Towards Optimizing Neural Network-Based Quantification for NMR Metabolomics.

Area of Science:

• Computational chemistry
• Biochemistry
• Machine learning

Background:

• Accurate, high-throughput quantification of metabolites from NMR spectra is crucial for metabolomics.
• Neural networks are underutilized in quantitative NMR metabolomics, despite their potential for speed and throughput.
• Conventional peak-fitting software can be slow and less effective for complex spectra.

Purpose of the Study:

• To investigate dataset and model development for metabolite quantification using neural networks directly from simulated NMR spectra.
• To compare the performance of multi-layered perceptron, convolutional neural network, and transformer models for NMR metabolomics.
• To optimize model architectures, training parameters, and datasets for accurate metabolite quantification.

Main Methods:

• Simulated 400-MHz 1H-NMR spectra of mixtures with 8, 44, or 86 metabolites were used.
• Three neural network models (MLP, CNN, Transformer) were trained and optimized.
• Models were validated on spectra simulated at 100-MHz and 800-MHz.

Main Results:

• The transformer model demonstrated superior performance in NMR metabolite quantification.
• The transformer was most effective with increasing metabolite numbers, low concentrations, or large dynamic ranges.
• Accurate quantification was achieved across a wide spectral range (100-MHz to 800-MHz).

Conclusions:

• Transformers show significant potential for accurate, real-time, fully automated metabolite quantification in NMR metabolomics.
• Further development with experimental data could lead to advanced automated quantitative NMR metabolomics software.
• This approach offers a promising alternative to conventional methods for complex spectral analysis.