EUVIMEDEUVIMED entdecken
Crossref

MuSe Net : a deep learning framework for trustworthy multiplet segmentation in 1D 1H NMR spectra

Giulia Fischetti, Nicolas Schmid, Simon Bruderer, Andreas Henrici, Heitmann, Bjoern, Scarso, Alessandro, Guido Caldarelli, DIRK WILHELM

Zürcher Hochschule für Angewandte Wissenschaften digital collection (Zurich University of Applied Sciences) · 02.07.2024

DOI: 10.21256/zhaw-31102 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Zürcher Hochschule für Angewandte Wissenschaften digital collection (Zurich University of Applied Sciences), erschienen am 02.07.2024. Verfasst von Giulia Fischetti und 9 weiteren Autor:innen.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Numerous research fields, including organic chemistry, metabolomics, drug discovery, battery technology, and environmental sciences, rely widely on one-dimensional proton NMR spectroscopy to investigate the structural properties of compounds, thanks to its simplicity and reduced acquisition times. However, spectra annotation can be time-consuming and prone to interpretation biases, due to artefacts and overlapping signals.

Aim

To address this issue, we exploited the potentiality of deep learning in pattern recognition and predictive uncertainty estimation methodologies to build a reliable tool to assist NMR practitioners in the annotation process.

Methods

We present MuSe Net (Multiplet Segmentation Network), a new supervised probabilistic deep learning framework designed to identify and categorize multiplets within phenotypes with up to four coupling constants based on the splitting pattern. Without any information on the compound, MuSe Net produces a segmentation of the spectral range, paired with a confidence score. The confidence score both assesses classification reliability and detects signals that do not fit in any other phenotype class: this is the case of overlapping multiplets. Our algorithm was trained on a wide collection of synthetic spectral segments, automatically produced and labelled to ensure consistency of the multiplet phenotypes represented.

Results

MuSe Net performance was evaluated against 48 experimental proton NMR spectra of small molecules annotated by experts. MuSe Net produced an accurate prediction in a large variety of experimental conditions, dealing effectively with solvent peaks, anomalies, and unclear signals while correctly classifying multiplets and detecting overlapping peaks. Interestingly, MuSe Net prediction and confidence score remained robust even at reduced signal-to-noise ratios.

Conclusions

In conclusion, MuSe Net holds the potential to significantly shorten the annotation process of one-dimensional proton NMR spectra, providing automatic reliable predictions of multiplet splitting patterns.

Schlagwörter

1H NMR spectrumdeep learningconfidence scoressignal-to-noise ratio robustness

Zitieren

APA

Fischetti, G., Schmid, N., Bruderer, S., Henrici, A., Heitmann, Bjoern, Scarso, Alessandro, Caldarelli, G., & WILHELM, D. (2024). MuSe Net : a deep learning framework for trustworthy multiplet segmentation in 1D 1H NMR spectra. Zürcher Hochschule für Angewandte Wissenschaften digital collection (Zurich University of Applied Sciences). https://doi.org/10.21256/zhaw-31102

Vancouver

Fischetti G, Schmid N, Bruderer S, Henrici A, Heitmann, Bjoern, et al. MuSe Net : a deep learning framework for trustworthy multiplet segmentation in 1D 1H NMR spectra. Zürcher Hochschule für Angewandte Wissenschaften digital collection (Zurich University of Applied Sciences). 2024. doi:10.21256/zhaw-31102

BibTeX
@article{Fischetti2024,
  author = {Fischetti, Giulia and Schmid, Nicolas and Bruderer, Simon and Henrici, Andreas and Heitmann and Bjoern and Scarso and Alessandro and Caldarelli, Guido and WILHELM, DIRK},
  title = {MuSe Net : a deep learning framework for trustworthy multiplet segmentation in 1D 1H NMR spectra},
  journal = {Zürcher Hochschule für Angewandte Wissenschaften digital collection (Zurich University of Applied Sciences)},
  year = {2024},
  doi = {10.21256/zhaw-31102}
}
Kostenlos starten →