tonydeep / molecule_generator

Code for "Multi-Objective De Novo Drug Design with Conditional Graph Generative Model" (https://arxiv.org/abs/1801.07299)

Conditional Molecule Generator

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).

Requirement

The model is built using Python 2.7, and utilizes the following packages:

• Tensorflow 1.3
• RDKit
• Numpy
• networkx 2.0

Todo list

The current repo contains only data sets and the source code, future updates will include:

• Models trained using the full dataset
• Add file descriptions and tutorials
• Activity data and trained predictor for GSK3b and JNK3

Usage

Run the train.py file for the network training. Set the environment variable TF_CPU_ALLOCATOR_USE_BFC

to true to avoid memory leak in CPU.

To train on a single GPU, use the single command:

python train.py single [config_dir]

Where [config_dir] is the directory containing model config files. Default configs are provided in the models folder (description of those files will be contained in the future release). For training with multiple GPUs, use multiple command:

# start parameter server
CUDA_VISIBLE_DEVICES="" nohup python train.py multiple [config_dir] ps 0

# start generator
CUDA_VISIBLE_DEVICES="" nohup python train.py multiple [config_dir] generator 0
CUDA_VISIBLE_DEVICES="" nohup python train.py multiple [config_dir] generator 1

# start worker
CUDA_VISIBLE_DEVICES=0 nohup python train.py multiple [config_dir] worker 0
CUDA_VISIBLE_DEVICES=1 nohup python train.py multiple [config_dir] worker 2
...
CUDA_VISIBLE_DEVICES=[n-1] nohup python train.py multiple [config_dir] worker [n-1]

Where n is the number of GPUs used during training. The default value for n is 4. To change n, you need to modify the corresponding config file.

To train with the entire dataset, use single-full and multiple-full respectively for single and multiple GPU.

It should be reminded that the default config requires a large amount of memory for training. Limit the batch size or k in the config file if the memory resource is limited.

Contact me if you have any questions. Email: [email protected] or [email protected]