(Reimplementation) Amortized Tree Generation for Bottom-up Synthesis Planning and Synthesizable Molecular Design

https://openreview.net/forum?id=FRxhHdnxt1

setup (conda environment)

# init conda, such as by activating .bashrc
source ~/.bashrc

conda create -n tree python=3.6 pip
conda activate tree

# rdkit can only be installed via conda, not pip
conda install -c rdkit rdkit -y

/path/to/conda/env/tree/bin/pip install -r requirements.txt

how to run

data preparation

Please run all the scripts in tree_gen_net/bash_scripts/data_prep/ following the numerical order from 1 to 10. The necessary input data has been included, obtained by following the paper. Rough timing benchmarks have also been included. Some routines run much faster on multiprocessing.

training

Please run the 4 scripts in tree_gen_net/bash_scripts/train/, each named after the model they train. Alternatively, you can just run bash tree_gen_net/bash_scripts/train/train_all.sh.

inference

For synthesis planning, please run bash tree_gen_net/bash_scripts/test/decode_trees.sh, which decodes synthetic trees greedily (k=1) conditional on product SMILES in the test set trees. Afterwards, please run bash tree_gen_net/bash_scripts/test/calc_simil.sh to calculate the pairwise Tanimoto similarity of the decoded SMILES versus those product SMILES in the test set trees.

For property optimization via genetic algorithm, please run bash tree_gen_net/bash_scripts/test/genetic_algorithm.sh

Of course, the arguments in all of these scripts can be modified and experimented with. None of the code is hardcoded and should be quite flexible. Currently, I have used either the settings from the paper, or otherwise scaled down some hyperparameters just to get working prototype out as fast as possible, given I have only spent 2 weekends (aka roughly 3-4 man days) on this project.

citation

all credits goes to the authors

Gao, Wenhao, Rocío Mercado, and Connor W. Coley. "Amortized Tree Generation for Bottom-up Synthesis Planning and Synthesizable Molecular Design." arXiv preprint arXiv:2110.06389 (2021).

potential next steps

• future experiments
  • experimental validation
  • filtering training trees by property of interest (property-biased neural networks)
  • analogue generation on best molecules from genetic algorithm, hope the analogues also have high property of interest
• unit tests