Phase Model-Driven Deep Learning for Robust Phase Correction in High-Throughput NMR-Based Metabolomics

High-throughput NMR, a key metabolomics tool, enables efficient, noninvasive profiling of large biological samples. Automatic data processing ensures scalable, consistent high-throughput NMR. A key workflow step is phase correction, critical for obtaining pure absorption-mode spectra necessary for accurate quantitative analysis. This study proposes the Phase Model-Driven Residual Attention Network (PD-RAN), a robust phase correction method that combines deep neural networks with a physically informed model. By learning low-dimensional phase features grounded in physical principles from one-dimensional NMR spectra containing thousands of data points (high-dimensional data representation), PD-RAN delivers precise and reliable phase correction. Experimental results show consistent superiority over conventional methods across diverse metabolomics samples, including brain extracts, plasma, and urine. The method demonstrates remarkable efficiency, processing 1,000 spectra in just 20 ms, rendering it highly suitable for high-throughput NMR metabolomics applications. Ablation studies further validate the effectiveness of the phase model-driven component and its robustness to noise and baseline distortions.