A deep learning framework for multiplet splitting classification in ¹H NMR.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Artificial Intelligence in Chemistry

Background:

• One-dimensional proton Nuclear Magnetic Resonance (1D ¹H NMR) is a rapid technique for chemical compound analysis.
• Manual annotation of 1D ¹H NMR spectra is time-consuming and can lead to inconsistent interpretations.
• Automated spectral analysis is crucial for streamlining chemical characterization and ensuring community-wide consistency.

Purpose of the Study:

• To introduce MuSe Net, a novel supervised probabilistic deep learning framework for automated 1D ¹H NMR spectral annotation.
• To emulate the capabilities of expert spectroscopists in identifying and classifying spectral features.
• To improve the efficiency and reliability of chemical compound characterization using NMR data.

Main Methods:

• Development of MuSe Net, a deep learning framework utilizing supervised probabilistic learning.
• The model analyzes 1D ¹H NMR spectra to detect and classify multiplets based on splitting patterns (up to four coupling constants).
• Implementation of uncertainty quantification to provide confidence scores for classification reliability and anomaly detection.

Main Results:

• MuSe Net successfully detects and classifies multiplets in 1D ¹H NMR spectra.
• The framework demonstrates robustness in handling spectral anomalies and overlapping peaks.
• Evaluation on 48 experimental spectra showed accurate classification and reliable confidence scoring compared to expert annotations.

Conclusions:

• MuSe Net offers an effective automated solution for 1D ¹H NMR spectral annotation.
• The deep learning approach enhances the consistency and efficiency of chemical structure elucidation.
• Uncertainty quantification in MuSe Net aids in assessing the reliability of spectral interpretations and identifying challenging signals.