Yield_log_RRIM

This is a repository for the paper "log-RRIM: Yield Prediction via Local-to-global Reaction Representation Learning and Interaction Modeling".

Data processing

The raw datasets of USPTO500MT and Buchwald-Hartwig are provided in the folder 'Data'. The Data processing step filters out reactions without identifiable reaction centers. The sample code of this step is provided in the file USPTO_data_process.ipynb. And the filtered datasets are also provided, Buchwald-Hartwig: BH_processed.csv, USPTO500MT: USPTO500MT_train_processed_100.csv, USPTO500MT_valid_processed_100.csv, USPTO500MT_test_processed_100.csv, sampled_CJHIF: sample_test_CJ_processed_final_5w_wo0.csv.

Feature Generation

This step generates the reactant, reagent, and product features needed for log-RRIM. The sample code is USPTO_features_process.ipynb. Each dataset's generated reactant, reagent, and product features are also provided in the repository.

Training and testing

In scripts, test_BH.pbs and test_USPTO.pbs contain the commands to train and test the log-RRIM on different datasets use the basic atom features or the learned atom features, which is controlled by parameter '--use_pretrain 0'.

Analysis

The comparison and the analyses between log-RRIM and T5chem can be found in the folder results_analysis.

Package requirements

Important python packages and their versions in our station are listed below:
python===3.8.18
torch===2.1.0+cu121
numpy===1.21.6
scipy===1.10.1
rdkit===2023.9.1

Contact with us

Email: [email protected]