Breaking the denoising dilemma in NMR spectroscopy by blind source separation

Abstract

Nuclear magnetic resonance (NMR) spectroscopy serves as a fundamental analytical technique finding valuable applications across chemistry, biology, and materials science, yet its widespread utility is frequently constrained by its inherently low sensitivity and concomitant noise, making noise reduction a critical processing step. Existing denoising methods typically impose a difficult trade-off, namely, traditional iterative algorithms are interpretable and widely adoptable but time-consuming, while deep learning approaches achieve fast denoising and better performance at the cost of generalizability and interpretability. To break this dilemma, we propose a denoising protocol, namely CCA-NMR, to unlock clean signals in noisy NMR spectra. This protocol exploits the idea of blind source separation to separate noise and real signals based on autocorrelation coefficients, delivering good noise reduction, fast computational time, and decent universality. Owing to these advantages, it enables reliable sensitivity benefits and recovers genuine peaks obscured by noise, thus reducing the necessary experimental acquisition times and facilitating subsequent quantitative analysis. Its performance is evidenced through experimental validation across a diverse range of NMR platforms, including relaxation and diffusion Laplace NMR, multidimensional protein NMR, and time-resolved NMR during real-time electrocatalytic reactions. As a consequence, these findings substantially expand the practical applicability of NMR spectroscopy, and indicate its considerable potential for broad chemical and biomedical applications.