ICVAE

ICVAE: Interpretable conditional variational autoencoder for de novo molecular design

Environment

• python = 3.6.13
• pytorch = 1.10.1
• RDKit
• numpy
• matplotlib
• jupyter notebook

How to run？

1. download dataset and save in your local path

Visit https://github.com/aksub99/molecular-vae/tree/master/ , and download the processed ZINC dataset from "data" fold. Then you can save it in your local path which contains the ICVAE code.

You should also download the normalized molecular property labels and origin property labels from the "prop_np" fold.

The "prop_np" file contains seven subfolds: molecular weight (weight), logP, sas, tpsa, qed, hba, hbd. In each fold, the y_train_norm.npy and y_test_norm.npy files is the normalized molecular property value ranging from 0 to 500. We set the big normalized range to make the latent value of each property has some distance, which make the sampling process more easily to generate smiles.

2. set up the pytorch environment

We recommend you to install anaconda and create a new environment by using the following command:

conda create --name icvae python=3.6

and then you can enter the enviroment by (windows) :

conda activate icvae

or (ubuntu) :

source activate icvae

and next, you need to install the package by:

conda install numpy, matplotlib, jupyter notebook, rdkit -c rdkit

Please note your must install the gpu support for pytorch. The detail can be found in this blog: https://medium.com/analytics-vidhya/4-steps-to-install-anaconda-and-pytorch-onwindows-10-5c9cb0c80dfe .

3. launch the jupyter notebook and run the code

launch the jupyter notebook by:

jupyter notebook

For training each molecular property, you can train ICVAE model by running train.ipynb.

4. generate the following latent image

you can just run the plot_latent.ipynb to obtain the molecular latent image.

5. sample the molecular with given the condition latent input

you can run the sampling.ipynb to generate the molecule with target property.

Issues

Please report all installation / usage issues by opening an issue at this repo.

References

Portions of the code have been re-used from the following repositories:

• topazape/molecular-VAE
• jaechanglim/CVAE