This repository contains a reaction condition selector.
python 3.10
To run the reaction condition selector you need to make sure you have installed anaconda. The version about pytorch and cudatoolkit should be depended on your machine.
conda create -n Cluster_Predictor python=3.10 \
conda activate Cluster_Predictor \
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116 \
pip install pandas==2.1.0 \
pip install chemprop==1.6.1 \
pip install torch-geometric==2.3.0 \
conda install rdkit=2023.3.3 -c rdkit \We also provide the corresponding environment file, You can reproduce the environment directly through the provided .yaml file.
conda env create -f environment.yamlPlease first download the Cluster Predictor program from GitHub. It contains some of the data,the pre-trained model and the code needed to replicate the work. :
git clone https://github.com/wzhstat/Reaction-Condition-Selector.git \
cd ./Reaction-Condition-Selector Before training, make sure you have the data file, which should contain two parts: the csv filesdata_train.csv, data_test.csv, data_val.csv and the labels folder. The labels file can be downloaded directly from data, the csv files can be download in https://www.dropbox.com/scl/fo/v1rhyes2wvead9dz3x4fb/h?rlkey=nqtst7azldcry3ixnoigmcv3v&dl=0. You should put them like this:
./data/data_train.csv
./data/data_test.csv
./data/data_val.csv
./data/labelsYou can run the preprocessing.py file to get the preprocessed data.
python preprocessing.pyYou can run train.sh files directly to get models of catalyst, solvents, and reagents, we recommend using GPUs for faster training.
bash ./train.shTo train model for a particular condition constitute, take solv0 for example, you can run:
chemprop_train --target_columns solv0 --data_path ./data/MPNN_data/GCN_data_train.csv --separate_val_path ./data/MPNN_data/GCN_data_val.csv --separate_test_path ./data/MPNN_data/GCN_data_test.csv --dataset_type multiclass --multiclass_num_classes 542 --save_dir ./models/GCN_solv0 --reaction --extra_metrics accuracy --epochs 35You end up with a models folder that contains the trained D-MPNN model. You can also find trained models directly from models.
Before making a prediction, run BuildConditionLibrary.sh to build the three types of reaction condition libraries needed for prediction, namely the r1 library, the r0 library, and the r0* library.
bash ./BuildConditionLibrary.shThis will generate a condition_library folder under the data folder, which contains the following libraries:
./data/condition_library/condition_library_r1.json.gz
./data/condition_library/condition_library_r0.json.gz
./data/condition_library/condition_library_r0_1.json.gz
./data/condition_library/classed_conditions_library_r1.json.gz
./data/condition_library/classed_conditions_library_r0.json.gz
./data/condition_library/classed_conditions_library_r0_1.json.gzYou can run make_predictions.sh to complete the model predictions, which generates a prediction folder containing predictions with and without clustering. The results are saved in the data folder by default
bash ./make_predictions.shAn examples of prediction results that include clustering are as follows:
class id: 50_4
class label [[], ['phosphine', 'halide']]
best condition: ['None', 'None', 'None', 'BrP(Br)Br', 'None', 'None']
score: 0.43502089128291327
condition score: [["['None', 'None', 'None', 'BrP(Br)Br', 'None', 'None']", 0.43502089128291327], ["['None', 'None', 'None', 'O=P(Br)(Br)Br', 'None', 'None']", 0.4005131881456007], ["['None', 'CN(C)C=O', 'None', 'BrP(Br)Br', 'None', 'None']", 0.02543110079025514], ["['None', 'CN(C)C=O', 'None', 'O=P(Br)(Br)Br', 'None', 'None']", 0.023413797956965517]... ["['None', 'None', 'None', 'CCOC(C)=O.O=C([O-])O.[Na+]', 'O=P(Cl)(Cl)Cl', 'None']", 6.239001680815093e-37]]
--------------------------------------------
class id: 50_5
class label [[], ['halide', 'aldehyde', 'amine']]
best condition: ['None', 'CN(C)C=O', 'None', 'O=S(Cl)Cl', 'None', 'None']
score: 6.458456286464104e-14
condition score: [["['None', 'CN(C)C=O', 'None', 'O=S(Cl)Cl', 'None', 'None']", 6.458456286464104e-14], ["['None', 'None', 'None', 'CN(C)C=O', 'O=S(Cl)Cl', 'None']", 1.2491413029613758e-18], ["['None', 'ClCCl', 'None', 'CN(C)C=O', 'O=S(Cl)Cl', 'None']", 7.428030276709379e-22], ["['None', 'Cc1ccccc1', 'None', 'CN(C)C=O', 'O=S(Cl)Cl', 'None']", 3.421287350994334e-22], ["['None', 'ClCCl', 'None', 'CN(C)C=O', 'O=C(Cl)C(=O)Cl', 'None']", 1.1259136045695308e-23],...["['None', 'CN(C)C=O', 'None', 'O=C(Cl)C(=O)Cl', 'None', 'None']", 7.935796084468542e-26], ["['None', 'CN(C)C=O', 'None', 'CCN(C=O)CC', 'O=S(Cl)Cl', 'None']", 4.955859832749846e-32]]
...
--------------------------------------------
class id: 50_0
class label [['ionic'], ['phosphine', 'halide']]
best condition: ['C[N+](C)(C)C.[Cl-]', 'None', 'None', 'O=P(Cl)(Cl)Cl', 'None', 'None']
score: 9.45609355368023e-28
condition score: [["['C[N+](C)(C)C.[Cl-]', 'None', 'None', 'O=P(Cl)(Cl)Cl', 'None', 'None']", 9.45609355368023e-28], ["['CC[N+](CC)(CC)CC.[Cl-]', 'CC#N', 'None', 'CN(C)c1ccccc1', 'O=P(Cl)(Cl)Cl', 'None']", 4.047334473315928e-40]]Please run Score.py to calculate the accuracy of the model predictions.
python Score.py --data_path ./data/data_test.csv --pred_path ./data/predictionThis may take a while, and the calculated result will be printed out：
Top-1 Accuracy: 0.44624893250539177
Top-3 Accuracy: 0.6360386179744381
Top-5 Accuracy: 0.706326805332407
Top-10 Accuracy: 0.7854645018599737
Cluster Top-1 Accuracy: 0.6553620796966144
Cluster Top-3 Accuracy: 0.8564770608073878
Cluster Top-5 Accuracy: 0.9074066032683428
Cluster Top-10 Accuracy: 0.9645355855660255Our actual route data was artificially extracted from the article, and the processed JMC_data.csv file can be found in the data folder.
You can run make_JMC_predictions.sh to complete the model predictions, which generates a prediction folder containing predictions with and without clustering.
bash ./make_JMC_predictions.shPlease run Score.py to calculate the accuracy of the model predictions.
python Score.py --data_path ./data/JMC_data.csv --pred_path ./data/JMC_prediction