Deep Learning Assisted Proton Pure Shift NMR Spectroscopy

Veera Mohana Rao Kakita,D. Flemming Hansen

DOI: https://doi.org/10.26434/chemrxiv-2024-ln3pk

2024-10-24

Abstract: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.

Chemistry

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the overlap issue in nuclear magnetic resonance spectroscopy (NMR) analysis, especially in proton (1H) NMR spectra, due to complex multiplet patterns. Although these multiplet patterns are very important for analysis, they can lead to significant signal overlap, thus affecting the resolution and sensitivity of the spectra. Traditional methods such as Pure Shift NMR techniques can improve the resolution to a certain extent, but usually at the expense of the sensitivity of the spectra.

To solve this problem, the paper proposes a deep - learning - based method, which can convert spin - echo - modulated 1H NMR spectra into high - sensitivity and high - resolution singlet NMR spectra, that is, virtual homonuclear - decoupled pure - shift spectra. This method not only improves the resolution of the spectra but also maintains high sensitivity and can predict the uncertainty of the conversion, thus supporting quantitative analysis. In addition, this method shows significant advantages in dealing with low - sensitivity samples. These samples cannot detect signals in traditional pure - shift spectra, and strong overlap hinders the analysis of conventional spectra.