Structural variation and indel detection by local assembly
Switch branches/tags
Clone or download

Build Status

SvABA - Structural variation and indel analysis by assembly

[Pronounced sah-bah. This project was formerly "Snowman"]

License: GNU GPLv3

Table of contents


We recommend compiling with GCC-4.8 or greater. We have successfully compiled on RHEL6, CentOS with GCC-4.8, 4.9 and 5.1, and MacOSX with Clang (Apple LLVM version 7.0.2 (clang-700.1.81))

git clone --recursive
cd svaba
make install

## QUICK START (eg run tumor / normal on Chr22, with 4 cores)
bin/svaba -t tumor.bam -n normal.bam -k 22 -G ref.fa -a test_id -p -4

## get help
svaba --help
svaba run --help

SvABA uses the SeqLib API for BAM access, BWA-MEM alignments, interval trees and operations, and several other auxillary operations.


SvABA is a method for detecting structural variants in sequencing data using genome-wide local assembly. Under the hood, SvABA uses a custom implementation of SGA (String Graph Assembler) by Jared Simpson, and BWA-MEM by Heng Li. Contigs are assembled for every 25kb window (with some small overlap) for every region in the genome. The default is to use only clipped, discordant, unmapped and indel reads, although this can be customized to any set of reads at the command line using VariantBam rules. These contigs are then immediately aligned to the reference with BWA-MEM and parsed to identify variants. Sequencing reads are then realigned to the contigs with BWA-MEM, and variants are scored by their read support.

SvABA is currently configured to provide indel and rearrangement calls (and anything "in between"). It can jointly call any number of BAM/CRAM/SAM files, and has built-in support for case-control experiments (e.g. tumor/normal, or trios or quads). In case/control mode, any number of cases and controls (but min of 1 case) can be input, and will jointly assemble all sequences together. If both a case and control are present, variants are output separately in "somatic" and "germline" VCFs. If only a single BAM is present (input with the -t flag), a single SV and a single indel VCF will be emitted.

A BWA-MEM index reference genome must also be supplied with -G.

Output file description


Raw, unfiltered variants. This file is parsed at the end to produce the VCF files. With the bps.txt.gz, one can define a new set of filteirng criteria (depending on sensitivity/specificity needs) using svaba refilter.


All assembly contigs as aligned to the reference with BWA-MEM. Note that this is an unsorted file. To view in IGV, it must be first sorted and indexed (e.g. samtools sort -m 8G id.contigs.bam id.sort && samtools index id.sort.bam)


Information on all clusters of discordant reads identified with 2+ reads.


Log file giving run-time information, including CPU and Wall time (and how it was partitioned among the tasks), number of reads retrieved and contigs assembled for each region.


An ASCII plot of variant-supporting contigs and the BWA-MEM alignment of reads to the contigs. This file is incredibly useful for debugging and visually inspecting the exact information SvABA saw when it performed the variant-calling. This file is typically quite large. The recommended usage is to identify the contig name of your variant of interest first from the VCF file (SCTG=contig_name). Then do gunzip -c id.alignment.txt.gz | grep contig_name > plot.txt. It is highly recommended that you view in a text editor with line truncation turned OFF, so as to not jumble the alignments.


VCF of rearrangements and indels parsed from bps.txt.gz and with a somatic_score == 1 (somatic) or 0 (germline) and quality == PASS. NOTE that the cutoff for rearrangement vs indel is taken from BWA-MEM, whether it produces a single gapped-alignment or two separate alignments. This is an arbitrary cutoff, just as there is no clear consensus distinction between what constitutes an "indel" and a "structural variant". The unfiltered VCF files include non-PASS variants.

Filtering and Refiltering

SvABA performs a series of log-likelihood calculations for each variant. The purpose is to first classify a variant as real vs artifact, and then to determine if the variant is somatic or germline. These log-likelihoods are output in the VCF and bps.txt.gz file and described here:

  • LOD (LO) - Log of the odds that variant is real vs artifact. For indels, the likelihood of an artifact read is proportional to the length of local repeats (repeating units up to 5 long per unit)
  • LR - Log of the odds that the variant has allelic fraction (AF) of 0 or >=0.5. This is used for somatic vs germline classification
  • SL - Scaled LOD. LOD scores is heuristically scaled as: (min(Mapping quality #1, Mapping quality #2) - 2 * NM) / 60 * LOD

SvABA can refilter the bps.txt.gz file to produce new VCFs with different stringency cutoffs. To run, the following are required:

  • -b - a BAM from the original run, which is used just for its header
  • -i - input bps.txt.gz file

Examples and recipes

Whole genome somatic SV and indel detection

wget "" ## get a DBSNP known indel file
CORES=8 ## set any number of cores
## -a is any string you like, which gives the run a unique ID
svaba run -t $TUM_BAM -n $NORM_BAM -p $CORES -D $DBSNP -a somatic_run -G $REF

Whole genome germline SV and indel detection

## Set -I to not do mate-region lookup if mates are mapped to different chromosome.
##   This is appropriate for germline-analysis, where we don't have a built-in control
##   to against mapping artifacts, and we don't want to get bogged down with mate-pair
##   lookups.
## Set -L to 6 which means that 6 or more mate reads must be clustered to 
##   trigger a mate lookup. This also reduces spurious lookups as above, and is more 
##   appropriate the expected ALT counts found in a germline sample 
##   (as opposed to impure, subclonal events in cancer that may have few discordant reads).
svaba run -t $GERMLINE_BAM -p $CORES -L 6 -I -a germline_run -G $REF

Targeted detection

## eg targets.bed is a set of exome capture regions
svaba run -t $BAM -k targets.bed -a exome_cap -G $REF

Targeted local assembly

## -k can be a chromosome, a samtools/IGV style string 
##     (e.g. 1:1,000,000-2,000,000), or a BED file
svaba run -t $TUM_BAM -n $NORM_BAM -p $CORES -k $k  -a TP53 -G $REF

Assemble all reads

## default behavior is just assemble clipped/discordant/unmapped/gapped reads
## This can be overridden with -r all flag
svaba run -t $BAM -r all -G $REF

Snapshot of where svaba run is currently operating

tail somatic_run.log

Debug a local assembly and produce the assembly graph

svaba run -t $BAM -a local_test -k $k --write-asqg

## plot the graph

View all of the ASCII alignments

## Make a read-only and no-line-wrapping version of emacs.
## Very useful for *.alignments.txt.gz files
function ev { 
  emacs $1 --eval '(setq buffer-read-only t)' -nw --eval '(setq truncate-lines t)';
ev somatic_run.alignments.txt.gz 

View a particular contig with read to contig alignments

gunzip -c somatic_run.alignments.txt.gz | grep c_1_123456789_123476789 > c_1_123456789_123476789.alignments.txt
ev c_1_123456789_123476789.alignments.txt

Make a function to sort and index contigs

function sai() {
  if [[ -f $1.contigs.bam ]]; then
     samtools sort -m 4G $1.contigs.bam -o $1.contigs.sort.bam
     mv $1.contigs.sort.bam $1.contigs.bam
     samtools index $1.contigs.bam
## for example, for somatic_run.contigs.bam:
sai somatic_run


SvABA is developed and maintained by Jeremiah Wala ( -- Rameen Berkoukhim lab -- Dana Farber Cancer Institute, Boston, MA.

This project was developed in collaboration with the Cancer Genome Analysis team at the Broad Institute. Particular thanks to:

Additional thanks to Jared Simpson for SGA, Heng Li for htslib and BWA, and for the other developers whose
code contributed to SeqLib.