augmented-GCN

This repository provides implementations introduced in "Deeply learning molecular structure property relationships using attention- and gate-augmented graph convolutional network".

Usage

We used several scripts and a 'Harvard Clean Energy Project (CEP)' dataset in https://github.com/HIPS/neural-fingerprint.

1. First, make results and save directories to save output files and save files, respectively.

mkdir results mkdir save

2. Convert smiles files to graph inputs at a database folder.

cd database python smilesToGraph.py ZINC 10000 1 python smilesToGraph.py CEP 1000 1

3. Also, enter below command to obtain logP, TPSA, QED and SAS.

python calcProperty.py

4. Training

python train.py model property #layers #epoch initial_learning_rate decay_rate

python train.py GCN logP 3 100 0.001 0.95

models : GCN, GCN+a, GCN+g, GCN+a+g, GGNN

property : logP, TPSA, QED, SAS (ZINC dataset) and pve (CEP dataset)

sparkingarthur / augmented-GCN