GAFSE

Reproducing of the paper entitled "An open unified deep graph learning framework for discovering drug leads"

• All rights reserved by Yueming Yin, Email: [email protected] (or [email protected]).

Environment

• Python 3.6.13
• Pytorch 1.10.2 (Up to your CUDA version)
• RDKit 2020.09.1.0
• Jupyter
• Tensorboard
• Tensorboardx
• Seaborn
• Matplotlib
• Scikit-learn

Installation

For a quick installation, please enter the following command in your terminal:

cd GAFSE
conda create --name GAFSE --file ./requirements.txt
conda activate GAFSE
jupyter notebook

Reproduction on GAFSE-HS (Hit screening)

In your jupyter notebook, re-run the corresponding notebook of "GAFSE-HS/GAFSE-HS/1_GAFSE_{Task ID}.ipynb" to reproduce the training process of GAFSE-HS. For example:

./GAFSE-HS/GAFSE-HS/1_GAFSE_1.ipynb

To reproduce the training process of GAFSE-MO on these HS tasks, please re-run the corresponding notebook of "GAFSE-HS/GAFSE-MO/1_GAFSE_{Task ID}.ipynb". For example:

./GAFSE-HS/GAFSE-MO/3_GAFSE_1.ipynb

Reproduction on GAFSE-MP (Molecular property prediction)

In your jupyter notebook, re-run the corresponding notebook of "GAFSE-MP/3C_GAFSE_Multi_Tasks_{Small, Medium, Big, Large}.ipynb" to reproduce the training process of GAFSE-MP. For example:

./GAFSE-MP/3C_GAFSE_Multi_Tasks_Small.ipynb

To readout the GAFSE-MP performance over all tasks, please re-run the notebook of "GAFSE-MP/3C_Performance_Readout.ipynb":

./GAFSE-MP/3C_Performance_Readout.ipynb

Reproduction on GAFSE-MO (Molecule optimization)

Reproduce GAFSE-MO on Molecular Activity

In your jupyter notebook, re-run the corresponding notebook of "GAFSE-MO/G_AFSE_{Task name}1/G_AFSE{Task name}_1.ipynb" to reproduce the training process of GAFSE-MO on molecular activity. For example:

./GAFSE-MO/G_AFSE_IC50_O43614_1/G_AFSE_IC50_O43614_1.ipynb

To test trained models, please re-run the corresponding notebook of "GAFSE-MO/G_AFSE_{Task name}1/G_AFSE{Task name}_1-Test.ipynb". For example:

./GAFSE-MO/G_AFSE_IC50_O43614_1/G_AFSE_IC50_O43614_1-Test.ipynb

Reproduce GAFSE-MO on Molecular Property

In your jupyter notebook, re-run the corresponding notebook of "GAFSE-MO/G_ADMET_M,T_C_{Task name}/G_ADMET_M,T_C_{Task name}.ipynb" to reproduce the training process of GAFSE-MO on molecular property. For example:

./GAFSE-MO/G_ADMET_M_C_CYP1A2_inhibitor/G_ADMET_M_C_CYP1A2_inhibitor.ipynb

To test trained models, please re-run the corresponding notebook of "GAFSE-MO/G_ADMET_M,T_C_{Task name}/G_ADMET_M,T_C_{Task name}-Test.ipynb". For example:

./GAFSE-MO/G_ADMET_M_C_CYP1A2_inhibitor/G_ADMET_M_C_CYP1A2_inhibitor-Test.ipynb

Reproduce GAFSE-MO on COVID-19

In your jupyter notebook, re-run the notebook of "GAFSE-MO/AID_1706/G_AFSE_AID_1706.ipynb" to reproduce the training process of GAFSE-MO on the COVID-19-related database:

./GAFSE-MO/AID_1706/G_AFSE_AID_1706.ipynb

For a quick view of the results, please check "GAFSE-MO/AID_1706/AID_1706_datatable_copy_generated_molecules.csv"

Tips

• Under the "Data" and "Module" folders are organized data and modules for review. For the convenience of operation, these data and modules are distributed to the "AttentiveFP" and "data" folders in subfolders.
• Some code may need to create an empty "result" folder in the current directory to store the run results.
• The "GRN" and "AFSE" classes imported "from AttentiveFP.AttentiveLayers_Sim_copy" are the most critical modules in this project.
• The "perturb_feature" and "modify_atoms" functions are the most critical functions in this project.
• The "AFSE" imported "from AttentiveFP.AttentiveLayers_Sim_copy" of GAFSE-MP is different from GAFSE-HS.
• The number of epochs for stopping "patience=30" is more suitable for GAFSE-MO, because too much training is not conducive to the optimization of molecules.
• If you encounter any problems reproducing our experiments, please contact me via [email protected] (or [email protected]).