robmacc / vaemols

Variational Autoencoder for Molecules

Variational Autoencoder for Molecules

Variational autoencoder for molecules in tensorflow.

Dependencies

1. Rdkit

conda install -c rdkit rdkit

2. Tensorflow

cpu-version

pip install tensorflow

gpu-version

pip install tensorflow-gpu

Preprocessing

1. Data

ChEBML 24 Database was used for SMILES data.

SMILES strings were padded with spaces to max_len(default=120) and strings larger than max_len were discarded. Remaining strings are labeled character by character(max_len labels in one string).

2. preprocess.py

Does the following steps:

1. Downloads chembl_24_1_chemreps.txt.gz
2. Preprocess SMILES strings
3. Saves processed data into numpy arrays.

Numpy arrays contains training data, testing data, dictionaries for character <-> label(integer) interchange.

Training

1. Model

Model consists of CNN encoder and CuDNNGRU decoder and defined in vae.py

2. train.py

Does the following steps:

1. Loads preprcessed data
2. trains with fit_generator using DataGenerator

Notebooks

Notebooks are here to help after training is done.

1. structure_variation.ipynb

This notebook helps to get variational structures when given a SMILES string.

2. visualize_latent_space.ipynb

This notebook helps visualizing learned latent space using a plot or tensorboard.

tensorboard visualization example:

3. find_top_k_mols_in_latent_space.ipynb

This notebook helps to get top_k similar molecules measured by euclidean distance in latent space.