DeepVariant module / subworkflow options

The DeepVariant tool can be run using the deepvariant/rundeepvariant subcommand, or the subworkflow deepvariant, which calls the subcommands makeexamples, callvariants and postprocessvariants. The subcommand rundeepvariant is simpler, but the subworkflow may be useful if you want to run callvariants on GPU.

Conda is not supported at the moment

The bioconda recipe is not fully working as expected.

See https://github.com/bioconda/bioconda-recipes/issues/30310 and https://github.com/nf-core/modules/issues/1754 for more information.

Hence, we are using the docker container provided by the authors of the tool:

• google/deepvariant

This image is mirrored on the nf-core quay.io for convenience.

DeepVariant subworkflow

You can use the subworkflow nf-core/deepvariant, which integrates the three processes to perform variant calling with common file formats.

These module subcommands incorporate the individual steps of the DeepVariant pipeline:

* makeexamples: Converts the input alignment file to a tfrecord format suitable for the deep learning model
* callvariants: Call variants based on input tfrecords. The output is also in
tfrecord format, and needs postprocessing to convert it to vcf.
* postprocessvariants: Convert variant calls from callvariants to VCF, and
also create GVCF files based on genomic information from makeexamples.

Recommended parameters

makeexamples

This process imports the data used for calling, and thus decides what information is available to the deep neural network. It's important to use the correct settings for the model you want to use for each step. The script run_deepvariant.py does this automatically. To figure out the flags needed for each model, you can run run_deepvariant.py while adding dry_run=true, to print out the command used for each step, as described here.

callvariants

It is mandatory to specify a model type. The models are available on the container filesystem in /opt/models - specify the one you want with the --checkpoint argument.

withName: "DEEPVARIANT_CALLVARIANTS" {
    ext.args = '--checkpoint "/opt/models/wgs'
}

The channels specified in the makeexamples process must match the model used for calling.