Rxitect was created for the purpose of experimenting with the development and implementation of retrosynthesis engines within molecular generators for the goal of de novo drug design; the focus of my Master's thesis. The code in this repository is based on DrugEx, released by Xuhan Liu (First Author) and Gerard J.P. van Westen (Correspondent Author) on March 8th, 2021. The same license terms apply for this repository, and can be found in the LICENSE file.

After cloning this repository, make sure you have a conda distribution installed. We recommend miniforge for licensing reasons, but anaconda/miniconda will work as well.

• conda env create -f environment.yml
• conda activate rxitect
• pip install -e .

Mamba is a drop-in replacement for conda, it's much faster at resolving environments and is recommended. You can install it by running conda install -c conda-forge mamba. In case you want to use this from the start, replace conda with mamba in the instructions above.

This project uses DVC to version control large data files and trained models on DagsHub. To use DVC, run the following commands:

• conda install -c conda-forge mamba
• mamba install -c conda-forge dvc

• black requires specific versions of typing-extensions, so you may need to run pip install typing-extensions --upgrade first.
• If you are developing on windows, we recommend to download pytorch manually to fit your needs, e.g.: conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch.
• If you run into issues after importing torch, try running conda install -c defaults intel-openmp -f.
• If you are developing on a mac, you may run into issues with xgboost. To fix this, you need to have cmake installed, which can be done by running the following commands (assuming you have brew installed): brew install gcc@11, followed by brew install cmake. Note that because RA Score has a hard dependency on tensorflow-gpu to run their pretrained models, development on a mac is currently limited to just the base functionality of Rxitect (unless you have a CUDA-compatible GPU).
• there's a recent (as of February 7, 2022) bug in rdkit where you could not import it or its modules properly if you have boost-cpp=1.74.0=h359cf19_6. If you run into this issue, update it by conda/mamba installing boost-cpp again.

The paper that accompanies the original DrugEx code can be found here.