guanjq / geom

GEOM: Energy-annotated molecular conformations

GEOM: Energy-annotated molecular conformations

GEOM is a dataset with over 30 million molecular conformations annotated by energy and statistical weight for over 400,000 molecules. If you use this data, please cite

Axelrod, S. and Gomez-Bombarelli, R., "GEOM: Energy-annotated molecular conformations for property prediction and molecular generation". arXiv preprint arXiv:2006.05531 (2020). Link: https://arxiv.org/pdf/2006.05531.pdf

Bibtex format:

@article{axelrod2020geom,
  title={GEOM: Energy-annotated molecular conformations for property prediction and molecular generation},
  author={Axelrod, Simon and Gomez-Bombarelli, Rafael},
  journal={arXiv preprint arXiv:2006.05531},
  year={2020}
}

Contents

This repository contains Jupyter notebook tutorials showing how to load the data and perform analysis.

Usage

This code was tested with the following dependencies:

python==3.7.5,
mgspack==1.0.0,
ipykernel==5.3.0,
rdkit==2020.03.2.0,
matplotlib==3.2.1,
e3fp==1.2.1

You can create an anaconda environment to manage dependencies. You can learn more about managing anaconda environments by reading this page. First create an environment with Python, RDKit, and Matplotlib:

conda upgrade conda
conda create -n geom python==3.7.5 rdkit==2020.03.2.0 matplotlib==3.2.1 -c rdkit -c conda-forge 

Next activate the environment and install msgpack, ipykernel, and e3fp:

conda activate geom
pip install msgpack==1.0.0 ipykernel==5.3.0 e3fp==1.2.1

To ensure that the geom environment is accessible through Jupyter, add the the geom display name:

python -m ipykernel install --user --name geom --display-name "Python [conda env:geom"]

Accessing the data

Language-agnostic data

The datasets are available here. There are four datasets that can be loaded by any programming language. They are drugs_crude.msgpack.tar.gz, drugs_featurized.msgpack.tar.gz, qm9_crude.msgpack.tar.gz, qm9_featurized.msgpack.tar.gz. The first tutorial gives instructions for extracting the files and loading their content using MessagePack. MessagePack is a binary serialization format that allows you to exchange information among different languages, like JSON, but it is faster and more compact.

Python-specific data

The featurized files contain bond and atom features as lists of dictionaries and are quite large. If you are using Python, it is far more convenient to use the folder rdkit_folder.tar.gz. This folder contains files in which the conformer coordinates are replaced by RDKit mol objects. These objects contain both the coordinates and all the connectivity information contained in the featurized files, but use far less disk space. Moreover, with RDKit you can generate your own 2D and 3D descriptors in a very straightforward way. The RDKit tutorial shows how to load the RDKit files, visualize conformers, generate additional descriptors, and export to PDB. If you are not familiar with RDKit, you can get started at the RDKit home page.

Finally, you may want to analyze only a few molecules based on certain properties (e.g., load 200 molecules that bind SARS-CoV 3CL protease, and 1000 that do not). However, you may not want to first load all molecules and then filter by properties. In this case you can load the files rdkit_folder/{drugs,qm9}_summary.json, which contain all the summary statistics for each molecule, but exclude conformer information. You can use these lightweight files to decide which molecules to load, and then load their RDKit pickle files one-by-one. This, too, is described in the RDKit tutorial.