Automated Determination of the Molecular Substructure from Nuclear Magnetic Resonance Spectra Using Neural Networks.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Nuclear Magnetic Resonance (NMR) spectroscopy is crucial for molecular structure determination.
• The increasing volume of NMR data, due to automated analysis, creates a bottleneck in structural characterization.
• Manual interpretation of NMR spectra is time-consuming and prone to errors.

Purpose of the Study:

• To investigate the potential of machine learning (ML) methods, specifically neural networks, to automate NMR spectral interpretation.
• To correlate NMR spectral features with molecular substructures.
• To evaluate different neural network architectures and molecular representations for NMR data analysis.

Main Methods:

• Three neural network architectures were explored: Multilayer Perceptron (MLP) + Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and MLP + Recurrent Neural Network (RNN).
• Molecular representations included functional groups and a novel neighbor-based method.
• Models were trained on experimental ¹³C and ¹H NMR spectra, with and without experimental metadata (field strength, temperature, solvent).

Main Results:

• The MLP + LSTM model achieved 88% accuracy on ¹³C NMR spectra when incorporating experimental metadata, a significant improvement over 77% without metadata.
• The CNN model demonstrated slightly lower accuracy (86%) but operated three times faster than the MLP + LSTM model.
• The CNN model was identified as the most practical choice considering accuracy, computational time, and cost.

Conclusions:

• Machine learning, particularly neural networks, can effectively automate the interpretation of NMR spectra.
• Incorporating experimental metadata enhances the accuracy of ML models for NMR analysis.
• The CNN model provides an efficient and accurate solution for the bottleneck in NMR-based structural characterization.