Supporting code to run inference experiments from the publication "Cell morphology-guided de novo hit design by conditioning GANs on phenotypic image features" by Paula A. Marin-Zapata and Oscar Méndez-Lucio et al. (2022) https://doi.org/10.1039/D2DD00081D
The CPMolGAN model requires Python 3.6, TensorFlow 1.15, selfies 1.0.4, and rdkit 2018.09.1. We provide an environment file for connivent installation using conda (environment.yml). For reproducibility, we also provide a more detailed environment including all dependencies of our development setup (environment_detailed.yml). Follow the instructions below to install cpmolgan:
Clone repository:
git clone https://github.com/Bayer-Group/CPMolGAN.git
cd CPMolGAN
Download pre-trained weights. Model weights are available for non-commercial use only. (The model weights are fitted on publicly available data only)
Download them directly from Google Drive and place them inside the directory cpmolgan/model_weights, or by executing the bash script:
chmod +x download_model_weights.sh
./download_model_weights.sh
Create conda environment:
conda env create -f environment.yml
Activate your environment:
conda activate cpmolgan
Install tensorflow without GPU support:
conda install tensorflow=1.15.0
or with GPU support:
conda install tensorflow-gpu=1.15.0
Install this project as python package (in editable mode -e if desired):
pip install -e .
Finally, install the environment kernel to run jupyter notebooks:
python -m ipykernel install --user --name cpmolgan
Test the inference model by running the Python script in the test folder with your activated environment. A successful test will print one confirmation message for the autoencoder and one for the GAN.
cd test
python test_inference_model.py
The molecular autoencoder is tested by computing molecular embeddings for the set of smiles in test/autoencoder_input_smiles.csv and comparing against a reference embeddings file (test/autoencoder_output_latents.npy). The GAN is tested by generating molecules conditioning of a set of reference morphological profiles (test/GAN_input_condition_profiles.csv) and comparing against reference generated molecules and classification scores in test/GAN_output_smiles.csv
This project only utilized publicly available data (as listed below) to train the machine learning algorithm.
Data can be downloaded by executing the bash script:
chmod +x download_data.sh
./download_data.sh
Alternatively, data is available for download from Google Drive. Make sure to unzip and place all files inside the data directory in order to run the scripts in the publication_notebooks directory.
The following files are provided:
ChEMBL_standardized_smiles.csv: Training set for the molecular autoencoder with standardized SMILEStrain_set_30kcpds_normalized_profiles.csv.gz: Training set for the GAN with normalized well profiles for ~30,000 compounds and standardized SMILES. Downloaded from gigadb.org.test_set_overexpression_normalized_profiles.csv: Well profiles from gene over-expression perturbations used as test set. Illumination corrected images were downloaded from idr.openmicroscopy using a web client. Profiles were extracted with the analysis.cppipe CellProfiler pipeline provided in [1].ExCAPE_ligands_agonist_noTrainSet.csv: Agonists towards the genes in the over-expression test set which were found in the ExCAPE database and are not part of the training set.sure_chembl_alerts.txt: sure ChEMBL alerts used to filter generated molecules with desired physicochemical properties
The examples directory provides notebooks exemplifying the usage of the pre-trained inference model:
examples/run_molecular_autoencoder_inference.ipynb: shows how to use the autoencoder model to compute molecular representations from SMILES. It also summarizes preprocessing steps applied to SMILES strings before being feed to the model.examples/run_GAN_inference.ipynb: shows how to use the GAN model to generate SMILES conditioning on morphological profiles. It also summarizes preprocessing steps applied to morphological profiles before being feed to the model.
Model parameters are available for non-commercial use only, under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
Code is available under BSD 3-Clause License.
For using the publicly available data as described in #Data, we refer to the original licenses of the data owners.
ChEMBL_standardized_smiles.csv: The ChEMBL data is made available on a Creative Commons Attribution-Share Alike 3.0 Unported License.train_set_30kcpds_normalized_profiles.csv.gz: Datasets associated with GigaScience journal articles are specifically released under a Creative Commons CC0 waiver unless stated otherwise.test_set_overexpression_normalized_profiles.csv: CC BY 4.0 https://creativecommons.org/licenses/by/4.0/.ExCAPE_ligands_agonist_noTrainSet.csv: Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/)sure_chembl_alerts.txt: See https://www.surechembl.org/terms/
The directory publication_notebooks provides a comprehensive set of notebooks to reproduce most publication results and figures.
Cell morphology-guided de novo hit design by conditioning GANs on phenotypic image features
P. A. Marin Zapata, O. Méndez-Lucio, T. Le, C. J. Beese, J. Wichard, D. Rouquié and D. Clevert,
Digital Discovery, 2023, Advance Article , DOI: 10.1039/D2DD00081D