Self-supervised molecular pretraining strategy for low-resource reaction prediction scenarios

This is the code for "Self-supervised molecular pretraining strategy for low-resource reaction prediction scenarios" paper.

Conda Environemt Setup

conda env create -f environment.yml

Dataset

The data for training, dev and testing of Baeyer-Villiger reaction are provided in data/Baeyer-Villiger reaction file. The data for training, dev and testing of Heck prediction are provided in data/Heck reaction file. The data for training, dev and testing of C-C bond formation reaction are provided in data/ C-C bond formation reaction file. The data for training, dev and testing of Functional group interconversion reaction are provided in data/ Functional group interconversion reaction file. The data for reaction transfer learning are provided in data/ USPTO reaction file. The molecular ChEMBL data can be found https://www.ebi.ac.uk/chembl/ The molecular ZINC data can be found http://zinc.docking.org/tranches/home/

Quickstart

Step 1: train the model in different reaction prediction

run train_transformer.py get a baseline model

Step 2. train the molecule-pretrained model in different reaction prediction

run train_downstream.py get the molecule-pretrained-Mass model

Step 3. test

run test.py