ScafVAE

ScafVAE, a scaffold-aware VAE framework for multi-objective molecular design.

Table of contents

• Installation
• Usage
  • Generating moleucules with properties of inerest
  • Train your own model
• License
• Citation

Installation

Required dependences

ScafVAE is implemented in PyTorch. All basic dependencies are listed in requirements.txt and most of them can be easily installed with pip install. We provide tested installation commands in install_cmd.txt, for your reference.

ScafVAE

pip install -e .

Model parameters

Please download the parameters from here (or here), and put it in the ScafVAE/params/.

Usage

Generating moleucules with properties of inerest

You can use ScafVAE to generate multi-objective molecules by following the notebook ScafVAE/demo/demo_molecular_design.ipynb.

The notebook contains:

1. Collecting molecules with their properties.
2. Training surrogate models.
3. Generating multi-objective molecules.

If you want to quickly test the molecular generation, you can simply run the following codes here (or run the ScafVAE/demo/demo_molecular_design.py) to generate the dual-target molecules.

from ScafVAE.app.app_utils import *

df_EGFR, df_HER2 = get_demo_properties(max_num=1000, name='binding')  # get molecules with their properties

generation_path = './output'  # path for saving data
tasks = [
  {
    # task name
    'name': 'EGFR_inhibition',

    # input data
    'data': df_EGFR,

    # task type: classification / regression
    'task_type': 'classification',

    # ML model for surrogate model
    'ML_model': 'RF',

    # property is need to be maximized or minimized: max / min
    'optimization_type': 'max',

    # pseudo weight for this property
    'pseudo_weight': 0.5,
  },
  {
    'name': 'HER2_inhibition',
    'data': df_HER2,
    'task_type': 'classification',
    'ML_model': 'RF',
    'optimization_type': 'max',
    'pseudo_weight': 0.5,
  },
]

base_model, surrogate_model = prepare_data_and_train(generation_path, tasks)
df_output = generate_mol(
  10,  # number of generated molecules
  generation_path, tasks, base_model, surrogate_model,
)
print(df_output)  # a dataframe contrains generated molecules with their properties

Train your own model

We provide a pipeline for training the ScafVAE based on user's data in ScafVAE/demo/demo_molecular_pretraining.ipynb.

License

Released under the MIT license.

Citation

@Article{D4SC08736D,
author ="Dong, Tiejun and You, Linlin and Chen, Calvin Yu-Chian",
title  ="Multi-objective drug design with a scaffold-aware variational autoencoder",
journal  ="Chem. Sci.",
year  ="2025",
pages  ="-",
publisher  ="The Royal Society of Chemistry",
doi  ="10.1039/D4SC08736D",}