This project explored generative methods for drug discovery using Variational Autoencoders. Our initial objective is to measure the ratio of syntactically valid molecules that the models can produce to determine which adjustments to the baseline model produce the largest improvements. We explore various training techniques, model improvements, and hard constraints imposed on the input format or latent space to achieve molecular generation that is more valid and chemically rich.
Our presentation can be found here: https://youtu.be/zZHGXTrgfrY
We discuss our results and ideas for further investigation in our paper, found here.
The inspiration and sources for this repository come from:
-
Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules
-
https://docs.microsoft.com/en-us/azure/machine-learning/quickstart-create-resources
-
https://github.com/wengong-jin/icml18-jtnn/tree/master/data/zinc
First we explored a dataset of 250K molecules to understand the dataset a little
better. None of us have a background in chemistry or biology so this was fairly
essential. The exploration can be reproduced by running Data_Exploration.ipynb
in a jupyter notebook environment (originally Google Colab).
Given the usage of Azure ML in these notebooks, the notebooks will create an image to run the notebook on a compute cluster. The notebooks will also automatically download the dataset required for training. The notebooks listed below will run in an Azure ML Instance in Azure creating a run of a experiment and log the results.
After exploring the data, we leveraged the deepchem library to train an instance of the AspuruGuzikAutoEncoder using our 250K molecule dataset. This work can be reproduced by running the notebook (inside an Azure ML Workspace):
Approach1_BaselineModel.ipynb
KL Annealing is used to prioritize reconstruction loss early in the training process and then later incorporate more of the Kullback–Leibler (KL) loss. The baseline model provided by deepchem and we wanted to test the impact of turning it off. Those results can be reproduced by running
Approach2_DisablingCostAnnealing.ipynb
Teacher forcing is a technique to help the decoder along when doing training so
that the decoding of the characters later in a sequence get a better chance to
train. For example, if I'm trying to reconstruct
C[NH+](C/C=C/c1ccco1)CCC(F)(F)F and my decoder messes up the first couple of
characters, the teacher-forcing model will propagate the correct character at
that iteration of the decoder so that the following characters have a better
chance of also being correct. This allows for the whole sequence being
trained/learned at the same time instead of depending on the initial decoding to
be correct in order to get the remaining characters. We attempted to add this
technique to the deepchem library, but ultimately our efforts were unsuccessful.
Ultimately we were able to use a VAE model with teacher forcing enabled using
the moses library. That implementation can be reproduced by running
moses_vae.ipynb
The results can be reproduced by running:
Approach3_Moses.ipynb
Up to this point, our models has used the SMILES representation to represent a molecule as a string. We were able to experiment with a different string representation of a molecule called SELFIES. Those results can be reproduced by running
Approach4_Selfies.ipynb.
Hopefully these notebooks are useful to you in your drug discovery journey!