Unlocking High-Fidelity Chemical NMR Spectral Information from Nonuniformly Sampled Experiments with Deep Learning

Accelerating multidimensional NMR acquisition underpins many advances in chemical and biomolecular research by enabling timely chemical insight, high-throughput spectral analysis, and access to transient or chemically evolving systems. Nonuniform sampling (NUS) offers a powerful route to accelerating multidimensional NMR experiments, but poses substantial challenges for accurate spectral reconstruction, especially in weak-peak regions. To overcome these limitations, we present a physics-guided deep learning framework, Consistency-guided Long-range Enhanced Attention for Reconstruction (CLEAR), which integrates convolutional layers with Transformer-based multi-head self-attention in a cascaded refinement architecture, enabling simultaneous modeling of local and global spectral correlations while enforcing strict data consistency with acquired samples. Comprehensive evaluations across multiple biomolecular NMR experiments demonstrate that CLEAR consistently outperforms state-of-the-art reconstruction methods, reducing reconstruction errors (RLNE) by approximately 16-25% while exhibiting overall superior or competitive performance across multiple quantitative metrics, including weak-peak preservation, under severe nonuniform sampling conditions (down to 5%). These results establish CLEAR as a robust and generalizable framework for high-fidelity NUS NMR spectral reconstruction.