A pilot study for fragment identification using 2D NMR and deep learning.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Spectroscopy

Background:

• Nuclear Magnetic Resonance (NMR) spectroscopy is crucial for chemical structure elucidation.
• Identifying substructures in complex mixtures using 2D NMR spectra presents significant challenges.
• Current methods may require extensive manual interpretation or specialized databases.

Purpose of the Study:

• To develop and validate an image-based convolutional neural network (CNN) for automated substructure identification in 2D NMR spectra.
• To assess the CNN's performance on both pure compounds and mixtures.
• To evaluate the utility of Heteronuclear Single Quantum Coherence (HSQC) and Heteronuclear Multiple Bond Correlation (HMBC) spectra, individually and combined.

Main Methods:

• A bespoke image-based convolutional neural network (CNN) application was developed.
• The CNN was trained and tested using 2D NMR data, specifically HSQC and HMBC spectra.
• The method was validated using pure compounds and subsequently applied to mixtures.

Main Results:

• The CNN reliably detected substructures in pure compounds with a simple network architecture.
• The application demonstrated successful substructure identification in mixtures when trained solely on pure compound data.
• HMBC spectra, and the combination of HMBC and HSQC spectra, yielded superior results compared to HSQC alone in this pilot study.

Conclusions:

• The developed CNN offers a promising proof-of-concept for automated substructure identification in 2D NMR spectra of mixtures.
• The AI-driven approach can simplify and accelerate the analysis of complex chemical samples.
• Further development and validation on diverse datasets are warranted to broaden its applicability.