A graph convolutional neural network for predicting NMR chemical shifts in molecules. This is the implementation for "Predicting Chemical Shifts with Graph Neural Networks" Z Yang, M Chakraborty, AD White 10.1101/2020.08.26.267971.
This code requires all molecules to be pre-processed into 256 atom fragments. Please use the updated model for general usage
data
: The TF Recordsgraphnmr
: The installable module containing model code
__init__.py
: The module init filedata.py
: functions for processing and loading datagcnmodel.py
: The main model codevalidation.py
: Validation code for checking correctness of dataparse
: Scripts for converting raw data into TF records for trainingscripts
: Scripts for running model
plot_gcn_comparison.py
: Script for plotting hyperparameters choices on gridtrain_hypers.py
: Script for running with variety of hyperparameterstrain_structural.py
: Main training scriptThe raw data is not in this repo due to the huge number of files. The processed records contain the parsed data.
numpy, matplotlib, tensorflow pre 2.0, graphviz, networkx, tqdm, gsd (conda-forge). If you want to do the data processing stuff, use the yml file included. Note
To run the scripts, you'll need to install the model code. Use pip install -e .
It will not attempt to install tensorflow, since
this is a system dependent task.