sustainable-processes/chemical-parameter-sharing

A meta learning approach to reaction condition prediction using the implications of named reactions

chemical-parameter-sharing

A parameter sharing approach to reaction condition prediction using chemical reactions clustered by reaction class.

Installation

First clone the repository using Git.

Then execute the following commands in the root of the repository (stolen from: https://github.com/MolecularAI/reaction_utils)

conda env create -f env-dev.yml
conda activate cps-env
poetry install

Sources of code:

Reaction difference FP: https://pubs.acs.org/doi/abs/10.1021/ci5006614

NN condition prediction (Gao model): https://pubs.acs.org/doi/full/10.1021/acscentsci.8b00357

Code workflow

1. Data
   • ORDerly-condition: https://openreview.net/forum?id=R8FQMsECIS&noteId=y1HEbmW7QZ
   • Source: USPTO
   • Format: SMILES reaction string
2. Reaction clustering
   • NameRxn from Nextmove software
3. NN condition prediction conditioned on reaction class
   • Hard vs soft parameter sharing: https://avivnavon.github.io/blog/parameter-sharing-in-deep-learning/
   • We introduced conditional layers for condition prediction