Skip to content

fbartusch/snakemake_tutorial

Repository files navigation

Snakemake tutorial workflow with SCIF

This repository implements the SnakeMake tutorial workflow and uses the Scientific Filesystem (SCIF) to provide a reproducible research environment. Just clone this repository and you can start to try out things. You can build s Singularity and/or Docker container to run the workflow.

Building containers

sudo singularity build snakemake.simg Singularity
docker build -t vanessa/snakemake.scif .

Run the workflow

The Snakefile specifies rules that should be executed. State a target file and Snakemake will build a directed acyclic graph (DAG) from the rules until the target file is reached. The rules are then executed to create the target file. The Snakefile hides the details of the environment provided by SCIF in the container as you can see from the commands below. All commands can be executed from outside the container or from within the container.

In all cases we bind the example dataset to the data directory provided by SCIF. SCIF provides the environment in which the workflow steps are executed. The workflow steps themself and therefore the interaction with SCIF is implemented in the Snakefile.

Inside container, Singularity

singularity shell --bind data/:/scif/data snakemake.simg
snakemake all

Inside container, Docker

To run the whole workflow we need the Snakefile which is not part of the container. To access the Snakefile from within the Docker container, we need an additional mount.

docker run -v $PWD/data:/scif/data:z -v $PWD:/working_dir -it -w /working_dir --entrypoint /bin/bash vanessa/snakemake.scif
snakemake all

Outside Container, Singularity

singularity exec --bind data/:/scif/data snakemake.simg snakemake all

Outside Container, Docker

Because Docker does not mount the current working directory, we mount it to a the SCIF data directory of the SCIF snakemake app. The snakemake app will execute snakemake in /scif/data/snakemake, thus snakemake can find the Snakefile.

docker run -v $PWD/data:/scif/data:z -v $PWD:/scif/data/snakemake:z -it vanessa/snakemake.scif run snakemake all

Generate graphical representation of the workflow

The whole workflow should finish within 5 seconds ... too fast to realize what is happening. A visualization of the whole workflow we just ran would be nice ...

The SCIF app graphviz_create_dag generates a directed acyclic graph of the workflow execution plan. This is a feature of Snakemake, but needs Graphviz as additional dependency. Therefore Graphviz is installed in the appinstall section of this app.

rm -r data/calls/ /data/mapped_reads/ data/sorted_reads/ data/report.html

Now we can generate the graph of the execution plan. The plan is computed from the Snakefile in our current working directory and the target file we specify. In this case we want to create the final report file report.html. The directed acyclic graph of the workflow should be saved atdata/dag.svg. The commands for Singularity and Docker are as follows:

Singularity

singularity run --bind data/:/scif/data snakemake.simg run graphviz_create_dag $PWD report.html dag.svg
> [graphviz_create_dag] executing /bin/bash /scif/apps/graphviz_create_dag/scif/runscript /home/fbartusch/github/snakemake_tutorial report.html dag.svg
> Building DAG of jobs...

Docker

In contrast to Singularity, Docker cannot access the current working directory. Therefore we mount the current working directory to the SCIF data folder of the graphviz_create_dag app.

docker run -v $PWD/data:/scif/data:z -v $PWD:/scif/data/graphviz_create_dag:z -it vanessa/snakemake.scif run graphviz_create_dag /scif/data/graphviz_create_dag report.html dag.svg
> [graphviz_create_dag] executing /bin/bash /scif/apps/graphviz_create_dag/scif/runscript /scif/data/graphviz_create_dag report.html dag.svg
> Building DAG of jobs...

Map with bwa mem

Instead of running the whole workflow at once with Snakemake, we can also use the SCIF environment to execute computations. The following examples also illustrate special SCIF syntax for working with environment variables or piping output to a file. First, we map the reads with bwa mem to the reference genome.

Inside container, Singularity

Note the use of [e] so we can easily pass the environment variable to the SCIF. Otherwise, it would be evaluated on the host.

singularity shell --bind data/:/scif/data snakemake.simg
mkdir /scif/data/mapped_reads
scif run bwa mem -o [e]SCIF_DATA/mapped_reads/A.sam [e]SCIF_DATA/genome.fa [e]SCIF_DATA/samples/A.fastq

Inside container, Docker

docker run -v $PWD/data:/scif/data:z -it --entrypoint /bin/bash vanessa/snakemake.scif
mkdir -p /scif/data/mapped_reads
scif run bwa mem -o [e]SCIF_DATA/mapped_reads/A.sam [e]SCIF_DATA/genome.fa [e]SCIF_DATA/samples/A.fastq

Outside container, Singularity

mkdir -p data/mapped_reads
singularity run --bind data:/scif/data snakemake.simg run bwa mem -o [e]SCIF_DATA/mapped_reads/A.sam [e]SCIF_DATA/genome.fa [e]SCIF_DATA/samples/A.fastq

Outside the container, Docker

mkdir -p data/mapped_reads
docker run -v $PWD/data:/scif/data:z vanessa/snakemake.scif run bwa mem -o [e]SCIF_DATA/mapped_reads/A.sam [e]SCIF_DATA/genome.fa [e]SCIF_DATA/samples/A.fastq

Sam to Bam Conversion

Inside the container, Singularity

Note the use of [out] as a substitute for >. If you wanted to use > you could put the entire thing in quotes.

singularity shell --bind data/:/scif/data snakemake.simg
scif run samtools view -Sb [e]SCIF_DATA/mapped_reads/A.sam [out] [e]SCIF_DATA/mapped_reads/A.bam    # or
scif run samtools 'view -Sb $SCIF_DATA/mapped_reads/A.sam > $SCIF_DATA/mapped_reads/A.bam'
> [samtools] executing /bin/bash /scif/apps/samtools/scif/runscript view -Sb $SCIF_DATA/mapped_reads/A.sam > $SCIF_DATA/mapped_reads/A.bam

Inside the container, Docker

docker run -v $PWD/data:/scif/data:z -it --entrypoint /bin/bash vanessa/snakemake.scif
scif run samtools view -Sb [e]SCIF_DATA/mapped_reads/A.sam [out] [e]SCIF_DATA/mapped_reads/A.bam    # or
scif run samtools 'view -Sb $SCIF_DATA/mapped_reads/A.sam > $SCIF_DATA/mapped_reads/A.bam'
> [samtools] executing /bin/bash /scif/apps/samtools/scif/runscript view -Sb $SCIF_DATA/mapped_reads/A.sam > $SCIF_DATA/mapped_reads/A.bam

Outside the container, Singularity

singularity run --bind data:/scif/data snakemake.simg run samtools view -Sb [e]SCIF_DATA/mapped_reads/A.sam [out] [e]SCIF_DATA/mapped_reads/A.bam   # or
singularity run --bind data:/scif/data snakemake.simg run samtools 'view -Sb $SCIF_DATA/mapped_reads/A.sam > $SCIF_DATA/mapped_reads/A.bam'

Outside the container, Docker

docker run -v $PWD/data:/scif/data vanessa/snakemake.scif run samtools view -Sb [e]SCIF_DATA/mapped_reads/A.sam [out] [e]SCIF_DATA/mapped_reads/A.bam   # or
docker run -v $PWD/data:/scif/data vanessa/snakemake.scif run samtools 'view -Sb $SCIF_DATA/mapped_reads/A.sam > $SCIF_DATA/mapped_reads/A.bam'

Interactive development

This can be done for Docker or Singularity, just with different commands to shell into the container!

docker run -it -v $PWD/data:/scif/data:z vanessa/snakemake.scif pyshell
singularity run --bind data/:/scif/data snakemake.simg pyshell
Found configurations for 4 scif apps
bwa
graphviz_create_dag
samtools
snakemake
[scif] /scif bwa | graphviz_create_dag | samtools | snakemake
Python 3.6.2 |Anaconda, Inc.| (default, Sep 22 2017, 02:03:08) 
[GCC 7.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
(InteractiveConsole)
>>> 

# Activate bwa 
client.activate('samtools')

# Environment variables active
client.environment

# Run bwa interactively
args = ['mem', '-o', '[e]SCIF_DATA/mapped_reads/a.sam', '[e]SCIF_DATA/genome.fa', '[e]SCIF_DATA/samples/A.fastq']
client.run('bwa', args=args)

# Run sam--bam interactively
args = ["view", "-Sb", "/scif/data/mapped_reads/r1_subset.sam", ">", "/scif/data/mapped_reads/r1_subset.bam"]
client.run('samtools', args=args)

This repo is also used as example on the SCIF GitHub repo, so take a look:

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published