Towards Optimizing Neural Network-Based Quantification for NMR Metabolomics

Background: Quantification of metabolites from nuclear magnetic resonance (NMR) spectra in an accurate, high-throughput manner requires effective data processing tools. Neural networks are relatively underexplored in quantitative NMR metabolomics despite impressive speed and throughput compared to more conventional peak-fitting metabolomics software. Methods: This work investigates practices for dataset and model development in the task of metabolite quantification directly from simulated NMR spectra for three neural network models: the multi-layered perceptron, the convolutional neural network, and the transformer. Model architectures, training parameters, and training datasets are optimized before comparing each model on simulated 400-MHz 1H-NMR spectra of complex mixtures with 8, 44, or 86 metabolites to quantify in spectra ranging from simple to highly complex and overlapping peaks. The optimized models were further validated on spectra at 100- and 800-MHz. Results: The transformer was the most effective network for NMR metabolite quantification, especially as the number of metabolites per spectra increased or target concentrations were low or had a large dynamic range. Further, the transformer was able to accurately quantify metabolites in simulated spectra from 100-MHz up to 800-MHz. Conclusions: The methods developed in this work reveal that transformers have the potential to accurately perform fully automated metabolite quantification in real-time and, with further development with experimental data, could be the basis for automated quantitative NMR metabolomics software.