/molecular-graph-descriptors

Codebase to accompany the paper A Look Inside the Black Box: Using Graph-Theoretical Descriptors to Interpret a Continuous-Filter Convolutional Neural Network (CF-CNN) trained on the Global and Local Minimum Energy Structures of Neutral Water Clusters.

Primary LanguagePython

Molecular Graph Descriptors

This codebase accompanies the paper A Look Inside the Black Box: Using Graph-Theoretical Descriptors to Interpret a Continuous-Filter Convolutional Neural Network (CF-CNN) trained on the Global and Local Minimum Energy Structures of Neutral Water Clusters.

Molecular graphs and structures of over 5 million water clusters can be found here.

Contents

data

.xyz files of the lowest-energy structure of each cluster size N=3-30 and a set of 15 clusters and their predited potential energy.

The formatted database for the 500k training and test sets used to train the published model can be obtained here: https://drive.google.com/file/d/1ZQNOhJnz0k_UWxc-CIIkYwE2d5o230Ad/view?usp=sharing

graphdescriptors

Code for generating graphs from molecular structures and computing various descriptors from .xyz files output from SchNetPack. The script get_descriptors.py collects a set of descriptors for all molecules in the .xyz file and outputs a csv file.

python get_descriptors.py --data_path ../data/test_predictions.xyz --output test_df.csv --min_dir ../data/

schnetpack

Code amended from SchNetPack to train and test the CF-CNN. See schnetpack/README.md for use.

Requirements

  • python 3
  • pytorch (>= 0.4.1)
  • h5py
  • ASE
  • networkx
  • pandas
  • numpy

References

Please cite our paper if you find the code and datasets useful.

  • Jenna A. Bilbrey, Joseph P. Heindel, Malachi Schram, Pradipta Bandyopadhyay, Sotiris S. Xantheas, and Sutanay Choudhury. "A look inside the black box: Using graph-theoretical descriptors to interpret a Continuous-Filter Convolutional Neural Network (CF-CNN) trained on the global and local minimum energy structures of neutral water clusters" J. Chem. Phys. 153, 024302 (2020).

BibTex

@article{Bilbrey2020Descriptors,
author = {Bilbrey,Jenna A.  and Heindel,Joseph P.  and Schram,Malachi  and Bandyopadhyay,Pradipta  and Xantheas,Sotiris S.  and Choudhury,Sutanay },
title = {A look inside the black box: Using graph-theoretical descriptors to interpret a Continuous-Filter Convolutional Neural Network (CF-CNN) trained on the global and local minimum energy structures of neutral water clusters},
journal = {The Journal of Chemical Physics},
volume = {153},
number = {2},
pages = {024302},
year = {2020},
doi = {10.1063/5.0009933}
}