Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit

Switching to term exclude-list where possible.

Git stats


Failed to load latest commit information.

Release Language

GRIDSS - the Genomic Rearrangement IDentification Software Suite

GRIDSS is a module software suite containing tools useful for the detection of genomic rearrangements. GRIDSS includes a genome-wide break-end assembler, as well as a structural variation caller for Illumina sequencing data. GRIDSS calls variants based on alignment-guided positional de Bruijn graph genome-wide break-end assembly, split read, and read pair evidence.

GRIDSS makes extensive use of the standard tags defined by SAM specifications. Due to the modular design, any step (such as split read identification) can be replaced by another implementation that also outputs using the standard tags. It is hoped that GRIDSS can serve as an exemplar modular structural variant pipeline designed for interoperability with other tools.

If you have any trouble running GRIDSS, please raise an issue using the Issues tab above. Based on feedback from users, a user guide will be produced outlining common workflows, pitfalls, and use cases.

Click here to download GRIDSS

Detailed documentation is being developed here

Quick start guide

Use the quick start guide to get up and running quickly


If you use GRIDSS, please cite:

Cameron DL, Schröder J, Penington JS, Do H, Molania R, Dobrovic A, Speed TP, Papenfuss AT. GRIDSS: sensitive and specific genomic rearrangement detection using positional de Bruijn graph assembly. Genome Research, 2017 Dec;27(12):2050-2060.

Cameron DL, Baber J, Shale C, Valle-Inclan JE, Besselink N, van Hoeck A, Janssen R, Cuppen E, Priestley P, Papenfuss AT. GRIDSS2: comprehensive characterisation of somatic structural variation using single breakend variants and structural variant phasing. Genome Biol. 2021 Jul 12;22(1):202.

The recommended minimal citations are:


To run GRIDSS the following must be installed:

  • java 1.8 or later
  • R 4.0 or later
    • gridss_somatic_filter and gridss_extract_overlapping_fragments require the following R libraries:
      • argparser
      • tidyverse
      • stringdist
      • testthat
      • stringr
      • StructuralVariantAnnotation
      • rtracklayer
      • BSgenome package for your reference genome (optional)
  • samtools 1.13 or later
  • bwa
  • bash
  • getopt(1) (part of util-linux)

To run VIRUSBreakend, kraken2, or repeatmasker annotations, the following additional software must be installed:

  • kraken2
    • Note that virusbreakend-build requires all kraken2-build dependencies
  • RepeatMasker
  • bcftools

Building GRIDSS

GRIDSS is mostly written in Java thus local building is not required. Just download the latest release and ensure you have the Pre-requistes installed.

If you wish to contribute to GRIDSS development, it can be built from source using maven with mvn package.

A prebuilt docker image is available as gridss/gridss:latest so building a docker image yourself is not necessary. If you do wish to build the docker image yourself, ensure that you perform a recursive git checkout. If you don't, you'll be missing htslib and gridsstools won't compile.

Building gridsstools

Some performance-critical steps are implemented in C using htslib. A precompiled version of gridsstools for linux is included as part of GRIDSS releases. If this precompiled version does not run on your system you will need to build it from source.

To build gridsstools from source run the following:

git clone --recurse-submodules
cd gridss/src/main/c/gridsstools/htslib/
autoreconf -i && ./configure && make
cd ..
autoreconf -i && ./configure && make all

Conda issues

Compiling with a conda environment active is likely to cause problems such as undefined reference to 'libdeflate_crc32'. This happens when the conda environment includes copies of the libraries used by htslib (z m bz2 lzma curl crypto pthread) without also including the headers for the libraries. This causes gridsstools to compile against the system headers, but link against the conda libraries, hence the error.

Run either conda install htslib or conda deactivate if you have problems compiling gridsstools in a conda environment.


Pre-compiled binaries are available at GRIDSS invokes external tools at multiple points during processing. By default this is bwa mem, but can be configured to use bowtie2 or another aligner.

The following programs are included in GRIDSS releases:

program description
gridss GRIDSS assembler and structural variant caller. Use this to generate a GRIDSS SV VCF.
gridss_extract_overlapping_fragments Extracts all alignments for read pairs with at least one aligment overlapping set of regions of interest. Correctly handles supplementary alignments. Use this script to extract reads of interest for targeted GRIDSS variant calling.
gridss_annotate_vcf_repeatmasker Annotates breakpoint and single breakend inserted sequences with the RepeatMasker classification of the sequence.
gridss_annotate_vcf_kraken2 Annotates breakpoint and single breakend inserted sequences with the Kraken2 classification of the sequence.
virusbreakend See VIRUSBreakend README
virusbreakend-build See VIRUSBreakend README
gridss_somatic_filter Somatic filtering script. Note that this has an equivalent java implementation in GRIPSS.
gridsstools C/htslib implementation of performance-critical steps. Currently used by virusbreakend and gridss_extract_overlapping_fragments

gridss command-line arguments

Usage: gridss --reference <reference.fa> --output <output.vcf.gz> --assembly <assembly.bam> [--threads n] [--jar gridss.jar] [--workingdir <directory>] [--jvmheap 30g] [--blacklist <exclude_list.bed>] [--steps All|PreProcess|Assemble|Call] [--configuration] [--maxcoverage 50000] [--labels input1,input2,...] input1.bam [input2.bam [...]]
required argument description
-o, --output output VCF
-r, --reference reference genome to use. Must have a .fai index file and a bwa index
optional argument description
-a, --assembly location of the GRIDSS assembly BAM. This file will be created by GRIDSS. The default filename adds a .assembly.bam suffix to the output file.
-t, --threads number of threads to use. Defaults to 8 cores.
-j, --jar location of GRIDSS jar. Can also be specified using the GRIDSS_JAR environment variable.
-b/--blacklist BED file containing regions to ignore. The ENCODE DAC blacklist is recommended for hg19. (Optional)
--jvmheap size of JVM heap for the high-memory part of the assembly and variant calling. Defaults to 30g to ensure GRIDSS runs on cloud instances with 32gb memory.
--otherjvmheap size of JVM heap for everything else. Useful to prevent java out of memory errors when using large (>4Gb) reference genomes. Note that some parts of assembly and variant calling use this heap size so if you get an OutOfMemory error during during these steps even after increasing jvmheap, you may need to increase otherjvmheap as well. (Default: 4gb)
--maxcoverage maximum coverage. Regions with coverage in excess of this are ignored. (Default: 50000)
--labels comma separated labels to use in the output VCF for the input files. Must have same number of entries as there are input files. Input files with the same label are aggregated (useful for multiple sequencing runs of the same sample). Labels default to input filenames, unless a single read group with a non-empty sample name exists in which case the read group sample name is used (which can be disabled by "useReadGroupSampleNameCategoryLabel=false" in the configuration file). If labels are specified, they must be specified for all input files.
--steps processing steps to run. Defaults to all steps. Multiple steps are specified using comma separators. Available steps are preprocess,assemble,call. Useful to improve parallelisation on a cluster as preprocess of each input file is independent, and can be performed in parallel, and has lower memory requirements than the assembly step.
--jobindex zero-based index of this assembly job node. Used to spread GRIDSS assembly across multiple compute nodes. Use only with -s assemble. Once all jobs have completed, a -s assemble or -s all job should be run to gather the results together.
--jobnodes total number of assembly jobs scheduled.

The following additional optional arguments may be useful if GRIDSS fails to run in your environment, or you want to run with non-standard parameters.

argument description
-c, --configuration configuration file use to override default GRIDSS settings
--externalaligner use the system version of bwa instead of the in-process version packaged with GRIDSS
--picardoptions additional standard Picard command line options. Useful options include VALIDATION_STRINGENCY=LENIENT and COMPRESSION_LEVEL=0. See
--useproperpair use SAM 'proper pair' flag to determine whether a read pair is discordant. Default: use library fragment size distribution to determine read pair concordance
--concordantreadpairdistribution portion of read pairs distribution considered concordantly mapped. Default: 0.995
--keepTempFiles keep intermediate files. Not recommended except for debugging due to the high disk usage.
--nojni do not use JNI native code acceleration libraries (snappy, GKL, ssw, bwa).

Warning: the somatic filtering script treats the first bam file as the matched normal, and all subsequent as tumour samples. If you are doing somatic calling, it is strongly recommended to follow this convention.

gridss steps

The following GRIDSS steps can be specified:

step description
setupreference Once-off setup generating additional files in the same directory as the reference. WARNING: multiple instances of GRIDSS attempting to perform setupreference at the same time will result in file corruption. Make sure these files are generated before running parallel GRIDSS jobs.
preprocess Pre-process input BAM files. Can be run per input file.
assemble Perform GRIDSS breakend assembly. Can split up across multiple nodes using --jobindex and --jobnodes.
call Perform variant calling.
all Run all steps (Default)

At present, command line valiation is performed independently of which steps are run. When splitting GRIDSS into multiple cluster jobs, the same command line parameters should be specified for every job except for:

  • --input on preprocess jobs (one input per file)
  • --jobindex and --jobnodes on assembly jobs


How do I run GRIDSS on multiple samples?

Just specify multiple BAMs on the command line. GRIDSS will perform joint calling and provide a per-BAM breakdown of support.

Should I do joint calling or run each sample individually?

Joint calling should always be used for related samples (e.g. tumour/normal or trio calling). Joint calling will ensure that a common variant near the single-sample threshold of detection will be reliably reported as a shared variant. This is not the case if the calling were done individually. Note that this particular behaviour is not specific to GRIDSS and is common to all variant callers (hence the joint calling support in many callers).

Joint calling allows for sensivity detection of variants that are present subclonally (or at low coverage) that would not be detected if called individually.

GRIDSS performs joint assembly then reports a per-sample breakdown. Joint calling has higher coverage of shared variants thus resulting in more reliable assembly of that variant.

Determining whether two SV calls in two different VCFs are actually the same call is non-trivial. Imprecise calls are especially problematic since the coordinates may differ between the VCF, or a call may be precise in one VCF and not in the other. A good example of why reconciling SV calls is so problematic is the case where call A (chrX:1-99->chrY:1-99) overlaps call B (chrX:50-149->chrY:1-99), call B overlaps call C (chrX:100-199->chrY:1-99), but A does not overlap C at all. Joint calling obviates this step.

How do I perform tumour/normal somatic variant calling?

Jointly call on all samples from the patient. It is strongly recommended that the normal be the first argument as that is what downstream steps expect. For example, gridss ... patientX_normal.bam patientX_primary.bam patientX_met.bam. To filter to somatic calls, use the gridss_somatic_filter script included in the GRIDSS release.

What aligner should I use?

The default of bwa mem is sufficient for most use cases.

Although GRIDSS aims to be aligner agnostic, not all aligners output BAM files suitable for processing by GRIDSS. GRIDSS requires:

  • One alignment per read. Supplementary (split read) alignments are ok, but secondary alignments are not.
    • This means that aligner settings such as the -a option of bwa mem and the -k and -a options of bowtie2 are unsuitable.
  • MAPQ to meaningfully follow the SAM specifications. Aligners that do not follow the specifications (e.g. subread) will have worse results.

Options such as the -Y option of bwa mem, or the fact that bowtie2 does not do split read alignment are not problematic as these differences are corrected in the GRIDSS preprocessing step.

How do I tell GRIDSS multiple BAMs are from the same sample?

Use the --labels command line option. Eg: --labels sample1,sample1,sample2 sample1_library1.bam sample1_library2.bam sample2.bam

Why are there ALT alleles with . in the output?

This is the VCF notation for single breakend variant calls. See section 5.4.9 of the specifications document. These calls indicate that a breakpoint was found at this location but the partner location could not be unambiguously determined.

How do I get output like the GRIDSS PURPLE LINX figures?

Run the integrated GRIDSS PURPLE LINX pipeline script or the docker image gridss/gridss-purple-linx:latest.

How do I do RepeatMasker annotation of breakend sequences?

Run gridss_annotate_vcf_repeatmasker on the GRIDSS output.

How do I do viral annotation?

Use VIRUSBreakend for viral annotations. See the VIRUSBreakend README for more details.

What does gridss_somatic_filter actually do?

See documentation at

The Hartwig Medical Foundation has reimplemented gridss_somatic_filter in Java as GRIPSS. GRIPSS is much faster, has additional features, and is the recommended tool for somatic filtering of GRIDSS output.

How do I create the panel of normals required by gridss_somatic_filter?

If you are using hg19 or hg38, then a PON based on Dutch samples is available from Make sure the reference you are using and the PON both use the same chromosome notation or nothing will get filtered (e.g. 1 vs chr1). If these are not appropriate, you'll need to create your own using gridss.GeneratePonBedpe

Here is an example that generates a PON from every VCF in the current directory:

mkdir -p pondir
java -Xmx8g \
	-cp ~/dev/gridss/target/gridss-2.10.2-gridss-jar-with-dependencies.jar \
	gridss.GeneratePonBedpe \
	$(ls -1 *.vcf.gz | awk ' { print "INPUT=" $0 }' | head -$n) \
	O=pondir/gridss_pon_breakpoint.bedpe \
	SBO=pondir/gridss_pon_single_breakend.bed \

The score fields of the bedpe/bed files is the count of the number of samples that variant was found in. I recommended filtering these output files to only variants found in 3+ samples.

Note that gridss_somatic_filter requires the files to be named gridss_pon_breakpoint.bedpe and gridss_pon_single_breakend.bed.

How do I merge PONs?

Merging of PONs is not supported but incremental updates are. Use INPUT_BEDPE and INPUT_BED files to gridss.GeneratePonBedpe. For example, you can add your samples to the Hartwig PONs by pointing INPUT_BEDPE and INPUT_BED to the Hartwig PONs and adding your VCFs as outlined above.

Should I process each input BAM separately or together?

Wherever possible, samples should be processed together. Joint calling enables the detection of low allelic fraction SVs, as well as making the downstream analysis much easier (a single VCF with a breakdown of support per sample is much easier to deal with than multiple VCFs - the matching logic required to determine if two SVs are equivalent is non-trivial).

GRIDSS joint calling has been tested on up 12 samples with ~1000x aggregate coverage. If you have hundreds of samples, joint assembly will likely be computationally prohibitive and you will need to perform assembly in batches, then merge the results together.

WARNING: multiple instances of GRIDSS generating reference files at the same time will result in file corruption. Make sure setupreference files have been generated before running parallel GRIDSS jobs.

How do I perform assembly on multiple nodes?

GRIDSS preprocessing and assembly can be spread across multiple nodes. Preprocess parallelisation is one job per input file. Assembly parallelisation is one thread per genomic region. To reduce wall times, regions can be distributed across multiple nodes using the --jobindex and --jobnodes parameters.

Here is an example:

gridss -s setupreference # once-off-setup
# in parallel:
gridss -s preprocess input1.bam
gridss -s preprocess input2.bam
gridss -s preprocess input3.bam
gridss -s preprocess input4.bam
# wait for all preprocessing jobs to complete
# Perform assembly parallel (across three nodes in this example)
gridss -s assemble --jobindex 0 --jobnodes 3 -a assembly.bam  input1.bam input2.bam input3.bam input4.bam
gridss -s assemble --jobindex 1 --jobnodes 3 -a assembly.bam input1.bam input2.bam input3.bam input4.bam
gridss -s assemble --jobindex 2 --jobnodes 3 -a assembly.bam input1.bam input2.bam input3.bam input4.bam
# wait for all assembly jobs to complete
# Gather the assembly results togther. This job is essentially a file copy and completes very quicky
gridss -s assemble -a assembly.bam input1.bam input2.bam input3.bam input4.bam
# perform variant calling
gridss -s call  -a assembly.bam input1.bam input1.bam input2.bam input3.bam input4.bam

How do I perform batched assembly?

If you can perform joint assembly, do so. If you run out of memory when performing joint assembly, or hit regularly encounter assembly timeouts, batched assembly is required. You are likely to encounter these issues at, or above, 1000x coverage.

To perform batched assembly, run the GRIDSS assemble step multiple times each with only a subset of the input files If you use input labels, all inputs with the same label must be in the same batch.

Here is an example:

gridss -s setupreference # once-off-setup
gridss -s preprocess input1.bam
gridss -s preprocess input2.bam
gridss -s preprocess input3.bam
gridss -s preprocess input4.bam
gridss -s assemble -a assembly12.bam input1.bam input2.bam
gridss -s assemble -a assembly34.bam input3.bam input4.bam
gridss -s call -a assembly12.bam -a assembly34.bam input1.bam input2.bam input3.bam input4.bam

Related samples should always be assembled together in the same batch.

I encountered an error. What should I do?

  • Check the bottom of this page for commonly encountered errors and their solutions

How many threads should I use?

GRIDSS has been optimised to run on a 8core/32gb cloud compute node.

If scaling above 8 cores, it is recommended to run multiple GRIDSS assembly processes and use the --jobindex and --jobnodes parameters. with each job allocated 8 cores/32gb.

How much memory should I give GRIDSS?

GRIDSS has been optimised to run on a 8core/32gb cloud compute node.

At least 4GB + 2GB per thread. It is recommended to run GRIDSS with max heap memory (-Xmx) of 8GB for single-threaded operation (WORKER_THREADS=1), 16GB for multi-core desktop operation, and 31GB for server operation. Note that due to Java's use of Compressed Oops, specifying a max heap size of between 32-48GB effectively reduces the memory available to GRIDSS so is strongly discouraged.

Why does GRIDSS use more CPU than the limit specified with --threads?

GRIDSS has been optimised to run on a 8core/32gb cloud compute node.

--threads specifies the size of the worker thread pool. IO, BAM decompression, and parsing are in their own thread pool which is not part of the worker thread pool. The pre-processing, variant calling, and annotation steps also perform some work that is executed in dedicated threads independent of the worker thread pool. Combined, this approach means that max CPU utilisation can exceed the thread count specified.

Asynchronous IO defaults can be changed by editing the jvm_args argument in gridss.

Should I include alt contigs in the reference?

GRIDSS relies on the aligner to determine the mapping location and quality of assembly contigs and to identify split read from soft clipped reads. GRIDSS considers alignments with low mapping quality (default mapq <= 10) to not be uniquely aligned and treats them as unaligned. If alt contigs are included in an aligner that is not alt-aware then hits to sequences that are in both the primary reference contigs and the alt contigs will be given a low mapq by the aligner and will be treated as unaligned by GRIDSS. By default, GRIDSS uses bwa mem for alignment so by including the bwa alt contig definition file, reads from regions with alt homology will be preferentially aligned to the reference with a correspondingly improved mapq. Whether or not to include alt contigs depends on what sort of downstream analysis you intend to perform and how your intend to handle structural variants involving alt contigs. That said, if your reference genome includes alt contigs and a bwa alt contig definition file is available for your genome, you should use it.

How do I process only my region of interest?

Extract all fragments overlapping your region of interest, using gridss_extract_overlapping_fragments then run gridss on the subset bam. gridss_extract_overlapping_fragments is almost identical to filtering using samtools view except that it extracts alignment records for any fragment overlapping a region of interest (ie mate reads and supplementary alignments). That is, all records with read names matching the read name of an alignment overlapping any region of interest.

How do I use GRIDSS to validate the calls from another caller?

  • Convert your calls to a BED file containing the start and end positions of all SVs called by your other caller.

  • Expand intervals by at least 10kbp. Too small a window will have a negative impact on GRIDSS QUAL scores (since they're emperically weighted, taking only regions with soft clipped reads will cause GRIDSS to massively downweight soft clips when scoring). Alternatively, for unbiased scoring, run gridss.CollectGridssMetrics on your input file and rename the .gridss.working directory to the name of your targeted bam file to enable the targeted bam to use the full bam metrics.

  • Process as per the region of interest processing outlined above.

Why does GRIDSS use centre-aligment?

'Normalised' SNV/indel calls are left-aligned. Why does GRIDSS not use this convention? There are two reasons:

For ++ or -- breakpoints, left-aligning the lower breakend will force right-alignment of the upper breakend. Similarly for right-alignment of the lower. This means that it is impossible to univerally left-align, or right-align breakpoints without resulting in an incorrect nominal call position. Centre-alignment is the option that does not cause this problem ( technically speaking, you still have an off-by-one problem for odd-length homology but that's less problematic).

Secondly, using left or right alignment for imprecise call will result in the nominal call being at the edge of the confidence interval bounds. Centre-aligning imprecise calls makes sense as it is (usually) the centre position that is the most likely to be correct.

What does gridss_annotate_kraken2 output?

gridss_annotate_kraken2 adds Kraken2 classifications to single breakend and breakpoint inserted sequences. The NCBI taxonomy ID for the inserted sequences is in the INSTAXID INFO field.

Why are all calls BND?

In VCF version 4.3 and earlier, the meaning of DEL, and DUP is ambiguous as it is unclear if the claim being made is a breakpoint claim, a copy number claim, or both. A DUP reported by a copy number caller has a very different meaning than a DUP reported by a breakpoint caller. A copy number DUP indicates there is at least one additional copy of the duplication region but makes no claim regarding where that extra copy is location. A breakpoint DUP indicates that the start and end of the 'duplicated' region are connected but makes no claim regarding the actual copy number of the duplicated region. A true DUP or DEL requires both a copy number change, a breakpoint of the correct orientation, and to not form part of a larger rearrangement.

Since GRIDSS is fundamentally a breakpoint (and single breakend) caller, we have made the choice to report all variants in BND notation to make it explicit exactly what it is that GRIDSS is detecting. For users only interested the analysis of simple events and for which the incorrect interpretation of complex rearrangements is acceptable, the example/simple-event-annotation.R will annotate GRIDSS calls with SIMPLE_TYPE and SVLEN fields.

Our approach of explicitly seperating the detection of the rearrangement building blocks (copy number, breakpoint, and single breakends) from the rearrangement events will be codified in the upcoming version 4.4 of the VCF specifications through the incorporation of a SVCLAIM field (to remove the ambiguity of DEL and DUP calls), and EVENT/EVENTTYPE fields (for linking of related variant calls into higher-order events such as chromothripsis.

For an example of why an event-based model that seperates detection from intepretation, see the LINX readme and preprint, and complex event visualisation examples.


GRIDSS takes a modular approach and the GRIDSS jar consists of a collection of separate tools. Each tool in the GRIDSS pipeline can be run independently. The following data flow diagram gives an overview of the GRIDSS pipeline used when running gridss.

GRIDSS data flow diagram


This tool runs every step in the variant calling pipeline. This entry point has been superceeded by gridss but is retained for backward compatibility with existing pipeline. gridss is preferred as it has lower peak memory usage and is slightly faster due to the use of samtools for sorting instead of htsjdk.


GRIDSS requires a number of input library metrics to be calculated. The metrics calculations programs are invoked in the same manner as Picard tools metrics. This program functions similarly to Picard tools CollectMultipleMetrics but, by default, only extracts the metrics required by GRIDSS.


Collects metrics regarding the size and type of read alignment CIGAR elements.


Collects generic library-level metrics such as the read and read pair mapping rates.


Collects metrics regarding distribution of read mapping quality scores.


Collects metrics regarding presence of SAM tags.


Extracts the subset of reads that provide potential support for structural variation. These reads fall into one or more of the following classes:

  • Reads aligned with an insertion or deletion in the read alignment
  • Soft clipped reads
  • Split reads (identified by the SA SAM Tag)
  • Discordant read pairs
  • Read pairs with only 1 read mapped
  • Unmapped reads


Identifies split reads by iterative realignment of soft clipped bases with a NGS aligner. By default, bwa mem is used for read alignment. The GRIDSS pipeline runs this on all input files, as well as the breakend assembly contigs generated by AssembleBreakends.


This step recomputes redundant information to ensure data consistency and writes additional standards-defined SAM tags by ppulates SAM tags required in downstream steps, softening hard clips, and correcting data inconsistencies in the input files. In particular, AssembleBreakends requires the R2 tag to be populated. Since R2 is populated from the mate sequence and softening hard clips requires all split read alignment records, this step requires all records with the same read name to be adjacent in the BAM file (ie read name grouped sort order). Data inconsistencies corrected include:

  • NM tag not matching edit distance to reference
  • Mate CIGAR and alignment field not matching actual alignment and CIGAR of mate (e.g. GATK indel realignment does not update mate information when realigning reads)
  • SA tag not matching supplementary alignment
  • Supplementary alignments not duplicated marked (markdup only marks primary alignments)
  • Supplementary alignments incorrectly flagged as secondary alignments

This step can recalculate the following SAM tags:

  • SA (default)
  • NM (default)
  • R2 (default)
  • Q2
  • MC (default)
  • MQ (default)
  • CC
  • CP
  • HI
  • IH
  • FI


Generates breakend assemblies from the input reads. Breakend assemblies are written as synthetic soft clipped reads. Each of these reads corresponds to a breakend assembly contig composed of reads anchored (either directly, or via the mate read) to the location of the breakend assembly.

Breakend contigs composed entirely of discordant read pairs or reads with unmapped mates cannot be uniquely placed as there exists an interval over which the contig is anchored. These alignments are written using a placeholder CIGAR alignment of the form XNX. For example, a breakend contig read with an alignment of 1X50N1X150S represents a 150bp contig in which the breakend is expected to occur at one of the 52 genomic positions given by the placeholder XNX alignment interval.

Breakpoints are identified by running SoftClipsToSplitReads on the breakend assembly contigs. High quality breakpoints are expected to have two independent breakend assemblies (one from each breakend of the breakpoint).


Identifies putative structural variants from the reads providing potential SV support, and the breakend assembly contigs.


Annotates breakpoint calls performing AllocateEvidence, AnnotateReferenceCoverage, AnnotateInexactHomology annotation.


Uniquely allocate reads/read pairs to identified breakpoints by ensuring that for each read/read pair, only the single best alignment is retained (relevant for input files containing multiple mapping locations for each read). Read counts and other summary statistics are collated.


Calculates the number of reads and read pairs providing support for the reference allele at each breakpoint.


Calculates the size of the inexact homology between the reference sequence and the breakpoint sequence. Breakpoints with long inexact homology are possibly due to alignment artifacts causing false positive breakpoint calls between regions of homologous sequence.


Finds potential mapping locations for single breakends and breakpoint insert sequences in the given reference. Used for RepeatMasker annotation, and viral integration detection.


This tool aggregates variants across multiple VCFs and counts the number of samples supporting each variant. Only the first sample per VCF is processed which is useful for generating a panel of normals (PON) from a cohort of cancer samples with matched normals. Output is a bedpe (breakpoint) and bed (single breakend) file suitable for use by gridss_somatic_filter.

Common Parameters

GRIDSS programs have a large number of parameters that can be be adjusted. The default parameter set has been tested with paired-end Illumina data ranging from 2x36bp to 2x250bp and should give a reasonably good result. Command line used parameters are listed below.

OUTPUT (Required)

Variant calling output file. Can be VCF or BCF.


Reference genome FASTA file. GRIDSS requires that the reference genome supplied exactly matches the reference genome of all input files. The reference genome must be in FASTA format and must have a tabix (.fai) index and an index for the NGS aligner (by default bwa). The NGS aligner index prefix must match the reference genome filename. For example, using the default setting against the reference file reference.fa, the following files must be present and readable:

File Description
reference.fa reference genome
reference.fa.fai Tabix index
reference.fa.amb bwa index
reference.fa.ann bwa index
reference.fa.bwt bwa index
reference.fa.pac bwa index bwa index

These can be created using samtools faidx reference.fa and bwa index reference.fa

A .dict sequence dictionary is also required but GRIDSS will automatically create one if not found.

INPUT (Required)

Input libraries. Specify multiple times (i.e. INPUT=file1.bam INPUT=file2.bam INPUT=file3.bam) to process multiple libraries together.

Input files must be coordinate sorted SAM/BAM/CRAM files.

GRIDSS considers all reads in each file to come from a single library.

Input files containing read groups from multiple different libraries should be split into an input file per-library.

The reference genome used for all input files should match the reference genome supplied to GRIDSS.


Labels to allocate inputs. The default label for each input file corresponds to the file name but can be overridden by specifying an INPUT_LABEL for each INPUT. The output for any INPUT files with the same INPUT_LABEL will be merged.

ASSEMBLY (Required)

File to write breakend assemblies to. It is strongly recommended that the assembly filename corresponds to the OUTPUT filename. Using ASSEMBLY=assembly.bam is problematic as (like the INPUT files) the assembly file is relative not to WORKING_DIR, but to the current directory of the calling process. This is likely to result in data corruption when the same assembly file name is used on different data sets (for example, writing assembly.bam to your home directory when running on a cluster).


BED exclude-list of regions to exclude from analysis. The ENCODE DAC blacklist is recommended when aligning against hg19.

Unlike haplotype assemblers such as TIGRA and GATK, GRIDSS does not abort assembly when complex assembly graphs are encountered. Processing of these graphs slows down the assembly process considerably, so if regions such as telomeric and centromeric regions are to be excluded from downstream analysis anyway, assembly of these regions is not required. It is recommended that a exclude-list such as the ENCODE DAC blacklist be used to filter such regions. Inclusion of additional mappability-based exclude-lists is not required as GRIDSS already considers the read mapping quality.


Portion (0.0-1.0) of read pairs to be considered concordant. Concordant read pairs are considered to provide no support for structural variation. Clearing this value will cause GRIDSS to use the 0x02 proper pair SAM flag written by the aligner to determine concordant pairing. Note that some aligners set this flag in a manner inappropriate for SV calling and set the flag for all reads with the expected orientation and strand regardless of the inferred fragment size.


Per input overrides for explicitly specifying fragment size interval to be considered concordant. As with INPUT_LABEL, these must be specified for all input files. Use null to indicate an override is not required for a particular input (e.g. INPUT=autocalc.bam INPUT_MIN_FRAGMENT_SIZE=null INPUT_MAX_FRAGMENT_SIZE=null INPUT=manual.bam INPUT_MIN_FRAGMENT_SIZE=100 INPUT_MAX_FRAGMENT_SIZE=300 )


Number of processing threads to use, including number of threads to use when invoking the aligner. Note that the number of threads spawned by GRIDSS is greater than the number of worker threads due to asynchronous I/O threads thus it is not uncommon to see over 100% CPU usage when WORKER_THREADS=1 as bam compression/decompression is a computationally expensive operation. This parameter defaults to the number of cores available.


Directory to write intermediate results directories. By default, intermediate files for each input or output file are written to a subdirectory in the same directory as the relevant input or output file. If WORKING_DIR is set, all intermediate results are written to subdirectories of the given directory.


This field is a standard Picard tools argument and carries the usual meaning. Temporary files created during processes such as sort are written to this directory.

samjdk defines

GRIDSS uses htsjdk as a SAM/BAM/CRAM/VCF parsing library. The following htsjdk java command-line options are strongly recommended for improved performance:

  • -Dsamjdk.use_async_io_read_samtools=true
  • -Dsamjdk.use_async_io_write_samtools=true
  • -Dsamjdk.use_async_io_write_tribble=true

Due to relatively poor performance of existing Java-based Smith-Waterman alignment packages, GRIDSS incorporates a JNI wrapper to the striped Smith-Waterman alignment library SSW. GRIDSS will attempt to load a precompiled version which is supplied as part of the GRIDSS package (a file will be created in the TMP_DIR when GRIDSS is run). If the precompiled version is not compatible with your linux distribution, or you are running a different operating system, recompilation of the wrapper from source will be required. When recompiling, ensure the correct is loaded using -Djava.library.path, or the LD_LIBRARY_PATH environment variable as per the JNI documentation.

If you have an older CPU that does not support SSE instructions, GRIDSS will terminate with a fatal error when loading the library. Library loading can be disabled by adding -Dsswjni.disable=true to the GRIDSS command line. If cannot be loaded, GRIDSS will fall back to a (50x) slower java implementation which will result in the GRIDSS inexact homology variant annotation step running very slowly.


GRIDSS uses a large number of configurable settings and thresholds which for ease of use are not included as command line arguments. Any of these individual settings can be overriden by specifying a configuration file to use instead. Note that this configuration file uses a different format to the Picard tools-compatible configuration file that is used instead of the standard command-line arguments.

When supplying a custom configuration, GRIDSS will use the overriding settings for all properties specified and fall back to the default for all properties that have not been overridden. Details on the meaning of each parameter can be found in the javadoc documentation of the classes.


GRIDSS is fundamentally a structural variation breakpoint caller. Variants are output as VCF breakends. Each call is a breakpoint consisting of two breakends, one from location A to location B, and a reciprocal record from location B back to A. Note that although each record fully defines the call, the VCF format requires both breakends to be written as separate records.

To assist in downstream analysis, the StructuralVariantAnnotation R BioConductor package is strongly recommended. Operations such as variant filter, annotation and exporting to other formats such as BEDPE can be easily accomplished using this package in conjuction with the BioConductor annotation packages.

Quality score

GRIDSS calculates quality scores according to the model outlined in the paper. As GRIDSS does not yet perform multiple test correction or score recalibration, QUAL scores are vastly overestimated for all variants. As a rule of thumb, variants that have QUAL >= 1000 and have assemblies from both sides of the breakpoint (AS > 0 & RAS > 0) are considered of high quality, variants with QUAL >= 500 but that can only be assembled from one breakend (AS > 0 | RAS > 0) are considered of intermediate quality, and variants with low QUAL score or lack any supporting assemblies are considered to be of low quality.

Non-standard INFO fields

GRIDSS writes a number of non-standard VCF fields. These fields are described in the VCF header.

BED/BEDPE export

The recommended way to convert GRIDSS output to BEDPE is via the BioConductor StructuralVariantAnntotation package.


vcf = readVcf("gridss.vcf")

# Export breakpoints to BEDPE
bpgr = breakpointRanges(vcf)
# TODO: add your event filtering here. The default GRIDSS output is very verbose/sensitive.
write.table(breakpointgr2bedpe(bpgr), file="gridss_breakpoints.bedpe", sep="\t", quote=FALSE, col.names=FALSE)
# Export single breakends to BED
begr = breakendRanges(vcf)
# TODO: add your event filtering here. The default GRIDSS output is very verbose/sensitive.
begr$score = begr$QUAL
export(begr, con="gridss_single_breakends.bed")

Visualisation of results

When performing downstream analysis on variant calls, it can be immensely useful to be able to inspect the reads that the variant caller used to make the variant calls. As part of the GRIDSS pipeline, the following intermediate files are generated:


The inputsv file contains all reads GRIDSS considered as providing putative support for any potential breakpoint, including breakpoints of such low quality that GRIDSS did not make any call. This file includes all soft clipped, indel-containing, and split reads, as well as all discordant read pairs and pairs with only one read mapped.

Split reads can be identified by the presence of a SA SAM tag.

GRIDSS treats breakend assemblies as synthetic soft clipped read alignments thus assemblies are displayed in the same manner as soft clipped/split reads.

  • No SA tags indicates a breakpoint could not be unambiguously identified from the breakend contig.
  • The source of the breakend contig can be identified by the read without the "Supplementary alignment" 0x800 SAM flag set.
  • More than two alignments for a breakend contig indicates the contig spans a complex event involving multiple breakpoints.

Intermediate Files

GRIDSS writes a large number of intermediate files. If rerunning GRIDSS with different parameters on the same input, all intermediate files must be deleted, or a different WORKING_DIR specified. All intermediate files are written to the WORKING_DIR directory tree, with the exception of temporary sort buffers which are written to TMP_DIR and automatically deleted at the conclusion of the sort operation.

File Description
gridss.* Temporary intermediate file
gridss.lock.breakend.vcf Lock directory to ensure that only one instance of GRIDSS is running for any given output file.
*.bai BAM index for coordinate sorted intermediate BAM file.
WORKING_DIR/file.gridss.working Working directory for intermediate files related to the given file.
WORKING_DIR/input.gridss.working/input*_metrics Various summary metrics for the given input file.
WORKING_DIR/input.gridss.working/input.realign.N.fq Split read identification fastq file requiring alignment by NGS aligner.
WORKING_DIR/input.gridss.working/input.realign.N.bam Result of NGS alignment.
WORKING_DIR/input.gridss.working/ Subset of input reads considered by GRIDSS. This file is useful for visualisation of the supporting reads contributing to structural variant calls. Note that this file includes split read alignments identified from soft clipped reads in the input file.
WORKING_DIR/assembly.gridss.working/ Assembly contigs represented as soft clipped or split reads. These are the assembly contigs GRIDSS uses for variant calling. Note that, as with split reads, GRIDSS uses the SA tag to encode split read alignments.
WORKING_DIR/output.gridss.working/output.breakpoint.vcf Raw unannotated variant calls before unique allocation of multi-mapping reads.

Building from source

Maven is used for build and dependency management which simplifies compile to the following steps:

  • git clone
  • cd gridss
  • mvn clean package

If GRIDSS was built successfully, a combined jar containing GRIDSS and all required libraries located at target/GRIDSS-VERSION-gridss-jar-with-dependencies.jar will have been created.

Error Messages

For some error messages, it is difficult to determine the root cause and what to do to fix it. Here is a list of key phrases of errors encountered by users and their solution

htsjdk.samtools.SAMFormatException: SAM validation error

Your input file does not conform to the SAM/BAM specifications. Solutions are to fix the input file so it conforms to the specifications (recommended) or add --picardoptions VALIDATION_STRINGENCY=LENIENT to ignore the error. Note that not all errors can be ignored.

Aborting since lock gridss.lock.OUTPUT already exists. GRIDSS does not support multiple simulatenous instances running on the same data.

Multiple instances of GRIDSS were run on the same data. GRIDSS does not yet support MPI parallelisation across multiple machines. Use the WORKER_THREADS parameter to specify the desired level of multi-threading. If using a cluster/job queuing system, a single non-MPI job should be submitted and either WORKER_THREADS explicitly set to the number of cores associated with the job requests, or the job should request the entire node.

If the lock directory exists and you know a GRIDSS process is not running (eg: the GRIDSS process was killed), then you can safely delete the lock directory.

Exception in thread "main" java.lang.UnsupportedClassVersionError: au/edu/wehi/idsv/Idsv : Unsupported major.minor version 52.0

You are attempting to run GRIDSS with an old Java version. GRIDSS requires Java 8 or later.

ExternalProcessFastqAligner Subprocess terminated with with exit status 1. Alignment failed for INPUT.realign.0.fq

The external aligner (bwa) could not be run. The most common causes of this are:

  • bwa is not on PATH
  • Does running "bwa" print out the bwa usage message? If you are using a cluster, you may have to add bwa to your PATH (eg module add bwa).
  • bwa index does not exist
  • bwa index has incorrect suffix
  • e.g. if the reference is ref.fa the index must be ref.fa.bwt not ref.bwt

Can you run the bwa command exactly as it appears in the error message?

(Too many open files)

GRIDSS has attempted to open too many files at once and the OS file handle limit has been reached. On linux 'ulimit -n' displays your current limit. This error likely to be encountered if you have specified a large number of input files or threads. The following solution is recommended:

  • Increase your OS limit on open file handles (eg ulimit -n _<larger number>_)
    • Note that many linux systems have a default hard limit on open file handles of 4096 which with many samples is frequently too still too few. Increasing the hard limit requires root access.
  • Increase the chunk size. The default chunk size is 10 million bases. This can be increased by adding a chunkSize=50000000 line a file and adding to the GRIDSS command line. Note that this will increase the number of bases processed by each job thus reduce the level of parallelisation possible.
  • Reduce number of worker threads. A large number of input files being processed in parallel results in a large number of files open at the same time.

If those options fail, your remaining options are:

  • Added -Dgridss.defensiveGC=true to the java command-line used for GRIDSS. Memory mapped file handles are not released to the OS until the buffer is garbage collected . This option add a request for garbage collection whenever a file handle is no longer used. This is a significant overhead and is not a good option for sparse data samples (such as exome or targetted sequencing) - increasing the chunk size is a much better option for these samples.
  • As a last-ditch effort, you can keep rerunning GRIDSS until it completes. If you are using the default entry point of gridss.CallVariants and have -Dgridss.gridss.output_to_temp_file=true, then you can rerun GRIDSS and it will continue from where it left off. Assuming it doesn't keep dying at the same spot, it will eventually complete.

Reference genome used by input.bam does not match reference genome reference.fa. The reference supplied must match the reference used for every input.

The reference genome used to align input.bam does not match the reference genome supplied to GRIDSS. If the differences are purely based on chromosome name and ordering, the Picard tools utility ReorderBam can be used to fix chromosome orderings.

Unable to use sswjni library - assembly will be very slow. Please ensure libsswjni for your OS and architecture can be found on java.library.path

The sswjni library could not be loaded as the precompiled version is not compatable with your environment. See the sswjni sections for details on how to disable libsswjni or recompile it for your system.

"Segmentation Fault", fatal JVM error, or no error message.

This is likely to be caused by a crash during alignment in libsswjni. See the sswjni sections for details on how to disable libsswjni or recompile it for your system.

Illegal Instruction

Your CPU does not support the SSE2 instruction set. See the sswjni sections for details on how to disable libsswjni.

Java HotSpot(TM) 64-Bit Server VM warning: INFO: os::commit_memory(0x00007fc36e200000, 48234496, 0) failed; error='Cannot allocate memory' (errno=12)

GRIDSS has run out of memory. Either not enough memory has been allocated to run GRIDSS or GRIDSS has attempted to memory map too many files (See "(Too many open files)"). In both cases, restart GRIDSS (increasing the memory available if required) and GRIDSS will continue from where it left off.

java.lang.AssertionError: java.lang.ClassNotFoundException:

You are running GRIDSS in multi-mapping mode using only a JRE instead of a full JDK. Update your PATH and JAVA_HOME to a Java 1.8+ JDK installation.

htsjdk.samtools.util.RuntimeIOException: No space left on device

Just like Picard tools and htsjdk libraries that GRIDSS uses, intermediate files are sorted according the the TEMP file location. On many system, /tmp does not have enough space to sort a BAM file so it is possible to run out of intermediate file storage even if you have plenty of space left on the file system the input and output files are stored on. Using the same command-line options as Picard tools, the intermediate files location can be set using the TMP_DIR command-line argument.

It's also possible that you've just run out of space.

Contributing to GRIDSS

Bug fix pull requests are always welcome.

If you have a feature you would like to implement, please first raise an issue outlining the problem and proposed solution so as to avoid any wasted or duplicated effort. To build GRIDSS, the development environment requires maven and all GRIDSS and VIRUSBreakend dependencies (see GRIDSS Dockerfile and VIRUSBreakend readme). To run all test cases, a ../ref/ directory is also required containing the following genomes (with all associated GRIDSS --setupreference files created):

  • hg19.fa
  • Homo_sapiens_assembly38.fasta
  • hg38.fa
  • Escherichia_coli_bl21_de3_.ASM956v1.dna.toplevel.fa

scripts/dev/ contains helpful utilities for GRIDSS development work.


GRIDSS uses the YourKit Java Profiler for Performance Tuning.

alt text

YourKit supports open source projects with innovative and intelligent tools for monitoring and profiling Java and .NET applications. YourKit is the creator of YourKit Java Profiler, YourKit .NET Profiler, YourKit YouMonitor.


GRIDSS: the Genomic Rearrangement IDentification Software Suite



Unknown, GPL-3.0 licenses found

Licenses found






No packages published