daiki-ko / Metric_MHG-VAE

Embedding of Molecular Structure Using Molecular Hypergraph Variational Autoencoder with Metric Learning

Core code for the paper "Embedding of Molecular Structure Using Molecular Hypergraph Variational Autoencoder with Metric Learning" (https://doi.org/10.1002/minf.202000203) by Daiki Koge, Naoaki ONO, Ming Huang, Md. Altaf-Ul-Amin, Shigehiko Kanaya.

Requirements

PyTorch
We have updated the code such that it is using version 0.4.1.

RDKit
version 2017.09.1.

Python
version 3.6.6.

Jupyter
version 1.0.0.

Running the Code

Prepare hypergraphs and descriptors

mol_smooth_embedding/extract_mhg.py

mol_smooth_embedding/Prepare_RDKit_descriptors.ipynb

Training mode

mol_smooth_embedding/Metric_MHG-VAE_with_DRL.ipynb

Evaluation of the models

using qm9 physical properties
mol_smooth_embedding/Evaluate_Model.ipynb

using rdkit descriptors
mol_smooth_embedding/Evaluate_Model_RDkit_desc.ipynb

About Scripts

As mentioned in our paper, the VAE architecture uses the same model as kajino's MHG-VAE (http://proceedings.mlr.press/v97/kajino19a/kajino19a.pdf). The code for the MHG-VAE can be found in mol_smooth_embedding/mhg. And kajino's original code can be found in https://github.com/ibm-research-tokyo/graph_grammar.