NMR spectroscopy provides atomic-level information about molecular structure, dynamics, and interactions. In this presentation, we will present new developments in spectra processing using artificial neural networks (NN) and deep learning.
We use an artificial neural network based on the WaveNet architecture (WNN), which is designed to grasp specific patterns in the NMR spectra. WNN produces high quality and robust reconstruction of the non-uniformly sampled (NUS) spectra by resolving specific point spread function (PSF) patterns produced by each spectral peak. When trained to perform virtual homo-decoupling, WNN outperforms traditional methods as demonstrated for methyl 1H-13C HMQC spectrum of 44 kDa protein MALT1. We show also that the artificial neural networks allow solving new problem that have not been addressed so far by the traditional signal processing algorithms. As an example, we show that a properly trained WNN can reconstruct high quality spectra using only Echo (or Anti-Echo) part of the NMR signal.