Core code for the paper "A Model to Search for Synthesizable Molecules" (https://arxiv.org/abs/1906.05221) by John Bradshaw, Brooks Paige, Matt J. Kusner, Marwin H. S. Segler, José Miguel Hernández-Lobato. Updated to now run on PyTorch 1.2.
- Install the requirements (listed in requirements.txt). We're using Python 3.6.
- Get the submodules, i.e.
git submodule initthengit submodule update - Unzip the data folder:
data.zip-- this zip folder of processed USPTO data comes from https://github.com/wengong-jin/nips17-rexgen ^*^ - Add the correct modules to the
PYTHONPATH, e.g. this can be done bysource set_up.sh
PyTorch
We have updated the code such that it is using version 1.2 (or 1.3).
RDKit
We recommend installing rdkit through Anaconda, see e.g.
https://anaconda.org/rdkit/rdkit:
conda install -c rdkit rdkit
pytest Is used for running tests.
rdfilters
The quality filters are computed using the rd_filters library. This can be found
here, and can be installed using:
pip install git+https://github.com/PatWalters/rd_filters.git.
^*^ The code for the paper: Jin, W., Coley, C., Barzilay, R. and Jaakkola, T., 2017. Predicting organic reaction outcomes with Weisfeiler-Lehman network. In Advances in Neural Information Processing Systems (pp. 2607-2616).
We have also provided a dockerfile.
This can be used to build a Docker image using eg docker build . -t molecule-chef
if you have the Molecular Transformer weights downloaded into this folder.
It can then be run by docker run -it molecule-chef.
An uploaded Docker image lives on Docker Hub here.
As mentioned in our paper we use the Molecular Transformer^†^ for reaction prediction. For this task we use the authors' official implementation for running this model. The weights that we use can be found on Figshare here.
shasum -a 256 molecular_transformer_weights.pt
## returns 93199b61da0a0f864e1d37a8a80a44f0ca9455645e291692e89d5405e786b450 molecular_transformer_weights.pt^†^ Schwaller, P., Laino, T., Gaudin, T., Bolgar, P., Hunter, C.A., Bekas, C. and Lee, A.A., 2019. Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction. ACS Central Science.
As shown below, Molecule Chef consists of an encoder and a decoder. The decoder samples a reactant bag, which can be
be fed into a reaction predictor (as discussed above we use the Molecular Transformer) to generate output molecules.
The code for the encoder and the decoder can be found in molecule_chef/model.
The various scripts in the scripts folder train the model and run experiments:
scripts/prepare_data/prepare_datasets.py: prepares the unzipped USPTO data and extracts reactant bags for training.scripts/train/train_molecule_chef_qed.py: is used for training Molecule Chef with the property predictor regressing from latent space to QED score.scripts/evaluate: this folder contains scripts to evaluate the learnt model. See the respective readmes on generation, local optimization and retrosynthesis for further details.