Introduction:

 

We have recently shown that sparsely sampled NMR spectra can be reconstructed using Deep Neural Networks: D. Flemming Hansen "Using Deep Neural Networks to Reconstruct Non-uniformly Sampled NMR Spectra" Journal of Biomolecular NMR (2019).

 

This scripts provided below will allow you to reconstruct sparsely sampled 2D spectra. The scripts have been tested under

1. Linux Ubuntu 18.04 using the python 2.7.15rc
2. tensorflow 1.14.0 (GPU)
3. keras 2.2.4
4. numpy 1.16.4
5. nmrglue 0.6

 

Thus, if not already installed, you will need to install the python libraries: tensorflow, keras, numpy, and nmrglue.

 

Keep in mind that the neural network can only reconstruct spectra that have been recorded with the sampling schedules that it has been trained on. Currently two sampling schedules are provided (Also provided in the tar file below):

1. Poisson-Gap Sampling: 32 complex points sampled out of 256
2. Random Sampling: 32 comple points sampled out of 256

 

The parameters used to train the network for the provided examples are tailored towards 15N-1H HSQC spectra. Spectra must be recorded with the first point of the FID being recorded at time=0 s, such that phases in the indirect dimension are p0 ~0 and p1 ~0.

 

Further instructions are provided in the README file within the tar file as well as comments within the scripts.

 

Download

 

DeepNUS.tar.gz (245 MB)