Single drug prediction Workflow

Table of Contents

• Single drug prediction Workflow
  • Table of Contents
  • Description
  • Contents
    • Building Blocks
    • Workflows
    • Resources
    • Tests
  • Instructions
    • Local machine
      • Requirements
      • Usage steps
    • MareNostrum 4
      • Requirements in MN4
      • Usage steps in MN4
  • License
  • Contact

Description

Complementarily, the workflow supports single drug response predictions to provide a baseline prediction in cases where drug response information for a given drug and cell line is not available. As an input, the workflow needs basal gene expression data for a cell, the drug targets (they need to be known for untested drugs) and optionally CARNIVAL features (sub-network activity predicted with CARNIVAL building block) and predicts log(IC50) values. This workflow uses a custom matrix factorization approach built with Google JAX and trained with gradient descent. The workflow can be used both for training a model, and for predicting new drug responses.

The workflow uses the following building blocks in order of execution (for training a model):

1. Carnival_gex_preprocess
   • Preprocessed the basal gene expression data from GDSC. The input is a matrix of Gene x Sample expression data.
2. Progeny
   • Using the preprocessed data, it estimates pathway activities for each column in the data (for each sample). It returns a matrix of Pathways x Samples with activity values for 11 pathways.
3. Omnipath
   • It downloads latest Prior Knowledge Network of signalling. This building block can be ommited if there exists already a csv file with the network.
4. TF Enrichment
   • For each sample, transcription factor activities are estimated using Dorothea.
5. CarnivalPy
   • Using the TF activities estimated before, it runs Carnival to obtain a sub-network consistent with the TF activities (for each sample).
6. Carnival_feature_merger
   • Preselect a set of genes by the user (if specified) and merge the features with the basal gene expression data.
7. ML Jax Drug Prediction
   • Trains a model using the combined features to predict IC50 values from GDSC.

For details on individual workflow steps, please check the scripts that use each individual building block in the workflow GitHub repository

Contents

Building Blocks

The BuildingBlocks folder contains the script to install the Building Blocks used in the Single Drug Prediction Workflow.

Workflows

The Workflow folder contains the workflows implementations.

Currently contains the implementation using PyCOMPSs.

Resources

The Resources folder contains a small dataset for testing purposes.

Tests

The Tests folder contains the scripts that run each Building Block used in the workflow for a small dataset. They can be executed individually without PyCOMPSs installed for testing purposes.

Instructions

Local machine

This section explains the requirements and usage for the Single Drug Prediction Workflow in a laptop or desktop computer.

Requirements

• permedcoe package
• PyCOMPSs
• Singularity

Usage steps

1. Clone this repository:

git clone https://github.com/PerMedCoE/single-drug-prediction-workflow.git

2. Install the Building Blocks required for the COVID19 Workflow:

single-drug-prediction-workflow/BuildingBlocks/./install_BBs.sh

3. Get the required Building Block images from the project B2DROP:

• Required images:
  • toolset.singularity
  • carnivalpy.singularity
  • ml-jax.singularity

The path where these files are stored MUST be exported in the PERMEDCOE_IMAGES environment variable.

:warning: TIP: These containers can be built manually as follows (be patient since some of them may take some time):

1. Clone the BuildingBlocks repository

   git clone https://github.com/PerMedCoE/BuildingBlocks.git
   

2. Build the required Building Block images

   cd BuildingBlocks/Resources/images
   ## Download new BB singularity files
   wget https://github.com/saezlab/permedcoe/archive/refs/heads/master.zip
   unzip master.zip
   cd permedcoe-master/containers
   ## Build containers
   cd toolset
   sudo /usr/local/bin/singularity build toolset.sif toolset.singularity
   mv toolset.sif ../../../
   cd ..
   cd carnivalpy
   sudo /usr/local/bin/singularity build carnivalpy.sif carnivalpy.singularity
   mv carnivalpy.sif ../../../
   cd ..
   cd ml-jax
   sudo /usr/local/bin/singularity build ml-jax.sif ml-jax.singularity
   mv ml-jax.sif ../../../tf-jax.sif
   cd ..
   cd ../..
   ## Cleanup
   rm -rf permedcoe-master
   rm master.zip
   cd ../../..
   

   :warning: TIP: The singularity containers can to be downloaded from: https://cloud.sylabs.io/library/pablormier

If using PyCOMPSs in local PC (make sure that PyCOMPSs in installed):

4. Go to Workflow/PyCOMPSs folder

   cd Workflows/PyCOMPSs
   

5. Execute ./run.sh

The execution is prepared to use the singularity images that MUST be placed into BuildingBlocks/Resources/images folder. If they are located in any other folder, please update the run.sh script setting the PERMEDCOE_IMAGES to the images folder.

TIP: If you want to run the workflow with a different dataset, please update the run.sh script setting the dataset variable to the new dataset folder and their file names.

MareNostrum 4

This section explains the requirements and usage for the Single Drug Prediction Workflow in the MareNostrum 4 supercomputer.

Requirements in MN4

• Access to MN4

All Building Blocks are already installed in MN4, and the Single Drug Prediction Workflow available.

Usage steps in MN4

1. Load the COMPSs, Singularity and permedcoe modules

   export COMPSS_PYTHON_VERSION=3
   module load COMPSs/3.1
   module load singularity/3.5.2
   module use /apps/modules/modulefiles/tools/COMPSs/libraries
   module load permedcoe
   

   TIP: Include the loading into your ${HOME}/.bashrc file to load it automatically on the session start.

   This commands will load COMPSs and the permedcoe package which provides all necessary dependencies, as well as the path to the singularity container images (PERMEDCOE_IMAGES environment variable) and testing dataset (SINGLE_DRUG_PREDICTION_WORKFLOW_DATASET environment variable).

2. Get a copy of the pilot workflow into your desired folder

   mkdir desired_folder
   cd desired_folder
   get_single_drug_prediction_workflow
   

3. Go to Workflow/PyCOMPSs folder

   cd Workflow/PyCOMPSs
   

4. Execute ./launch.sh

This command will launch a job into the job queuing system (SLURM) requesting 2 nodes (one node acting half master and half worker, and other full worker node) for 20 minutes, and is prepared to use the singularity images that are already deployed in MN4 (located into the PERMEDCOE_IMAGES environment variable). It uses the dataset located into ../../Resources/data folder.

:warning: TIP: If you want to run the workflow with a different dataset, please edit the launch.sh script and define the appropriate dataset path.

After the execution, a results folder will be available with with Single Drug Prediction Workflow results.

License

Apache 2.0

Contact

This software has been developed for the PerMedCoE project, funded by the European Commission (EU H2020 951773).