A graph neural network model for predicting reaction products. Based on "A graph-convolutional neural network model for the prediction of chemical reactivity." (DOI: 10.1039/C8SC04228D)
Using Anaconda is the easiest way, as RDKit is a required dependency which must be installed with conda, or built from source. First install Anaconda.
Next, it is recommended to create a conda environment specifically for this project. The code is developed using Python 3.6, and hasn't been tested for any other versions at this time. Create a conda environment with
conda create --name rxntorch python=3.6
Then, activate the new conda environment with
conda activate rxntorch
Next, install RDKit
conda install -c rdkit rdkit
Installing PyTorch depends on whether you will be using CUDA or not. For a CUDA enabled version
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
otherwise, for a CPU only version
conda install pytorch torchvision cpuonly -c pytorch
Next, install scikit-learn with
conda install scikit-learn
Finally, clone this repository to your local machine and install
git clone [email protected]:nsf-c-cas/rxntorch.git
cd rxntorch
Currently, only the first portion of the network is built, which predicts likely bond changes. There is a script in the root project directory named train_reactivity.py which loads the data, builds the model, and runs the training procedure. The dataset can be found in rxntorch/data/ in a compressed archive. First, expand the archived data with
tar -zxf train.txt.proc.tar.gz
for the training data, and similarly for the test and validation data. To run the network, a minimal example with the default hyperparameters is as follows:
python train_reactivity.py -c "train.txt.proc" -o "reactivity.model"