Skip to content

6. Workflow Structure

There are three directories in a Nextflow workflow repository that have a special purpose:

6.1 ./bin

The bin directory (if it exists) is always added to the $PATH for all tasks. If the tasks are performed on a remote machine, the directory is copied across to the new machine before the task begins. This Nextflow feature is designed to make it easy to include accessory scripts directly in the workflow without having to commit those scripts into the container. This feature also ensures that the scripts used inside the workflow move on the same revision schedule as the workflow itself.

It is important to know that Nextflow will take care of updating $PATH and ensuring the files are available wherever the task is running, but will not change the permissions of any files in that directory. If a file is called by a task as an executable, the workflow developer must ensure that the file has the correct permissions to be executed.

For example, let's say we have a small R script that produces a csv and a tsv:

1
2
3
4
5
6
#!/usr/bin/env Rscript
library(tidyverse)

plot <- ggplot(mpg, aes(displ, hwy, colour = class)) + geom_point()
mtcars |> write_tsv("cars.tsv")
ggsave("cars.png", plot = plot)

We'd like to use this script in a simple workflow:

process PlotCars {
    container 'rocker/tidyverse:latest'

    output:
    path("*.png"), emit: "plot"
    path("*.tsv"), emit: "table"

    script:
    """
    cars.R
    """
}

workflow {
    PlotCars()

    PlotCars.out.table | view { "Found a tsv: $it" }
    PlotCars.out.plot | view { "Found a png: $it" }
}

To do this, we can create the bin directory, write our R script into the directory. Finally, and crucially, we make the script executable:

mkdir -p bin
cat << EOF > bin/cars.R
#!/usr/bin/env Rscript
library(tidyverse)

plot <- ggplot(mpg, aes(displ, hwy, colour = class)) + geom_point()
mtcars |> write_tsv("cars.tsv")
ggsave("cars.png", plot = plot)
EOF
chmod +x bin/cars.R

Warning

Always ensure that your scripts are executable. The scripts will not be available to your Nextflow processes without this step.

Let's run the script and see what Nextflow is doing for us behind the scenes:

1
2
3
4
5
6
7
8
9
cat << EOF > nextflow.config
profiles {
    docker {
        docker.enabled = true
    }
}
EOF
rm -r work
nextflow run . -profile docker

and then inspect the .command.run file that Nextflow has generated

code work/*/*/.command.run

You'll notice a nxf_container_env bash function that appends our bin directory to $PATH:

nxf_container_env() {
cat << EOF
export PATH="\$PATH:/workspace/gitpod/nf-training/advanced/structure/bin"
EOF
}

When working on the cloud, Nextflow will also ensure that the bin directory is copied onto the virtual machine running your task in addition to the modification of $PATH.

Warning

Always use a portable shebang line in your bin directory scripts.

In the R script example shown above, I may have the Rscript program installed at (for example) /opt/homebrew/bin/Rscript. If I hard-code this path into my cars.R, everything will work when I'm testing locally outside of the docker container, but will fail when running with docker/singularity or in the cloud as the Rscript program may be installed in a different location in those contexts.

It is strongly recommended to use #!/usr/bin/env when setting the shebang for scripts in the bin directory to ensure maximum portability.

6.2 ./templates

If a process script block is becoming too long, it can be moved to a template file. The template file can then be imported into the process script block using the template method. This is useful for keeping the process block tidy and readable. Nextflow's use of $ to indicate variables also allows for directly testing the template file by running it as a script.

The structure directory already contains an example template - a very simple python script. We can add a new process that uses this template:

1
2
3
4
5
process SayHiTemplate {
    debug true
    input: val(name)
    script: template 'adder.py'
}

6.3 ./lib

In the previous chapter, we saw the addition of small helper Groovy functions to the main.nf file. It may at times be helpful to bundle functionality into a new Groovy class. Any classes defined in the lib directory are available for use in the workflow - both main.nf and any imported modules.

Classes defined in lib directory can be used for a variety of purposes. For example, the nf-core/rnaseq workflow uses five custom classes:

  1. NfcoreSchema.groovy for parsing the schema.json file and validating the workflow parameters.
  2. NfcoreTemplate.groovy for email templating and nf-core utility functions.
  3. Utils.groovy for provision of a single checkCondaChannels method.
  4. WorkflowMain.groovy for workflow setup and to call the NfcoreTemplate class.
  5. WorkflowRnaseq.groovy for the workflow-specific functions.

The classes listed above all provide utility executed at the beginning of a workflow, and are generally used to "set up" the workflow. However, classes defined in lib can also be used to provide functionality to the workflow itself.

6.3.1 Making a Metadata Class

Let's consider an example where we create our own custom class to handle metadata. We can create a new class in ./lib/Metadata.groovy. We'll extend the built-in HashMap class, and add a simple method to return a value:

1
2
3
4
5
class Metadata extends HashMap {
    def hi() {
        return "Hello, workshop participants!"
    }
}

We can then use this class in our workflow:

1
2
3
4
5
workflow {
    Channel.of("Montreal")
    | map { new Metadata() }
    | view
}

We can use the new hi method in the workflow:

1
2
3
4
5
workflow {
    Channel.of("Montreal")
    | map { new Metadata() }
    | view { it.hi() }
}

At the moment, the Metadata class is not making use of the "Montreal" being passed into the closure. Let's change that by adding a constructor to the class:

1
2
3
4
5
6
7
8
9
class Metadata extends HashMap {
    Metadata(String location) {
        this.location = location
    }

    def hi() {
        return this.location ? "Hello, from ${this.location}!" : "Hello, workshop participants!"
    }
}

Which we can use like so:

1
2
3
4
5
workflow {
    Channel.of("Montreal")
    | map { place -> new Metadata(place) }
    | view { it.hi() }
}

We can also use this method when passing the object to a process:

process UseMeta {
    input: val(meta)
    output: path("out.txt")
    script: "echo '${meta.hi()}' | tee out.txt"
}

workflow {
    Channel.of("Montreal")
    | map { place -> new Metadata(place) }
    | UseMeta
    | view
}

Why might this be helpful? You can add extra classes to the metadata which can be computed from the existing metadata. For example, we might want to grab the adapter prefix:

1
2
3
def getAdapterStart() {
    this.adapter?.substring(0, 3)
}

Which we might use like so:

process UseMeta {
    input: val(meta)
    output: path("out.txt")
    script: "echo '${meta.adapter} prefix is ${meta.getAdapterStart()}' | tee out.txt"
}

workflow {
    Channel.of("Montreal")
    | map { place -> new Metadata(place) }
    | map { it + [adapter:"AACGTAGCTTGAC"] }
    | UseMeta
    | view
}

You might even want to reach out to external services such as a LIMS or the E-utilities API. Here we add a dummy "getSampleName()" method that reaches out to a public API:

1
2
3
4
5
6
7
8
9
def getSampleName() {
    def get = new URL('https://postman-echo.com/get?sampleName=Fido').openConnection()
    def getRC = get.getResponseCode();
    if (getRC.equals(200)) {
        JsonSlurper jsonSlurper = new JsonSlurper()
        def json = jsonSlurper.parseText(get.getInputStream().getText())
        return json.args.sampleName
    }
}

Which we can use like so:

1
2
3
4
5
6
process UseMeta {
    input: val(meta)
    output: path("out.txt")
    script:
    "echo '${meta.adapter} prefix is ${meta.getAdapterStart()} with sampleName ${meta.getSampleName()}' | tee out.txt"
}

Nextflow caching

When we start passing custom classes through the workflow, it's important to understand a little about the Nextflow caching mechanism. When a task is run, a unique hash is calculated based on the task name, the input files/values, and the input parameters. Our class extends from HashMap, which means that the hash will be calculated based on the contents of the HashMap. If we add a new method to the class, or amend a class method, this does not change the value of the objects in the hash, which means that the hash will not change.

Exercise

Can you show changing a method in our Metadata class does not change the hash?

Solution

We might increase the length of the adapter prefix to 5 characters:

1
2
3
    def getAdapterStart() {
        this.adapter?.substring(0, 5)
    }

Changing this method and resuming the workflow will not change the hash, and the existing method will be used.

We are not limited to using or extending the built-in Groovy classes. Let's start by creating a Dog class in ./lib/Dog.groovy:

1
2
3
4
class Dog {
    String name
    Boolean isHungry = true
}

We can create a new dog at the beginning of the workflow:

1
2
3
4
workflow {
    dog = new Dog(name: "fido")
    log.info "Found a new dog: $dog"
}

We can pass objects of our class through channels. Here we take a channel of dog names and create a channel of dogs:

1
2
3
4
5
workflow {
    Channel.of("Argente", "Absolon", "Chowne")
    | map { new Dog(name: it) }
    | view
}

If we try to use this new class in a resumed process, no caches will be used.

Exercise

Show that the Dog class is not cached when resuming a workflow.

6.3.2 Making a ValueObject

Nextflow has provided a decorator to help serialize your custom classes. By adding @ValueObject to the class definition, Nextflow will automatically serialize the class and cache it. This is useful if you want to pass a custom class through a channel, or if you want to use the class in a resumed workflow.

Let's add the decorator to our Dog class:

1
2
3
4
5
6
7
import nextflow.io.ValueObject

@ValueObject
class Dog {
    String name
    Boolean isHungry = true
}

Lastly, we will need to register the class with Kryo, the Java serialization framework. Again, Nextflow provides a helper method to do this. We can add the following to the main.nf file:

1
2
3
import nextflow.util.KryoHelper

KryoHelper.register(Dog)

Exercise

Show that the Dog class can now be used in processes and cached correctly.