Abstract
NMR spectroscopy is ubiquitous in many areas of science, including chemistry, material science, biophysics, and structural biology. Heteronuclear two-dimensional (2D) NMR experiments form the basis of several NMR-based investigations, especially those involving biomolecules like proteins. However, in solution, as the molecule of interest becomes larger, transverse relaxation times of the spins of interest become shorter. This makes it difficult to record 2D (1)H-(13)C or (1)H-(15)N correlation maps of large protein molecules using standard Fourier transform-based experiments as they contain transfer delays that are long compared to the short relaxation times. Herein, we explore the possibility of leveraging deep neural networks (DNNs) to obtain 2D correlation maps of proteins while eliminating transfer delays from experiments. In this study, we show that 2D methyl (1)H-(13)C correlation maps can be obtained from experiments containing only a single (1)H excitation pulse followed by off-resonance (13)C continuous-wave decoupling. A DNN was successfully trained to reconstruct the 2D (1)H-(13)C correlation map from datasets recorded with two (13)C decoupling fields using protein samples enriched with isoleucine, leucine, and valine (ILV) (13)CHD(2) methyl-groups in a (2)H background. The efficacy of this strategy is demonstrated by using the DNN to reconstruct methyl (1)H-(13)C correlation maps of the ~8 kDa FF domain from human HYPA/FBP11, the ~18 kDa T4 phage lysozyme, the ~360 kDa α(7)α(7) particle from the Thermoplasma acidophilum proteasome and a ~530 kDa ancestrally reconstructed L(8)S(8) Rubisco complex. This study illustrates the potential for improving the sensitivity and resolution of NMR spectra by designing experiments tailored for use with DNNs.