𝐷𝑟𝑢𝑔𝐺𝐸𝑁-𝐺𝐴𝑁

𝑂𝑣𝑒𝑟𝑣𝑖𝑒𝑤

The model integrates GAN with reinforcement learning. These two models are trained separately but are used jointly to generate novel targeted chemical libraries.

The model and data of this project are partially referenced by Drug-Discovery-using-GANs. I mainly modified the relevant code details and added new visualization functions

The repository is organised as follows:

_model contains:

• hyperparams.json Data input path and related parameter settings.
• disdata.py``gendata.py``model.py``TargetLSTM.py``text_CNN.py Implementation of GAN model
• rollout_policy.py Implementation of RL
• mol_metrics.py Implementation of the evaluation module

environment.yml Environment configuration file

run.py Script for running the main program

Visualization Module： SMIpro1.py File merging and normalization of generated molecules SMIpro2.py Check whether visualization is possible SMIshow.py 2D visualization and 3D structure file generation

𝐺𝑒𝑡𝑡𝑖𝑛𝑔 𝑆𝑡𝑎𝑟𝑡𝑒𝑑

𝑆𝑒𝑡𝑡𝑖𝑛𝑔 𝑢𝑝 𝑒𝑛𝑣𝑖𝑟𝑜𝑛𝑚𝑒𝑛𝑡

cd DrugGAN

conda env create -f environment.yml

conda activate py36

𝑆𝑡𝑎𝑟𝑡𝑖𝑛𝑔 𝑡ℎ𝑒 𝑡𝑟𝑎𝑖𝑛𝑖𝑛𝑔

Before you start the training, you can specify the TRAIN_FILE and OBJECTIVE in hyperparams.json yourself. This allows you to upload your own data and specify the potential molecular properties of the molecules you wish to generate.

cd DrugGAN

python run.py

𝑉𝑖𝑠𝑢𝑎𝑙𝑖𝑧𝑒

cd DrugGAN

python SMIpro1.py

python SMIpro2.py

python SMIshow.py

𝑅𝑒𝑠𝑢𝑙𝑡𝑠