Electronic Theses and Dissertations

Date

2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Biomedical Engineering

Committee Chair

Aaryani Tipirneni-Sajja

Committee Member

Amy de Jongh Curry

Committee Member

Deepak Venugopal

Committee Member

Marie van der Merwe

Committee Member

Melissa Puppa

Abstract

NMR-based metabolomics is an increasingly important tool for understanding organisms and biological systems; however, data processing is an often-tedious process and access to high-field NMR spectrometers is limited. Conventional data processing for metabolite profiling relies on slow, user-dependent analysis, yet modern artificial intelligence techniques have scantly been explored in quantitative NMR applications despite their industry transforming effects in many major data processing applications. This work aims to develop neural network approaches to fully automate the identification and quantification of metabolites from NMR spectra while increasing speed and throughput compared to conventional analysis. Our initial investigation implements neural networks for the quantification of lipids in NMR spectra acquired as part of a dietary metabolomics study. The black box nature of a neural network approach can be a hindrance to widespread acceptance; however, this work seeks to increase confidence in neural network-based metabolite profiling by presenting methods in both explainable artificial intelligence techniques and uncertainty quantification. Access to instrumentation is a major limitation in NMR spectroscopy, but improvements in low-field, benchtop instrumentation over the past decade have made strides towards alleviating this burden. However, the inherently lower sensitivity and greater peak dispersion of low-field NMR increase spectral complexity and complicate data analysis. This study explores novel neural network approaches for optimizing benchtop NMR spectroscopy for metabolite profiling by both directly quantifying metabolites in low-field spectra and by first converting spectra to high-field prior to neural network quantification. Finally, this work investigates ways to enhance the performance of neural networks for NMR metabolite profiling in complex spectra by implementing approaches like transformer encoder networks for quantification, exploratory data generation techniques, and hyperparameter optimization. Overall, this novel research implements budding technologies in artificial intelligence to advance the potential application and accessibility of NMR-based metabolomics.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Open access.

Download

Share

COinS