SynCluster: A reaction type clustering and recommendation strategy for synthesis planning

This research proposes an innovative framework named SynCluster to investigate the impact of fine-grained reaction-type labels, rendered by clustering, on downstream tasks. The directory contains source code of the unpublished article: Tiantao et al's SynCluster: A reaction type clustering and recommendation strategy for synthesis planning Note: 1-77 clustering types refer to 0-76 labels in the running.

Download

git clone https://github.com/Yoko1030/SynCluster.git
cd SynCluster

Dataset

USPTO-50K: https://github.com/Hanjun-Dai/GLN
USPTO-500-MT: https://github.com/HelloJocelynLu/t5chem
We have sorted out the original data and the training weights, which can be downloaded from the following address: https://drive.google.com/drive/folders/19pFDvQ31NJMT2kmccFtpFVZ3pfWlU6lf?usp=sharing. The data should be put in clustering/scripts/data. "classifier_best.pth" should be put into Classifier/models. "SynCluster_best.pt" should be put into model (not combined with augment and alignment).

Environment

conda env create -f environment.yaml
conda activate SynCluster

Clustering

cd clustering/scripts
python 01_data_preparation.py
python 02_template_correction.py
python 03_cluster.py
python 04_output.py

Classifier (not included in reagent's prediciton)

cd ../../Classifier/fp
python Train2.py
python out.py

Tansformer data preparation (non-augment)

cd ../../preprocess
python preprocess.py

Tansformer Training (non-augment)

cd ..
sh train.sh

Inference (non-augment)

sh inference.sh

The results will be saved in output/output