GNN_PersLay

Contributors

• Jason Apostol (CS): Worked on training and implementing multi-graph representation for protein structures.
• Anna Su (CBB): Worked on persistent homology computation and persistence layer generation.
• Vasilije Pantelic (MB&B): Worked on the molecular geometry computations.

Project Workflow

1. Construct Data Files:

   • Run the run_scripts.sh bash script to generate files in the data_files directory.
   • The script implements Python scripts in calculate_features to create the following files for each subdirectory:
     1. Reconstructed PDB file (id_rec.pdb)
     2. Backbone structure file (id_rec_bb.txt)
     3. Bond angle file (id_ba.txt)
     4. Bond length file (id_bl.txt)
     5. Dihedral angle file (id_da.txt)

   Directory Structure:

   data_files
   └── 1by7
       ├── 1by7_ba.txt
       ├── 1by7_bl.txt
       ├── 1by7_da.txt
       └── 1by7_rec.txt
   

2. Construct PH Files:

   • Run the run_ph.sh script to generate the id_rec_bb.txt files from Step 1.
   • This script utilizes ph_functions.py to compute persistence diagrams, vectors, and landscapes for each backbone structure. The results are stored in the ph_files directory.

3. Run geo_prep.py:

   • Execute geo_prep.py to create an index for your data and place it in the datasets_for_geo directory.

4. Clone PaddleHelix

   • Clone the PaddleHelix Repository into the root of the current directory.
   • Replace GEM in PaddleHelix/apps/pretrained_compound/ChemRL with the GEM folder in this repository.

5. Install Dependencies:

   • Ensure all dependencies listed in requirements.txt are installed.

6. Run knottrain.sh:

   • Locate the script in the forked PaddleHelix repository located at PaddleHelix/apps/pretrained_compound/ChemRL/GEM/scripts.
   • Modify the data_path in the script to point to the JSON files generated by geo_prep.py.
   • Execute via sh scripts/knottrain.sh. This will both preprocess the data and train the models described in the paper.