Graph neural network for predicting NMR chemical shifts

This library is the code and a pre-trained model to predict NMR chemical shifts from protein structures and organic molecules. It relies on the nmrdata package which includes embeddings and NMR parameters.

Install

Install using pip

pip install nmrgnn

Colab

To use this package without installing, use this colab

Command Line Usage

Available commands are

• nmrgnn eval-struct to predict chemical shifts of structure via MDAnalysis library as coordinate reader
• nmrgnn train to train a model
• nmrgnn hyper to tune hyperparameters
• nmrgnn eval-tfrecords to evaluate model on records in format from nmrdata package

Predict NMR Chemical Shfits

Note: This model is trained on models with no solvent, so remove that before use. For small molecules, the model was trained mostly on water solutions. You should only expect agreement in relative chemical shifts between atoms depending on your solvent and reference.

To predict NMR chemical shifts via the MDAnalysis library as a reader:

nmrgnn eval-struct [struct-file] [output-csv]

where struct-file could be a pdb file or equivalent. Example:

nmrgnn eval-struct 108M.pdb 108M-predicted.csv

For a trajectory, try

nmrgnn eval-struct 108M.pdb 108M.trr 108M-predicted.csv --stride 5

which computes shifts every 5 frames.

Warning about Peaks

If you receive a warning about peaks being poor, you likely have no hydrogens in your protein. You can add using online tools or use these commands to fix quickly by using OpenMM

conda install -y -c omnia openmm
pip install nmrdata[parse]@git+git://github.com/ur-whitelab/nmrgnn.git
nmrparse clean-pdb [your-pdb] [your-pdb]-H.pdb

Library Usage

Available functions are

• load_model to load the included pre-trained model or specify a path to a trained model
• universe2graph to convert an MDAnalysis universe into a tuple of atoms, neighbor list, edges, inverse_degree.
• check_peaks to estimate validity of predicted peaks

The example below predicts peaks and estimates (True/False) if the peaks are valid. Examples of why peaks are not valid are that the elements are not inlcuded in training data (e.g., oxygen shifts) or unusual chemistries or you forgot to remove solvent.

import MDAnalysis as md
import nmrgnn

model = nmrgnn.load_model()
u = md.Universe('108M.pdb')
g = nmrgnn.universe2graph(u)
peaks = model(g)
# check_peaks only uses first element of tuple (atom identities)
confident = nmrgnn.check_peaks(g[0], peaks)

You should not trust peaks coming from model without checking

Analyzing Trajectories

Here is an example for analzying a trajectory

import MDAnalysis as md
import nmrgnn

model = nmrgnn.load_model()

u = md.Universe(PATH_TO_FILES)
for ts in u.trajectory:
    x = nmrgnn.universe2graph(u)
    peaks = model(x)
    nmrgnn.check_peaks(x[0], peaks)    
    # do something with peaks
    ....

Citation

Please cite Predicting Chemical Shifts with Graph Neural Networks

@article{yang2021predicting,
  title={Predicting Chemical Shifts with Graph Neural Networks},
  author={Yang, Ziyue and Chakraborty, Maghesree and White, Andrew D},
  journal={Chemical Science},
  year={2021},
  publisher={Royal Society of Chemistry}
}

Model Performance

Here is the included model performance on proteins (P prefix) and organic molecules (Mol prefix). r is correlation coefficient and rmsd is root mean square deviation. These results vary from paper values because they are evaluated on whole proteins instead of 256 atom fragments.