG002 15kb CCS dataset, we are tracking pbsv performance with respect to the genome in a bottle annotation version 0.6.

Step-by-step reproducibility and benchmarks can be found at github.com/PacificBiosciences/sv-benchmark.

FAQ

To where do I report bugs and ask questions about the pre-release version of pbsv?

Please refer to our official pbbioconda page to report bugs and ask questions.

Where can I find an example dataset to try pbsv?

For Genome in a Bottle sample HG002:

• 10-fold CLR coverage
• 28-fold CCS coverage

The binary does not work on my linux system!

If you get Illegal instruction upon execution of pbsv, then your CPU is not supported. A modern (post-2008) CPU with support for SSE4.1 instructions is required.

Why do I have to use --median-filter for CLR data?

I'm not sure how deep your knowledge goes for the PacBio technology. You might know that each ZMW contains a polymerase that sequences one SMRTbell, which consists of your piece of DNA, we call insert, and PacBio's hairpin adapters. This SMRTbell is being processed once the acquisition starts and ends when either the polymerase dies or the acquisition ends. Depending how large your insert and how long the movie time is, it might happen that you read the same molecule multiple times. In this case, a ZMW has multiple subreads (also called CLR), originating from the same insert, from the forward and reverse strands. If we were to align all subreads to the genome and call SVs, it will happen that we call SVs that have support from multiple subreads, but all of them are from the same ZMW. In NGS you would call that PCR duplicates or some would refer to them as technical replicates. One does not want to account evidence more than once per ZMW aka per molecule. The median filter picks one subread per ZMW, to be precise the subread of median length, to have exactly one evidence per molecule. If you don't use it, you will get false positive SV calls and the genotypes will be wrong.

Why does the VCF contain no ambiguous IUPAC REF codes?

Starting with version 2.4.0, pbsv constrains all bases in the reference to be A,C,T,G,N. Bases unknown (e.g. IUPAC) will be automatically converted to N. Toggling --preserve-non-acgt will retain non A,C,T,G,N bases in the reference.

Full Changelog

• 2.9.0:

  • Update --min-mapq to 5
  • Allow spaces in sample name

• 2.8.0:

  • Deprecate copy number variation calling.

• 2.7.1:

  • Fix END info field for inversions.

• 2.7.0:

  • Add HiFi preset to 'discover'
  • Change default values for --call-min-reads-all-samples and --call-min-reads-one-sample options
  • Handle '#' header lines in BED input files.

• 2.6.2:

  • Fix: Truncate long svsig lines

• 2.4.1:

  • Increased specificity for BNDs
  • Added field to svsig file
  • Modified depth logic for edge case

• 2.4.0:

  • Public release in SMRT Link 10.0.0
  • Ensure identical output for one giant or multiple small svsig files as inputs
  • Add --preserve-non-acgt
  • Fix a problem where signatures in call were dropped before target window
  • Support non-spanning inversions
  • Use incremental VCF ids

• 2.3.0:

  • Public release in SMRT Link 9.0.0
  • New CLI UX
  • Change svsig compression to bgzip to enable indexing via tabix

• 2.2.2:

  • Public release in SMRT Link 8.0.0
  • Fix --sample, used to override input type

• 2.2.1:

  • Public release in SMRT Link 7.0.0
  • Add —call-min-bnd-reads-all-samples

• 2.2.0:

  • Add duplications and copy number variations
  • Improved sensitivity for larger insertions and deletions
  • Simplified parameters and relaxed inversion calling criteria
  • Add --ccs support for pbsv call
  • ~5x pbsv discover speedup and minimize memory footprint
  • Add maximum variant limits for duplications and insertions, the latter improves memory footprint of pbsv call

• 2.1.1:

  • Improve error output if reference and svsig contigs do not match
  • Add IMPRECISE to VCF header
  • Separate filter with semicolon

• 2.1.0:

  • Algorithmic improvements to increase recall and sensitivity across all SV lengths
  • Add SAC Stranded Allel Counts for subread input
  • Remove pbsv fasta and rely on pbmm2 input
  • Add MATEDIST, distance between two breakends of a translocation if the two breakends are on the same contig
  • Fix VCF POS/REF/ALT and sorting to pass GATK ValidateVariants and allow VCF bgzip and tabix
  • Allow multiple sample names and multiple BAM inputs for pbsv discover
  • Allow length KMG suffix
  • Flag NearContigEnd
  • Allow breakend calling from single chromosome svsig.gz
  • Allow custom --annotations for known sequence types
  • Loop termination fix to properly fix the issue in the 2.0.2 patch
  • Decrease memory overhead in pbsv discover

• 2.0.2: Fix rare pbsv call abort, because of missing coverage

• 2.0.0: Drop RC for conda release

• 2.0.0-RC2: First public release candidate for SMRT Link 6.0.0

DISCLAIMER

THIS WEBSITE AND CONTENT AND ALL SITE-RELATED SERVICES, INCLUDING ANY DATA, ARE PROVIDED "AS IS," WITH ALL FAULTS, WITH NO REPRESENTATIONS OR WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY, SATISFACTORY QUALITY, NON-INFRINGEMENT OR FITNESS FOR A PARTICULAR PURPOSE. YOU ASSUME TOTAL RESPONSIBILITY AND RISK FOR YOUR USE OF THIS SITE, ALL SITE-RELATED SERVICES, AND ANY THIRD PARTY WEBSITES OR APPLICATIONS. NO ORAL OR WRITTEN INFORMATION OR ADVICE SHALL CREATE A WARRANTY OF ANY KIND. ANY REFERENCES TO SPECIFIC PRODUCTS OR SERVICES ON THE WEBSITES DO NOT CONSTITUTE OR IMPLY A RECOMMENDATION OR ENDORSEMENT BY PACIFIC BIOSCIENCES.

FOR RESEARCH USE ONLY. NOT FOR USE IN DIAGNOSTICS PROCEDURES.