Neural network-based NMR metabolomics software for automated lipid and aqueous metabolite quantification

Background

NMR spectroscopy has emerged as a powerful tool for metabolite profiling, widely used in biomedical research to quantify small molecules in biological samples. Metabolomics research often requires analysis of both lipid and aqueous fractions to gain comprehensive insights into cellular metabolism, disease biomarkers, and therapeutic monitoring. However, the complexity of one-dimensional 1H NMR spectra, with dozens of overlapping resonances and subtle chemical shifts, makes manual interpretation time-consuming and prone to variability. As demand grows for large-scale studies and real-time metabolic screening, there is a pressing need for automated, high-throughput tools that can rapidly deconvolute complex spectra and deliver accurate, reproducible concentration data across diverse sample types.

Existing NMR metabolomics workflows typically rely on manual peak fitting or semi-automated routines using separate software platforms for aqueous and lipid analyses. Manual deconvolution requires significant expertise to choose fitting parameters, identify overlapping signals, and normalize to internal standards, leading to operator-dependent bias and long processing times. Commercial solutions often support only limited metabolite libraries or focus on one sample class, forcing researchers to juggle multiple tools and protocols. This fragmentation hampers throughput, scalability, and consistency, particularly in studies demanding simultaneous quantification of both lipid and aqueous metabolites under standardized conditions.

Technology

Two multilayer perceptron neural networks—one trained on lipid extracts and the other on aqueous samples—automate the analysis of 1D proton NMR spectra acquired on a 400 MHz spectrometer. Users prepare lipid fractions in deuterated chloroform/methanol/D₂O at 0 °C (single-pulse sequence) and aqueous fractions in D₂O at pH 7.4 at room temperature (1D-NOESY sequence), then upload their spectra via a graphical interface. Reference spectra of individual standards are linearly combined into large synthetic training datasets for each network. During analysis, overlapping peaks are deconvoluted, signals are normalized to dimethyl sulfone (lipids) or TSPP (aqueous), and a comprehensive table of identified metabolites with absolute concentrations is generated, enabling fully automated, high-throughput quantification of both fractions in a single workflow.

This technology is distinguished by its unified handling of lipid and aqueous metabolomes within one platform, eliminating the need for separate tools. By employing MLP architectures instead of more complex convolutional networks, it achieves faster training and inference, while synthetic dataset generation ensures robust performance across diverse sample mixtures. Its end-to-end automation— from spectral deconvolution to concentration reporting—significantly reduces manual intervention and analysis time compared to existing solutions, delivering rapid, reproducible results for academic and industrial metabolomics applications.

Benefits

• Fully automated identification and quantification of lipid and aqueous metabolites in a single platform
• High-throughput processing enables rapid analysis of large sample sets
• Accurate deconvolution of overlapping NMR peaks using trained neural networks
• Consistent quantitative results via normalization to internal standards
• User-friendly graphical interface simplifies spectrum upload and analysis
• Efficient MLP-based models allow faster training and processing

Applications

• Drug development metabolite profiling
• Clinical biomarker quantification
• Food safety metabolomics testing
• Environmental contamination analysis
• Biotech fermentation monitoring

Additional info

A software platform uses two multilayer perceptron neural networks—one for lipid extracts (chloroform/methanol/D2O, 0 °C, single-pulse sequence) and one for aqueous samples (D2O, pH 7.4, 1D-NOESY)—to deconvolute overlapping peaks in 400 MHz 1D proton NMR spectra. Reference spectra are linearly combined to generate synthetic training sets. Users upload spectra, select a network, and receive fully automated, high-throughput metabolite concentration tables normalized to internal standards (dimethyl sulfone for lipids, TSPP for aqueous).