Deep Learning-Based Method for Compound Identification in NMR Spectra of Mixtures

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful, unbiased tool for analyzing small molecule mixtures.
• Compound identification in NMR spectra is difficult due to chemical shift variations and overlapping peaks.

Purpose of the Study:

• To present a novel pseudo-Siamese convolutional neural network (pSCNN) method for accurate compound identification in NMR spectra mixtures.
• To address challenges of chemical shift variation and peak overlapping in NMR spectral analysis.

Main Methods:

• Implemented a data augmentation technique involving superposition of NMR spectra with random noise.
• Trained, validated, and tested a pSCNN model using augmented and experimental datasets.
• Utilized convolutional neural networks' translational invariance to handle chemical shift variations.

Main Results:

• The pSCNN model achieved high accuracy on augmented data (ACC=99.80%) and experimental datasets (flavor mixtures: ACC=97.62%, additional flavor mixture: ACC=91.67%).
• Demonstrated robustness to chemical shift variations inherent in NMR spectra analysis.
• Achieved excellent performance metrics including high True Positive Rate (TPR) and low False Positive Rate (FPR).

Conclusions:

• The pSCNN method provides an accurate and robust solution for compound identification in NMR spectroscopy mixtures.
• The approach effectively overcomes key challenges, making it a valuable tool for spectral analysis.
• pSCNN is an off-the-shelf method suitable for real-world NMR applications.