High-Quality NMR Pure Shift Spectra Empowered by Self-Attention Deep Learning

Pure shift spectra improve the resolution of NMR spectra, aiding in the structural analysis of substances. To obtain high-quality pure shift spectra, a network called Spin Echo to Pure Shift Network (SE2PSNet) is proposed to process spin echo spectra combined with chemical shift information. SE2PSNet consists of three modules: the spectral information association module, the self-attention mechanism module, and the pure shift spectrum recovery module, which are used for encoding, dynamic adjustment of feature weights, and decoding, respectively. The spin echo spectra help the network capture the relationships between signal and phase variations, while the chemical shift spectrum enhances the network's attention to features near the signals. Combining these two types of information as input will facilitate the network in obtaining a high quality pure shift spectrum with accurate quantitative information. The newly designed network can process spin echo spectra of several typical samples, effectively resolve the issue of signal overlap without introducing artifacts. Additionally, it is capable of processing spectra with low S/N (signal to noise ratio), achieving sensitivity comparable to conventional one-dimensional proton spectra and it can provide accurate quantitative information, making this approach widely applicable in chemical analysis.