QBMG: Quasi-Biogenic Molecule Generator with Deep Recurrent Network

Introduction

This is a PyTorch implementation of the paper: QBMG: quasi-biogenic molecule generator with deep recurrent neural network The user can (i) obtain a certain number of unbiased biogenic-like molecules and (ii) obtain a certain number of focused chemotype derivatives with transfer learning.We thank the previous work by the Olivecrona team.The code in this repository is inspired on REINVENT.

Usage

• sample.py: script that is used to generate a certain number of molecules (both unbiased or certain chemotype biogenic-like molecules) with trained model.

  Example:sampling 1000 molecules from unbiased model and focused chemotype model respectively

   %run sample.py ./data/biogenic.ckpt 1000

   %run sample.py ./data/coumarin.ckpt 1000

• transfer_learning.py: script that is used to train focused chemotype biogenic moleucles and obtain focused chemotype derivatives. The user can provide own focused chemotype molecules to generate new derivatives.

  Example:training a focused chemotype library

   %run transfer_learning.py ./data/coumarin.smi 

Requirments

This package requires:

• Python 3.6
• PyTorch 0.4.0
• RDKit
• tqdm
• jupyter notebook

Contact

Welcome to contact us. http://www.rcdd.org.cn/home/

Use QBMG to generate fragment-like molecules

ZINC Fragment database was used to train prior model so that it can be used to generate fragment-like molecules.

Welcome to contact me: gkxiao(at)molcalx.com, replace the (at) with @.

More detailed information in Chinese:

http://blog.molcalx.com.cn/2019/02/24/deep-learning-fragment-generator.html