DrugHIVE: Structure-based drug design with a deep hierarchical generative model

This repository is the official implementation of DrugHIVE, a deep hierarchical variational autoencoder developed for structure-based drug design. JCIM paper.

<div> <a href="https://github.com/jssweller/drughive#gh-light-mode-only"> <img width="100%" style="aspect-ratio: auto; max-width: 1000px" src="https://github.com/jssweller/DrugHIVE/blob/main/img/hive.png#gh-light-mode-only"/> </a> <a href="https://github.com/jssweller/drughive#gh-dark-mode-only"> <img width="100%" style="aspect-ratio: auto; max-width: 1000px" src="https://github.com/jssweller/DrugHIVE/blob/main/img/hive_dark.png#gh-dark-mode-only"/> </a> </div>

Installation

1. Dependencies

The code has been tested in the following environment: | Software | Version | |-------------------|-----------| | Python | 3.9.16 | | CUDA | 11.6 | | OpenBabel | 3.1.1 | | PyTorch | 1.12.1 | | PyTorch Lightning | 2.0.0 | | RDKit | 2021.09.5 |

Install via conda

Install dependencies using the listed requirements in requirements.txt:

conda create -n drughive -c conda-forge -c pytorch -c nvidia -c rdkit --file requirements.txt

2. Git clone the repository

git clone https://github.com/jssweller/DrugHIVE

Sampling

Pre-trained models

Pre-trained model weights can be downloaded from Zenodo:

wget -P model_checkpoints/ https://zenodo.org/records/12668687/files/drughive_model_ch9.ckpt

Ligand generation

To sample from DrugHIVE, first adjust the parameters in the generate.yml example configuration file. Then, run the following command:

python generate_molecules.py config/generate.yml

To sample from the prior, set zbetas: 1. in the configuration file.
To sample from the posterior, set zbetas: 0. in the configuration file.
To sample between the prior and posterior, set the values of zbetas between 0. and 1..

Substructure modification (scaffold hopping)

To generate molecules with substructure modification, first adjust the parameters in the generate_spatial.yml example configuration file. Then, run the following command:

python generate_molecules.py config/generate_spatial.yml

Ligand optimization

Install QuickVina 2

Before running the optimization process, the QuickVina 2 docking tool must be installed:

• download (or compile) the QuickVina2 docking tool from https://qvina.github.io
• place qvina2.1 in DrugHIVE/ or in a directory in listed in your PATH variable (e.g., /usr/bin/)

Run optimization

To optimize molecules with DrugHIVE, first adjust the parameters in the generate_optimize.yml example configuration file. Then, run the following command:

python generate_optimize.py config/generate_optimize.yml

Training

1. Download PDBbind dataset

Download and extract the PDBbind refined dataset from http://www.pdbbind.org.cn/

2. Download ZINC molecules

Download ZINC molecules from https://zinc20.docking.org/ in SDF or MOL2 format. Place them in a single directory.

3. Process the datasets for training

To process the PDBbind dataset, run:

python process_pdbbind_data.py <path/to/PDBbind/directory>

To process the ZINC dataset, run:

python process_zinc_data.py <path/to/ZINC/directory> -o data/zinc_data/zinc_data.h5 -ext <file_extension>

Here, <file_extension> can be one of sdf or mol2.

4. Run training

First, adjust the training parameters in the config/train.yml example configuration file. Make sure to set data_path_pdb and data_path_zinc to the locations of your datasets. Then, run the following command:

python train.py config/train.yml