MRGVAE

This repository contains our implementation of "A Deep Molecular Generative Model Based on Multi-Resolution Graph Variational Autoencoders"

Requirements

• Python(version >= 3.6)
• Pytorch(version >= 1.1.0)
• RDKit (version >= 2019.03)
• networkx (version >= 2.4 )
• numpy (version >= 1.18 )

We highly recommend you to use conda for package management.

Vocab Extraction

Fragments and interchangeable fragments are extracted by the following commond. (replace ChEMBL_Subset4Training.txt with your own molecular dataset)

unzip data/ChEMBL_Subset4Training.zip 
python vocab_extract.py --mol_data ChEMBL_Subset4Training.txt --vocab_file chembl_vocab_frag_rd2.txt --bond_file chembl_vocab_bond_rd2.txt --radius 2 --ncpu 8

Molecule Processing

Molecules are converted into three-level hierarchical graphs using the fragment vocabulary from the first step.

mkdir chembl_train
python mol_preprocess.py --mol_data ChEMBL_Subset4Training.txt --vocab chembl_vocab_frag_rd2.txt --bond chembl_vocab_bond_rd2.txt --mol_folder chembl_train/ --radius 2 --ncpu 8

Model Training

Train the model with KL regularization, use

mkdir chembl_model
python train.py --mol_folder chembl_train --vocab chembl_vocab_frag_rd2.txt --bond chembl_vocab_bond_rd2.txt --model_folder chembl_model/  --epoch 10

Molecule Generation

To generate molecules, use

python generator.py --vocab chembl_vocab_frag_rd2.txt --bond chembl_vocab_bond_rd2.txt --model_folder chembl_model/model.0 --fragment_num 5  --num_decode 100  --batch_size 20 --gen_file 'gen.txt'

To generate molecules with fixed scaffold, use

python generatorwithscaffold.py --vocab chembl_vocab_frag_rd2.txt --bond chembl_vocab_bond_rd2.txt --model_folder chembl_model/model.0 --fragment_num 5  --num_decode 100  --batch_size 20 --radius 2 --gen_file 'gen.txt' --scaffold '*[CH:1]1CCn2cncc21'