recommendation_engine

Repo for the recommendation engine that was part of the DRP project

Recommender Pipeline

Steps to implement a recommender pipeline (Specific implementation of this pipeline is available in ./recommender/recommender_pipeline.py)

1. Generate reaction features

   • Get the chemicals in a reaction. For DRP these are referred to as triples
   • Generate descriptors for each of the chemicals in the reaction
   • Generate a sampling grid of reaction parameters
   • Expand grid by associating descriptors with each point on the grid

2. Run trained models with the reaction Sieve

   • Get a trained machine learning model
   • Filter sampling grid by running it through the ML model
   • Make a list of all the potentially successful reactions as predicted by the ML model

3. Recommend reactions

   • Calculate the mutual information of the potential reactions as compared to the already completed reactions
   • Select the top 'k' reactions with the highest MI

Progress

• Generate Reaction features
• Reaction Sieve
• Reaction Recommender
• Test and evaluate against Nature paper