Pocket Detection part of "Drug The Whole Genome" project

Two methods for pocket detection are implemented in this repository:

Pocket Detection with Generative model Refinement

We first use Fpocket to detect pockets in size of 10A, while may not be accurate enough. Then we use our SBDD model to generate ligands in the 10A pocket and relax the protein-ligand complex. We use the relaxed ligand to better detect the pocket with size 6A.

• Fpocket code : ./SBDD_AUG/fpocket.py

• SBDD model code (with the code fork by MolCRAFT): ./SBDD_AUG/sbdd

• • to use the code, please refer to the MolCRAFT repository:
• • data for training out SBDD model and the trained checkpoint can be found in google drive: https://drive.google.com/file/d/1aUGnX_QV3MQKChosdjS0_zXPn7b0YMUP/view?usp=drive_link
• • to use the trained model, please download the checkpoint and put it in the ./SBDD_AUG/sbdd/checkpoints folder

• Relaxation code: ./SBDD_AUG/relax.py

• Extraction 6A pocket code with the relaxed ligand: ./SBDD_AUG/extract_pocket.py

Pocket Detection with template matching

In this method, we use the PDBBind database as the template database, and use TM-align to align the protein of human. If the TM-score and pocket matching IOU is higher than a threshold, we consider a new pocket is detected.

• TM-align code: ./template_matching/tmalign.py this code calculates the TM-score between two proteins and save the rotation matrix

• Pocket matching code: ./template_matching/pocket_match.py this code filters the pocket with the rotation matrix and calculate the IOU

• clash_detection code: ./template_matching/clash_detection.ipynb this code remove the cases with clashes