This document describes the integration of Sentieon software with CWL and Toil. This integration enables to easily run CWL-defined Sentieon pipelines on a variety of HPC and cloud environments, either on a single node or distributed onto multiple nodes for rapid turnaround processing. Please see Sentieon Distribution Application Note on distribution mode for more information on that topic.
CWL, also known as Common Workflow Language, is a specification for describing analysis workflows and tools. The objective of CWL standards is to make analysis workflows portable and scalable on a variety of hardware and software environments. CWL enables application developers to focus on the pipeline development without having to worry about deployment details or distribution implementation. For more information about CWL, please refer to its website.
Toil is a cross-platform pipeline management system and execution engine that implements the CWL standards. Toil enables executing pipelines defined in CWL on different job management systems. It is written in Python. Toil supports a number of scheduler systems, including Grid Engine, Slurm, Torque and LSF as well as workflow execution on public clouds including Amazon Web Services, Google Cloud, and Microsoft's Azure. For more details about Toil, please refer to its website.
Sentieon has developed a set of reference pipelines following CWL specification. Users can modify these pipelines to fit their needs for deployment in different environments.
Both CWL and Toil are under rapid development, and Sentieon is an active contributor to both open source projects. As CWL and Toil develop, Sentieon will continue to improve these templates to make them more user, and deployment-friendly. We welcome users' comment and feedback.
To run the CWL pipelines provided by Sentieon, the following software is required on all work nodes:
-
Python 2.7 or above
-
Node.js
-
Toil-supported scheduler, for instance: Grid Engine, Slurm, Torque and LSF. In this App note, we will use LSF as an example (IBM Spectrum LSF Suite Community Edition 10.1. See IBM's LSF website for more details)
-
Cwltoil with versions and installation process below
To install cwltoil, first, clone the Sentieon's cwltoil repository from github, and install the python library source:
git clone https://github.com/Sentieon/toil.git
cd toil
git checkout tags/sentieon-toil-20180522
pip install .[all]
You should get cwltoil version as shown below:
cwltoil --version
DEBUG:rdflib:RDFLib Version: 4.2.2
3.16.0a1
You will also need the CWL template package provided by Sentieon. The package contains templates that define Sentieon algorithm modules and typical pipelines with commonly used settings.
Checkout this github repository:
git clone https://github.com/Sentieon/Sentieon-cwl.git
There are three sub-directories:
- algo: CWL definitions of all Sentieon algorithm modules and corresponding example input YAML files.
- stage: CWL definitions of typical stages as building blocks for commonly used pipelines. An example is deduplication stage. This stage requires using two algo modules: LocusCollector and Dedup.
- pipeline: CWL definitions of commonly used pipelines and corresponding example input YAML files.
In typical CWL workflows, the CWL file defines the general pipeline, while a YAML file defines the input arguments of the pipeline. In the case of Sentieon's DNAseq germline variant calling pipeline from BAM to VCF, YAML defines the following:
- Input BAM file
- Reference file
- Interval file or bed file, if necessary
- dbSNP VCF
- Known sites VCFs for indel realignment and BQSR
- Intermediate output file names
- Number of threads
- Shards, which define how the whole reference is cut into small segments for distribution
Below is an example YAML file:
reference: /path/to/ucsc.hg19.fasta
input_bam:
- sorted.bam
realign_known_sites:
- /path/to/Mills_and_1000G_gold_standard.indels.hg19.sites.vcf.gz
- /path/to/dbsnp_135.hg19.vcf.gz
bqsr_known_sites:
- /path/to/Mills_and_1000G_gold_standard.indels.hg19.sites.vcf.gz
- /path/to/dbsnp_135.hg19.vcf.gz
dbsnp: /path/to/dbsnp_135.hg19.vcf.gz
qcal_output_file: recal_table-pipeline.txt
dedup_output_bam: deduped-pipeline.bam
dedup_metrics_output_file: dedup-metrics.txt
realign_output_bam: realigned-pipeline.bam
output_file: output_hc.vcf.gz
interval: interval.txt
shard: ["chr1", "chr2", "chr3"]
threads: 16
You should first test the initial CWL installation and integration with your chosen scheduler. You can follow the instructions below to execute a simple single-stage (algorithm Readwriter) CWL pipeline.
You will find a directory called cwl with a number of cwl and yaml files. For this test run, we will use rw.cwl
and rw.yaml
.
To get started, the user will need prepare a set of directories:
- Job store: a centralized directory for all the temporary files for the workflow. Toil will initiate the job by copying all the necessary files to this location. The user will only need to create the parent directory for the job store. The job store must be accessible by all potential job nodes. Toil will create this directory if it is not present. After the job is complete, the job store directory will be cleared.
- Working directory: Absolute path to the directory where temporary files generated during the workflow will be placed. The working directory can be in the local storage of the node. The user needs to create this directory before starting the job.
- Output directory: Absolute path to the directory where the output of each stage and the final output will be stored. The user needs to create this directory before starting the job.
After creating the above directories, copy the rw.yaml to your own job directory, and modify the reference and input bam file path. Example of the rw.yaml
is shown below.
reference: /path/to/references/hg19/ucsc.hg19.fasta
input_bam:
- /path/to/jobdir/cwl/sorted.bam
output_file: out.bam
threads: 8
Then, the user needs to prepare Sentieon-specific environment variables:
- Add
SENTIEON_INSTALL_DIR/bin
toPATH
- Make sure
SENTIEON_LICENSE
is properly set - Set
SENTIEON_TMPDIR
to a fast local disk drive, instead of in a NFS to reduce IO.
Start the job with following command:
workDIR=/path/to/workDir
jobStore=/path/to/jobStore
outDIR=/path/to/outDir
templateDir=/path/to/Sentieon_CWL_TEMPLATE
TMPDIR=$workDIR cwltoil --preserve-environment SENTIEON_LICENSE PATH SENTIEON_TMPDIR --workDir $workDIR --outdir $workDIR --batchSystem=lsf --jobStore file:$jobStore --logDebug --logFile run.log $templateDir/rw.cwl rw.yaml
Some additional comments:
-
Here we use LSF as the batch system. Simply set the option
--batchSystem
to use other batching systems. Please refer to the Toil manual for more details. -
--logDebug
defines the log level to DEBUG, and--logFile
specifies a file path to write the logging output to.
After the job is finished successfully, you should be able to see out.bam
in the output directory, as defined in the job setup --outDir
. This shows the Toil setup is working as expected.
Running a distributed Sentieon CWL pipeline requires defining how the jobs are divided up among different nodes. Sentieon software natively supports sharded mode, where small regions of the genome, called shards, are processed in parallel. This allows computational work to be distributed across multiple servers or compute nodes. The shards are predefined in the yaml file. Without the shard option, only a single node is used as in the above example of rw.yaml
.
Shards can be defined as below:
- Chromosome name only, eg. chr1, chr2, etc.
- Chromosome name and locus range, eg., chr1:1-1234556
- A combination of above separated by comma, eg., chrM,chr1:1-123456,...
In the YAML file, the shard option defines an array of shards, below is an example:
shard: [
"1:1-50000000",
"1:50000001-100000000",
"1:100000001-150000000",
"1:150000001-200000000",
"1:200000001-249250621,2:1-749379",
"2:749380-50749379",
"2:50749380-100749379",
"2:100749380-150749379",
"2:150749380-200749379",
"2:200749380-243199373,3:1-7550006",
"3:7550007-57550006",
"3:57550007-107550006",
"3:107550007-157550006",
"3:157550007-198022430,4:1-9527576",
"4:9527577-59527576",
"4:59527577-109527576",
"4:109527577-159527576",
...
]
You can use the shell script script/determine_shards.sh
to divide the whole genome into uniform shards. The "NO_COOR" shard processes unmapped reads.
determine_shards.sh $BAM $NUM_PARTS
The user can use the above shell script to generate the set of shards for distributed jobs, completing the yaml file. With the yaml file updated, you can run the pipeline as in the above simple example.
Similar to rw.cwl
for single-node CWL pipeline testing, Sentieon also provides rw-distr.cwl
for distribution test-run.
Copy the rw-distr.yaml
to a local directory, modify the input parameters of the file. Please note that the only difference between rw.yaml
and rw-distr.yaml
is the shard option.
Reference: /path/to/references/hg19/ucsc.hg19.fasta
input_bam:
- /path/to/jobdir/cwl/sorted.bam
output_file: out.bam
shard: ["chr1", "chr2", "chr3"]
threads: 8
Prepare the necessary directories as explained in the single-node test-run example, then run the pipeline as below:
workDIR=/path/to/workDir
jobStore=/path/to/jobStore
outDIR=/path/to/outDir
templateDir=/path/to/Sentieon_CWL_TEMPLATE
TMPDIR=$workDIR cwltoil --preserve-environment SENTIEON_LICENSE PATH SENTIEON_TMPDIR --workDir $workDIR --outdir $workDIR --batchSystem=lsf --jobStore file:$jobStore --logDebug --logFile run.log $templateDir/rw-distr.cwl rw-distr.yaml
Please note that Sentieon can achieve the best performance by utilizing all available virtual cores. The scheduler needs to properly interpret the CPU resource request as virtual cores, instead of physical cores.
Running distribution pipeline from fastq requires a few extra steps to ensure identical result as running on a single server.
Run the following command to create fastq file index:
sentieon fqidx -K chunk_size -o fastq_index sample1_1.fq.gz sample1_2.fq.gz
where:
-K chunk_size
: Defines the chunk size in bases. Same meaning as chunk size in bwa-mem. Using fixed chunk size makes sure consistent result. Example: -K 10000000. Please use the same chunk size here aschunk_size
defined in the YAML file shown below.-o fastq_index
: Defines the output index file name.sample1_1.fq.gz
andsample1_2.fq.gz
: input pair-ended fastq files.
Please refer to Sentieon's App note - Distributed mode for more details of the fqidx
.
Below is an example of the pipeline-fastq2vcf-distr.yaml
file.
input_reads:
- /data/NA11932_1000G_Exome_ERR034544_1.fastq.gz
- /data/NA11932_1000G_Exome_ERR034544_2.fastq.gz
input_reads_index_file: /data/NA11932_1000G_Exome_ERR034544.fq.idx
reference: /path/to/references/hg19/ucsc.hg19.fasta
realign_known_sites:
- /path/to/Mills_and_1000G_gold_standard.indels.hg19.sites.vcf.gz
- /path/to/dbsnp_138.hg19.vcf.gz
bqsr_known_sites:
- /path/to/Mills_and_1000G_gold_standard.indels.hg19.sites.vcf.gz
- /path/to/dbsnp_138.hg19.vcf.gz
dbsnp: /path/to/dbsnp_138.hg19.vcf.gz
sort_output_bam: sorted-pipeline.bam
qcal_output_file: recal_table-pipeline.txt
dedup_output_bam: deduped-pipeline.bam
dedup_metrics_output_file: dedup-metrics.txt
output_file: output_hc.vcf.gz
shard: ["chrM,chr1,chr2,chr3,chr4,chr5:1-165000000", \
"chr5:165000001-180915260,chr6,chr7,chr8,chr9,chr10,chr11,chr12,chr13:1-8000000",\
"chr13:8000001-115169878,chr14,chr15,chr16,chr17,chr18,chr19,chr20,chr21,chr22,chrX,chrY,chr1_gl000191_random,chr1_gl000192_random,chr4_ctg9_hap1,chr4_gl000193_random,chr4_gl000194_random,chr6_apd_hap1,chr6_cox_hap2,chr6_dbb_hap3,chr6_mann_hap4,chr6_mcf_hap5,chr6_qbl_hap6,chr6_ssto_hap7,chr7_gl000195_random,chr8_gl000196_random,chr8_gl000197_random,chr9_gl000198_random,chr9_gl000199_random,chr9_gl000200_random,chr9_gl000201_random,chr11_gl000202_random,chr17_ctg5_hap1,chr17_gl000203_random,chr17_gl000204_random,chr17_gl000205_random,chr17_gl000206_random,chr18_gl000207_random,chr19_gl000208_random,chr19_gl000209_random,chr21_gl000210_random,chrUn_gl000211,chrUn_gl000212,chrUn_gl000213,chrUn_gl000214,chrUn_gl000215,chrUn_gl000216,chrUn_gl000217,chrUn_gl000218,chrUn_gl000219,chrUn_gl000220,chrUn_gl000221,chrUn_gl000222,chrUn_gl000223,chrUn_gl000224,chrUn_gl000225,chrUn_gl000226,chrUn_gl000227,chrUn_gl000228,chrUn_gl000229,chrUn_gl000230,chrUn_gl000231,chrUn_gl000232,chrUn_gl000233,chrUn_gl000234,chrUn_gl000235,chrUn_gl000236,chrUn_gl000237,chrUn_gl000238,chrUn_gl000239,chrUn_gl000240,chrUn_gl000241,chrUn_gl000242,chrUn_gl000243,chrUn_gl000244,chrUn_gl000245,chrUn_gl000246,chrUn_gl000247,chrUn_gl000248,chrUn_gl000249,NO_COOR"]
extract_chunks: ["0-373", "373-746", "746-1119"]
chunk_size: 10000000
readgroup: "group2"
platform: "ILLUMINA"
sample: "sample"
library: "library"
mark_secondary: true
threads: 8
In the above YAML file, extract_chunk
field defines how bwa stage is distributed. Use the script script/determine_chunks.sh
to find out the exact range for distribution:
script/determine_chunks.sh $INPUT_FASTQ.IDX $NUM_PARTS
where INPUT_FASTQ.IDX is the index file generated above with fqidx command, and NUM_PARTS is the number of parts the fastq file will be sliced into. Please note NUM_PARTS defines the number of subjobs the bwa distribution will generate.
The shell script determine_chunks.sh
will call sentieon fqidx query
as shown above to get the total number chunks from the fastq index file, and then divide it into equal-size chunks. In this case, with 1119 chunks distributed into 3 parts, the above script will give the following output. Enter these chunk ranges into the extract_chunk field of the yaml file.
["0-373","373-746","746-1119]
Then use the script script/determine_shards.sh
to calculate the shards from the fai file for shard
field in the YAML file. Run the script as shown below:
script/determine_shards.sh hg19.fasta.fai $NUM_PARTS
where you should use the FAI index file for same reference as defined in the YAML file, and NUM_PARTS defines the number of subjobs the CWL distribution will generate.
With these preparation, user can run the cwltoil pipeline with the yaml input file and get the same result as in the single server.
Sentieon CWL pipelines support specifying the path of input data files either as a string or file. Below is an example:
# Specifying path as a string
reference: /path/to/references/hg19/ucsc.hg19.fasta
# Specifying path as an array of string
realign_known_sites:
- /path/to/Mills_and_1000G_gold_standard.indels.hg19.sites.vcf.gz
- /path/to/dbsnp_138.hg19.vcf.gz
# Specifying path as an array of File
input_reads:
- class: File
path: NA11932_1000G_Exome_ERR034544_1.fastq.gz
- class: File
path: NA11932_1000G_Exome_ERR034544_2.fastq.gz
# Speficying path as File
input_reads_index_file:
- class: File
path: NA11932_1000G_Exome_ERR034544.fq.idx
# Only specify the output file name as a string for output data files. Absolute path is not accepted.
output_bam: output.bam
Sentieon recommends the following for best performance and compatibility:
- When specifying the file path as a string, use only absolute path that can be used on all computing nodes.
- For input files that will be used repeatedly throughout the pipeline, like reference, known sites, interval files, etc. use string representation so that the files are not copied to local working directory for every stage of the pipeline.
- When specifying the file path as a file, one can either use absolute or relative path. However, when relative file path is used, the file is assumed to be in the same directory of the YAML file.
- Put workDir, outDir and jobStore on the same device so that either symbolic link or hard link will be created for stage input/output to reduce file IO operation. Set
SENTIEON_TMPDIR
environment variable to a local scratch directory for temporary files of each stage and pass it to CWLtoil with--preserve-environment
argument. - User can only specify the file name for output file. All output files will be saved at
outDir
, and user cannot dictate where the file is saved by specifying the output file path in the YAML file.
As mentioned above, both CWL and Toil are under active and rapid development, and new version may not be always backward compatible. Sentieon is actively improving cwltoil as well as Sentieon pipelines, to make them more efficient and user-friendly.
Currently, we have found the following known issues:
- In cwltoil, the input/output for each stage must be read from/written to the job store. With large files, this can lead to significant runtime spent on unnecessary file copying between compute nodes.
- All input VCF files should be in
.vcf.gz
format with.tbi
index files. Uncompressed VCF with .idx index files are not supported at the moment.