NMRCID: a deep learning framework for automated metabolite profiling from 1D NMR spectra

Nuclear Magnetic Resonance (NMR) spectroscopy plays a central role in metabolomics by enabling the profiling of small molecules. Because metabolites exhibit unique spectral fingerprints, many compounds can be detected in complex mixtures, yet manual interpretation of these spectra remains a challenge. To address this, we developed NMRCID, a deep learning software package for automated analysis of 1D NMR spectra. The system employs a convolutional neural network trained to identify up to 30 metabolites simultaneously, achieving over 95% accuracy on both synthetic and experimental spectra. Large-scale training datasets were constructed through data augmentation of pure compound spectra, incorporating variations in chemical shift and spectral noise. The augmentation process also applies tailored scaling ranges to compound groups to capture intensity variations. In addition, NMRCID supports training multiple models with distinct datasets and combines their outputs through quorum-based prediction, enhancing robustness and reliability. Moreover, the software reports individual accuracy metrics for each metabolite, ensuring interpretable results. These results highlight the potential of NMRCID to reduce manual effort and improve consistency in NMR-based metabolomics.