DM-NMR: A universal model for non-uniform sampling reconstruction of NMR spectra

Multidimensional nuclear magnetic resonance (NMR) is a widely utilized technique for protein studies of structures, dynamics, and interactions, providing crucial atomic-level insight. Non-uniform sampling (NUS) technology is an effective approach to greatly reduce sampling time, and advanced reconstruction algorithms is needed to obtain high-quality NMR spectra. Traditional algorithms require manual parameter tuning, often leading to suboptimal results. The emergence of deep learning (DL) provides a new approach for reconstruction. However, when the sampling scheme or sampling rate employed in experimental settings differs substantially from those represented in the training data, existing DL models must be retrained to preserve performance, which severely limits their general applicability. Here, a diffusion-based model for NMR non-uniform sampling reconstruction, referred to as DM-NMR, is proposed. By learning the characteristics of fully sampled data in the frequency domain and incorporating time domain low-rank constraints during reconstruction, compared to current state-of-art both traditional and DL methods, DM-NMR achieves superior reconstruction performance across various sampling rates and sampling schemes.