jidushanbojue / GraphRXN

GraphRXN

GraphRXN

Source code for our paper "A deep learning framework for accurate reaction prediction and its application on high-throughput experimentation data". The code was built based on CMPNN (https://github.com/SY575/CMPNN), DeepReac (https://github.com/bm2-lab/DeepReac), YieldBert (https://github.com/bm2-lab/DeepReac). Thanks a lot for their sharing.

Figure 1. Model architecture of GraphRXN

Figure 2. General workflow of HTE process

Figure 3. Reaction scheme and substrate scope

Figure 4. Distribution of Ratio(UV), where A represents amine, and B represents bromide

Figure 5. The scatter plots of GraphRXN on the entire dataset

Model Performance of three public datasets over ten-fold CV on test set

Dataset	Methods	R2	MAE	RMSE
Dataset1	GraphRXN-concat	0.951	4.30	5.98
Dataset1	GraphRXN-sum	0.937	4.85	6.80
Dataset1	Yield-BERT	0.951	4.00	6.03
Dataset1	DeepReac+	0.922	5.25	7.54
Dataset2	GraphRXN-concat	0.844	7.94	11.08
Dataset2	GraphRXN-sum	0.838	8.09	11.29
Dataset2	Yield-BERT	0.815	8.13	12.08
Dataset2	DeepReac+	0.827	8.06	11.65
Dataset3	GraphRXN-concat	0.892	0.16	0.23
Dataset3	GraphRXN-sum	0.881	0.18	0.24
Dataset3	Yield-BERT	0.886	0.16	0.24
Dataset3	DeepReac+	0.853	0.18	0.25

Model performance of in-house dataset over 5-fold CV on test set

Groupe	Size	methods	R2	MAE	RMSE
Entire	1558	GraphRXN-concat	0.713	0.06	0.09
Entire	1558	GraphRXN-sum	0.704	0.06	0.09
Entire	1558	Yield-BERT	0.645	0.10	0.07
Entire	1558	DeepReac+	0.610	0.07	0.10
G1	317	GraphRXN-concat	0.661	0.08	0.11
G1	317	GraphRXN-sum	0.462	0.11	0.14
G1	317	Yield-BERT	0.718	0.07	0.10
G1	317	DeepReac+	0.551	0.09	0.13
G2	419	GraphRXN-concat	0.629	0.05	0.07
G2	419	GraphRXN-sum	0.592	0.06	0.07
G2	419	Yield-BERT	0.512	0.06	0.08
G2	419	DeepReac+	0.528	0.06	0.08
G3	401	GraphRXN-concat	0.802	0.06	0.08
G3	401	GraphRXN-sum	0.775	0.06	0.08
G3	401	Yield-BERT	0.785	0.06	0.08
G3	401	DeepReac+	0.745	0.07	0.09
G4	421	GraphRXN-concat	0.459	0.08	0.12
G4	421	GraphRXN-sum	0.419	0.09	0.12
G4	421	Yield-BERT	0.503	0.08	0.11
G4	421	DeepReac+	0.23	0.10	0.14

Quick start

GraphRXN

conda env create -f GraphRXN.yaml ### Create GraphRXN env
conda activate GraphRXN
python reaction_train.py  --data_path data_scaler/Buchward-Hartwig/random_split/FullCV_01_train_temp_scaler.csv
                          --separate_test_path data_scaler/Buchward-Hartwig/random_split/FullCV_01_test_temp_scaler.csv
                          --dataset_type regression 
                          --num_folds 1 
                          --gpu 0 
                          --epochs 100 
                          --batch_size 128 
                          --save_dir ./result/Buchward/concat_01_temp
                          --metric r2 
                          --reaction_agg_method concat
Note: If choosing summation aggregation method, please specify --reaction_agg_method sum

DeepReac+

cd DeepReac
conda env create -f DeepReact.yaml ### Create DeepReac+ env
conda activate DeepReact
### train and predict
python DeepReac_train.py -train data_scaler/Buchward-Hartwig/random_split/FullCV_01_train_temp_scaler.csv
                         -test data_scaler/Buchward-Hartwig/random_split/FullCV_01_test_temp_scaler.csv
                         -epochs 100
                         -stats ./result_scaler/Buchward_01_test_stats.csv

Yield-BERT

cd Yield-BERT
conda env create -f rxnyields.yaml ### create Yield-BERT env
conda activate rxnyields

cd yield-BERT_baseline

### For Dataset 1 (Buchwald) training
python launch_buchwald_hartwig_training.py

### For Dataset 2 (Suzuki) training
python lauch_suzuki_miyaura_training.py

### For Dataset 3 (Denmark) training
python data3_training_10cv.py

### For in-house dataset training
python inhouse_data_transform.py