obelix-ml-pipeline

This Github repo contains the ML pipeline used in the paper '...'. For the exact 'production' versions used to generate the figures in the paper, visit ...

First, homogeneous catalyst structures are featurized using our OBeLiX workflow. Various other representations are created for these structures and used as input for machine learning models. The models are trained to predict the enantioselectivity or conversion.

Installation

First clone the repository:

git clone https://github.com/EPiCs-group/obelix-ml-pipeline

Next, install or load conda and create a new environment from the environment.yml file:

conda env create -f environment.yml

Then activate the environment:

conda activate obelix-ml-pipeline

Afterwards, install the package:

pip install -e .

Then the notebooks can be run to run examples of the pipeline. In your anaconda prompt, run:

cd obelix-ml-pipeline/notebooks
jupyter notebook

Repository structure

In the paper, 4 different prediction tasks are performed. Fully out-of-domain, partially out-of-domain, in-domain and monte-carlo in-domain. The functions to perform these tasks are kept in their own files. Example use cases are shown in the notebooks. The ligand representations, substrate representations and experimental response are loaded from the data folder. A more detailed description of the files is given below.

Data

• Experimental response/
  Contains the experimental response for the different substrates and solvents.

  • jnjdata_sm12378_MeOH_16h: contains the experimental response for 16 hours with one solvent.
  • jnjdata_sm12378_MeOH: contains the experimental response for at 1 hour for SM1/2/3 and 16 hours for the other substrates with one solvent.

• Ligand representations/
  Contains the various representations of the ligands.

  • raw_data_processing: contains the raw data and scripts needed to create each representation.

• Substrate representations/
  Contains the various representations of the substrates.

  • raw_data_processing: contains the raw data and scripts needed to create each representation.

Code

representation_variables.py: contains a selection of features for representations of the ligand or substrate. When the file itself is run, it will create correlation plots for the features in the file.

utilities.py: contains functions for data loading and general utilities.

load_representations.py: contains functions for loading representations of the ligand or substrate.

machine_learning.py: contains functions for machine learning, training and testing.

data_classes.py: contains classes for the data returned in the ML pipeline.

predictions_on_unseen_substrate.py: contains the functions to perform the fully out-of-domain prediction task.

predictions_on_unseen_substrate_filtered.py: contains the functions to fully out-of-domain prediction task, except in this case if a classification task is performed, ligands that are in the same class across all training substrates will be removed from the set. This was done to test how well the models work if ligands that always perform well are removed.

predictions_on_partially_unseen_substrate.py: contains the functions to perform the partially out-of-domain prediction task.

predictions_within_substrate_class.py: contains the functions to perform the in-domain prediction task.

predictions_within_substrate_class_for_random_subset.py: contains the functions to perform the monte-carlo in-domain prediction task.

Notes