ACoVAE

Attention-Based Conditional Variational Autoencoder.

Project overview

A new attention-based conditional variational autoencoder neural network architecture based on recent developments in attention-based methods.

Prerequisites

• Operating system : Linux with CUDA support
• Hardware : GPU compatible with tensorflow
• Python environment manager : anaconda, miniconda...

Project structure

.
└── ACoVAE/
    ├── doc/
    │   └── Pharms_transformer.pdf # A visual representation of the ACoVAE neural network 
    ├── utils/
    │   ├── __init__.py # Imports
    │   ├── data_preparation_umaps.sh # data preparation - only for GTM universal maps. 
    │   ├── layers.py # Network implementation
    │   └── utils.py # Utilities (SMILESParser class)
    ├── training_data/
    │   └── README.md
    ├── LICENSE # GNU General Public License v3.0 license
    ├── README.md
    ├── main.py
    ├── model_parameters_standard.yaml # Standard parameters for model training
    └── requirements.txt # Requirements to create the python environment

Installation

Clone the repository:

git clone https://github.com/Laboratoire-de-Chemoinformatique/ACoVAE.git
cd ACoVAE
mkdir model

Install dependencies

conda create --name ACoVAE --file requirements.txt
pip install adabelief-tf CGRtools

(Optional) Check installation

conda list | grep tensorflow
conda list | grep cudatoolkit
conda list | grep cudnn
python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

Usage

Input file

Input file description

The input file is a descriptor matrix file under the .svm format.

It can be generated using the ISIDA/Fragmentor software (Open access, request the software using the form : https://infochim.u-strasbg.fr/-DOWNLOADS-SOFTWARE-.html).

Please refer to the ISIDA/Fragmentor documentation to generate the descriptor matrix.

Input file format

• Column 1 : SMILES
• Column 2-end : Descriptor matrix in a .svm file following the libSVM format. These columns consist in a pair of values separated by a ":". The first value identifies the fragment's index in the header file (.hdr file created by ISIDA/Fragmentor), the second value is the fragment count.

Example dataset

An example dataset can be downloaded here : https://entrepot.recherche.data.gouv.fr/dataset.xhtml?persistentId=doi:10.57745/ILWSLF

Commands

Model training

conda activate ACoVAE
python3 main.py -i ./training_data/chembl23_umap1.svm -m ./model/model_name -mp model_parameters_standard.yaml --log log_19-07-2022

For extended functions, consult the help command:

python main.py --help

Results and selected model

Check the log file for model quality scores : epoch, loss, mask_acc, rec_rate, val_loss, val_mask_acc, val_rec_rate

val_mask_acc (accuracy - character-specific reconstruction rate), val_rec_rate (reconstruction rate) are the values to follow.

Select the model which suits your needs in the /model folder.

Generate compounds

First, generate the descriptor vector for a known compound, using ISIDA/Fragmentor. Note : The first column in the output file must be the ID of the compound, not the SMILES. (see ./training_data/1.svm for an example).

Then, use the generated vector as seeds for new compounds generation.

mkdir sampled_smi
python main.py -f ./training_data/1.svm -n 1000 -m ./model/model_name_99_0.98 -sp ./model/model_name_smi_parser.pkl -mp model_parameters_standard.yaml -o ./sampled_smi/known_compound_vector_sampled.smi

• sp : SMILES parser pickle object created during the network training. It is needed for sampling.
• m : the model created during the network training.
• mp : the model yaml created during the network training.
• n : Number of batches sampled per query.

Credits

• Arkadii Lin, Daniyar Mazitov, William Bort, Timur Madzhidov and Alexandre Varnek
• Kazan Federal University, Russia
• University of Strasbourg, France

License

Distributed under the GNU GENERAL PUBLIC LICENSE Version 3. See the LICENSE link in the additional resources for more information.

Additional Resources

• GitHub
• License
• Publication