Plixer is a generative model for drug-like molecules that bind in protein pockets. Provide a PDB file and a pocket location and the model generates SMILES strings for small molecule binders.

Paper accepted to GenBio @ ICML 2025

Read the preprint here

Installation

1. Clone the repository:

   git clone git@github.com:JudeWells/plixer.git
   cd plixer

2. Install dependencies: It is recommended to use a virtual environment. Plixer was developed and tested using python 3.11

   python3.11 -m venv venvPlixer
   source venvPlixer/bin/activate
   pip install -r requirements.txt

   or

   pip install -r requirements_no_gpu.txt
   

3. Download Model Checkpoints: The necessary model checkpoints are hosted on Hugging Face Hub. Run the following script to download them into the checkpoints/ directory.

   python download_checkpoints.py

4. Check instalation by running

python inference/generate_smiles_from_pdb.py

Usage

The main use-case for plixer is generating SMILES strings from a PDB file which contains your target protein inference/generate_smiles_from_pdb.py

Secondary use cases include: scoring / ranking molecules from pre-defined list inference/rank_smiles.py

generating structural analogs of a 3D ligand inference/sample_smiles_from_3d_ligand.py

These secondary use cases are not extensively tested, and it is likely that traditional chemoinformatics methods will perform better at these tasks.

Generating novel molecules for a PDB

To see the configurable options for inference run:

python inference/generate_smiles_from_pdb.py --help

Training poc2mol

The original model was trained using bfloat16. If you want to replicate this modify this file in your dock2grid install (optional): venvPlixer/lib/python3.11/site-packages/docktgrid/config.py

DTYPE = torch.bfloat16

python src/train.py experiment=example_train_poc2mol_plinder

Training vox2smiles on zinc molecules

python src/train.py experiment=example_train_vox2smiles_zinc

Fine-tuning Vox2Smiles on Poc2Mol Outputs AND zinc

python src/train.py experiment=train_vox2smiles_combined

This repository contains three models:

1. Poc2Mol: Generates voxelized ligands from protein voxel inputs
2. Vox2Smiles: Decodes voxelized ligands into SMILES strings
3. CombinedProtein2Smiles combines 1 & 2 for end-to-end pipeline

Project Structure

├── configs/                  # Configuration files
│   ├── data/                 # Data configuration
│   ├── model/                # Model configuration
│   └── experiments/          # Launch training runs with this
├── src/                      # Source code
│   ├── data/                 # Data processing modules
│   │   ├── common/           # Common data utilities
│   │   │   ├── tokenizers/   # SMILES tokenizers
│   │   │   └── voxelization/ # Unified voxelization code
│   │   ├── poc2mol/          # Poc2Mol data modules
│   │   ├── vox2smiles/       # Vox2Smiles data modules
│   │   └── poc2smiles/       # Combined data modules
│   ├── models/               # Model definitions
│   └── utils/                # Utility functions
└── scripts/                  # Random scripts

Model Architecture

Poc2Mol

The Poc2Mol model is a 3D U-Net that takes protein voxels as input and generates ligand voxels as output.

Vox2Smiles

The Vox2Smiles model is a Vision Transformer (ViT) that takes ligand voxels as input and generates SMILES strings as output. It uses a transformer encoder to process the voxel patches and a GPT-style decoder to generate the SMILES tokens.

Combined Model

The combined CombinedProtein2Smiles model connects the Poc2Mol and Vox2Smiles models in an end-to-end fashion. It takes protein voxels as input, generates ligand voxels using Poc2Mol, and then generates SMILES strings using Vox2Smiles.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgements

• The voxelization code is based on DockTGrid.