Resolution Enhancement of Metabolomic J-Res NMR Spectra Using Deep Learning.

Area of Science:

• Metabolomics
• Nuclear Magnetic Resonance (NMR) Spectroscopy
• Deep Learning

Background:

• J-Resolved (J-Res) NMR spectroscopy is crucial for metabolomics but suffers from peak overlap in low-resolution (LR) experiments.
• High-resolution (HR) experiments are time-consuming, presenting a trade-off between speed and data quality.

Purpose of the Study:

• To develop a deep learning model for enhancing the resolution of 2D NMR J-Res spectra.
• To improve peak resolvability in metabolomic analyses using J-Res NMR data.

Main Methods:

• Introduction of J-RESRGAN, a generative adversarial network (GAN) adapted for NMR spectral super-resolution (SR).
• Training the model on simulated HR J-Res spectra and their LR counterparts generated via blurring and down-sampling.
• Incorporation of a novel symmetric loss function exploiting the vertical symmetry of J-Res spectra.

Main Results:

• J-RESRGAN demonstrated significant improvements in peak pair resolvability across various sample types.
• 100% of peak pairs showed enhanced resolution in simulated plasma data.
• High resolution enhancement was observed in experimental plasma (80.8-100%), urine (85.0-96.7%), milk (94.4-98.9%), and orange juice (82.6-91.7%).

Conclusions:

• Deep learning, specifically J-RESRGAN, effectively enhances NMR metabolomic data resolution.
• The model is versatile, independent of sample type, spectrometer, or field strength, and provides rapid analysis.
• J-RESRGAN advances precision in NMR-based metabolomics by elucidating overlapping peaks.