MMV model training image

Basic install and config

• Install docker. Install git if you're under windows as well.

  • https://docs.docker.com/install/
  • https://git-scm.com/downloads

You may need to configure Docker to let it use all available system memory if you're running this under Mac or Windows.

On Mac:

On Windows (click on the whale icon on the windows tray icons area):

On Windows, you will need to explicitly configure volume sharing across your host and the container:

Build the image and train the model

Run the following commands to set the docker image up and train the model. Use a terminal in either Linux or Mac and Windows Power Shell on Windows.

• Clone the repository, cd into the directory and build the docker image. Be sure you have write rights on the directory (ex: avoid clonning to c:\WINDOWS\system32):

  git clone https://github.com/chembl/mmv_train_image.git
  cd mmv_train_image
  docker build -t chemblgroup/mmv_train .
  

• In case you experience any kind of difficulties building the image behind a proxy, you can skip the 'docker build' command and instead of it run:

  docker pull chemblgroup/mmv_train
  

• Training dataset have to be named training_set.csv. It must copied in model_train folder and formatted like model_train/training_set_sample.csv.

• Configure a model by editing the configuration file(model_train/model_configs.json) if needed. Available descriptors are:

  • Fingerprints: Select between fcfp or ecfp. Set the radius.
  • Physicochemical: alogp, mw, hba, hbd, rtb, n_h_atoms

• Run the Docker container using the previously generated image. The container will map your host's model_train directory to the container's /model_train directory allowing the collection of all output files.

  docker run -v /full/path/to/model_train:/model_train chemblgroup/mmv_train
  
  ex mac: docker run -v /Users/efelix/projects/mmv_train_image/model_train:/model_train chemblgroup/mmv_train
  ex linux: docker run -v /home/efelix/projects/mmv_train_image/model_train:/model_train chemblgroup/mmv_train
  ex windows: docker run -v C:\Users\efelix\projects\mmv_train_image\model_train:/model_train chemblgroup/mmv_train
  

The container will generate 5 files(description) for each model in model_train/outputs folder:

• modelX.json: the dump of the model.
• coverage_values_modelX.json: json file with a list of coverage values for each molecule in eMolecules dataset. Boxplots like the ones from the original publication will be generated with them.
• eMolecules_predictions_modelX.csv: eMolecules dataset predictions.
• internal_validation_report_modelX.json: classification metrics with training data using a 5 k-fold cross-validation.
• external_validation_report_modelX.json: classification metrics report with eMolecules predictions.

Send the required files to the EBI.