Deep Learning Assisted Proton Pure Shift NMR Spectroscopy

Nuclear magnetic resonance spectroscopy (NMR) plays a key role for the analysis of a plethora of molecules, including natural products and drug-like organic molecules. For such cases 1H NMR spectra have proven imperative because of their high sensitivity. However, these spectra are complicated by complex multiplet patterns that, although important for the analysis, lead to substantial overlap. Here we show a deep-learning approach, which transforms spin-echo modulated 1H NMR spectra into highly sensitive and high-resolution singlet NMR spectra, that is, virtual homonuclear decoupled pure shift spectra. The approach was evaluated on experimental NMR spectra of complex organic compounds, where it outperforms current methods. The method also predicts uncertainties of the transformation and therefore allows for quantifications. We believe that our approach will provide significant advantages when characterizing low sensitivity samples, where no signals are observed in traditional pure-shift spectra and strong overlaps are hampering analysis from conventional spectra.