Skip to content

usage.md

Latest commit

 

History

History
303 lines (206 loc) · 18.5 KB

usage.md

File metadata and controls

303 lines (206 loc) · 18.5 KB

yocra3/NetActivityTrain: Usage

Introduction

NetActivityTrain trains a sparsely-connected autoencoder to encode gene expression measurements into gene set activity scores. NetActivityTrain requires four elements: a gene expression dataset, a mapping between genes and gene sets and a network architecture and its configuration.

Gene expression dataset

NetActivity requires that the gene expression data is normalized (counts are not accepted) and saved in a HDF5SummarizedExperiment. You can easily obtain these objects from R using DESeq2 and HDF5Array packages:

library(DESeq2)
library(HDF5Array)

dds <- DESeqDataSetFromMatrix(countData = counts,
                              colData = phenotype,
                              design = ~ 1)
# Normalize data with Variant Stabilizing Transformation
vst <- vst(dds, blind=FALSE)

## Create HDF5SummarizedExperiment
saveHDF5SummarizedExperiment(vst, prefix = "dataset_")

This command will generate two files: dataset_assays.h5 and dataset_se.rds. They can be passed to NetActivityTrain with:

--data_prefix 'dataset_'

Group-based train-test split

NetActivityTrain divides the dataset in samples for training and samples for testing. NetActivityTrain can perform the splitting by balancing samples based on a variable. For instance, when training the model in GTEx, we decided to balance samples by tissue in train and test. If we would like NetActivityTrain to consider a variable for the splitting, we can add a column to the colData called Group:

vst$Group <- phenotype_of_interest

Gene set mapping

NetActivityTrain requires a tab-delimited file with the mapping between genes and gene sets. This file should contain the columns Gene and GeneSetID:

GeneSetID Gene
GO:0006734 ENSG00000248746
GO:0006734 ENSG00000136872

You can find an example in the data folder. Notice that the identifiers of this file should map the identifiers in the rownames of the gene expression dataset.

Network definition

NetActivityTrain requires a python file with the network definition and another file with the network configuration.

The network definition file consists of a the definition of a function called model_generator_train, with four arguments: x_train, y_train, gene_mask and params:

def model_generator_train(x_train, y_train, gene_mask, params):

params represents an object with the parameters that can be edited with the network configuration file.

Then, we have the model definition using keras. Here, the user can specify different network architectures. Nonetheless, the user should use the class tfmot.sparsity.keras.ConstantSparsity to achieve the sparsely-connected layer:

  model = Sequential()
  model.add(tfmot.sparsity.keras.prune_low_magnitude(Dense(gene_mask.shape[1],
    input_dim = x_train.shape[1], activation = params.activation),
    tfmot.sparsity.keras.ConstantSparsity(0.5, 1000000, end_step =  1000000, frequency = 100)))
  if params.dense1_dropout:
        model.add(Dropout(params.dense1_dropout))
  model.add(Dense(x_train.shape[1]))

The additional layer of the tfmot.sparsity.keras.ConstantSparsity layer should be modified in order to set to 0 those connections of genes not present in a gene set:

  ## Add gene mask
  w = model.get_weights()
  w[2] = gene_mask
  model.set_weights(w)

Finally, we just need to define the optimization algorithm, the less and compile the model:

  opt = Adam(learning_rate = params.alpha)

  model.compile(loss='mse',
    optimizer = opt,
    metrics = ['mse'])
  return model

Notice that all dependencies states in this file should be explictly loaded at the beggining of the file. An example of a network definition file can be found in the data folder. This definition file defines a sparsely-connected autoencoder with one layer. We recommend the users to use it as a template for other structure definitions.

Finally, we can set the value of any variable defined using params. using the params file. This is very useful when testing different networks configurations or when aiming to test different trainings.

Running the pipeline

The typical command for running the pipeline is as follows:

nextflow run yocra3/NetActivityTrain --data_prefix SE_h5 --gene_mask gene_mask.txt --network network.py --network_params params.py --outdir <OUTDIR> -profile docker

This will launch the pipeline with the docker configuration profile. See below for more information about profiles.

Note that the pipeline will create the following files in your working directory:

work                # Directory containing the nextflow working files
<OUTDIR>            # Finished results in specified location (defined with --outdir)
.nextflow_log       # Log file from Nextflow
# Other nextflow hidden files, eg. history of pipeline runs and old logs.

Updating the pipeline

When you run the above command, Nextflow automatically pulls the pipeline code from GitHub and stores it as a cached version. When running the pipeline after this, it will always use the cached version if available - even if the pipeline has been updated since. To make sure that you're running the latest version of the pipeline, make sure that you regularly update the cached version of the pipeline:

nextflow pull yocra3/NetActivityTrain

Reproducibility

It is a good idea to specify a pipeline version when running the pipeline on your data. This ensures that a specific version of the pipeline code and software are used when you run your pipeline. If you keep using the same tag, you'll be running the same version of the pipeline, even if there have been changes to the code since.

First, go to the yocra3/NetActivityTrain releases page and find the latest pipeline version - numeric only (eg. 1.3.1). Then specify this when running the pipeline with -r (one hyphen) - eg. -r 1.3.1. Of course, you can switch to another version by changing the number after the -r flag.

This version number will be logged in reports when you run the pipeline, so that you'll know what you used when you look back in the future. For example, at the bottom of the MultiQC reports.

Core Nextflow arguments

NB: These options are part of Nextflow and use a single hyphen (pipeline parameters use a double-hyphen).

-profile

Use this parameter to choose a configuration profile. Profiles can give configuration presets for different compute environments.

Several generic profiles are bundled with the pipeline which instruct the pipeline to use software packaged using different methods (Docker, Singularity) - see below.

We highly recommend the use of Docker or Singularity containers for full pipeline reproducibility, however when this is not possible, Conda is also supported.

The pipeline also dynamically loads configurations from https://github.com/nf-core/configs when it runs, making multiple config profiles for various institutional clusters available at run time. For more information and to see if your system is available in these configs please see the nf-core/configs documentation.

Note that multiple profiles can be loaded, for example: -profile test,docker - the order of arguments is important! They are loaded in sequence, so later profiles can overwrite earlier profiles.

If -profile is not specified, the pipeline will run locally and expect all software to be installed and available on the PATH. This is not recommended, since it can lead to different results on different machines dependent on the computer enviroment.

  • test
    • A profile with a complete configuration for automated testing
    • Includes links to test data so needs no other parameters
  • docker
    • A generic configuration profile to be used with Docker
  • singularity
    • A generic configuration profile to be used with Singularity

-resume

Specify this when restarting a pipeline. Nextflow will use cached results from any pipeline steps where the inputs are the same, continuing from where it got to previously. For input to be considered the same, not only the names must be identical but the files' contents as well. For more info about this parameter, see this blog post.

You can also supply a run name to resume a specific run: -resume [run-name]. Use the nextflow log command to show previous run names.

-c

Specify the path to a specific config file (this is a core Nextflow command). See the nf-core website documentation for more information.

Custom configuration

Resource requests

Whilst the default requirements set within the pipeline will hopefully work for most people and with most input data, you may find that you want to customise the compute resources that the pipeline requests. Each step in the pipeline has a default set of requirements for number of CPUs, memory and time. For most of the steps in the pipeline, if the job exits with any of the error codes specified here it will automatically be resubmitted with higher requests (2 x original, then 3 x original). If it still fails after the third attempt then the pipeline execution is stopped.

For example, if the nf-core/rnaseq pipeline is failing after multiple re-submissions of the STAR_ALIGN process due to an exit code of 137 this would indicate that there is an out of memory issue:

[62/149eb0] NOTE: Process `NFCORE_RNASEQ:RNASEQ:ALIGN_STAR:STAR_ALIGN (WT_REP1)` terminated with an error exit status (137) -- Execution is retried (1)
Error executing process > 'NFCORE_RNASEQ:RNASEQ:ALIGN_STAR:STAR_ALIGN (WT_REP1)'

Caused by:
    Process `NFCORE_RNASEQ:RNASEQ:ALIGN_STAR:STAR_ALIGN (WT_REP1)` terminated with an error exit status (137)

Command executed:
    STAR \
        --genomeDir star \
        --readFilesIn WT_REP1_trimmed.fq.gz  \
        --runThreadN 2 \
        --outFileNamePrefix WT_REP1. \
        <TRUNCATED>

Command exit status:
    137

Command output:
    (empty)

Command error:
    .command.sh: line 9:  30 Killed    STAR --genomeDir star --readFilesIn WT_REP1_trimmed.fq.gz --runThreadN 2 --outFileNamePrefix WT_REP1. <TRUNCATED>
Work dir:
    /home/pipelinetest/work/9d/172ca5881234073e8d76f2a19c88fb

Tip: you can replicate the issue by changing to the process work dir and entering the command `bash .command.run`

For beginners

A first step to bypass this error, you could try to increase the amount of CPUs, memory, and time for the whole pipeline. Therefor you can try to increase the resource for the parameters --max_cpus, --max_memory, and --max_time. Based on the error above, you have to increase the amount of memory. Therefore you can go to the parameter documentation of rnaseq and scroll down to the show hidden parameter button to get the default value for --max_memory. In this case 128GB, you than can try to run your pipeline again with --max_memory 200GB -resume to skip all process, that were already calculated. If you can not increase the resource of the complete pipeline, you can try to adapt the resource for a single process as mentioned below.

Advanced option on process level

To bypass this error you would need to find exactly which resources are set by the STAR_ALIGN process. The quickest way is to search for process STAR_ALIGN in the nf-core/rnaseq Github repo. We have standardised the structure of Nextflow DSL2 pipelines such that all module files will be present in the modules/ directory and so, based on the search results, the file we want is modules/nf-core/software/star/align/main.nf. If you click on the link to that file you will notice that there is a label directive at the top of the module that is set to label process_high. The Nextflow label directive allows us to organise workflow processes in separate groups which can be referenced in a configuration file to select and configure subset of processes having similar computing requirements. The default values for the process_high label are set in the pipeline's base.config which in this case is defined as 72GB. Providing you haven't set any other standard nf-core parameters to cap the maximum resources used by the pipeline then we can try and bypass the STAR_ALIGN process failure by creating a custom config file that sets at least 72GB of memory, in this case increased to 100GB. The custom config below can then be provided to the pipeline via the -c parameter as highlighted in previous sections.

process {
    withName: 'NFCORE_RNASEQ:RNASEQ:ALIGN_STAR:STAR_ALIGN' {
        memory = 100.GB
    }
}

NB: We specify the full process name i.e. NFCORE_RNASEQ:RNASEQ:ALIGN_STAR:STAR_ALIGN in the config file because this takes priority over the short name (STAR_ALIGN) and allows existing configuration using the full process name to be correctly overridden.

If you get a warning suggesting that the process selector isn't recognised check that the process name has been specified correctly.

Updating containers (advanced users)

The Nextflow DSL2 implementation of this pipeline uses one container per process which makes it much easier to maintain and update software dependencies. If for some reason you need to use a different version of a particular tool with the pipeline then you just need to identify the process name and override the Nextflow container definition for that process using the withName declaration. For example, in the nf-core/viralrecon pipeline a tool called Pangolin has been used during the COVID-19 pandemic to assign lineages to SARS-CoV-2 genome sequenced samples. Given that the lineage assignments change quite frequently it doesn't make sense to re-release the nf-core/viralrecon everytime a new version of Pangolin has been released. However, you can override the default container used by the pipeline by creating a custom config file and passing it as a command-line argument via -c custom.config.

  1. Check the default version used by the pipeline in the module file for Pangolin

  2. Find the latest version of the Biocontainer available on Quay.io

  3. Create the custom config accordingly:

    • For Docker:

      process {
    withName: PANGOLIN {
        container = 'quay.io/biocontainers/pangolin:3.0.5--pyhdfd78af_0'
    }
}

    • For Singularity:

      process {
    withName: PANGOLIN {
        container = 'https://depot.galaxyproject.org/singularity/pangolin:3.0.5--pyhdfd78af_0'
    }
}

    • For Conda:

      process {
    withName: PANGOLIN {
        conda = 'bioconda::pangolin=3.0.5'
    }
}

NB: If you wish to periodically update individual tool-specific results (e.g. Pangolin) generated by the pipeline then you must ensure to keep the work/ directory otherwise the -resume ability of the pipeline will be compromised and it will restart from scratch.

nf-core/configs

In most cases, you will only need to create a custom config as a one-off but if you and others within your organisation are likely to be running nf-core pipelines regularly and need to use the same settings regularly it may be a good idea to request that your custom config file is uploaded to the nf-core/configs git repository. Before you do this please can you test that the config file works with your pipeline of choice using the -c parameter. You can then create a pull request to the nf-core/configs repository with the addition of your config file, associated documentation file (see examples in nf-core/configs/docs), and amending nfcore_custom.config to include your custom profile.

See the main Nextflow documentation for more information about creating your own configuration files.

If you have any questions or issues please send us a message on Slack on the #configs channel.

Azure Resource Requests

To be used with the azurebatch profile by specifying the -profile azurebatch. We recommend providing a compute params.vm_type of Standard_D16_v3 VMs by default but these options can be changed if required.

Note that the choice of VM size depends on your quota and the overall workload during the analysis. For a thorough list, please refer the Azure Sizes for virtual machines in Azure.

Running in the background

Nextflow handles job submissions and supervises the running jobs. The Nextflow process must run until the pipeline is finished.

The Nextflow -bg flag launches Nextflow in the background, detached from your terminal so that the workflow does not stop if you log out of your session. The logs are saved to a file.

Alternatively, you can use screen / tmux or similar tool to create a detached session which you can log back into at a later time. Some HPC setups also allow you to run nextflow within a cluster job submitted your job scheduler (from where it submits more jobs).

Nextflow memory requirements

In some cases, the Nextflow Java virtual machines can start to request a large amount of memory. We recommend adding the following line to your environment to limit this (typically in ~/.bashrc or ~./bash_profile):

NXF_OPTS='-Xms1g -Xmx4g'