Performance validation of neural network based (13)c NMR prediction using a publicly available data source

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• 13C NMR spectroscopy is crucial for chemical structure elucidation.
• Accurate prediction of 13C NMR chemical shifts aids in spectral assignment and structure determination.
• Neural network algorithms offer a promising approach for chemical shift prediction.

Purpose of the Study:

• To validate the performance of a neural network-based 13C NMR prediction algorithm.
• To assess prediction accuracy using a large, open-source chemical shift database (NMRShiftDB).
• To compare algorithm performance with and without overlap between training and test data.

Main Methods:

• Validation of a neural network algorithm using NMRShiftDB (ca. 214,000 chemical shifts).
• Analysis of performance on two subsets: 'included shift set' (ca. 121,000 shifts) and 'excluded shift set' (ca. 93,000 shifts).
• Comparison of results with Robien's CNMR Neural Network Predictor.

Main Results:

• The algorithm achieved a mean error of 1.59 ppm across the entire NMRShiftDB.
• Mean deviations were 1.47 ppm for the 'included shift set' and 1.74 ppm for the 'excluded shift set'.
• Performance was comparable to other reported algorithms.

Conclusions:

• The neural network algorithm demonstrates reliable performance for 13C NMR chemical shift prediction.
• The validation using NMRShiftDB provides a robust assessment of the algorithm's accuracy.
• This tool can aid researchers in spectral analysis and structure elucidation.