Graph NMR

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

Layout

• data: The TF Records
• graphnmr: The installable module containing model code
  • __init__.py: The module init file
  • data.py: functions for processing and loading data
  • gcnmodel.py: The main model code
  • validation.py: Validation code for checking correctness of data
• parse: Scripts for converting raw data into TF records for training
• scripts: Scripts for running model
  • plot_gcn_comparison.py: Script for plotting hyperparameters choices on grid
  • train_hypers.py: Script for running with variety of hyperparameters
  • train_structural.py: Main training script

Data

The raw data is not in this repo due to the huge number of files. The processed records contain the parsed data.

Preequisites

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

Install

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.