Reactivity predictions for substitution reactions

This repository contains code and model for predicting regio-selectivity for substitution reactions as described in Regio-Selectivity Prediction with a Machine-Learned Reaction Representation and On-the-Fly Quantum Mechanical Descriptors

Requirements

python 3.7
tensorflow 2.0.0
rdkit
qmdesc (python package for predicting QM descriptors on the fly)

Conda environment

To set up a conda environment:

conda env create --name <env-name> --file environment.yml

Data

In order to train the model, you must provide training data containing reactions (as reaction SMILES with mapped atom number) and potential products (as molecular SMILES strings with mapped atom number).

The data file must be a CSV file that must include reaction_id, rxn_smiles, and products_run in the header row. An example input file is provided as in data_example.csv

,reaction_id,rxn_smiles,products_run
0,244c06c0151311ea81f9b7db9d39a498,[Br:1][Br:2].[OH:3][c:4]1[cH:5][cH:6][cH:7][cH:8][c:9]1[F:10]>>[Br:2][c:5]1[c:4]([OH:3])[c:9]([F:10])[cH:8][cH:7][cH:6]1,[Br:1][c:5]1[c:4]([OH:3])[c:9]([F:10])[cH:8][cH:7][cH:6]1.[Br:1][c:6]1[cH:5][c:4]([OH:3])[c:9]([F:10])[cH:8][cH:7]1.[Br:1][c:7]1[cH:6][cH:5][c:4]([OH:3])[c:9]([F:10])[cH:8]1.[Br:1][c:8]1[cH:7][cH:6][cH:5][c:4]([OH:3])[c:9]1[F:10]
1,2182e760151311ea81f9b7db9d39a498,[F:1][F:2].[NH2:3][c:4]1[cH:5][cH:6][nH:7][c:8](=[O:9])[n:10]1>>[F:2][c:5]1[c:4]([NH2:3])[n:10][c:8](=[O:9])[nH:7][cH:6]1,[F:1][c:5]1[c:4]([NH2:3])[n:10][c:8](=[O:9])[nH:7][cH:6]1.[F:1][c:6]1[cH:5][c:4]([NH2:3])[n:10][c:8](=[O:9])[nH:7]1

in which, rxn_smiles are the reaction SMILES. And products_run are the potential products (major.minor1.minor2.....).

USPTO demo data

We provide three classes of reactions (aromatic CH functionalization, aromatic CX substitution, and other substitution reactions) to demonstrate our eventual model for the regio-selectivity predictions (Figure 7 in our paper). The .csv file for three classes of reactions are provided in the uspto_demo_data directory.

Training

This repo provides two model architectures as described in the paper.

GNN

A conventional graph neural network that relies only on the machine learned reaction representation of a given reaction. To train the model, run:

python reactivitiy.py -m GNN --data_path <path to the .csv file> --model_dir <directory to save the trained model>

For example, to train the model on CH functionalization reactions to predict the regio-selectivity:

python reactivitiy.py -m GNN --data_path uspto_demo_data/uspto_CH.csv --model_dir trained_model/GNN_uspto_CH

A checkpoint file, best_model.hdf5, will be saved in the trained_model/GNN_uspto_CH directory.

ml-QM-GNN

This is the novel fusion model introduced in the paper, which combines machine learned reaction representation and on-the-fly calculated QM descriptors. To use this architecture, the Chemprop-atom-bond must be installed. To train the model, run:

python reactivitiy.py --data_path <path to the .csv file> --model_dir <directory to save the trained model>

The reactivity.py use ml-QM-GNN mode by default. The workflow first predict QM atomic/bond descriptors for all reactants found in reactions. The predicted descriptors will then be scaled between [0, 1] through min-max scaler. A dictionary containing scikit-learn scaler object will be saved as scalers.pickle in the model_dir for later predicting task. A checkpoint file, best_model.hdf5 will also be saved in the model_dir

For example:

python reactivitiy.py --data_path uspto_demo_data/uspto_CH.csv --model_dir trained_model/ml_QM_GNN_uspto_CH

Predicting

To use the trained model, run:

python reactivitiy -m <mode> --data_path <path to the predicting .csv file> --model_dir <directory containing the trained model> -p

where data_path is path to the predicting .csv file, whose format is the same as the one discussed. model_dir is the directory holding the trained model. The model must be named as best_model.hdb5 and stores parameters only. The model_dir must also include a scalers.pickle under ml_QM_GNN mode as discussed in the training session.

We provide models trained on Pistachio regio-selective reactions, which are stored in the trained_model. For example:

python reactivitiy.py -m ml_QM_GNN --data_path uspto_demo_data/uspto_CH.csv --model_dir trained_model/ml_QM_GNN_CH -p

The predicted result will be saved as a '.csv' file in the user specified output directory by the flag --output_dir, which is output by default. An example of the predicted selectivity:

,rxn_id,predicted
0,86,"[0.9857026934623718, 0.004333616700023413, 0.00996393896639347]"
1,126,"[0.9388353824615479, 0.06116437166929245]"
2,7220,"[0.9834543466567993, 0.016545617952942848]"

where predicted column is the softmaxed selectivity score for each potential products.

yanfeiguan/reactivity_predictions_substitution