HaotianZhangAI4Science/Causal-Delete

🌈Structure-based Deep Molecular Generation Paradigm Enables Efficient Design of Potent CLIP1-LTK Inhibitors

Environment

First Approach

Install via conda .yml file (cuda 11.3)

conda install mamba
mamba env create -f delete_environment.yml -n delete
conda activate delete

Of note, the mamba could be replaced by conda.

Second Approach

We also offer a Conda package for easy installation, which is available for download from Zenodo. After downloading, simply unzip the package in your conda environments directory. In my case, the directory is located at ~/.conda/envs. We extend our sincere gratitude to the team at Zenodo, who provide a valuable free platform for storing large files for scholarly purposes.

mkdir ~/.conda/envs/surfgen
tar -xzvf delete.tar.gz -C ~/.conda/envs/delete
conda activate delete

Please remember to replace "my case" with your actual username or replace it depending on your requirements.

Docking Environment (Optional)

conda create -n docking rdkit ipykernel jupyter biopython python=3.8 easydict -c rdkit -c conda-forge
# Then download mgltools locally, do not use the conda! 

Embed Causal-Delete in the Real-world Applications

Causal-Delete is based on the deconstruction-reconstruction approach, which means that you can delete the undesired part of molecules for the subsequent model suggestions. This strategy has successfully resulted in several successful drug design campaigns in history. Now, this classical approach is powered by the structure-based deep learning method, allowing chemists to take multiple-choice instead of doing fill-in-the-blank questions.

To illustrate, you can load the complex structures, obtained from the Protein Data Bank or from the docking results, in the pymol, as illustrated in the first piece of the following figure. Then delete partial structures. In our example, we want to modify the fragment that extends into the protein pocket, thus we delete atoms as shown in the second piece, getting the hit fragments.

Then, suggest novel molecules with the following code. The pre-trained checkpoint can be downloaded [here](https://doi.org/10.5281/zenodo.7985201).

python -u design_mols.py --surf_path ./example/LTK_pocket_8.0.ply --frag_path ./example/partial_ltk.sdf --check_point ./ckpt/delete.pt --outdir ./outputs --suboutdir ./example

The examples of generated molecules are listed below.

Combining Structure-based and Ligand-based Paradigms

Causal-Delete is the integration of structure- and ligand-based method. In the LTK case, a novel target that has no reported specific ligands recently. To emulate the real-world scenario, where chemists can use the inspiration from a similar target to design its specific inhibitors, the causal-inference strategy is adopted to borrow the chemical space from the ALK target, which shares a high sequence similarity with LTK and contains more than 2000 verified active molecules. In your case, you can perform a similar approach to embed the prior knowledge into the structure-based approach. You can follow the subsequent workflow to build your own case.

Private data preparation

Prepare your own data in the ./causal_inference/SDF, with each .sdf file being a molecule of interest.

Label preparation (Optional)

Then get the label of each molecule. For example, predicted affinity, IC50/Kd, or the docking score.

# If you do not have the affinity label, docking is an alternative choice. 
python ./causal_inference/docking.py 

Then, each *.sdf file will be processed as the *_out.sdf file.

LMDB data preparation

Follow the guidance in the ./causal_inference_data.ipynb

Firstly, create the protein-ligand index files. Secondly, create the LMDB database with the index.pkl Thirdly, transform each data in the LMDB database to filter the invalid train sample, leaving them out of the name2id dictionary. For the train_val_split, the first twenty samples are treated as validation sets.

If you run them successfully, you will get the causal_inference_data.lmdb, causal_inference_data.lmdb-lock, causal_inference_data_name2id.pt, causal_inference_data_split_name.pt in the ./data

Transfer the chemical space to the structure-based model

# Of note, change the base model in the config file, and make sure the data path is correct. 
python causual_inference_transfer.py --config ./causual_inference.yml

Train Base model

You can train the base causal delete model from the scratch.

python delete_train.py --config ./configs/base_train.yml

The data used for training can be found at CrossDock, including lmdb, lmdb-lock, name2id, split_by_name.

Make surface data on your own

create the base Python environment

To generate the surface data, you need a conda environment and several installed software.

conda create -n surface pymesh2 jupyter scipy joblib biopython rdkit plyfile -c conda-forge
# or you can use the yml file to create the environment
# conda env create -f surf_maker_environment.yml

Install APBS Toolkits

When the base python environment was created, then install APBS-3.0.0, pdb2pqr-2.1.1 on your computer. Then set the msms_bin, apbs_bin, pdb2pqr_bin, and multivalue_bin path directly in your ./bashrc or just set them in the scripts when creating the surface file from the pdb file. (Like ./utils/masif/generate_prot_ply.py )

Try Generate Surface Now !

Having successfully set up all the necessary environments, you can now proceed to generate surface data. Please follow the instructions in ./data/surf_maker for this process. Alternatively, to test the successful configuration of your environment, you can execute the ./data/surf_maker/surf_maker_test.py script.

python ./data/surf_maker/surf_maker_test.py

If the surface is generated, you will find the .ply file in the ./data/surf_maker