Deep Learning and Artificial Intelligence for NMR data Analysis

A method termed as Direct Detection in Solution State NMR of biomolecules is not only used for the detection of 1H but has also become popular for the detection of ¹³C and ¹⁵N due to the absence of solvent signals, higher chemical shift dispersion of signals, and favorable relaxation properties. This has provided opportunities to understand the structure and behavior of biomolecules. However, the direction detection of uniformly labeled ¹³C suffers the loss in sensitivity and resolution due to the presence of large one bone ¹³C-¹³C scalar coupling. A typical solution to this problem is the use of virtual decoupling which involves recording multiple spectra and taking linear combinations. Here researchers from the University College London give an alternative method of virtual decoupling using deep neural networks which requires the acquisition of a single spectrum and gives a significant enhancement in resolution while reducing the minimum effective phase cycles of the experiments by at least a factor of 2. The work funded by UKRI, BBSRC and published in the Journal of the American Chemical Society shows how the authors have successfully applied this methodology to virtually decouple in-phase CON (¹³CO–¹⁵N) protein NMR spectra, ¹³C–¹³C correlation spectra of protein side chains, and ¹³Cα-detected protein ¹³Cα–¹³CO spectra where two large homonuclear couplings are present.