Clean PSYCHE Pure Shift Spectra Obtained by Deep Learning.

Proton (1H) nuclear magnetic resonance (NMR) is the most extensively utilized NMR technique due to its high sensitivity and information content. Nonetheless, its narrow chemical shift range and complex multiplet splitting often lead to severe spectral overlap, complicating spectral analysis. Pure shift methods address this challenge by converting weakly coupled multiplets into singlets, thereby significantly improving spectral resolution. Current pure shift methods struggle, however, with strongly coupled spin systems, yielding both main peaks that can be slightly displaced from the exact chemical shift and extra peaks, often at frequencies intermediate between the shifts of strongly coupled protons. These "strong coupling artifacts" can complicate spectral analysis. In this work, a multiscale residual network model is developed to postprocess PSYCHE pure shift spectra to remove strong coupling artifacts. Within the applicable ranges of spectral parameters, the model can correctly retain the desired pure shift signals and recover pure singlets. The method introduces no additional acquisition-related sensitivity loss by postprocessing PSYCHE spectra and exhibits superior performance compared to TSE-PSYCHE in suppressing strong coupling artifacts. Additionally, it preserves the parent PSYCHE method's direct proportionality between signal intensity and analyte concentration, enabling the monitoring of concentration changes for different analytes in a mixture. The new method offers a novel strategy for obtaining clean pure shift spectra and is expected to facilitate molecular structure elucidation and composition analysis in chemistry research.