This repository contains the source code and data sets for the graph based molecule generator discussed in the article "Multi-Objective De Novo Drug Design with Conditional Graph Generative Model" (https://arxiv.org/abs/1801.07299).

Briefly speaking, we used conditional graph convolution to structure the generative model. The properties of output molecules can then be controlled using the conditional code.

This repo is built using Python 2.7, and utilizes the following packages:

• MXNet == 1.3.1
• RDKit == 2018.03.3
• Numpy
• Scikit-learn (for the predictive model)

To ease the installation process, please use the dockerfile environment defined in the Dockerfile.

• train.py: main training script.
• mx_mg: package for the molecule generative model:
  • data: packages for data processing workflows:
    • conditionals.py: callables used to generate the conditional codes for molecules
    • data_struct.py: defines atom types and bond types
    • dataloaders.py , datasets.py and samplers.py: data loading logics
    • utils.py: utility functions
  • models: library for graph generative models
    • modules.py: define modules (or blocks) such as graph convolution
    • networks.py: define networks (MolMP, MolRNN and CMolRNN)
    • functions.py: autograd.Function objects and operations
  • builders.py: utilities for building molecules using generative models
• rdkit_contrib: functions used to calculate QED and SAscore (for older version of rdkit)
• example.ipynb: tutorial

To train the model, first unpackdatasets.tar.gz (download here) to the current directory, and call:

./train.py {molmp|molrnn|scaffold|prop|kinase} path/to/output

Where {molmp|molrnn|scaffold|prop|kinase} are model types, and path/to/output is the directory where you want to save the model's checkpoint file and log files. The following call:

./train.py {molmp|molrnn|scaffold|prop|kinase} -h

gives help for each model type.

Please contact me. Email: [email protected] or [email protected]