Antibiotic Discovery Using Graph Deep Learning

Goal

Getting an understanding of drug discovery using machine learning and getting experience using Chemprop. Chemprop is a tool which allows for modeling of molecules using message passing graph neural networks.

This is done as the final project in the course ID2223 Scalable Machine Learning and Deep Learning at KTH.

Data

The datasets used in this project are available in the data folder. We used article_training.csv for training our model and assessing it. And then used the obtained model in order to predict novel antibiotic candidates using the coco_test.csv dataset.

Our training dataset comes from the article which we used as inspiration for this project: A Deep Learning Approach to Antibiotic Discovery.

Our dataset used for predictions consists of SMILES strings collected from COCONUT, a database of natural compounds, by filtering for compounds with 0-5 carbons. Our conversion script sdf_to_csv.py is then used in order to convert the COCONUT data from SDF to CSV.

Training and Prediction

At first, we tried to make a notebook in Google Colab, then DeepNote and Databricks:, but we faced severe issues while trying to install and/or use Miniconda on them. We should have tried to make a minimum viable product first, locally. In the end, we decided to do the project locally, using the terminal and Jupyter Notebooks.

Report

The project report, along with a list of our antibiotic candidates, is available in PDF format in this repository.

Reproducibility

The commands that we ran in order to obtain our results can be found in the commands.md file.

Environment Setup

Install Miniconda and activate conda:

Please have a look at these pages in case you don't have a working conda installation already.

https://docs.conda.io/en/latest/miniconda.html

https://conda.io/projects/conda/en/latest/user-guide/install/index.html

Install packages

Make sure to also have a look at the most recent Chemprop installation instructions, just in case the instructions here have become out-dated. We are going to install it from PyPi.

Create a new environment for chemprop and activate it.

conda create -n chemprop python=3.8
conda activate chemprop

Make sure that the current pip is from the current environment.

which pip

Install Chemprop and other necessary packages.

conda install -c conda-forge rdkit
pip install git+https://github.com/bp-kelley/descriptastorus  
pip install chemprop

You should now be able to use Chemprop's CLI, with commands starting with chemprop_*. (need to reboot the terminal? or will they be available right away?)

From here, you should be able to start a Jupyter notebook session, and start using Chemprop. In our case, however, we've had to also install jupyter notebook within the current conda environment, so that the chemprop python module would get discovered inside the notebook.

For solving the problem with import chemprop within Jupyter notebooks, the following steps will help.

conda install ipykernel  notebook
conda install -c conda-forge ipywidgets
jupyter notebook .

Test your installation by entering import chemprop.

Chemprop will show some warnings the first time it gets imported, but we were unable to find any solutions to them, and they seemed to be okay to ignore.

Closing and deactivating the environment

jupyter notebook list
jupyter notebook stop <port number>
conda deactivate