SAND

SAND performs automated decomposition and quantification of NMR spectra to enable accurate metabolite quantification in untargeted NMR metabolomics.

Key Features:

• Automated Time-Domain Modeling: Performs automatic time-domain modeling to quantify entire NMR spectra without manual intervention.
• Hybrid Optimization Techniques: Integrates Markov chain Monte Carlo (MCMC) with subsampling in both time and frequency domains to improve robustness and accuracy of spectral decomposition.
• Overfitting Mitigation: Randomly divides time-domain data into training and validation sets to reduce overfitting and improve generalization.
• High Accuracy and Correlation: Achieves a correlation coefficient of approximately 0.9 with ground truth across highly overlapped simulated datasets, two-compound mixtures, and urine samples spiked with a four-compound mixture.
• Automated Annotation via Correlation Networks: Constructs correlation networks from decomposed spectral peaks and recovers approximately 74% of peaks per compound within single clusters.
• Scalability and Flexibility: Relies on time-domain subsampling using random subsets of time-domain points, enabling extensions to higher-dimensional data and nonuniformly sampled datasets.

Scientific Applications:

• Untargeted metabolomics: Supports quantitative profiling of complex biological sample sets through automated spectral decomposition and quantification.
• Metabolite identification and characterization: Facilitates recovery and annotation of compound-specific peaks via correlation networks from decomposed spectra.
• Biomarker discovery and pathway analysis: Provides quantitative spectral measures suitable for metabolic pathway analysis and biomarker discovery in biological samples such as urine.

Methodology:

Automatic time-domain modeling and automated quantification of entire NMR spectra; hybrid optimization combining Markov chain Monte Carlo (MCMC) with subsampling in time and frequency domains; random time-domain subsampling using random subsets of points; random division of time-domain data into training and validation sets to mitigate overfitting; construction of correlation networks from decomposed spectral peaks for automated annotation.