Virtual Homonuclear Decoupling in Direct Detection NMR Experiments using Deep Neural Networks

Nuclear magnetic resonance (NMR) experiments are frequently complicated by the presence of homonuclear scalar couplings. For the growing body of biomolecular 13C-detected NMR methods, one-bond 13C-13C couplings significantly reduce sensitivity and resolution. The solution to this problem has typically been to perform virtual decoupling by recording multiple spectra and taking linear combinations. Here, we propose an alternative method of virtual decoupling using deep neural networks, which only requires a single spectrum and gives a significant boost in resolution while reducing the effective phase cycles of the experiments by at least a factor of two. We successfully apply 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. The deep neural network approach effectively decouples spectra with a high degree of flexibility, including in cases where existing methods fail, facilitates the use of simpler pulse sequences, and yields spectra with comparable quality to traditional virtual decoupling schemes in half the time or less.