Home

Awesome

RELATION: REceptor-LigAnd interacTION

A Deep Generative Model for Structure-based De Novo Drug Design

overview of the architecture of RELATION

Overview

This repository contains the source of RELATION, a software for DL-based de novo drug design.

Requirements

if utilizing GPU accelerated model training

Creat a new environment in conda

conda env create -f env.yml

Running RELATION

Prepare molecular dataset

To train the RELATION network, the source dataset and target dataset (akt1 and cdk2) must by converted to a 4D-tensor-(19,16,16,16), which means the 3D gird with 19 channels.

Source dataset

python model/data_prepare.py --input ./data/zinc/zinc.csv --output ./data/zinc/zinc.npz --mode 0

Target dataset

python model/data_prepare.py --input ./data/akt1 --output ./data/akt1/akt_pkis.npz --pkidir ./data/akt1.csv --mode 1

Training RELATION

Load sourch dataset (./data/zinc/zinc.npz) and target dataset (./data/akt1/akt_pkis.npzor ./data/cdk2/cdk2_pkis.npz).

python model/train.py --epoches 150 --steps 5000 --target akt1 --batchsize 256 --decive 0

Sampling

Load the ./akt1_relation.pth or ./cdk2_relation.pth generative model, and typing the following codes:

python sample.py --method 0 --numbers 500 --output ./output --target akt1

or you can also use the bayesian optimization in sampling process:

python sample.py --method 1 --numbers 500 --output ./output --target akt1

Or you can use our online RELATION-based server http://cadd.zju.edu.cn/relation/remode/