Using wildcards to generalize your rules

As we showed in Chapter 6, when you have repeated substrings between input and output, you can extract them into wildcards - going from a rule that makes specific outputs from specific inputs into rules that operate on any input/output sets that match a pattern.

For example, the following code creates a single sourmash sketch from a specific genome:

rule sketch_genomes_1:
    input:
        "genomes/GCF_000017325.1.fna.gz",
    output:
        "GCF_000017325.1.fna.gz.sig",
    shell: """
        sourmash sketch dna -p k=31 {input} --name-from-first
    """

While this rule does the same for any genome ending in .fna.gz!

rule sketch_genomes_1:
    input:
        "genomes/{accession}.fna.gz",
    output:
        "{accession}.fna.gz.sig",
    shell: """
        sourmash sketch dna -p k=31 {input} \
            --name-from-first
    """

Here, {accession} is a wildcard that "fills in" as needed for any filename that is under the genomes/ directory and ends with .fna.gz.

Snakemake uses simple pattern matching to determine the value of {accession} - if asked for a filename ending in .fna.gz.sig, snakemake takes the prefix, and then looks for the matching input file genomes/{accession}.fna.gz, and fills in {input} accordingly.

Wildcards are incredibly useful and using them means that in many cases you can write a single rule that can generate hundreds or thousands of files! However, there are a few subtleties to consider. In this chapter, we're going to cover the most important of those subtleties, and provide links where you can learn more.

Rules for wildcards

First, let's go through some basic rules for wildcards.

Wildcards are determined by the desired output

The first and most important rule of wildcards is this: snakemake fills in wildcard values based on the filename it is asked to produce.

Consider the following rule:

# targets: result1.a.out

rule a:
    output: "{prefix}.a.out"
    shell: "touch {output}"

The wildcard in the output block will match any file that ends with .a.out, and the associated shell command will create it! This is both powerful and constraining: you can create any file with the suffix .a.out - but you also need to ask for the file to be created.

This means that in order to make use of this rule, there needs to be another rule that has a file that ends in .a.out as a required input. (You can also explicitly ask for such a file on the command line. CTB doc link.) There's no other way for snakemake to guess at the value of the wildcard: snakemake follows the dictum that explicit is better than implicit, and it will not guess at what files you want created.

For example, the above rule could be paired with another rule that asks for one or more filenames ending in .a.out:

rule make_me_a_file:
    input:
        "result1.a.out",
        "result2.a.out",

This also means that once you put a wildcard in a rule, you can no longer run that rule by the rule name - you have to ask for a filename, instead. If you try to run a rule that contains a wildcard but don't tell it what filename you want to create, you'll get:

Target rules may not contain wildcards.

One common way to work with wildcard rules is to have another rule that uses expand to construct a list of desired files; this is often paired with a glob_wildcards to load a list of wildcards. See the recipe for renaming files by prefix, below, or the chapter on Using expand to generate filenames.

All wildcards used in a rule must match to wildcards in the output: block

snakemake uses the wildcards in the output: block to fill in the wildcards elsewhere in the rule, so you can only use wildcards mentioned in one or more outputs.

This means that every wildcard used in the input: block needs to be present in output:. Consider the following example, where the input block contains a wildcard analysis that is not used in the output block:

# this does not work:

rule analyze_sample:
    input: "{sample}.x.{analysis}.in"
    output: "{sample}.out"

This doesn't work because snakemake doesn't know how to fill in the analysis wildcard in the input block, and you will get an error that says so:

WildcardError in line 1 of ...
Wildcards in input files cannot be determined from output files:
'analysis'

Think about it this way: if this worked, there could be multiple different input files for the same output, and snakemake would have no way to choose which input file to use to produce the desired output; moreover, the outputs would presumably be different depending on the inputs used, leading to irreproducibility.

Every wildcard in the input: block does need to be in the output: block. However, there are situations where wildcards in the output: block do not need to be in the input: block - see "Using wildcards to determine parameters to use in the shell block", below, on using wildcards to determine parameters for the shell block!

Wildcards are local to each rule

Wildcard names only need to match within a rule block; wildcards are not shared between rules. You can use the same wildcard names in multiple rules for consistency and readability, but snakemake will treat them as independent wildcards, and wildcard values will not be shared.

So, for example, these two rules use the same wildcard a in both rules -

rule analyze_this:
    input: "{a}.first.txt"
    output: "{a}.second.txt"

rule analyze_that:
    input: "{a}.second.txt"
    output: "{a}.third.txt"

but this is equivalent to these next two rules, which use different wildcards a and b in the separate rules:

rule analyze_this:
    input: "{a}.first.txt"
    output: "{a}.second.txt"

rule analyze_that:
    input: "{b}.second.txt"
    #        ^-- different - 'b' instead of 'a' in the first rule
    output: "{b}.third.txt"
    #        ^-- different - 'b' instead of 'a' in the first rule

There is one exception to the rule that wildcards are independent: when you use global wildcard constraints to limit wildcard matching by wildcard name, the constraints apply across all uses of that wildcard name in the Snakefile. However, the values of the wildcards remain independent - it's just the constraint that is shared by all wildcards of the same name.

While wildcards are independent in values and you could use different wildcards in every rule, it is a good convention to choose wildcards to have the same semantic meaning across the Snakefile - e.g. always use sample consistently to refer to a sample identifier, or accession to refer to a database ID. This makes reading the Snakefile easier!

One interesting addendum: because wildcards are local to each rule, you are free to match different parts of patterns in different rules! See "Mixing and matching wildcards", below.

The wildcard namespace is implicitly available in input: and output: blocks, but not in other blocks.

Within the input: and output: blocks in a rule, you can refer to wildcards directly by name. If you want to use wildcards in most other parts of a rule you need to use the wildcards prefix; the only exception to this rule is params: blocks (see the chapter params: blocks and {params}). Here, wildcards is a namespace, which we will talk about more later. (CTB)

Consider this Snakefile:

# this does not work:

rule analyze_this:
    input: "{a}.first.txt"
    output: "{a}.second.txt"
    shell: "analyze {input} -o {output} --title {a}"

Here you will get an error,

NameError: The name 'a' is unknown in this context. Did you mean 'wildcards.a'?

As the error suggests, you need to use wildcards.a in the shell block instead:

rule analyze_this:
    input: "{a}.first.txt"
    output: "{a}.second.txt"
    shell: "analyze {input} -o {output} --title {wildcards.a}"

Wildcards match as broadly as possible, unless constrained in some way

Wildcard pattern matching chooses the longest possible match to any characters, which can result in slightly confusing behavior. Consider:

rule all:
    input:
        "x.y.z.gz"

rule something:
    input: "{prefix}.{suffix}.txt"
    output: "{prefix}.{suffix}.gz"
    shell: "gzip -c {input} > {output}"

In the something rule, for the desired output file x.y.z.gz, {prefix} will currently be x.y and {suffix} will be z. But it would be equally valid for {prefix} to be x and suffix to be y.z.

A more extreme example shows the greedy matching even more clearly:

rule all:
    input:
        "longer_filename.gz"

rule something:
    input: "{prefix}{suffix}.txt"
    output: "{prefix}{suffix}.gz"
    shell: "gzip -c {input} > {output}"

where {suffix} is reduced down to a single character, e, and {prefix} is longer_filenam!

Two simple rules for wildcard matching are:

  • all wildcards must match at least one character.
  • after that, wildcards will match greedily: each wildcard will match everything it can before the next wildcard is considered.

This is why it's good practice to use wildcard constraints to limit wildcard matching. See "Constraining wildcards to avoid subdirectories and/or periods", below, for some examples, and see the wildcard constraints chapter for more details!

Some examples of wildcards

Running one rule on many files

Wildcards can be used to run the same simple rule on many files - this is one of the simplest and most powerful uses for snakemake!

Consider this Snakefile for compressing many files:

rule all:
    input:
        "compressed/F3D141_S207_L001_R1_001.fastq.gz",
        "compressed/F3D141_S207_L001_R2_001.fastq.gz",
        "compressed/F3D142_S208_L001_R1_001.fastq.gz",
        "compressed/F3D142_S208_L001_R2_001.fastq.gz"

rule gzip_file:
    input:
        "original/{filename}"
    output:
        "compressed/{filename}.gz"
    shell:
        "gzip -c {input} > {output}"

This Snakefile specifies a list of compressed files that it wants produced, and relies on wildcards to do the pattern matching required to find the input files and fill in the shell block.

See Replacing for loops with Snakefiles for more examples of this powerful pattern!

That having been said, this Snakefile is inconvenient to write and is somewhat error prone:

  • writing out the files individually is annoying if you have many of them!
  • to generate the list of files, you have to hand-rename them, which is error prone!

Snakemake provides several features that can help with these issues. You can load the list of files from a text file or spreadsheet, or get the list directly from the directory using glob_wildcards; and you can use expand to rename them in bulk. Read on for some examples!

Why is this better than using gzip directly?

It is possible to accomplish the same task by using gzip -k original/*, although you'd have to move the files into their final location, too.

How is using gzip -k original/* different from using snakemake? And is it better?

First, while the results aren't different - both approaches will compress the set of input files, which is what you want! - the gzip -k command runs in serial and will not run in parallel - that is, gzip will by default compress one file at a time. The Snakefile will run the rule gzip_file in parallel, using as many processors as you specify with -j. That means that if you had many, many such files - a common problem in bioinformatics! - the snakemake version could potentially run many times faster.

Second, specifying many files on the command line with gzip -k original/* works with gzip but not with every shell command. Some commands only run on one file at a time; gzip just happens to work whether you give it one or many files. Many other programs do not work on multiple input files; e.g. the fastp program for preprocessing FASTQ files runs on one dataset at a time. (It's also worth mentioning that snakemake gives you a way to flexibly write custom command lines; for some examples, see the chapter on Input and Output Blocks.)

Third, in the Snakefile we are being explicit about which files we expect to exist after the rules are run, while if we just ran gzip -k original/* we are asking the shell to compress every file in original/. If we accidentally deleted a file in the original subdirectory, then gzip would not know about it and would not complain - but snakemake would. This is a theme that will come up repeatedly - it's often safer to be really explicit about what files you expect, so that you can be alerted to possible mistakes.

And, fourth, the Snakefile approach will let you rename the output files in interesting ways - with gzip -k original/*, you're stuck with the original filenames. This is a feature we will explore in the next subsection!

Renaming files by prefix using glob_wildcards

Consider a set of files named like so:

F3D141_S207_L001_R1_001.fastq
F3D141_S207_L001_R2_001.fastq

within the original/ subdirectory.

Now suppose you want to rename them all to get rid of the _001 suffix before .fastq. This is very easy with wildcards!

The below Snakefile uses glob_wildcards to load in a list of files from a directory and then make a copy of them with the new name under the renamed/ subdirectory. Here, glob_wildcards extracts the {sample} pattern from the set of available files in the directory:

# first, find matches to filenames of this form:
files = glob_wildcards("original/{sample}_001.fastq")

# next, specify the form of the name you want:
rule all:
    input:
        expand("renamed/{sample}.fastq", sample=files.sample)

# finally, give snakemake a recipe for going from inputs to outputs.
rule rename:
    input:
        "original/{sample}_001.fastq",
    output:
        "renamed/{sample}.fastq"
    shell:
        "cp {input} {output}"

This Snakefile also makes use of expand to rewrite the loaded list into the desired set of filenames. This means that we no longer have to write out the list of files ourselves - we can let snakemake do it! expand is discussed further in Using expand to generate filenames.

Note that here you could do a mv instead of a cp and then glob_wildcards would no longer pick up the changed files after running.

This Snakefile loads the list of files from the directory itself, which means that if an input file is accidentally deleted, snakemake won't complain. When renaming files, this is unlikely to cause problems; however, when running workflows, we recommend loading the list of samples from a text file or spreadsheet to avoid problems

Also note that this Snakefile will find and rename all files in original/ as well as any subdirectories! This is because glob_wildcards by default includes all subdirectories. See the next section below to see how to use wildcard constraints to prevent loading from subdirectories.

Constraining wildcards to avoid subdirectories and/or periods

Wildcards match to any string, including '/', and so glob_wildcards will automatically find files in subdirectories and will also "stretch out" to match common delimiters in filenames such as '.' and '-'. This is commonly referred to as "greedy matching" and it means that sometimes your wildcards will match to far more of a filename than you want! You can limit wildcard matches using wildcard constraints.

Two common wildcard constraints are shown below, separately and in combination. The first constraint avoids files in subdirectories, and the second constraint avoids periods.

# match all .txt files - no constraints
all_files = glob_wildcards("{filename}.txt").filename

# match all .txt files in this directory only - avoid /
this_dir_files = glob_wildcards("{filename,[^/]+}.txt").filename

# match all files with only a single period in their name - avoid .
prefix_only = glob_wildcards("{filename,[^.]+}.txt").filename

# match all files in this directory with only a single period in their name
# avoid / and .
prefix_and_dir_only = glob_wildcards("{filename,[^./]+}.txt").filename

See Wildcard constraints for more information and details.

Advanced wildcard examples

Renaming files using multiple wildcards

The first renaming example above works really well when you want to change just the suffix of a file and can use a single wildcard, but if you want to do more complicated renaming you may have to use multiple wildcards.

Consider the situation where you want to rename files from the form of F3D141_S207_L001_R1_001.fastq to F3D141_S207_R1.fastq. You can't do that with a single wildcard, unfortunately - but you can use two, like so:

# first, find matches to filenames of this form:
files = glob_wildcards("original/{sample}_L001_{r}_001.fastq")

# next, specify the form of the name you want:
rule all:
    input:
        expand("renamed/{sample}_{r}.fastq", zip,
               sample=files.sample, r=files.r)

# finally, give snakemake a recipe for going from inputs to outputs.
rule rename:
    input:
        "original/{sample}_L001_{r}_001.fastq",
    output:
        "renamed/{sample}_{r}.fastq"
    shell:
        "cp {input} {output}"

We're making use of three new features in this code:

First, glob_wildcards is matching multiple wildcards, and puts the resulting values into a single result variable (here, files).

Second, the matching values are placed in two ordered lists, files.sample and files.r, such that values extracted from file names match in pairs.

Third, when we use expand, we're asking it to "zip" the two lists of wildcards together, rather than the default, which is to make all possible combinations with product. See Using expand to generate filenames for more information on zip vs product.

Also - as with the previous example, this Snakefile will find and rename all files in original/ as well as any subdirectories!

Links:

Mixing and matching strings

A somewhat nonintuitive (but also very useful) consequence of wildcards being local to rules is that you can do clever string matching to mix and match generic rules with more specific rules.

Consider this Snakefile, in which we are mapping reads from multiple samples to multiple references (rule map_reads_to_reference) as well as converting SAM to BAM files:

rule all:
    input:
        "sample1.x.ecoli.bam",
        "sample2.x.shewanella.bam",
        "sample1.x.shewanella.bam"

rule map_reads_to_reference:
    input:
        reads="{sample}.fq",
        reference="{genome}.fa",
    output:
        "{reads}.x.{reference}.sam"
    shell: "minimap2 -ax sr {input.reference} {input.reads} > {output}"
        
rule convert_sam_to_bam:
    input:
        "{filename}.sam"
    output:
        "{filename}.bam"
    shell: "samtools view -b {input} -o {output}

Here, snakemake is happily using different wildcards in each rule, and matching them to different parts of the pattern! So,

  • Rule convert_sam_to_bam will generically convert any SAM file to a BAM file based solely on the .bam and .sam suffixes.

  • However, map_reads_to_references will only produce mapping files that match the pattern of {sample}.x.{reference}, which in turn depend on the existence of {reference}.fa and {sample}.fastq.

This works because, ultimately, snakemake is just matching strings and does not "know" anything about the structure of the strings that it's matching. And it also doesn't remember wildcards across rules. So snakemake will happily match one set of wildcards in one rule, and a different set of wildcards in another rule!

Using wildcards to determine parameters to use in the shell block.

You can also use wildcards to build rules that produce output files where the parameters used to generate the contents are based on the filename; for example, consider this example of generating subsets of FASTQ files:

rule all:
    input:
        "big.subset100.fastq"

rule subset:
    input:
        "big.fastq"
    output:
        "big.subset{num_lines}.fastq"
    shell: """
        head -{wildcards.num_lines} {input} > {output}
    """

Here, the wildcard is only in the output filename, not in the input filename. The wildcard value is used by snakemake to determine how to fill in the number of lines for head to select from the file!

This can be really useful for generating files from giving many different parameters to a shell command - what we call "parameter sweeps". More about this later!

How to think about wildcards

Wildcards (together with expand and glob_wildcards) are among the most powerful and useful features in snakemake: they permit generic application of rules to an arbitrary number of files, based entirely on simple patterns.

However, with that power comes quite a bit of complexity!

Ultimately, wildcards are all about strings and patterns. Snakemake is using pattern matching to extract patterns from the desired output files, and then filling those matches in elsewhere in the rule. Most of the ensuing complexity comes avoiding ambiguity in matching and filling in patterns, along with the paired challenge of constructing all the names of the files you actually want to create.

Additional references

See also: the snakemake docs on wildcards.