HyFactor is an open-source architecture for structure generation using graph-based approaches. It is based on the new type of molecular graph - Hydrogen-count Labelled Graph (HLG). This graph, similar to the InChI linear notation, considers the number of hydrogens attached to the heavy atoms instead of the bond types. Additionally, with HyFactor we add ReFactor architecture, which is based molecular graph-based architecture with defactorization procedure adopted from the reported DEFactor architecture.

This repository includes official implementations of HyFactor, ReFactor and functions for translation molecular graph to HLG and back.

For more details please refer to the paper

If you are using this repository in your paper, please cite us as:

HyFactor: A Novel Open-Source, Graph-Based Architecture for Chemical Structure Generation
Tagir Akhmetshin, Arkadii Lin, Daniyar Mazitov, Yuliana Zabolotna, Evgenii Ziaikin, Timur Madzhidov, and Alexandre Varnek
Journal of Chemical Information and Modeling 2022 62 (15), 3524-3534
DOI: 10.1021/acs.jcim.2c00744

All materials used in the publication are availible on Figshare project page

The standardized data sets and training/validation splits:

1. ZINC 250K standardized data set
2. ChEMBL v.27 standardized data set
3. The MOSES data set was used as it is

The original data sets were taken from:

1. Original ZINC 250K data set
2. ChEMBL page
3. MOSES benchmarking GitHub repository

The weights of Autoencoders from the experiments are available on Figshare

First, download the repository on your machine. Then, create conda enviroment with the folowing code:

conda env create -f enviroment.yml

The enviroment file is made for use on GPU with CUDA version = 11.3. If you have different versions of drivers or want to use a CPU version, please modify this file before the installation. For additional support, please, visit the tensorflow documentation page.

When your enviroment is ready, activate it and execute command to install the architecture:

python3 setup.py install

In this case you should create enviroment folder anywhere you prefer, install here the enviroment and activate it:

mkdir hyfactor_env
python3 -m venv hyfactor_env/
source hyfactor_env/bin/activate

Then, similarly as with conda, you just run the folowing code:

python3 setup.py install

This tool works in two modes: command-line and as usual python package. In both ways you should specify config file which will be used for every task. The examples of config file you can find in the folder examples/configs.

Once you specified your config file, execute the AutoEncoder with folowing command:

hyfactor -cfg YOUR_CONFIG_FILE.yaml

Here you can simply import the HYFactor package in folowing way:

from HYFactor import task_preparer
import yaml

with open('YOUR_CONFIG_FILE.yaml', 'r') as file:
    config = yaml.load(file, Loader=yaml.SafeLoader)

task_preparer(config)

The code for conversion of the HLG to molecular graph is implemented in function HYFactor.nn_utils.reconstruction_hyfactor.

Here is a custom example of HLG conversion:

from HYFactor.nn_utils import recompile_rules
from HYFactor.mol_utils import reconstruct_molecule_hyfactor

atoms = [...]  # list of atoms indices according to tuple atom_types
hydrogens = [...]  # list of hydrogens attached to heavy atoms from 0 to 4
adjs = [[...]]  # list of lists or binary matrix of connectivity between atoms

recompiled_rules = recompile_rules(atom_types)
molecule = reconstruct_molecule_hyfactor(atoms, hydrogens, adjs, atom_types, recompiled_rules)

We welcome contributions, in the form of issues or pull requests.

If you have a question or want to report a bug, please submit an issue.

To contribute with code to the project, follow these steps:

1. Fork this repository.
2. Create a branch: git checkout -b <branch_name>.
3. Make your changes and commit them: git commit -m '<commit_message>'
4. Push to the remote branch: git push
5. Create the pull request.

• Tagir Akhmetshin tagirshin@gmail.com
• Arkadii Lin arkadiyl18@gmail.com
• Timur Madzhidov tmadzhidov@gmail.com
• Alexandre Varnek varnek@unistra.fr