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

Abstract

One-dimensional ¹H Nuclear Magnetic Resonance (NMR) stands out as the quickest and simplest among various NMR experimental setups. Unfortunately, it suffers from lengthy annotation times and does not always have a clear and unique interpretation. From NMR discovery, efforts have been dedicated to introducing an automated approach to streamline the characterization of chemical compounds while ensuring consistency of the results across the scientific community. Nonetheless, this remains an ongoing challenge that has garnered renewed interest with the emergence of deep learning techniques. Here, we present MuSe Net, a novel supervised probabilistic deep learning framework that can emulate the tasks performed by an expert spectroscopist in annotating one-dimensional NMR spectra generated by small molecules. Considering only the spectrum, MuSe Net detects and classifies multiplets with up to four coupling constants for their splitting phenotype, providing a segmentation of the spectral range. We exploit uncertainty quantification to produce a confidence score to both assess classification reliability and to detect signals that do not fit into any other phenotype class. The results of the evaluation against 48 experimental ¹H NMR spectra of small molecules annotated by experts demonstrate that MuSe Net can deal with anomalies and unclear signals while correctly classifying multiplets and detecting overlapping peaks.