Automated metabolic assignment: Semi-supervised learning in metabolic analysis employing two dimensional Nuclear...

Area of Science:

• Biomedical diagnostics
• Metabolomics
• Nuclear Magnetic Resonance (NMR) spectroscopy

Background:

• Metabolomics aids in understanding disease mechanisms by revealing metabolic reprogramming.
• 1D NMR spectral analysis presents challenges in metabolite assignment due to spectral resolution issues.
• 2D NMR techniques, like total correlation spectroscopy (TOCSY), offer improved spectral resolution for complex metabolic profiling.

Purpose of the Study:

• To develop an automated method for metabolite identification and assignment from 1H-1H TOCSY spectra.
• To apply this automated approach to real breast cancer tissue samples.
• To evaluate the performance of semi-supervised classifiers for NMR-based metabolic profiling.

Main Methods:

• Utilized customized and extended semi-supervised classifiers (KNFST, SVM, PC3, PC4) for metabolite assignment.
• Focused on vertical and horizontal frequencies from 1H-1H TOCSY spectra for assignment.
• Implemented semi-supervised classifiers to achieve a fully automatic procedure for TOCSY signal assignment and deconvolution.

Main Results:

• KNFST and SVM classifiers demonstrated high performance with good accuracy and low mislabeling rates, even with limited initial training data.
• Polynomial classifiers (PC3, PC4) exhibited lower accuracy and higher mislabeling rates, failing at very low training data percentages.
• Successfully deduced and assigned 27 metabolites from TOCSY spectra, consistent with 1D NMR analysis.

Conclusions:

• The developed automated approach using semi-supervised classifiers significantly advances NMR metabolic profiling.
• KNFST and SVM show promise for accurate and efficient metabolite identification in complex biological samples.
• This method provides a robust tool for analyzing metabolic alterations in diseases like breast cancer.