An Efficient Approach for Obtaining Small and Macro-molecular 1H NMR Spectra Based on Neural Network

Metabolomics plays a vital role in comprehending cellular and organismal metabolic processes. In NMR-based metabolomics studies, specific NMR pulse sequences such as the standard 1D nuclear Overhauser effect spectroscopy (NOESY), 1D Carr-Purcell-Meiboom-Gill (CPMG), and 1D diffusion-edited sequences are commonly utilized to detect distinctive NMR characteristics of small molecule and macromolecule metabolites in plasma or serum samples. However, conducting NMR experiments on multiple samples in metabolomics can be time-consuming. This study introduces the Spectrum-Edited Neural Network (SENNet) for efficient and accurate separation of spectral signals from both macromolecules and small molecules in 1H NMR spectra. The proposed model provides an end-to-end mapping of the entire metabolome NMR spectrum to the macromolecular and small molecule NMR spectra. To validate and optimize the model's hyperparameters, we employed a total of 113 serum samples. Furthermore, the SENNet method was applied to post-process 1D NOESY-presat spectra obtained from 120 plasma samples and 463 serum samples, which were then compared with the corresponding 1D CPMG spectra and 1D diffusion-edited spectra. Our results demonstrate the effective extraction of small molecule signals using the proposed method, as confirmed by comparison with experimental spectra. Principal component analysis (PCA) performed on the macromolecule and small molecule signals reveals comparable statistical information to analyses conducted using experimental data, indicating the efficiency of the SENNet method for signal extraction. This high-throughput NMR post-processing method holds substantial potential for metabolomics research. Additionally, the SENNet method serves as a valuable reference for separating signals from both macro and small molecules in NMR samples.